<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Korea Semiconductor Watch on Korea Invest Insights</title><link>https://koreainvestinsights.com/series/korea-semiconductor-watch/</link><description>Recent content in Korea Semiconductor Watch on Korea Invest Insights</description><generator>Hugo -- gohugo.io</generator><language>en</language><copyright>koreainvestinsights.com · @korea_invest_insights</copyright><lastBuildDate>Tue, 10 Nov 2026 09:15:24 +0900</lastBuildDate><atom:link href="https://koreainvestinsights.com/series/korea-semiconductor-watch/feed.xml" rel="self" type="application/rss+xml"/><item><title>Samsung's Share Price Pays for Less Than One Year of the Memory Boom</title><link>https://koreainvestinsights.com/post/samsung-price-pays-for-less-than-one-boom-year-2028-gives-242k-280k-2026-11-10/</link><pubDate>Tue, 10 Nov 2026 09:13:00 +0900</pubDate><guid>https://koreainvestinsights.com/post/samsung-price-pays-for-less-than-one-boom-year-2028-gives-242k-280k-2026-11-10/</guid><description>&lt;img src="https://koreainvestinsights.com/" alt="Featured image of post Samsung's Share Price Pays for Less Than One Year of the Memory Boom" /&gt;&lt;p&gt;Samsung Electronics closed at &lt;strong&gt;₩189,900&lt;/strong&gt; on November 9. We rebuilt our valuation model of Samsung from the ground up and asked it one question: how many years of boom profit does that price pay for?&lt;/p&gt;
&lt;p&gt;The answer is less than one. In our model, a share is worth &lt;strong&gt;₩155,600 to ₩162,000&lt;/strong&gt; if memory profit went back to a normal year in 2027. Counting 2027&amp;rsquo;s boom profit lifts the value to &lt;strong&gt;₩200,200 to ₩215,800&lt;/strong&gt;. Today&amp;rsquo;s price sits between the two. It pays for 52% to 77% of 2027&amp;rsquo;s excess profit and for nothing after that.&lt;/p&gt;
&lt;p&gt;Each further year you are willing to count adds &lt;strong&gt;₩40,000 to ₩64,000&lt;/strong&gt; a share. Count through 2028 and the value is &lt;strong&gt;₩242,400 to ₩279,900&lt;/strong&gt;, 28% to 47% above the close.&lt;/p&gt;
&lt;p&gt;On November 8 we wrote that Samsung&amp;rsquo;s price &lt;a class="link" href="https://koreainvestinsights.com/post/samsung-188700-prices-three-bear-cases-at-once-2026-11-08/" &gt;only works if three bear cases hit at once&lt;/a&gt;. The rebuilt model gives a simpler reading. One assumption is enough: the boom ends after 2027.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Our view: we count 2028, and we stop there. The inventory ledger is the reason. Even if the industry adds 22% more DRAM bits in 2027, as brokers expect, supplier inventory in our model only rises to 7.6 weeks by the end of 2028, still below a normal 9 weeks. We do not count 2029 and later, because those inputs have no source and they would need memory spending at three to four times what demand can sustain. Our working range of ₩233,000 to ₩291,000 does not change. These numbers are provisional; data due in December will reset them.&lt;/strong&gt;&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;Samsung&amp;rsquo;s value is a staircase. The floor is a normal year, worth ₩155,600 to ₩162,000. Each step is one more year of boom profit, worth ₩40,000 to ₩64,000.&lt;/li&gt;
&lt;li&gt;The market stands partway up the first step. Under harsh capital assumptions it reaches the second step. It never reaches the third.&lt;/li&gt;
&lt;li&gt;2027 profit splits on one input: how much DRAM the industry adds. Our bottom-up build gives +10.7% and ₩724.7 trillion of operating profit. The sell-side +22% gives ₩616.9 trillion, next to the ₩606.6 trillion consensus.&lt;/li&gt;
&lt;li&gt;2028 profit holds in our model even at +22% supply (₩621.2 trillion), because inventory never gets back to normal. The model cannot draw a crash, though, and that is its largest blind spot.&lt;/li&gt;
&lt;li&gt;Micron&amp;rsquo;s supply view in mid-December comes first. The first data that can tell the two supply paths apart is TrendForce&amp;rsquo;s second-quarter 2027 contract-price outlook in late March.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id="samsungs-value-is-a-staircase"&gt;Samsung&amp;rsquo;s value is a staircase
&lt;/h2&gt;&lt;p&gt;&lt;img alt="Samsung value per share in ₩ thousand by the last year of boom profit counted, base capital conventions. Normal from 2027: 155.6 to 162.0. Count 2027: 200.2 to 215.8. Count through 2028: 242.4 to 279.9. Count through 2029: 282.4 to 342.2. The November 9 close of 189.9 sits between the first two columns. Each extra year adds 40.0 to 64.1." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/samsung-price-pays-for-less-than-one-boom-year-2028-gives-242k-280k-2026-11-10/ladder.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;Picture Samsung&amp;rsquo;s value as a staircase. The ground floor is what Samsung would be worth if memory went back to a normal year right away. Each step up is one more year of boom profit that you believe in.&lt;/p&gt;
&lt;p&gt;The ground floor is low. In our normal year, Samsung&amp;rsquo;s operating profit settles at ₩116 trillion to ₩122 trillion, against about ₩387 trillion estimated for 2026. Across our range of normal-year assumptions, conventional DRAM prices fall to about 0.3 times their third-quarter 2026 level. That floor is worth ₩155,600 to ₩162,000 a share, below today&amp;rsquo;s price.&lt;/p&gt;
&lt;p&gt;The steps are large. Counting 2027 adds ₩44,600 to ₩53,800. Counting 2028 adds another ₩42,200 to ₩64,100. Counting 2029 adds ₩40,000 to ₩62,300. The two ends of each bar are the two supply paths we explain below.&lt;/p&gt;
&lt;p&gt;Now read the price against the staircase. ₩189,900 is above the floor and below the first step. On the sell-side supply path it pays for 77% of 2027&amp;rsquo;s excess profit. On our structural path it pays for 52%. Put another way, to justify today&amp;rsquo;s price while counting 2027 to 2029, Samsung&amp;rsquo;s operating profit would have to settle at ₩200 trillion to ₩254 trillion a year. That is 52% to 66% of this year&amp;rsquo;s estimated profit.&lt;/p&gt;
&lt;h2 id="how-we-built-the-model"&gt;How we built the model
&lt;/h2&gt;&lt;p&gt;&lt;img alt="Six layers of the model, each passing the same quarterly state to the next. 1 Industry inventory ledger: chips made minus shipped is supplier stock, shipped minus used is buyer stock, prices rise when stock runs thin. 2 Samsung memory profit. 3 Company profit and EPS. 4 Capital and cash books. 5 Normal year: what buyers can keep paying once the shortage ends. 6 Value per share by two methods that must agree. Consensus is a check, never an input. Main judgment calls sit on 8 named dials. Across 4,212 settings the two methods differ by at most 8 times 10 to the minus 16." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/samsung-price-pays-for-less-than-one-boom-year-2028-gives-242k-280k-2026-11-10/chain.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;Our earlier Samsung work came from four families of models. They came closest on 2027 and split on most other things. When we lined them up, much of the disagreement had nothing to do with the future. The models used different share counts (6.519, 6.521 and 6.649 billion), different exchange rates (₩1,340, ₩1,364 and ₩1,417 per dollar), different tax rates, different discount timing and different treatment of cash.&lt;/p&gt;
&lt;p&gt;So we rebuilt the work as one chain with one rulebook. Six layers pass the same quarterly state down the line.&lt;/p&gt;
&lt;ol&gt;
&lt;li&gt;&lt;strong&gt;Industry inventory ledger.&lt;/strong&gt; Every quarter, chips made minus chips shipped is the change in supplier inventory, and chips shipped minus chips used is the change in buyer inventory. Prices rise when inventory runs thin and fall when it piles up. This layer turns supply and demand into a price path.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Samsung memory profit.&lt;/strong&gt; Bits times price for conventional DRAM, HBM and NAND, minus costs. Fixed costs are 55% of memory costs, a share that reproduces how Samsung&amp;rsquo;s costs moved in the 2019 and 2023 downturns.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Company profit and EPS.&lt;/strong&gt; The model adds the rest of Samsung, takes tax at 23.8% and divides by one share count, 6.6486 billion.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Capital and cash books.&lt;/strong&gt; Capex, depreciation, payouts and cash. Half of free cash flow goes back to shareholders. The rest piles up and earns 3% before tax, about 2.3% after tax.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Normal year.&lt;/strong&gt; What buyers can keep paying once the shortage ends. We assume memory can keep 25% of big-cloud capex, plus demand from phones and PCs, and we check that budget against what the industry can supply.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Value per share.&lt;/strong&gt; The model computes value two ways on the same books: residual income (book value plus profit above the cost of capital) and the cash paid to shareholders.&lt;/li&gt;
&lt;/ol&gt;
&lt;p&gt;Rules keep the chain honest. Consensus is a check, never an input, so we do not tune the model to land on brokers&amp;rsquo; numbers. The main judgment calls sit on eight named dials, such as 2027 supply or how long boom profit lasts, so the same judgment cannot enter twice. Other assumptions, such as starting inventory, are stated and tested one at a time; there is no calibration factor. And the two value methods must agree. Across 4,212 dial settings, the largest gap between them was 8×10⁻¹⁶, a rounding error.&lt;/p&gt;
&lt;p&gt;We also tested the chain against outside benchmarks. It reproduces the memory profit of our older engines within 0.4%. It runs against Samsung&amp;rsquo;s reported quarters since 2022. And a second AI model from another vendor reran the code on its own and reproduced nine headline numbers. That review raised five issues. We fixed three, recorded one as a weakness and withdrew one conclusion that our formulas did not support.&lt;/p&gt;
&lt;h2 id="the-first-fork-how-much-dram-the-industry-adds-in-2027"&gt;The first fork: how much DRAM the industry adds in 2027
&lt;/h2&gt;&lt;p&gt;&lt;img alt="2027 DRAM bit supply growth. Wafers +10.5%. With finer processes (+9%), +20.4% before HBM. After wafers move to HBM, conventional DRAM +6.6%. Adding HBM bits (8.5% of bits, +55%) gives +10.7% in total, our structural path. Citi estimates +18.8%, which needs wafers +17.1%. The sell-side estimates +22%, which needs wafers +19.7%. The top of our wafer input range is +13%." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/samsung-price-pays-for-less-than-one-boom-year-2028-gives-242k-280k-2026-11-10/supply.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;2027 profit turns on one number: how many more DRAM bits the industry ships. Brokers use about +22%, and Citi uses +18.8%. We built the number from the bottom up.&lt;/p&gt;
&lt;p&gt;Wafer capacity grows about 10.5% in 2027, and finer processes add about 9% more bits per wafer. Together that is +20.4% before HBM. HBM, the stacked memory inside AI accelerators, takes a growing share of those wafers. After the move, conventional DRAM supply grows only 6.6%. Adding HBM&amp;rsquo;s own bits (8.5% of the total, up 55%) gives +10.7% for the industry.&lt;/p&gt;
&lt;p&gt;To reach +22%, wafers would have to grow 19.7%. To reach +18.8%, they would have to grow 17.1%. The top of our wafer input range is 13%. The top-down figures sit close to our number before HBM takes its wafers, but we could not confirm how they are defined. We note the match and do not draw a conclusion from it. Micron&amp;rsquo;s guidance in mid-December should settle the question.&lt;/p&gt;
&lt;p&gt;&lt;img alt="Samsung operating profit in ₩ trillion for 2027 and 2028. Structural supply +10.7%: 724.7 and 889.2, EPS 83,142 and 101,853. Citi supply +18.8%: 635.3 and 735.2. Supply +22%: 616.9 and 621.2, EPS 70,844 and 71,300. FnGuide consensus: 606.6 and 668.8, EPS 73,055 and 81,587. The 2026 estimate is 387." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/samsung-price-pays-for-less-than-one-boom-year-2028-gives-242k-280k-2026-11-10/profit.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;The two supply paths give very different years. On the structural path, 2027 operating profit is ₩724.7 trillion and EPS is ₩83,142. At +22% it is ₩616.9 trillion and ₩70,844, close to the FnGuide consensus of ₩606.6 trillion and ₩73,055. All 26 profit combinations that pass the model&amp;rsquo;s checks put 2027 at ₩616.9 trillion or more. Consensus sits just below the bottom of our range.&lt;/p&gt;
&lt;p&gt;&lt;img alt="Change in 2027 operating profit from the base case of ₩724.7 trillion, one input at a time, ₩ trillion. Buyer stock at end-2026 of 6 weeks: +157.4. 2027 supply +22%: −107.8. Supplier stock of 3.3 weeks: −98.1. Supplier minimum stock of 4 weeks: −79.5. Weaker price response to falling stock: −68.1. Buyer stock of 14 weeks: −52.5. Won at 1,417: +48.9. HBM 2027 price +42%: −31.0. Won at 1,300: −25.3. Price rise capped at 40%: −25.2." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/samsung-price-pays-for-less-than-one-boom-year-2028-gives-242k-280k-2026-11-10/fy27.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;2027 is less settled than it looks. Fourth-quarter exit prices do carry into next year. The inventory that Samsung&amp;rsquo;s customers will hold at the end of this year, however, has not been measured. If buyers hold 6 weeks instead of our assumed 11, 2027 profit rises by ₩157 trillion. If suppliers hold 3.3 weeks instead of 4.5, it falls by ₩98 trillion. These swings are as large as the supply question.&lt;/p&gt;
&lt;h2 id="why-we-count-2028-inventory-never-gets-back-to-normal"&gt;Why we count 2028: inventory never gets back to normal
&lt;/h2&gt;&lt;p&gt;&lt;img alt="Weeks of DRAM inventory by quarter end, model output, 4Q26 to 4Q28. Supplier stock on the structural path falls from 4.5 to the 3-week floor by 3Q27. Supplier stock at +22% supply rises from 4.5 to 7.6 by 4Q28, below the normal 9 weeks. Buyer stock falls from 11 weeks to 3.0 on the structural path and to 7.3 at +22%. Quarterly DRAM price change in 2Q27: +11.2% structural, −1.4% at +22%. In 2028: flat versus −1.1% to −1.4% a quarter." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/samsung-price-pays-for-less-than-one-boom-year-2028-gives-242k-280k-2026-11-10/inventory.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;The market prices 2028 as the year the boom ends. Our inventory ledger disagrees, even with the brokers&amp;rsquo; supply number.&lt;/p&gt;
&lt;p&gt;On the structural path, 2027 drains both sides. Supplier inventory falls to the 3-week floor by the third quarter, and buyer inventory falls from 11 weeks to about 6 by the end of 2027. In 2028, conventional DRAM supply grows 13.5% and demand grows 12.9%, so prices hold flat.&lt;/p&gt;
&lt;p&gt;On the +22% path, supplier inventory does build, from 4.5 weeks at the end of 2026 to 7.6 weeks at the end of 2028. It still does not reach the 9 weeks we treat as normal. Prices slip only 1.1% to 1.4% a quarter, and 2028 operating profit stays at ₩621.2 trillion, about where 2027 was.&lt;/p&gt;
&lt;p&gt;This is why we side with counting 2028. The model says the profit survives, and the price says it does not. That gap is worth ₩42,000 to ₩64,000 a share.&lt;/p&gt;
&lt;p&gt;The evidence against this view deserves space. First, spot prices and lead times have started to soften. TrendForce&amp;rsquo;s module table showed the spot price of a 32GB DDR5 server module down 6.5% in the week of November 2. That is one week of data, and it mostly reflects a shrinking module premium while chip spot prices sit near their highs. TrendForce&amp;rsquo;s AI server parts tracker put DRAM lead times at 18 weeks on November 9, down from 20 two weeks earlier, as Chinese cloud firms cut high-capacity orders and U.S. demand stayed firm. Contract prices are still rising. Second, the model has no order cancellations and no unwinding of double orders, so it cannot draw a crash. The absence of a crash in the model says nothing about the odds of one. Third, the 2019 and 2023 downturns both came with only three major suppliers in the market, so a concentrated industry gives no protection on its own.&lt;/p&gt;
&lt;p&gt;We also share new analyses on Telegram as they are published. You are welcome to join us at &lt;a class="link" href="https://t.me/korea_invest_insights" target="_blank" rel="noopener"
 &gt;@korea_invest_insights&lt;/a&gt;.&lt;/p&gt;
&lt;h2 id="what-189900-is-really-saying"&gt;What ₩189,900 is really saying
&lt;/h2&gt;&lt;p&gt;&lt;img alt="Value per share in ₩ thousand by capital convention and the last year of boom profit counted. Base with structural supply: 162.0, 215.8, 279.9, 342.2. Base with supply +22%: 155.6, 200.2, 242.4, 282.4. Discount rate 12% with structural supply: 130.7, 182.6, 243.8, 302.6. Cautious payout with +22%: 149.6, 187.1, 223.0, 257.3. Conservative with structural supply: 124.7, 168.1, 219.8, 270.2. Conservative with +22%: 119.9, 155.9, 190.0, 222.3. Friendly with structural supply: 167.2, 227.8, 299.0, 367.3. Cells within 5% of the 189.9 close: 182.6, 187.1 and 190.0." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/samsung-price-pays-for-less-than-one-boom-year-2028-gives-242k-280k-2026-11-10/conventions.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;The staircase depends on capital assumptions: the discount rate, how much cash Samsung returns and how much we discount the cash it keeps. So we redrew it under several sets.&lt;/p&gt;
&lt;p&gt;Under our base set (a 10% discount rate, half of free cash flow returned and a 15% haircut on retained cash), the price pays for less than one year. With a 12% discount rate, it pays for about one year. Under a conservative set (12%, a third returned and a 30% haircut) on the +22% supply path, it pays for two years: counting through 2028 gives ₩190,000.&lt;/p&gt;
&lt;p&gt;So the price reads two ways: about one year of boom on normal assumptions, or two years on harsh ones. The two readings share one feature. No row gives any value to boom profit in 2029 or later. Every &amp;ldquo;through 2029&amp;rdquo; cell is at least 17% above the price.&lt;/p&gt;
&lt;p&gt;How much cash Samsung returns matters for this reading. We covered that link in &lt;a class="link" href="https://koreainvestinsights.com/post/samsung-two-engines-profit-and-payouts-equity-roe-2026-11-08/" &gt;Samsung&amp;rsquo;s two engines&lt;/a&gt;. The board&amp;rsquo;s return policy in January gives the first answer.&lt;/p&gt;
&lt;h2 id="why-we-do-not-treat-the-big-numbers-as-proof"&gt;Why we do not treat the big numbers as proof
&lt;/h2&gt;&lt;p&gt;&lt;img alt="Industry memory revenue implied by the model versus a sustainable spending budget, $ trillion. Budget: 0.51 in 2027, 0.56 in 2028, 0.60 in 2029, 0.63 in 2030. Supply +22% path: 1.57, 1.61, 1.66, 2.31. Structural path: 1.79, 2.17, 2.26, 2.45. The paths run at 2.8 to 3.9 times the budget." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/samsung-price-pays-for-less-than-one-boom-year-2028-gives-242k-280k-2026-11-10/budget.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;Our full grid of 4,212 dial settings gives ₩198,900 to ₩834,900 a share, and no setting falls below today&amp;rsquo;s price. That sounds like a strong signal. We do not use it as one.&lt;/p&gt;
&lt;p&gt;The reason is how the grid is built. Its shortest setting for the life of boom profit still counts every year through 2029, so the whole grid starts above the price by construction. The question the price asks lives below the grid, which is why we built the staircase separately.&lt;/p&gt;
&lt;p&gt;The grid&amp;rsquo;s base cell, worth ₩427,800, is an upper value in our reading, and we do not use it as a center. To get there, industry memory revenue would stay at 3.5 to 3.9 times a sustainable budget every year through 2030 before normalizing in 2033. The 2029 and 2030 inputs behind it have no source. By 2033 Samsung would also hold ₩2,047 trillion of retained cash, so much of that value depends on a 50% payout lasting seven years.&lt;/p&gt;
&lt;h2 id="what-moves-the-value"&gt;What moves the value
&lt;/h2&gt;&lt;p&gt;&lt;img alt="Swing in value per share when counting through 2028, ₩ thousand. Does 2028 boom profit survive: 42 to 64, settled by 2028 contracts for HBM and DRAM in 2H 2027. Height of the normal year: 64 to 66, settled by measured memory spending against cloud capex in 2027. Capital conventions: 52 to 60, settled by the board return policy in January 2027. 2027 supply definition: 38, settled by Micron’s supply view in mid-December." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/samsung-price-pays-for-less-than-one-boom-year-2028-gives-242k-280k-2026-11-10/levers.en.svg"&gt;
&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Variable&lt;/th&gt;
 &lt;th&gt;Swing per share&lt;/th&gt;
 &lt;th&gt;Where things stand&lt;/th&gt;
 &lt;th&gt;What would flip it&lt;/th&gt;
 &lt;th&gt;When&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Does 2028 boom profit survive?&lt;/td&gt;
 &lt;td&gt;₩42,000 to ₩64,000&lt;/td&gt;
 &lt;td&gt;The model says yes. The price says no.&lt;/td&gt;
 &lt;td&gt;2028 contract prices for HBM and DRAM, buyers cancelling orders, supplier inventory above 5 weeks and rising&lt;/td&gt;
 &lt;td&gt;2H 2027; inventory from late January 2027&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;How high is the normal year?&lt;/td&gt;
 &lt;td&gt;₩64,000 to ₩66,000&lt;/td&gt;
 &lt;td&gt;We assume memory keeps 25% of cloud capex (range 20% to 30%)&lt;/td&gt;
 &lt;td&gt;Measured memory spending against cloud capex, contract prices in 2H 2027&lt;/td&gt;
 &lt;td&gt;2027&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Capital conventions&lt;/td&gt;
 &lt;td&gt;₩52,000 to ₩60,000&lt;/td&gt;
 &lt;td&gt;Open&lt;/td&gt;
 &lt;td&gt;The payout ratio and buyback share in the board&amp;rsquo;s return policy&lt;/td&gt;
 &lt;td&gt;January 2027&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;2027 supply definition&lt;/td&gt;
 &lt;td&gt;₩38,000&lt;/td&gt;
 &lt;td&gt;Open. Consensus sits at +22%&lt;/td&gt;
 &lt;td&gt;Micron&amp;rsquo;s view of 2027 industry DRAM bit supply&lt;/td&gt;
 &lt;td&gt;Mid-December 2026&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Our view rests on the first row. We side with 2028 profit surviving because of the inventory ledger, and we listed the evidence against it above. The second row is as large, and nobody can observe it before 2027. The supply definition has the smallest swing, but it is the first to get an answer.&lt;/p&gt;
&lt;h2 id="where-the-model-is-weak"&gt;Where the model is weak
&lt;/h2&gt;&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Weakness&lt;/th&gt;
 &lt;th&gt;Where it shows&lt;/th&gt;
 &lt;th&gt;Size&lt;/th&gt;
 &lt;th&gt;Next fix&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;No order cancellations or double-order unwinding&lt;/td&gt;
 &lt;td&gt;The model cannot draw a 2028 crash&lt;/td&gt;
 &lt;td&gt;Not measured&lt;/td&gt;
 &lt;td&gt;Add buyer cancellations to the inventory ledger&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;No demand-budget check for 2027 to 2030&lt;/td&gt;
 &lt;td&gt;Base-path revenue at 3.5 to 3.9 times the sustainable budget&lt;/td&gt;
 &lt;td&gt;Upper values inflated&lt;/td&gt;
 &lt;td&gt;Make yearly budget use a gate&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Prices react more to inventory changes than to levels&lt;/td&gt;
 &lt;td&gt;2027 profit moves with starting inventory&lt;/td&gt;
 &lt;td&gt;±₩50 trillion to ₩100 trillion in 2027&lt;/td&gt;
 &lt;td&gt;Add a level term, reset starting stock in late January&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;The 2029 cleanroom wave is missing&lt;/td&gt;
 &lt;td&gt;Conventional supply grows only 8.0% in 2029 and 3.8% in 2030&lt;/td&gt;
 &lt;td&gt;The whole path after 2029&lt;/td&gt;
 &lt;td&gt;Feed the cleanroom schedule into the ledger&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Weak tax fit to history&lt;/td&gt;
 &lt;td&gt;Past effective tax median 11.1% against our 23.8%&lt;/td&gt;
 &lt;td&gt;Fits 2026-level profit only&lt;/td&gt;
 &lt;td&gt;Fix tax and share rules from the third-quarter report&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;NAND path&lt;/td&gt;
 &lt;td&gt;NAND prices fall to 0.25 times the 2026 average by 2030&lt;/td&gt;
 &lt;td&gt;NAND revenue falls from ₩177 trillion to ₩66 trillion&lt;/td&gt;
 &lt;td&gt;Review supply-cut response&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;The missing cleanroom wave matters most for the years we do not count. We described that wave in &lt;a class="link" href="https://koreainvestinsights.com/post/dram-glut-2019-rhymes-in-2029-hbm-takes-the-wafers-2026-11-08/" &gt;the DRAM glut post&lt;/a&gt;.&lt;/p&gt;
&lt;h2 id="what-we-watch-next"&gt;What we watch next
&lt;/h2&gt;&lt;p&gt;&lt;img alt="Timeline of observations. Mid-December 2026: Micron’s 2027 industry DRAM supply view. Plus 20% or more with no inventory build means the +22% path; plus 12% or less means the structural path. Late December 2026: TrendForce’s first-quarter 2027 contract outlook, which does not pick a path. Late January 2027: Samsung’s fourth-quarter results, inventory and board return policy, which move value by minus 52,000 to plus 19,000 won. Late March 2027: TrendForce’s second-quarter outlook, the first real fork, +11.2% versus −1.4% in the model. Second half of 2027: 2028 contracts for HBM and conventional DRAM decide whether 2028 stays counted." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/samsung-price-pays-for-less-than-one-boom-year-2028-gives-242k-280k-2026-11-10/timeline.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;These checkpoints mark where we will look at the view again. They are not automatic trading rules.&lt;/p&gt;
&lt;h2 id="data-notes"&gt;Data notes
&lt;/h2&gt;&lt;ul&gt;
&lt;li&gt;We collected no new external data for this post. Figures come from our model run of November 10 and our dated model files and analysis notes.&lt;/li&gt;
&lt;li&gt;The share price is the KRX close on November 9. Consensus is FnGuide as of November 6. We did not refresh either for this post.&lt;/li&gt;
&lt;li&gt;All values are provisional and come before Micron&amp;rsquo;s December guidance and TrendForce&amp;rsquo;s late-December outlook. We will rerun the same input file when the data arrives.&lt;/li&gt;
&lt;li&gt;Grid cell counts and ranges are not probabilities. We did not weight scenarios.&lt;/li&gt;
&lt;li&gt;End-2026 inventory weeks, price-response coefficients and buyers&amp;rsquo; target inventory are assumptions. Together they move 2027 profit by ±₩50 trillion to ₩150 trillion.&lt;/li&gt;
&lt;li&gt;EPS uses a 23.8% tax rate. With the 20.3% rate implied by consensus, the structural path gives ₩86,961 for 2027 and ₩106,532 for 2028.&lt;/li&gt;
&lt;li&gt;Exchange rates are ₩1,337.4 per dollar for the fourth quarter of 2026 (actual through October 30) and ₩1,340 from 2027.&lt;/li&gt;
&lt;li&gt;Spot and lead-time figures come from TrendForce tables as of November 2 and November 9.&lt;/li&gt;
&lt;li&gt;The second-model review checked calculations and formulas. It did not re-verify the facts or freshness of the original sources.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;&lt;strong&gt;How many years of Samsung&amp;rsquo;s memory boom would you pay for today? Reply with A, B, C or D, and the one number that decides it for you.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;A. Less than one. The price is right, and the boom fades during 2027.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;B. One. Count 2027, then a normal year.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;C. Two. Inventory stays thin through 2028.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;D. Three or more. HBM contracts carry excess profit past 2029.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;This is for research and information only, not investment advice. Value ranges are analytical assumptions, not trading instructions or individual portfolio decisions. Figures are dated data. We built and ran the model with Anthropic&amp;rsquo;s Claude, and an OpenAI GPT model reviewed the calculations without making judgments. Samsung&amp;rsquo;s memory demand depends on AI capital spending, and both companies belong to the industry on the buying side. We did not assess the size of that effect.&lt;/p&gt;</description></item><item><title>HBM's Fight Moved to a 2.7% Part: TSMC and NVIDIA Win on Every Path, and Samsung's Fate Rides on One Google Decision</title><link>https://koreainvestinsights.com/post/hbm-base-die-fight-tsmc-nvidia-win-every-path-samsung-rides-on-google-2026-11-08/</link><pubDate>Sun, 08 Nov 2026 20:35:00 +0900</pubDate><guid>https://koreainvestinsights.com/post/hbm-base-die-fight-tsmc-nvidia-win-every-path-samsung-rides-on-google-2026-11-08/</guid><description>&lt;img src="https://koreainvestinsights.com/" alt="Featured image of post HBM's Fight Moved to a 2.7% Part: TSMC and NVIDIA Win on Every Path, and Samsung's Fate Rides on One Google Decision" /&gt;&lt;p&gt;Open an HBM stack and the DRAM chips get the attention. Under them sits a logic chip called the &lt;strong&gt;base die&lt;/strong&gt;, which handles the link to the GPU. Until now, memory makers designed it, and the contest between them was about DRAM cores: capacity, yield, speed. Now NVIDIA, the JEDEC standards body, TSMC, Marvell and Samsung are all putting more logic into it, and the contest is moving with it.&lt;/p&gt;
&lt;p&gt;The catch is the size. In 2028 the base die is worth about &lt;strong&gt;$7.6 billion&lt;/strong&gt;, 2.7% of an HBM market we assume at $282 billion. What it controls is larger. A 10% move in the price of NVIDIA&amp;rsquo;s HBM volume is &lt;strong&gt;$10.5 billion to $16.4 billion&lt;/strong&gt;. The base die is a lever, and the prize is who sets the price.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Our view: TSMC and NVIDIA win on every path we can draw. SK hynix loses part of its NVIDIA premium and AMD falls behind on every path. Broadcom, Google and Samsung rise or fall together on one variable: whether Google keeps its TPU memory outside NVIDIA&amp;rsquo;s NVLink Fusion. On our assumptions Samsung sells 51% to 76% of its HBM to non-NVIDIA buyers, the highest of the three makers, and Google alone is 23% to 47% of it. So Samsung&amp;rsquo;s loss is small if NVIDIA&amp;rsquo;s standard stays inside NVIDIA&amp;rsquo;s own volume, and large if Google crosses over. In our 2028 math that is a loss of ₩3.8 trillion to ₩6.4 trillion in the base case and ₩23.3 trillion (3.8% of operating profit) in the worst case. Even then, the first-order driver of Samsung&amp;rsquo;s profit is still commodity DRAM price.&lt;/strong&gt;&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;The base die is a $7.6 billion market. A 10% price move on NVIDIA&amp;rsquo;s HBM volume is bigger, which is why NVIDIA wants the design.&lt;/li&gt;
&lt;li&gt;Non-NVIDIA is no longer one camp. NVIDIA took a $2 billion stake in Marvell (March 31) and bought $3.5 billion of MediaTek convertible bonds (August 31). The real fault line is NVIDIA&amp;rsquo;s ecosystem against Broadcom&amp;rsquo;s camp.&lt;/li&gt;
&lt;li&gt;Our subjective odds for 2028: the two-track world holds at 55%, design partners and TSMC spread at 25%, NVIDIA&amp;rsquo;s standard spreads to Google or Meta at 20%.&lt;/li&gt;
&lt;li&gt;If memory makers must win back a 10% price cut with market share, SK hynix needs 5.8 points, Samsung 3.1 and Micron 2.2. The three needs add to 111%, so someone loses share.&lt;/li&gt;
&lt;li&gt;Samsung&amp;rsquo;s worst path costs 4 to 6 times its base case, and that asymmetry sits with Google.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id="a-chip-nobody-watched-is-now-where-the-fight-is"&gt;A chip nobody watched is now where the fight is
&lt;/h2&gt;&lt;p&gt;An HBM stack is a tower of DRAM chips on a logic chip. For HBM4 and earlier generations, the memory makers designed the logic chip. Several designs now compete to take it over:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;NVIDIA&amp;rsquo;s NVHBM (announced August 26) moves the memory controller into the base die. It replaces the wide parallel bus of the JEDEC standard with a narrow serial link between dies, and it comes inside NVLink Fusion, NVIDIA&amp;rsquo;s program for connecting other companies&amp;rsquo; chips to its network. Amazon&amp;rsquo;s Trainium4 is the first partner. NVIDIA&amp;rsquo;s performance claims are company statements.&lt;/li&gt;
&lt;li&gt;JEDEC&amp;rsquo;s own SPHBM4 (July 13) goes the same direction. It keeps the HBM4 DRAM dies, adds a new-interface base die with 4:1 serialization of 512 signals, and allows an organic substrate.&lt;/li&gt;
&lt;li&gt;TSMC and GUC offer a custom base die on N3P with about twice the efficiency (company materials via press). Marvell sells custom HBM4E with 2nm to 3nm logic in the base die. Samsung builds its own 4nm base die.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;All of these narrow the parallel interface, so more logic moves into the base die. Whoever designs that logic controls the margin and the specification.&lt;/p&gt;
&lt;h2 id="who-designs-the-base-die-decides-who-keeps-the-margin"&gt;Who designs the base die decides who keeps the margin
&lt;/h2&gt;&lt;p&gt;&lt;img alt="Who designs the base die decides who keeps the margin. Five buyer camps. NVIDIA: memory maker designs until 2027, NVIDIA from 2028; made at TSMC, SK hynix 12nm-class, Samsung 4nm; memory-maker power weakest. Broadcom camp (Google TPU, Meta MTIA, OpenAI Titan): memory maker designs standard HBM4/4E; Samsung 4nm and 2nm I/O or SK hynix at TSMC; power strong. Design houses (AWS Trainium, Microsoft Maia): design partner; TSMC N3P, Marvell 2nm and 3nm; power middle. AMD: memory maker; Samsung 4nm; power strong. HBM avoiders (Qualcomm, Intel): not applicable, outside HBM pricing." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/hbm-base-die-fight-tsmc-nvidia-win-every-path-samsung-rides-on-google-2026-11-08/control.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;The question is who holds the pen. If the buyer designs the base die, as NVIDIA will from 2028, the memory makers become suppliers of a core die to a fixed design, and they are easy to swap. If a design partner designs it, the logic value moves to the partner and to TSMC. If the memory maker keeps it, as with standard HBM4 and 4E or Samsung&amp;rsquo;s turnkey offer, bargaining power stays with the memory maker.&lt;/p&gt;
&lt;p&gt;TSMC is paid in all three cases. NVHBM&amp;rsquo;s base die, SK hynix&amp;rsquo;s 12nm-class base die, Micron&amp;rsquo;s HBM4E, C-HBM4E on N3P and CoWoS packaging all run through TSMC. The routes around TSMC are Samsung&amp;rsquo;s 4nm and 2nm base dies, Intel&amp;rsquo;s EMIB-T packaging, and SPHBM4 on organic substrates. All three sit on the non-NVIDIA side.&lt;/p&gt;
&lt;h2 id="the-money-a-76-billion-part-sets-the-price-of-the-rest"&gt;The money: a $7.6 billion part sets the price of the rest
&lt;/h2&gt;&lt;p&gt;&lt;img alt="The base die is a $7.6B market; 10% of NVIDIA’s HBM price is bigger. 2028, $ billion. Base die whole market 7.6, 2.7% of the stack value. 10% price on Google’s HBM purchases 10.2, 36% of 2027 demand. 10% price on NVIDIA’s HBM purchases 10.5 to 16.4, 37% to 58% of 2027 demand." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/hbm-base-die-fight-tsmc-nvidia-win-every-path-samsung-rides-on-google-2026-11-08/stake.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;At a $282 billion HBM market (JPMorgan, secondary) and a 2.7% base-die value share, the base die is $7.6 billion. NVIDIA buys $105 billion to $164 billion of HBM in 2028 at a 37% to 58% demand share, so 10% of its volume price is $10.5 billion to $16.4 billion. Google&amp;rsquo;s 36% share makes 10% of its price worth $10.2 billion.&lt;/p&gt;
&lt;p&gt;This is our reading of why NVIDIA wants the design. The reward is not base-die revenue. It is the ability to make three suppliers build the same part and compete on price. NVIDIA&amp;rsquo;s stake in that price is larger than TSMC&amp;rsquo;s whole base-die market.&lt;/p&gt;
&lt;h2 id="nvidias-play-own-the-standard"&gt;NVIDIA&amp;rsquo;s play: own the standard
&lt;/h2&gt;&lt;p&gt;NVIDIA&amp;rsquo;s plan runs in three steps.&lt;/p&gt;
&lt;ol&gt;
&lt;li&gt;&lt;strong&gt;2026 to 2027, standard HBM4 and 4E.&lt;/strong&gt; Memory makers design the base die. SK hynix&amp;rsquo;s work with TSMC gave it a performance edge, and SK hynix holds 54% of NVIDIA&amp;rsquo;s HBM in 2026 (JPMorgan, secondary).&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;2028, NVHBM.&lt;/strong&gt; NVIDIA designs the controller and the die-to-die link, and a Seoul press report says the three makers build the same design on TSMC 3nm (secondary). SK hynix&amp;rsquo;s TSMC edge and Samsung&amp;rsquo;s 4nm edge both disappear on NVIDIA volume. What is left to compete on is core 1c yield, stacking and capacity. In our model, a 10% price cut on NVIDIA volume costs ₩4.9 trillion at Samsung, ₩6.3 trillion at SK hynix and ₩2.8 trillion at Micron in 2028 operating profit.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;2029 and later, logic under the memory.&lt;/strong&gt; A report on November 6 says TSMC is looking at 3D DRAM (business plan unconfirmed). SK hynix is also hiring in San Jose for 3D DRAM on logic and co-designing a logic die with a U.S. customer (secondary). If NVIDIA and TSMC design and make the logic layer under the cells, the memory maker&amp;rsquo;s share narrows to the cell array.&lt;/li&gt;
&lt;/ol&gt;
&lt;p&gt;NVIDIA is also growing tiers outside HBM: SOCAMM (LPDDR) for inference, KV flash and Groq-style SRAM designs. A buyer that holds the design and the alternatives has the stronger hand.&lt;/p&gt;
&lt;h2 id="non-nvidia-is-not-one-camp"&gt;Non-NVIDIA is not one camp
&lt;/h2&gt;&lt;p&gt;&lt;img alt="The real fault line: NVIDIA’s ecosystem versus Broadcom’s camp. NVIDIA ecosystem (NVLink Fusion): AWS Trainium4 (first NVHBM partner), Marvell ($2B equity, March 31), MediaTek ($3.5B convertible, August 31), Alchip (Fusion partner since 2025). Broadcom camp: Google TPU (largest non-NVIDIA buyer, 36% of 2027 HBM), Meta MTIA and TPU rental, OpenAI Titan, Broadcom (no base-die answer yet per Counterpoint). Samsung sits between as the turnkey base-die supplier. The swing variable is whether Google keeps TPU memory outside NVLink Fusion." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/hbm-base-die-fight-tsmc-nvidia-win-every-path-samsung-rides-on-google-2026-11-08/camps.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;We first split non-NVIDIA demand four ways: Broadcom-centered hyperscalers, design-house custom chips, AMD, and HBM avoiders. Then two deals changed the map. NVIDIA took a $2 billion stake in Marvell on March 31, and on August 31 it bought $3.5 billion of MediaTek&amp;rsquo;s $3.9 billion convertible issue. MediaTek adopted NVLink Fusion for its own customers. Alchip, MediaTek and Marvell were also on Fusion&amp;rsquo;s first list of custom-compute partners in 2025. JPMorgan reads the Marvell deal as confirming Marvell as AWS&amp;rsquo;s Fusion integration partner (analyst reading).&lt;/p&gt;
&lt;p&gt;So design houses are no longer an independent camp. The practical split is between NVIDIA&amp;rsquo;s ecosystem and Broadcom&amp;rsquo;s camp.&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Broadcom&amp;rsquo;s camp is the biggest non-NVIDIA demand.&lt;/strong&gt; Alphabet alone is 36% of 2027 HBM demand (Morgan Stanley). Broadcom&amp;rsquo;s CoWoS allocation grows from 230,000 to 350,000 wafers in 2027 (JPMorgan, reported). This camp still uses standard HBM designed by the memory maker, which is where Samsung&amp;rsquo;s turnkey offer sells best. Reports have Samsung with more than 60% of Google&amp;rsquo;s HBM3E, an exclusive on OpenAI&amp;rsquo;s Titan HBM4, and a 2nm memory I/O die on MediaTek&amp;rsquo;s next Google TPU (all secondary, claims). Counterpoint calls Broadcom the last big design house NVIDIA has not joined, and says it has no base-die answer (analyst view). That is the gap Samsung&amp;rsquo;s 4nm and 2nm base dies can fill.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;AMD is the biggest single-chip buyer of standard HBM and Samsung&amp;rsquo;s first customer.&lt;/strong&gt; The MI455X carries 432GB of HBM4 on 12 stacks. The March 18 AMD and Samsung agreement is the only primary-source supply deal in this story, and it includes talks on foundry cooperation. AMD is 12.1% of 2027 demand (Morgan Stanley).&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;HBM avoiders cap the HBM premium.&lt;/strong&gt; Qualcomm&amp;rsquo;s AI200 carries 768GB of LPDDR per card, and Intel&amp;rsquo;s Crescent Island carries 160GB of LPDDR5X. The bits stay in DRAM but outside HBM pricing. Orders are still small.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;Google is the swing. It splits suppliers (Samsung, SK hynix), design (Broadcom, MediaTek) and packaging (TSMC, Intel EMIB-T from 2027, with a reported 2028 booking of more than 3 million TPUs). MediaTek, which designs Google&amp;rsquo;s next TPU, is now inside Fusion, and Alphabet joined the same bond offer. That opens a path to NVIDIA&amp;rsquo;s standard. It is not a decision, so we keep our odds where they were.&lt;/p&gt;
&lt;p&gt;We also share new analyses on Telegram as they are published. You are welcome to join us at &lt;a class="link" href="https://t.me/korea_invest_insights" target="_blank" rel="noopener"
 &gt;@korea_invest_insights&lt;/a&gt;.&lt;/p&gt;
&lt;h2 id="samsung-is-the-most-exposed-to-non-nvidia-buyers"&gt;Samsung is the most exposed to non-NVIDIA buyers
&lt;/h2&gt;&lt;p&gt;&lt;img alt="Samsung is the most non-NVIDIA of the three. Share of each maker’s HBM that goes to non-NVIDIA buyers across four assumption sets, percent. Samsung 51 to 76, of which Google 23 to 47. SK hynix 34 to 61. Micron 31 to 49. The order is the same in all four sets." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/hbm-base-die-fight-tsmc-nvidia-win-every-path-samsung-rides-on-google-2026-11-08/exposure.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;The formula is each buyer&amp;rsquo;s 2027 demand share times each maker&amp;rsquo;s split inside that buyer. The splits are assumptions fixed from press reports: NVIDIA (Samsung 28%, SK hynix 52%, Micron 20%), Google (55, 40, 5), AMD (60, 20, 20) and others (35, 45, 20). We ran two NVIDIA demand shares, 37% (Morgan Stanley) and 58%, and let Google&amp;rsquo;s volume either stay or move toward SK hynix.&lt;/p&gt;
&lt;p&gt;The order never changes. Non-NVIDIA is the largest part of Samsung&amp;rsquo;s HBM, then SK hynix&amp;rsquo;s, then Micron&amp;rsquo;s. The shares these assumptions imply (Samsung 33% to 43%, SK hynix 43% to 50%, Micron 15% to 17%) come from the assumptions and do not test them.&lt;/p&gt;
&lt;p&gt;NVHBM&amp;rsquo;s reach follows from the same table. NVIDIA alone is 37% of demand. Adding Amazon (about 5%, assumed) makes 42%, and adding Google makes 78%. If Google adopts NVIDIA&amp;rsquo;s standard, about 75% of Samsung&amp;rsquo;s HBM sits under it, against 25% with NVIDIA alone. Across the four cases the range is 70% to 81%. This assumes each buyer&amp;rsquo;s whole demand moves.&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;SK hynix&lt;/strong&gt; loses the most premium when NVIDIA neutralizes suppliers, because its NVIDIA share is high.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Samsung&amp;rsquo;s&lt;/strong&gt; risk sits with Google and Broadcom. The three triggers are whether Google adopts NVHBM, whether TPU HBM4 and 4E allocation tilts to SK hynix, and whether Broadcom&amp;rsquo;s custom base die goes to TSMC N3P.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Micron&lt;/strong&gt; has high NVIDIA exposure, but as the number three supplier it also gains from standard parts that lower qualification barriers.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id="what-each-path-costs-samsung"&gt;What each path costs Samsung
&lt;/h2&gt;&lt;p&gt;&lt;img alt="What each path costs Samsung in 2028 operating profit. Change versus today’s structure, ₩ trillion. Two-track holds (55%): minus 6.4 to minus 3.8, 0.6% to 1.0%. Design partners and TSMC spread (25%): minus 16.0 to minus 8.0, 1.3% to 2.6%. NVIDIA standard spreads to Google (20%): minus 23.3, 3.8%. Probability-weighted: minus 10.5, about 1.7%. Offset if demand moves 5 to 10 points to non-NVIDIA: plus 3.1 to plus 6.2." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/hbm-base-die-fight-tsmc-nvidia-win-every-path-samsung-rides-on-google-2026-11-08/paths.en.svg"&gt;
&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Path&lt;/th&gt;
 &lt;th&gt;What happens&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Our odds&lt;/th&gt;
 &lt;th&gt;Why&lt;/th&gt;
 &lt;th&gt;Signal&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Two-track holds&lt;/td&gt;
 &lt;td&gt;NVIDIA and Amazon use NVHBM. Google, Meta, OpenAI and AMD stay on standard or turnkey HBM&lt;/td&gt;
 &lt;td style="text-align: right"&gt;55%&lt;/td&gt;
 &lt;td&gt;Google and Broadcom have their own interconnects and Broadcom&amp;rsquo;s Meta contract runs to 2029, so they have little reason to hand NVIDIA the spec. AMD has little reason to adopt a standard its rival controls&lt;/td&gt;
 &lt;td&gt;Google and Meta stay silent on NVHBM. Broadcom&amp;rsquo;s custom HBM keeps the memory maker&amp;rsquo;s design&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Design partners and TSMC spread&lt;/td&gt;
 &lt;td&gt;Non-NVIDIA buyers also adopt C-HBM4E or Marvell-style custom parts, so memory makers supply core dies everywhere&lt;/td&gt;
 &lt;td style="text-align: right"&gt;25%&lt;/td&gt;
 &lt;td&gt;N3P&amp;rsquo;s roughly 2x efficiency and Marvell&amp;rsquo;s view that HBM exceeds 50% of an XPU&amp;rsquo;s cost push buyers toward custom&lt;/td&gt;
 &lt;td&gt;Broadcom or Google picks a TSMC N3P custom base die. Meta&amp;rsquo;s MTIA gets custom HBM&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;NVIDIA standard spreads&lt;/td&gt;
 &lt;td&gt;Google or Meta adopts NVLink Fusion and NVHBM&lt;/td&gt;
 &lt;td style="text-align: right"&gt;20%&lt;/td&gt;
 &lt;td&gt;Amazon already did, and NVIDIA wants Google too (secondary). Against that, Google is widening supply with EMIB-T, MediaTek and Samsung&amp;rsquo;s 2nm&lt;/td&gt;
 &lt;td&gt;Partner list at the December 2 event. Google&amp;rsquo;s next TPU memory spec&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Because Marvell and MediaTek are now inside Fusion, much of the second path can play out inside NVIDIA&amp;rsquo;s ecosystem. We keep the odds, but Samsung&amp;rsquo;s two bad paths, 45% together, are best read as one risk.&lt;/p&gt;
&lt;p&gt;The cost on each path, using $99 billion of Samsung 2028 HBM revenue ($282 billion market, 35% share), ₩2.66 trillion of operating profit per point of share, a 10% price cut on affected volume, and a base die worth 2.7% of the stack at a 60% contribution margin:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Two-track holds:&lt;/strong&gt; 29% to 48% of Samsung&amp;rsquo;s HBM (NVIDIA and Amazon volume) loses 10% of price. That is −₩3.8 trillion to −₩6.4 trillion, 0.6% to 1.0% of 2028 operating profit.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Design partners spread:&lt;/strong&gt; the above, plus ₩1.55 trillion of base-die foundry contribution profit, plus 1 to 3 points of share lost with the turnkey advantage. That is −₩8.0 trillion to −₩16.0 trillion, 1.3% to 2.6%.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;NVIDIA standard spreads:&lt;/strong&gt; 75% of volume loses 10% of price (₩10.0 trillion), Google&amp;rsquo;s share of Samsung falls from 55% to 40% (5.4 points of market, ₩14.4 trillion), Google&amp;rsquo;s base die is worth ₩1.0 trillion, and we subtract ₩2.1 trillion for double-counting the price loss on volume already lost. That is about −₩23.3 trillion, 3.8%.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Probability-weighted,&lt;/strong&gt; using range midpoints, about −₩10.5 trillion, 1.7%.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;The base case is small because the loss lands where Samsung&amp;rsquo;s share is lowest (28% of NVIDIA&amp;rsquo;s volume) and the growth lands where its share is highest (51% of non-NVIDIA volume). If demand moves 5 points from NVIDIA to non-NVIDIA buyers, Samsung&amp;rsquo;s share rises 1.2 points and operating profit rises ₩3.1 trillion. A 10-point move gives 2.3 points and ₩6.2 trillion. Counterpoint expects HBM demand for ASIC compute to grow many times over by 2028 (secondary). If that shift happens, it offsets part or all of the base-case price loss. If it does not, the base case is negative.&lt;/p&gt;
&lt;p&gt;The worst path is the tail. If Google moves to NVIDIA&amp;rsquo;s standard, or tilts HBM4E toward SK hynix, the damage is about 4 to 6 times the base case.&lt;/p&gt;
&lt;p&gt;There is a foundry leg too. In the base case Samsung keeps the base-die foundry business for non-NVIDIA volume, about ₩2.6 trillion of internal revenue in 2028 (excluding Amazon) and about ₩1.55 trillion of contribution profit. That equals about 40% of our estimated 2026 foundry and LSI loss of ₩3.92 trillion. On the second and third paths that revenue shrinks, and one support for the foundry breakeven case weakens.&lt;/p&gt;
&lt;p&gt;All three paths are 0% to 4% of 2028 operating profit. The first-order driver of Samsung&amp;rsquo;s profit is still commodity DRAM price, and HBM structure is second-order. It is, however, one of the biggest items separating Samsung from SK hynix in relative terms.&lt;/p&gt;
&lt;p&gt;What changes for Samsung&amp;rsquo;s HBM story is its aim. With HBM4, Samsung used a 1c core and its own 4nm base die, more advanced logic than SK hynix&amp;rsquo;s TSMC 12nm-class die, and its HBM share rose from 21% in 1Q26 to 33% in 2Q26 (Counterpoint). Under NVHBM the base die is TSMC 3nm for everyone, and SK hynix moves to 1c from HBM4E. On NVIDIA volume Samsung&amp;rsquo;s weapons shrink to core yield, capacity and stacking. The story becomes standard and turnkey supply to non-NVIDIA buyers, and that business only exists outside NVIDIA.&lt;/p&gt;
&lt;h2 id="who-wins-on-which-path"&gt;Who wins on which path
&lt;/h2&gt;&lt;p&gt;&lt;img alt="Who wins on which path. Our judgment from strong gain to strong loss. NVIDIA: plus, zero, plus plus. TSMC: plus, plus plus, plus. Broadcom: plus, minus, minus minus. Marvell and MediaTek: zero, plus, plus. Google: plus, plus, minus. Amazon: plus, zero, plus. Samsung: minus, minus minus, minus minus minus. SK hynix: minus, zero, minus. Micron: zero, zero, plus. AMD: minus on all three." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/hbm-base-die-fight-tsmc-nvidia-win-every-path-samsung-rides-on-google-2026-11-08/scorecard.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;TSMC and NVIDIA do not lose on any path. AMD is behind on all three. Broadcom and Samsung move together: both are safe in the base case and lose in the other two.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Company&lt;/th&gt;
 &lt;th&gt;What it gets&lt;/th&gt;
 &lt;th&gt;What it risks&lt;/th&gt;
 &lt;th&gt;What flips it&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;NVIDIA&lt;/td&gt;
 &lt;td&gt;Control of the base-die spec. Three makers on one design lower switching cost. Binds Amazon, Marvell, MediaTek and Alchip through Fusion&lt;/td&gt;
 &lt;td&gt;More dependence on TSMC 3nm. If non-NVIDIA buyers stay on memory-maker designs or SPHBM4, NVHBM stays inside its own camp. First-generation serial-link execution risk&lt;/td&gt;
 &lt;td&gt;Google and Meta absent from the December 2 partner list limits the win to its own volume&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;TSMC&lt;/td&gt;
 &lt;td&gt;NVHBM 3nm, SK hynix 12nm-class, Micron HBM4E, C-HBM4E N3P and CoWoS. The base die ties customers closer&lt;/td&gt;
 &lt;td&gt;Google&amp;rsquo;s EMIB-T and Intel packaging, SPHBM4 on organic substrates, Samsung&amp;rsquo;s 2nm I/O die. All of these are on the non-NVIDIA side&lt;/td&gt;
 &lt;td&gt;Intel confirms EMIB-T for Google. Broadcom&amp;rsquo;s custom base die goes to Samsung&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Broadcom&lt;/td&gt;
 &lt;td&gt;The last big design house outside NVIDIA&amp;rsquo;s orbit (Counterpoint view). The main option for Google, Meta and OpenAI to stay outside NVIDIA&lt;/td&gt;
 &lt;td&gt;No base-die answer. If the base die becomes the differentiator it depends on Samsung&amp;rsquo;s turnkey offer or TSMC. MediaTek, Google&amp;rsquo;s second TPU designer, is inside Fusion&lt;/td&gt;
 &lt;td&gt;Google or Meta joins Fusion. Broadcom names a base-die partner on its fiscal fourth-quarter call&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Marvell, MediaTek, Alchip, GUC&lt;/td&gt;
 &lt;td&gt;Marvell and MediaTek have NVIDIA money and Fusion access and still design custom chips for others. GUC has C-HBM4E&lt;/td&gt;
 &lt;td&gt;Marvell&amp;rsquo;s own custom HBM competes with NVHBM for the same customers. MediaTek must manage trust with Google and NVIDIA at once&lt;/td&gt;
 &lt;td&gt;How MediaTek&amp;rsquo;s Icefish connects, over Fusion or Google&amp;rsquo;s own network&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Google&lt;/td&gt;
 &lt;td&gt;The largest non-NVIDIA buyer with every supply layer split in two or more, which builds its leverage&lt;/td&gt;
 &lt;td&gt;If NVHBM becomes the de facto standard, keeping its own spec costs more. Allocation priority in a shortage&lt;/td&gt;
 &lt;td&gt;Joining Fusion gets cost and gives up spec independence&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Meta&lt;/td&gt;
 &lt;td&gt;Three options: NVIDIA GPUs, Google TPUs (rental reported, 2027 purchase talks) and MTIA with Broadcom through 2029&lt;/td&gt;
 &lt;td&gt;MTIA volume is small, so it is weak in custom HBM allocation&lt;/td&gt;
 &lt;td&gt;A confirmed TPU purchase widens Broadcom&amp;rsquo;s camp. Joining Fusion moves it to NVIDIA&amp;rsquo;s path&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Microsoft&lt;/td&gt;
 &lt;td&gt;Little. NVIDIA&amp;rsquo;s HBM cost savings may or may not reach GPU prices&lt;/td&gt;
 &lt;td&gt;Maia 200 reportedly uses SK hynix HBM3E alone. Small own-chip volume gives it the weakest custom-HBM leverage among hyperscalers. Maia 300 is 8 stacks and 288GB of HBM4 in 2027, supplier undisclosed&lt;/td&gt;
 &lt;td&gt;Maia 300 suppliers and design partner disclosed&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Amazon&lt;/td&gt;
 &lt;td&gt;Trainium4 adopts Fusion and NVHBM: stable supply, cost, NVIDIA rack compatibility&lt;/td&gt;
 &lt;td&gt;Hands memory-controller design to NVIDIA and loses spec independence. Keeps UALink in parallel as a partial hedge&lt;/td&gt;
 &lt;td&gt;Whether Trainium5 keeps NVHBM&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Samsung&lt;/td&gt;
 &lt;td&gt;51% to 76% non-NVIDIA. The one memory maker that can fill Broadcom&amp;rsquo;s base-die gap. AMD agreement (primary). Share recovery room on NVIDIA volume if suppliers are neutralized&lt;/td&gt;
 &lt;td&gt;Google&amp;rsquo;s choice (the −₩23.3 trillion path). Loss of turnkey advantage on the design-partner path. MediaTek inside Fusion&lt;/td&gt;
 &lt;td&gt;Broadcom adopts a Samsung base die (gain). Google or Meta joins Fusion (threat)&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;SK hynix&lt;/td&gt;
 &lt;td&gt;First place at NVIDIA (54%, 2026). TSMC partnership. Reported sole supplier for Microsoft&amp;rsquo;s chips. 3D DRAM on logic co-design&lt;/td&gt;
 &lt;td&gt;NVHBM erases the TSMC edge. A 10% price cut on NVIDIA volume is worth $5.5 billion to $8.5 billion. Share converges toward capacity and yield, so holding 52% is hard&lt;/td&gt;
 &lt;td&gt;Google HBM4E tilts to SK hynix. SK hynix leads on 1c yield&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Micron&lt;/td&gt;
 &lt;td&gt;Standard parts lower qualification barriers. TSMC base die fits any camp&lt;/td&gt;
 &lt;td&gt;Small Google and Broadcom-camp exposure, so a small share of non-NVIDIA growth. Lower NVIDIA volume price&lt;/td&gt;
 &lt;td&gt;Custom HBM4E with Google appears on the December call&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;AMD&lt;/td&gt;
 &lt;td&gt;The only primary-source supply link, with Samsung. Standard HBM4, 12 stacks&lt;/td&gt;
 &lt;td&gt;NVHBM&amp;rsquo;s claimed 30% more bandwidth and 15% less power, plus cost, widens NVIDIA&amp;rsquo;s lead. Strong supplier power makes AMD a weaker buyer&lt;/td&gt;
 &lt;td&gt;AMD adopts a custom HBM with a Samsung base die&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Intel gets a packaging bypass through Google&amp;rsquo;s EMIB-T, with small volume and SK hynix qualification pending. Qualcomm benefits from the HBM premium because its LPDDR design avoids it, though cheaper NVIDIA HBM narrows that advantage.&lt;/p&gt;
&lt;h2 id="the-memory-makers-neutralization-does-not-hit-evenly"&gt;The memory makers: neutralization does not hit evenly
&lt;/h2&gt;&lt;p&gt;&lt;img alt="Winning back a 10% price cut: the shares do not add up. Share of NVIDIA’s HBM volume today versus share needed to keep revenue flat. Samsung 28 to 31.1. SK hynix 52 to 57.8. Micron 20 to 22.2. All three together 100 to 111.1, above the 100% ceiling." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/hbm-base-die-fight-tsmc-nvidia-win-every-path-samsung-rides-on-google-2026-11-08/offset.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;Under NVHBM the three makers build the same design, so share inside NVIDIA converges toward core 1c yield and capacity. A 10% price cut on NVIDIA volume costs Samsung $3.0 billion to $4.6 billion, SK hynix $5.5 billion to $8.5 billion and Micron $2.1 billion to $3.3 billion, across the 37% to 58% demand range. To make up the revenue with share alone, Samsung needs 3.1 more points, SK hynix 5.8 and Micron 2.2.&lt;/p&gt;
&lt;p&gt;Those needs add to 111.1 points. That cannot happen, because shares add to 100. Someone gives up share, and convergence on yield and capacity points at SK hynix, which holds the largest share (52%) and so has the most to give back. In other words, SK hynix is likely to lose on price and on share, while Samsung and Micron can cover a price cut with small share gains. This is revenue only. It ignores margins and capacity limits, and in the 2027 to 2028 shortage allocation sets share.&lt;/p&gt;
&lt;h2 id="what-would-change-our-reading"&gt;What would change our reading
&lt;/h2&gt;&lt;p&gt;&lt;img alt="What to check next. December 2, 2026: NVIDIA GTC DC, whether Google, Meta or MediaTek customers join Fusion or NVHBM. Mid-December (estimated): Broadcom fiscal fourth quarter, who builds the custom HBM base die. December 17 (estimated): Micron fiscal first quarter, custom HBM4E customers. January 14, 2027 (customary): TSMC fourth quarter. January 27 to 28, 2027: SK hynix and Samsung fourth quarter. Any time: Meta confirms TPU purchases, Microsoft names Maia 300 suppliers." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/hbm-base-die-fight-tsmc-nvidia-win-every-path-samsung-rides-on-google-2026-11-08/calendar.en.svg"&gt;
&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Date&lt;/th&gt;
 &lt;th&gt;What to read&lt;/th&gt;
 &lt;th&gt;Default&lt;/th&gt;
 &lt;th&gt;What would move us&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;December 2, 2026&lt;/td&gt;
 &lt;td&gt;NVIDIA GTC DC: NVHBM and Fusion partners, whether Google, Meta or MediaTek customers are in, and Feynman memory&lt;/td&gt;
 &lt;td&gt;No new large partner&lt;/td&gt;
 &lt;td&gt;Google or Meta named: the NVIDIA-standard path rises about 15 points (our estimate), and Broadcom and Samsung are marked down together&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Mid-December (estimated)&lt;/td&gt;
 &lt;td&gt;Broadcom fiscal fourth quarter: AI revenue, custom chips by customer, HBM and packaging remarks&lt;/td&gt;
 &lt;td&gt;HBM structure not mentioned&lt;/td&gt;
 &lt;td&gt;A Samsung base die named supports the non-NVIDIA standard camp. A custom base die with TSMC reopens the design-partner path&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;December 17 (estimated)&lt;/td&gt;
 &lt;td&gt;Micron fiscal first quarter: custom HBM4E customers, ASIC share&lt;/td&gt;
 &lt;td&gt;ASIC customer growth&lt;/td&gt;
 &lt;td&gt;Google named: we revise the allocation assumptions&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;January 14, 2027 (customary)&lt;/td&gt;
 &lt;td&gt;TSMC fourth quarter: CoWoS allocation, C-HBM4E adoption, 3D DRAM question&lt;/td&gt;
 &lt;td&gt;Not mentioned&lt;/td&gt;
 &lt;td&gt;More custom base-die customers: the design-partner path rises about 5 points (our estimate)&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;January 27 to 28, 2027&lt;/td&gt;
 &lt;td&gt;SK hynix and Samsung fourth quarter: custom HBM4E wins, 2nm base die&lt;/td&gt;
 &lt;td&gt;Customer names withheld&lt;/td&gt;
 &lt;td&gt;A Samsung win with Google or Broadcom supports the two-track path&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;h2 id="data-notes"&gt;Data notes
&lt;/h2&gt;&lt;ul&gt;
&lt;li&gt;Supplier splits inside each buyer come from anonymous press reports, and Google&amp;rsquo;s 60% for Samsung dates from the HBM3E era. We do not know the HBM4 or 4E split. Read the results as order and range.&lt;/li&gt;
&lt;li&gt;NVIDIA&amp;rsquo;s 2027 demand share is estimated anywhere from 37% to 61%, because estimates mix bit demand, contracted volume and shipments. We computed with the 37% and 58% ends and did not pick a midpoint.&lt;/li&gt;
&lt;li&gt;Many items were confirmed through search summaries, not by opening the pages. The AMD and Samsung agreement is the only primary-source supply item. We did not open the original announcements for NVIDIA&amp;rsquo;s investments in Marvell and MediaTek, and the Marvell and Microsoft Maia 300 link is unconfirmed. Meta&amp;rsquo;s TPU purchase is at the negotiation stage in reports.&lt;/li&gt;
&lt;li&gt;Market size, base-die value share and demand shares are assumptions carried through the whole analysis. Paths, odds and the signs in the scorecard are our subjective judgment. The scorecard was made by one model family and was not cross-reviewed. We recomputed the 16 exposure cells and the price lines, and they match our research note.&lt;/li&gt;
&lt;li&gt;Counterpoint&amp;rsquo;s growth multiple for ASIC HBM demand comes from a headline we could not open (bot check), so we state it only as many times over.&lt;/li&gt;
&lt;li&gt;We found no Broadcom statement on base dies, no list of SPHBM4 supporters, and no confirmed HBM4 supplier for Meta or Microsoft Maia 300.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;Sources:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;a class="link" href="https://newsroom.amd.com/news/samsung-and-amd-expand-strategic-collaboratio/" target="_blank" rel="noopener"
 &gt;AMD: Samsung and AMD expand strategic collaboration&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://blogs.nvidia.com/blog/2026/08/26/nvlink-fusion-nvhbm/" target="_blank" rel="noopener"
 &gt;NVIDIA blog: NVLink Fusion and NVHBM (2026-08-26)&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.tomshardware.com/pc-components/dram/nvidia-custom-nvhbm-promises-30-percent-higher-bandwidth-15-percent-lower-power-than-commodity-hbm4e-custom-base-die-and-phy-will-be-available-to-nvlink-fusion-partners" target="_blank" rel="noopener"
 &gt;Tom&amp;rsquo;s Hardware: NVIDIA custom NVHBM, 30% more bandwidth, 15% less power&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.storagereview.com/news/nvidia-nvhbm-moves-the-memory-controller-into-the-hbm-stack-with-amazons-trainium4-first-in-line" target="_blank" rel="noopener"
 &gt;StorageReview: NVHBM moves the memory controller into the HBM stack, Trainium4 first&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://counterpointresearch.com/cn/insights/the-trillion-dollar-bottleneck-nvidias-nvhbm-fight-for-base-die-blog" target="_blank" rel="noopener"
 &gt;Counterpoint: The trillion-dollar bottleneck, NVIDIA&amp;rsquo;s NVHBM fight for the base die&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://counterpointresearch.com/en/insights/HBM-Demand-for-AI-Server-Compute-ASICs-to-Grow-35x-by-2028" target="_blank" rel="noopener"
 &gt;Counterpoint: HBM demand for AI server compute ASICs&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://blocksandfiles.com/ai-ml/2026/10/26/nvidia-sidesteps-jedec-hbm-standard/5305323" target="_blank" rel="noopener"
 &gt;Blocks and Files: NVIDIA sidesteps the JEDEC HBM standard&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.tomshardware.com/pc-components/dram/jedec-releases-new-sphbm4-standard-to-slash-ai-memory-costs-narrow-512-bit-interface-enables-dropping-expensive-interposers-for-organic-substrates" target="_blank" rel="noopener"
 &gt;Tom&amp;rsquo;s Hardware: JEDEC SPHBM4 standard&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.tomshardware.com/pc-components/dram/hbm-undergoes-major-architectural-shakeup-as-tsmc-and-guc-detail-hbm4-hbm4e-and-c-hbm4e-3nm-base-dies-to-enable-2-5x-performance-boost-with-speeds-of-up-to-12-8gt-s-by-2027" target="_blank" rel="noopener"
 &gt;Tom&amp;rsquo;s Hardware: TSMC and GUC on HBM4, HBM4E and C-HBM4E base dies&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.tomshardware.com/tech-industry/nvidia-invests-2-billion-in-marvell-to-deepen-nvlink-fusion-partnership" target="_blank" rel="noopener"
 &gt;Tom&amp;rsquo;s Hardware: NVIDIA invests $2 billion in Marvell&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://datacenterdynamics.com/en/news/nvidia-invests-2bn-in-marvell-as-part-of-wider-nvlink-fusion-deal" target="_blank" rel="noopener"
 &gt;DCD: NVIDIA invests $2bn in Marvell&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.networkworld.com/article/4220808/nvidia-invests-3-5b-in-mediatek-to-extend-its-grip-on-ai.html" target="_blank" rel="noopener"
 &gt;Network World: NVIDIA invests $3.5B in MediaTek&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://hothardware.com/news/nvidia-nvlink-fusion-major-boost-mediatek-deal" target="_blank" rel="noopener"
 &gt;HotHardware: NVLink Fusion and the MediaTek deal&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.servethehome.com/nvidia-nvlink-fusion-tapped-for-future-aws-trainium4-deployments/" target="_blank" rel="noopener"
 &gt;ServeTheHome: NVLink Fusion tapped for AWS Trainium4&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.techradar.com/pro/the-inference-inflection-has-arrived-nvidia-pumps-usd2-billion-into-chipmaker-marvell-to-boost-its-ai-factories-to-the-next-level-so-does-this-mean-itll-be-working-with-amazon-trainium-soon" target="_blank" rel="noopener"
 &gt;TechRadar: NVIDIA pumps $2 billion into Marvell&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.tomshardware.com/pc-components/dram/marvell-develops-custom-hbm-solution-that-offers-higher-performance-in-a-smaller-physical-space" target="_blank" rel="noopener"
 &gt;Tom&amp;rsquo;s Hardware: Marvell custom HBM solution&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.tomshardware.com/tech-industry/semiconductors/custom-ai-asics-examined-from-broadcom-to-mtia" target="_blank" rel="noopener"
 &gt;Tom&amp;rsquo;s Hardware: Custom AI ASICs examined, from Broadcom to MTIA&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.tomshardware.com/tech-industry/semiconductors/meta-reveals-four-new-mtia-chips-built-for-ai-inference" target="_blank" rel="noopener"
 &gt;Tom&amp;rsquo;s Hardware: Meta&amp;rsquo;s four new MTIA chips&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.techradar.com/pro/meta-and-google-could-be-about-to-sign-a-mega-ai-chip-deal-and-it-could-change-everything-in-the-tech-space" target="_blank" rel="noopener"
 &gt;TechRadar: Meta and Google could sign a mega AI chip deal&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://techenclave.com/news/items/microsoft-announces-maia-300-ai-chip-and-multi-gigawatt-deployment-plans-for-2027" target="_blank" rel="noopener"
 &gt;TechEnclave: Microsoft announces Maia 300&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.koreatimes.co.kr/business/companies/20260127/sk-hynix-becomes-sole-supplier-of-hbm-for-microsoft-ai-chips-sources" target="_blank" rel="noopener"
 &gt;Korea Times: SK hynix becomes sole HBM supplier for Microsoft AI chips&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.trendforce.com/news/2025/12/01/news-samsung-reportedly-supplies-60-of-google-tpu-hbm3e-set-to-remain-primary-supplier-in-2026/" target="_blank" rel="noopener"
 &gt;TrendForce: Samsung reportedly supplies 60% of Google TPU HBM3E&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.trendforce.com/news/2025/12/03/news-samsung-hbm4-reportedly-beats-expectations-in-broadcom-test-set-to-lead-2026-google-tpu-supply/" target="_blank" rel="noopener"
 &gt;TrendForce: Samsung HBM4 beats expectations in Broadcom test&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.trendforce.com/news/2025/06/20/news-sk-hynix-reportedly-leads-custom-hbm4e-push-with-nvidia-microsoft-samsung-trails-with-hbm4/" target="_blank" rel="noopener"
 &gt;TrendForce: SK hynix leads custom HBM4E push&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.trendforce.com/presscenter/news/20260918-13241.html" target="_blank" rel="noopener"
 &gt;TrendForce news (2026-09-18)&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.tomshardware.com/tech-industry/google-reportedly-books-intel-for-more-than-3-million-tpus-in-2028" target="_blank" rel="noopener"
 &gt;Tom&amp;rsquo;s Hardware: Google reportedly books Intel for more than 3 million TPUs in 2028&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.sammobile.com/news/samsung-will-exclusively-power-your-future-chatgpt-conversations/" target="_blank" rel="noopener"
 &gt;SamMobile: Samsung and OpenAI HBM supply&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://nand-research.com/research-note-qualcomm-introduces-ai200-ai250-for-data-center-inference/" target="_blank" rel="noopener"
 &gt;NAND Research: Qualcomm AI200 and AI250&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.tomshardware.com/pc-components/gpus/intel-unveils-crescent-island-an-inference-only-gpu-with-xe3p-architecture-and-160gb-of-memory" target="_blank" rel="noopener"
 &gt;Tom&amp;rsquo;s Hardware: Intel Crescent Island&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://profesionalreview.com/2026/10/02/nvidia-controlara-37-memoria-hbm-2027" target="_blank" rel="noopener"
 &gt;Profesional Review: NVIDIA to control 37% of HBM in 2027&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;&lt;strong&gt;Two questions for the comments.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;1. Who will NVIDIA name as an NVHBM or Fusion partner at GTC DC on December 2? A. Neither Google nor Meta. B. Meta only. C. Google only. D. Both.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;2. Which memory maker has the most to lose from standardized base dies? A. SK hynix. B. Samsung. C. Micron. D. None of them, because this is all price power at NVIDIA.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;This is for research and information only, not investment advice. Probabilities, ranges and signs are analytical assumptions, not trading instructions or individual portfolio decisions. Figures are dated data. Anthropic, which makes the Claude model we used to collect sources and edit this post, is a large user of Google TPUs and AWS Trainium, both discussed here. We used public material only and do not judge the size of that effect.&lt;/p&gt;</description></item><item><title>If a 2019-Style DRAM Glut Returns, It Lands in 2029: Shrinks Add Half the Bits They Did, and HBM Takes 30% of 2027 Wafers</title><link>https://koreainvestinsights.com/post/dram-glut-2019-rhymes-in-2029-hbm-takes-the-wafers-2026-11-08/</link><pubDate>Sun, 08 Nov 2026 19:15:00 +0900</pubDate><guid>https://koreainvestinsights.com/post/dram-glut-2019-rhymes-in-2029-hbm-takes-the-wafers-2026-11-08/</guid><description>&lt;img src="https://koreainvestinsights.com/" alt="Featured image of post If a 2019-Style DRAM Glut Returns, It Lands in 2029: Shrinks Add Half the Bits They Did, and HBM Takes 30% of 2027 Wafers" /&gt;&lt;p&gt;Memory downturns used to follow one script. In a boom, makers add supply. When demand growth slows, node shrinks keep adding bits anyway. The gap shows up as falling prices. 2018–2019 was the textbook case. According to Micron remarks (secondary), industry DRAM bits grew &lt;strong&gt;20%+&lt;/strong&gt; in 2018 with almost no new fabs. When the 2019 demand outlook slowed to the mid-teens, Micron&amp;rsquo;s average DRAM selling price fell &lt;strong&gt;about 30%&lt;/strong&gt;.&lt;/p&gt;
&lt;p&gt;The market fears the same script now. Our answer: 2028 is different, and 2029 may not be. Two numbers changed since 2018. Shrinks now add &lt;strong&gt;6–11%&lt;/strong&gt; of bits a year, about half the 2018 pace (our estimate). And HBM is set to use &lt;strong&gt;about 30%&lt;/strong&gt; of DRAM wafers in 2027 (TrendForce). HBM needs about three times the wafer area of conventional DRAM for the same bits.&lt;/p&gt;
&lt;p&gt;So the variable that sets supply this cycle is HBM bit demand, more than shrinks. Conventional DRAM supply comes from the wafers HBM leaves behind. If 2028 HBM bits grow 29% to 43%, conventional DRAM supply grows about 11% to 20%, overlapping or below a demand range of 12% to 20% (our assumption). The risky year is 2029, when new cleanrooms bunch up and HBM growth slows. On that path conventional supply grows 21% to 34%, as much as in 2018 or more.&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;On the supply side, shrinks drove the 2019 glut. Per Micron remarks (secondary), bits grew 20%+ with almost no new wafers. The gap to demand was about 5 points on the outlook at the time, yet prices fell about 30%.&lt;/li&gt;
&lt;li&gt;Shrinks now add 6–11% a year. New cleanrooms supply more than half of the growth, and a new cleanroom adds supply only when its owner decides to install tools.&lt;/li&gt;
&lt;li&gt;2028 HBM bits grow +29% in our central case, and memory makers&amp;rsquo; wafer allocation to HBM sets that number. Conventional supply then grows 14% to 20%, overlapping the demand range of 12% to 20%.&lt;/li&gt;
&lt;li&gt;If 2028 HBM bit growth comes in below +20%, the glut moves into 2028. That +20% line is the first thing to watch.&lt;/li&gt;
&lt;li&gt;There is a counterpoint. Micron&amp;rsquo;s inventory days are already above 2018, and on an annual basis they equal the 147 days of fiscal 2019.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id="the-old-explanation-fit-2019-and-it-only-half-fits-this-cycle"&gt;The Old Explanation Fit 2019, and It Only Half Fits This Cycle
&lt;/h2&gt;&lt;p&gt;We first checked the old explanation, &amp;ldquo;demand slows, shrinks keep adding bits, and the gap becomes a price drop,&amp;rdquo; line by line.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Claim&lt;/th&gt;
 &lt;th&gt;2019&lt;/th&gt;
 &lt;th&gt;This cycle&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Makers add supply into a boom&lt;/td&gt;
 &lt;td&gt;True. They invested in shrinks in 2017–2018&lt;/td&gt;
 &lt;td&gt;True. They build cleanrooms for 2027–2029, but tools go in step by step as demand allows&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Shrinks keep adding bits&lt;/td&gt;
 &lt;td&gt;True. Bits grew 20%+ with almost no new wafers&lt;/td&gt;
 &lt;td&gt;Weaker. 6–11% a year, about half of 2018&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;The demand-supply gap becomes a price drop&lt;/td&gt;
 &lt;td&gt;True. Demand mid-teens, average selling price about −30%&lt;/td&gt;
 &lt;td&gt;True, but how many wafers HBM takes sets the size of conventional supply&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;&amp;ldquo;Supply is constrained this time&amp;rdquo; proves this time is different&lt;/td&gt;
 &lt;td&gt;The same words were said in 2018&lt;/td&gt;
 &lt;td&gt;Words prove nothing. The size of the shrink engine and the HBM draw are different numbers&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;NAND is the same&lt;/td&gt;
 &lt;td&gt;&lt;/td&gt;
 &lt;td&gt;NAND still adds many bits through stacking (Micron G11 has about 2x the bits of G9), so the old explanation still fits&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;What changed is the engine that adds bits and the demand that takes those bits first. The sections below go through each.&lt;/p&gt;
&lt;h2 id="in-2019-bits-grew-20-with-almost-no-new-fabs"&gt;In 2019, Bits Grew 20%+ With Almost No New Fabs
&lt;/h2&gt;&lt;p&gt;Start with what happened in 2019. Industry bits are the outlook at the time from copies of Micron&amp;rsquo;s calls (secondary). Prices and inventory come from Micron&amp;rsquo;s filings (primary).&lt;/p&gt;
&lt;p&gt;&lt;img alt="The 2019 cycle. Industry bit growth: 2018 supply 20%+, 2018 demand about 20%, 2019 demand mid-teens (outlook at the time). Micron DRAM average selling price: FY2018 about +35%, FY2019 about −30%. Micron inventory days: FY2017 96, FY2018 105, FY2019 147." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/dram-glut-2019-rhymes-in-2029-hbm-takes-the-wafers-2026-11-08/cycle-2019.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;Supply kept coming from shrinks, 20%+ without new fabs. In 2018 Micron said it would hold wafer capacity and invest in node transitions (call copy, secondary). On the demand side, cloud customers digested inventory, smartphones were weak and CPUs were short. Micron&amp;rsquo;s fiscal 2019 annual report (10-K) attributed the price drop to a supply-demand imbalance, customer inventory adjustments and CPU shortages.&lt;sup id="fnref:1"&gt;&lt;a href="#fn:1" class="footnote-ref" role="doc-noteref"&gt;1&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;On the outlook at the time, the gap between demand and supply was about 5 points. Prices still fell about 30%. Producer inventory absorbed the gap. In fiscal 2019 Micron&amp;rsquo;s inventory rose 42% while cost of goods sold rose 2%.&lt;sup id="fnref:2"&gt;&lt;a href="#fn:2" class="footnote-ref" role="doc-noteref"&gt;2&lt;/a&gt;&lt;/sup&gt; Producers cut wafer starts and capex only after prices had already fallen.&lt;/p&gt;
&lt;p&gt;One more sentence is worth noting. In September 2018 Micron said industry supply growth from node transitions was slowing structurally (call copy, secondary). The supply-constraint story existed in 2018 too. So &amp;ldquo;supply is constrained this time&amp;rdquo; is not by itself evidence that this time is different. To say that, the numbers have to differ.&lt;/p&gt;
&lt;h2 id="shrinks-now-add-half-the-bits-they-did-in-2018"&gt;Shrinks Now Add Half the Bits They Did in 2018
&lt;/h2&gt;&lt;p&gt;A shrink makes circuits smaller so one wafer holds more memory cells. Bits grow without a new fab. That is why, as in 2018, shrink-driven supply is hard to stop once it starts. Split bit growth by source for 2018 and 2028, and it looks like this.&lt;/p&gt;
&lt;p&gt;&lt;img alt="Where industry DRAM bit growth comes from. 2018: shrinks 15–20% and almost no new wafers, for industry bits of 20%+. 2028: new wafers +15.2% and shrinks 6–11%, about +24% before HBM. About 30% of 2027 wafers go to HBM." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/dram-glut-2019-rhymes-in-2029-hbm-takes-the-wafers-2026-11-08/engine-shift.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;Total bits can still grow in the 20s in 2028. Micron also sees 2028 industry DRAM bit growth in the low 20s. But most of that growth comes from new wafers. New wafers need new cleanrooms and tools, and HBM takes part of them first.&lt;/p&gt;
&lt;p&gt;The density gain per node went from 23–25% for 1x to 1y up to above 30% for 1γ (company figure), and falls back to the mid-20s for 1δ, one of the last 6F² generations. The bigger change is the tooling each transition needs.&lt;/p&gt;
&lt;p&gt;&lt;img alt="Bit density gain per node and EUV layers. Density gain: 1x to 1y (2018) 23–25%, 1γ above 30% (company), 1δ mid-20s%. EUV layers: none in 2018, Micron 1γ 1, SK hynix 1b 4, 1c 5–6, Samsung 1c 6–7. Bits added by shrinks in one year fell from 15–20% in 2017–2019 to 6–11% in 2026–2028." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/dram-glut-2019-rhymes-in-2029-hbm-takes-the-wafers-2026-11-08/node-engine.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;EUV lithography tools are expensive and take a lot of cleanroom space. More layers mean more tools and floor space per transition, and slower transitions. Process steps also differ more than ever between products made on the same node (HBM, mobile LPDDR, server DDR5). Micron&amp;rsquo;s chief technology officer said the process gap between products on one node is the largest in DRAM history.&lt;/p&gt;
&lt;p&gt;The makers say the same. Micron said node transitions yield less than they used to.&lt;sup id="fnref:3"&gt;&lt;a href="#fn:3" class="footnote-ref" role="doc-noteref"&gt;3&lt;/a&gt;&lt;/sup&gt; SK hynix said there is a limit to adding bits through shrinks. Samsung said higher capex in 2027 does not translate well into 2028 bit growth. The shared background is that the 6F² cell structure is close to its limit.&lt;/p&gt;
&lt;h2 id="supply-now-comes-from-new-cleanrooms-and-a-cleanroom-adds-supply-only-when-tools-go-in"&gt;Supply Now Comes From New Cleanrooms, and a Cleanroom Adds Supply Only When Tools Go In
&lt;/h2&gt;&lt;p&gt;With shrinks weaker, supply has to come from new wafers. Put the main new fab milestones on one page, and they bunch up from the second half of 2028 into 2029.&lt;/p&gt;
&lt;p&gt;&lt;img alt="Main new DRAM fab milestones. Micron ID1 first wafer mid-2027, Tongluo ships 2H27, ID2, Hiroshima and Singapore 2H28. SK hynix M15X full capacity end-2026, Yongin phase 1 starts 2027, Cheongju M17 cleanroom December 2028, Yongin phase 2 cleanroom June 2029. Samsung P5 late 2027–2028 (reports differ). CXMT from 320k wafers a month to 420k in 2027 and a 500–600k target by end-2028." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/dram-glut-2019-rhymes-in-2029-hbm-takes-the-wafers-2026-11-08/cleanrooms.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;Wafer capacity at the four memory makers, including CXMT, is estimated to grow 15.2% on average in 2028 (based on SemiAnalysis estimates). More than half of supply growth comes from new wafers.&lt;/p&gt;
&lt;p&gt;That creates one difference from 2018. In 2018 node transitions in existing fabs pushed bits out, and that supply was hard to stop. Now tool move-in sets supply. In its new fab investment disclosure, SK hynix said construction comes first and tools go in step by step as demand allows. Micron said more than half of its fiscal 2027 capex increase is cleanroom construction for use after late 2028. Producers can leave a building empty. That discretion sets the size of the 2029 supply wave.&lt;/p&gt;
&lt;p&gt;We also share new analyses on Telegram as they are published. You are welcome to join us at &lt;a class="link" href="https://t.me/korea_invest_insights" target="_blank" rel="noopener"
 &gt;@korea_invest_insights&lt;/a&gt;.&lt;/p&gt;
&lt;h2 id="hbm-uses-about-three-times-the-wafer-area-for-the-same-bits"&gt;HBM Uses About Three Times the Wafer Area for the Same Bits
&lt;/h2&gt;&lt;p&gt;HBM, high-bandwidth memory, stacks DRAM dies and sits right next to an AI accelerator. It needs far more wafer area than conventional DRAM for the same capacity.&lt;/p&gt;
&lt;p&gt;&lt;img alt="Bits from one wafer. One conventional DRAM wafer makes 3 units of bits; one HBM wafer makes 1. Through-silicon vias, wide I/O circuits and more banks make the die bigger. The ratio is about 3x for HBM3E and above 4x for HBM4E (Micron)." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/dram-glut-2019-rhymes-in-2029-hbm-takes-the-wafers-2026-11-08/trade-ratio.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;The ratio rises with each generation. Micron said it exceeds 4x for HBM4E, and that vertically stacked HBM again carries a very high ratio. The same HBM bits take more wafers every year.&lt;/p&gt;
&lt;p&gt;The result is the pair of numbers below. In 2027 HBM is set to use 30% of wafers and make 13% of bits.&lt;/p&gt;
&lt;p&gt;&lt;img alt="HBM share of the three makers’ DRAM wafer starts: 18% in 2025, 22% in 2026, 30% in 2027. HBM’s share of 2027 bits is about 13%. HBM makes 13 units of bits on 30 units of wafers, conventional DRAM makes 87 on 70. For the same bits HBM uses about 2.9x the wafers." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/dram-glut-2019-rhymes-in-2029-hbm-takes-the-wafers-2026-11-08/hbm-share.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;The HBM wafer share is a TrendForce estimate.&lt;sup id="fnref:4"&gt;&lt;a href="#fn:4" class="footnote-ref" role="doc-noteref"&gt;4&lt;/a&gt;&lt;/sup&gt; The 13% bit share for 2027 is our model value. The ratio between them drives the arithmetic in the next section.&lt;/p&gt;
&lt;h2 id="conventional-dram-supply-is-what-hbm-leaves-behind"&gt;Conventional DRAM Supply Is What HBM Leaves Behind
&lt;/h2&gt;&lt;p&gt;The arithmetic is simple. Set end-2027 wafers at 100: 30 for HBM and 70 for conventional DRAM. In 2028 wafers grow 15.2% to 115.2. HBM takes its share first. The wafers left over, times the shrink gain (8% assumed), give conventional DRAM supply.&lt;/p&gt;
&lt;p&gt;&lt;img alt="End-2027 wafers of 100 (HBM 30, conventional 70) grow to 115.2 in 2028. With HBM bits +43%: HBM 42.9, conventional 72.3, conventional wafers +3.3%, conventional bit supply +11.5%. With +29%: HBM 38.7, conventional 76.5, +9.3%, +18.0%. With +10%: HBM 33.0, conventional 82.2, +17.4%, +26.8%." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/dram-glut-2019-rhymes-in-2029-hbm-takes-the-wafers-2026-11-08/wafer-split.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;If HBM bits grow 43% (the same level as TrendForce&amp;rsquo;s 2028 shipment outlook), conventional wafers grow only 3.3%. With shrinks added, conventional supply grows 11.5%. If HBM bits grow only 10%, conventional wafers grow 17.4% and conventional supply grows 26.8%. Same fabs, same wafers, and HBM demand alone more than doubles the range of conventional supply.&lt;/p&gt;
&lt;p&gt;Draw the relationship as a line, and it looks like this. The green band is our assumption that conventional DRAM demand grows 12% to 20%.&lt;/p&gt;
&lt;p&gt;&lt;img alt="2028 HBM bit growth against 2028 conventional DRAM bit supply growth. On the 8% shrink line: HBM +10% gives conventional +26.8%, +29% gives +18.0%, +43% (TrendForce shipment outlook) gives +11.5%. The 6% and 11% shrink lines sit below and above. The band shows conventional demand of +12–20%, and a dashed line marks +20% HBM growth." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/dram-glut-2019-rhymes-in-2029-hbm-takes-the-wafers-2026-11-08/residual-curve.en.svg"&gt;
&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;2028 HBM bit growth&lt;/th&gt;
 &lt;th&gt;2028 HBM wafer share&lt;/th&gt;
 &lt;th&gt;Conventional supply (shrinks 6%)&lt;/th&gt;
 &lt;th&gt;Conventional supply (shrinks 8%)&lt;/th&gt;
 &lt;th&gt;Conventional supply (shrinks 11%)&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;+43%&lt;/td&gt;
 &lt;td&gt;37.2%&lt;/td&gt;
 &lt;td&gt;+9.5%&lt;/td&gt;
 &lt;td&gt;+11.5%&lt;/td&gt;
 &lt;td&gt;+14.6%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;+29%&lt;/td&gt;
 &lt;td&gt;33.6%&lt;/td&gt;
 &lt;td&gt;+15.8%&lt;/td&gt;
 &lt;td&gt;+18.0%&lt;/td&gt;
 &lt;td&gt;+21.3%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;+20%&lt;/td&gt;
 &lt;td&gt;31.3%&lt;/td&gt;
 &lt;td&gt;+19.9%&lt;/td&gt;
 &lt;td&gt;+22.2%&lt;/td&gt;
 &lt;td&gt;+25.6%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;+10%&lt;/td&gt;
 &lt;td&gt;28.6%&lt;/td&gt;
 &lt;td&gt;+24.5%&lt;/td&gt;
 &lt;td&gt;+26.8%&lt;/td&gt;
 &lt;td&gt;+30.3%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;0%&lt;/td&gt;
 &lt;td&gt;26.0%&lt;/td&gt;
 &lt;td&gt;+29.0%&lt;/td&gt;
 &lt;td&gt;+31.5%&lt;/td&gt;
 &lt;td&gt;+35.1%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;We computed the table values with the same formula. The +43% row puts TrendForce&amp;rsquo;s 2028 shipment outlook on a bit path. The table holds HBM bits per wafer at the 2027 level. As HBM4E lowers bits per wafer, the same HBM bits take more wafers, and conventional supply comes in below the table.&lt;/p&gt;
&lt;p&gt;For supply to match conventional demand growth of 12% to 20%, HBM bits must grow 20% to 37% with 6% shrinks. TrendForce sees 2028 HBM shipments up 42.9% (as cited by Forbes Japan, secondary). On that path, 2028 conventional supply sits inside the demand range or below it. For a 2019-style glut to arrive in 2028, HBM bit growth has to fall below +20%.&lt;/p&gt;
&lt;h2 id="2028-hbm-bits-are-set-by-memory-makers-wafer-allocation"&gt;2028 HBM Bits Are Set by Memory Makers&amp;rsquo; Wafer Allocation
&lt;/h2&gt;&lt;p&gt;What sets 2028 HBM bits, then? HBM bits are the number of accelerators times the HBM in each one. Each accelerator already carries a lot, and every generation carries more.&lt;/p&gt;
&lt;p&gt;&lt;img alt="HBM per accelerator. AMD MI455X 432 GB (twelve 12-high HBM4 stacks), NVIDIA Rubin 288 GB (eight 12-high HBM4 stacks), Google TPU v8i 288 GB, Microsoft Maia 300 288 GB, Google TPU v8t 216 GB. NVIDIA Rubin Ultra (2027) planned 12-high HBM4E; 8-high options are under review since the third quarter, and the design is not final." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/dram-glut-2019-rhymes-in-2029-hbm-takes-the-wafers-2026-11-08/accelerators.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;We split 2028 HBM bits into two formulas. The first is a demand ceiling: advanced packaging wafers (TSMC CoWoS and others), which cap how many accelerators can be built, times GB per stack. Analysts estimate TSMC CoWoS capacity at 130,000 wafers a month at end-2026 and 260,000 at end-2028 (not TSMC guidance).&lt;sup id="fnref:5"&gt;&lt;a href="#fn:5" class="footnote-ref" role="doc-noteref"&gt;5&lt;/a&gt;&lt;/sup&gt; The second is supply: the share of new DRAM wafers allocated to HBM times HBM bits per wafer. The bits that reach the market are the lower of the two.&lt;/p&gt;
&lt;p&gt;&lt;img alt="2028 HBM bit growth. Demand ceiling (advanced packaging times GB per stack): central +53%, range +21% to +92%. Supply (HBM wafer allocation times bits per wafer): central +29%, range +7% to +53%. Realized is the lower of the two, +29%. Below +20%, the conventional glut moves into 2028." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/dram-glut-2019-rhymes-in-2029-hbm-takes-the-wafers-2026-11-08/hbm28-limits.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;In the base case supply binds, because the packaging ceiling (+53%) sits above supply (+29%). So the value that most directly sets 2028 HBM bits is the share of new wafers the three makers give to HBM. Our central assumption is 70%. At 50% HBM bits grow 19%, and at 90% they grow 38% (with bits per wafer at −5%). On the central path HBM wafers grow 35% and conventional supply grows about 14%. Matching Micron&amp;rsquo;s industry bit outlook (low 20s) needs a larger shrink gain for conventional DRAM, and conventional supply then grows about 20%. So we put conventional supply on the central path at 14% to 20%. One Korean broker puts 2028 HBM bit growth at about +50% for SK hynix and +55% for Samsung. At industry averages that would need more wafers for HBM than all the new wafers combined, so it sits outside our range.&lt;/p&gt;
&lt;p&gt;We think profit per wafer sets the allocation. Makers give wafers to the product that earns more per wafer. TrendForce said that in the first quarter of 2026 a 64 GB server DDR5 module earned more revenue per wafer than HBM, and that this is the basis for 2027 HBM price increases.&lt;sup id="fnref1:4"&gt;&lt;a href="#fn:4" class="footnote-ref" role="doc-noteref"&gt;4&lt;/a&gt;&lt;/sup&gt; If 2028 HBM contract prices keep per-wafer profit at or above conventional DRAM, the allocation holds.&lt;/p&gt;
&lt;p&gt;How should NVIDIA&amp;rsquo;s look at 8-high stacks be read? TrendForce reported that Rubin Ultra is evaluating 8-high stacks instead of 12-high.&lt;sup id="fnref:6"&gt;&lt;a href="#fn:6" class="footnote-ref" role="doc-noteref"&gt;6&lt;/a&gt;&lt;/sup&gt; 8-high stacks let the same HBM bits serve more GPUs. When HBM is short, accelerator counts rise as per-unit content falls, and total HBM bits hold. The risk is per-unit content falling while accelerator counts do not rise. By our arithmetic, demand starts to set HBM bits only if accelerator orders fall about 16% short of packaging plans. That is an AI-spending slowdown path, and its first signal would be cuts to hyperscaler 2028 capex.&lt;/p&gt;
&lt;p&gt;Micron said it can run several products on the same line and shift the mix toward demand.&lt;sup id="fnref1:3"&gt;&lt;a href="#fn:3" class="footnote-ref" role="doc-noteref"&gt;3&lt;/a&gt;&lt;/sup&gt; Wafers that HBM absorbs can return to conventional DRAM. Because process steps differ by product, we expect a switch to take a quarter or more (our estimate).&lt;/p&gt;
&lt;h2 id="2029-looks-like-2019"&gt;2029 Looks Like 2019
&lt;/h2&gt;&lt;p&gt;In 2029 the fab schedule above overlaps. Micron&amp;rsquo;s ID2, Hiroshima and Singapore (2H28), SK hynix&amp;rsquo;s Cheongju M17 (December 2028) and Yongin phase 2 (June 2029), and CXMT&amp;rsquo;s end-2028 target of 500,000–600,000 wafers a month (secondary report) all overlap. We set 2029 wafer growth at 15% and ran paths where HBM growth slows.&lt;/p&gt;
&lt;p&gt;&lt;img alt="Conventional DRAM bit supply growth. HBM bits +43% in 2028 and +30% in 2029: conventional +11.5% in 2028 and +20.8% in 2029 (2029 HBM wafer share 39%). +29% then +20%: +18.0% and +26.3% (32%). +29% then +5%: +18.0% and +33.9% (28%). A dashed line marks 2018 industry supply growth of 20%+." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/dram-glut-2019-rhymes-in-2029-hbm-takes-the-wafers-2026-11-08/rhyme-2029.en.svg"&gt;
&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Path (2028 / 2029 HBM bits)&lt;/th&gt;
 &lt;th&gt;2028 conventional supply&lt;/th&gt;
 &lt;th&gt;2029 conventional supply&lt;/th&gt;
 &lt;th&gt;2029 HBM wafer share&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;+43% / +30%&lt;/td&gt;
 &lt;td&gt;+11.5%&lt;/td&gt;
 &lt;td&gt;+20.8%&lt;/td&gt;
 &lt;td&gt;39%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;+29% / +20%&lt;/td&gt;
 &lt;td&gt;+18.0%&lt;/td&gt;
 &lt;td&gt;+26.3%&lt;/td&gt;
 &lt;td&gt;32%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;+29% / +5%&lt;/td&gt;
 &lt;td&gt;+18.0%&lt;/td&gt;
 &lt;td&gt;+33.9%&lt;/td&gt;
 &lt;td&gt;28%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;On all three paths, 2029 conventional supply grows as much as 2018 supply (20%+) or more. 2029 is the year with the highest risk of a 2019-style glut.&lt;/p&gt;
&lt;p&gt;There are buffers. Producers can delay tool move-in. SK hynix said long-term agreements with 14 customers cover more than 50% of its 2027–2028 capacity (call summary, secondary). Conventional demand also responds to price. High prices have cut memory content in smartphones and PCs, so lower prices can lift content again and absorb part of a glut.&lt;/p&gt;
&lt;p&gt;The order of events is the problem. In 2019 Micron cut wafer starts only after prices fell. If the order repeats, 2029 prices move first. Long-term agreements do not remove the risk either. Take-or-pay contracts move demand risk into buyer inventory. We covered how buyer inventory is part of the DRAM shortage in &lt;a class="link" href="https://koreainvestinsights.com/post/dram-shortage-part-inventory-buyers-stop-when-cpus-arrive-2026-10-30/" &gt;an earlier post&lt;/a&gt;.&lt;/p&gt;
&lt;h2 id="the-counterpoint-producer-inventory-is-already-high"&gt;The Counterpoint: Producer Inventory Is Already High
&lt;/h2&gt;&lt;p&gt;Some argue that producer inventory is low this time, so this time is different. We recomputed Micron&amp;rsquo;s inventory days from SEC filings.&lt;/p&gt;
&lt;p&gt;&lt;img alt="Micron inventory days (fiscal year-end inventory divided by annual cost of goods sold, times 365). FY16 107, FY17 96, FY18 105, FY19 147, FY20 132, FY21 95, FY22 144, FY23 181, FY24 166, FY25 136, FY26 147. On fourth-quarter cost of goods sold: FY18 99, FY19 134, FY26 142." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/dram-glut-2019-rhymes-in-2029-hbm-takes-the-wafers-2026-11-08/inventory-days.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;At the end of fiscal 2026 inventory days were 147 on an annual basis, equal to the end of fiscal 2019 and above the end of fiscal 2018 (105). This method reads high in a year when cost of goods sold grows fast, so we also used fourth-quarter cost of goods sold. On that basis fiscal 2026 ended at 142 days, still above 99 at the end of fiscal 2018.&lt;sup id="fnref1:2"&gt;&lt;a href="#fn:2" class="footnote-ref" role="doc-noteref"&gt;2&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;HBM has a long process, so more of it sits in work in progress. Without that, inventory days would be lower. Even so, the claim that low producer inventory makes this time different is weak.&lt;/p&gt;
&lt;h2 id="the-process-engine-may-return-around-2030"&gt;The Process Engine May Return Around 2030
&lt;/h2&gt;&lt;p&gt;This analysis rests on a weak process engine. We also need to know when that premise changes.&lt;/p&gt;
&lt;p&gt;&lt;img alt="DRAM, HBM and NAND technology timing. DRAM 6F² shrinks: 1γ (2026), 1δ/1d (2H27–2028). New structures: 4F²/VCT (Samsung 10a target 2028, uncertain), 3D DRAM (around 2030, estimated). HBM: HBM4 (2026), HBM4E (2027), HBM5 and vertical HBM (2028). NAND: G11 with 400+ layers and about 2x bits (2H27–2028)." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/dram-glut-2019-rhymes-in-2029-hbm-takes-the-wafers-2026-11-08/roadmap.en.svg"&gt;
&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;4F² and VCT (vertical channel transistor): Samsung targets mass production of its 10a process in 2028 and claims 30% to 50% higher cell density. The schedule is uncertain. Cells and peripheral circuits are made on separate wafers and bonded, so bits per wafer start may gain less than cell area.&lt;/li&gt;
&lt;li&gt;3D DRAM: cells are stacked vertically. Samsung reportedly plans to apply it from 10d, and we estimate the timing at around 2030. Once in mass production, it creates an engine that adds bits by adding layers, as in NAND. That is when the old explanation, process transitions push supply, fits DRAM again.&lt;/li&gt;
&lt;li&gt;NAND: the old explanation still fits NAND today. Micron&amp;rsquo;s G11 (400+ layers) has about twice the bits of G9. The spot price of a 512 Gb TLC NAND wafer fell 30%, from $23.00 on March 18 to $16.15 on November 2.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;SK hynix&amp;rsquo;s position is to stretch 6F² with EUV as long as possible. So the process engine is likely to stay weak in 2027–2028.&lt;/p&gt;
&lt;h2 id="what-would-flip-this-call-and-what-to-check-first"&gt;What Would Flip This Call, and What to Check First
&lt;/h2&gt;&lt;p&gt;These conditions would flip the call.&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;If 2028 HBM bit growth comes in below +20%, a 2019-style glut moves into 2028. That happens if lower HBM per accelerator is confirmed and accelerator counts do not rise to match, or if 2028 HBM contracts are weak.&lt;/li&gt;
&lt;li&gt;If tool move-in at the 2H28 cleanrooms is delayed, the 2029 supply wave gets smaller.&lt;/li&gt;
&lt;li&gt;If 4F² reaches mass production in 2028 as planned, the process engine returns early and 2029 supply exceeds the table above.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;&lt;img alt="What to check, in order. Mid-December: Micron FQ1 FY27 (date not set) for the 2028 HBM and conventional bit outlook, 2028 contracts and tool move-in. Mid-January: TSMC 4Q26, whether CoWoS is around 200,000 wafers a month at end-2027. Late January: SK hynix and Samsung 4Q26, whether 2028 HBM capacity grows about 30%. First half of 2027: Rubin Ultra HBM configuration. Second half of 2027: 2028 HBM contracts, whether bits under +20% come with price cuts." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/dram-glut-2019-rhymes-in-2029-hbm-takes-the-wafers-2026-11-08/signposts.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;The first check is Micron&amp;rsquo;s earnings, expected in mid-December (date not set). There we look for 2028 HBM bit growth and progress on 2028 contracts. Then, in late January, SK hynix and Samsung results show whether 2028 HBM capacity grows about 30%. Above +40% points to a larger HBM allocation, and +20% or less points to a smaller one.&lt;/p&gt;
&lt;p&gt;This call revisits, from the supply side, the conclusion in our &lt;a class="link" href="https://koreainvestinsights.com/post/ai-demand-provisional-verdict-2027-supply-capped-2028-bill-2026-11-01/" &gt;provisional verdict on AI demand&lt;/a&gt; that the bill lands in 2028. If the demand bill lands in 2028 orders, the bill for conventional DRAM prices lands in 2029 supply.&lt;/p&gt;
&lt;h2 id="data-notes"&gt;Data Notes
&lt;/h2&gt;&lt;ul&gt;
&lt;li&gt;Industry bit supply and demand for 2017–2020 come from copies of Micron call remarks (secondary) and mix the outlook at the time with actual results. We did not find industry wafer capacity growth for those years.&lt;/li&gt;
&lt;li&gt;The 15–20% shrink contribution in 2018 is inferred from &amp;ldquo;bits up 20%+ with almost no new wafers.&amp;rdquo; If other makers added wafers, the shrink contribution was smaller.&lt;/li&gt;
&lt;li&gt;HBM wafer share, the 2028 HBM shipment outlook, SK hynix&amp;rsquo;s long-term agreements and Samsung&amp;rsquo;s 10a schedule come from secondary reports. TrendForce&amp;rsquo;s 2028 HBM +42.9% is on a shipment basis, not bits. CoWoS capacity figures are analyst and industry estimates, not TSMC guidance.&lt;/li&gt;
&lt;li&gt;The supply model moves about ±6 points with the wafer, shrink and HBM ratio assumptions. 2029 wafer growth of 15% is an assumption. The 2029 math also lets HBM bits per wafer improve 8%, the same as shrinks, while the 2028 math holds them flat. The HBM allocation math uses four-maker wafer growth including CXMT as an approximation; CXMT wafers do not go into Western HBM. Conventional demand growth of 12% to 20% is our assumption.&lt;/li&gt;
&lt;li&gt;Micron&amp;rsquo;s fiscal fourth-quarter 2026 call remarks come from a third-party transcript.&lt;sup id="fnref2:3"&gt;&lt;a href="#fn:3" class="footnote-ref" role="doc-noteref"&gt;3&lt;/a&gt;&lt;/sup&gt; We computed inventory days from SEC XBRL data (retrieved November 8).&lt;/li&gt;
&lt;li&gt;If substrate shortages cap advanced packaging output, the HBM demand ceiling falls. In our arithmetic the ceiling still sits above supply, so the call is the same.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;&lt;strong&gt;Will 2028 HBM bit growth land above or below the +20% line? Reply with the one piece of evidence that decides it for you.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;This is for research and information only, not investment advice. Supply paths and thresholds are analytical assumptions, not trading instructions or individual portfolio decisions. Figures are dated data.&lt;/p&gt;
&lt;p&gt;We built the supply model with AI and used AI for research and editing, working with Anthropic&amp;rsquo;s Claude and OpenAI GPT models. The accelerator chart includes the TPU of Alphabet (Google), an investor in Anthropic, and the Maia of Microsoft, an investor in OpenAI; Anthropic is also a large user of Google TPUs. This post does not assess those companies&amp;rsquo; businesses or value. It shows only their disclosed HBM content.&lt;/p&gt;
&lt;div class="footnotes" role="doc-endnotes"&gt;
&lt;hr&gt;
&lt;ol&gt;
&lt;li id="fn:1"&gt;
&lt;p&gt;Micron Technology, &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/723125/000072312519000094/a2019q4.htm" target="_blank" rel="noopener"
 &gt;Form 10-K, fiscal year 2019&lt;/a&gt;. FQ1 FY2019 earnings release: &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/723125/000072312518000152/a2019q1exhibit991-pressrel.htm" target="_blank" rel="noopener"
 &gt;8-K Exhibit 99.1&lt;/a&gt;. FQ4 FY2019 earnings release: &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/723125/000072312519000068/a2019q4exhibit991-pressrel.htm" target="_blank" rel="noopener"
 &gt;8-K Exhibit 99.1&lt;/a&gt;.&amp;#160;&lt;a href="#fnref:1" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:2"&gt;
&lt;p&gt;SEC, &lt;a class="link" href="https://data.sec.gov/api/xbrl/companyfacts/CIK0000723125.json" target="_blank" rel="noopener"
 &gt;XBRL companyfacts: Micron Technology&lt;/a&gt;. Computed from InventoryNet and CostOfGoodsAndServicesSold, retrieved 2026-11-08.&amp;#160;&lt;a href="#fnref:2" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:2" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:3"&gt;
&lt;p&gt;The Motley Fool, &lt;a class="link" href="https://www.fool.com/earnings/call-transcripts/2026/10/01/micron-mu-q4-2026-earnings-call-transcript/" target="_blank" rel="noopener"
 &gt;Micron (MU) Q4 2026 earnings call transcript&lt;/a&gt;, 2026-10-01.&amp;#160;&lt;a href="#fnref:3" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:3" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref2:3" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:4"&gt;
&lt;p&gt;TrendForce, &lt;a class="link" href="https://www.trendforce.com/presscenter/news/20260602-13074.html" target="_blank" rel="noopener"
 &gt;HBM wafer share and revenue per wafer&lt;/a&gt;, 2026-06-02. 2027 HBM outlook: &lt;a class="link" href="https://www.trendforce.com/presscenter/news/20260929-13255.html" target="_blank" rel="noopener"
 &gt;2026-09-29 release&lt;/a&gt;.&amp;#160;&lt;a href="#fnref:4" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:4" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:5"&gt;
&lt;p&gt;Wccftech, &lt;a class="link" href="https://wccftech.com/report-tsmc-to-double-cowos-capacity-by-2028-as-ai-chip-shortage-spills-over-to-rivals/" target="_blank" rel="noopener"
 &gt;TSMC to double CoWoS capacity by 2028&lt;/a&gt; (via udn, 2026-09-11); DIGITIMES, &lt;a class="link" href="https://www.digitimes.com/news/a20260710PD226/tsmc-cowos-2027-packaging-capacity.html" target="_blank" rel="noopener"
 &gt;TSMC CoWoS 2027 packaging capacity&lt;/a&gt; (2026-07-10).&amp;#160;&lt;a href="#fnref:5" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:6"&gt;
&lt;p&gt;TrendForce, &lt;a class="link" href="https://www.trendforce.com/presscenter/news/20260804-13166.html" target="_blank" rel="noopener"
 &gt;Rubin Ultra HBM configuration review&lt;/a&gt;, 2026-08-04, and &lt;a class="link" href="https://www.trendforce.com/news/2026/10/27/news-nvidias-memory-shake-up-rubin-ultra-rethinks-12-hi-hbm4e-for-8-hi-hbm4-rubin-gpu-bandwidth-cut-13/" target="_blank" rel="noopener"
 &gt;8-high HBM4 report&lt;/a&gt;, 2026-10-27.&amp;#160;&lt;a href="#fnref:6" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;/ol&gt;
&lt;/div&gt;</description></item><item><title>Samsung's Equity Passes Berkshire's by 2028 at a 50% Payout: Two Engines, Profit and Payouts, Decide Whether the Stock Follows</title><link>https://koreainvestinsights.com/post/samsung-two-engines-profit-and-payouts-equity-roe-2026-11-08/</link><pubDate>Sun, 08 Nov 2026 18:41:00 +0900</pubDate><guid>https://koreainvestinsights.com/post/samsung-two-engines-profit-and-payouts-equity-roe-2026-11-08/</guid><description>&lt;img src="https://koreainvestinsights.com/" alt="Featured image of post Samsung's Equity Passes Berkshire's by 2028 at a 50% Payout: Two Engines, Profit and Payouts, Decide Whether the Stock Follows" /&gt;&lt;p&gt;Samsung Electronics is worth &lt;strong&gt;₩1,230 trillion&lt;/strong&gt; at the &lt;strong&gt;₩188,700&lt;/strong&gt; close on November 6. On our base path, with half of free cash flow paid out, equity attributable to owners builds to &lt;strong&gt;₩1,140 trillion&lt;/strong&gt; by the end of 2028. That is about $840 billion, more than Berkshire Hathaway&amp;rsquo;s $747.9 billion, the largest equity base we found among listed companies.&lt;/p&gt;
&lt;p&gt;A big equity base does not buy a big market value. Four of the eight largest equity bases belong to China&amp;rsquo;s big four banks. Their equity of $463.6 billion to $647.0 billion is larger than Samsung&amp;rsquo;s current $442.5 billion. Their market values, $331 billion to $467 billion, are roughly a third to a half of Samsung&amp;rsquo;s $907 billion, and they trade at 0.56 to 0.82 times book value.&lt;/p&gt;
&lt;p&gt;Two engines decide where Samsung&amp;rsquo;s stock goes. The &lt;strong&gt;business engine&lt;/strong&gt; sets how much the company earns. The &lt;strong&gt;payout engine&lt;/strong&gt; sets how much of that profit stays inside the company. Together they set return on equity (ROE), and whether ROE clears investors&amp;rsquo; 10% required return sets the multiple the market pays. The market decides the first engine. The company decides the second, and its first answer comes at the board meeting in late January.&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;Market value is equity times price-to-book (P/B), and P/B depends on how far ROE sits above the 10% required return.&lt;/li&gt;
&lt;li&gt;On the base path net income falls from ₩448 trillion in 2027 to ₩90 trillion in 2030 (plus interest on retained cash) while equity grows from ₩914 trillion to ₩1,357 trillion, so ROE drops from 56% to 7.7%.&lt;/li&gt;
&lt;li&gt;About ₩650 trillion of the 2030 equity is retained cash earning roughly 3% after tax. That cash drags ROE below the required return.&lt;/li&gt;
&lt;li&gt;Raising the payout from 50% to 100% of free cash flow leaves 2030 equity at ₩847 trillion and lifts ROE to 11.1%, above 10%. The payout rate is the one big variable the company sets directly.&lt;/li&gt;
&lt;li&gt;Payouts do not create value. In our model, value rises only about ₩10,000 per share when the payout goes from 50% to 100%. What changes a lot is the frame the market uses to price the stock.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id="the-largest-equity-in-the-world-does-not-buy-the-largest-multiple"&gt;The Largest Equity in the World Does Not Buy the Largest Multiple
&lt;/h2&gt;&lt;p&gt;&lt;img alt="Equity attributable to owners of the largest equity bases, $ billion, with price-to-book and ROE. Samsung 2028 estimate 840 to 1,017, P/B 0.89 to 1.08, normalized ROE 7.7% to 11.1%. Berkshire 747.9, P/B 1.63, ROE 11.5%. Alphabet 677.5, 5.15x, 34.8%. ICBC 647.0, 0.56x, 8.6%. Amazon 613.6, 4.57x, 22.0%. China Construction Bank 569.7, 0.82x, 9.5%. Agricultural Bank of China 497.6, 0.77x, 10.1%. Microsoft 470.5, 8.37x, 30.3%. Bank of China 463.6, 0.71x, 8.7%. Samsung today 442.5, 2.1x, boom-year ROE about 50%. Aramco 436.0, 3.87x, ROACE 22.1%. JPMorgan 372.4, 2.39x, 17.7%." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/samsung-two-engines-profit-and-payouts-equity-roe-2026-11-08/peers.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;The bars show equity. The two right-hand columns show the multiple the market paid for that equity and the return the equity earns. Bar length and P/B follow no rule. The multiple follows ROE. China&amp;rsquo;s big four banks earn 8.6% to 10.1% and trade at 0.56 to 0.82 times book. Berkshire earns 11.5% and trades at 1.63 times. Alphabet, Amazon and Microsoft earn 22% to 35% and trade at 4.6 to 8.4 times.&lt;/p&gt;
&lt;p&gt;Samsung sits between the Chinese banks and Berkshire. If normalized ROE lands at 7.7% to 11.1%, the market multiple for similar returns is about one times book. Today, in a boom year, Samsung&amp;rsquo;s ROE is about 50%, yet it trades at 2.1 times book (on equity reported at September 30), below Alphabet&amp;rsquo;s 5.15 times at a 34.8% ROE.&lt;/p&gt;
&lt;h2 id="the-stock-price-is-a-chain-profit-equity-multiple"&gt;The Stock Price Is a Chain: Profit, Equity, Multiple
&lt;/h2&gt;&lt;p&gt;Think of a bathtub. Profit is the tap, payouts are the drain and equity is the water level. Investors do not pay for how high the water stands. They pay for how hard that water works. Unrolled, the chain reads like this.&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;ROE = net income ÷ equity&lt;/li&gt;
&lt;li&gt;P/B = (ROE − growth) ÷ (required return − growth). With growth set to 0 and a 10% required return, as in our model, P/B = ROE ÷ 10%.&lt;/li&gt;
&lt;li&gt;Market value = equity × P/B&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;When ROE equals the required return, P/B is 1. Put in ₩1, earn 10%, and investors ask for 10%: that ₩1 is worth exactly ₩1. If ROE is lower, the ₩1 is priced below ₩1. If it is higher, the ₩1 is priced at several won. China&amp;rsquo;s big banks are the first case. Alphabet and Microsoft are the second.&lt;/p&gt;
&lt;p&gt;&lt;img alt="Two engines move Samsung’s share price. Engine 1, business, sets normalized operating profit of ₩117 trillion in 2030 and is decided by the market. Engine 2, payouts, sets a payout rate of 33%, 50% or 100% of free cash flow and is decided by the company. Business becomes net income and payouts shape equity. Net income divided by equity is ROE, ROE divided by 10% is price-to-book, and equity times price-to-book is market value." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/samsung-two-engines-profit-and-payouts-equity-roe-2026-11-08/engines.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;The business engine moves the numerator. After the 2027 to 2029 boom, normalized operating profit settles near ₩117 trillion (2030). Memory prices and supply discipline set that level. The market sets it, not the company.&lt;/p&gt;
&lt;p&gt;The payout engine moves the denominator. Cash paid out does not pile up as equity. The payout rate is the variable the company sets directly.&lt;/p&gt;
&lt;h2 id="the-boom-piles-up-equity-and-roe-falls-with-it"&gt;The Boom Piles Up Equity, and ROE Falls With It
&lt;/h2&gt;&lt;p&gt;On the base path, the 2027 to 2029 boom profits pile up as equity. After normalization, operating profit returns to ₩117 trillion. The numerator shrinks while the denominator grows. This is the result with the current 50% payout policy kept in place.&lt;/p&gt;
&lt;p&gt;&lt;img alt="Year-end equity and ROE on the base path at a 50% payout. Equity is ₩914 trillion in 2027, ₩1,140 trillion in 2028, ₩1,293 trillion in 2029 and ₩1,357 trillion in 2030. ROE is 56%, 38%, 22% and 7.7%. Net income is ₩448 trillion, ₩377 trillion, ₩255 trillion and ₩90 trillion plus interest. The 2030 ROE of 7.7% sits below the 10% required return." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/samsung-two-engines-profit-and-payouts-equity-roe-2026-11-08/roe-path.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;The 2030 ROE of 7.7% sits below the 10% required return. The equity base is among the largest in the world, and it earns less than its cost.&lt;/p&gt;
&lt;h2 id="nearly-half-of-that-equity-is-cash-earning-about-3"&gt;Nearly Half of That Equity Is Cash Earning About 3%
&lt;/h2&gt;&lt;p&gt;Of the ₩1,357 trillion of equity in 2030, about ₩650 trillion is retained cash that the business does not use. That cash earns roughly 3% after tax. The capital that does work in the business earns far more (about 17% ROIC in the model).&lt;/p&gt;
&lt;p&gt;&lt;img alt="Makeup of ₩1,357 trillion of equity at the end of 2030. Operating capital earns far more, about 17% ROIC in the model. Retained cash of about ₩650 trillion earns about 3% after tax. On total equity the ROE is 7.7%, below the 10% required return." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/samsung-two-engines-profit-and-payouts-equity-roe-2026-11-08/retained-cash.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;Put on total book equity, the mix pulls ROE down to 7.7%. Every extra ₩1 trillion of retained cash earns only 3% while it enlarges the denominator. Left alone, the largest equity base turns into capital that does not earn its cost.&lt;/p&gt;
&lt;p&gt;We also share new analyses on Telegram as they are published. You are welcome to join us at &lt;a class="link" href="https://t.me/korea_invest_insights" target="_blank" rel="noopener"
 &gt;@korea_invest_insights&lt;/a&gt;.&lt;/p&gt;
&lt;h2 id="the-payout-engine-is-the-one-big-variable-the-company-sets-itself"&gt;The Payout Engine Is the One Big Variable the Company Sets Itself
&lt;/h2&gt;&lt;p&gt;Cash paid out does not pile up as equity. Here is the 2030 result when only the payout rate changes and the operating profit path stays the same.&lt;/p&gt;
&lt;p&gt;&lt;img alt="2030 ROE by payout rate of free cash flow. A 50% payout (current policy) gives equity of ₩1,357 trillion, ROE of 7.7% and a fair P/B of 0.77x. A 75% payout gives equity of ₩1,102 trillion, ROE of 9.0% and a fair P/B of 0.90x. A 100% payout gives equity of ₩847 trillion, ROE of 11.1% and a fair P/B of 1.11x. The required return is 10%." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/samsung-two-engines-profit-and-payouts-equity-roe-2026-11-08/payout-roe.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;At a 100% payout equity stops at ₩847 trillion, ROE reaches 11.1% above the required return, and P/B sits above 1. At 50%, equity is ₩1,357 trillion, ROE is 7.7% and P/B sits below 1. That is how a smaller equity base can carry a higher multiple. It is also why today&amp;rsquo;s ₩1,230 trillion market value is about one times 2028 equity. The market prices Samsung as a company with a large equity base that earns about its cost of capital.&lt;/p&gt;
&lt;p&gt;Equity at the end of 2028 depends on the payout rate and the profit path.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Equity attributable to owners, end of 2028, ₩ trillion&lt;/th&gt;
 &lt;th style="text-align: right"&gt;33% payout&lt;/th&gt;
 &lt;th style="text-align: right"&gt;50% payout&lt;/th&gt;
 &lt;th style="text-align: right"&gt;100% payout&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Lower profit path (independent anchor model)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;1,236&lt;/td&gt;
 &lt;td style="text-align: right"&gt;1,126&lt;/td&gt;
 &lt;td style="text-align: right"&gt;803&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Base profit path (sustainable price-and-volume model)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;1,255&lt;/td&gt;
 &lt;td style="text-align: right"&gt;1,140&lt;/td&gt;
 &lt;td style="text-align: right"&gt;803&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Broker consensus, FY27-28&lt;/td&gt;
 &lt;td style="text-align: right"&gt;1,360&lt;/td&gt;
 &lt;td style="text-align: right"&gt;1,219&lt;/td&gt;
 &lt;td style="text-align: right"&gt;803&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Higher profit path (cash-flow model)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;1,380&lt;/td&gt;
 &lt;td style="text-align: right"&gt;1,234&lt;/td&gt;
 &lt;td style="text-align: right"&gt;803&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;The claim that equity is at least ₩1,200 trillion even on conservative numbers comes with conditions. At a 50% payout only the consensus path and above clear ₩1,200 trillion, and our base path gives ₩1,140 trillion. At a 33% payout every path is ₩1,236 trillion or more. At a 100% payout every path ends at ₩803 trillion, below Berkshire. Today&amp;rsquo;s ₩1,230 trillion market value is about 0.9 to 1.1 times 2028 equity.&lt;/p&gt;
&lt;p&gt;The 2024 to 2026 return policy (50% of free cash flow) ends in 2026. The late-January board meeting sets the policy from 2027. Cash arriving in 2027 to 2029 comes to about ₩960 trillion of free cash flow. That one decision places 2030 equity somewhere between ₩847 trillion and ₩1,357 trillion, and 2030 ROE somewhere between 7.7% and 11.1%.&lt;/p&gt;
&lt;h2 id="the-business-engine-has-the-higher-ceiling-but-the-company-does-not-set-it"&gt;The Business Engine Has the Higher Ceiling, but the Company Does Not Set It
&lt;/h2&gt;&lt;p&gt;There are only two ways to raise ROE: shrink the denominator or grow the numerator. Payouts shrink the denominator. Higher profit and reinvestment of retained cash grow the numerator.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Lever&lt;/th&gt;
 &lt;th&gt;How it works&lt;/th&gt;
 &lt;th&gt;2030 ROE effect (base path)&lt;/th&gt;
 &lt;th&gt;Certainty&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;① Payouts (smaller denominator)&lt;/td&gt;
 &lt;td&gt;Cash goes back, so equity does not pile up&lt;/td&gt;
 &lt;td&gt;50% to 100% payout: 7.7% to 11.1%&lt;/td&gt;
 &lt;td&gt;High. The company decides&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;② Higher profit (bigger numerator)&lt;/td&gt;
 &lt;td&gt;Price discipline (λ), share, bit growth&lt;/td&gt;
 &lt;td&gt;λ 0.5 to 1: normalized operating profit ₩117 trillion to ₩129 trillion, ROE about +0.7 points. Larger if the 2029 to 2033 profit path is higher&lt;/td&gt;
 &lt;td&gt;Low. The market decides&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;③ Reinvest retained cash (bigger numerator)&lt;/td&gt;
 &lt;td&gt;Put 3% cash into a business earning 17% ROIC&lt;/td&gt;
 &lt;td&gt;All ₩650 trillion at 17%: ROE of 14% or more in theory&lt;/td&gt;
 &lt;td&gt;Lowest. No obvious projects, poor track record&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;&lt;img alt="2030 ROE by lever. Today’s path 7.7%. Raising the payout to 100% gives 11.1% (decided by the board, certainty high). Raising price discipline λ from 0.5 to 1 gives about 8.4% (decided by the price cycle, certainty low). Reinvesting ₩650 trillion of retained cash at 17% gives 14% or more on paper (needs ₩650 trillion of projects, certainty lowest). The required return is 10%." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/samsung-two-engines-profit-and-payouts-equity-roe-2026-11-08/levers.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;Profit (②) is set by the market. The one step our model tested, λ from 0.5 to 1, lifts normalized operating profit from ₩117 trillion to ₩129 trillion and ROE by about 0.7 points. On a path where the profit level itself is higher, the effect is much larger. If operating profit doubled, ROE would rise far more than any payout change. But the price cycle sets that variable.&lt;/p&gt;
&lt;p&gt;Reinvestment (③) works only if there are places to put ₩650 trillion at 17%. Memory capex already runs at ₩120 trillion a year, and more than that creates oversupply that cuts prices and hurts ②. Foundry and acquisitions have earned low returns in the past. So ③ is more likely to burn capital than to grow the numerator. That is why our model puts a 15% loss rate on retained cash (the windfall loss rate).&lt;/p&gt;
&lt;p&gt;Payouts are not the biggest way to lift ROE. The profit level can be the biggest lever, but Samsung does not set it, and reinvestment is large on paper but likely to fail. What remains is the payout. The accurate claim is narrower: payouts are the only reliable method the company can choose by itself.&lt;/p&gt;
&lt;h2 id="the-same-equity-is-worth-103000-or-300000-a-share-depending-on-the-multiple"&gt;The Same Equity Is Worth ₩103,000 or ₩300,000 a Share, Depending on the Multiple
&lt;/h2&gt;&lt;p&gt;Assume 2028 equity of ₩1,200 trillion and change only the multiple.&lt;/p&gt;
&lt;p&gt;&lt;img alt="Per-share value of ₩1,200 trillion of equity at reference multiples, ₩ thousand. China’s big-bank multiples of 0.56x to 0.82x give ₩103,000 to ₩151,000 a share (₩670 trillion to ₩980 trillion). The November 6 close of ₩188,700 equals a ₩1,230 trillion market value. Berkshire’s 1.63x gives ₩300,000 a share (₩1,960 trillion)." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/samsung-two-engines-profit-and-payouts-equity-roe-2026-11-08/price-ladder.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;China&amp;rsquo;s big-bank multiples of 0.56 to 0.82 give ₩670 trillion to ₩980 trillion (₩103,000 to ₩151,000 a share). Berkshire&amp;rsquo;s 1.63 gives ₩1,960 trillion (₩300,000 a share). The November 6 close of ₩188,700 sits in between. Samsung&amp;rsquo;s normalized ROE range of 7.7% to 11.1% runs from the Chinese banks (8.6% to 10.1%) up to just below Berkshire (11.5%).&lt;/p&gt;
&lt;p&gt;Equity above ₩1,200 trillion is not a reason for the market value to rise. It is a risk that ROE gets diluted. The payout rate and the return on retained cash decide whether that risk grows or shrinks.&lt;/p&gt;
&lt;h2 id="payouts-do-not-create-value-they-change-how-the-market-prices-it"&gt;Payouts Do Not Create Value; They Change How the Market Prices It
&lt;/h2&gt;&lt;p&gt;A payout returns ₩1 as ₩1. In our model, value per share rises only about ₩10,000 when the payout goes from 50% to 100%, because the model counts retained cash at 85% of face value. What changes a lot is accounting ROE and P/B, that is, the frame the market prices the stock in, and the risk that cash leaks into low-return investment. At the same value, a company that looks like it earns below its cost of capital trades differently from one that looks like it earns above. That is the gap between the Chinese banks and Berkshire.&lt;/p&gt;
&lt;p&gt;Payouts have more limits. A buyback raises value per share only when the price is below value. ₩188,700 is below our model&amp;rsquo;s value of ₩276,000, so repurchases help now. Buying above value would hurt. A dividend has no such problem, but it is taxed.&lt;/p&gt;
&lt;p&gt;There is also a ceiling. At a 100% payout ROE is 11.1%, only one point above the required return. To go beyond that, normalized profit itself must grow.&lt;/p&gt;
&lt;h2 id="what-the-late-january-board-can-show-and-what-it-cannot"&gt;What the Late-January Board Can Show, and What It Cannot
&lt;/h2&gt;&lt;p&gt;The best case uses both engines. Supply discipline keeps λ up and supports the numerator. Cash that is still left over goes back and pins the denominator. The late-January board can set the payout rate and the size of capex. It cannot set memory prices. So the January announcement is the first look at how fast the company&amp;rsquo;s own engine is running.&lt;/p&gt;
&lt;h2 id="data-notes"&gt;Data Notes
&lt;/h2&gt;&lt;ul&gt;
&lt;li&gt;We collected no new external data for this post. Figures come from our model files and the analysis note of November 8.&lt;/li&gt;
&lt;li&gt;The share price and market values are as of November 6. The 1.17 times book value in our earlier post used 12-month-forward equity. The current P/B of 2.1 in this post uses equity reported at September 30. On projected end-2028 equity, P/B is about 0.9 to 1.1.&lt;/li&gt;
&lt;li&gt;For global equity, US company values were rechecked directly in SEC XBRL data. The Chinese banks and Aramco come from filed reports. Berkshire and the Chinese banks have no September 30 report yet, so their June 30 values are used. All values are converted at the November 6 exchange rate, and reporting dates differ from pricing dates.&lt;/li&gt;
&lt;li&gt;Aramco&amp;rsquo;s ROE slot shows ROACE (return on average capital employed). Samsung&amp;rsquo;s boom-year ROE of about 50% is an approximation.&lt;/li&gt;
&lt;li&gt;Fair P/B is simplified to 0% growth and a 10% required return. ROE effects are measured inside our model and depend on its settings.&lt;/li&gt;
&lt;li&gt;The 2030 net income of ₩90 trillion has interest on retained cash added to it. ROIC divides by operating invested capital and ROE divides by book equity, so a simple weighted average of the 17% ROIC and the 3% cash return does not give 7.7%.&lt;/li&gt;
&lt;li&gt;Profit paths and payout rates are assumptions. The payout rate is set by the board and has not been announced.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;&lt;strong&gt;If you sat on Samsung&amp;rsquo;s board in January, would you vote for a 50% or a 100% payout? Reply with the one number that decides it for you.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;This is for research and information only, not investment advice. Value ranges and scenarios are analytical assumptions, not trading instructions or individual portfolio decisions. Figures are dated data. We built the profit model with AI and used AI for research and editing, working with Anthropic&amp;rsquo;s Claude and OpenAI GPT models. The comparison table includes Alphabet and Amazon, investors in Anthropic, and Microsoft, an investor in OpenAI. This post does not assess those companies&amp;rsquo; businesses or value. It compares only their equity and ROE figures.&lt;/p&gt;</description></item><item><title>Samsung at ₩188,700 Only Works If Three Bear Cases Hit at Once</title><link>https://koreainvestinsights.com/post/samsung-188700-prices-three-bear-cases-at-once-2026-11-08/</link><pubDate>Sun, 08 Nov 2026 01:24:00 +0900</pubDate><guid>https://koreainvestinsights.com/post/samsung-188700-prices-three-bear-cases-at-once-2026-11-08/</guid><description>&lt;img src="https://koreainvestinsights.com/" alt="Featured image of post Samsung at ₩188,700 Only Works If Three Bear Cases Hit at Once" /&gt;&lt;p&gt;Samsung Electronics closed at &lt;strong&gt;₩188,700&lt;/strong&gt; on November 6. That is 2.6 to 2.9 times 12-month forward earnings and 1.17 times book value. On the same day, the average target price of the 21 brokers tracked by FnGuide was &lt;strong&gt;₩472,381&lt;/strong&gt;, and 20 of the 21 rated the stock a buy. The average target is 2.5 times the share price.&lt;/p&gt;
&lt;p&gt;To explain that gap, we built two separate valuation models of Samsung. The first is a scenario grid that splits the timing and depth of the memory price turn into 96 cases. The second is a cash-flow model that projects cash flows and returns on capital out to 2040. They came out at &lt;strong&gt;₩232,600&lt;/strong&gt; and &lt;strong&gt;₩291,300&lt;/strong&gt; per share, ₩59,000 apart.&lt;/p&gt;
&lt;p&gt;When we took the gap apart one assumption at a time, the picture became simple. Both models and the consensus put 2027 operating profit near ₩600 trillion. They split on 2028 prices, on how long excess profits last after 2029, and on what happens to the cash that piles up. Load the bear assumptions into the cash-flow model at the same time and it lands at &lt;strong&gt;₩195,300&lt;/strong&gt;, almost exactly today&amp;rsquo;s price.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Our view: today&amp;rsquo;s price only makes sense if a 2028 profit decline, the end of excess profits in 2030 and a low return on retained cash all come true at once. We read Samsung as conditionally undervalued. Our working range is ₩233,000 to ₩291,000, with a center near ₩260,000, 38% above the close. The condition is the shareholder-return policy from the board meeting in late January 2027. If returns stay undecided and capex rises 30% or more, we move our view toward ₩196,000.&lt;/strong&gt;&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;2027 operating profit converges at ₩600 trillion to ₩630 trillion. Fourth-quarter exit prices and long-term contracts tie the numbers down. Eight brokers cluster between ₩604 trillion and ₩644 trillion.&lt;/li&gt;
&lt;li&gt;2028 splits from ₩532 trillion (−11%) to ₩684 trillion (+9%). Among brokers the spread runs from ₩337 trillion to ₩939 trillion. Nobody has yet observed when prices turn.&lt;/li&gt;
&lt;li&gt;67% to 72% of the value comes from 2029 onward. So the gap between the models comes more from the life of excess profits and the treatment of retained cash than from 2027 to 2028 profit.&lt;/li&gt;
&lt;li&gt;The ₩59,000 gap is 2028 profit, the life of excess profits, a permanent excess return and retained cash, each worth ₩30,000 to ₩40,000, pulling in opposite directions.&lt;/li&gt;
&lt;li&gt;Retained cash is the first variable we will observe. The board meeting in late January 2027 gives the first answer.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id="the-same-company-carries-per-share-values-from-164000-to-472000"&gt;The same company carries per-share values from ₩164,000 to ₩472,000
&lt;/h2&gt;&lt;p&gt;&lt;img alt="Samsung value ladder, ₩ thousand per share. November 6 close 188.7. Replaced first model 164.4, replaced earlier grid 201.4, scenario grid 232.6 (10th to 90th percentile 183.8 to 283.1), cash-flow model 291.3 (downside 107.1 to upside 525.4), five brokers aligned to present value 424.1 (331.3 to 554.1), FnGuide average target of 21 brokers 472.4 (270 to 690). Our working range 233 to 291." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/samsung-188700-prices-three-bear-cases-at-once-2026-11-08/value-ladder.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;Before comparing numbers, align the definitions. Our two models give a present value as of today. Broker targets are 12-month prices, so the ₩472,381 average discounts to about ₩429,000 over one year.&lt;/p&gt;
&lt;p&gt;So we restated the targets of five brokers (Hana, IBK, KB, Shinhan and iM) to the same date, the same dividend rights and a 12% discount rate. Their average present value is ₩424,100, from ₩331,300 at IBK to ₩554,100 at KB. FnGuide targets run from ₩270,000 at BNK (hold) to ₩690,000 at Yuanta.&lt;/p&gt;
&lt;p&gt;The horizontal bars are each model&amp;rsquo;s own stated range, not confidence intervals. The grid bar is the 10th to 90th percentile of its 96 cases (₩183,800 to ₩283,100). The cash-flow bar runs from a downside of ₩107,100, with the bad assumptions stacked, to an upside of ₩525,400.&lt;/p&gt;
&lt;p&gt;The hollow markers are numbers we threw away. ₩164,400 came from a first model built on one broker&amp;rsquo;s detailed estimate table, which we replaced with a skeleton built from company filings and industry data. ₩201,400 came from an earlier grid that drew more downside paths than upside paths; we replaced it when we put up and down cases on the same axes. Every model still standing has its center above today&amp;rsquo;s price.&lt;/p&gt;
&lt;h2 id="everyone-uses-the-same-2027-profit"&gt;Everyone uses the same 2027 profit
&lt;/h2&gt;&lt;p&gt;&lt;img alt="Operating profit by year, ₩ trillion. Actuals: 2025 43.7, first nine months of 2026 254.1. 2026 is 383 to 388 for all three. In 2027 the scenario grid has 600.2, the cash-flow model 629.5 and the FnGuide consensus 606.6, with individual broker estimates between 380 and 750. In 2028 they split: grid 532.3, cash-flow model 684.0, consensus 668.8, seven brokers between 337 and 939. In 2029 the grid has 640.8 and the cash-flow model 593.2. The shaded band is the grid’s 10th to 90th percentile." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/samsung-188700-prices-three-bear-cases-at-once-2026-11-08/profit-path.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;The only reported results are 2025 operating profit of ₩43.7 trillion and ₩254.1 trillion for the first nine months of 2026 (₩57.2 trillion, ₩89.5 trillion and ₩107.4 trillion by quarter). Everything after that is an estimate.&lt;/p&gt;
&lt;p&gt;All three put 2026 at ₩383 trillion to ₩388 trillion, adding ₩129 trillion to ₩134 trillion for the fourth quarter. The company&amp;rsquo;s shipment guidance and TrendForce&amp;rsquo;s fourth-quarter contract-price outlook (+10% to 15%) tie that number down.&lt;/p&gt;
&lt;p&gt;2027 converges too. The scenario grid has ₩600.2 trillion, the cash-flow model ₩629.5 trillion and the FnGuide consensus ₩606.6 trillion, all within ±5% of one another. Eight brokers cluster between ₩604 trillion and ₩644 trillion. At Hana, 86% of the 2027 estimate is the fourth-quarter exit price run for a full year. Most of the 2027 estimates are tied to this year&amp;rsquo;s closing prices.&lt;/p&gt;
&lt;p&gt;The quality of that profit overlaps as well. The share of operating profit left as profit for shareholders is 76% to 77% (76.9% in the grid, a 76.3% broker median). Several sources put 2027 HBM revenue at ₩115 trillion to ₩117 trillion.&lt;/p&gt;
&lt;p&gt;The exception is Kiwoom. It alone has profit falling, to ₩380 trillion in 2027 and ₩337 trillion in 2028. We could not find the price table behind it in public material.&lt;/p&gt;
&lt;h2 id="from-2028-the-numbers-split-because-nobody-knows-when-prices-turn"&gt;From 2028 the numbers split, because nobody knows when prices turn
&lt;/h2&gt;&lt;p&gt;2028 operating profit is ₩532.3 trillion in the grid (−11.3%), ₩684.0 trillion in the cash-flow model (+8.7%) and ₩668.8 trillion in the consensus (+10.2%). Among brokers it runs from ₩337.4 trillion at Kiwoom (−11%) to ₩939.3 trillion at IBK (+32%).&lt;/p&gt;
&lt;p&gt;The consensus +10% is an agreement in appearance only. It averages a range from −11% to +32%. The cash-flow model&amp;rsquo;s +9% sits on that average, but it was built a different way.&lt;/p&gt;
&lt;p&gt;The split comes from one fact: no one has yet observed when memory prices turn. The grid uses the length of past upcycles as its base rate. The last four lasted 9, 3, 3 and 5 quarters. This one reaches its sixth quarter in the fourth quarter of 2026, longer than three of the four. So the grid puts 85% of its weight on prices turning by the first quarter of 2028. That weight is a subjective judgment, not a calibrated probability.&lt;/p&gt;
&lt;p&gt;The cash-flow model and the brokers extend the company&amp;rsquo;s call for &amp;ldquo;a deeper shortage in 2027&amp;rdquo; and its long-term contracts into 2028. The cash-flow model&amp;rsquo;s ₩684 trillion assumes conventional DRAM prices −10%, NAND −20% and flat HBM prices, with DRAM shipments +18%, HBM +35%, NAND +20% and cost per bit −5%. That keeps the conventional memory operating margin almost unchanged, from 87.9% to 87.2%. In the grid&amp;rsquo;s distribution, ₩684 trillion sits in roughly the top quarter.&lt;/p&gt;
&lt;p&gt;The grid&amp;rsquo;s expected quarterly path peaks at ₩154.7 trillion in the second quarter of 2027 and bottoms at ₩128.9 trillion in the second quarter of 2028. If prices turn early, in the first quarter of 2027, profit falls to ₩94.8 trillion by the fourth quarter of 2027. If prices do not turn through 2028, it climbs to ₩222.3 trillion by the fourth quarter of 2028.&lt;/p&gt;
&lt;p&gt;Broker quarterly tables do not answer the question either. Hana, DS and Daishin go from +4% to 16% in the first quarter of 2027 to −1% to +1% by the fourth quarter. None of them shows a price fall the size of past downturns.&lt;/p&gt;
&lt;h2 id="but-value-is-decided-from-2029-on"&gt;But value is decided from 2029 on
&lt;/h2&gt;&lt;p&gt;In the grid, 72% of the value comes from the present value of book equity at the end of 2029. In the cash-flow model, 67% of operating value comes from 2029 onward. The fight over 2027 and 2028 profit is a fight over about a third of the value.&lt;/p&gt;
&lt;p&gt;&lt;img alt="How the two models treat operating profit after 2028, ₩ trillion. The grid goes from 641 in 2029 straight to a steady state from 2030 (net profit ₩167 trillion, about 220 in operating-profit terms). The cash-flow model steps down over five years, 593 in 2029, 501 in 2030, 407 in 2031, 310 in 2032 and 211 in 2033, then holds at 183 in 2036 and 202 in 2040. The area between the lines and the permanent excess return are the gap between the models." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/samsung-188700-prices-three-bear-cases-at-once-2026-11-08/long-run.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;The grid drops from ₩641 trillion in 2029 to a steady state in 2030 in one step. Net profit falls 66%, from ₩492 trillion to ₩167 trillion. In the steady state, return on equity (ROE) of 11% equals the 11% discount rate, so excess profit is zero.&lt;/p&gt;
&lt;p&gt;The cash-flow model fades excess profit on conventional memory over five years and on high-value products such as HBM over eight. After that, it lets capital used in production earn 18% forever, against a 12% required return. From 2036 on, the two models show similar profit levels. The gap is the area between them in 2030 to 2033 and that permanent excess return.&lt;/p&gt;
&lt;p&gt;Samsung&amp;rsquo;s own record answers part of the question. Its ROE from 2017 to 2025 averaged 12.7% with a median of 10.8%. The longest run of years above 17% is two. Micron&amp;rsquo;s longest is three.&lt;/p&gt;
&lt;p&gt;&lt;img alt="Samsung Electronics ROE attributable to owners, DART consolidated, percent. 2017 21.0, 2018 19.6, 2019 8.7, 2020 10.0, 2021 13.9, 2022 17.1, 2023 4.1, 2024 9.0, 2025 10.8. Average 12.7%, median 10.8%. Lines mark the terminal ROE today’s price requires (8.1%), the grid’s terminal ROE (11%) and the lowest terminal ROE the average target requires (19.5%)." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/samsung-188700-prices-three-bear-cases-at-once-2026-11-08/roe-history.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;The cash-flow model&amp;rsquo;s 18% is a return on operating capital (ROIC) that excludes non-operating cash, so it does not compare directly with ROE. Even so, a permanent 6-point excess return is more optimistic than Samsung&amp;rsquo;s nine-year record. Micron&amp;rsquo;s average ROE depends on the sample: 14.0% for fiscal 2011 to 2025, 18.7% with fiscal 2026 included.&lt;/p&gt;
&lt;p&gt;What happens to the cash that piles up is another large assumption. At a 33% payout ratio, equity attributable to owners grows from ₩608 trillion in 2026 to ₩1,521 trillion in 2029, 2.5 times. The grid lets that money earn 11% on the books. The cash-flow model returns half at once and half after five years, earning 2.5% in the meantime. Most brokers do not state this assumption.&lt;/p&gt;
&lt;p&gt;Both models point at the same variables with similar weight. Variables after 2029 (the life of excess profits, retained cash and the discount rate) are larger than 2027 to 2028 profit.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Variable (change from base, ₩ thousand)&lt;/th&gt;
 &lt;th&gt;Scenario grid (base 232.6)&lt;/th&gt;
 &lt;th&gt;Cash-flow model (base 291.3)&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Life of excess profits&lt;/td&gt;
 &lt;td&gt;0 to 3 years +31.2, 0 to 10 years +112.8, 0 to 15 years +152.4&lt;/td&gt;
 &lt;td&gt;3 and 5 years −33.6, 8 and 12 years +40.5&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Conventional memory excess profit ends&lt;/td&gt;
 &lt;td&gt;Not modeled&lt;/td&gt;
 &lt;td&gt;Zero from 2029 −55.2, losses in 2030 to 2031 −83.0&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Retained cash&lt;/td&gt;
 &lt;td&gt;Half of retained capital earns 3%: −36.7&lt;/td&gt;
 &lt;td&gt;Return delayed 10 years −38.2, returned at once +59.5&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Terminal return&lt;/td&gt;
 &lt;td&gt;ROE 9% −30.5, 13% +30.5&lt;/td&gt;
 &lt;td&gt;ROIC 12% −33.9, 24% +33.9&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Discount rate&lt;/td&gt;
 &lt;td&gt;12% −19.4, 10% +23.2&lt;/td&gt;
 &lt;td&gt;14% −36.4, 10% +52.0&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;2027 to 2028 profit&lt;/td&gt;
 &lt;td&gt;16 weightings of price-turn timing and depth −16.3 to +18.8&lt;/td&gt;
 &lt;td&gt;2028 at grid value −31.4, at consensus −3.2&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Other&lt;/td&gt;
 &lt;td&gt;Long-term contract price band (unconfirmed) +11.8, FX −6.2 to +11.9&lt;/td&gt;
 &lt;td&gt;Early-return share 35% or 65% ∓17.9&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Reweighting the timing and depth of the price turn 16 different ways moves the grid&amp;rsquo;s value by −₩16,300 to +₩18,800. Extending the life of excess profits by just three years moves it by ₩31,200.&lt;/p&gt;
&lt;p&gt;We also share new analyses on Telegram as they are published. You are welcome to join us at &lt;a class="link" href="https://t.me/korea_invest_insights" target="_blank" rel="noopener"
 &gt;@korea_invest_insights&lt;/a&gt;.&lt;/p&gt;
&lt;h2 id="the-59000-gap-is-levers-cancelling-each-other-out"&gt;The ₩59,000 gap is levers cancelling each other out
&lt;/h2&gt;&lt;p&gt;We kept the cash-flow model&amp;rsquo;s code as it is and switched its assumptions, one at a time, to the grid&amp;rsquo;s settings.&lt;/p&gt;
&lt;p&gt;&lt;img alt="Bridge from the cash-flow model’s ₩291,300 to the scenario grid’s ₩232,600, ₩ thousand. 2028 operating profit 684 to 532 trillion −31.4; immediate normalization in 2030 −28.5; permanent ROIC 18% to 12% −36.2, which brings the running total to 195.3, almost today’s 188.7. Retained cash returned at once +38.6; no discount on investment assets +2.9; required return and steady-state ROIC both 11% +4.1; 6.519 billion shares +4.8; structural residual −13.1; scenario grid 232.6." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/samsung-188700-prices-three-bear-cases-at-once-2026-11-08/bridge.en.svg"&gt;
&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Step&lt;/th&gt;
 &lt;th&gt;Assumption changed&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Change (₩ thousand)&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Running total (₩ thousand)&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Start&lt;/td&gt;
 &lt;td&gt;Cash-flow model as built&lt;/td&gt;
 &lt;td style="text-align: right"&gt;&lt;/td&gt;
 &lt;td style="text-align: right"&gt;291.3&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;1&lt;/td&gt;
 &lt;td&gt;2028 operating profit ₩684 trillion to ₩532 trillion&lt;/td&gt;
 &lt;td style="text-align: right"&gt;−31.4&lt;/td&gt;
 &lt;td style="text-align: right"&gt;259.9&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;2&lt;/td&gt;
 &lt;td&gt;Excess profit kept through 2029, zero from 2030&lt;/td&gt;
 &lt;td style="text-align: right"&gt;−28.5&lt;/td&gt;
 &lt;td style="text-align: right"&gt;231.4&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;3&lt;/td&gt;
 &lt;td&gt;Steady-state ROIC 18% to 12% (no permanent excess return)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;−36.2&lt;/td&gt;
 &lt;td style="text-align: right"&gt;195.3&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;4&lt;/td&gt;
 &lt;td&gt;Retained cash: half after five years to all at once&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+38.6&lt;/td&gt;
 &lt;td style="text-align: right"&gt;233.8&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;5&lt;/td&gt;
 &lt;td&gt;Remove the 20% discount on investment assets&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+2.9&lt;/td&gt;
 &lt;td style="text-align: right"&gt;236.8&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;6&lt;/td&gt;
 &lt;td&gt;Required return and steady-state ROIC 12% to 11%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+4.1&lt;/td&gt;
 &lt;td style="text-align: right"&gt;240.8&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;7&lt;/td&gt;
 &lt;td&gt;Shares 6.649 billion to 6.519 billion (buybacks)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+4.8&lt;/td&gt;
 &lt;td style="text-align: right"&gt;245.6&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;8&lt;/td&gt;
 &lt;td&gt;Structural differences (2027 and 2029 profit, treatment of cash and investments)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;−13.1&lt;/td&gt;
 &lt;td style="text-align: right"&gt;232.6&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;After step three the running total is ₩195,300, almost the share price. At that point retained cash still follows the cash-flow model&amp;rsquo;s own assumption: half comes back five years late. Today&amp;rsquo;s price is therefore the price at which an 11% profit decline in 2028, excess profit vanishing in 2030, no permanent excess return and late return of retained cash all hold together.&lt;/p&gt;
&lt;p&gt;The grid finds the same spot. If half of retained capital sits as cash earning 3%, it gives ₩195,900. A terminal ROE of 8.1% also reproduces today&amp;rsquo;s price. That is below Samsung&amp;rsquo;s nine-year median of 10.8%.&lt;/p&gt;
&lt;p&gt;The order of the steps changes the size of each bar. Switch the valuation assumptions first and immediate normalization is −₩38,600, the permanent excess return −₩36,200, cash +₩43,200 and 2028 profit −₩26,600. The shape is the same: each lever is ₩30,000 to ₩40,000, and they point in opposite directions.&lt;/p&gt;
&lt;h2 id="each-price-asks-for-a-different-future"&gt;Each price asks for a different future
&lt;/h2&gt;&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Price&lt;/th&gt;
 &lt;th&gt;What has to be true&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Today, ₩188,700&lt;/td&gt;
 &lt;td&gt;Grid: terminal ROE of 8.1% or a 13.6% discount rate, or half of retained capital earning 3% (₩195,900). Cash-flow model: 2028 at ₩532 trillion, immediate normalization in 2030, no permanent excess return and half the cash five years late (₩195,300). With its original assumptions, a 20.0% required return&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Scenario grid, ₩232,600&lt;/td&gt;
 &lt;td&gt;₩600 trillion in 2027, ₩532 trillion in 2028, ₩641 trillion in 2029, then ROE equal to the 11% discount rate from 2030. All accumulated cash earns 11%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Cash-flow model, ₩291,300&lt;/td&gt;
 &lt;td&gt;2028 +8.7%, excess profits fading over five and eight years, permanent ROIC of 18%, half the cash five years late, 12% required return&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Average target ₩472,381 (present value ₩424,100)&lt;/td&gt;
 &lt;td&gt;In the grid&amp;rsquo;s framework, a terminal ROE of 19.5% to 26.7%, or consensus profit with a 15-year fade (₩475,800), or a 10% discount rate, 20% ROE and 3% growth (₩485,400). The same level as the cash-flow model&amp;rsquo;s upside (₩525,400)&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Most broker targets are book value per share times a price-to-book multiple: 4.6 times at KB, 4.3 at IBK, 3.2 at Shinhan and 2.3 at iM. That puts a normal multiple on peak profit. Hana raised its profit estimate by 20% and lowered its multiple to keep a ₩480,000 target.&lt;/p&gt;
&lt;p&gt;&lt;img alt="FnGuide average target price and Samsung’s closing price, won. June 30: 465,208 versus 334,000. July 31: 493,542 versus 262,500. August 31: 491,875 versus 260,000. September 30: 493,864 versus 268,500. October 30: 490,227 versus 225,000. November 6: 472,381 versus 188,700. Targets stayed between ₩465,000 and ₩494,000 while the stock fell 43.5%." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/samsung-188700-prices-three-bear-cases-at-once-2026-11-08/targets-vs-price.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;Since the end of June the average target has stayed between ₩465,000 and ₩494,000, while the stock fell 43.5%, from ₩334,000 to ₩188,700. The record runs the other way too. The average target of ₩139,385 on December 30, 2025, was far below the ₩334,000 the stock reached six months later. Targets tend to lag in upcycles and stay put after the peak.&lt;/p&gt;
&lt;p&gt;The 12-month horizon has not expired, so we do not treat this as proof the targets are wrong. On the same profit path, changing only the valuation assumptions gives ₩161,500 to ₩485,400, which reaches both our numbers and the targets. We do not rule on who is wrong. We say where today&amp;rsquo;s price sits, below the center of all three, and which variables make the gap.&lt;/p&gt;
&lt;h2 id="our-view-conditionally-undervalued-but-we-do-not-center-on-291300"&gt;Our view: conditionally undervalued, but we do not center on ₩291,300
&lt;/h2&gt;&lt;p&gt;If each model gives up one assumption, they meet. Give the grid a three-year fade in excess profits and it reaches ₩263,800. Give the cash-flow model the grid&amp;rsquo;s 2028 profit and it reaches ₩259,900. So our working range is ₩233,000 to ₩291,000 with a center near ₩260,000. That center approximates two points; it is not a probability-weighted expected value.&lt;/p&gt;
&lt;p&gt;There are reasons not to center on the cash-flow model&amp;rsquo;s ₩291,300. Its 2028 profit of ₩684 trillion sits in roughly the top quarter of the grid&amp;rsquo;s distribution and keeps an 87% conventional memory margin through 2028. A permanent 6-point excess return does not fit Samsung&amp;rsquo;s nine-year median ROE of 10.8%. Its 2027 starting point is the broker median, so it is not independent evidence. Its outside reviewer also approved it on the condition that the model as a whole is not called conservative.&lt;/p&gt;
&lt;p&gt;There are reasons not to keep the grid&amp;rsquo;s ₩232,600 either. The cliff from ₩492 trillion of net profit in 2029 to ₩167 trillion in 2030 is an input, not a result. The grid started from the possibility of a profit decline, and it says so: it leans toward downside stress.&lt;/p&gt;
&lt;p&gt;The loss paths are real. In the cash-flow model, if conventional memory excess profit is zero from 2029, value is ₩236,100; if conventional memory loses money in 2030 and 2031, it is ₩208,300. With the bad assumptions stacked, it is ₩107,100. In the grid, 12.5% of cases sit below today&amp;rsquo;s price, and in the cases where prices turn in the first quarter of 2027, value falls to ₩157,000 to ₩184,000. A center above the price does not mean the risk of loss is small.&lt;/p&gt;
&lt;p&gt;We also have a warning sign of our own. Over November 7 and 8, our own value estimate rose from ₩201,400 to ₩218,000, ₩225,000, ₩232,600 and then ₩291,300, while the stock was falling. The reasons were a corrected HBM input and a more symmetric scenario design. Still, a number moving against the price looks like the broker targets that never followed the stock down. So we will recalculate only after first-hand observations come in.&lt;/p&gt;
&lt;h2 id="the-board-meeting-in-late-january-gives-the-first-verdict"&gt;The board meeting in late January gives the first verdict
&lt;/h2&gt;&lt;p&gt;Retained cash is the first variable we will observe. The two models bet −₩36,700 (grid) and −₩38,200 to +₩59,500 (cash-flow model) on this one lever.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Date&lt;/th&gt;
 &lt;th&gt;Event&lt;/th&gt;
 &lt;th&gt;Variable it settles&lt;/th&gt;
 &lt;th&gt;Signal toward undervalued&lt;/th&gt;
 &lt;th&gt;Signal toward today&amp;rsquo;s price&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Nov 12, 2026&lt;/td&gt;
 &lt;td&gt;FnGuide recheck&lt;/td&gt;
 &lt;td&gt;Whether targets or EPS move first&lt;/td&gt;
 &lt;td&gt;2027 EPS of ₩73,055 holds within ±3%&lt;/td&gt;
 &lt;td&gt;At or below ₩69,402 (−5%), an early warning&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Late Dec 2026&lt;/td&gt;
 &lt;td&gt;TrendForce outlook for first-quarter 2027 conventional DRAM contract prices&lt;/td&gt;
 &lt;td&gt;When prices turn&lt;/td&gt;
 &lt;td&gt;+6% or more lowers the weight on an early turn&lt;/td&gt;
 &lt;td&gt;0% or less raises it and widens the downside tail&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Jan 28, 2027 (est.)&lt;/td&gt;
 &lt;td&gt;Fourth-quarter results, board return policy, 2027 capex&lt;/td&gt;
 &lt;td&gt;Retained cash&lt;/td&gt;
 &lt;td&gt;Returns of 50% or more of free cash flow made regular from 2027, with a cancellation schedule and a disclosed long-term contract price band: center to ₩260,000 or higher&lt;/td&gt;
 &lt;td&gt;Returns undecided, capex up 30% or more, renegotiation clauses: toward ₩196,000&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Jul 31, 2027 (est.)&lt;/td&gt;
 &lt;td&gt;Second-quarter 2027 results, the first quarter after prices turn&lt;/td&gt;
 &lt;td&gt;Life of excess profits&lt;/td&gt;
 &lt;td&gt;Quarterly net profit at 60% or more of the peak quarter, cash conversion intact&lt;/td&gt;
 &lt;td&gt;30% or less, inventory and working capital rising&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;These lines mark where we will look at the view again. They are not automatic trading rules.&lt;/p&gt;
&lt;h2 id="data-notes"&gt;Data notes
&lt;/h2&gt;&lt;ul&gt;
&lt;li&gt;We collected no new external data for this post. Figures come from our dated model files and analysis notes.&lt;/li&gt;
&lt;li&gt;The share price is the KRX close on November 6. Consensus and targets are FnGuide as of November 6. Broker estimates come from each broker&amp;rsquo;s original report.&lt;/li&gt;
&lt;li&gt;We did not obtain the original reports from Korea Investment, Samsung Securities, NH, Meritz or BNK, or Kiwoom&amp;rsquo;s price table. We have no ROE baseline for SK hynix or TSMC.&lt;/li&gt;
&lt;li&gt;Scenario weights and probabilities in both models are subjective and uncalibrated.&lt;/li&gt;
&lt;li&gt;The bridge depends on the order of steps. We did not split the final structural residual of −₩13,100 further.&lt;/li&gt;
&lt;li&gt;We did not value the share-count difference (6.52 billion versus 6.65 billion) or preferred-share rights precisely.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;&lt;strong&gt;Which bear assumption in today&amp;rsquo;s price breaks first? Reply with A, B, C or D, and one data point that backs your pick.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;A. 2028 profit: prices hold through 2028 and profit grows.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;B. Life of excess profits: HBM and long-term contracts keep above-average returns after 2030.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;C. Retained cash: the January board makes returns regular and the cash comes back to shareholders.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;D. None: the bear assumptions really do arrive together.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;This is for research and information only, not investment advice. Value ranges and scenarios are analytical assumptions, not trading instructions or individual portfolio decisions. Figures are dated data. We built the two valuation models with AI: the scenario grid with Anthropic&amp;rsquo;s Claude and the cash-flow model with an OpenAI GPT model, and we used Claude for research and editing. Neither company is directly related to the subject of this post.&lt;/p&gt;</description></item><item><title>Every AI Agent Needs Its Own Computer. The Four Stocks Behind That Are Down 43-60%</title><link>https://koreainvestinsights.com/post/ai-agent-cloud-every-agent-gets-its-own-computer-docn-akam-stx-2026-11-07/</link><pubDate>Sat, 07 Nov 2026 21:55:00 +0900</pubDate><guid>https://koreainvestinsights.com/post/ai-agent-cloud-every-agent-gets-its-own-computer-docn-akam-stx-2026-11-07/</guid><description>&lt;img src="https://koreainvestinsights.com/" alt="Featured image of post Every AI Agent Needs Its Own Computer. The Four Stocks Behind That Are Down 43-60%" /&gt;&lt;p&gt;On November 5, the Philadelphia Semiconductor Index (SOX) fell 12.25% in one day. By the November 6 close, DigitalOcean was &lt;strong&gt;60.0%&lt;/strong&gt; below its 12-month high, Seagate &lt;strong&gt;57.9%&lt;/strong&gt;, Akamai &lt;strong&gt;48.4%&lt;/strong&gt; and Micron &lt;strong&gt;43.1%&lt;/strong&gt;.&lt;/p&gt;
&lt;p&gt;Now look at what these companies sell. DigitalOcean says revenue from its AI customers is running at &lt;strong&gt;$376 million&lt;/strong&gt; a year, up 288%. Akamai signed a seven-year, $11.6 billion deal with Anthropic. AMD&amp;rsquo;s data center sales rose 97% in the third quarter, and Intel&amp;rsquo;s rose 78%. The demand is showing up in the numbers, and the stocks fell anyway.&lt;/p&gt;
&lt;p&gt;The reason is a change in what a cloud computer is. When millions of people use a website, they all share one program. AI agents do not work that way. Each agent job gets a small sealed computer of its own, for a few seconds or a few days. That needs more server chips, more memory and more storage, and it needs them on top of the graphics chips (GPUs) that investors already watch.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Our call: the new layers of software that agents need (sealed computers, model routers, training arenas) mostly belong to private companies and the big cloud providers. Public-market investors have a chance in two places. One is older cloud platforms that are rebuilding themselves as agent clouds, DigitalOcean and Akamai. The other is storage and memory suppliers, Seagate and Micron. AMD and Cloudflare sell the right product, but their prices already count on it.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;We start with the idea in plain words, then go company by company, then say what would prove us wrong. At the end we ask which side you take.&lt;/p&gt;
&lt;h2 id="an-agent-cloud-in-plain-words"&gt;An agent cloud, in plain words
&lt;/h2&gt;&lt;p&gt;Think of a restaurant. The old cloud is one kitchen that feeds thousands of diners from the same stoves. Its goal is to keep every burner busy. An agent cloud is closer to a workshop that hands each contractor a locked workbench. The contractor finishes the job, the bench is wiped and taken away, and the next contractor gets a fresh one.&lt;/p&gt;
&lt;p&gt;The locks matter because of what the contractor does. An AI agent writes code a few seconds before it runs it, and no human has read that code. If it misbehaves, it must not touch anyone else&amp;rsquo;s data. So each job runs inside its own sealed virtual computer. Amazon&amp;rsquo;s agent service, for example, gives each session its own isolated processor, memory and file system. A session can last up to eight hours and is shut down after 15 idle minutes (&lt;a class="link" href="https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-sessions.html" target="_blank" rel="noopener"
 &gt;AWS documentation&lt;/a&gt;).&lt;/p&gt;
&lt;p&gt;Making this cheap is an engineering feat. Firecracker, the open-source technology behind Amazon&amp;rsquo;s serverless services, boots a sealed computer in under 125 milliseconds and adds less than 5 MB of memory overhead to each one (&lt;a class="link" href="https://www.usenix.org/system/files/nsdi20-paper-agache.pdf" target="_blank" rel="noopener"
 &gt;Firecracker paper&lt;/a&gt;).&lt;/p&gt;
&lt;p&gt;&lt;img alt="From one shared app to one sealed computer per job. In the traditional cloud one program serves millions of people at once, the code is written and reviewed by developers, and the scarce thing is server utilization. In the agent cloud each job gets a sealed computer for seconds to days, the code was written by a model seconds ago, and the scarce things are isolation, memory and start-up time." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/ai-agent-cloud-every-agent-gets-its-own-computer-docn-akam-stx-2026-11-07/old-vs-new.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;The idea comes from Jamin Ball&amp;rsquo;s &lt;a class="link" href="https://cloudedjudgement.substack.com/p/clouded-judgement-11626-the-agent" target="_blank" rel="noopener"
 &gt;Clouded Judgement issue of November 6&lt;/a&gt;, which argues that agent cloud infrastructure will end up dwarfing classic cloud infrastructure. That is his opinion. The issue gives no usage numbers.&lt;/p&gt;
&lt;p&gt;The old cloud was built to squeeze more work out of each server. The agent cloud is built to keep jobs apart, which uses more CPU, memory and storage for the same work.&lt;/p&gt;
&lt;h2 id="why-this-needs-chips-and-which-ones"&gt;Why this needs chips, and which ones
&lt;/h2&gt;&lt;p&gt;Here is how that turns into chip demand.&lt;/p&gt;
&lt;p&gt;First, a sealed computer needs processor cores (CPUs) and server memory (DRAM). It does not need a GPU. The GPU is busy somewhere else, running the AI model itself.&lt;/p&gt;
&lt;p&gt;Second, when an agent pauses, its memory is written out to fast flash storage (SSD) so the job can resume later. Pausing does not remove the need for processors and memory. It moves part of it to storage. Logs, saved sessions and the traces used to train agents pile up in bulk storage.&lt;/p&gt;
&lt;p&gt;Third, more of these computers sit next to every AI chip. &lt;a class="link" href="https://insights.trendforce.com/p/agentic-ai-cpu-gpu" target="_blank" rel="noopener"
 &gt;TrendForce&lt;/a&gt; wrote in April that today&amp;rsquo;s ratio of server CPUs to AI chips is roughly 1 to 4 up to 1 to 8, and that agentic AI is expected to push it to between 1 to 1 and 1 to 2. We take the end of each range with the fewest CPUs, 1 to 8 and 1 to 2, so the arithmetic is not the most generous reading. For every million AI chips, that is 125,000 CPUs before and 500,000 after. The difference, 375,000 CPUs, is about 188,000 two-socket servers.&lt;/p&gt;
&lt;p&gt;&lt;img alt="Server CPUs needed per 1 million AI chips. Before, at 1 CPU per 8 AI chips, 125,000. In the agent era, at 1 CPU per 2 AI chips, 500,000. The difference is 375,000 CPU sockets, about 188,000 two-socket servers. Ratios from TrendForce (April 14, 2026): today about 1 to 4 up to 1 to 8, expected 1 to 1 to 1 to 2. We use the ends with the fewest CPUs. Arithmetic ours, assuming 2 CPU sockets per server." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/ai-agent-cloud-every-agent-gets-its-own-computer-docn-akam-stx-2026-11-07/cpu-attach.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;A size check keeps the story honest. If 10 million consumer agents each held an 8 GB computer at the same moment, that would be 80 petabytes of server memory. That is small next to the chip market. The big numbers come from the ratio above and from always-on agents that work for companies around the clock.&lt;/p&gt;
&lt;p&gt;That last piece is the largest gap in our analysis. The CPU ratio rides on the number of AI chips installed, so if AI chip spending slows, this demand slows with it. Always-on company agents are the only part that can grow independently, and nobody has measured how big they are.&lt;/p&gt;
&lt;h2 id="who-gets-the-money"&gt;Who gets the money
&lt;/h2&gt;&lt;p&gt;The new layers of software that agents need are mostly outside the stock market.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Layer&lt;/th&gt;
 &lt;th&gt;What it does, in plain words&lt;/th&gt;
 &lt;th&gt;Who captures the value&lt;/th&gt;
 &lt;th&gt;Listed stocks tied to it&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Sealed computers (sandboxes)&lt;/td&gt;
 &lt;td&gt;Give each job an isolated computer&lt;/td&gt;
 &lt;td&gt;Amazon, Microsoft and Google&amp;rsquo;s own services; private start-ups E2B, Daytona and Modal&lt;/td&gt;
 &lt;td&gt;Cloudflare, Akamai, DigitalOcean (their own versions)&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Saved state&lt;/td&gt;
 &lt;td&gt;Pause, resume and copy a job&amp;rsquo;s memory and files&lt;/td&gt;
 &lt;td&gt;Private start-ups and the big clouds&amp;rsquo; storage&lt;/td&gt;
 &lt;td&gt;Dropbox (buying Tigris), Cloudflare; flash and disk makers Sandisk, Micron, Seagate, Western Digital&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Running the AI model&lt;/td&gt;
 &lt;td&gt;Generate the answers&lt;/td&gt;
 &lt;td&gt;Big clouds, private inference firms&lt;/td&gt;
 &lt;td&gt;Nvidia, AMD, CoreWeave, Nebius, Oracle&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Routers&lt;/td&gt;
 &lt;td&gt;Pick the best model for each job on cost and quality&lt;/td&gt;
 &lt;td&gt;Private gateways such as OpenRouter, Cloudflare&lt;/td&gt;
 &lt;td&gt;Cloudflare (a small revenue pool)&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Training arenas&lt;/td&gt;
 &lt;td&gt;Practice grounds where agents learn by running code and tests&lt;/td&gt;
 &lt;td&gt;The AI labs themselves, private firms&lt;/td&gt;
 &lt;td&gt;AMD, Intel, Arm (the CPUs); Seagate, Western Digital (logs)&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Base layer&lt;/td&gt;
 &lt;td&gt;Servers, storage and network for everything above&lt;/td&gt;
 &lt;td&gt;Suppliers with pricing power while supply is tight&lt;/td&gt;
 &lt;td&gt;AMD, Intel, Arm, Micron, Sandisk, Seagate, Western Digital, Arista, Astera Labs&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Private companies are already priced richly. Press reports say Modal raised $750 million at a $15.75 billion valuation this fall (secondary source). Public investors are left with two routes. The first is platform companies that already own servers, customers and networks, and are adding agent products on top. The second is suppliers that sell scarce chips and storage to everyone in the table.&lt;/p&gt;
&lt;p&gt;The November 5 sell-off helped the first route get cheaper. Our reading, written up on November 6, is that the drop re-priced how long supplier profits can last and unwound leverage. It did not cut earnings. Agent demand lengthens that duration. Yet the market sold the platform companies together with the suppliers and the borrowers. DigitalOcean fell 7.2% on November 4 and about 16% on November 5.&lt;/p&gt;
&lt;h2 id="a-big-drop-is-not-the-same-as-cheap"&gt;A big drop is not the same as cheap
&lt;/h2&gt;&lt;p&gt;&lt;img alt="A big drop is not the same as cheap. Fall from the 12-month high to the November 6 close. Look-first group: DigitalOcean minus 60.0 percent, 4.4 times 2027 sales (our estimate); Seagate minus 57.9 percent, 12.5 times annualized profit guidance; Akamai minus 48.4 percent, 12.6 times 2026 profit guidance; Micron minus 43.1 percent, 4.5 times annualized profit guidance. Fell too, but price or another risk is the issue: Arm minus 60.3 percent, 84 times; Intel minus 60.1 percent, 56 times; AMD minus 30.9 percent, about 42 times (our estimate); Cloudflare minus 25.5 percent, 28 times annualized revenue." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/ai-agent-cloud-every-agent-gets-its-own-computer-docn-akam-stx-2026-11-07/drop-vs-value.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;Arm and Intel fell as much as DigitalOcean, and they are not cheap. A drop tells you what happened to the price. The valuation check tells you what the price now asks of the company. The next section goes through the first four.&lt;/p&gt;
&lt;h2 id="the-four-we-would-look-at-first"&gt;The four we would look at first
&lt;/h2&gt;&lt;p&gt;This is the order in which we would study them. It is not a recommendation to buy.&lt;/p&gt;
&lt;h3 id="digitalocean-docn-the-plain-cloud-for-small-developers"&gt;DigitalOcean (DOCN): the plain cloud for small developers
&lt;/h3&gt;&lt;p&gt;&lt;strong&gt;What it does.&lt;/strong&gt; DigitalOcean rents cloud servers by the hour to developers, start-ups and small businesses. Its pitch is simple prices and simple tools, against Amazon&amp;rsquo;s huge menu. It now sells Managed Agents and Agent Droplets, its name for ready-made sealed computers for agents.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Why there is a chance.&lt;/strong&gt; The people building AI agents are mostly start-ups and small teams, the customers DigitalOcean already serves. In its third quarter report, AI customers (those using one or more of its AI offerings) ran at $376 million of annual revenue, up 288%, and the volume of AI tokens handled by its inference service grew roughly tenfold from the previous quarter. The CEO says the company expects 60% or more revenue growth in 2027. The market value is $8.55 billion, about 4.4 times our estimate of 2027 revenue of $1.94 billion (the middle of its 2026 guidance, $1.21 billion, grown by exactly 60%). The stock is 60.0% below its high and fell 41.5% in a month, because it was sold with AI-infrastructure borrowers although it mostly rents ordinary servers.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;What could go wrong.&lt;/strong&gt; Its finance-lease obligations, the contracts it uses to pay for equipment over time, rose to $719 million from $99 million at the start of the year for the long-term part alone. With the part due within a year, they are about $924 million against $131 million. That is debt-like. If AI start-up customers leave or cut spending, the equipment stays and the payments continue.&lt;/p&gt;
&lt;h3 id="akamai-akam-the-network-company-that-became-a-cloud"&gt;Akamai (AKAM): the network company that became a cloud
&lt;/h3&gt;&lt;p&gt;&lt;strong&gt;What it does.&lt;/strong&gt; Akamai started by speeding up websites and protecting them from attacks, using servers placed close to users around the world. It then bought the cloud provider Linode in 2022 to rent out servers too. It has now launched Agent Sandbox, a sealed computer for agents, and inference services.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Why there is a chance.&lt;/strong&gt; At $83.09 the stock trades at 12.6 times the profit it guides for 2026 ($6.47 to $6.68 per share). Cloud infrastructure revenue grew 34% to $109 million in the quarter, and the company aims for $1.1 billion to $1.3 billion in 2027. The Anthropic deal, announced September 24, is $11.6 billion over seven years. Akamai expects to spend about $5.5 billion on equipment for it, and Anthropic receives a warrant for up to about 5% of Akamai&amp;rsquo;s shares, which vests in steps as its purchases grow. Akamai&amp;rsquo;s CEO says total revenue is projected to exceed $7 billion in 2028.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;What could go wrong.&lt;/strong&gt; One customer carries much of the growth, and the build cost will show up as depreciation. We do not judge the size of the Anthropic deal. Anthropic is the maker of the AI model that helped us write this post, so treat the Akamai view with that in mind and check the contract terms yourself.&lt;/p&gt;
&lt;h3 id="seagate-stx-the-cheapest-place-to-keep-what-agents-write-down"&gt;Seagate (STX): the cheapest place to keep what agents write down
&lt;/h3&gt;&lt;p&gt;&lt;strong&gt;What it does.&lt;/strong&gt; Seagate makes hard disk drives, the spinning-disk storage in data centers. Most of the world&amp;rsquo;s data sits on disks because they cost the least per terabyte.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Why there is a chance.&lt;/strong&gt; Agents leave a trail: logs, saved sessions and the traces used to train the next agent. That data has to be kept, and it is not worth paying flash prices for. On its October 27 call it said it shipped a record 198 exabytes of data center drives in the quarter (secondary transcript). It guides fiscal second-quarter earnings of $9.20 per share, plus or minus $0.20. Multiplied by four, that is $36.8, and at $460.99 the stock trades at 12.5 times. It is 57.9% below its high and fell 40.5% in a month.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;What could go wrong.&lt;/strong&gt; A low multiple on a peak profit is how cyclical stocks look near the top. If the profit lasts, the price is cheap. If it fades soon, it is not. Pricing gains Seagate has made could reverse if demand cools.&lt;/p&gt;
&lt;h3 id="micron-mu-the-memory-every-sealed-computer-needs"&gt;Micron (MU): the memory every sealed computer needs
&lt;/h3&gt;&lt;p&gt;&lt;strong&gt;What it does.&lt;/strong&gt; Micron makes DRAM, the working memory of servers, and NAND flash, the memory in SSDs.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Why there is a chance.&lt;/strong&gt; Every sealed computer needs working memory, and paused ones are written to flash. At $690.00 and its guidance of $38.15 of earnings per share (plus or minus $1.00) for the November quarter, the stock trades at 4.5 times the annualized figure. Micron said on its June 24 call that its long-term supply agreements are take-or-pay with binding volume commitments, typically five years from 2026 through 2030. Take-or-pay means the customer pays for the agreed volume even if it does not take delivery.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;What could go wrong.&lt;/strong&gt; Memory prices swing, and the high-bandwidth memory part of Micron&amp;rsquo;s business rides on AI chip spending. Micron also moves with the same drivers as Samsung Electronics and SK hynix, so it adds little if you already hold Korean memory.&lt;/p&gt;
&lt;h2 id="the-price-already-counts-on-the-story-amd-and-cloudflare"&gt;The price already counts on the story: AMD and Cloudflare
&lt;/h2&gt;&lt;p&gt;&lt;strong&gt;AMD&lt;/strong&gt; designs the CPUs inside servers (EPYC) and AI chips. Third-quarter data center revenue was $8.546 billion, up 97%. At $449.04 and our estimate of $10.6 in annualized earnings, the stock trades near 42 times. For the stock to rise 50% at a 30-times multiple, earnings per share would need to reach $22.5, 2.1 times our estimate. The agent CPU story can be true and still be mostly in the price.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Cloudflare&lt;/strong&gt; has the best product fit of any listed name. It sells sealed computers, storage, AI model access, a model gateway and a global network. But its enterprise value is about $95 billion by our calculation, or 28 times its fourth-quarter revenue guidance annualized (the guidance range is $845 million to $846 million). It has fallen the least, 25.5%, and the price carries the product lead.&lt;/p&gt;
&lt;h2 id="real-demand-but-another-risk-decides-the-outcome"&gt;Real demand, but another risk decides the outcome
&lt;/h2&gt;&lt;p&gt;&lt;strong&gt;Intel&lt;/strong&gt; grew its data center and AI segment 78% to $7.3 billion. Its CEO says the priority is to unlock more supply for &amp;ldquo;extraordinary demand&amp;rdquo; for its Xeon server chips, and its CFO said on the October 29 call that supply stays tight through at least the first half of 2027 (secondary transcript). The result depends on whether Intel can make enough chips, not on demand. At $56.28 the stock trades at 56 times its annualized profit guidance.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Arm&lt;/strong&gt; licenses CPU designs. It trades at 84 times annualized profit guidance (secondary source). Press reports say SoftBank has a $25 billion margin loan backed by Arm shares. SoftBank&amp;rsquo;s May filings showed about 769 million Arm shares, which the November 6 close values at about $134 billion, so the loan is about 19% of that value (the share count is a May figure). That risk has nothing to do with agent demand, and we pass.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;CoreWeave, Nebius and Oracle&lt;/strong&gt; rent AI chips and depend on a small number of AI labs and on financing. Our October 26 view on them has not changed. CoreWeave reports on November 10 and Nebius on November 12.&lt;/p&gt;
&lt;h2 id="what-would-prove-us-wrong"&gt;What would prove us wrong
&lt;/h2&gt;&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;If this happens&lt;/th&gt;
 &lt;th&gt;Then&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;The big cloud providers bundle agent computers into their standard plans&lt;/td&gt;
 &lt;td&gt;DigitalOcean and Akamai lose their edge. Only the chip and storage suppliers keep their chance.&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;DigitalOcean&amp;rsquo;s AI-customer revenue grows less than 10% from the previous quarter in its February report&lt;/td&gt;
 &lt;td&gt;We drop it from the first group.&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Akamai lowers its $1.1 billion to $1.3 billion 2027 cloud infrastructure outlook&lt;/td&gt;
 &lt;td&gt;We drop it from the first group.&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Lead times for server CPUs return to 20 weeks or less&lt;/td&gt;
 &lt;td&gt;The CPU path weakens.&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;The first row is the test that matters most. If agent jobs stay on independent platforms and customers keep paying, the platform companies earn the profit. If the big clouds absorb them, the chance moves to the suppliers.&lt;/p&gt;
&lt;h2 id="what-to-check-next"&gt;What to check next
&lt;/h2&gt;&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;When&lt;/th&gt;
 &lt;th&gt;What&lt;/th&gt;
 &lt;th&gt;Why it matters&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;November 10&lt;/td&gt;
 &lt;td&gt;CoreWeave third-quarter results&lt;/td&gt;
 &lt;td&gt;Cuts or financing delays would harden our view on AI-chip renters, and we keep them apart from the platform companies&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;November 12&lt;/td&gt;
 &lt;td&gt;Nebius third-quarter results. Check how DigitalOcean and Akamai trade the next day&lt;/td&gt;
 &lt;td&gt;Shows whether the market still sells the platform companies together with the borrowers&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Late November to early December&lt;/td&gt;
 &lt;td&gt;Everpure (formerly Pure Storage) and NetApp earnings&lt;/td&gt;
 &lt;td&gt;Any wording on demand for saved agent state&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Mid-December&lt;/td&gt;
 &lt;td&gt;Micron fiscal first-quarter results&lt;/td&gt;
 &lt;td&gt;Number of long-term agreements and server DRAM wording&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Late January to early February&lt;/td&gt;
 &lt;td&gt;Intel and AMD fourth-quarter results&lt;/td&gt;
 &lt;td&gt;2027 server CPU guidance and lead times&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;February 2027&lt;/td&gt;
 &lt;td&gt;DigitalOcean and Akamai fourth-quarter results&lt;/td&gt;
 &lt;td&gt;The two tests in the table above&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;h2 id="what-this-means-for-korean-investors"&gt;What this means for Korean investors
&lt;/h2&gt;&lt;p&gt;Micron and the two Korean memory makers share the same drivers, so Micron is a read-through for them, not a separate bet. The agent path adds a source of ordinary server DRAM demand that does not depend on high-bandwidth memory. It also makes the 2028 oversupply risk we have written about a little smaller. We do not treat it as a reason to change that view.&lt;/p&gt;
&lt;h2 id="data-notes"&gt;Data notes
&lt;/h2&gt;&lt;ul&gt;
&lt;li&gt;Prices are US regular-session closes on November 6, 2026. &amp;ldquo;Annualized&amp;rdquo; means the latest company quarterly guidance multiplied by four. We used no analyst consensus, because we could not find a dated source.&lt;/li&gt;
&lt;li&gt;The AMD earnings figure is our estimate: its fourth-quarter revenue guidance times the third-quarter profit margin, divided by shares outstanding.&lt;/li&gt;
&lt;li&gt;DigitalOcean&amp;rsquo;s value is market value, because we could not confirm its convertible debt. Its 2027 revenue is our assumption.&lt;/li&gt;
&lt;li&gt;Seagate&amp;rsquo;s and Micron&amp;rsquo;s guidance comes from their 8-K press releases. Seagate&amp;rsquo;s exabyte figure and Intel&amp;rsquo;s remark that supply is tight come from secondary call transcripts, and Arm&amp;rsquo;s results from a secondary summary. The Modal figure and the SoftBank loan come from press reports.&lt;/li&gt;
&lt;li&gt;We did not read the question-and-answer part of AMD&amp;rsquo;s third-quarter call.&lt;/li&gt;
&lt;li&gt;A single AI vendor reviewed this analysis. No second vendor checked it.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;&lt;strong&gt;Where does the money end up? Reply with A, B or C, and one data point that backs your pick.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;A. In agent clouds: DigitalOcean and Akamai keep the customers.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;B. In storage and memory: Seagate and Micron sell to everyone.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;C. In neither: the big cloud providers bundle it and the chance disappears.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;This is for research and information only, not investment advice. Estimates and scenarios are analytical assumptions, not trading instructions or individual portfolio decisions. Figures are dated data, and some reported figures are secondary compilations. We used Anthropic&amp;rsquo;s Claude for research and editing. Anthropic appears in this post as Akamai&amp;rsquo;s largest contract counterparty. We report the contract as announced and do not weigh its size.&lt;/p&gt;</description></item><item><title>SK hynix -61%, 2027 Profit Estimates Up: AI Stocks Stopped Asking Whether Demand Is Real and Started Asking Who Pays</title><link>https://koreainvestinsights.com/post/ai-narrative-demand-to-money-to-permission-credit-decides-2026-11-07/</link><pubDate>Sat, 07 Nov 2026 20:50:00 +0900</pubDate><guid>https://koreainvestinsights.com/post/ai-narrative-demand-to-money-to-permission-credit-decides-2026-11-07/</guid><description>&lt;img src="https://koreainvestinsights.com/" alt="Featured image of post SK hynix -61%, 2027 Profit Estimates Up: AI Stocks Stopped Asking Whether Demand Is Real and Started Asking Who Pays" /&gt;&lt;p&gt;Since September 23, the KOSPI is down &lt;strong&gt;28.5%&lt;/strong&gt;. From their June highs, Samsung Electronics is down &lt;strong&gt;47.9%&lt;/strong&gt; and SK hynix &lt;strong&gt;61.2%&lt;/strong&gt;. The Philadelphia Semiconductor Index (SOX) fell &lt;strong&gt;38.4%&lt;/strong&gt; from June 22 to its November 5 low.&lt;/p&gt;
&lt;p&gt;Over the same weeks, the profit side got stronger. Samsung booked KRW 107.4 trillion of operating profit in the third quarter. TrendForce expects fourth-quarter conventional DRAM contract prices to rise another 10-15% (&lt;a class="link" href="https://www.trendforce.com/presscenter/news/20260930-13258.html" target="_blank" rel="noopener"
 &gt;TrendForce, September 30&lt;/a&gt;). All four big US hyperscalers raised their 2027 capex on third-quarter earnings calls. Korea&amp;rsquo;s October chip exports reached $56.44 billion, up 259.4% from a year earlier.&lt;/p&gt;
&lt;p&gt;Both facts are true because the market&amp;rsquo;s question kept changing. In late September it asked whether AI infrastructure could keep its pricing power at high rates. In mid-October it asked how long record profits last. By late October it asked who pays. This week it asked whether politics will keep allowing the build-out. Each question sits further from the one thing that improved, which is profit.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Our call: the price verdict is bear. The 2027 profit verdict is bull. The variable that connects them is credit. The high-yield spread closed at 348bp on November 5, 2bp below the 350bp line we watch.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;Both verdicts can be tested against lines you can watch: 330bp and 350bp on the high-yield spread. At the end we ask which side you take.&lt;/p&gt;
&lt;h2 id="prices-fell-like-a-bear-market-while-profit-estimates-rose"&gt;Prices fell like a bear market while profit estimates rose
&lt;/h2&gt;&lt;p&gt;&lt;img alt="Drawdown from the stated high to the latest close. SK hynix -61.2% versus its June 22 high, Samsung Electronics -47.9% versus June 18, the SOX chip index -38.4% versus June 22, the KOSPI -28.5% versus its September 23 rebound high, the Nasdaq-100 -13.6% and the S&amp;P 500 -5.7% versus October 6. Same weeks on the profit side: Samsung Q3 operating profit KRW 107.4 trillion, FY27 EPS estimates up 0.7% for Samsung and 2.0% for SK hynix between October 1 and 23, and Korea’s October chip exports of $56.44 billion, up 259.4% from a year earlier." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/ai-narrative-demand-to-money-to-permission-credit-decides-2026-11-07/drawdowns.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;Korean memory&amp;rsquo;s big drop started in June, not October. Samsung peaked on June 18 and SK hynix on June 22. October added a second leg: from October 6 to 15, Samsung fell 22.8%, SK hynix 28.8% and the KOSPI 20.0%.&lt;/p&gt;
&lt;p&gt;The US looks different. The S&amp;amp;P 500 is 5.7% below its October 6 record close, and the Nasdaq-100 is 13.6% below its high. The gap comes from how sensitive each company&amp;rsquo;s revenue is to the business cycle. The side that buys AI demand fell least. The side that sells components fell most.&lt;/p&gt;
&lt;h2 id="the-question-moved-four-times-each-time-away-from-profit"&gt;The question moved four times, each time away from profit
&lt;/h2&gt;&lt;p&gt;&lt;img alt="The market’s question moved four times. Sep 18 to Oct 6: can AI infrastructure keep its pricing power at high rates. Oct 7 to 16: how long do record profits last and what does capital cost. Oct 17 to 31: demand is real, so who pays for it. Nov 1 to 7: will politics keep allowing it." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/ai-narrative-demand-to-money-to-permission-credit-decides-2026-11-07/question-ladder.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;The October turn shows it best. Samsung&amp;rsquo;s record KRW 107.4 trillion landed 1.2% above one consensus (LSEG, KRW 106.1 trillion) and 4.7% below another (Visible Alpha, KRW 112.73 trillion), and the two use different bases. The KOSPI still fell 2.62% that day. The argument had moved from the size of profit to its duration.&lt;/p&gt;
&lt;p&gt;Then it moved to money. On October 14, CoreWeave fell 15.5% and Oracle 8.8% in US trading. Both fell more than Micron (-4.1%). Oracle&amp;rsquo;s five-year credit default swap (CDS) stood at 288bp, up 26bp. CoreWeave&amp;rsquo;s 2032 bond yielded 14.8%. The market was selling the financing chain, and memory sat inside it.&lt;/p&gt;
&lt;p&gt;By the third quarter calls the demand evidence was hard to argue with. Amazon booked most of 2027 and a large part of 2028. Google&amp;rsquo;s API token volume rose 40% quarter on quarter. North American data center vacancy stood at 1.4%. But the five companies&amp;rsquo; combined third-quarter free cash flow was -$31.8 billion, and capex in the latest quarter ran at 116% of operating cash flow. External funding needs for 2027 come to about $200 billion, with guarantee paths of $48 billion at Alphabet and $41.5 billion at Meta.&lt;/p&gt;
&lt;h2 id="the-bull-case-2027-profit-is-already-sealed"&gt;The bull case: 2027 profit is already sealed
&lt;/h2&gt;&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Argument&lt;/th&gt;
 &lt;th&gt;Evidence&lt;/th&gt;
 &lt;th&gt;Source&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;2027 profit is locked&lt;/td&gt;
 &lt;td&gt;SK hynix long-term deals above 50% of volume. Three makers&amp;rsquo; 2027 volume commitments unchanged. FY27 EPS estimates up between Oct 1 and Oct 23 (Samsung +0.7%, SK hynix +2.0%). Q4 DRAM contract prices seen +10-20%&lt;/td&gt;
 &lt;td&gt;Company data, consensus&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Korean memory activity is strong&lt;/td&gt;
 &lt;td&gt;October chip exports $56.44bn (+259.4%), memory $50.14bn (+360.5%), 14.4% above the Q3 monthly average. The September-to-October dip repeats last year&amp;rsquo;s calendar pattern&lt;/td&gt;
 &lt;td&gt;Korean trade ministry, Nov 1&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Supply, not demand, sets the 2027 ceiling&lt;/td&gt;
 &lt;td&gt;Google API token growth: +60% (Q1), +37% (Q2), +40% (Q3, quarter on quarter). All four hyperscalers raised 2027 capex. Amazon booked most of 2027&lt;/td&gt;
 &lt;td&gt;Earnings calls, our model&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Cheap&lt;/td&gt;
 &lt;td&gt;FY27 P/E: Samsung 3.12x, SK hynix 2.81x (Oct 26). 12-month forward P/B: 1.25x and 1.27x (Nov 6)&lt;/td&gt;
 &lt;td&gt;Consensus, secondary&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Fear is high and shorts are crowded&lt;/td&gt;
 &lt;td&gt;VIX 28.18 on Nov 5, the highest since Mar 31. FINRA short interest on Oct 15 versus Sep 30: EWY +57.3%, SOXX +35.5%&lt;/td&gt;
 &lt;td&gt;FINRA, secondary&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Credit has not followed stocks down&lt;/td&gt;
 &lt;td&gt;Nov 3 to 5: SOX -15.0% while the high-yield ETF HYG rose 0.4%. Spread 340bp to 348bp. The won held at 1,341 on Nov 5&lt;/td&gt;
 &lt;td&gt;FRED, Yahoo Finance&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Macro pressure eased&lt;/td&gt;
 &lt;td&gt;US 10-year 5.31% (Oct 5) to 4.859% (Nov 6). Brent front month $113.43 (Oct 29) to $97.76 (Nov 6)&lt;/td&gt;
 &lt;td&gt;Market panel&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;h2 id="the-bear-case-money-is-the-weakest-link"&gt;The bear case: money is the weakest link
&lt;/h2&gt;&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Argument&lt;/th&gt;
 &lt;th&gt;Evidence&lt;/th&gt;
 &lt;th&gt;Source&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;The policy cushion thinned&lt;/td&gt;
 &lt;td&gt;Democrats won both chambers on Nov 3. Texas&amp;rsquo;s TCEQ paused permits on Sep 21, and a winning candidate has promised new halts. New York Executive Order 62 covers 50MW-plus projects for up to a year. 38-40 local restriction bills passed. 56% of voters call AI and data centers an important issue&lt;/td&gt;
 &lt;td&gt;Executive order, secondary tallies&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Funding is the weakest link&lt;/td&gt;
 &lt;td&gt;High yield 266bp to 348bp (+82bp). CCC 1,075bp to 1,356bp (Oct 23 high). Investment grade 77bp to 93bp. Oracle CDS reported above 300bp (secondary, not confirmed first-hand). CoreWeave unsecured debt priced at 15% or more. OpenAI&amp;rsquo;s funding target cut from $30bn to $20bn (secondary)&lt;/td&gt;
 &lt;td&gt;FRED, secondary&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;The Fed still leans toward hiking&lt;/td&gt;
 &lt;td&gt;Hiked on Sep 16, held on Oct 28 with two dissents for a hike. December hike odds 67.7-81% (secondary sources conflict). October payrolls +165,000 versus +65,000 expected. September core PCE 3.1%&lt;/td&gt;
 &lt;td&gt;&lt;a class="link" href="https://www.federalreserve.gov/newsevents/pressreleases/monetary20260916a1.htm" target="_blank" rel="noopener"
 &gt;Federal Reserve&lt;/a&gt;, secondary&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;AI labs face a profit scissors&lt;/td&gt;
 &lt;td&gt;Anthropic&amp;rsquo;s quarterly revenue growth fell from +144.2% to +46.7% quarter on quarter, compute cost rose from 48.1% to 58.2% of revenue, and Q3 showed a GAAP operating loss (&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/2130357/000162828026068676/anthropicpbc-sx1ax1.htm" target="_blank" rel="noopener"
 &gt;S-1/A&lt;/a&gt;). Blended token prices fell 26.8% in four weeks while B200 rental rose 10.6%&lt;/td&gt;
 &lt;td&gt;SEC filing, one non-representative sample&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;2028 supply and the bullwhip&lt;/td&gt;
 &lt;td&gt;The big four plan 15-17% more 2028 wafers (CXMT included). Buyer inventories sit 1-5 weeks above normal, 8-13% of 2027 shipments&lt;/td&gt;
 &lt;td&gt;Model&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Memory spot turned for the first time&lt;/td&gt;
 &lt;td&gt;DDR5 16Gb spot from $61.50 to $60.167 (-2.17%) by Nov 6. DDR4 8Gb from $47.107 (Oct 23) to $45.607. NAND and DDR4 fell together, as in March-April&lt;/td&gt;
 &lt;td&gt;Market panel&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;The policy row needs one more line. The market used to treat Washington as a backstop for the AI build-out, an &amp;ldquo;AI put&amp;rdquo;. Our reading is that it is weaker at the federal level and now has state-level brakes. The same election also pulled Brent and the 10-year yield down, which lowered macro risk. Counting &amp;ldquo;AI put weaker&amp;rdquo; and &amp;ldquo;rates and oil lower&amp;rdquo; as independent factors double counts one event.&lt;/p&gt;
&lt;h2 id="why-credit-decides"&gt;Why credit decides
&lt;/h2&gt;&lt;p&gt;&lt;img alt="Four variables, one that decides. V1 2027 profit delivery leans bull. V2 funding cost and credit leans bear and is marked as the deciding variable. V3 political permission leans bear and moves slowly. V4 positioning is bearish over 1-2 weeks and bullish over 1-3 months." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/ai-narrative-demand-to-money-to-permission-credit-decides-2026-11-07/four-variables.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;Political limits and AI-lab profitability both reach 2028 orders through one channel: financing. When the policy cushion thins, return on invested capital and the durability of free cash flow become the only backstops. The market then asks harder about credit. This is why strong 2027 profit cannot lift the price discount yet. Price will not release its duration discount until credit settles.&lt;/p&gt;
&lt;p&gt;&lt;img alt="High-yield spread in basis points. 266 on Sep 21, 323 on Oct 15, 344 on Oct 30, 340 on Nov 3 and 348 on Nov 5. Two lines are drawn: 350bp, above which for five sessions confirms a credit squeeze, and 330bp, at or below which for five sessions reverses it. The latest value is 2bp below the 350 line. CCC spreads rose from 1,075bp to 1,356bp and investment grade from 77bp to 93bp." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/ai-narrative-demand-to-money-to-permission-credit-decides-2026-11-07/credit-gate.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;The chart carries the best bull argument and the best bear argument at once. In the two sessions when SOX fell 15%, spreads widened only 8bp, so credit has not followed equities down. But the spread has widened 82bp since September 21, and it sits 2bp from the line where the funding story stops being a worry and becomes a fact (&lt;a class="link" href="https://fred.stlouisfed.org/series/BAMLH0A0HYM2" target="_blank" rel="noopener"
 &gt;FRED&lt;/a&gt;).&lt;/p&gt;
&lt;h2 id="what-flips-the-call"&gt;What flips the call
&lt;/h2&gt;&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;If this happens&lt;/th&gt;
 &lt;th&gt;Then&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Nvidia&amp;rsquo;s fiscal third-quarter report holds its fourth-quarter guidance and shows supply commitments of $279 billion or more, and the November issuance window clears with high yield at or below 330bp for five sessions&lt;/td&gt;
 &lt;td&gt;The call turns bull on price. Credit unlocks and 2027 profit takes over the price.&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;High yield stays above 350bp for five sessions&lt;/td&gt;
 &lt;td&gt;Bear confirmed: the credit-squeeze scenario is on.&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Any one hyperscaler cuts its 2027 capex&lt;/td&gt;
 &lt;td&gt;Bear confirmed: the demand-slowdown scenario is on.&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;The late-December outlook for first-quarter 2027 conventional contract prices turns flat or lower&lt;/td&gt;
 &lt;td&gt;Bear confirmed, and it breaks the 2027 profit leg itself.&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;The two bull conditions must arrive together. One is not enough. The bear conditions work alone.&lt;/p&gt;
&lt;h2 id="where-we-could-be-wrong"&gt;Where we could be wrong
&lt;/h2&gt;&lt;p&gt;&lt;strong&gt;Objection one: &amp;ldquo;A 15% two-day drop in chips is forced selling, not a financing story.&amp;quot;&lt;/strong&gt; Partly right, and it is right for the next one to two weeks. Trend-following funds and retail leverage dominate that window. Our credit claim is about one to three months, when crowded shorts turn into rebound fuel instead of pressure.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Objection two: &amp;ldquo;Fear at 28 and a P/E under 3.2x say buy.&amp;quot;&lt;/strong&gt; Fear and low multiples describe the price. They say nothing about who finances 2028. We do not treat them as evidence for a bottom by themselves.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Objection three: &amp;ldquo;2027 profit is sealed, so 2028 is far away.&amp;quot;&lt;/strong&gt; Stocks price duration, not one year. Buyer inventories sit above normal and the big four plan 15-17% more wafers for 2028. If orders for 2028 depend on credit, the 2027 profit does not release the price.&lt;/p&gt;
&lt;h2 id="what-to-check-next"&gt;What to check next
&lt;/h2&gt;&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;When&lt;/th&gt;
 &lt;th&gt;What&lt;/th&gt;
 &lt;th&gt;Where&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Every trading day&lt;/td&gt;
 &lt;td&gt;High-yield, CCC and investment-grade spreads&lt;/td&gt;
 &lt;td&gt;FRED&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;By November 10&lt;/td&gt;
 &lt;td&gt;Nvidia&amp;rsquo;s fiscal third-quarter report date (our files show November 18, 19 and 20)&lt;/td&gt;
 &lt;td&gt;Nvidia investor relations&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;November 11 and 19&lt;/td&gt;
 &lt;td&gt;Weekly DDR5 and DDR4 spot. A weekly DDR5 average at or below $57.44 is our alert line&lt;/td&gt;
 &lt;td&gt;Spot price panel&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;December FOMC&lt;/td&gt;
 &lt;td&gt;Hike odds (about two-thirds or more on secondary sources, which conflict)&lt;/td&gt;
 &lt;td&gt;CME FedWatch&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Late December&lt;/td&gt;
 &lt;td&gt;First-quarter 2027 conventional contract-price outlook&lt;/td&gt;
 &lt;td&gt;TrendForce&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;h2 id="what-this-means"&gt;What this means
&lt;/h2&gt;&lt;p&gt;Korean memory prices already paid for a cyclical bear market, while 2027 profit estimates went up. The gap between them closes when credit decides, and not before.&lt;/p&gt;
&lt;p&gt;Over one to two weeks, selling by trend-following funds and by retail leverage dominates. Over one to three months, crowded shorts increase the size of any rebound. The S&amp;amp;P 500 has already broken 7,544, a mid-term line that trend-following funds watch. Their long-term line sits at 7,040.&lt;/p&gt;
&lt;p&gt;This is not a call to buy. Fear gauges and low multiples alone are not a reason.&lt;/p&gt;
&lt;h2 id="data-notes"&gt;Data notes
&lt;/h2&gt;&lt;ul&gt;
&lt;li&gt;We collected no new external data for this post. Figures come from our dated market panel and notes. Where a note marked a figure as secondary, we kept that label.&lt;/li&gt;
&lt;li&gt;Brent differs by contract month across sources. We used only the front-month series in our panel ($113.43 on Oct 29, $97.76 on Nov 6).&lt;/li&gt;
&lt;li&gt;The November 6 KOSPI close is the KRX value, 5,059.69. A Yahoo Finance value of 5,042.88 appears in other notes.&lt;/li&gt;
&lt;li&gt;Nvidia&amp;rsquo;s report date differs across our files. Check it at the source.&lt;/li&gt;
&lt;li&gt;The only SK hynix September-end close we hold is the September 30 panel value, and we did not carry Micron&amp;rsquo;s October 1 earnings figures.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;&lt;strong&gt;Which breaks first? Reply with A, B or C, and one data point that backs your pick.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;A. Credit: high yield stays above 350bp for five sessions.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;B. Capex: a hyperscaler cuts its 2027 plan.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;C. Neither: Nvidia&amp;rsquo;s guidance holds, the November issuance window clears and high yield falls to 330bp or lower.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;This is for research and information only, not investment advice. Estimates and scenarios are analytical assumptions, not trading instructions or individual portfolio decisions. Figures are dated data, and some reported figures are secondary compilations. We used Anthropic&amp;rsquo;s Claude for research and editing. Anthropic appears in this post as an AI lab whose filed S-1/A figures we report as filed and do not weigh.&lt;/p&gt;</description></item><item><title>Marvell Already Paid for Lumentum. For the Next Year, the Stock Trades on AI Optics.</title><link>https://koreainvestinsights.com/post/marvell-lumentum-terms-priced-ai-optics-cycle-2026-11-02/</link><pubDate>Mon, 02 Nov 2026 22:22:00 +0900</pubDate><guid>https://koreainvestinsights.com/post/marvell-lumentum-terms-priced-ai-optics-cycle-2026-11-02/</guid><description>&lt;img src="https://koreainvestinsights.com/" alt="Featured image of post Marvell Already Paid for Lumentum. For the Next Year, the Stock Trades on AI Optics." /&gt;&lt;p&gt;On Monday, November 2, Marvell Technology agreed to buy Lumentum in an all-stock deal. Lumentum holders get &lt;strong&gt;4.5 Marvell shares for each share&lt;/strong&gt;, a 23.9% premium on unaffected prices. The deal needs no new debt. Marvell promises more than $1 billion a year of cost synergies and expects to close in the first half of 2028. The terms are better than the market feared. The stock did not bounce. It fell again. In pre-market trading at 8:27 a.m. New York time, Marvell was down 3.1% and Lumentum was down 5.8%.&lt;/p&gt;
&lt;p&gt;The common read is that investors hate the price. Our read is that they have already charged for it, and then some. On October 29, the day the talks leaked, Marvell fell 11.8 percentage points more than the Philadelphia semiconductor index (SOX). That gap is about &lt;strong&gt;$26.7 billion&lt;/strong&gt; of market value. The premium Marvell agreed to pay is about &lt;strong&gt;$22.8 billion&lt;/strong&gt;. The promised synergies, after tax and valued at 25 times, are worth about &lt;strong&gt;$21.3 billion&lt;/strong&gt;, close to the full premium. The stock price gives them nothing.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Our judgment: for the next 6 to 12 months, the AI optics spending cycle will drive Marvell&amp;rsquo;s stock more than the deal terms will.&lt;/strong&gt; The exchange ratio is fixed, so Marvell and Lumentum now move together, and Marvell&amp;rsquo;s sensitivity to optics rises. &lt;strong&gt;We lean neutral to mildly positive.&lt;/strong&gt; The price and the dilution are settled, and the stock already carries them. The synergies are as large as the premium, and the stock gives them zero value. That puts the asymmetry on the upside. Lumentum&amp;rsquo;s business is also speeding up. Its revenue guidance for the December quarter is 27% above the September quarter, and its second co-packaged optics (NPO/CPO) win comes with more than $5 billion of purchase commitments. The variable that could flip this view is customer neutrality.&lt;/p&gt;
&lt;p&gt;This article is dated November 2, 2026, Korea time. Closing prices are Nasdaq closes through October 30. Pre-market quotes are from 8:27 a.m. ET on November 2, before the regular session. We read Marvell&amp;rsquo;s press release and investor presentation, the merger agreement Marvell filed on Form 8-K this morning, and Lumentum&amp;rsquo;s first-quarter results on Form 8-K. Our &lt;a class="link" href="https://koreainvestinsights.com/post/marvell-lumentum-talks-laser-scarcity-cpo-stock-deal-2026-10-29/" target="_blank" rel="noopener"
 &gt;first take on the reported talks&lt;/a&gt; and our &lt;a class="link" href="https://koreainvestinsights.com/post/marvell-acquisition-playbook-lumentum-outside-it-2026-10-30/" target="_blank" rel="noopener"
 &gt;look at Marvell&amp;rsquo;s deal record&lt;/a&gt; set up the tests we use below.&lt;/p&gt;
&lt;p&gt;&lt;img alt="Marvell’s value lost against the SOX on October 29 was about $26.7 billion. The premium to Lumentum holders is about $22.8 billion. The value of $1 billion a year of cost synergies, after a 15% tax and at 25 times, is about $21.3 billion, and the stock price gives it zero." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/marvell-lumentum-terms-priced-ai-optics-cycle-2026-11-02/price-paid-synergy-free.en.png"&gt;
&lt;/p&gt;
&lt;h2 id="what-to-take-away"&gt;What to take away
&lt;/h2&gt;&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;The terms beat the fear on funding, meet it on price and miss it on timing.&lt;/strong&gt; All stock and zero debt remove the credit-rating risk. A 23.9% premium sits between our 20% &amp;ldquo;good&amp;rdquo; line and our 30% &amp;ldquo;bad&amp;rdquo; line. A close in the first half of 2028 is longer than the 9 to 12 months that precedents suggested.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;The stock already paid more than the premium.&lt;/strong&gt; The October 29 loss beyond the SOX ($26.7 billion) is larger than the premium ($22.8 billion). The synergy target is worth about the same as the premium and is priced at zero. The confirmed terms brought no rebound before the open.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;For a year, Marvell is an optics stock.&lt;/strong&gt; With a fixed ratio, Lumentum trades as 4.5 Marvell shares less a spread. On current guidance, Lumentum would supply about 38% of the combined company&amp;rsquo;s earnings. On October 30, Marvell fell about as much as the optics group did.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Lumentum is accelerating into the deal.&lt;/strong&gt; First-quarter revenue grew 25.5% from the prior quarter. Guidance points to about $1.6 billion next quarter at an operating margin of 45% or more. Every dollar Lumentum earns before closing is bought with a fixed number of Marvell shares.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Customer neutrality decides the long-term value.&lt;/strong&gt; If Nvidia walks away from its purchase commitments or a large module maker qualifies Broadcom&amp;rsquo;s signal processors, we turn negative.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id="the-terms-beat-the-fear-on-funding-and-miss-it-on-timing"&gt;The terms beat the fear on funding and miss it on timing
&lt;/h2&gt;&lt;p&gt;The question for this table is simple. Compared with what the market expected before Monday, were the announced terms better or worse?&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Item&lt;/th&gt;
 &lt;th&gt;Expected before Monday&lt;/th&gt;
 &lt;th&gt;Announced&lt;/th&gt;
 &lt;th&gt;Verdict&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Payment&lt;/td&gt;
 &lt;td&gt;All stock was the likely case (CNBC, October 30)&lt;/td&gt;
 &lt;td&gt;All stock, no new debt&lt;/td&gt;
 &lt;td&gt;As expected. The credit-rating risk is gone&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Premium&lt;/td&gt;
 &lt;td&gt;&amp;ldquo;Not a big premium&amp;rdquo; (CNBC). Our lines: 20% or less is good, above 30% is bad&lt;/td&gt;
 &lt;td&gt;23.9%&lt;/td&gt;
 &lt;td&gt;Middle&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;New Marvell shares&lt;/td&gt;
 &lt;td&gt;About +42% to +44% in our earlier cash-and-stock case&lt;/td&gt;
 &lt;td&gt;About +50.6% (470 million new shares on 930 million)&lt;/td&gt;
 &lt;td&gt;More, because there is no cash&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Earnings per share&lt;/td&gt;
 &lt;td&gt;Analysts warned of dilution&lt;/td&gt;
 &lt;td&gt;Company: higher non-GAAP EPS in the first year after closing&lt;/td&gt;
 &lt;td&gt;Better than feared, but the gain depends on synergies&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Synergies&lt;/td&gt;
 &lt;td&gt;Our bull line: $1 billion a year, with a date&lt;/td&gt;
 &lt;td&gt;More than $1 billion a year of run-rate cost synergies, no date&lt;/td&gt;
 &lt;td&gt;The size is met. The date is missing&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Closing&lt;/td&gt;
 &lt;td&gt;9 to 12 months on precedents&lt;/td&gt;
 &lt;td&gt;First half of 2028. The outside date is November 1, 2027, and it can extend to August 1, 2028 for regulatory approvals&lt;/td&gt;
 &lt;td&gt;Longer&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Nvidia and neutral supply&lt;/td&gt;
 &lt;td&gt;A plan for Nvidia&amp;rsquo;s stake and for neutral supply&lt;/td&gt;
 &lt;td&gt;Nvidia&amp;rsquo;s Lumentum preferred stock can convert into matching Marvell preferred. Nothing on purchase commitments. Both module businesses stay&lt;/td&gt;
 &lt;td&gt;Not met&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Lumentum results&lt;/td&gt;
 &lt;td&gt;Consensus (as of October 28): $1.253 billion for the September quarter, $1.480 billion for the December quarter&lt;/td&gt;
 &lt;td&gt;$1.263 billion actual. Guidance midpoint $1.600 billion&lt;/td&gt;
 &lt;td&gt;Guidance is strong&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;By our own test from Thursday, this deal looks less like Inphi, Marvell&amp;rsquo;s best deal. We listed four signals that would mark an Inphi-type deal: a pledge to run Lumentum as a neutral laser supplier, a plan for the module business, a premium of 20% or less, and $1 billion of synergies with a date. Only the synergy size arrived. &lt;strong&gt;The deal still carries the customer risk we flagged. What changed is the price of the stock.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;The merger agreement adds detail. Marvell owes Lumentum $5.8 billion if Marvell&amp;rsquo;s board changes its recommendation or takes a better offer. Lumentum owes $2.9 billion in the reverse case. If regulators block the deal, or the outside date passes while only regulatory approvals are missing, Marvell pays a $3.2 billion regulatory fee. That is about 2.8% of the $115 billion enterprise value. The public text does not name China. The list of required approvals beyond U.S. antitrust review sits in a disclosure letter that is not public. Marvell does not have to accept remedies that would be material to the combined company.&lt;/p&gt;
&lt;h2 id="the-stock-already-paid-more-than-the-premium"&gt;The stock already paid more than the premium
&lt;/h2&gt;&lt;p&gt;The chart at the top answers one question: has the market charged Marvell for the premium yet?&lt;/p&gt;
&lt;p&gt;The loss is Marvell&amp;rsquo;s October 29 return minus the SOX return, applied to Marvell&amp;rsquo;s market value on October 28. Marvell fell 13.2%, and the SOX fell 1.4%. Part of the gap may reflect Alphabet&amp;rsquo;s results that day. Marvell designs custom chips for Google. Broadcom, Google&amp;rsquo;s larger chip partner, fell only 1.9%, so most of the gap belongs to the deal. The premium is Lumentum&amp;rsquo;s 104.5 million fully diluted shares times the difference between the $1,132.40 deal value and the $913.95 unaffected price.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;The market took about $3.9 billion more than the premium.&lt;/strong&gt; That extra discount is the market&amp;rsquo;s charge for customer loss and for a fabless company that will now run laser fabs. The synergies got nothing. On our numbers, it takes about $1.07 billion a year of pre-tax synergy to earn back a $22.8 billion premium at 25 times after-tax earnings. Marvell&amp;rsquo;s target, more than $1 billion, is close to that line. If Marvell delivers it, its holders break even against the unaffected price. Today&amp;rsquo;s stock price assumes zero. Any credible proof of synergy is upside from here.&lt;/p&gt;
&lt;p&gt;Marvell&amp;rsquo;s fall also cut the price of Lumentum. At Marvell&amp;rsquo;s October 30 close of $196.40, a Lumentum share is worth $883.80 in the deal. That is 3.3% below Lumentum&amp;rsquo;s unaffected price. In earnings terms, Marvell&amp;rsquo;s guided non-GAAP EPS for its fiscal third quarter is $1.10. At four times that, Marvell trades at 44.6 times earnings. Lumentum&amp;rsquo;s guided EPS for the December quarter is $6.05. At four times that, the deal values Lumentum at 36.5 times. &lt;strong&gt;Marvell pays with a currency at 45 times earnings for earnings at about 37 times.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;That gap explains a split in the earnings math. Both columns below are our estimates.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Basis&lt;/th&gt;
 &lt;th style="text-align: right"&gt;No synergy&lt;/th&gt;
 &lt;th style="text-align: right"&gt;$0.5 billion synergy&lt;/th&gt;
 &lt;th style="text-align: right"&gt;$1 billion synergy&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Calendar 2027, consensus as of October 28&lt;/td&gt;
 &lt;td style="text-align: right"&gt;-4.3%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+0.1%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+4.5%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Calendar 2027, if Lumentum earns $32 a share (our assumption)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;-1.0%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+3.4%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+7.7%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Current guidance, annualized (Marvell $1.10, Lumentum $6.05 a quarter)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;about +6% to +7%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;n/a&lt;/td&gt;
 &lt;td style="text-align: right"&gt;n/a&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;On current run rates, the deal adds to Marvell&amp;rsquo;s EPS before any synergy. Consensus expects Marvell&amp;rsquo;s earnings to grow much faster than Lumentum&amp;rsquo;s into 2027, so the 2027 math shows dilution before synergies. That consensus may be stale on the Lumentum side. Lumentum&amp;rsquo;s first-quarter EPS ($4.57) and second-quarter guidance ($6.05) already total $10.62. Consensus for fiscal 2027, which ends in June, is $21.80. That asks for only $11.18 in the second half, less than two quarters at the December-quarter guidance ($12.10). If that number rises, the 2027 dilution shrinks toward -1%.&lt;/p&gt;
&lt;h2 id="for-the-next-year-marvell-is-an-optics-stock"&gt;For the next year, Marvell is an optics stock
&lt;/h2&gt;&lt;p&gt;This is the part of the deal the market has not fully absorbed. A fixed exchange ratio with no collar ties the two stocks. Lumentum now trades as 4.5 Marvell shares less a merger spread. Marvell holders, in turn, will own 67% of a company in which Lumentum supplies about 38% of current run-rate earnings. Marvell&amp;rsquo;s own optical signal processors also sell into the same transceiver build. The combined earnings base leans on optical demand much more than Marvell&amp;rsquo;s did alone.&lt;/p&gt;
&lt;p&gt;&lt;img alt="Daily price changes on October 29 and 30. On October 29 Marvell fell 13.2% while Lumentum rose 0.6% and the SOX fell 1.4%. On October 30 Marvell fell 10.1%, Lumentum 8.6%, Coherent 11.3%, Applied Optoelectronics 12.8%, Broadcom 5.5% and the SOX 7.2%." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/marvell-lumentum-terms-priced-ai-optics-cycle-2026-11-02/two-days-two-causes.en.png"&gt;
&lt;/p&gt;
&lt;p&gt;The two days after the leak show the shift. October 29 was the deal: Marvell fell 13.2% and the optics names barely moved. October 30 was the cycle. Alphabet&amp;rsquo;s quarterly cash burn of $9.1 billion, against about $3.5 billion expected, spread fear about hyperscaler spending down the supply chain (24/7 Wall St.). Marvell fell 10.1%. The optics group (Lumentum, Coherent, Applied Optoelectronics) fell 10.9% on average. Marvell traded like an optics stock that day. We expect that link to hold until the deal closes.&lt;/p&gt;
&lt;p&gt;Nothing about the deal itself resolves within a year. The S-4 registration, the shareholder votes and the regulatory reviews are process steps. The synergies can only be proven after closing in 2028. In the meantime, the numbers that move the stock will be hyperscaler spending guidance, optical orders and optical prices. That is a tug of war. On one side sits Alphabet&amp;rsquo;s cash burn. On the other sits Lumentum&amp;rsquo;s order book.&lt;/p&gt;
&lt;p&gt;Lumentum&amp;rsquo;s own price says the market is not betting on the deal. At 8:27 a.m., Lumentum traded at $791.87 before the open, 8.2% below the $856.62 that 4.5 Marvell shares were worth. Spread over the time to a 2028 close, that is about 6% a year. The spread is modest, and it says little about closing odds, because Lumentum&amp;rsquo;s standalone value and the deal value are close to each other.&lt;/p&gt;
&lt;h2 id="lumentum-is-accelerating-into-the-deal"&gt;Lumentum is accelerating into the deal
&lt;/h2&gt;&lt;p&gt;This chart asks whether Marvell is buying a business at its peak or one that is still gaining speed.&lt;/p&gt;
&lt;p&gt;&lt;img alt="Lumentum revenue by fiscal quarter: $534 million in the September 2025 quarter at an 18.7% non-GAAP operating margin, $1,006 million in the June 2026 quarter at 36.6%, $1,263 million in the September 2026 quarter at 42.3%, and a $1,600 million guidance midpoint for the December 2026 quarter at 45% to 46%." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/marvell-lumentum-terms-priced-ai-optics-cycle-2026-11-02/lumentum-acceleration.en.png"&gt;
&lt;/p&gt;
&lt;p&gt;Lumentum reported its fiscal first quarter on Monday, three days earlier than expected. Revenue was $1.263 billion, up 25.5% from the prior quarter and 136.5% from a year ago. The non-GAAP operating margin rose to 42.3%. For the December quarter, Lumentum guides to $1.575 to $1.625 billion of revenue and a 45% to 46% margin. Chief executive Michael Hurlston said the company expects to add about $350 million of revenue in one quarter. He noted that this is more than the whole company&amp;rsquo;s quarterly revenue two years ago.&lt;/p&gt;
&lt;p&gt;Two lines in the release matter more for Marvell. Lumentum secured its &lt;strong&gt;second major NPO/CPO win, backed by purchase commitments of more than $5 billion&lt;/strong&gt;. Demand visibility for optical circuit switches has also grown, and Lumentum expects them to become a multi-billion-dollar yearly business. Both point to the scale-up optics market that Marvell says is the reason for the deal.&lt;/p&gt;
&lt;p&gt;The fixed ratio makes this growth valuable to Marvell holders. Marvell issues a set number of shares no matter what Lumentum earns before closing. Each upgrade to Lumentum&amp;rsquo;s earnings before 2028 lowers the effective price.&lt;/p&gt;
&lt;h2 id="customer-neutrality-is-the-variable-that-flips-us"&gt;Customer neutrality is the variable that flips us
&lt;/h2&gt;&lt;p&gt;The risk sits in the source of the synergies. Marvell&amp;rsquo;s presentation describes them as run-rate synergies, &amp;ldquo;primarily COGS&amp;rdquo;, meaning cost of goods sold. We think much of that will come from internal sourcing: Marvell signal processors, drivers and amplifiers inside Lumentum modules, and Lumentum lasers inside Marvell optical engines. This is our inference, because the company has not published a breakdown.&lt;/p&gt;
&lt;p&gt;The same mechanism gives outside customers reasons to leave. Module makers that compete with Lumentum can move to Broadcom&amp;rsquo;s signal processors. Stifel estimates Marvell&amp;rsquo;s share of PAM4 signal processors at 60% to 65%, and that share is exposed. Laser customers that compete with Marvell can move to Coherent. Coherent traded 1.6% higher before Monday&amp;rsquo;s open, which fits that bet. The announcement kept both module businesses and said nothing about Nvidia&amp;rsquo;s purchase commitments. Eighteen months to closing is enough time for a customer to qualify a second source.&lt;/p&gt;
&lt;p&gt;We would turn negative on any of these signals:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;Nvidia declines to carry over its Lumentum purchase commitments, or moves volume to Coherent.&lt;/li&gt;
&lt;li&gt;A large module maker qualifies Broadcom&amp;rsquo;s signal processor for its next platform.&lt;/li&gt;
&lt;li&gt;A regulator, most likely China&amp;rsquo;s, asks for conditions that Marvell refuses. The agreement lets Marvell refuse material remedies, which would raise the chance of a break and the $3.2 billion fee.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;We would turn more positive if the S-4 shows a synergy breakdown with dates, or if Nvidia confirms its commitments.&lt;/p&gt;
&lt;h2 id="what-to-watch"&gt;What to watch
&lt;/h2&gt;&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;When&lt;/th&gt;
 &lt;th&gt;What to check&lt;/th&gt;
 &lt;th&gt;Our line&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;November 2 to 4&lt;/td&gt;
 &lt;td&gt;Marvell against the SOX over three trading days from the October 30 close&lt;/td&gt;
 &lt;td&gt;Within -5%: the market accepts the terms. Worse than -10%: rejection&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;November 14&lt;/td&gt;
 &lt;td&gt;Nvidia&amp;rsquo;s 13F filing: its Lumentum and Marvell holdings&lt;/td&gt;
 &lt;td&gt;A cut would be a customer-conflict signal&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Early December (date not confirmed)&lt;/td&gt;
 &lt;td&gt;Marvell&amp;rsquo;s fiscal third-quarter results and fourth-quarter guidance&lt;/td&gt;
 &lt;td&gt;Custom silicon pace and scale-up optics revenue&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;About 4 to 8 weeks after signing (precedent)&lt;/td&gt;
 &lt;td&gt;The S-4 registration statement&lt;/td&gt;
 &lt;td&gt;Synergy breakdown and timing, management projections for Lumentum&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;November 1, 2027, extendable to August 1, 2028&lt;/td&gt;
 &lt;td&gt;Outside date&lt;/td&gt;
 &lt;td&gt;The $3.2 billion regulatory fee if approvals fail&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;&lt;strong&gt;Over the next year, what will move Marvell more: proof that the $1 billion of synergies is real, or the next hyperscaler spending number? Tell us which one you are watching.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Research and information only; not investment advice. Estimates and scenarios are analytical assumptions, not company guidance, trade instructions or personalized portfolio decisions. Event returns use Nasdaq closing prices, price only. Pre-market quotes are not final prices. Figures are date-bound. AI tools assisted research and editing.&lt;/em&gt;&lt;/p&gt;</description></item><item><title>The 2027–28 AI Bottleneck Is ABF Substrates, Not HBM</title><link>https://koreainvestinsights.com/post/ai-bottleneck-abf-not-hbm-2026-11-02/</link><pubDate>Mon, 02 Nov 2026 14:00:00 +0900</pubDate><guid>https://koreainvestinsights.com/post/ai-bottleneck-abf-not-hbm-2026-11-02/</guid><description>&lt;p&gt;The consensus says: &amp;ldquo;If AI buildout overshoots, supply will flood the market, and the bottleneck is already easing at HBM.&amp;rdquo;&lt;/p&gt;
&lt;p&gt;Public numbers move the bottleneck. At TPCA on October 20, 2026, Unimicron&amp;rsquo;s chairman put 2028 capacity versus 2024 at CoWoS 6.8x, hyperscaler capex 6.3x, IC substrate 2.0x, and DRAM 1.8x (&lt;a class="link" href="https://www.trendforce.com/news/2026/10/20/news-unimicron-warns-of-ic-substrate-crunch-up-to-30-month-equipment-lead-times-20-material-gap-in-2027/" target="_blank" rel="noopener"
 &gt;TrendForce write-up, secondary, Delayed&lt;/a&gt;). In the same remarks, critical tool lead times stretch to 30 months, 2027 supply of key materials (low-CTE glass fiber, CCL, ABF) is short by about 20%, and a new substrate fab needs 2.5–3 years from groundbreaking to shipment, so the next capacity wave arrives no earlier than 2029.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Our view: The binding bottleneck for 2027–28 AI accelerator shipments is ABF and IC substrate, not HBM stacking. HBM, CoWoS-class packaging, and substrate ship as one set. The slowest part sets the set ceiling. A substrate ceiling does not erase AI demand. It delays shipments and lengthens the memory cycle into 2028–29. We are wrong if, by June 2027, Unimicron, Ibiden, and the three Taiwan substrate names all print two consecutive quarters of YoY revenue growth below +20%, and the Morgan Stanley (reported) ABF shortage assumption for 2027 is cut below 5%.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;On October 30, our &lt;a class="link" href="https://koreainvestinsights.com/post/the-substrate-shortage-just-got-longer/" target="_blank" rel="noopener"
 &gt;note on a longer substrate shortage&lt;/a&gt; separated Ibiden, SEMCO, and Daeduck around Daeduck&amp;rsquo;s Q3. Below keeps only &lt;strong&gt;how a set bottleneck lengthens the memory cycle&lt;/strong&gt; in the same supply chain.&lt;/p&gt;
&lt;h2 id="set-ceiling-three-parts-must-leave-together"&gt;Set ceiling: three parts must leave together
&lt;/h2&gt;&lt;p&gt;One AI GPU or ASIC roughly needs HBM, advanced packaging (CoWoS-class), and an ABF substrate. If one leg is short, finished units are capped there.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Axis (2024→2028)&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Multiple&lt;/th&gt;
 &lt;th&gt;Definition / as-of&lt;/th&gt;
 &lt;th&gt;Source&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;CoWoS capacity&lt;/td&gt;
 &lt;td style="text-align: right"&gt;6.8×&lt;/td&gt;
 &lt;td&gt;Unimicron chairman, 2026-10-20&lt;/td&gt;
 &lt;td&gt;TrendForce path, secondary·Delayed&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Hyperscaler capex&lt;/td&gt;
 &lt;td style="text-align: right"&gt;6.3×&lt;/td&gt;
 &lt;td&gt;Same remarks&lt;/td&gt;
 &lt;td&gt;Same&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;IC substrate capacity&lt;/td&gt;
 &lt;td style="text-align: right"&gt;2.0×&lt;/td&gt;
 &lt;td&gt;Same remarks&lt;/td&gt;
 &lt;td&gt;Same&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;DRAM capacity&lt;/td&gt;
 &lt;td style="text-align: right"&gt;1.8×&lt;/td&gt;
 &lt;td&gt;Same remarks&lt;/td&gt;
 &lt;td&gt;Same&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;The table answers one question: &lt;strong&gt;which axis throttles set buildout.&lt;/strong&gt; When CoWoS and capex grow more than 6x, substrate grows 2x and DRAM 1.8x. HBM can be tight and substrate can still be the slower multiplier. Sold-out 2027 HBM is evidence of demand and contracts. It is not proof that HBM is the set ceiling.&lt;/p&gt;
&lt;p&gt;One cross-path in numbers: if the largest AI substrate grows from 80mm on a side (2025) to 110mm (2028 roadmap), area scales by (110/80)² ≈ 1.89x (Ibiden roadmap assumption, fallback). If substrate &amp;ldquo;capacity&amp;rdquo; doubles on an &lt;strong&gt;area&lt;/strong&gt; basis, unit supply of the largest package is about 2.0 / 1.89 ≈ 1.06x. Whether disclosed capacity is area or unit count is Missing, so treat this as an upper-bound estimate only. The point is simple. As packages grow, the same area capacity holds fewer units.&lt;/p&gt;
&lt;h2 id="abf-balance-shortage-rates-rise-prices-follow"&gt;ABF balance: shortage rates rise, prices follow
&lt;/h2&gt;&lt;p&gt;Morgan Stanley estimates (2026-10-14, via Postation, &lt;strong&gt;secondary·unverified&lt;/strong&gt;) raised the ABF substrate shortage to 9.3% in 2027 (from 6.2%) and 13.7% in 2028 (from 9.4%). The same secondary path cites substrate prices up 25%+ in 2027 and an about +30% ABF film hike in 3Q26 for Ajinomoto (95%+ share) (&lt;a class="link" href="https://www.postation.jp/news/abf-substrate-shortage-ibiden-ajinomoto" target="_blank" rel="noopener"
 &gt;Postation&lt;/a&gt;, &lt;a class="link" href="https://www.vlsi.kr/en/abf-substrate-crunch-mb-en/" target="_blank" rel="noopener"
 &gt;VLSI.kr summary&lt;/a&gt;).&lt;/p&gt;
&lt;p&gt;Primary confirmation sits with suppliers.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Observation&lt;/th&gt;
 &lt;th&gt;Value&lt;/th&gt;
 &lt;th&gt;Period / definition&lt;/th&gt;
 &lt;th&gt;Source&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Ibiden electronics FY3/27 OP guidance&lt;/td&gt;
 &lt;td&gt;¥114bn (~2.5x prior year)&lt;/td&gt;
 &lt;td&gt;Re-raised 2026-10-29&lt;/td&gt;
 &lt;td&gt;Company IR / filings path (primary)&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Ibiden 1H OP&lt;/td&gt;
 &lt;td&gt;+96.7% YoY&lt;/td&gt;
 &lt;td&gt;FY3/27 first half&lt;/td&gt;
 &lt;td&gt;Company disclosure path&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Unimicron Sep revenue&lt;/td&gt;
 &lt;td&gt;+64.7% YoY&lt;/td&gt;
 &lt;td&gt;TWSE, 2026-09&lt;/td&gt;
 &lt;td&gt;TWSE OpenAPI (primary)&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Nanya PCB Sep revenue&lt;/td&gt;
 &lt;td&gt;+57.8% YoY&lt;/td&gt;
 &lt;td&gt;Same&lt;/td&gt;
 &lt;td&gt;TWSE (primary)&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Kinsus Sep revenue&lt;/td&gt;
 &lt;td&gt;+44.5% YoY&lt;/td&gt;
 &lt;td&gt;Same&lt;/td&gt;
 &lt;td&gt;TWSE (primary)&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;September Taiwan revenues are not ABF-only. Still, +40–65% YoY at substrate names is a physical signal that price, mix, and utilization moved together. Ibiden&amp;rsquo;s 2.5x guide is primary evidence that scarcity is showing up in earnings.&lt;/p&gt;
&lt;h2 id="hbm-is-locked-packaging-growth-decelerates-in-2028"&gt;HBM is locked; packaging growth decelerates in 2028
&lt;/h2&gt;&lt;p&gt;A substrate bottleneck does not cancel HBM demand. Samsung Electronics&amp;rsquo; Q3 2026 IR shows memory revenue from KRW120.8tn to KRW140.7tn (+16.5% QoQ). On the call, DRAM ASP was about +20% and bit shipments were low- to mid-single digits (&lt;a class="link" href="https://news.samsung.com/global/samsung-electronics-announces-third-quarter-2026-results" target="_blank" rel="noopener"
 &gt;Samsung IR&lt;/a&gt;, &lt;a class="link" href="https://images.samsung.com/is/content/samsung/assets/global/ir/docs/2026_3Q_conference_eng.pdf" target="_blank" rel="noopener"
 &gt;conference PDF&lt;/a&gt;, primary). A log decomposition puts roughly one-quarter on bits and about three-quarters on price and mix (fallback). HBM4 was about 60% of HBM revenue, HBM4 revenue more than 3x QoQ, long-term agreements stood at 10 signed and 10 in negotiation, and 2027 HBM targets were sold out.&lt;/p&gt;
&lt;p&gt;Packaging has a different slope. Goldman estimates via TrendForce (&lt;strong&gt;secondary·unverified&lt;/strong&gt;, as_of 2026-10-27) put CoWoS capacity at 1.275m wafers in 2026 → 2.73m in 2027 (+114%) → 3.48m in 2028 (+27%). 2027 ramps hard. 2028 growth slows to about one-quarter of that rate. Reports of a 10–20% CoWoS price review and a 3–6% wafer price increase from January 2027 sit in the same bundle (&lt;a class="link" href="https://www.trendforce.com/news/2026/10/27/news-tsmc-reportedly-eyes-20-cowos-price-hike-skips-equipment-price-renegotiations-amid-ai-demand-surge/" target="_blank" rel="noopener"
 &gt;TrendForce path&lt;/a&gt;).&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Year&lt;/th&gt;
 &lt;th&gt;What binds the set&lt;/th&gt;
 &lt;th&gt;Implication for memory&lt;/th&gt;
 &lt;th&gt;Observation that flips the rank&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;2027&lt;/td&gt;
 &lt;td&gt;HBM LTAs / sell-out + substrate material and tool lead times&lt;/td&gt;
 &lt;td&gt;Shipment cap; price/mix holds&lt;/td&gt;
 &lt;td&gt;LTA renegotiation failure and substrate revenue slowdown together&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;2028&lt;/td&gt;
 &lt;td&gt;Lagging substrate multiples + CoWoS +27%&lt;/td&gt;
 &lt;td&gt;Deferred demand lengthens the cycle vs bit supply that outruns it&lt;/td&gt;
 &lt;td&gt;2028 bit-supply growth far above CoWoS and substrate growth&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;2029&lt;/td&gt;
 &lt;td&gt;New substrate fabs ship (2.5–3 years after groundbreaking)&lt;/td&gt;
 &lt;td&gt;Possible easing of the set ceiling&lt;/td&gt;
 &lt;td&gt;Fab schedules slip by a year or more&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;One line from the table: &lt;strong&gt;2027 is a locked-volume year. Substrate stretches shipments later more than it &amp;ldquo;upgrades&amp;rdquo; scarcity to a stronger label.&lt;/strong&gt;&lt;/p&gt;
&lt;h2 id="collision-with-the-bubble-oversupply-story"&gt;Collision with the bubble (oversupply) story
&lt;/h2&gt;&lt;p&gt;The oversupply story says capex runs ahead of demand and product floods the market. The multiples above point the other way. When capex and CoWoS outgrow substrate and DRAM, the physical result is a queue or delayed delivery, not empty racks. Microsoft&amp;rsquo;s September-quarter 2026 capex was about $45.2bn versus an internal plan near $50bn. The CFO pointed to third-party (neocloud) GPU-as-a-service delays into 1H 2027 (earnings path / secondary arithmetic). Azure still grew 48%, and commercial RPO was $677bn (+73%). The miss is procurement and delivery, not vanished demand.&lt;/p&gt;
&lt;p&gt;Reports that NVIDIA Vera Rubin HBM4 bandwidth eased from 22 to 19.2 TB/s per GPU (about −13%). Product-page edits and coverage from 27–29 October 2026 are Delayed. That change can raise effective HBM4 supply. It is closer to one more supply-side reason to keep 2028 memory supply risk at &amp;ldquo;moderate&amp;rdquo; than a break of 2027 volume locks. Substrate need not be labeled &amp;ldquo;severe&amp;rdquo; to still push the cycle into 2028–29.&lt;/p&gt;
&lt;p&gt;Taiwan&amp;rsquo;s September export orders hit a monthly record at $114.21bn (+62.4% YoY). The ministry said ASIC servers are growing faster than GPU servers (&lt;a class="link" href="https://www.taipeitimes.com/News/biz/archives/2026/10/21/2003866318" target="_blank" rel="noopener"
 &gt;Taipei Times path, Delayed&lt;/a&gt;). ASIC acceleration widens HBM demand beyond one GPU vendor and still consumes the same substrate and packaging set.&lt;/p&gt;
&lt;h2 id="name-frame"&gt;Name frame
&lt;/h2&gt;&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Name&lt;/th&gt;
 &lt;th&gt;What this thesis says&lt;/th&gt;
 &lt;th&gt;What is public&lt;/th&gt;
 &lt;th&gt;Missing / Partial&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Samsung Electronics&lt;/td&gt;
 &lt;td&gt;A lid on shipment speed and high reliance of profit on price/mix&lt;/td&gt;
 &lt;td&gt;Q3'26 bits low, price high; 2027 HBM sold out and LTAs&lt;/td&gt;
 &lt;td&gt;2028 contract price floor&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Samsung Electro-Mechanics&lt;/td&gt;
 &lt;td&gt;Candidate beneficiary of FC-BGA / AI high-layer substrate tightness&lt;/td&gt;
 &lt;td&gt;Moves with the industry revenue and price cycle&lt;/td&gt;
 &lt;td&gt;AI high-layer share (Partial)&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Daeduck Electronics&lt;/td&gt;
 &lt;td&gt;A price and utilization cycle bet&lt;/td&gt;
 &lt;td&gt;Tied to industry tightness&lt;/td&gt;
 &lt;td&gt;Share of top-spec AI substrates (Missing)&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Frame only. How much of any name to hold is outside this piece.&lt;/p&gt;
&lt;h2 id="counter-evidence"&gt;Counter-evidence
&lt;/h2&gt;&lt;ol&gt;
&lt;li&gt;&lt;strong&gt;HBM is still tight.&lt;/strong&gt; Sold-out 2027 and LTAs explain why markets talk HBM first. Calling substrate binding means &amp;ldquo;slower in the set,&amp;rdquo; not &amp;ldquo;HBM is loose.&amp;rdquo;&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;CoWoS +114% in 2027 is a real counter to packaging scarcity.&lt;/strong&gt; 2028&amp;rsquo;s +27% and substrate&amp;rsquo;s 2.0x remain. If packaging eases and substrate does not, the set stays bound.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;If 2028 memory bit supply (greenfield, CXMT, and peers) outruns deferred demand,&lt;/strong&gt; the outcome is price pressure, not a longer cycle. Then &amp;ldquo;lengthens the cycle&amp;rdquo; is wrong.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;MS ABF shortage rates and the Ajinomoto +30% print are secondary and unverified.&lt;/strong&gt; We do not paste paid primary text. If those figures are cut, they meet the falsifiers above.&lt;/li&gt;
&lt;/ol&gt;
&lt;h2 id="what-to-watch-next"&gt;What to watch next
&lt;/h2&gt;&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Checkpoint&lt;/th&gt;
 &lt;th&gt;Signal that strengthens the view&lt;/th&gt;
 &lt;th&gt;Signal that weakens it&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;December TWSE substrate trio revenue&lt;/td&gt;
 &lt;td&gt;YoY still ≥ +40%&lt;/td&gt;
 &lt;td&gt;YoY slows below +20%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Second ABF film price reports&lt;/td&gt;
 &lt;td&gt;Further hikes or tighter allocation&lt;/td&gt;
 &lt;td&gt;Hike reversed; inventory builds&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Ibiden / Unimicron guidance&lt;/td&gt;
 &lt;td&gt;OP or capacity raised&lt;/td&gt;
 &lt;td&gt;Demand cancellations; utilization cut&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;2027 hyperscaler capex guides&lt;/td&gt;
 &lt;td&gt;Held or raised, with delivery delays named&lt;/td&gt;
 &lt;td&gt;Two or more names cut explicitly&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;2028 memory wafers and LTAs&lt;/td&gt;
 &lt;td&gt;LTAs extended; wafer schedules slip&lt;/td&gt;
 &lt;td&gt;LTAs fail and wafer guides rise together&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;h2 id="what-remains-in-one-line"&gt;What remains in one line
&lt;/h2&gt;&lt;p&gt;To judge whether AI buildout is &amp;ldquo;too much,&amp;rdquo; watch how fast the &lt;strong&gt;set&lt;/strong&gt; actually ships, not only total capex. In the 2024→2028 multiples, the slowest published axis is IC substrate. That ceiling is not evidence against demand. It is evidence that shipments slip later and the memory price-and-mix window can last longer.&lt;/p&gt;</description></item><item><title>Provisional Verdict on AI Demand: Supply Caps 2027, the Bill Lands in 2028, and the Market Sold the Wrong Year</title><link>https://koreainvestinsights.com/post/ai-demand-provisional-verdict-2027-supply-capped-2028-bill-2026-11-01/</link><pubDate>Sun, 01 Nov 2026 23:47:00 +0900</pubDate><guid>https://koreainvestinsights.com/post/ai-demand-provisional-verdict-2027-supply-capped-2028-bill-2026-11-01/</guid><description>&lt;img src="https://koreainvestinsights.com/" alt="Featured image of post Provisional Verdict on AI Demand: Supply Caps 2027, the Bill Lands in 2028, and the Market Sold the Wrong Year" /&gt;&lt;p&gt;The market&amp;rsquo;s story is short: &amp;ldquo;AI demand is rolling over. That is why SK hynix is down 51% from its high.&amp;rdquo;&lt;/p&gt;
&lt;p&gt;We take the other side. Over the past two weeks this series asked the same question from several angles: &lt;a class="link" href="https://koreainvestinsights.com/post/dram-shortage-part-inventory-buyers-stop-when-cpus-arrive-2026-10-30/" &gt;part of the DRAM shortage is buyer inventory&lt;/a&gt;, &lt;a class="link" href="https://koreainvestinsights.com/post/ai-bullwhip-grid-cloud-memory-2026-10-31/" &gt;the AI bullwhip map&lt;/a&gt;, &lt;a class="link" href="https://koreainvestinsights.com/post/ai-roi-beyond-token-price-2026-10-31/" &gt;AI ROI clears the bar through volume and other ledgers&lt;/a&gt;, &lt;a class="link" href="https://koreainvestinsights.com/post/ai-hardware-selloff-supplier-profit-shelf-life-2026-11-01/" &gt;the market sold the shelf life of supplier profits&lt;/a&gt;, and &lt;a class="link" href="https://koreainvestinsights.com/post/samsung-per-3x-prices-60-percent-2028-profit-drop-2026-11-01/" &gt;what Samsung&amp;rsquo;s 3x multiple prices into 2028&lt;/a&gt;. Today we tie those pieces into one call. It is provisional, and the conditions that would flip it are listed with dates.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Our view: In 2027, AI compute demand exceeds supply, and it stays that way all year. Supply, not demand, sets the 2027 token count. So &amp;ldquo;demand slowdown&amp;rdquo; is not the 2027 risk. The risk sits where the money flows. Token prices fall 45-55% a year while compute rents hold. That squeeze lands on labs that rent their compute, and the real variable is whether it travels through finance into 2028 orders. The bullwhip is real but small. It sits in buyers&amp;rsquo; commodity DRAM buffers, and its bill comes in 2028. Because the supply chain is behaving conservatively, a 2028 commodity DRAM correction is coming, but it will be shallower than what the market has priced. The market sold 2028 deep and has not bought any of 2027&amp;rsquo;s earnings.&lt;/strong&gt;&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;We put 80% odds on paying demand exceeding capacity for all of 2027. Flat growth, announced list-price hikes and sold-out capacity are the shape of a ceiling.&lt;/li&gt;
&lt;li&gt;The odds that 2027 memory profits break through the demand path are 10% or less. The odds that the finance path cuts 2027 shipments are 15% or less. The squeeze lands on 2028 orders.&lt;/li&gt;
&lt;li&gt;Non-repeating buyer-buffer demand is 8-13% of 2027 shipments. Moved into 2028 demand, it is about 10%, one scenario notch.&lt;/li&gt;
&lt;li&gt;We put 55% odds on a 2028-29 commodity DRAM correction. Because the big four memory makers add only 15-17% more wafers in 2028, we expect it to be shallow.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;The reference date is November 1, 2026. All probabilities are our subjective estimates. We move them only in 5-point steps, and only on a primary source or a threshold break.&lt;/p&gt;
&lt;h2 id="the-numbers-draw-a-ceiling-not-fatigue"&gt;The numbers draw a ceiling, not fatigue
&lt;/h2&gt;&lt;p&gt;In a market with tiring demand, quarterly growth keeps slowing and prices and bookings fall with it. Every current number points the other way.&lt;/p&gt;
&lt;p&gt;&lt;img alt="A ceiling, not fatigue. Google first-party API tokens per minute grew 60%, 37% and 40% QoQ in Q1, Q2 and Q3 2026, flat near +40%. Combined quarterly revenue of two frontier labs was $10.4bn, $18.2bn and $26.4bn, up 45% QoQ despite falling prices. Per Epoch, fixed-price demand grows about 10x a year against 3.4x for inference supply, a gap of about 3x. Google announced API list-price hikes, North American data center vacancy is 1.4%, H100 one-year contract prices are up 38% YoY, and most of Amazon’s 2027 capacity is already booked." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/ai-demand-provisional-verdict-2027-supply-capped-2028-bill-2026-11-01/ceiling.en.svg"&gt;
&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Indicator&lt;/th&gt;
 &lt;th&gt;Value&lt;/th&gt;
 &lt;th&gt;How to read it&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Google first-party API tokens per minute, QoQ&lt;/td&gt;
 &lt;td&gt;Q1 2026 +60%, Q2 +37%, Q3 +40%. Two quarters annualized: 3.7-4.0x&lt;/td&gt;
 &lt;td&gt;It stepped down once and then flattened. It is not bending every quarter&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Google API list prices&lt;/td&gt;
 &lt;td&gt;Hike announced (effective Dec 1 and Jan 1, 2027)&lt;/td&gt;
 &lt;td&gt;Demand fatigue would push prices down. They are going up&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Two frontier labs, combined quarterly revenue&lt;/td&gt;
 &lt;td&gt;Q1 2026 $10.4bn, Q2 $18.2bn, Q3 $26.4bn&lt;/td&gt;
 &lt;td&gt;+45% QoQ despite falling unit prices&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Amazon Bedrock tokens (YoY), Microsoft Foundry&lt;/td&gt;
 &lt;td&gt;17x, 5x&lt;/td&gt;
 &lt;td&gt;Wherever capacity frees up, it fills&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;North American DC vacancy, H100 one-year contract price&lt;/td&gt;
 &lt;td&gt;1.4%, +38% YoY&lt;/td&gt;
 &lt;td&gt;The scarce thing is energized GW&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Amazon 2027 capacity&lt;/td&gt;
 &lt;td&gt;Mostly booked; a sizable share of 2028 too&lt;/td&gt;
 &lt;td&gt;2027 is already sold&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Epoch: fixed-price demand vs inference supply&lt;/td&gt;
 &lt;td&gt;About 10x vs 3.4x a year&lt;/td&gt;
 &lt;td&gt;A ~3x gap, toward shortage&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Google&amp;rsquo;s token growth is flat near +40%. Over the same stretch Google announced list-price hikes and Amazon closed out 2027. That combination is the shape of a supply curve, not a demand curve. The Q3 token figure is in Alphabet&amp;rsquo;s earnings release (&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1652044/000165204426000080/googexhibit991q32026.htm" target="_blank" rel="noopener"
 &gt;Alphabet 8-K, Q3 2026&lt;/a&gt;).&lt;/p&gt;
&lt;p&gt;Anthropic&amp;rsquo;s second quarter showed the same thing. When capacity freed up, pent-up demand appeared at once. We used Anthropic&amp;rsquo;s Claude for research and editing, so we carry Anthropic&amp;rsquo;s numbers as reported and do not judge their impact.&lt;/p&gt;
&lt;h2 id="solve-2027-token-volume-with-the-supply-formula"&gt;Solve 2027 token volume with the supply formula
&lt;/h2&gt;&lt;p&gt;When demand is pinned against a ceiling, demand surveys cannot tell you 2027 volume. The supply formula can. Multiply supply GW growth of 40-50% a year by Nvidia Rubin&amp;rsquo;s realized efficiency.&lt;/p&gt;
&lt;p&gt;If Rubin delivers 5-7x Blackwell on real workloads, hardware alone gives 2.5-3.1x the 2026 level. Add 1.3x from software and the ceiling is 3.3-4.0x. Our previous base range of 2.5-3.0x sat at the bottom of that ceiling. We move the base to 2.8-3.3x. That range is GW +45% times 1.9-2.3x tokens per GW. Because volume is supply-bound, GW demand is also near the same +45%.&lt;/p&gt;
&lt;p&gt;&lt;img alt="The supply formula for 2027 token volume. Hardware alone: 2.5-3.1x. Ceiling with 1.3x software: 3.3-4.0x. Old base: 2.5-3.0x. New base: 2.8-3.3x. Scenario weights move from bear 30 / base 50 / bull 20 to bear 20 / base 55 / bull 25. This is a sensitivity example that excludes power, HBM, CPU and substrate bottlenecks." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/ai-demand-provisional-verdict-2027-supply-capped-2028-bill-2026-11-01/supply-math.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;Demand does not decide anything above 3.5x. Rubin does. That is why the bull case struggles to go above 20-25% no matter how strong demand is. A result below 2.5x has two routes. One is demand breaking. The other is smaller throughput from a Rubin delay, weak realized efficiency, or power and substrate bottlenecks. In that second case unit prices and rents hold, which is the opposite signal from weak demand. There is no evidence that demand has broken, so we cut the bear case from 30% to 20%. We track the supply route separately through measured Rubin efficiency and power.&lt;/p&gt;
&lt;p&gt;One condition withdraws this call. If Alphabet&amp;rsquo;s Q4 API tokens per minute grow less than 15% QoQ, the problem is demand, not price, and that number alone is enough. If token volume in corporate spend samples like Ramp shrinks, we review; three months of no growth plus another drop in the top 1% sends us back to the previous base. The first check is the Ramp data on November 13. The decisive check is Alphabet&amp;rsquo;s results in early February 2027.&lt;/p&gt;
&lt;h2 id="the-risk-lives-where-the-money-flows"&gt;The risk lives where the money flows
&lt;/h2&gt;&lt;p&gt;Demand can grow without money growing at the same rate. We call this the token-compute scissors. Over the same four weeks, the blended token price in corporate card data fell 26.8% and Nvidia B200 rental rates rose 10.6%.&lt;/p&gt;
&lt;p&gt;&lt;img alt="The token-compute scissors. Over the same four weeks, B200 rental rose 10.6% and the Ramp blended token price fell 26.8%. The space between the lines is where lab margins get pinched. Annualized, token prices fall 45-55% while compute rents hold." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/ai-demand-provisional-verdict-2027-supply-capped-2028-bill-2026-11-01/scissors.en.svg"&gt;
&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Variable&lt;/th&gt;
 &lt;th&gt;Value&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Ramp blended token price, 4 weeks&lt;/td&gt;
 &lt;td&gt;−26.8%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;B200 rental, same period&lt;/td&gt;
 &lt;td&gt;+10.6%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;2027 lab revenue (4Q27/4Q26, previous base of 2.5-3.0x volume)&lt;/td&gt;
 &lt;td&gt;+40-80%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;2027 GW demand (previous base)&lt;/td&gt;
 &lt;td&gt;+25-90%, center +53%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Lab revenue at the new 2.8-3.3x base and unit prices −40-45%&lt;/td&gt;
 &lt;td&gt;+54-98%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;GW demand at 2.8x Rubin efficiency and 4.2x volume&lt;/td&gt;
 &lt;td&gt;+50%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Lab revenue per GW under the same conditions&lt;/td&gt;
 &lt;td&gt;+12-40%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Compute scarcity protects rental rates. It does not protect token prices, because labs chase volume with cheap tiers. So the squeeze hits the margins of labs that rent compute before it hits hyperscalers that own it.&lt;/p&gt;
&lt;p&gt;If lab funding stalls, the backlog that leans on labs wobbles. Oracle&amp;rsquo;s prepayments, the OpenAI share of Microsoft&amp;rsquo;s and Oracle&amp;rsquo;s backlog, and the neoclouds sit there. Neocloud and developer problems reach hyperscaler balance sheets through guarantees: Alphabet&amp;rsquo;s $48 billion data center payment guarantees and Meta&amp;rsquo;s $41.5 billion residual value guarantees.&lt;/p&gt;
&lt;p&gt;&lt;img alt="Where the risk lives. Token prices falling 45-55% a year squeeze the margins of labs that rent compute. That pressure runs into Oracle prepayments, the OpenAI share of Microsoft and Oracle backlog, and neoclouds, and then into hyperscaler balance sheets through Alphabet’s $48bn data center guarantees and Meta’s $41.5bn residual value guarantees. Odds: neocloud partial funding block 35%, hyperscaler funding block 15%. 2027 shipments are locked by LTAs, prepayments and the power cap, so the odds the finance path cuts them are under 15%; the squeeze lands on 2028 orders. +100bp is 0.28% of hyperscaler operating cash flow; 2027 funding need is about $200bn." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/ai-demand-provisional-verdict-2027-supply-capped-2028-bill-2026-11-01/money-path.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;We put 15% odds on hyperscaler funding stalling and 35% on neoclouds being partly blocked. This path can arrive inside 2027. LTAs carry price bands, volume-cut and undelivered-volume termination clauses, and SK hynix&amp;rsquo;s contracts cover more than 50% of 2027-28 capacity, not all of it. Still, hyperscalers are the ones buying 2027 shipments, and prepayments and the power cap tie up volume. So this path mainly cuts 2028 orders rather than 2027 shipments. How firmly volume is locked follows the counterparty&amp;rsquo;s cash ranking: hyperscalers first, then labs, then neoclouds.&lt;/p&gt;
&lt;h3 id="more-tokens-do-not-mean-proportionally-more-memory-bits"&gt;More tokens do not mean proportionally more memory bits
&lt;/h3&gt;&lt;p&gt;Here is the paradox. Even if tokens grow faster than expected, GW and memory bits do not grow at the same rate. At 4.2x token volume, GW demand is +50%, the same as the center of the base case. What rises is lab revenue per GW.&lt;/p&gt;
&lt;p&gt;&lt;img alt="Token upside is not memory-bit upside. 4.2x token volume equals 1.5x GW demand times 2.8x Rubin efficiency. If tokens beat, lab revenue per GW rises 12-40%, lab margins recover and finance-path risk shrinks. If tokens miss, 2027 GW stays put because LTAs, prepayments and the power cap bind it, so 2027 memory profit is insensitive to the token count." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/ai-demand-provisional-verdict-2027-supply-capped-2028-bill-2026-11-01/token-gw.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;So token upside does not arrive as memory-bit upside. It arrives as recovering lab margins, and it shrinks the finance-path risk. In reverse, token downside cannot cut 2027 GW by much either, because most 2027 volume is tied up by long-term agreements (LTAs), prepayments and the power cap. 2027 memory profit is insensitive to the token count. That is the precise meaning of &amp;ldquo;2027 is locked.&amp;rdquo;&lt;/p&gt;
&lt;p&gt;So here is where we stand. The odds that 2027 memory profits break through the demand path are 10% or less. The odds that the finance path cuts 2027 shipments are 15% or less. In 2027, the squeeze will show up in lab margins and in Oracle and neocloud spreads, not in memory earnings.&lt;/p&gt;
&lt;p&gt;Rates are second-order. A 100bp rise costs hyperscalers 0.28% of operating cash flow. The issue is funding volume, not the rate level. The 2027 funding need is about $200 billion. As our &lt;a class="link" href="https://koreainvestinsights.com/post/ackman-ai-rates-fed-mistake-marginal-dollar-is-debt-2026-11-01/" &gt;same-day Ackman post&lt;/a&gt; showed, Big Tech&amp;rsquo;s combined Q3 free cash flow is already negative. The marginal AI dollar is debt, and the variable is whether that debt gets funded on time.&lt;/p&gt;
&lt;h2 id="the-bullwhip-is-small-and-it-sits-in-buyers-commodity-dram"&gt;The bullwhip is small, and it sits in buyers&amp;rsquo; commodity DRAM
&lt;/h2&gt;&lt;p&gt;The bullwhip has a different character at each stage. Paper orders are largest in grid connection requests and Chinese GPU orders. But they build no inventory, and forecasts already filter them out (&lt;a class="link" href="https://koreainvestinsights.com/post/ai-bullwhip-grid-cloud-memory-2026-10-31/" &gt;AI bullwhip map&lt;/a&gt;). Committed forward demand is largest in multi-year lab-to-cloud commitments (RPO). Its downside path is renegotiation, not cancellation.&lt;/p&gt;
&lt;p&gt;The bullwhip that builds inventory is in memory, specifically buyers&amp;rsquo; buffers of commodity server and PC DRAM. The cause is the server CPU shortage. Without CPUs, buyers cannot build servers, yet they bought DRAM first (&lt;a class="link" href="https://koreainvestinsights.com/post/dram-shortage-part-inventory-buyers-stop-when-cpus-arrive-2026-10-30/" &gt;part of the DRAM shortage is inventory&lt;/a&gt;).&lt;/p&gt;
&lt;p&gt;&lt;img alt="The bullwhip differs by stage. Paper orders (grid requests, China GPUs) are the most inflated but build no inventory and are already filtered, so the bill is near zero. Committed forward demand (lab-to-cloud RPO) is largest in dollars, its downside path is renegotiation, and its bill is 2028 orders. The inventory-building whip (commodity server and PC DRAM buyers) is a 3-5 week buffer; buyers stop when CPUs and substrates land, costing about 10% of 2028 demand and about 10 points of price. 8-13% of 2027 shipments do not repeat. The first to take an HBM-side whip is Nvidia’s $279bn of supply commitments, 96% in FY27-29." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/ai-demand-provisional-verdict-2027-supply-capped-2028-bill-2026-11-01/bullwhip.en.svg"&gt;
&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Item&lt;/th&gt;
 &lt;th&gt;Value&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Buyer DRAM buffer, excess over normal&lt;/td&gt;
 &lt;td&gt;1-5 weeks estimated (cumulative shipments minus installs 2-9 weeks, center 5; end-2025 opening stock unconfirmed)&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Non-repeating buffer demand in 2027 shipments&lt;/td&gt;
 &lt;td&gt;About 8-13%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Share of 2028 demand taken by a 5-week excess buffer&lt;/td&gt;
 &lt;td&gt;About 10%. One scenario notch, about 10 points of price&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Threshold for a deep correction&lt;/td&gt;
 &lt;td&gt;Excess buffer of 8-12 weeks at end-2027, 2028 orders down 15-23%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Trigger for unwinding&lt;/td&gt;
 &lt;td&gt;2H27 server CPU and substrate supply additions&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;First to take an HBM-side whip&lt;/td&gt;
 &lt;td&gt;Nvidia supply commitments of $279bn, 96% in FY27-29 (&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1045810/000104581026000075/nvda-20260726.htm" target="_blank" rel="noopener"
 &gt;NVIDIA 10-Q&lt;/a&gt;, Note 10)&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;It is small, and we know where it sits. The 2027 profit figure does not change. The bill comes in 2028, and its size is about the difference between a balanced 2028 and a mildly oversupplied one.&lt;/p&gt;
&lt;h2 id="a-conservative-supply-chain-keeps-the-2028-correction-shallow"&gt;A conservative supply chain keeps the 2028 correction shallow
&lt;/h2&gt;&lt;p&gt;The big four memory makers grow 2028 wafers by 15-17%. Samsung runs five-year rolling LTAs, SK hynix typically five years, and the three makers&amp;rsquo; $38 billion of prepayments stretch to 2030 (&lt;a class="link" href="https://koreainvestinsights.com/post/samsung-sk-hynix-lta-volume-locked-terms-flexible-2026-10-29/" &gt;LTAs lock in volume, not profit&lt;/a&gt;).&lt;/p&gt;
&lt;p&gt;The 38% drop in the equipment basket discounts lower equipment orders in 2027-28. If that is the main cause, 2028 supply is more disciplined than the market fears. HBM is only about 10% of DRAM revenue at all three makers (estimates: SK hynix 13-14%, Samsung 11-12%, Micron 7-10%), so they ration capacity like commodity-cycle companies.&lt;/p&gt;
&lt;p&gt;Power points the same way. The US can build at most 30-40GW of data centers a year. Transformers are about 15% short. Gas turbines ordered today deliver in 2029-31. So 2027 supply cannot grow beyond the build-out already set (GW +40-50%), and extra supply arrives only modestly in 2028.&lt;/p&gt;
&lt;p&gt;&lt;img alt="The supply chain is conservative. Memory LTAs roll for five years and $38bn of prepayments run to 2030. Big-four wafers grow only 15-17% in 2028. On power, the US build cap of 30-40GW a year and a 15% transformer shortfall run through 2028, and gas turbines deliver in 2029-31. Buyer buffers build through 1H27 and are drawn down after CPUs arrive in 2H27. Amazon has booked most of 2027 and a sizable share of 2028. Our call: 55% odds of a 2028-29 commodity DRAM correction, shallow, and 20% odds of an HBM glut." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/ai-demand-provisional-verdict-2027-supply-capped-2028-bill-2026-11-01/lock-timeline.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;Here is how we see 2028. A commodity DRAM correction is coming; we put the cumulative 2028-29 odds at 55%. If CPUs and substrates arrive in 2H27, buffer building stops and installs absorb the stock, it will not be a deep 2021-22-style correction, because the supply increment is small and volume is tied up by prepayments and LTAs.&lt;/p&gt;
&lt;p&gt;The shallow call is conditional. In 2021 the excess buffer was 3-7 weeks, which overlaps today&amp;rsquo;s 1-5 week estimate. If the excess buffer reaches 8-12 weeks by end-2027, 2028 orders fall 15-23%, a collapse-level cut. The threshold is buyer inventory days rising all through 2027 and not turning in 2H27.&lt;/p&gt;
&lt;p&gt;We put the odds of an HBM glut at 20%. Samsung, with its larger commodity share, is more exposed. The first signal is mobile DRAM contract prices.&lt;/p&gt;
&lt;h2 id="the-market-sold-2028-deep-and-bought-none-of-2027"&gt;The market sold 2028 deep and bought none of 2027
&lt;/h2&gt;&lt;p&gt;This is where the market is wrong. From September 21 to October 30, about 83% of foreign net selling in Taiwan was AI hardware. Financials, telecoms and steel were not sold. What got sold was the shelf life of supplier profits, not AI demand (&lt;a class="link" href="https://koreainvestinsights.com/post/ai-hardware-selloff-supplier-profit-shelf-life-2026-11-01/" &gt;our November 1 post&lt;/a&gt;).&lt;/p&gt;
&lt;p&gt;Prices make it sharper. Samsung trades at 3.1x 2027 expected earnings. Plug in a normal 8x multiple, and the price assumes 2028 earnings about 60% below 2027 (&lt;a class="link" href="https://koreainvestinsights.com/post/samsung-per-3x-prices-60-percent-2028-profit-drop-2026-11-01/" &gt;the Samsung 3x post&lt;/a&gt;). Almost none of 2026-27&amp;rsquo;s earnings are in the price; 2028 is priced deep. The 2028 we see is shallower.&lt;/p&gt;
&lt;p&gt;Reading the drops in Intel and Chinese AI stocks as a demand slowdown today is not supported by current demand data. The gate to the demand-slowdown scenario (S3) is a single one: hyperscalers slowing capex. On their Q3 calls, all four raised 2027 capex. Amazon called 2027 a &amp;ldquo;meaningful increase,&amp;rdquo; and Meta even turned to equity financing. Prices alone cannot rule out that the market is pricing a future demand slowdown starting with suppliers. The tell is the order in which equipment and memory relative strength turn.&lt;/p&gt;
&lt;p&gt;Our call rests on 2027 earnings that are not in the price. The fastest way that gap closes is through flows, not fundamentals. Deleveraging is at the start of stage 4 of 6, and forced selling by leveraged retail investors is still ahead. That means one more shakeout before the bottom is possible.&lt;/p&gt;
&lt;h2 id="the-stance-in-numbers-less-demand-risk-more-credit-risk"&gt;The stance in numbers: less demand risk, more credit risk
&lt;/h2&gt;&lt;p&gt;&lt;img alt="The stance in numbers. 2027 token-volume weights move from bear 30 / base 50 / bull 20 to bear 20 / base 55 / bull 25, and the base range rises from 2.5-3.0x to 2.8-3.3x. Memory-cycle scenarios move from S1 rebound 55 / S2 credit squeeze 15 / S3 demand slowdown 15 / S4 macro shock 15 to S1 55 / S2 20 / S3 10 / S4 15, shifting 5 points from S3 to S2. New calls: 80% that paying demand exceeds capacity all of 2027, 10% or less that demand breaks 2027 memory profits, 55% for a shallow 2028-29 commodity DRAM correction." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/ai-demand-provisional-verdict-2027-supply-capped-2028-bill-2026-11-01/weights.en.svg"&gt;
&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Call&lt;/th&gt;
 &lt;th&gt;Previous&lt;/th&gt;
 &lt;th&gt;This stance&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;2027 token volume&lt;/td&gt;
 &lt;td&gt;Bull 20 / base 50 / bear 30, base 2.5-3.0x&lt;/td&gt;
 &lt;td&gt;Bull 25 / base 55 / bear 20, base 2.8-3.3x&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Memory-cycle scenarios&lt;/td&gt;
 &lt;td&gt;S1 55 / S2 15 / S3 15 / S4 15&lt;/td&gt;
 &lt;td&gt;S1 55 / S2 20 / S3 10 / S4 15&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Paying demand exceeds capacity all of 2027&lt;/td&gt;
 &lt;td&gt;New&lt;/td&gt;
 &lt;td&gt;80%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;2027 memory profits break via demand&lt;/td&gt;
 &lt;td&gt;New&lt;/td&gt;
 &lt;td&gt;10% or less&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;2028-29 commodity DRAM correction&lt;/td&gt;
 &lt;td&gt;2028 45%, cumulative 55%&lt;/td&gt;
 &lt;td&gt;Cumulative 55%, shallow&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Character of the bullwhip&lt;/td&gt;
 &lt;td&gt;Temporary 35 / structural 50 / negligible 15&lt;/td&gt;
 &lt;td&gt;Unchanged, leaning structural&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;The scenario names follow our &lt;a class="link" href="https://koreainvestinsights.com/post/chip-selloff-four-scenarios-week-ahead-2026-10-24/" &gt;October 24 post&lt;/a&gt;: S1 rebound, S2 credit squeeze, S3 demand slowdown, S4 macro shock.&lt;/p&gt;
&lt;p&gt;Moving 5 points from S3 to S2 has one meaning: we move the risk&amp;rsquo;s address from demand to credit. After Amazon&amp;rsquo;s results we also weighed raising S2 to 35%, but we kept the move smaller. The company borrowing cash fastest is also the one with the longest booked backlog, and we give those two facts equal weight.&lt;/p&gt;
&lt;h2 id="answering-the-obvious-objections-first"&gt;Answering the obvious objections first
&lt;/h2&gt;&lt;p&gt;&lt;strong&gt;&amp;ldquo;People use tokens because they got cheap. That is not real demand.&amp;quot;&lt;/strong&gt; Cheap tokens did pull in demand. That is exactly why we put the risk in the money flow. But demand that rests only on price would give Google no reason to announce list-price hikes. In quarters when unit prices fell, volume still grew 40% and lab revenue grew 45%.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;&amp;ldquo;Intel and Chinese AI stocks fell. Demand is cooling.&amp;quot;&lt;/strong&gt; The gate that decides a demand slowdown is hyperscalers&amp;rsquo; 2027 capex. All four raised it in Q3. While that gate is shut, drops in peripheral stocks are explained by flows and discount rates.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;&amp;ldquo;If Rubin is 7x more efficient, there will be spare GW.&amp;quot;&lt;/strong&gt; Higher efficiency raises the ceiling on 2027 token volume. It does not leave GW idle. 2027 GW is already tied up by LTAs and the power cap, and paying demand fills it. If Rubin stays at 2-4x, volume comes in smaller and unit prices hold. Either way, 2027 memory profits barely move.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;&amp;ldquo;When the bullwhip snaps, it will be as deep as 2021-22.&amp;quot;&lt;/strong&gt; What differs is supply&amp;rsquo;s hands. 2028 wafer growth is 15-17%, and volume is tied up in prepayments and LTAs. A 5-week excess buffer takes about 10% of 2028 demand. That is one scenario notch, not a whole cycle. We also wrote down when this answer fails: an excess buffer of 8-12 weeks by end-2027 means a deep correction.&lt;/p&gt;
&lt;h2 id="what-would-flip-this-call"&gt;What would flip this call
&lt;/h2&gt;&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Area&lt;/th&gt;
 &lt;th&gt;Signal&lt;/th&gt;
 &lt;th&gt;Threshold&lt;/th&gt;
 &lt;th&gt;When&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Demand&lt;/td&gt;
 &lt;td&gt;Alphabet API tokens per minute, QoQ&lt;/td&gt;
 &lt;td&gt;Under +15% means a demand problem. Under +30% means review a lower base range. +40% or more for two quarters confirms a higher range&lt;/td&gt;
 &lt;td&gt;Early Feb 2027&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Demand&lt;/td&gt;
 &lt;td&gt;Token volume and spend in samples like Ramp&lt;/td&gt;
 &lt;td&gt;Volume +20% a month plus a spend rebound locks in the 25% bull case. Three months of no growth plus another drop in the top 1% restores the bear case&lt;/td&gt;
 &lt;td&gt;Nov 13&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Finance&lt;/td&gt;
 &lt;td&gt;High-yield spread, Oracle CDS, CoreWeave bonds&lt;/td&gt;
 &lt;td&gt;350bp+, 300bp+ and below 75 cents together mean a shift to S2&lt;/td&gt;
 &lt;td&gt;Early-November issuance window; CoreWeave Nov 11&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Finance&lt;/td&gt;
 &lt;td&gt;Two labs&amp;rsquo; combined quarterly revenue, QoQ&lt;/td&gt;
 &lt;td&gt;Below +9% for two straight quarters means the 2030 payback path breaks&lt;/td&gt;
 &lt;td&gt;Every quarter in 2027&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Supply&lt;/td&gt;
 &lt;td&gt;Rubin customer-measured tokens per MW&lt;/td&gt;
 &lt;td&gt;5x or more raises the volume ceiling. Stuck at 2-4x means smaller 2027 volume and firm unit prices&lt;/td&gt;
 &lt;td&gt;Nov 19 earnings is the starting point; independent checks in 1H27&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Bullwhip&lt;/td&gt;
 &lt;td&gt;Rising buyer DRAM inventory days, falling same-generation rents, slowing follow-on orders&lt;/td&gt;
 &lt;td&gt;All three together move 2028 correction depth from &amp;ldquo;shallow&amp;rdquo; to &amp;ldquo;moderate&amp;rdquo;&lt;/td&gt;
 &lt;td&gt;Quarterly&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;2028&lt;/td&gt;
 &lt;td&gt;Mobile DRAM contract prices, 2028 HBM specs&lt;/td&gt;
 &lt;td&gt;A turn to falling contract prices pulls the commodity correction forward&lt;/td&gt;
 &lt;td&gt;Late-October contract prices onward; Nov 19 specs&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;h2 id="limits-and-interests"&gt;Limits and interests
&lt;/h2&gt;&lt;p&gt;Google&amp;rsquo;s mid-2025 monthly tokens, Microsoft Foundry, OpenRouter charts and Anthropic&amp;rsquo;s listing-document figures are secondary reports. Token definitions differ by metric. The capacity formula is a sensitivity example that leaves out power, HBM, CPU and substrate bottlenecks. The equipment drawdown mixes China exposure and position liquidation, so we do not read all of it as supply discipline. Without the end-2025 opening inventory, we cannot yet tell whether the buyer buffer is one week or five.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Do you think demand or supply sets the 2027 ceiling? Which number will shake this call first, the November 13 Ramp data or Alphabet&amp;rsquo;s token growth in February 2027? Tell us why.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;For research and information only. Not investment advice. Probabilities and scenarios are analytical assumptions, not trading instructions or individual portfolio decisions. Figures carry reference dates, and some come from public reports and secondary tallies. We used Anthropic&amp;rsquo;s Claude for research and editing. Anthropic appears in the AI demand chain discussed here as a customer and a compute buyer.&lt;/p&gt;</description></item><item><title>Samsung at 3x Earnings Already Prices a 60% Profit Drop in 2028; the Realistic Floor Is ₩160,000-184,000</title><link>https://koreainvestinsights.com/post/samsung-per-3x-prices-60-percent-2028-profit-drop-2026-11-01/</link><pubDate>Sun, 01 Nov 2026 23:20:00 +0900</pubDate><guid>https://koreainvestinsights.com/post/samsung-per-3x-prices-60-percent-2028-profit-drop-2026-11-01/</guid><description>&lt;img src="https://koreainvestinsights.com/" alt="Featured image of post Samsung at 3x Earnings Already Prices a 60% Profit Drop in 2028; the Realistic Floor Is ₩160,000-184,000" /&gt;&lt;p&gt;Samsung Electronics closed at ₩220,500 on October 30. That is &lt;strong&gt;4.8x&lt;/strong&gt; 2026 expected EPS and &lt;strong&gt;3.1x&lt;/strong&gt; 2027 consensus EPS. The closing high was ₩362,500 on June 18.&lt;/p&gt;
&lt;p&gt;A 3x multiple invites the line &amp;ldquo;this has to be the bottom.&amp;rdquo; Run the same price backward and you get a different picture. Plug in a normal 8x multiple, and the market is pricing 2028 EPS &lt;strong&gt;about 60% below&lt;/strong&gt; the 2027 estimate. The market does not doubt AI memory profits. It prices 2026-27 as a two-year spike, after which more than half of those profits disappear.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Our view: The stock can fall further. Liquidation flows and a softer Q1 2027 contract price alone can take it to a realistic floor of ₩160,000-184,000 (−17% to −27% from here). The ₩100,000s (1x end-2026 book value) open only if the price floors in long-term agreements (LTAs) actually break and hyperscalers cut 2027 capex, and we see no path for that evidence before the second half of 2027. If Q1 2027 commodity DRAM contract prices keep rising and HBM4 prices step up, this floor call stops mattering; if contract prices turn flat or down, the ₩160,000s open.&lt;/strong&gt;&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;Today&amp;rsquo;s price embeds a roughly 60% earnings drop in 2028. The October 28 FnGuide 2027 PER of 3.2-3.4x gives 57-60%.&lt;/li&gt;
&lt;li&gt;In the past two cycles, the bottom formed at about 6x peak earnings and about 1x book, after the drop was confirmed. Today the stock is already below that, before any drop is visible.&lt;/li&gt;
&lt;li&gt;The market&amp;rsquo;s conviction has a basis on direction. It is weak on size and timing, because it sets LTA floors and annual HBM pricing to zero.&lt;/li&gt;
&lt;li&gt;The deciding variables are the direction of Q1 2027 commodity DRAM contract prices and what the late-January board meeting discloses on LTA coverage and HBM4 annual pricing.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id="run-3x-backward-and-you-get-a-60-profit-drop-in-2028"&gt;Run 3x backward and you get a 60% profit drop in 2028
&lt;/h2&gt;&lt;p&gt;Start with the earnings each price assumes. The 2026 consensus is from Naver Finance, checked November 1. The 2027 EPS of about ₩72,000 is backed out from FnGuide&amp;rsquo;s 2027 PER of 3.2x on October 28 (October 29 close: ₩230,500). The 2027 operating profit of ₩555 trillion is Kiwoom Securities&amp;rsquo; October 8 estimate.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Basis&lt;/th&gt;
 &lt;th style="text-align: right"&gt;EPS&lt;/th&gt;
 &lt;th style="text-align: right"&gt;PER&lt;/th&gt;
 &lt;th&gt;Source&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;2026 consensus&lt;/td&gt;
 &lt;td style="text-align: right"&gt;₩45,983&lt;/td&gt;
 &lt;td style="text-align: right"&gt;4.8x&lt;/td&gt;
 &lt;td&gt;Naver Finance compilation&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;2027 consensus (OP ₩555T)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;~₩72,000&lt;/td&gt;
 &lt;td style="text-align: right"&gt;3.1x&lt;/td&gt;
 &lt;td&gt;Backed out from FnGuide, Oct 28&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;2027 bull case (OP ₩700T)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;~₩85,000&lt;/td&gt;
 &lt;td style="text-align: right"&gt;2.6x&lt;/td&gt;
 &lt;td&gt;Author estimate&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Divide the price by a normal multiple, and you get the earnings the market is paying for.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Reading&lt;/th&gt;
 &lt;th&gt;Formula&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Result&lt;/th&gt;
 &lt;th&gt;Meaning&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Normal 8x&lt;/td&gt;
 &lt;td&gt;220,500 ÷ 8&lt;/td&gt;
 &lt;td style="text-align: right"&gt;2028 EPS ~₩27,600&lt;/td&gt;
 &lt;td&gt;~62% below 2027E&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Normal 10x&lt;/td&gt;
 &lt;td&gt;220,500 ÷ 10, × 6.74B shares ÷ 0.78 conversion&lt;/td&gt;
 &lt;td style="text-align: right"&gt;Normalized OP ~₩190T&lt;/td&gt;
 &lt;td&gt;~3.2x the 2018 peak of ₩58.9T&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;At 8x, the market&amp;rsquo;s implied 2028 operating profit is about ₩240-250 trillion (EPS ₩27,600-28,800 × 6.74 billion shares ÷ 0.78). If a 2019-style collapse hits from ₩555 trillion in 2027, 2028 operating profit lands at about ₩210-325 trillion. Today&amp;rsquo;s price already sits inside that range.&lt;/p&gt;
&lt;p&gt;The market is not saying &amp;ldquo;AI memory profits are fake.&amp;rdquo; It is saying those profits last two years and then more than half of them go away.&lt;/p&gt;
&lt;h2 id="past-bottoms-came-at-6x-after-the-drop-showed-today-it-is-48x-before-it-shows"&gt;Past bottoms came at 6x after the drop showed; today it is 4.8x before it shows
&lt;/h2&gt;&lt;p&gt;To see how hard the market is discounting, compare it with past cycles. The historical figures below are the author&amp;rsquo;s approximations. We did not recheck them against primary data for this piece.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Date&lt;/th&gt;
 &lt;th style="text-align: right"&gt;PER on peak earnings&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Price to book&lt;/th&gt;
 &lt;th&gt;Situation then&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Jan 2019 low (~₩36,850)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;~6.1x (2018 EPS ₩6,024)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;~1.0x&lt;/td&gt;
 &lt;td&gt;After the Q4 profit drop was confirmed&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Sep 2022 low (~₩51,800)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;~6.4x (2022 EPS ₩8,057)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;~1.0x&lt;/td&gt;
 &lt;td&gt;After the Q3 guidance cut&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Nov 2017 high (~₩57,220)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;~8x 2018E&lt;/td&gt;
 &lt;td style="text-align: right"&gt;&lt;/td&gt;
 &lt;td&gt;The market doubted the peak, and was right&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Jan 2021 high (~₩96,800)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;~17x 2021E&lt;/td&gt;
 &lt;td style="text-align: right"&gt;&lt;/td&gt;
 &lt;td&gt;The market was optimistic, and was wrong&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Oct 30, 2026&lt;/td&gt;
 &lt;td style="text-align: right"&gt;4.8x 2026E&lt;/td&gt;
 &lt;td style="text-align: right"&gt;2.56x&lt;/td&gt;
 &lt;td&gt;Profits still rising (Q3 ₩107T, Q4 consensus ₩116T)&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;The two past bottoms formed at about 6x peak earnings, after the drop was visible. Today the stock is at 4.8x before any drop is visible. The discount is deeper than in November 2017 (8x), when the market doubted the peak and was right.&lt;/p&gt;
&lt;h2 id="flows-alone-can-reach-160000-184000"&gt;Flows alone can reach ₩160,000-184,000
&lt;/h2&gt;&lt;p&gt;We built the downside ladder on PER and book value per share (BPS). End-2026 BPS is about ₩104,000: Q2 BPS of ₩86,052, plus Q3 retained earnings (EPS ₩13,504 − dividend ₩4,601), plus Q4 retained earnings (consensus OP ₩116.3T × 0.78 ÷ 6.74B shares ≈ ₩13,460, − dividend of about ₩4,600). End-2027 BPS adds 2027E EPS of ₩72,000 and subtracts ₩12,000-20,000 in dividends (2027 policy not yet announced), for about ₩156,000-164,000. These are rough figures that exclude translation gains and losses and the effect of buybacks.&lt;/p&gt;
&lt;p&gt;&lt;img alt="Samsung Electronics downside ladder from the October 30 close of ₩220,500. The October 15 low of ₩204,500-207,500 (−6% to −7%), 1.2x end-2027 book at ₩187,000-197,000 (−11% to −15%), and 4.0x 2026E earnings at ₩184,000 (−17%) are reachable on liquidation flows alone. 3.5x 2026E earnings, equal to 1.0x end-2027 book, at ₩156,000-164,000 (−26% to −29%) opens when Q1 2027 contract prices turn flat or down. 1.0x end-2026 book at ₩104,000 (−53%) requires LTA floors to fail and hyperscalers to cut capex." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/samsung-per-3x-prices-60-percent-2028-profit-drop-2026-11-01/ladder.en.svg"&gt;
&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Level&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Price&lt;/th&gt;
 &lt;th style="text-align: right"&gt;vs. current&lt;/th&gt;
 &lt;th&gt;What it takes&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Oct 15 low (close / intraday)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;₩207,500 / ₩204,500&lt;/td&gt;
 &lt;td style="text-align: right"&gt;−6% to −7%&lt;/td&gt;
 &lt;td&gt;Continued liquidation alone&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;1.2x end-2027 book&lt;/td&gt;
 &lt;td style="text-align: right"&gt;~₩187,000-197,000&lt;/td&gt;
 &lt;td style="text-align: right"&gt;−11% to −15%&lt;/td&gt;
 &lt;td&gt;Deeper liquidation, retail leverage capitulates&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;4.0x 2026E earnings&lt;/td&gt;
 &lt;td style="text-align: right"&gt;~₩184,000&lt;/td&gt;
 &lt;td style="text-align: right"&gt;−17%&lt;/td&gt;
 &lt;td&gt;Same&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;3.5x 2026E earnings ≈ 1.0x end-2027 book&lt;/td&gt;
 &lt;td style="text-align: right"&gt;~₩156,000-164,000&lt;/td&gt;
 &lt;td style="text-align: right"&gt;−26% to −29%&lt;/td&gt;
 &lt;td&gt;Q1 2027 contract prices visibly turn flat or down&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;1.0x end-2026 book (old-style bottom)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;~₩104,000&lt;/td&gt;
 &lt;td style="text-align: right"&gt;−53%&lt;/td&gt;
 &lt;td&gt;LTA floors fail and hyperscaler capex cuts are confirmed&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;We also priced different 2028 earnings drops, using 2027E EPS of ₩72,000.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;2028 earnings drop&lt;/th&gt;
 &lt;th style="text-align: right"&gt;2028 EPS&lt;/th&gt;
 &lt;th style="text-align: right"&gt;At 6x&lt;/th&gt;
 &lt;th style="text-align: right"&gt;At 8x&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;−30%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;₩50,400&lt;/td&gt;
 &lt;td style="text-align: right"&gt;₩302,000&lt;/td&gt;
 &lt;td style="text-align: right"&gt;₩403,000&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;−60% (what today&amp;rsquo;s price embeds)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;₩28,800&lt;/td&gt;
 &lt;td style="text-align: right"&gt;₩173,000&lt;/td&gt;
 &lt;td style="text-align: right"&gt;₩230,000&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;−75%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;₩18,000&lt;/td&gt;
 &lt;td style="text-align: right"&gt;₩108,000&lt;/td&gt;
 &lt;td style="text-align: right"&gt;₩144,000&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;The ₩160,000-184,000 zone is realistic because flows and sentiment alone can reach it. Going below it requires pricing a 2028 earnings drop of 75% or more. 2027 HBM is sold out at agreed prices, and Q4 commodity DRAM contract prices are still rising 10-15%. In that setting, we see no path for evidence of a 75% drop before the second half of 2027.&lt;/p&gt;
&lt;p&gt;Time also lifts the floor. Book value grows by about ₩9,000 per quarter. So the old-style bottom of 1x book rises every quarter.&lt;/p&gt;
&lt;h2 id="the-markets-conviction-is-grounded-on-direction-not-on-size-and-timing"&gt;The market&amp;rsquo;s conviction is grounded on direction, not on size and timing
&lt;/h2&gt;&lt;p&gt;Direction rests on the base rate. By the author&amp;rsquo;s record of past cycles, every memory upcycle of the last 30 years ended with operating profit down 50-85%. The stock turned about 9-15 months before the profit peak (the November 2017 high came before the Q3 2018 profit peak; the January 2021 high came before the first-half 2022 peak). This cycle&amp;rsquo;s closing high was June 18. Apply the same lead, and the market is calling the profit peak for mid-2027. That is the implicit claim in the price.&lt;/p&gt;
&lt;p&gt;These facts support it.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Fact&lt;/th&gt;
 &lt;th&gt;Status&lt;/th&gt;
 &lt;th&gt;Source&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Buyers are building DRAM inventory: 2-9 weeks above normal by end-Q4&lt;/td&gt;
 &lt;td&gt;Confirmed&lt;/td&gt;
 &lt;td&gt;TrendForce (cloud and PC makers), HPE 10-Q&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Consumer NAND and some DDR4 prices have turned down&lt;/td&gt;
 &lt;td&gt;Confirmed&lt;/td&gt;
 &lt;td&gt;Market price survey (Oct 16)&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Hyperscaler funding is shifting from operating cash to debt and outside capital (Alphabet free cash flow −$9.1B, Meta $97.5B raised, Amazon free cash flow negative)&lt;/td&gt;
 &lt;td&gt;Confirmed&lt;/td&gt;
 &lt;td&gt;Company 10-Qs and bond filings&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;FX: about −₩10T to Q3 operating profit, a similar headwind in Q4&lt;/td&gt;
 &lt;td&gt;Confirmed&lt;/td&gt;
 &lt;td&gt;Samsung Newsroom&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;86% of Samsung&amp;rsquo;s DRAM revenue is commodity, so it is more exposed to a contract-price turn than SK hynix&lt;/td&gt;
 &lt;td&gt;Confirmed&lt;/td&gt;
 &lt;td&gt;Author analysis (Oct 30)&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;CXMT 4F² prototype validated&lt;/td&gt;
 &lt;td&gt;Partly confirmed (single report)&lt;/td&gt;
 &lt;td&gt;November 1 report&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;The facts against it carry the same weight.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Fact&lt;/th&gt;
 &lt;th&gt;Status&lt;/th&gt;
 &lt;th&gt;Source&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;DDR5 16Gb spot flat at $61.50, not falling&lt;/td&gt;
 &lt;td&gt;Confirmed&lt;/td&gt;
 &lt;td&gt;Spot price, Oct 31&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Q4 commodity DRAM contract prices still rising 10-15%&lt;/td&gt;
 &lt;td&gt;Confirmed&lt;/td&gt;
 &lt;td&gt;&lt;a class="link" href="https://www.trendforce.com/presscenter/news/20260930-13258.html" target="_blank" rel="noopener"
 &gt;TrendForce, Sep 30&lt;/a&gt;&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;2027 estimates raised after Q3 results&lt;/td&gt;
 &lt;td&gt;Confirmed&lt;/td&gt;
 &lt;td&gt;Broker estimate compilation&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;2027 HBM sold out at agreed prices; 10 LTAs signed and 10 in talks; shortage through 2028. SK hynix has sold through 2029&lt;/td&gt;
 &lt;td&gt;Confirmed (company statements)&lt;/td&gt;
 &lt;td&gt;Samsung Q3 IR, SK hynix Newsroom&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;LTA price floors and annual HBM pricing, which did not exist in 2018 or 2022&lt;/td&gt;
 &lt;td&gt;Confirmed&lt;/td&gt;
 &lt;td&gt;July 30 earnings call transcript&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Here is where we stand. The 60% drop priced for 2028 is too much, because it sets LTA floors and the annual contract structure to zero. But no floor has ever held through a downturn. Solar polysilicon prepayment contracts ended in renegotiation and lawsuits. It is rational for the market not to pay for those floors before they are tested. LTAs have locked volume, but nobody has proven they lock a profit floor (&lt;a class="link" href="https://koreainvestinsights.com/post/samsung-sk-hynix-lta-volume-locked-terms-flexible-2026-10-29/" &gt;our LTA analysis&lt;/a&gt;).&lt;/p&gt;
&lt;p&gt;So the call is tied to a calendar.&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;What flips it up: Q1 2027 contract prices keep rising and HBM4 prices step up. That kills the &amp;ldquo;2026 peak&amp;rdquo; thesis, and the stock can return to 5x 2027E or more (₩360,000+).&lt;/li&gt;
&lt;li&gt;What confirms the downside: any one of flat or falling Q1 2027 contract prices, buyer inventory above 9 weeks, or hyperscaler 2027 capex cuts opens the ₩160,000-184,000 zone.&lt;/li&gt;
&lt;li&gt;Until then, declines belong more to flows than to earnings revisions. Chip liquidation is under way, but retail leverage has not capitulated. Hedge fund short leverage is at the 99th percentile of five years in Goldman Sachs prime brokerage data, and US retail bought $9.3 billion of the 3x chip ETF (SOXL) in two weeks.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id="contract-prices-decide-the-path-to-the-end-of-q1-2027"&gt;Contract prices decide the path to the end of Q1 2027
&lt;/h2&gt;&lt;p&gt;We split the next five months into paths with signals and thresholds. The base paths are A and B. C needs a contract-price turn to open, and D is the tail.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Path&lt;/th&gt;
 &lt;th&gt;Signals&lt;/th&gt;
 &lt;th&gt;Threshold&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;A. Flow-driven touch of ₩184,000-207,000, then recovery&lt;/td&gt;
 &lt;td&gt;Liquidation in progress, retail leverage not capitulated, the October 30 US chip selloff hits Korea on November 2&lt;/td&gt;
 &lt;td&gt;Break of the Oct 15 intraday low of ₩204,500&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;B. Holds above ₩200,000, rerates after the January board and Q1 contract prices&lt;/td&gt;
 &lt;td&gt;Q4 contract prices +10-15%, 2027 estimates raised, three days of foreign net buying before Oct 29&lt;/td&gt;
 &lt;td&gt;Q1 2027 contract prices +5% or more&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;C. Downturn confirmed, ₩160,000-184,000&lt;/td&gt;
 &lt;td&gt;Consumer NAND and DDR4 declines spread, buyer inventory above 9 weeks&lt;/td&gt;
 &lt;td&gt;Q1 2027 contract prices flat or down&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;D. LTA floors fail and capex is cut, below ₩144,000&lt;/td&gt;
 &lt;td&gt;Hyperscalers cut 2027 capex guidance&lt;/td&gt;
 &lt;td&gt;Reports of floor renegotiation&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;h2 id="the-calendar-that-decides-it"&gt;The calendar that decides it
&lt;/h2&gt;&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Event&lt;/th&gt;
 &lt;th&gt;When&lt;/th&gt;
 &lt;th&gt;Baseline&lt;/th&gt;
 &lt;th&gt;Upside call&lt;/th&gt;
 &lt;th&gt;Downside call&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Q4 preliminary results&lt;/td&gt;
 &lt;td&gt;Early Jan 2027&lt;/td&gt;
 &lt;td&gt;OP ₩116.3T (consensus, secondhand)&lt;/td&gt;
 &lt;td&gt;₩120T+ despite FX headwind&lt;/td&gt;
 &lt;td&gt;Below ₩110T&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Late-January board meeting&lt;/td&gt;
 &lt;td&gt;Late Jan 2027&lt;/td&gt;
 &lt;td&gt;₩30T+ cancelled out of ₩55-75T remaining&lt;/td&gt;
 &lt;td&gt;Comes with LTA coverage and HBM4 pricing disclosure&lt;/td&gt;
 &lt;td&gt;Cancellation below ₩30T, preferred share weight 12% or less&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Q1 2027 commodity DRAM contract prices&lt;/td&gt;
 &lt;td&gt;Late Dec 2026 to Jan 2027&lt;/td&gt;
 &lt;td&gt;+5-10% (author estimate)&lt;/td&gt;
 &lt;td&gt;Keep rising&lt;/td&gt;
 &lt;td&gt;Flat or down&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Buyer inventory&lt;/td&gt;
 &lt;td&gt;Quarterly 10-Qs, TrendForce&lt;/td&gt;
 &lt;td&gt;2-9 weeks above normal&lt;/td&gt;
 &lt;td&gt;Shrinking&lt;/td&gt;
 &lt;td&gt;Above 9 weeks&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Hyperscaler 2027 capex&lt;/td&gt;
 &lt;td&gt;Late Jan to Feb 2027 results&lt;/td&gt;
 &lt;td&gt;Still growing&lt;/td&gt;
 &lt;td&gt;Growth accelerates&lt;/td&gt;
 &lt;td&gt;Cuts&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;November 2 close&lt;/td&gt;
 &lt;td&gt;Nov 2, 2026&lt;/td&gt;
 &lt;td&gt;Reflects the US chip selloff&lt;/td&gt;
 &lt;td&gt;Holds above ₩207,000, foreign net buying&lt;/td&gt;
 &lt;td&gt;Breaks the Oct 15 low&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;h2 id="a-6-7-payout-and-9000-of-book-value-per-quarter-support-the-floor"&gt;A 6-7% payout and ₩9,000 of book value per quarter support the floor
&lt;/h2&gt;&lt;p&gt;The 2026 shareholder return plan of ₩90-110 trillion is ₩13,400-16,300 per share, a 6.1-7.4% yield at today&amp;rsquo;s price. The January board meeting sets the 2027 policy. We covered the board timing and buyback calendar in &lt;a class="link" href="https://koreainvestinsights.com/post/samsung-buyback-calendar-preferred-shares-2026-10-29/" &gt;our shareholder return analysis&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;As shown above, book value grows by about ₩9,000 per quarter. Even if the stock stays flat, price to book falls every quarter. The longer the decline lasts, the closer the 1x-book floor comes to the stock price.&lt;/p&gt;
&lt;h2 id="sk-hynix-and-micron-watch-the-same-contract-price"&gt;SK hynix and Micron watch the same contract price
&lt;/h2&gt;&lt;p&gt;SK hynix has a larger HBM share, so under the same drop assumption its PER is lower (3.4x 2027). Micron&amp;rsquo;s fiscal year differs, which makes a direct comparison hard. For all three, commodity DRAM contract prices drive 2026-27 profits. So the direction of Q1 2027 contract prices is the shared deciding variable. We covered fair multiples on normalized earnings in &lt;a class="link" href="https://koreainvestinsights.com/post/samsung-hynix-fair-per-normalized-earnings-2026-10-25/" &gt;Samsung at 5-7x, SK hynix at 6-8x&lt;/a&gt;, and buyer inventory in &lt;a class="link" href="https://koreainvestinsights.com/post/dram-shortage-part-inventory-buyers-stop-when-cpus-arrive-2026-10-30/" &gt;part of the DRAM shortage is buyer inventory&lt;/a&gt;.&lt;/p&gt;
&lt;h2 id="limits-of-this-math"&gt;Limits of this math
&lt;/h2&gt;&lt;ul&gt;
&lt;li&gt;The past-cycle bottom multiples (Jan 2019 ~6.1x and 1.0x book, Sep 2022 ~6.4x and 1.0x book, Nov 2017 8x 2018E, Jan 2021 17x 2021E) and the 9-15 month lead of the stock over the profit peak are the author&amp;rsquo;s approximations. We did not recheck them against primary data.&lt;/li&gt;
&lt;li&gt;The 2027 EPS of about ₩72,000 is backed out from FnGuide&amp;rsquo;s 2027 PER. Naver Finance does not publish 2027 estimates.&lt;/li&gt;
&lt;li&gt;The BPS estimates are rough. They simplify translation gains and losses, buyback dilution, and the 2027 dividend policy.&lt;/li&gt;
&lt;li&gt;Prices are October 30 closes. Korea trades the October 30 US chip selloff on November 2.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;&lt;strong&gt;Which way do you think Q1 2027 commodity DRAM contract prices will go? Pick one of still rising, flat, or falling, and tell us why.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;This is for research and information only, not investment advice. Estimates and scenarios are analytical assumptions, not trading instructions or individual portfolio decisions. Figures are dated data, and consensus and some reported figures are secondary compilations. We used Anthropic&amp;rsquo;s Claude for research and editing. Anthropic appears in the AI memory demand chain discussed here as a compute buyer.&lt;/p&gt;</description></item><item><title>S&amp;P 500 Down 3%, SK hynix Down 51%: The Market Sold the Shelf Life of Supplier Profits, Not AI Demand</title><link>https://koreainvestinsights.com/post/ai-hardware-selloff-supplier-profit-shelf-life-2026-11-01/</link><pubDate>Sun, 01 Nov 2026 21:57:00 +0900</pubDate><guid>https://koreainvestinsights.com/post/ai-hardware-selloff-supplier-profit-shelf-life-2026-11-01/</guid><description>&lt;img src="https://koreainvestinsights.com/" alt="Featured image of post S&amp;P 500 Down 3%, SK hynix Down 51%: The Market Sold the Shelf Life of Supplier Profits, Not AI Demand" /&gt;&lt;p&gt;At the October 30 close, the S&amp;amp;P 500 was &lt;strong&gt;3.1%&lt;/strong&gt; below its high. That is not even a textbook correction (-10%). On the same day, the KOSPI was &lt;strong&gt;37.1%&lt;/strong&gt; below its June high, Samsung Electronics &lt;strong&gt;37.9%&lt;/strong&gt;, and SK hynix &lt;strong&gt;51.4%&lt;/strong&gt;. Applied Materials (AMAT), the chip equipment maker, was down &lt;strong&gt;46.0%&lt;/strong&gt;.&lt;/p&gt;
&lt;p&gt;Earnings moved the other way. Samsung Electronics earned KRW 107.4 trillion of operating profit in the third quarter, and SK hynix ran an 80% operating margin. TrendForce expects conventional DRAM contract prices to rise another 10-15% in the fourth quarter (&lt;a class="link" href="https://www.trendforce.com/presscenter/news/20260930-13258.html" target="_blank" rel="noopener"
 &gt;TrendForce, September 30&lt;/a&gt;). Between October 1 and 23, 2027 EPS estimates rose 0.7% for Samsung and 2.0% for SK hynix.&lt;/p&gt;
&lt;p&gt;Earnings are rising, yet the stocks fell by bear-market size. So the question is simple. Is this a passing correction, or the entrance to a structural bear market that ends the AI investment cycle?&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Our view: this drop comes from a higher price of money (rates and funding costs) and from large investors unwinding positions. The market did not sell today&amp;rsquo;s AI demand. It sold the belief that the high prices memory and equipment makers receive now will still hold in 2028. US indexes are in a correction, and Korean memory has already paid for a cyclical bear market. Of the switches that would make this a structural bear market, none was on as of November 1. If two switches from different groups turn on together, this view is wrong.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;The reference date is November 1, 2026. Prices are October 30 closes (Yahoo Finance daily data, local currency). Foreign flows cover September 21 to October 30. Credit spreads are FRED values for October 29, and US short interest is the FINRA October 15 settlement.&lt;/p&gt;
&lt;h2 id="the-same-market-is-in-two-different-phases"&gt;The same &amp;ldquo;market&amp;rdquo; is in two different phases
&lt;/h2&gt;&lt;p&gt;&amp;ldquo;Is this a correction?&amp;rdquo; has a different answer depending on which market you mean. US large-cap indexes peaked on October 6 and are still above their 200-day moving averages. Chips and Korea peaked in late June and are already far past the bear-market line (-20%).&lt;/p&gt;
&lt;p&gt;&lt;img alt="Drawdown from the 2026 high to the October 30 close. The S&amp;P 500 at -3.1% and the Nasdaq at -7.8% are inside the -10% correction line. The SOX at -27.0%, KOSPI -37.1%, KOSDAQ -40.9%, Samsung Electronics -37.9%, SK hynix -51.4% and AMAT -46.0% are far past the -20% bear-market line." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/ai-hardware-selloff-supplier-profit-shelf-life-2026-11-01/drawdowns.en.svg"&gt;
&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Index or stock&lt;/th&gt;
 &lt;th&gt;2026 high (date)&lt;/th&gt;
 &lt;th style="text-align: right"&gt;From high&lt;/th&gt;
 &lt;th&gt;Low after high (date)&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Bounce from low&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Vs 200-day MA&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;S&amp;amp;P 500&lt;/td&gt;
 &lt;td&gt;7,818.93 (10/6)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;-3.1%&lt;/td&gt;
 &lt;td&gt;7,580.07 (10/30)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;0.0%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+3.7%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Nasdaq Composite&lt;/td&gt;
 &lt;td&gt;27,599.79 (10/6)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;-7.8%&lt;/td&gt;
 &lt;td&gt;25,437.00 (10/30)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;0.0%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+1.5%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Philadelphia Semiconductor (SOX)&lt;/td&gt;
 &lt;td&gt;14,634.72 (6/22)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;-27.0%&lt;/td&gt;
 &lt;td&gt;10,447.49 (7/29)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+2.3%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;-0.9%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;KOSPI&lt;/td&gt;
 &lt;td&gt;9,114.55 (6/22)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;-37.1%&lt;/td&gt;
 &lt;td&gt;5,469.47 (10/26)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+4.8%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;-11.7%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;KOSDAQ&lt;/td&gt;
 &lt;td&gt;1,226.18 (4/27)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;-40.9%&lt;/td&gt;
 &lt;td&gt;644.78 (7/30)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+12.3%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;-25.5%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Samsung Electronics&lt;/td&gt;
 &lt;td&gt;KRW 362,500 (6/18)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;-37.9%&lt;/td&gt;
 &lt;td&gt;KRW 207,000 (7/30)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+8.7%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;-5.2%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;SK hynix&lt;/td&gt;
 &lt;td&gt;KRW 2,919,000 (6/22)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;-51.4%&lt;/td&gt;
 &lt;td&gt;KRW 1,223,000 (10/20)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+15.9%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;-5.0%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;The KOSPI fell 16.2% in October alone. On October 30 it dropped 3.48%, and foreigners sold KRW 1.6 trillion that day (&lt;a class="link" href="https://en.sedaily.com/news/2026/10/30/kospi-plunges-35-percent-on-alphabet-shock-samsung-chip" target="_blank" rel="noopener"
 &gt;Seoul Economic Daily, October 30&lt;/a&gt;). Only 17.5% of Korean stocks trade above their 50-day average, and 15.6% above their 200-day average. By the numbers, Korea is in the middle of a bear market.&lt;/p&gt;
&lt;p&gt;So the question needs two parts. For the US, we ask whether the correction spreads into a bear market. For Korea, we ask whether the bear market that already came ends in one cycle, as in 2018-19, or drags on, as in 2022.&lt;/p&gt;
&lt;h2 id="the-layers-pushing-prices-now-are-the-price-of-money-and-the-unwind"&gt;The layers pushing prices now are the price of money and the unwind
&lt;/h2&gt;&lt;p&gt;Chip stocks are priced by four layers at once: earnings, long-run supply and demand, the price of money, and positioning. Which layer moves prices now separates a correction from a structural bear market.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Layer&lt;/th&gt;
 &lt;th&gt;What it asks&lt;/th&gt;
 &lt;th&gt;State now&lt;/th&gt;
 &lt;th&gt;How much is priced in&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;A. 2026-27 earnings&lt;/td&gt;
 &lt;td&gt;How much will memory earn this year and next?&lt;/td&gt;
 &lt;td&gt;Strong. Q4 contract prices seen higher, 2027 estimates up, 2027 HBM sold out&lt;/td&gt;
 &lt;td&gt;Barely&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;B. Supply and demand after 2028&lt;/td&gt;
 &lt;td&gt;Do high prices survive new capacity and China&amp;rsquo;s CXMT?&lt;/td&gt;
 &lt;td&gt;Uncertain. Big-4 wafers +15-17%, CXMT share 7.6% to 9.5%&lt;/td&gt;
 &lt;td&gt;Heavily&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;C. Price of money (rates, funding)&lt;/td&gt;
 &lt;td&gt;How expensive do rates and credit make investment?&lt;/td&gt;
 &lt;td&gt;Worsening. Fed hiking cycle, 10-year at 5%, hyperscaler cash burn&lt;/td&gt;
 &lt;td&gt;Partly&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;D. Positioning&lt;/td&gt;
 &lt;td&gt;Who is selling, and how much is left?&lt;/td&gt;
 &lt;td&gt;Institutional longs mostly unwound, shorts crowded in ETFs&lt;/td&gt;
 &lt;td&gt;Largely&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;The pressure on chips now comes from layers C and D. Company earnings are growing, but money got more expensive and large investors cut positions. A drop like this ends when rate pressure and the unwind stop. A structural bear market starts when the layer A earnings outlook turns, and that signal has not appeared.&lt;/p&gt;
&lt;p&gt;Stock prices, however, move before earnings. Strong layer A earnings cannot rule out a structural bear market. They only lower its odds. This point comes back in the historical comparison below.&lt;/p&gt;
&lt;p&gt;The test is one question: does the pressure in layer C spread into layer B? If higher capital costs make hyperscalers cut 2027-28 capex, the decline turns structural. As of November 1, Alphabet and Amazon kept their 2027 growth plans, and Microsoft trimmed its 2026 plan slightly (&lt;a class="link" href="https://www.cnbc.com/2026/10/29/microsoft-google-ai-capex-estimates.html" target="_blank" rel="noopener"
 &gt;CNBC, October 29&lt;/a&gt;).&lt;/p&gt;
&lt;h2 id="why-stocks-fall-while-earnings-rise-the-up-loop-and-the-down-loop"&gt;Why stocks fall while earnings rise: the up loop and the down loop
&lt;/h2&gt;&lt;p&gt;AI investment heated the economy, and the hot economy pushed rates up. When rates rise, investors turn cautious even if companies are fine. Korean macro commentator Oh Geon-young describes it as a shift from a time when money was abundant to a time when money is scarce. When money is scarce, investors pick and choose. The bottom of the profitability ladder (neoclouds, equipment, commodity memory) is sold first.&lt;/p&gt;
&lt;p&gt;&lt;img alt="Two loops. The up loop runs from AI capex to growth, rising stocks, wealth-effect spending, rising inflation and yields, and Fed hikes. The down loop runs from doubt on AI returns to falling AI stocks, slower spending, slower growth and lower yields; it worked on October 22 and 30. Brent at 113 dollars caps the fall in yields. If the drop comes from AI demand, bonds hedge; if from oil, energy hedges; in a full deleveraging, only the dollar hedges." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/ai-hardware-selloff-supplier-profit-shelf-life-2026-11-01/two-loops.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;The up loop ends at the Fed. The Fed hiked in September and held at 3.75-4.00% on October 28. The vote was 10 to 2, and the two dissenters (Cook and Hammack) wanted a hike (&lt;a class="link" href="https://www.cnbc.com/2026/10/28/fed-interest-rate-decision-october-2026.html" target="_blank" rel="noopener"
 &gt;CNBC, October 28&lt;/a&gt;). Futures price an 85.7% chance of a hike on December 9 (CME). The US 10-year yield touched 5.03% on the day of the meeting (&lt;a class="link" href="https://www.cnbc.com/2026/10/28/us-treasurys-bonds-yields-federal-reserve.html" target="_blank" rel="noopener"
 &gt;CNBC&lt;/a&gt;) and closed October 30 at 4.93%. The 30-year is at 5.31%.&lt;/p&gt;
&lt;p&gt;The pressure reaches the companies that fund AI. Three of the four big hyperscalers now have negative free cash flow. Amazon is at -$42.4 billion over the trailing twelve months (&lt;a class="link" href="https://www.cnbc.com/2026/10/29/amazon-amzn-q3-earnings-report-2026.html" target="_blank" rel="noopener"
 &gt;CNBC, October 29&lt;/a&gt;), Alphabet at -$9.1 billion for the quarter, and Meta is also negative for the quarter (reported figures with different period bases). Microsoft has signaled bond market funding. When capex moves from cash to debt, rates become a variable in the capex plan.&lt;/p&gt;
&lt;p&gt;The down loop runs the other way. When AI stocks fall, spending and growth slow, and yields fall. This loop actually worked on October 22 and 30: on the days the Nasdaq fell, Treasury prices rose. In the down loop, bonds hedge stocks.&lt;/p&gt;
&lt;p&gt;The problem is oil. Brent for January delivery was $113.43 on October 30, held up by the Iran war and the risk to the Strait of Hormuz. US PCE inflation is 3.6%, and core is 3.1%. While oil holds inflation up, the Fed cannot easily cut even if growth weakens. In June 2008, the Fed statement also cited oil-driven inflation pressure (&lt;a class="link" href="https://www.federalreserve.gov/newsevents/pressreleases/monetary20080625a.htm" target="_blank" rel="noopener"
 &gt;FOMC statement, June 25, 2008&lt;/a&gt;).&lt;/p&gt;
&lt;p&gt;The contrast with 2018 is clear. In Q4 2018, the S&amp;amp;P 500 fell 19.8% from September 20 to December 24. Oil (WTI) collapsed from $76 to $43, and on January 4, 2019, Powell could talk about &amp;ldquo;patience&amp;rdquo; and change course. That exit is narrow now. This is why we see a fatter tail in this episode than in an ordinary correction.&lt;/p&gt;
&lt;p&gt;Consumption has a slower loop too. Of the 3.4% US GDP growth in Q3, 1.61 percentage points came from inventories, and final domestic sales grew 1.7%. The saving rate is 3.8%, the lowest since October 2022 (BEA). Most stocks are held by the top of the distribution, so flat stock prices alone can weaken spending. This loop works more slowly than capex, so we track it separately through the late-November saving rate and real spending.&lt;/p&gt;
&lt;h2 id="the-market-sold-the-shelf-life-of-supplier-profits"&gt;The market sold the shelf life of supplier profits
&lt;/h2&gt;&lt;p&gt;If the price of money is the backdrop, flows show what was sold. We collected stock-level foreign trading in Taiwan and Korea from September 21 to October 30.&lt;/p&gt;
&lt;p&gt;&lt;img alt="Drawdown by position in the AI supply chain. The demand side is shallow: S&amp;P 500 -3.0%, Nvidia -14.5%, TSMC -8.3%. The supply side is deep: Micron -27.1%, Samsung Electronics -37.2%, SK hynix -51.4%, Tokyo Electron -39.6%, Lam Research -44.8%, AMAT -46.0%. The lower bars show the AI hardware share of foreign net selling: Taiwan 82.5%, Korea 72.3%." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/ai-hardware-selloff-supplier-profit-shelf-life-2026-11-01/supply-chain-slope.en.svg"&gt;
&lt;/p&gt;
&lt;h3 id="foreigners-sold-ai-hardware-in-taiwan-and-korea-and-little-else"&gt;Foreigners sold AI hardware in Taiwan and Korea, and little else
&lt;/h3&gt;&lt;p&gt;Taiwan Stock Exchange (TWSE) stock-level data show foreign net selling of TWD 955.9 billion over 25 trading days. TSMC was 41% of it, MediaTek 18%, and Delta Electronics 8%. The top 11 AI hardware names add up to 82.5%. Foreigners were net buyers of Chunghwa Telecom, the Fubon, Cathay and CTBC financial groups, and China Steel.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Stock&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Foreign net buying (TWD 100m)&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Share of net selling&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Domestic funds net buying (TWD 100m)&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;TSMC&lt;/td&gt;
 &lt;td style="text-align: right"&gt;-3,905&lt;/td&gt;
 &lt;td style="text-align: right"&gt;40.9%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+569&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;MediaTek&lt;/td&gt;
 &lt;td style="text-align: right"&gt;-1,744&lt;/td&gt;
 &lt;td style="text-align: right"&gt;18.2%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+162&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Delta Electronics&lt;/td&gt;
 &lt;td style="text-align: right"&gt;-738&lt;/td&gt;
 &lt;td style="text-align: right"&gt;7.7%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+105&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;UMC&lt;/td&gt;
 &lt;td style="text-align: right"&gt;-410&lt;/td&gt;
 &lt;td style="text-align: right"&gt;4.3%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;-295&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;AVC&lt;/td&gt;
 &lt;td style="text-align: right"&gt;-253&lt;/td&gt;
 &lt;td style="text-align: right"&gt;2.6%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+134&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Nanya&lt;/td&gt;
 &lt;td style="text-align: right"&gt;-240&lt;/td&gt;
 &lt;td style="text-align: right"&gt;2.5%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;-94&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Chunghwa Telecom, three financials, China Steel&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+120, +123, +51&lt;/td&gt;
 &lt;td style="text-align: right"&gt;Net buying&lt;/td&gt;
 &lt;td style="text-align: right"&gt;Mostly net buying&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Foreign total&lt;/td&gt;
 &lt;td style="text-align: right"&gt;-9,559&lt;/td&gt;
 &lt;td style="text-align: right"&gt;100%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;Domestic funds total +2,092&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;TSMC&amp;rsquo;s 41% is close to its weight in the Taiwan index, so that is close to proportional selling. MediaTek&amp;rsquo;s 18% is far above its single-digit index weight. Taiwan&amp;rsquo;s domestic funds bought TSMC worth TWD 56.9 billion over the same period.&lt;/p&gt;
&lt;p&gt;Korea looks the same. Naver&amp;rsquo;s stock-level investor data (a secondary compilation) show foreigners sold KRW 28.5 trillion of KOSPI stocks over the period. SK hynix accounted for KRW 16.5 trillion and Samsung Electronics for KRW 4.2 trillion, together 72.3%. That is far above the two stocks&amp;rsquo; combined market cap weight (around 40%).&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Stock&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Foreign net buying (KRW tn)&lt;/th&gt;
 &lt;th&gt;Foreign ownership 9/18 to 10/30&lt;/th&gt;
 &lt;th style="text-align: right"&gt;From high&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;SK hynix&lt;/td&gt;
 &lt;td style="text-align: right"&gt;-16.47&lt;/td&gt;
 &lt;td&gt;50.23% to 48.67%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;-51.4%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Samsung Electronics&lt;/td&gt;
 &lt;td style="text-align: right"&gt;-4.15&lt;/td&gt;
 &lt;td&gt;46.46% to 46.15%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;-37.2%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Doosan Enerbility&lt;/td&gt;
 &lt;td style="text-align: right"&gt;-0.58&lt;/td&gt;
 &lt;td&gt;23.12% to 22.11%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;-41.8%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;KB Financial&lt;/td&gt;
 &lt;td style="text-align: right"&gt;-0.39&lt;/td&gt;
 &lt;td&gt;78.95% to 78.17%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;-16.5%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Hyundai Motor&lt;/td&gt;
 &lt;td style="text-align: right"&gt;-0.34&lt;/td&gt;
 &lt;td&gt;24.36% to 24.10%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;-51.5%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;KOSPI total&lt;/td&gt;
 &lt;td style="text-align: right"&gt;-28.52&lt;/td&gt;
 &lt;td&gt;&lt;/td&gt;
 &lt;td style="text-align: right"&gt;-37.1%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;If foreigners were leaving the country, they would have sold banks, telecom and steel too, and the won would have weakened. Instead, banks and telecom saw net buying, and USD/KRW fell 13.1% from 1,539 on June 22 to 1,338 on October 30. The won gained 15%. The foreign selling in both markets reads as a cut in one factor, AI hardware. We see no trace of a country exit.&lt;/p&gt;
&lt;h3 id="the-further-down-the-supply-chain-the-deeper-the-drop"&gt;The further down the supply chain, the deeper the drop
&lt;/h3&gt;&lt;p&gt;Line up drawdowns from the highest close in the June 22 to October 30 window by supply chain position, and a slope appears. Because of the window, Samsung&amp;rsquo;s figure differs slightly from the table above (June 18 high, -37.9%).&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Chain position&lt;/th&gt;
 &lt;th&gt;What sets profit&lt;/th&gt;
 &lt;th&gt;Stocks and drawdown&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Creates demand&lt;/td&gt;
 &lt;td&gt;Own service revenue&lt;/td&gt;
 &lt;td&gt;S&amp;amp;P 500 -3.0%, QQQ -9.1%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Aggregates demand&lt;/td&gt;
 &lt;td&gt;Total customer orders&lt;/td&gt;
 &lt;td&gt;Nvidia -14.5%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Long contracts, monopoly&lt;/td&gt;
 &lt;td&gt;Contracts and technology lock-in&lt;/td&gt;
 &lt;td&gt;TSMC -8.3% (US ADR -16.4%)&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Memory&lt;/td&gt;
 &lt;td&gt;Selling price&lt;/td&gt;
 &lt;td&gt;Micron -27.1%, Samsung -37.2%, SK hynix -51.4% (average of five, -31.6%)&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Equipment&lt;/td&gt;
 &lt;td&gt;Customer orders&lt;/td&gt;
 &lt;td&gt;AMAT -46.0%, Lam Research -44.8%, Disco -44.0%, Tokyo Electron -39.6% (average of six, -38.4%)&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;The order matches how cyclical each group&amp;rsquo;s profit is. An equipment maker loses sales when orders stop. A memory maker loses profit when prices turn. TSMC defends its prices with long contracts and a technology lead. Nvidia sits where demand is aggregated, and hyperscalers create the demand.&lt;/p&gt;
&lt;p&gt;Think of restaurants and farms. If customers were leaving a neighborhood, the restaurants that serve them would suffer first, and the farms and farm-equipment makers that supply them would suffer later. Today it is the opposite. The farm-equipment makers fell most, and the restaurants barely moved. What the market doubts is less the number of customers than whether the farm can keep charging today&amp;rsquo;s high prices in 2028.&lt;/p&gt;
&lt;p&gt;There are exceptions. SK hynix at -51% fell more than any equipment stock, a mix of Korea-specific and company-specific factors. TSMC&amp;rsquo;s local shares fell less than Nvidia. Peak timing splits in two as well. Most suppliers peaked in late June and fell for four months. The demand side, US indexes and TSMC peaked on October 6 and fell for one month. The core of this factor cut is that supplier stocks fell from late June to October 6 while demand-side stocks rose. The slope held through the broad market drop after October 6 (AMAT -26.4%, SK hynix -20.0%, Micron -15.4%, Nvidia -14.5%, TSMC -8.3%).&lt;/p&gt;
&lt;h3 id="in-korea-a-domestic-unwind-sits-on-top"&gt;In Korea, a domestic unwind sits on top
&lt;/h3&gt;&lt;p&gt;Korea needs one more explanation. Hyundai Motor (-51.5%), Hanwha Ocean (-57.1%), HD Hyundai Heavy Industries (-42.7%), Doosan Enerbility (-41.8%) and Kia (-40.5%) fell as much as memory. Yet foreign net selling in each was under KRW 0.6 trillion. Foreigners did not sell them.&lt;/p&gt;
&lt;p&gt;This part is a domestic unwind. Margin loans fell from KRW 38.6 trillion in June to KRW 32.75 trillion on October 15 (&lt;a class="link" href="https://en.sedaily.com/finance/2026/10/18/margin-debt-up-4-trillion-won-as-stock-market-cash-falls-7" target="_blank" rel="noopener"
 &gt;Seoul Economic Daily, October 18&lt;/a&gt;). Forced selling of margin positions reached KRW 89.3 billion on October 14 alone (&lt;a class="link" href="https://biz.heraldcorp.com/article/10904276" target="_blank" rel="noopener"
 &gt;Herald Business&lt;/a&gt;). Korea&amp;rsquo;s drop is the global AI hardware factor plus a domestic unwind of the whole 2026 leadership basket. We estimate that about half of Korea&amp;rsquo;s drawdown came from this market structure: concentrated foreign selling, substitution into overseas products, and forced margin selling. These are flows that can reverse, which supports the cyclical case.&lt;/p&gt;
&lt;h3 id="what-the-demand-breakdown-hypothesis-would-need-to-show"&gt;What the demand-breakdown hypothesis would need to show
&lt;/h3&gt;&lt;p&gt;The counterargument is clear. The market may have priced a future demand slowdown early and sold the suppliers with the most operating leverage first. Prices alone cannot rule this out. So we checked current demand indicators separately.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Indicator&lt;/th&gt;
 &lt;th&gt;If demand were breaking&lt;/th&gt;
 &lt;th&gt;Now&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Demand-side and aggregator stocks&lt;/td&gt;
 &lt;td&gt;Hyperscalers, Nvidia and TSMC fall with or before suppliers&lt;/td&gt;
 &lt;td&gt;S&amp;amp;P 500 -3%, Nvidia -15%, TSMC -8% vs equipment -45%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Capex&lt;/td&gt;
 &lt;td&gt;At least one hyperscaler slows&lt;/td&gt;
 &lt;td&gt;All four keep 2027 growth&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;AI server shipments&lt;/td&gt;
 &lt;td&gt;Taiwan server ODM sales slow&lt;/td&gt;
 &lt;td&gt;Six Taiwan AI server makers: September sales +57% YoY&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Physical build-out&lt;/td&gt;
 &lt;td&gt;Data center construction falls&lt;/td&gt;
 &lt;td&gt;US August +73%, eight straight months of growth&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Currency&lt;/td&gt;
 &lt;td&gt;Supplier-country currency weakens (exports fall)&lt;/td&gt;
 &lt;td&gt;Won 15% stronger&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;None of these shows up now. The way to tell the two stories apart is sequence. If demand-side shipment, utilization and capex outlooks are cut first, the demand hypothesis is right. If demand outlooks hold while only supplier price and margin outlooks are cut, the supplier-profit hypothesis is right. Taiwan memory makers&amp;rsquo; October sales on November 10 and Nvidia&amp;rsquo;s guidance on November 17-18 will show the order.&lt;/p&gt;
&lt;h2 id="four-months-paid-for-a-full-cycles-drop"&gt;Four months paid for a full cycle&amp;rsquo;s drop
&lt;/h2&gt;&lt;p&gt;Korean memory and global equipment have already matched the full drawdowns of the last two downcycles.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Stock&lt;/th&gt;
 &lt;th&gt;2018 downcycle&lt;/th&gt;
 &lt;th&gt;2021-22 downcycle&lt;/th&gt;
 &lt;th&gt;This time (June high to 10/30, four months)&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;SK hynix&lt;/td&gt;
 &lt;td&gt;2018-04 to 2018-11, -35%&lt;/td&gt;
 &lt;td&gt;2021-01 to 2022-11, -47%&lt;/td&gt;
 &lt;td&gt;-51%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Samsung Electronics&lt;/td&gt;
 &lt;td&gt;2017-09 to 2018-11, -30%&lt;/td&gt;
 &lt;td&gt;2021-01 to 2022-08, -36%&lt;/td&gt;
 &lt;td&gt;-37%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Micron&lt;/td&gt;
 &lt;td&gt;2018-05 to 2018-10, -35%&lt;/td&gt;
 &lt;td&gt;2021-12 to 2022-09, -46%&lt;/td&gt;
 &lt;td&gt;-27%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;AMAT&lt;/td&gt;
 &lt;td&gt;2018-02 to 2018-10, -43%&lt;/td&gt;
 &lt;td&gt;2021-12 to 2022-09, -48%&lt;/td&gt;
 &lt;td&gt;-46%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Past cycles use monthly closes and this one uses daily closes, so the bases differ slightly. The table compares size only. It does not say the drop is fully priced. The same drawdown means different things depending on the starting valuation and the normal earnings outlook.&lt;/p&gt;
&lt;p&gt;That narrows the meaning. For prices to fall further, the earnings normalization in this downcycle must be deeper than in 2018 and 2022. A CXMT-driven oversupply in 2028 is that path. Or the normal earnings outlook must keep falling. As long as 2027 estimates hold, Samsung at 3.1 times 2027 earnings and SK hynix at 2.8 times are close to prices that have already paid for a downcycle. At a normal P/E of 8-12 times, Samsung&amp;rsquo;s current price equals a 57-72% fall in earnings (our conditional calculation). That is deeper than the 2019 downcycle&amp;rsquo;s operating profit decline (-53%).&lt;/p&gt;
&lt;h3 id="memory-stocks-moved-two-to-four-quarters-ahead-of-prices"&gt;Memory stocks moved two to four quarters ahead of prices
&lt;/h3&gt;&lt;p&gt;In the 2017-18 cycle, stock peaks (Samsung in November 2017, SK hynix and Micron in May 2018) came two to four quarters before the DRAM contract price peak (Q3 2018). DRAM revenue hit a record in Q3 2018 and began to fall in Q4 (&lt;a class="link" href="https://www.trendforce.com/presscenter/news/20181115-10071.html" target="_blank" rel="noopener"
 &gt;TrendForce, November 2018&lt;/a&gt;). The stock low (December 2018) came while contract prices were still falling. PC DRAM contract prices fell about 30% in Q1 2019 (secondary source). 2000 followed the same order: Micron peaked in August 2000, and prices collapsed in the second half of that year.&lt;/p&gt;
&lt;p&gt;Apply this order to today, and the June stock peak reads as pricing a 2027 first-half price peak two to four quarters early. Industry tallies show quarterly contract price increases slowing from about 80% in Q2 to about 40% in Q3 and about 20% in Q4 (forecast). Prices are still rising, but more slowly, and stocks sell the slowdown first.&lt;/p&gt;
&lt;p&gt;This sequence has an uncomfortable implication for investors. In 2018, the stock low also came two to three quarters before the price low. A plan to wait for fundamental confirmation and buy back ends up buying 30-50% above the low.&lt;/p&gt;
&lt;h2 id="who-sold-and-who-is-left"&gt;Who sold, and who is left
&lt;/h2&gt;&lt;p&gt;Sellers move prices directly. Split them into stages and the asymmetry ahead becomes visible.&lt;/p&gt;
&lt;p&gt;&lt;img alt="Three stages of positioning. First, institutional longs were unwound from September 24 to October 15, with hedge fund tech selling at a 10-year high. Second, shorts moved to ETFs: QQQ short interest rose 16%, SOXX 36% and EWY 57%. Third, the sellers left are CTAs and retail leverage under water (SOXL, 18.3 billion dollars). Foreigners in Korea sold 28.5 trillion won in cash stocks but net bought 3.2 trillion won of futures. Price-triggered selling leads for one to two weeks; crowded shorts amplify a rebound over one to three months." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/ai-hardware-selloff-supplier-profit-shelf-life-2026-11-01/positioning.en.svg"&gt;
&lt;/p&gt;
&lt;p&gt;First, large investors emptied their longs between late September and mid-October. Prime broker data showed hedge fund net selling of tech at its largest since 2016, and net leverage at a one-year low (secondary reports). Not much institutional length is left to sell.&lt;/p&gt;
&lt;p&gt;Second, shorts moved into index ETFs. In the FINRA October 15 settlement, QQQ short interest rose 16%, the SOXX semiconductor ETF 36%, and the EWY Korea ETF 57% (&lt;a class="link" href="https://www.finra.org/finra-data/browse-catalog/equity-short-interest" target="_blank" rel="noopener"
 &gt;FINRA short interest&lt;/a&gt;). Micron single-name short interest fell 13%. Shorts betting on company fundamentals left, and index-level hedges remained. If shorts cover, the buying concentrates in chips and Korea.&lt;/p&gt;
&lt;p&gt;Third, the sellers still in place are trend-following funds (CTAs) that sell with price, and retail leverage under water. SOXL, the 3x leveraged chip ETF, has about $18.3 billion of net assets. These sellers sell more as prices fall, but their size shrinks with price. Their selling is finite.&lt;/p&gt;
&lt;p&gt;In Korea, foreigners sold KRW 28.5 trillion of cash stocks but net bought KRW 3.2 trillion of domestic futures. They are cutting longs, not building shorts. They were net buyers for three days from October 27 to 29 and sold again on the 30th. The forced-sale share of margin positions fell from 6.6% to about 1% (October 26, secondary compilation).&lt;/p&gt;
&lt;p&gt;So the time axis splits. Over the next one to two weeks, price-triggered mechanical selling has the upper hand. Over one to three months, crowded shorts favor an outsized rebound. Short covering, though, does not come from falling prices alone. It needs news. The candidates are November 10 inflation and short interest data, CoreWeave earnings on November 11, and Nvidia earnings on November 17-18.&lt;/p&gt;
&lt;p&gt;Sentiment gauges show no panic. The VIX at 19.42 is below the three-month VIX at 20.78, the normal shape in which near-term fear is smaller than longer-term worry. But in BofA&amp;rsquo;s October fund manager survey, cash was 3.9% and the net overweight in tech was 37%, the largest since August 2020 (&lt;a class="link" href="https://finance.yahoo.com/markets/articles/investor-bullishness-surges-bofa-survey-141800790.html" target="_blank" rel="noopener"
 &gt;Yahoo Finance summary&lt;/a&gt;). The survey ran October 2-8, so it misses the mid-October drop. US institutions are still heavy in tech, and that remains a tail risk.&lt;/p&gt;
&lt;h2 id="what-history-says-us-indexes-mostly-held-chips-fell-further"&gt;What history says: US indexes mostly held, chips fell further
&lt;/h2&gt;&lt;p&gt;Since 1995, there have been 13 days on which the Philadelphia Semiconductor Index was more than 20% below its 52-week high while the S&amp;amp;P 500 was within 5% of its high (&lt;a class="link" href="https://www.thetrading.tools/insights/semiconductor-bear-market-sp500-near-record" target="_blank" rel="noopener"
 &gt;thetrading.tools, secondary source&lt;/a&gt;).&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Measure&lt;/th&gt;
 &lt;th&gt;Value&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Cases where the S&amp;amp;P 500 fell into a 20%+ bear market&lt;/td&gt;
 &lt;td&gt;2 of 13 (April 2000, December 2007)&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Cases where it did not&lt;/td&gt;
 &lt;td&gt;11 of 13&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Median SOX return after the signal&lt;/td&gt;
 &lt;td&gt;-3.4% at six months, +19.2% at twelve months&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Median further SOX drop within twelve months&lt;/td&gt;
 &lt;td&gt;-21.3%. In 8 of 13 cases, SOX fell another 20% or more&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;2026&lt;/td&gt;
 &lt;td&gt;July 17 (SOX -20.2%, S&amp;amp;P 500 -2.0%) and October 22 (SOX -26.7%, S&amp;amp;P 500 -2.0%). The source treats them as one episode&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;The first reading is the resilience of US indexes. In 11 of 13 cases, US large caps did not fall into a bear market. In the two that did (2000 and 2007), the earnings cycle was already turning: DRAM prices collapsed in the second half of 2000, and financial sector profits broke in 2007. Today, earnings estimates are rising.&lt;/p&gt;
&lt;p&gt;The second reading is the further drop in chips. After the signal, the semiconductor index itself fell another 20% or more in 8 of 13 cases. US index resilience does not make a chip-heavy account safe. The rule in this study (-20% from a 52-week high) differs slightly from today (-27% from the 2026 high), and it comes from a single secondary source.&lt;/p&gt;
&lt;p&gt;The precedents for a decline during tightening are Q4 2018 and 2022. In Q4 2018, the S&amp;amp;P 500 fell 19.8% and stopped at the bear-market threshold. In 2022, it fell 25.4%, from 4,796.56 to 3,577.03, and the Nasdaq fell 35%. Both ended when the Fed changed direction and credit markets absorbed the shock. Today the Fed is at the start of a hiking cycle, and oil blocks a turn. This is why we keep the 2022-type path open.&lt;/p&gt;
&lt;h2 id="the-switches-that-would-change-the-call"&gt;The switches that would change the call
&lt;/h2&gt;&lt;p&gt;To avoid changing the call on feelings, we set the conditions in advance. The switches fall into a demand group (T1, T2), a financial group (T3, T4) and currency (T5).&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Switch&lt;/th&gt;
 &lt;th&gt;Turns on when&lt;/th&gt;
 &lt;th&gt;Status on November 1&lt;/th&gt;
 &lt;th&gt;When to check&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;T1 Earnings&lt;/td&gt;
 &lt;td&gt;Samsung or SK hynix 2027 or 2028 consensus turns to annual cuts, Q1 2027 contract prices are forecast to fall, or order cancellations and delivery delays are confirmed at first hand&lt;/td&gt;
 &lt;td&gt;Off. Estimates still rising (October 23), Q4 prices seen higher&lt;/td&gt;
 &lt;td&gt;November 5, November 10, late December&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;T2 Capex&lt;/td&gt;
 &lt;td&gt;One hyperscaler cuts or defers 2027 capex, or Nvidia misses guidance or its $279 billion supply commitments shrink&lt;/td&gt;
 &lt;td&gt;Amber. Alphabet &amp;ldquo;significantly higher in 2027,&amp;rdquo; Amazon &amp;ldquo;meaningfully more,&amp;rdquo; Microsoft trimmed 2026 by $5 billion&lt;/td&gt;
 &lt;td&gt;November 17-18, late January&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;T3 Credit&lt;/td&gt;
 &lt;td&gt;High-yield spread at 350bp or more, pulled deals or a jump in funding costs, Oracle 5-year CDS at 300bp, CoreWeave 2032 bonds below 75 cents&lt;/td&gt;
 &lt;td&gt;Off. High yield 326bp, investment-grade client buy/sell ratio 1.22&lt;/td&gt;
 &lt;td&gt;November 7, November 11, daily&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;T4 Fed and oil&lt;/td&gt;
 &lt;td&gt;A December 9 hike, Brent holding $110 and the 10-year back above 5.3%, all together&lt;/td&gt;
 &lt;td&gt;Off. Hike 85.7% priced, Brent $113, 10-year 4.93%&lt;/td&gt;
 &lt;td&gt;November 6, November 10, December 9&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;T5 Currency&lt;/td&gt;
 &lt;td&gt;KOSPI down 5% or more over five sessions together with USD/KRW up 3% or more over five sessions (or above 1,400)&lt;/td&gt;
 &lt;td&gt;Off. In the 1,330s, won stronger&lt;/td&gt;
 &lt;td&gt;Daily&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;The rule is this. Two switches in the same group (T1 and T2, or T3 and T4) count the same shock twice. The call changes to &amp;ldquo;structural bear market&amp;rdquo; when two or more switches from different groups turn on together. If only one turns on, we watch more closely. If a single signal is large, such as an Nvidia guidance miss or high yield at 400bp, we reassess at once.&lt;/p&gt;
&lt;p&gt;As of November 1, no switch is on. There are amber signals. Microsoft trimmed its 2026 capex slightly. Intel&amp;rsquo;s explanation also deserves attention. Intel cut its Q4 guidance and named memory prices as a reason for weaker PC demand, and its stock fell 29% (&lt;a class="link" href="https://www.cnbc.com/2026/10/31/intel-not-a-buy-after-drop.html" target="_blank" rel="noopener"
 &gt;CNBC, October 31&lt;/a&gt;). It is the first case of higher memory prices raising customers&amp;rsquo; costs and eroding commodity demand. Because this is how memory strength can eat into its own demand, we record it as a leading signal for T1.&lt;/p&gt;
&lt;p&gt;Items we could not check stay as &amp;ldquo;not yet seen.&amp;rdquo; We do not count them as &amp;ldquo;off.&amp;rdquo; These include consensus revisions after the October 29 earnings calls, short interest for the October 30 settlement, and first-hand Oracle CDS quotes.&lt;/p&gt;
&lt;p&gt;We also mark the path on the credit side. The high-yield spread is low at 326bp, but both ends of the rating scale are widening. AA spreads are at 65bp, the 98th percentile of the last three years, and CCC spreads are at 1,325bp (&lt;a class="link" href="https://fred.stlouisfed.org/series/BAMLH0A0HYM2" target="_blank" rel="noopener"
 &gt;FRED&lt;/a&gt;). At the top, heavy hyperscaler issuance creates supply pressure; at the bottom, spreads price default. If pressure from the price of money spreads into demand, the entrance is likely the tail of neoclouds and the lowest ratings. Hyperscaler high-grade bonds would be the last to shake.&lt;/p&gt;
&lt;h2 id="the-strongest-objections-and-our-answers"&gt;The strongest objections, and our answers
&lt;/h2&gt;&lt;p&gt;&lt;strong&gt;Objection one: &amp;ldquo;Earnings are this strong, so the drop is just an opportunity.&amp;quot;&lt;/strong&gt; We also see the cyclical path as the base case. But stocks move before earnings. In 2000 and 2007, earnings were still good when the decline began. Strong earnings cannot close the structural path. We keep the credit-crunch and demand-slowdown paths open and check whether they actually open with the switches.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Objection two: &amp;ldquo;The market simply priced a future demand slowdown first.&amp;quot;&lt;/strong&gt; That is possible. Suppliers with the most operating leverage also fall first when a demand slowdown is priced early. Sequence settles it: are demand-side outlooks cut first, or only supplier price outlooks? Equipment versus memory relative strength is another clue. If equipment rebounds first with rising estimates, expansion continues, and the 2028 supply fear stays valid. If memory rebounds first, supply discipline is winning.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Objection three: &amp;ldquo;The US is not even in a correction, yet only Korea fell. Isn&amp;rsquo;t Korea simply oversold?&amp;quot;&lt;/strong&gt; Partly, yes. A large part of Korea&amp;rsquo;s drop came from its market structure: concentrated foreign selling and forced margin selling. That part is a flow that can reverse. If non-AI leaders recover before memory, the domestic factor is easing. The part left in memory will ease only when evidence of 2028 supply discipline appears.&lt;/p&gt;
&lt;h2 id="what-this-means-for-investors"&gt;What this means for investors
&lt;/h2&gt;&lt;p&gt;First, tie the call to the switches. Prices have already paid for a cycle-sized drop, and the selling left is finite, mechanical selling. But if two switches from different groups turn on, defense comes first, even if prices look cheap.&lt;/p&gt;
&lt;p&gt;Second, waiting for fundamental confirmation has a cost. If 2018&amp;rsquo;s order repeats, the stock low comes before the price low. Waiting for confirmation means buying late. Buying now, on the other hand, means sitting through one to two weeks of mechanical selling.&lt;/p&gt;
&lt;p&gt;Third, the resilience of US indexes does not protect a portfolio concentrated in chips and Korea. In the history table, the chip index fell another 20% or more in 8 of 13 cases. Now is the time to check whether you can carry the size of your concentration.&lt;/p&gt;
&lt;h2 id="what-to-watch-over-the-next-six-weeks"&gt;What to watch over the next six weeks
&lt;/h2&gt;&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Date (Korea time)&lt;/th&gt;
 &lt;th&gt;What to check&lt;/th&gt;
 &lt;th&gt;How to read it&lt;/th&gt;
 &lt;th&gt;Switch&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;November 2&lt;/td&gt;
 &lt;td&gt;Foreign cash and futures flows in Korea, Samsung and SK hynix relative strength&lt;/td&gt;
 &lt;td&gt;Selling both cash and futures speeds the drop; selling cash while buying futures means the position cut continues&lt;/td&gt;
 &lt;td&gt;T5&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;November 3-4&lt;/td&gt;
 &lt;td&gt;US midterms, Iran strikes or a deal&lt;/td&gt;
 &lt;td&gt;Strikes mean an oil shock (Brent above $110); a deal means an oil drop and a tech rebound&lt;/td&gt;
 &lt;td&gt;T4&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;November 5&lt;/td&gt;
 &lt;td&gt;Samsung and SK hynix 2027 consensus&lt;/td&gt;
 &lt;td&gt;A turn to annual cuts means T1&lt;/td&gt;
 &lt;td&gt;T1&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;November 6&lt;/td&gt;
 &lt;td&gt;US October payrolls&lt;/td&gt;
 &lt;td&gt;Under 50,000 jobs with a falling 10-year points to a demand slowdown&lt;/td&gt;
 &lt;td&gt;T4&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;November 7&lt;/td&gt;
 &lt;td&gt;Absorption of bond supply&lt;/td&gt;
 &lt;td&gt;Investment-grade client buy/sell below 1.0 and high yield at 350bp mean T3&lt;/td&gt;
 &lt;td&gt;T3&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;November 10&lt;/td&gt;
 &lt;td&gt;US CPI, FINRA short interest, Taiwan October sales&lt;/td&gt;
 &lt;td&gt;Falling ETF shorts mean covering has started; month-on-month declines at Nanya and Winbond mark the peak in price momentum&lt;/td&gt;
 &lt;td&gt;T1, T4&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;November 11&lt;/td&gt;
 &lt;td&gt;CoreWeave earnings&lt;/td&gt;
 &lt;td&gt;Funding, utilization, cash flow, 2032 bonds at 75 cents&lt;/td&gt;
 &lt;td&gt;T3&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;November 17-18&lt;/td&gt;
 &lt;td&gt;Nvidia earnings and guidance&lt;/td&gt;
 &lt;td&gt;A guidance miss means T2&lt;/td&gt;
 &lt;td&gt;T2&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;December 9&lt;/td&gt;
 &lt;td&gt;FOMC&lt;/td&gt;
 &lt;td&gt;Whether it hikes, and the statement wording&lt;/td&gt;
 &lt;td&gt;T4&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;The daily tells are equipment versus memory relative strength, Korean non-AI leaders versus memory, and whether foreign selling spreads into banks and telecom. If foreign selling spreads into banks and telecom while the won weakens, that is a country exit.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Which switch do you think flips the call first? Pick one of Taiwan memory sales on November 10, Nvidia guidance on November 17-18, or a 350bp high-yield spread, and tell us why.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;This is for research and information only, not investment advice. Estimates and scenarios are analytical assumptions, not trading instructions or individual portfolio decisions. Figures are dated data, and Naver investor data and some reported figures are secondary compilations. We used Anthropic&amp;rsquo;s Claude for research and editing. Anthropic appears in the AI demand chain discussed here as a customer and compute buyer.&lt;/p&gt;</description></item><item><title>AI ROI Clears the Bar Through Volume and Other Ledgers, Not Token Prices</title><link>https://koreainvestinsights.com/post/ai-roi-beyond-token-price-2026-10-31/</link><pubDate>Sat, 31 Oct 2026 23:00:00 +0900</pubDate><guid>https://koreainvestinsights.com/post/ai-roi-beyond-token-price-2026-10-31/</guid><description>&lt;p&gt;The consensus says: &amp;ldquo;Token prices keep falling, so AI capex will not be recovered through token revenue.&amp;rdquo;&lt;/p&gt;
&lt;p&gt;Public numbers undercut the premise. In the Ramp AI Index, which uses corporate card data, the token unit-price measure fell from 1.10 to 0.64 (−42%) between August 2 and October 11, 2026. Over the same period, volume rose 65% and corporate AI spend rose 1.5% (Ramp AI Index, raw CSV not verified). The price collapsed. Total spend did not. In the same quarter, Alphabet disclosed that its first-party model APIs process nearly 32 billion tokens per minute, up 40% QoQ and more than 4x YoY (&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1652044/000165204426000080/googexhibit991q32026.htm" target="_blank" rel="noopener"
 &gt;Alphabet 8-K Exhibit 99.1, Q3 2026&lt;/a&gt;). Meta sold no tokens, yet grew Q3 ad revenue 26%, with average price per ad up 14% (&lt;a class="link" href="https://investor.atmeta.com/investor-news/press-release-details/2026/Meta-Reports-Third-Quarter-2026-Results/default.aspx" target="_blank" rel="noopener"
 &gt;Meta Q3 results&lt;/a&gt;).&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Our view: The path past current AI ROI is not token price (P). It is volume (Q) that grows faster than unit prices fall, plus enabler revenue that never shows up as tokens. We are wrong if, by June 2027, corporate AI spend growth in the Ramp data fails to beat the token unit-price decline for two consecutive quarters (elasticity stuck at or below 1), and Meta&amp;rsquo;s price-per-ad growth drops below +5% YoY for two quarters.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;If this holds, the money lands first on the books of compute owners and incumbent platforms, not frontier labs. That is why &amp;ldquo;AI revenue ÷ capex&amp;rdquo; understates the payback, and also why it cannot block overstated claims: AI-driven ad revenue is not disclosed separately.&lt;/p&gt;
&lt;h2 id="the-pq-scissors-growth-rate-gaps-not-price-levels-decide-roi"&gt;The P×Q scissors: growth-rate gaps, not price levels, decide ROI
&lt;/h2&gt;&lt;p&gt;The condition is one line. Growth in token spend (P×Q) must beat growth in capex. A falling P is not the problem by itself.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Metric&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Value&lt;/th&gt;
 &lt;th&gt;Period / definition&lt;/th&gt;
 &lt;th&gt;Source&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Token unit-price measure&lt;/td&gt;
 &lt;td style="text-align: right"&gt;1.10 → 0.64 (−42%)&lt;/td&gt;
 &lt;td&gt;2026-08-02 → 10-11, corporate spend&lt;/td&gt;
 &lt;td&gt;Ramp AI Index (not verified)&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Token volume&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+65%&lt;/td&gt;
 &lt;td&gt;Same period&lt;/td&gt;
 &lt;td&gt;Ramp AI Index&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Corporate AI spend&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+1.5%&lt;/td&gt;
 &lt;td&gt;Same period&lt;/td&gt;
 &lt;td&gt;Ramp AI Index&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Alphabet first-party API tokens&lt;/td&gt;
 &lt;td style="text-align: right"&gt;~32B per minute&lt;/td&gt;
 &lt;td&gt;Q3 2026, +40% QoQ, &amp;gt;4x YoY&lt;/td&gt;
 &lt;td&gt;Alphabet 8-K (primary)&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Top-5 hyperscaler ROIC&lt;/td&gt;
 &lt;td style="text-align: right"&gt;27.8% → 20.1%&lt;/td&gt;
 &lt;td&gt;2024-12 → 2026-03, chart read ±0.5pp&lt;/td&gt;
 &lt;td&gt;Apollo chart (secondary)&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Simple multiplication gives 0.58 × 1.65 ≈ 0.96, so spend should have fallen about 4%. Actual spend rose 1.5%. The gap may be a mix shift toward pricier models, but public data does not confirm it. What is clear is that price elasticity is about 1: volume barely offsets the price decline. If that sticks, revenue stalls while capex grows, and ROIC falls. The ROIC decline above shows that denominator effect.&lt;/p&gt;
&lt;p&gt;So the volume paths matter most. Long-horizon work is the example. At a 99% per-step success rate, the chance of finishing a 100-step task is 0.99¹⁰⁰ ≈ 36.6%. At 99.5%, it is 60.6% (assumed formula). A 0.5pp gain per step makes completion 1.7x as likely. When more work can be delegated, tokens per task and the number of tasks rise together. This path lifts Q regardless of price declines.&lt;/p&gt;
&lt;h2 id="path-by-path-verdict-which-routes-clear-the-roi-bar"&gt;Path-by-path verdict: which routes clear the ROI bar
&lt;/h2&gt;&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Path&lt;/th&gt;
 &lt;th&gt;Verdict&lt;/th&gt;
 &lt;th&gt;Public evidence&lt;/th&gt;
 &lt;th&gt;Observation that flips it&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;① P: premium model prices hold or rise&lt;/td&gt;
 &lt;td&gt;NO&lt;/td&gt;
 &lt;td&gt;No primary case of a lab list-price increase. Top-tier output price cut from $75 to $25 (2025-11)&lt;/td&gt;
 &lt;td&gt;Top-tier list price frozen 2 quarters + paid revenue accelerating&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;① P: rising costs normalize low-end token prices&lt;/td&gt;
 &lt;td&gt;NO (near term)&lt;/td&gt;
 &lt;td&gt;B200 rental ~+39% YTD while token index −54% (end-May → Sep, Silicon Data)&lt;/td&gt;
 &lt;td&gt;2+ low-end API list-price increases&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;① Q: spread to office and daily work&lt;/td&gt;
 &lt;td&gt;YES (volume)&lt;/td&gt;
 &lt;td&gt;Ramp paid AI adoption 55.9% (Oct). Microsoft Copilot 30M paid seats (reported, secondary)&lt;/td&gt;
 &lt;td&gt;Adoption flat for 2 quarters&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;① Q: token-maxing in office work&lt;/td&gt;
 &lt;td&gt;CONDITIONAL&lt;/td&gt;
 &lt;td&gt;No direct tokens-per-task data&lt;/td&gt;
 &lt;td&gt;Seat price frozen while tokens rise, squeezing only lab margins&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;① Q: better long-horizon capability&lt;/td&gt;
 &lt;td&gt;YES (dominant)&lt;/td&gt;
 &lt;td&gt;Alphabet 32B tokens/min, &amp;gt;4x YoY (primary). Agent work time 6x human (Anthropic report 10-16; report withholds causality)&lt;/td&gt;
 &lt;td&gt;Completion rates flat for 2 quarters&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;② Chinese lab ROI worsens, cheap supply shrinks&lt;/td&gt;
 &lt;td&gt;NO&lt;/td&gt;
 &lt;td&gt;Counter-evidence: Tencent, Alibaba, Baidu fund AI capex from operating cash, cash, and equity (company results). No Chinese lab P&amp;amp;L disclosed&lt;/td&gt;
 &lt;td&gt;Alibaba fails another raise, Baidu cash burn accelerates&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;② Open-weight catch-up slows&lt;/td&gt;
 &lt;td&gt;NO (12 months)&lt;/td&gt;
 &lt;td&gt;Open-vs-top gap projected 12 → 3 points (AA index, Kimi K3.1 Pro); lag ~4 months (Epoch)&lt;/td&gt;
 &lt;td&gt;Gap widens on long-task benchmarks; tighter export controls&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;③ Infra demand rises even with open weights&lt;/td&gt;
 &lt;td&gt;YES&lt;/td&gt;
 &lt;td&gt;Large open MoE weights need 1.1–1.4TB memory&lt;/td&gt;
 &lt;td&gt;Distillation and smaller models cut compute per task sharply&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;③ Open weights + services secure ROI&lt;/td&gt;
 &lt;td&gt;CONDITIONAL&lt;/td&gt;
 &lt;td&gt;No service-provider margin disclosed&lt;/td&gt;
 &lt;td&gt;Service-provider margin disclosure&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;④ Paid delegated agents (per task)&lt;/td&gt;
 &lt;td&gt;CONDITIONAL (2027+)&lt;/td&gt;
 &lt;td&gt;Coding agent run-rate $2.5B+ (Feb, secondary). No outcome-based billing revenue disclosed&lt;/td&gt;
 &lt;td&gt;Disclosed commercial outcome-billing revenue&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;⑤ Free delegated agents (ads, fees)&lt;/td&gt;
 &lt;td&gt;NO (minor before 2028)&lt;/td&gt;
 &lt;td&gt;Chatbot ads ~$1B annualized (Digiday, secondary). 4% in-chat checkout fee withdrawn&lt;/td&gt;
 &lt;td&gt;Disclosed agent-routed transaction fees&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;⑥ AI as enabler in existing services&lt;/td&gt;
 &lt;td&gt;YES (dominant)&lt;/td&gt;
 &lt;td&gt;Meta ads +26%, price per ad +14%, impressions +10% (primary)&lt;/td&gt;
 &lt;td&gt;Price per ad below +5% YoY for 2 quarters&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;The table&amp;rsquo;s conclusion is one line. The two axes that clear the bar are ⑥ enabler and ① long-horizon volume. Public evidence does not support the price (P) path the question hoped for.&lt;/p&gt;
&lt;h2 id="enabler-the-biggest-payback-never-shows-up-as-token-revenue"&gt;Enabler: the biggest payback never shows up as token revenue
&lt;/h2&gt;&lt;p&gt;Meta&amp;rsquo;s Q3 total revenue was $64.6B, up 26% YoY. Ad impressions rose 10% and average price per ad rose 14% (&lt;a class="link" href="https://investor.atmeta.com/investor-news/press-release-details/2026/Meta-Reports-Third-Quarter-2026-Results/default.aspx" target="_blank" rel="noopener"
 &gt;Meta IR, primary&lt;/a&gt;). Meta&amp;rsquo;s engineering blog said its GEM ad-ranking model lifted conversions 18.9% (October 13, company statement). Price rising faster than impressions means advertisers paid more for the same exposure. The filings do not attribute the full price increase to AI, and we do not upgrade correlation to causation.&lt;/p&gt;
&lt;p&gt;Counter-evidence sits in the same company. Meta&amp;rsquo;s Q3 free cash flow was −$2.12B, and net cash shrank to about $1.5B (Meta 8-K). Enabler revenue is strong, but capex is growing faster. That is why Meta&amp;rsquo;s price-per-ad growth is a falsification condition for this view.&lt;/p&gt;
&lt;h2 id="open-weights-they-hit-lab-pricing-not-infrastructure-volume"&gt;Open weights: they hit lab pricing, not infrastructure volume
&lt;/h2&gt;&lt;p&gt;Large open-weight MoE models need 1.1–1.4TB of memory just to load weights (Nemotron 3 Ultra BF16, Kimi K3 4-bit). Enterprises that self-host need the same GPUs and memory. What open weights change is who owns the profit: lab model margins shrink, compute volume stays with whoever owns the hardware.&lt;/p&gt;
&lt;p&gt;The checkpoint is November 4, when Moonshot AI plans to release Kimi K3.1 Pro weights. If the Artificial Analysis index then confirms a 3-point open-vs-top gap, the lab premium window shortens and path ① P weakens further. China&amp;rsquo;s draft revision of its technology export list (comments close November 15) excludes public open source, so it is not yet evidence that catch-up will slow.&lt;/p&gt;
&lt;h2 id="delegated-agents-paid-and-free-sit-on-different-ledgers"&gt;Delegated agents: paid and free sit on different ledgers
&lt;/h2&gt;&lt;p&gt;Paid delegation (per task) comes from corporate budgets. Some run-rates are public, such as coding agents, but &lt;strong&gt;no company has yet disclosed&lt;/strong&gt; outcome-based revenue, paid only when a task is completed. Liability and incident-compensation rules are also unsettled.&lt;/p&gt;
&lt;p&gt;Free delegation (everyday use) must cover costs with ads and transaction fees. &lt;strong&gt;No company discloses&lt;/strong&gt; agent-routed transaction fees as revenue either. Some platforms, including Amazon, block outside agents (reported). Free agents add users and inference volume, but we expect little ROI contribution before 2028.&lt;/p&gt;
&lt;h2 id="counter-evidence"&gt;Counter-evidence
&lt;/h2&gt;&lt;ol&gt;
&lt;li&gt;&lt;strong&gt;Elasticity near 1:&lt;/strong&gt; Taken at face value, Ramp shows volume barely offsetting price declines. If that persists, revenue stalls, capex grows, and ROIC keeps falling (27.8% → 20.1%).&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Cost of capital:&lt;/strong&gt; Reports say Anthropic&amp;rsquo;s credit rating is expected in BB territory after listing (FT, October 30, secondary). Hyperscaler IG spreads are 119bp versus 86bp for the index (FRED, October 29). Capital costs in the denominator could break ROI before demand does.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Narrowing open gap:&lt;/strong&gt; A 3-point gap trims the top of premium pricing and narrows the lab revenue path.&lt;/li&gt;
&lt;/ol&gt;
&lt;h2 id="what-to-check-next"&gt;What to check next
&lt;/h2&gt;&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;When&lt;/th&gt;
 &lt;th&gt;What&lt;/th&gt;
 &lt;th&gt;Strengthening signal&lt;/th&gt;
 &lt;th&gt;Weakening signal&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Nov 4&lt;/td&gt;
 &lt;td&gt;Kimi K3.1 Pro weight release&lt;/td&gt;
 &lt;td&gt;Re-measured gap stays ≥5 points&lt;/td&gt;
 &lt;td&gt;3-point gap confirmed&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Nov 15&lt;/td&gt;
 &lt;td&gt;China export-list comment deadline&lt;/td&gt;
 &lt;td&gt;Tighter limits on overseas local deployment&lt;/td&gt;
 &lt;td&gt;Draft unchanged (open source excluded)&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;November&lt;/td&gt;
 &lt;td&gt;Ramp AI Index, November edition&lt;/td&gt;
 &lt;td&gt;Spend growth &amp;gt; price decline&lt;/td&gt;
 &lt;td&gt;Spend flat, elasticity ≤1&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;TBD&lt;/td&gt;
 &lt;td&gt;Anthropic amended S-1&lt;/td&gt;
 &lt;td&gt;Improved run-rate and gross margin&lt;/td&gt;
 &lt;td&gt;Revenue slowing versus compute commitments&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Late Jan–early Feb 2027&lt;/td&gt;
 &lt;td&gt;Meta and Alphabet Q4 results&lt;/td&gt;
 &lt;td&gt;Price per ad holds ≥+5% YoY; tokens/min keeps rising QoQ&lt;/td&gt;
 &lt;td&gt;Price per ad &amp;lt;+5% YoY for two quarters; token QoQ growth slows&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;h2 id="the-supply-chain-side-lives-in-the-bullwhip-piece"&gt;The supply-chain side lives in the bullwhip piece
&lt;/h2&gt;&lt;p&gt;Inflation in power requests, cloud commitments, and memory inventory was covered in &lt;a class="link" href="https://koreainvestinsights.com/post/ai-bullwhip-grid-cloud-memory-2026-10-31/" &gt;The AI Bullwhip Map&lt;/a&gt;. On the final-payer side, two questions remain: who pays, and on whose books the money lands.&lt;/p&gt;
&lt;p&gt;&lt;small&gt;This is analysis based on public sources and not investment advice. Secondary-report figures are marked in the text.&lt;/small&gt;&lt;/p&gt;</description></item><item><title>The AI Bullwhip Map: Inflation Gets Filtered at the Grid, Losses Stay Where Cash Is Tied Up</title><link>https://koreainvestinsights.com/post/ai-bullwhip-grid-cloud-memory-2026-10-31/</link><pubDate>Sat, 31 Oct 2026 12:00:00 +0900</pubDate><guid>https://koreainvestinsights.com/post/ai-bullwhip-grid-cloud-memory-2026-10-31/</guid><description>&lt;p&gt;The consensus story goes like this: &amp;ldquo;The AI bubble is in grid connection requests. Requests are several times real demand, so that is where it pops.&amp;rdquo;&lt;/p&gt;
&lt;p&gt;Half of that is right. Power requests are the most inflated number in the chain. But that inflation is already measured in public filings, and it is being filtered out. On September 29, 2026, Virginia utility Dominion told PJM, the eastern U.S. grid operator, that it had removed 19,044 MW (58.6%) of the 32,473 MW of data center requests at the early engineering stage from its own load forecast. A number that gets filtered out does not become a loss.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Our view: rank AI supply chain loss risk by &amp;ldquo;cancellation cost × cash already tied up,&amp;rdquo; not by how inflated a number looks. On that test, buyer-held, already-paid component inventory ranks first, supplier-backed multi-year cloud commitments second, and grid requests last. The ranking flips if any of the conditions in the table below shows up by the first half of 2027.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;Memory inventory itself was covered yesterday in &lt;a class="link" href="https://koreainvestinsights.com/post/dram-shortage-part-inventory-buyers-stop-when-cpus-arrive-2026-10-30/" &gt;Part of the DRAM Shortage Is Inventory&lt;/a&gt;. Below are the power and cloud numbers, and the path by which delays in both change memory orders.&lt;/p&gt;
&lt;h2 id="why-ranking-by-inflation-gets-it-wrong"&gt;Why ranking by inflation gets it wrong
&lt;/h2&gt;&lt;p&gt;The bullwhip effect is order volatility that grows as you move upstream, because each stage reads the orders below it as demand and adds a safety margin (Lee, Padmanabhan, Whang, 1997). On top of the classic causes (duplicate requests, padded orders under allocation, buying ahead of price increases), AI adds two more. Servers are complements: GPU, CPU, memory, and power all have to be there. And suppliers backstop customer demand with their own credit.&lt;/p&gt;
&lt;p&gt;So the same &amp;ldquo;inflation&amp;rdquo; ends differently. Turning it into a loss takes three things at once:&lt;/p&gt;
&lt;ol&gt;
&lt;li&gt;Unwinding costs money (cancellation cost).&lt;/li&gt;
&lt;li&gt;Cash has already gone out or hit the financial statements (paid or recognized).&lt;/li&gt;
&lt;li&gt;The unwind lands when supply is growing (timing).&lt;/li&gt;
&lt;/ol&gt;
&lt;p&gt;Grid requests fail both 1 and 2. Below, we put numbers on all three, stage by stage.&lt;/p&gt;
&lt;h2 id="power-the-most-inflated-stage-and-the-one-being-measured-and-filtered"&gt;Power: the most inflated stage, and the one being measured and filtered
&lt;/h2&gt;&lt;p&gt;Dominion&amp;rsquo;s filing breaks down requests by contract stage (as of July 2026):&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Stage&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Capacity&lt;/th&gt;
 &lt;th&gt;Nature&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Electric service agreement (ESA)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;11,997 MW&lt;/td&gt;
 &lt;td&gt;Most firm&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Construction letter of authorization (CLOA)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;9,356 MW&lt;/td&gt;
 &lt;td&gt;Construction stage&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Engineering letter of authorization (ELOA)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;32,473 MW&lt;/td&gt;
 &lt;td&gt;Earliest stage&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Total&lt;/td&gt;
 &lt;td style="text-align: right"&gt;53,826 MW&lt;/td&gt;
 &lt;td&gt;&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;ELOA excluded from forecast&lt;/td&gt;
 &lt;td style="text-align: right"&gt;19,044 MW&lt;/td&gt;
 &lt;td&gt;58.6% of ELOA&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Source: &lt;a class="link" href="https://www.pjm.com/-/media/DotCom/committees-groups/subcommittees/las/2026/20260929/20260929-item-5ai---dominion_energy.pdf" target="_blank" rel="noopener"
 &gt;Dominion PJM filing, Sept. 29, 2026&lt;/a&gt;, pp. 5-6.&lt;/p&gt;
&lt;p&gt;One caveat. The 58.6% is the exclusion rate for early-stage requests only, not a &amp;ldquo;phantom rate&amp;rdquo; for all data centers. The 21,353 MW at ESA and CLOA stays in. Since September 2026, PJM also requires utilities to ask customers whether they filed the same load elsewhere (&lt;a class="link" href="https://insidelines.pjm.com/pjm-kicks-off-annual-process-of-vetting-large-load-requests/" target="_blank" rel="noopener"
 &gt;PJM, Sept. 29&lt;/a&gt;). PJM has not published an actual duplication rate.&lt;/p&gt;
&lt;p&gt;The national numbers look even more inflated. ERCOT&amp;rsquo;s 2032 peak forecast of 367.8 GW is about 4.3 times its all-time record of 85.5 GW (Texas Tribune, June 17, 2026). Bernstein judges only 33-35% of the 456 GW U.S. data center power pipeline credible (as cited in press reports; we did not see the original). These numbers are large, which is exactly why regulators and utilities cut them first.&lt;/p&gt;
&lt;p&gt;The real money in power is in equipment, not requests. GE Vernova&amp;rsquo;s 122 GW of gas turbine exposure at the end of Q3 is 64 GW of firm backlog plus 58 GW of slot reservations. In the quarter, 13 GW of reservations converted to firm orders, against 8 GW of new reservations (&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1996810/000199681026000162/gevpressrelease3q26.htm" target="_blank" rel="noopener"
 &gt;GE Vernova Q3 results, Oct. 28&lt;/a&gt;). Reservations turning into orders with deposits faster than new ones arrive means paid demand, not paper demand, is flowing into equipment.&lt;/p&gt;
&lt;p&gt;So the number to watch in power is not &amp;ldquo;requested GW falling.&amp;rdquo; If requests fall while ESA, CLOA, and turbine conversions hold, real demand is intact. If cancellations show up at the ESA or CLOA stage, that is no longer paper being filtered. That is a loss.&lt;/p&gt;
&lt;h2 id="cloud-the-same-capacity-counted-three-times-and-the-credit-behind-it"&gt;Cloud: the same capacity counted three times, and the credit behind it
&lt;/h2&gt;&lt;p&gt;Cloud commitments are the largest layer by dollars. Microsoft&amp;rsquo;s and Oracle&amp;rsquo;s remaining performance obligations were reported at about $677 billion and $664 billion (press reports of company filings, secondary). The problem is structure, not size.&lt;/p&gt;
&lt;p&gt;The same capacity gets counted more than once. Nebius&amp;rsquo;s deal with Meta pairs $12 billion of dedicated capacity with Meta&amp;rsquo;s commitment to buy up to $15 billion of capacity Nebius cannot sell to third parties (&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1513845/000110465926027886/tm268879d1_6k.htm" target="_blank" rel="noopener"
 &gt;Nebius 6-K, March 16, 2026&lt;/a&gt;). CoreWeave&amp;rsquo;s 2025 deal with NVIDIA is a similar unsold-capacity purchase with an initial value of $6.3 billion (&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1769628/000176962825000047/crwv-20250909.htm" target="_blank" rel="noopener"
 &gt;CoreWeave 8-K, Sept. 2025&lt;/a&gt;). Add third-party sales and the backstop maximum as demand and you count the same GPU twice. Add the subcontracting chain from AI lab to hyperscaler to neocloud and it is three times.&lt;/p&gt;
&lt;p&gt;Revenue does not tell you utilization. 98% of CoreWeave&amp;rsquo;s Q2 2026 revenue came from committed contracts (&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1769628/000176962826000366/crwv-20260630.htm" target="_blank" rel="noopener"
 &gt;CoreWeave 10-Q&lt;/a&gt;). Revenue paid regardless of use can hold up while GPUs sit idle.&lt;/p&gt;
&lt;p&gt;Risk is moved, not removed. Backstops are not fraud and not fake revenue. They do mean that capacity can be built on credit from inside the supply chain before an independent end customer shows up. If usage falls short, the cost is split by contract among the lab, the backstop provider, and the lenders, not just neocloud shareholders. Anthropic&amp;rsquo;s S-1 discloses $265.6 billion of fixed and minimum purchase commitments at June 30 (terms of 1 to 20 years) against first-half revenue of $16.29 billion (&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/2130357/000162828026067168/anthropicpbc-sx1.htm" target="_blank" rel="noopener"
 &gt;Anthropic S-1, Oct. 26&lt;/a&gt;). Dividing one by the other and calling it &amp;ldquo;16x over-ordering&amp;rdquo; is wrong, because the two cover different time spans. What matters is the payment schedule by year and actual delivery dates.&lt;/p&gt;
&lt;p&gt;Right now, the shortage is real. On October 28, Microsoft reported 48% Azure growth and demand exceeding supply (&lt;a class="link" href="https://www.microsoft.com/en-us/investor/events/fy-2027/earnings-fy-2027-q1" target="_blank" rel="noopener"
 &gt;Microsoft earnings, Oct. 28&lt;/a&gt;). On September 17, CoreWeave said it signed new 3-6 month contracts at about $40 million of annualized revenue per MW of cluster power (&lt;a class="link" href="https://investors.coreweave.com/news/news-details/2026/CoreWeave-Continues-to-Contract-New-Compute-Capacity-at-Higher-Prices/default.aspx" target="_blank" rel="noopener"
 &gt;CoreWeave, Sept. 17&lt;/a&gt;). High prices on short contracts show scarcity today.&lt;/p&gt;
&lt;p&gt;So the cloud risk is not &amp;ldquo;glut now.&amp;rdquo; It is the gap between long fixed commitments and short-term market prices. Once short-term prices turn, whoever holds long commitments on the same GPU generation eats the difference. This layer ranks second rather than first for one reason: the adjustment gets spread over years through renegotiation and term extensions. A February 2026 report that OpenAI&amp;rsquo;s compute spending target through 2030 was reset from about $1.4 trillion to about $600 billion (secondary) is an example of how that works.&lt;/p&gt;
&lt;h2 id="memory-briefly-yesterdays-conclusion-and-the-new-link"&gt;Memory: briefly, yesterday&amp;rsquo;s conclusion and the new link
&lt;/h2&gt;&lt;p&gt;Yesterday&amp;rsquo;s point: U.S. cloud providers&amp;rsquo; DRAM inventory rose in Q2 2026 because of server CPU shortages (&lt;a class="link" href="https://www.trendforce.com/presscenter/news/20260709-13140.html" target="_blank" rel="noopener"
 &gt;TrendForce, July 9&lt;/a&gt;). That inventory is already booked as supplier revenue, and it unwinds just as Intel&amp;rsquo;s server CPU supply increase arrives in the second half of 2027 and memory wafer capacity grows in 2028 (industry estimate: 15-17%).&lt;/p&gt;
&lt;p&gt;The new numbers are on the buyer&amp;rsquo;s books. Dell&amp;rsquo;s production materials inventory rose 137%, from $6.70 billion at the end of January 2026 to $15.89 billion at the end of July (&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1571996/000157199626000046/dell-20260731.htm" target="_blank" rel="noopener"
 &gt;Dell 10-Q&lt;/a&gt;). It is not all memory, but it is parts the buyer has already paid for. On the supply side, NVIDIA&amp;rsquo;s supply and capacity commitments rose from $119 billion to $279 billion in one quarter (&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1045810/000104581026000075/nvda-20260726.htm" target="_blank" rel="noopener"
 &gt;NVIDIA 10-Q&lt;/a&gt;, Note 10). Both are high cancellation cost with a lot of cash tied up.&lt;/p&gt;
&lt;h2 id="the-cross-stage-path-usage-can-rise-while-orders-halve"&gt;The cross-stage path: usage can rise while orders halve
&lt;/h2&gt;&lt;p&gt;The core of this piece is the link between stages. When power and CPUs are late, buyers stock memory first. When the bottleneck clears, that stock replaces new orders. A hypothetical example for a single part (not a market estimate):&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Item&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Quarter A&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Quarter B&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Actual usage (shipments)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;100&lt;/td&gt;
 &lt;td style="text-align: right"&gt;110&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Opening inventory&lt;/td&gt;
 &lt;td style="text-align: right"&gt;20&lt;/td&gt;
 &lt;td style="text-align: right"&gt;60&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Target closing inventory&lt;/td&gt;
 &lt;td style="text-align: right"&gt;60&lt;/td&gt;
 &lt;td style="text-align: right"&gt;20&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Orders = usage + target − opening&lt;/td&gt;
 &lt;td style="text-align: right"&gt;140&lt;/td&gt;
 &lt;td style="text-align: right"&gt;70&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Usage rose 10%. Orders fell 50%. Power delays and CPU shortages push inventory targets up, as in Quarter A. Clearing the bottleneck pulls them down, as in Quarter B. So &amp;ldquo;token usage keeps rising, so chip orders are fine&amp;rdquo; does not follow. In Ramp&amp;rsquo;s linked sample, token usage in the week ended October 23 was up 37.1% from four weeks earlier, while spending rose only 8.9% (&lt;a class="link" href="https://ramp.com/data/ai-index" target="_blank" rel="noopener"
 &gt;Ramp AI Index&lt;/a&gt;; sample limits apply). Usage and money are already moving at different speeds.&lt;/p&gt;
&lt;p&gt;&lt;img alt="Inflation size versus loss risk by AI supply chain stage. Grid requests are the most inflated but carry little risk; buyer memory inventory is moderately inflated but carries the most risk. Our qualitative assessment." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/ai-bullwhip-grid-cloud-memory-2026-10-31/bullwhip-stages.en.svg"&gt;
&lt;/p&gt;
&lt;h2 id="the-ranking-and-what-would-flip-it"&gt;The ranking, and what would flip it
&lt;/h2&gt;&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Rank&lt;/th&gt;
 &lt;th&gt;Stage&lt;/th&gt;
 &lt;th&gt;Cancellation cost&lt;/th&gt;
 &lt;th&gt;Cash tied up&lt;/th&gt;
 &lt;th&gt;Unwind overlaps supply growth&lt;/th&gt;
 &lt;th&gt;What would flip this rank&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;1&lt;/td&gt;
 &lt;td&gt;Buyer component and memory inventory&lt;/td&gt;
 &lt;td&gt;Already paid&lt;/td&gt;
 &lt;td&gt;Large (Dell materials +137%)&lt;/td&gt;
 &lt;td&gt;H2 2027 CPU supply + 2028 memory supply&lt;/td&gt;
 &lt;td&gt;If buyer inventory falls as CPU supply rises in H1 2027 and DRAM spot holds above contract → it was assembly backlog, not over-ordering; drops to 3&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;2&lt;/td&gt;
 &lt;td&gt;Supplier-backed cloud commitments&lt;/td&gt;
 &lt;td&gt;High on paper, renegotiated in practice&lt;/td&gt;
 &lt;td&gt;Cash only for usage, rest is commitment&lt;/td&gt;
 &lt;td&gt;Spread over years&lt;/td&gt;
 &lt;td&gt;If new short-term prices for the same GPU generation fall two quarters in a row, or backstops are actually drawn → rises to 1&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;3&lt;/td&gt;
 &lt;td&gt;Power generation equipment&lt;/td&gt;
 &lt;td&gt;Low for reservations, deposits for orders&lt;/td&gt;
 &lt;td&gt;Medium&lt;/td&gt;
 &lt;td&gt;2027-2028&lt;/td&gt;
 &lt;td&gt;If GE Vernova new reservations exceed conversions two quarters in a row and cancellations or deferrals are disclosed → rises to 2&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;4&lt;/td&gt;
 &lt;td&gt;Grid connection requests&lt;/td&gt;
 &lt;td&gt;Near zero&lt;/td&gt;
 &lt;td&gt;Near zero&lt;/td&gt;
 &lt;td&gt;N/A&lt;/td&gt;
 &lt;td&gt;If cancellations hit firm stages (ESA, CLOA) → it is a loss, not paper; rises with equipment&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;This ranking is our judgment, not a statistical ranking of over-order rates on a common denominator. What sets the order is the three right-hand columns, not the size of the inflation.&lt;/p&gt;
&lt;h2 id="counter-evidence-and-limits"&gt;Counter-evidence and limits
&lt;/h2&gt;&lt;p&gt;The evidence against a broad glut is clear. Azure growth and short-term contract prices show a shortage today. On its October 15 call, TSMC said it checks demand with its customers&amp;rsquo; customers and visits U.S. data center construction and power sites (&lt;a class="link" href="https://investor.tsmc.com/english/encrypt/files/encrypt_file/reports/2026-10/eba4fac72873bf546b669bd2b5d973df5c93215d/TSMC%203Q26%20Transcript.pdf" target="_blank" rel="noopener"
 &gt;TSMC Q3 transcript&lt;/a&gt;, pp. 4-5). This piece is not about &amp;ldquo;a bubble now.&amp;rdquo; It is about who cuts orders first in the next cycle.&lt;/p&gt;
&lt;p&gt;Some things we could not measure. Cloud providers&amp;rsquo; DRAM weeks of inventory and PJM&amp;rsquo;s actual duplication rate are not disclosed. ERCOT, Bernstein, Microsoft and Oracle obligations, and the OpenAI figure come from press reports. Anthropic figures are from its S-1 only.&lt;/p&gt;
&lt;h2 id="what-to-check-next"&gt;What to check next
&lt;/h2&gt;&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Timing&lt;/th&gt;
 &lt;th&gt;What&lt;/th&gt;
 &lt;th&gt;Strengthens this view&lt;/th&gt;
 &lt;th&gt;Weakens it&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Mid-Nov&lt;/td&gt;
 &lt;td&gt;NVIDIA October quarter&lt;/td&gt;
 &lt;td&gt;Supply commitments outgrow revenue&lt;/td&gt;
 &lt;td&gt;Commitments slow, inventory falls&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Late Nov&lt;/td&gt;
 &lt;td&gt;Dell Q3 10-Q&lt;/td&gt;
 &lt;td&gt;Materials inventory outgrows revenue&lt;/td&gt;
 &lt;td&gt;Inventory falls while revenue rises&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Nov-Dec&lt;/td&gt;
 &lt;td&gt;PJM 2027 load forecast&lt;/td&gt;
 &lt;td&gt;More early-stage exclusions, firm stages intact&lt;/td&gt;
 &lt;td&gt;Cancellations at ESA or CLOA&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Next quarter&lt;/td&gt;
 &lt;td&gt;CoreWeave, Nebius results&lt;/td&gt;
 &lt;td&gt;Short-term prices fall, backstops drawn&lt;/td&gt;
 &lt;td&gt;Independent third-party usage rises&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Mid-Dec&lt;/td&gt;
 &lt;td&gt;Micron November quarter&lt;/td&gt;
 &lt;td&gt;Customer inventory comments plus supplier inventory rebound&lt;/td&gt;
 &lt;td&gt;Customer inventory falling&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Late Jan 2027&lt;/td&gt;
 &lt;td&gt;GE Vernova Q4&lt;/td&gt;
 &lt;td&gt;(Power rank rises) new reservations &amp;gt; conversions&lt;/td&gt;
 &lt;td&gt;Conversions stay above new reservations&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;&lt;strong&gt;Which cell in this ranking do you think flips first? If you have a number we are missing, tell us in the comments.&lt;/strong&gt;&lt;/p&gt;
&lt;h2 id="key-sources"&gt;Key sources
&lt;/h2&gt;&lt;ul&gt;
&lt;li&gt;&lt;a class="link" href="https://www.pjm.com/-/media/DotCom/committees-groups/subcommittees/las/2026/20260929/20260929-item-5ai---dominion_energy.pdf" target="_blank" rel="noopener"
 &gt;Dominion Energy, PJM filing, Sept. 29, 2026&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://insidelines.pjm.com/pjm-kicks-off-annual-process-of-vetting-large-load-requests/" target="_blank" rel="noopener"
 &gt;PJM, large load vetting process, Sept. 29, 2026&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1996810/000199681026000162/gevpressrelease3q26.htm" target="_blank" rel="noopener"
 &gt;GE Vernova Q3 results, Oct. 28, 2026&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1513845/000110465926027886/tm268879d1_6k.htm" target="_blank" rel="noopener"
 &gt;Nebius 6-K, March 16, 2026&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1769628/000176962825000047/crwv-20250909.htm" target="_blank" rel="noopener"
 &gt;CoreWeave 8-K, Sept. 15, 2025&lt;/a&gt; · &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1769628/000176962826000366/crwv-20260630.htm" target="_blank" rel="noopener"
 &gt;10-Q, Q2 2026&lt;/a&gt; · &lt;a class="link" href="https://investors.coreweave.com/news/news-details/2026/CoreWeave-Continues-to-Contract-New-Compute-Capacity-at-Higher-Prices/default.aspx" target="_blank" rel="noopener"
 &gt;Short-term contracts, Sept. 17&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/2130357/000162828026067168/anthropicpbc-sx1.htm" target="_blank" rel="noopener"
 &gt;Anthropic S-1, Oct. 26, 2026&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1045810/000104581026000075/nvda-20260726.htm" target="_blank" rel="noopener"
 &gt;NVIDIA 10-Q, July 26, 2026&lt;/a&gt; · &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1571996/000157199626000046/dell-20260731.htm" target="_blank" rel="noopener"
 &gt;Dell 10-Q, July 31, 2026&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.microsoft.com/en-us/investor/events/fy-2027/earnings-fy-2027-q1" target="_blank" rel="noopener"
 &gt;Microsoft FY27 Q1 earnings, Oct. 28&lt;/a&gt; · &lt;a class="link" href="https://investor.tsmc.com/english/encrypt/files/encrypt_file/reports/2026-10/eba4fac72873bf546b669bd2b5d973df5c93215d/TSMC%203Q26%20Transcript.pdf" target="_blank" rel="noopener"
 &gt;TSMC Q3 transcript, Oct. 15&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.trendforce.com/presscenter/news/20260709-13140.html" target="_blank" rel="noopener"
 &gt;TrendForce, U.S. cloud DRAM inventory, July 9, 2026&lt;/a&gt; · &lt;a class="link" href="https://ramp.com/data/ai-index" target="_blank" rel="noopener"
 &gt;Ramp AI Index&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.texastribune.org/2026/06/17/texas-ercot-data-center-energy-grid/" target="_blank" rel="noopener"
 &gt;Texas Tribune, ERCOT large loads, June 17, 2026&lt;/a&gt; · &lt;a class="link" href="https://finance.yahoo.com/news/openai-sees-compute-spend-around-223950561.html" target="_blank" rel="noopener"
 &gt;OpenAI compute spending report (secondary)&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://koreainvestinsights.com/post/dram-shortage-part-inventory-buyers-stop-when-cpus-arrive-2026-10-30/" &gt;KII, Part of the DRAM Shortage Is Inventory, Oct. 30, 2026&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;&lt;em&gt;This analysis is based on public information and is not investment advice.&lt;/em&gt;&lt;/p&gt;</description></item><item><title>Korea's October Chip Exports: A $53 Billion Forecast and a Rebound Test</title><link>https://koreainvestinsights.com/post/korea-chip-exports-november-rebound-test-2026-10-30/</link><pubDate>Fri, 30 Oct 2026 17:30:00 +0900</pubDate><guid>https://koreainvestinsights.com/post/korea-chip-exports-november-rebound-test-2026-10-30/</guid><description>&lt;img src="https://koreainvestinsights.com/" alt="Featured image of post Korea's October Chip Exports: A $53 Billion Forecast and a Rebound Test" /&gt;&lt;p&gt;&lt;strong&gt;I expect Korea&amp;rsquo;s October semiconductor exports to be about $53 billion, within a base range of $50–57 billion.&lt;/strong&gt; That would be roughly 12% below September but 238% above the previous October. This is my forecast before the release, not a reported result or market consensus.&lt;/p&gt;
&lt;p&gt;The early-November export release could catalyze a rebound in Korean semiconductor shares. It can show that higher prices are turning into export receipts and that shipments held up late in the month. But &lt;strong&gt;a headline of more than 200% yearly growth is not enough new information&lt;/strong&gt;: growth through October 20 was already 222.5%.&lt;/p&gt;
&lt;p&gt;My first constructive reference point is $58 billion, above the base range. The stronger test is approximately $62 billion. Assuming 22 working days in October, that would also maintain September&amp;rsquo;s daily export pace. These are analytical reference points, not thresholds that guarantee a share-price rally.&lt;/p&gt;
&lt;p&gt;&lt;em&gt;As of October 30, 2026. This previews October trade data expected around November 1. I have not verified the official time for this particular release; Customs and the trade ministry can publish on different dates. November 1 is a Sunday, so equity-market confirmation would come from November 2 onward, depending on publication timing.&lt;/em&gt;&lt;/p&gt;
&lt;h2 id="yearly-growth-is-strong-the-recent-daily-pace-is-softer"&gt;Yearly growth is strong. The recent daily pace is softer.
&lt;/h2&gt;&lt;p&gt;Korea Customs Service reported semiconductor exports of $27.516 billion for October 1–20, versus $34.128 billion for September 1–20. That is a 19.4% decline in total value. Working days also fell from 15.5 to 13, making the raw decline a poor measure of demand. &lt;a class="link" href="https://www.korea.kr/common/download.do?fileId=198574612&amp;amp;tblKey=GMN" target="_blank" rel="noopener"
 &gt;October Customs PDF&lt;/a&gt;, &lt;a class="link" href="https://www.korea.kr/common/download.do?fileId=198556844&amp;amp;tblKey=GMN" target="_blank" rel="noopener"
 &gt;September Customs PDF&lt;/a&gt;&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;First 20 calendar days&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Chip exports, $bn&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Working days&lt;/th&gt;
 &lt;th style="text-align: right"&gt;$bn per working day&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;July 2026&lt;/td&gt;
 &lt;td style="text-align: right"&gt;22.113&lt;/td&gt;
 &lt;td style="text-align: right"&gt;14.5&lt;/td&gt;
 &lt;td style="text-align: right"&gt;1.525&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;August 2026&lt;/td&gt;
 &lt;td style="text-align: right"&gt;26.032&lt;/td&gt;
 &lt;td style="text-align: right"&gt;14.0&lt;/td&gt;
 &lt;td style="text-align: right"&gt;1.859&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;September 2026&lt;/td&gt;
 &lt;td style="text-align: right"&gt;34.128&lt;/td&gt;
 &lt;td style="text-align: right"&gt;15.5&lt;/td&gt;
 &lt;td style="text-align: right"&gt;2.202&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;October 2026&lt;/td&gt;
 &lt;td style="text-align: right"&gt;27.516&lt;/td&gt;
 &lt;td style="text-align: right"&gt;13.0&lt;/td&gt;
 &lt;td style="text-align: right"&gt;2.117&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Sources: Customs &lt;a class="link" href="https://www.customs.go.kr/common/nttFileDownload.do?fileKey=85da4d4aa472a42f8a2be178eaea98eb" target="_blank" rel="noopener"
 &gt;July&lt;/a&gt;, &lt;a class="link" href="https://www.customs.go.kr/common/nttFileDownload.do?fileKey=9e0f2c18a1f61b82dfa34a7f821c07c7" target="_blank" rel="noopener"
 &gt;August&lt;/a&gt;, and the September/October provisional releases above. Daily values are calculated. This is the Customs semiconductor commodity category, not HBM alone.&lt;/p&gt;
&lt;p&gt;&lt;img alt="Korea’s first-20-day semiconductor exports per working day. October is about 3.9% below September." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/korea-chip-exports-november-rebound-test-2026-10-30/daily.en.png"&gt;
&lt;/p&gt;
&lt;p&gt;After adjusting for working days, the sequential decline is about 3.9%. That is much less alarming than 19.4%, but it also does not establish continued acceleration. These are dollar values, not bit shipments. Prices, product mix and quantity cannot be separated from this aggregate alone.&lt;/p&gt;
&lt;h2 id="the-53-billion-forecast-assumes-a-35-late-month-uplift"&gt;The $53 billion forecast assumes a 35% late-month uplift
&lt;/h2&gt;&lt;p&gt;Extending the observed daily rate across October would produce only about $46.6 billion. A forecast should also consider the concentration of customs clearance and export recognition toward month-end.&lt;/p&gt;
&lt;p&gt;The model, in US$ billions, is:&lt;/p&gt;

 &lt;blockquote&gt;
 &lt;p&gt;October = 27.516 + (27.516 / 13) × 9 remaining working days × 1.35 late-month multiplier&lt;/p&gt;
&lt;p&gt;Calculated result: $53.23bn. Rounded central forecast: $53bn.&lt;/p&gt;

 &lt;/blockquote&gt;
&lt;p&gt;The remaining nine working days are an explicit assumption: eight weekdays plus two Saturdays counted at half a day each, from October 21–31. Added to the official first-period count of 13, that gives 22. The final release must be checked against this count. A one-day change moves the base estimate by about $2.86 billion.&lt;/p&gt;
&lt;p&gt;The 1.35 multiplier is grounded in two recent non-quarter-end months. Bridging Customs&amp;rsquo; interim numbers to the ministry&amp;rsquo;s monthly semiconductor totals implies a remaining-period daily rate about 30% higher in July and 39% higher in August. I use 35% between those observations. September&amp;rsquo;s implied uplift was much larger, about 98%. Repeating that concentration would produce about $65.2 billion for October. &lt;a class="link" href="https://www.korea.kr/common/download.do?fileId=198563931&amp;amp;tblKey=GMN" target="_blank" rel="noopener"
 &gt;Monthly totals and prior-year base: ministry September release, page 20&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;This bridge has a limitation: Customs&amp;rsquo; interim commodity classification and the ministry&amp;rsquo;s major-export-product classification are not guaranteed to be identical. Classification and provisional-data revisions may enter the multiplier. &lt;strong&gt;It is an approximation linking releases, not a pure physical shipment-rate estimate.&lt;/strong&gt; Two calibration months do not make a statistically validated forecasting model.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Assumption&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Late-month multiplier&lt;/th&gt;
 &lt;th style="text-align: right"&gt;October exports&lt;/th&gt;
 &lt;th&gt;Interpretation&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;No extra concentration&lt;/td&gt;
 &lt;td style="text-align: right"&gt;1.00x&lt;/td&gt;
 &lt;td style="text-align: right"&gt;$46.6bn&lt;/td&gt;
 &lt;td&gt;Observed daily pace continues&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Base&lt;/td&gt;
 &lt;td style="text-align: right"&gt;1.35x&lt;/td&gt;
 &lt;td style="text-align: right"&gt;$53.2bn&lt;/td&gt;
 &lt;td&gt;Reference: July and August&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;September-style concentration&lt;/td&gt;
 &lt;td style="text-align: right"&gt;About 1.98x&lt;/td&gt;
 &lt;td style="text-align: right"&gt;$65.2bn&lt;/td&gt;
 &lt;td&gt;Upside sensitivity, not the base case&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;&lt;img alt="October export scenarios split observed first-20-day exports from the modeled remainder." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/korea-chip-exports-november-rebound-test-2026-10-30/scenarios.en.png"&gt;
&lt;/p&gt;
&lt;p&gt;Multipliers of 1.20–1.55 yield about $50.4–57.0 billion, rounded to a $50–57 billion base range. This is not a confidence interval or probability distribution. A repeat of September&amp;rsquo;s concentration could exceed it substantially.&lt;/p&gt;
&lt;h2 id="dram-prices-support-the-case-they-are-not-an-extra-export-multiplier"&gt;DRAM prices support the case. They are not an extra export multiplier.
&lt;/h2&gt;&lt;p&gt;The supplied SK Securities research summary describes further Q4 PC DRAM price increases, at a slower rate. TrendForce&amp;rsquo;s October 30 public abstract also describes continued increases, some negotiations moving into November, weaker notebook demand and rising buyer inventories. I could not independently reconcile the detailed prices or the cited 13–18% quarterly increase against the paywalled full report. It should not be presented as a fully settled quarter. &lt;a class="link" href="https://www.trendforce.com.tw/research/download/RP261030JL" target="_blank" rel="noopener"
 &gt;TrendForce October contract-price abstract&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;These selected figures are arithmetic checks on the supplied memo, not independently verified primary price observations.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Product&lt;/th&gt;
 &lt;th style="text-align: right"&gt;September → October, US$&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Calculated change&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;DDR5 16GB module&lt;/td&gt;
 &lt;td style="text-align: right"&gt;247 → 275&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+11.3%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;DDR4 8GB module&lt;/td&gt;
 &lt;td style="text-align: right"&gt;148 → 170&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+14.9%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;DDR5 16Gb chip&lt;/td&gt;
 &lt;td style="text-align: right"&gt;48 → 50&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+4.2%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;DDR4 8Gb chip&lt;/td&gt;
 &lt;td style="text-align: right"&gt;26 → 27.5&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+5.8%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Source: user-supplied TrendForce summary attributed to SK Securities&amp;rsquo; semiconductor and overseas-equity research team; &lt;a class="link" href="https://t.me/skitteam" target="_blank" rel="noopener"
 &gt;research channel&lt;/a&gt;. GB module capacity and Gb chip capacity are different units; transaction terms are not assumed identical.&lt;/p&gt;
&lt;p&gt;Multiplying September&amp;rsquo;s entire chip exports by a PC-memory price increase would be a mistake. Korean semiconductor exports include server memory, HBM, NAND and non-memory products. October&amp;rsquo;s observed export value already incorporates prices and mix through October 20. Applying another broad DRAM uplift to this model could partly double-count the same effect.&lt;/p&gt;
&lt;p&gt;The claimed narrowing of the DDR5 discount also depends on the SKU. Using the supplied monthly prices, 8GB DDR5&amp;rsquo;s discount to DDR4 narrows from 10.1% to 2.9%; at 16GB it widens from 10.2% to 11.3%. The memo&amp;rsquo;s separate Q3-average figure of 6% could not be reconciled to a verified quarterly denominator. One product and one measurement window cannot represent the entire DDR5 market.&lt;/p&gt;
&lt;p&gt;Spot prices are not uniformly rising either. The October 30 public board shows some DDR5 items up and some DDR4 items flat or down. A separate Specialty DRAM abstract discusses spot prices retreating from peaks. Product segments and contract/spot series must remain distinct. &lt;a class="link" href="https://www.trendforce.com.tw/price/dram/dram_contract" target="_blank" rel="noopener"
 &gt;Public price board&lt;/a&gt;, &lt;a class="link" href="https://www.trendforce.com/research/download/RP261030HH" target="_blank" rel="noopener"
 &gt;Specialty DRAM abstract&lt;/a&gt;&lt;/p&gt;
&lt;h2 id="58-billion-would-beat-this-model-62-billion-would-answer-a-harder-question"&gt;$58 billion would beat this model. $62 billion would answer a harder question.
&lt;/h2&gt;&lt;p&gt;The useful surprise is evidence against recent deceleration, rather than another large annual growth headline. I have not obtained a formal market consensus for semiconductor exports. This table compares the release with my own model.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Reported result&lt;/th&gt;
 &lt;th&gt;My interpretation&lt;/th&gt;
 &lt;th&gt;Rebound relevance&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Below $50bn&lt;/td&gt;
 &lt;td&gt;Below the base range; weak late-month recovery is possible&lt;/td&gt;
 &lt;td&gt;Exports alone offer limited support&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;$50–57bn&lt;/td&gt;
 &lt;td&gt;Strong yearly growth alongside a sequential slowdown&lt;/td&gt;
 &lt;td&gt;Confirmation; could support an oversold argument&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;$58bn or more&lt;/td&gt;
 &lt;td&gt;Above the base range; better late-month receipts than assumed&lt;/td&gt;
 &lt;td&gt;A plausible short-run catalyst&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;About $62bn or more&lt;/td&gt;
 &lt;td&gt;Also maintains/exceeds September&amp;rsquo;s daily pace on a 22-day assumption&lt;/td&gt;
 &lt;td&gt;Stronger evidence against recent deceleration&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;These are rounded reference points on a continuous scale, not mechanical boundaries. There is no special cliff between $57 billion and $58 billion, and the $62 billion case also satisfies the $58 billion condition.&lt;/p&gt;
&lt;p&gt;September&amp;rsquo;s monthly semiconductor exports were $60.295 billion over 21.5 working days. Matching that daily pace over 22 October working days requires about $61.7 billion. A small new monthly record above $60.3 billion could therefore coexist with a lower daily pace.&lt;/p&gt;
&lt;p&gt;Shares could rebound around $53 billion if investors had priced in greater damage. They could also fall after $62 billion if rates, currencies, US chip stocks or forced selling dominate. On the first session after the release, I would look for semiconductor relative strength, retention of early gains and easing foreign-investor selling. Those observations would strengthen the link between a good physical-data release and actual buying.&lt;/p&gt;
&lt;h2 id="strong-exports-do-not-settle-the-buyer-inventory-debate"&gt;Strong exports do not settle the buyer-inventory debate
&lt;/h2&gt;&lt;p&gt;The strongest counterargument concerns where the memory goes after customs. Export value does not tell us whether a customer installed the chips in servers or placed them in inventory. Purchases pulled forward by fear of higher prices can raise today&amp;rsquo;s exports while borrowing from future orders.&lt;/p&gt;
&lt;p&gt;TrendForce has discussed inventory accumulation alongside PC demand pressure. A strong October release and concern about future buyer inventories can both be valid. &lt;a class="link" href="https://www.trendforce.com/presscenter/news/20260930-13258.html" target="_blank" rel="noopener"
 &gt;TrendForce&amp;rsquo;s September outlook&lt;/a&gt;, &lt;a class="link" href="https://koreainvestinsights.com/post/dram-shortage-part-inventory-buyers-stop-when-cpus-arrive-2026-10-30/" target="_blank" rel="noopener"
 &gt;our earlier buyer-inventory analysis&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;My assessment leaves room for a short-term rebound. It does not turn one export release into proof of 2027–2028 final demand, lasting supply shortages, or company-specific shipments at Samsung and SK hynix. Memory/non-memory details, subsequent customs data and customer inventories remain the follow-through tests.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;If the number is $53 billion, would you emphasize “still exceptionally strong” or “weaker than September”?&lt;/strong&gt; Leave the export figure that would change your view, and why. I will look for both a result above the base range and evidence that buyers respond.&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Forecast frozen using information available on October 30. The base range is not a statistical confidence interval or an individualized trading recommendation. Publication timing, working-day assumptions, classification and revisions can change the calculation.&lt;/em&gt;&lt;/p&gt;</description></item><item><title>Part of the DRAM Shortage Is Inventory. The Day CPUs Arrive, Buyers Stop Buying.</title><link>https://koreainvestinsights.com/post/dram-shortage-part-inventory-buyers-stop-when-cpus-arrive-2026-10-30/</link><pubDate>Fri, 30 Oct 2026 11:29:00 +0900</pubDate><guid>https://koreainvestinsights.com/post/dram-shortage-part-inventory-buyers-stop-when-cpus-arrive-2026-10-30/</guid><description>&lt;img src="https://koreainvestinsights.com/" alt="Featured image of post Part of the DRAM Shortage Is Inventory. The Day CPUs Arrive, Buyers Stop Buying." /&gt;&lt;p&gt;The market story is simple. DRAM stays short through 2028, and the three memory makers have locked 2027-28 volume into long-term agreements. On its call, SK hynix said customers asked for more than twice the volume it can supply (call transcript, secondary source). &lt;a class="link" href="https://koreainvestinsights.com/post/samsung-sk-hynix-lta-volume-locked-terms-flexible-2026-10-29/" &gt;KII, Oct 29&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;We think part of that shortage is buyers building inventory, not end consumption. A server needs a CPU, DRAM and a substrate. Right now CPUs and substrates are short, so the DRAM that arrived first sits in a warehouse waiting for assembly. Yesterday Intel named the substrate shortage as a limit on server CPU output and said most of its substrate suppliers&amp;rsquo; new capacity arrives in the second half of 2027. &lt;a class="link" href="https://koreainvestinsights.com/post/substrate-shortage-longer-stocks-fell-daeduck-readthrough-2026-10-30/" &gt;KII, today&amp;rsquo;s substrate post&lt;/a&gt; That is when DRAM buyers start using what they stockpiled and cut new orders.&lt;/p&gt;
&lt;p&gt;This setup has a precedent. In the fourth quarter of 2021, PC makers held 11-13 weeks of DRAM. TrendForce wrote that DRAM had become a component &amp;ldquo;in relative surplus compared to other components,&amp;rdquo; and that PC makers would lower DRAM inventory as component gaps were partially resolved. It forecast first-quarter 2022 DRAM average prices 8-13% below the prior quarter. &lt;a class="link" href="https://www.trendforce.com/presscenter/news/20211213-11050.html" target="_blank" rel="noopener"
 &gt;TrendForce, Dec 13, 2021&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Our view: 2027 memory profits are strong. But we estimate that 8-13% of 2027 DRAM shipments refill buyer inventory, and that share does not repeat. The risk sits in commodity server and PC DRAM more than in HBM. The trigger is CPU and substrate supply in the second half of 2027, which overlaps with new fab output from the memory makers. The size of the buffer cannot be measured precisely. There is also no evidence that it is smaller than in 2021.&lt;/strong&gt;&lt;/p&gt;
&lt;h2 id="buyers-are-stockpiling-dram-now"&gt;Buyers are stockpiling DRAM now
&lt;/h2&gt;&lt;p&gt;The build-up is in the filings and reports.&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Cloud:&lt;/strong&gt; TrendForce wrote that US cloud providers&amp;rsquo; DRAM inventory rose in the second quarter of 2026 because of server CPU shortages. &lt;a class="link" href="https://www.trendforce.com/presscenter/news/20260709-13140.html" target="_blank" rel="noopener"
 &gt;TrendForce, Jul 9&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;PCs:&lt;/strong&gt; TrendForce wrote that PC makers are procuring aggressively because they expect shortages in 2027. &lt;a class="link" href="https://www.trendforce.com/presscenter/news/20260930-13258.html" target="_blank" rel="noopener"
 &gt;TrendForce, Sep 30&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Server makers:&lt;/strong&gt; HPE&amp;rsquo;s quarterly report says &amp;ldquo;our current efforts to secure memory components and SSDs in advance to meet forecasted demand may further increase our inventory levels in the medium term.&amp;rdquo; Its unconditional purchase obligations are about $30.4bn, including services. &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1645590/000164559026000080/hpe-20260731.htm" target="_blank" rel="noopener"
 &gt;HPE 10-Q&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Dell:&lt;/strong&gt; Raw materials inventory rose 3.4 times in a year, from $4.68bn to $15.89bn. Purchase commitments rose 67% in one quarter, from $20.8bn to $34.8bn. &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1571996/000157199626000046/dell-20260731.htm" target="_blank" rel="noopener"
 &gt;Dell 10-Q&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Apple:&lt;/strong&gt; Components inventory rose 3.6 times in nine months, from $2.12bn to $7.65bn. Manufacturing purchase obligations are $57.0bn, with $56.2bn payable within 12 months. &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/320193/000032019326000020/aapl-20260627.htm" target="_blank" rel="noopener"
 &gt;Apple 10-Q&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;&lt;img alt="Latest-quarter growth in inventory and revenue for Apple, HPE, HP Inc., Lenovo and Xiaomi. Inventory: Apple +87%, HPE +65%, HP Inc. +24%, Lenovo +80%, Xiaomi +29%. Revenue: +16%, +34%, +13%, +43% and -6%. Lenovo and Xiaomi are secondary data." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/dram-shortage-part-inventory-buyers-stop-when-cpus-arrive-2026-10-30/buyer-inventory-vs-sales.en.png"&gt;
&lt;/p&gt;
&lt;p&gt;Device and server makers&amp;rsquo; inventories grew roughly 2 to 5 times faster than their sales. Inventory value mixes in memory price increases, GPUs and finished servers, so these numbers should not be read as DRAM units. The direction is clear. Supplier inventory is near historical lows (TrendForce as cited in press reports, secondary source), and inventory has moved from suppliers to buyers. Supplier revenue is booked on shipment, so DRAM piling up in buyers&amp;rsquo; warehouses still shows up as growth in suppliers&amp;rsquo; results.&lt;/p&gt;
&lt;h2 id="why-it-piles-up-orders-equal-consumption-plus-the-change-in-target-inventory"&gt;Why it piles up: orders equal consumption plus the change in target inventory
&lt;/h2&gt;&lt;p&gt;A buyer&amp;rsquo;s orders equal consumption plus the change in target inventory. Target inventory is target weeks times weekly consumption. So even with flat consumption, raising target weeks makes orders jump.&lt;/p&gt;
&lt;p&gt;The arithmetic is simple. In a year when a buyer lifts its target from six weeks to eight, orders run about 4% above consumption even with no change in use (2 weeks ÷ 52). In the year it cuts the target back to six weeks, orders run about 4% below consumption. Without any drop in consumption, orders swing by about 8 percentage points between those two years.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Buyer rule&lt;/th&gt;
 &lt;th&gt;How it works&lt;/th&gt;
 &lt;th&gt;What we see now&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Kitting delay&lt;/td&gt;
 &lt;td&gt;The scarcest part sets how many servers get built; spare DRAM waits&lt;/td&gt;
 &lt;td&gt;Cloud DRAM inventory rose on CPU shortages (TrendForce, July)&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Higher safety stock&lt;/td&gt;
 &lt;td&gt;Longer lead times and more uncertainty raise target weeks&lt;/td&gt;
 &lt;td&gt;PC makers&amp;rsquo; aggressive procurement (TrendForce, September)&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Buying ahead of price&lt;/td&gt;
 &lt;td&gt;Buy early when next quarter&amp;rsquo;s increase exceeds holding cost&lt;/td&gt;
 &lt;td&gt;Purchase obligations and inventory jumps at Apple, HPE and Dell&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Allocation games&lt;/td&gt;
 &lt;td&gt;The lower the allocation rate, the more buyers over-request&lt;/td&gt;
 &lt;td&gt;Customer requests above twice supply (SK hynix call, secondary)&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Buyers do not cancel the DRAM that is waiting. It is tied up in long-term agreements and prepayments, they will need it when CPUs arrive, and they fear losing their allocation. So the build-up is quiet, and the unwind comes all at once.&lt;/p&gt;
&lt;h2 id="size-there-is-no-evidence-it-is-smaller-than-2021"&gt;Size: there is no evidence it is smaller than 2021
&lt;/h2&gt;&lt;p&gt;We estimate the size as shipments minus installations. DRAM bit shipments grow 20-25% in 2026 (Micron and broker estimates, secondary). Weighting separate installation growth rates for servers, HBM, mobile, PCs and graphics gives installation growth of 6-15%. The difference stays in buyers&amp;rsquo; warehouses.&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;The net build in 2026 alone equals 2-9 weeks of end-2026 consumption (center about 5 weeks).&lt;/li&gt;
&lt;li&gt;The excess over normal inventory is 1-5 weeks on the base path. It depends on buyers&amp;rsquo; starting inventory at end-2025, and it is near zero if installation runs fast.&lt;/li&gt;
&lt;li&gt;In 2021, PC makers held 11-13 weeks of DRAM in total. With a 6-8 week normal, the excess was 3-7 weeks.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;&lt;img alt="Weeks of DRAM held by buyers. 2021 PC makers: 11-13 weeks total, 3-7 weeks above a 6-8 week normal. End-2026 buyers: a 2026 net build of 2-9 weeks in our model (center 5) and 1-5 weeks above normal." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/dram-shortage-part-inventory-buyers-stop-when-cpus-arrive-2026-10-30/buffer-weeks-vs-2021.en.png"&gt;
&lt;/p&gt;
&lt;p&gt;Today&amp;rsquo;s excess may be smaller than 2021&amp;rsquo;s, or about the same. There is no consistent series with the same definition, so we cannot tell. Not knowing the size is no reason for comfort.&lt;/p&gt;
&lt;p&gt;The speed worries us more. In our model, fourth-quarter 2026 shipments run about 17% above installations. As long as that level gap remains, the buffer keeps growing every quarter in 2027 even if the two growth rates converge. The model puts about 13% of 2027 DRAM shipments into buffer refill. If installation grows 24% and server assembly speeds up, that drops to about 8%. The moment buffer building stops, that share disappears from orders even with unchanged consumption.&lt;/p&gt;
&lt;h2 id="this-is-how-2021-ended"&gt;This is how 2021 ended
&lt;/h2&gt;&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Stage&lt;/th&gt;
 &lt;th&gt;2021-22&lt;/th&gt;
 &lt;th&gt;2026-27&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Scarce part&lt;/td&gt;
 &lt;td&gt;Components other than DRAM (DRAM in relative surplus)&lt;/td&gt;
 &lt;td&gt;Server CPUs and advanced EMIB substrates&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Where DRAM piled up&lt;/td&gt;
 &lt;td&gt;PC makers, 11-13 weeks total (4Q21)&lt;/td&gt;
 &lt;td&gt;US cloud, server makers, PC makers, size undisclosed&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Trigger&lt;/td&gt;
 &lt;td&gt;Component gaps partially resolved&lt;/td&gt;
 &lt;td&gt;Substrate suppliers add about 40%, mostly in 2H27 (Intel call, secondary)&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Price&lt;/td&gt;
 &lt;td&gt;1Q22 DRAM average prices forecast down 8-13% QoQ (TrendForce)&lt;/td&gt;
 &lt;td&gt;Not yet known&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;In 2021, PC makers were expected to lower DRAM inventory once component gaps partially closed. Buyers first use the DRAM they stockpiled, and when they decide the shortage is over they also cut target inventory. Both moves pull orders below consumption. Today&amp;rsquo;s trigger is the same kind. Intel said the industry stays supply-constrained throughout 2027 and into 2028. The second half of 2027, when substrates start to loosen, is the earliest window for the unwind.&lt;/p&gt;
&lt;h2 id="commodity-dram-is-riskier-than-hbm"&gt;Commodity DRAM is riskier than HBM
&lt;/h2&gt;&lt;p&gt;Nvidia takes the HBM bullwhip first. Its supply and capacity commitments rose from $119bn to $279bn in a single quarter. 96% falls in fiscal 2027-2029, and they are &amp;ldquo;primarily memory and manufacturing facilities.&amp;rdquo; &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1045810/000104581026000075/nvda-20260726.htm" target="_blank" rel="noopener"
 &gt;Nvidia 10-Q&lt;/a&gt; If GPU demand turns, these commitments show up first as Nvidia write-downs, before memory makers see cancelled orders. There is a precedent. In the first quarter of fiscal 2026, Nvidia took &amp;ldquo;a $4.5 billion charge associated with H20 excess inventory and purchase obligations,&amp;rdquo; and it still carries a $2.1bn liability for excess inventory purchase obligations.&lt;/p&gt;
&lt;p&gt;Commodity server and PC DRAM has no such buffer. Dell, HPE and cloud providers buy directly from the memory makers. When the buffer unwinds, that volume turns straight into fewer new orders for memory makers. The bigger a maker&amp;rsquo;s commodity DRAM mix, the larger its exposure. In Korea, Samsung Electronics sits closer to this risk than SK hynix.&lt;/p&gt;
&lt;h2 id="when-this-warning-would-be-wrong"&gt;When this warning would be wrong
&lt;/h2&gt;&lt;ul&gt;
&lt;li&gt;If server unit growth catches up with server DRAM bit growth in the first half of 2027, today&amp;rsquo;s buffer was just assembly delay. The risk shrinks.&lt;/li&gt;
&lt;li&gt;If buyer inventory days fall while CPU supply rises and DRAM spot prices hold above contract prices, the buffer went into real servers.&lt;/li&gt;
&lt;li&gt;There is already a counter-signal. In September, TrendForce wrote that &amp;ldquo;improving server CPU availability has prompted CSPs and server OEMs to increase bit procurement.&amp;rdquo; As CPUs loosened, DRAM procurement rose instead of falling. If that holds through 2027, the buffer was assembly delay.&lt;/li&gt;
&lt;li&gt;If buyer inventories rise while same-generation GPU rental rates and follow-on orders both weaken, the unwind has come early. When inventory, rental rates and follow-on orders move the same way, the warning becomes reality.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;Our 2027 memory profit view stands. The warning targets the share of that profit that will not repeat, and the timing of when it disappears. Long-term agreements can lock volume, but they cannot lock buyers&amp;rsquo; inventory decisions.&lt;/p&gt;
&lt;h2 id="what-to-watch-next"&gt;What to watch next
&lt;/h2&gt;&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;When&lt;/th&gt;
 &lt;th&gt;What&lt;/th&gt;
 &lt;th&gt;How to read it&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Early Nov&lt;/td&gt;
 &lt;td&gt;Apple and Supermicro September quarters&lt;/td&gt;
 &lt;td&gt;Whether purchase obligations and inventory keep outgrowing sales&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Mid-Nov&lt;/td&gt;
 &lt;td&gt;Nvidia October quarter&lt;/td&gt;
 &lt;td&gt;If supply commitments and excess inventory provisions rise together, the HBM buffer is thinning&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Late Nov&lt;/td&gt;
 &lt;td&gt;Dell third quarter&lt;/td&gt;
 &lt;td&gt;Raw materials inventory and purchase commitments&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Early Dec&lt;/td&gt;
 &lt;td&gt;HPE October quarter&lt;/td&gt;
 &lt;td&gt;Whether inventory and unconditional purchase obligations turn down&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Mid-Dec&lt;/td&gt;
 &lt;td&gt;Micron November quarter&lt;/td&gt;
 &lt;td&gt;Whether customer inventory adjustment comments spread to servers&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Late Dec&lt;/td&gt;
 &lt;td&gt;TrendForce 1Q27 contract price outlook&lt;/td&gt;
 &lt;td&gt;Whether buyer accumulation language turns into inventory adjustment language&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Late Jan 2027&lt;/td&gt;
 &lt;td&gt;Intel fourth quarter&lt;/td&gt;
 &lt;td&gt;Whether the substrate relief moves earlier than 2H27&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;&lt;strong&gt;Is today&amp;rsquo;s buyer DRAM inventory a sensible hedge against real demand, or a 2021-style bullwhip? Tell us which number would change your mind.&lt;/strong&gt;&lt;/p&gt;
&lt;h2 id="key-sources"&gt;Key sources
&lt;/h2&gt;&lt;ul&gt;
&lt;li&gt;&lt;a class="link" href="https://www.trendforce.com/presscenter/news/20211213-11050.html" target="_blank" rel="noopener"
 &gt;TrendForce, PC makers&amp;rsquo; DRAM inventory and 1Q22 price outlook, Dec 13, 2021&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.trendforce.com/presscenter/news/20220217-11128.html" target="_blank" rel="noopener"
 &gt;TrendForce, buyers&amp;rsquo; inventories still flush, Feb 17, 2022&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1571996/000157199626000046/dell-20260731.htm" target="_blank" rel="noopener"
 &gt;Dell 10-Q, quarter ended Jul 31&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/320193/000032019326000020/aapl-20260627.htm" target="_blank" rel="noopener"
 &gt;Apple 10-Q, quarter ended Jun 27&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.trendforce.com/presscenter/news/20260709-13140.html" target="_blank" rel="noopener"
 &gt;TrendForce, US cloud DRAM inventory, Jul 9, 2026&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.trendforce.com/presscenter/news/20260930-13258.html" target="_blank" rel="noopener"
 &gt;TrendForce, 4Q DRAM contract prices and buyer procurement, Sep 30, 2026&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1645590/000164559026000080/hpe-20260731.htm" target="_blank" rel="noopener"
 &gt;HPE 10-Q, quarter ended Jul 31&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1045810/000104581026000075/nvda-20260726.htm" target="_blank" rel="noopener"
 &gt;Nvidia 10-Q, quarter ended Jul 26&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://d1io3yog0oux5.cloudfront.net/_47986fe477a6ac84fe34efbcbf38dd13/intel/db/887/9273/prepared_remarks/Q3%2726_Prepared_Remarks.pdf" target="_blank" rel="noopener"
 &gt;Intel CEO and CFO prepared remarks, Oct 29&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;HP Inc. SEC XBRL data; Lenovo and Xiaomi aggregator data (secondary)&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;For research and information only, not investment advice. Buffer sizes and shipment shares are estimates from a shipments-minus-installations model, not company guidance or trading instructions. Figures are dated. AI tools were used in research and editing.&lt;/p&gt;</description></item><item><title>The Substrate Shortage Just Got Longer. The Substrate Stocks Fell 9-15% Anyway.</title><link>https://koreainvestinsights.com/post/substrate-shortage-longer-stocks-fell-daeduck-readthrough-2026-10-30/</link><pubDate>Fri, 30 Oct 2026 10:35:00 +0900</pubDate><guid>https://koreainvestinsights.com/post/substrate-shortage-longer-stocks-fell-daeduck-readthrough-2026-10-30/</guid><description>&lt;img src="https://koreainvestinsights.com/" alt="Featured image of post The Substrate Shortage Just Got Longer. The Substrate Stocks Fell 9-15% Anyway." /&gt;&lt;p&gt;This week the substrate makers and their biggest customer reported one after another. Samsung Electro-Mechanics (SEMCO) reported on October 28, Ibiden on the afternoon of October 29, and Intel after the US close the same day. They pointed the same way. Advanced chip substrates are a bottleneck for AI and server chip supply, and the shortage runs past 2027.&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Intel:&lt;/strong&gt; The CEO said &amp;ldquo;an industry wide shortage of advanced organic substrates for our EMIB packaging is also impacting our ability to scale output.&amp;rdquo; The CFO expects industry server CPU units to grow more than 30% in both 2026 and 2027, and higher &amp;ldquo;if additional industry supply were available.&amp;rdquo; &lt;a class="link" href="https://d1io3yog0oux5.cloudfront.net/_47986fe477a6ac84fe34efbcbf38dd13/intel/db/887/9273/prepared_remarks/Q3%2726_Prepared_Remarks.pdf" target="_blank" rel="noopener"
 &gt;Intel prepared remarks&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Ibiden:&lt;/strong&gt; It expects high-end substrate demand in 2030 at 3.5 times the 2026 level, well above industry capacity. It raised its fiscal 2027 operating profit guidance by JPY 15bn without changing revenue. &lt;a class="link" href="https://www.ibiden.co.jp/ir/items/kessannsetsumei2026Q2.pdf" target="_blank" rel="noopener"
 &gt;Ibiden results presentation&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;SEMCO:&lt;/strong&gt; Package Solution revenue rose 22% from the prior quarter. Its fourth-quarter outlook says AI data-center customers keep asking for more supply. &lt;a class="link" href="https://www.samsungsem.com/resources/file/kr/ir/earnings_release/3Q26_Earnings_Release_kor.pdf" target="_blank" rel="noopener"
 &gt;SEMCO 3Q IR&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;Then on the morning of October 30, &lt;strong&gt;Ibiden hit its daily limit-down (-15.2%), SEMCO fell 9.4% and Daeduck Electronics fell 13.0%&lt;/strong&gt;. The KOSPI was down 4.7%.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Our view: today&amp;rsquo;s selling is not a verdict on the substrate cycle. The market left 2027 substrate profits alone and cut the multiple it pays for 2028 and beyond. Beta explains most of SEMCO&amp;rsquo;s drop. Daeduck has about 7.5 percentage points that beta does not explain. The signal that reaches Daeduck runs through memory substrate pricing, not the server bottleneck, and today&amp;rsquo;s 3Q release should be read on two things: fourth-quarter pricing and who pays for its new capacity.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;On October 28 we said Daeduck&amp;rsquo;s results should be read on pricing, yield and cash collection. &lt;a class="link" href="https://koreainvestinsights.com/post/korea-substrates-semco-daeduck-2026-10-28/" &gt;KII, Oct 28&lt;/a&gt; Two days later Ibiden and Intel showed the industry benchmark for two of those: pricing and customer money.&lt;/p&gt;
&lt;p&gt;We wrote this at 10:35 a.m. KST on October 30, before Daeduck&amp;rsquo;s 3Q numbers came out. Korean and Japanese prices are intraday at 10:30 a.m. KST on the same day and provisional.&lt;/p&gt;
&lt;h2 id="intels-shock-reads-differently-for-pcs-and-servers"&gt;Intel&amp;rsquo;s shock reads differently for PCs and servers
&lt;/h2&gt;&lt;p&gt;Intel&amp;rsquo;s third-quarter revenue was $16.5bn, up 21% from a year earlier, and non-GAAP EPS was $0.52. Data Center and AI revenue was $7.3bn, up 78%. The problem was the fourth-quarter guide: revenue of $15.0-16.0bn and non-GAAP EPS of $0.25, well below market expectations of about $17.1bn and $0.44. The stock fell 23% after hours. &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/50863/000005086326000188/q326earningsrelease.htm" target="_blank" rel="noopener"
 &gt;Intel 8-K&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;The causes split between PCs and servers.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Business&lt;/th&gt;
 &lt;th&gt;What Intel said&lt;/th&gt;
 &lt;th&gt;What it means for substrates&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;PC&lt;/td&gt;
 &lt;td&gt;PC units fall about 12% in 2026 and about 10% in 2027, with no recovery &amp;ldquo;until memory availability improves and pricing begins to normalize&amp;rdquo;&lt;/td&gt;
 &lt;td&gt;Negative for PC substrate volume&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Server&lt;/td&gt;
 &lt;td&gt;Server CPU units grow more than 30% in both 2026 and 2027, but the EMIB substrate shortage limits output. The industry stays supply-constrained &amp;ldquo;throughout 2027 and into 2028&amp;rdquo;&lt;/td&gt;
 &lt;td&gt;Positive for pricing and utilization of large server substrates&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Intel&amp;rsquo;s 8-K risk factors also cite &amp;ldquo;current industry-wide substrate and memory shortages.&amp;rdquo; On the call, management said substrate suppliers are adding about 40% capacity, mostly in the second half of 2027 (call transcript, secondary source). That makes the first half of 2027 the tightest stretch for server substrates.&lt;/p&gt;
&lt;p&gt;The server side is the one that matters for substrate makers. The world&amp;rsquo;s largest server CPU supplier named substrates as a reason it cannot ship more. The rent from that scarcity goes to the makers with large, high-layer FC-BGA capacity: Ibiden, Shinko, Unimicron and SEMCO. Intel did not name its suppliers, and we have no primary evidence that SEMCO supplies Intel.&lt;/p&gt;
&lt;h2 id="ibiden-raised-profit-without-volume-and-its-customers-pay-for-the-plants"&gt;Ibiden raised profit without volume, and its customers pay for the plants
&lt;/h2&gt;&lt;p&gt;Ibiden&amp;rsquo;s fiscal second quarter (July to September) brought revenue of JPY 139.3bn (+42.1%) and operating profit of JPY 33.2bn (+122.5%). The electronics segment margin rose to 30.2% from 27.6% in the first quarter. Ibiden does not disclose standalone second-quarter figures, so these are first-half totals minus the first quarter. &lt;a class="link" href="https://contents.xj-storage.jp/xcontents/AS00538/659141d4/403c/4aa6/9e22/ba1f0e7d725a/140120261028557691.pdf" target="_blank" rel="noopener"
 &gt;Ibiden 2Q earnings release&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;The shape of the guidance says more than the quarter.&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Fiscal 2027 operating profit guidance went from JPY 150bn to JPY 165bn, with revenue held at JPY 650bn.&lt;/strong&gt; Profit rose JPY 15bn on the same sales. Ibiden did not give a breakdown. We read it as pricing and mix. In its first-quarter Q&amp;amp;A, Ibiden said average substrate prices rose more than expected because of the industry shortage.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Customer advances went from JPY 80.95bn at the end of March to JPY 199.75bn at the end of September.&lt;/strong&gt; That is 2.47 times in six months. Customers are paying before the plants are built.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;A new investment of about JPY 151bn.&lt;/strong&gt; The earnings release describes it as high-performance IC package substrates for CPUs, coming online in stages from fiscal 2029. Nikkei reported that the president said the plan came out of talks with Intel. Ibiden says it will decide on further expansion (Cell 7 and 9) by March 2027, after it settles terms with customers.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;&lt;img alt="Demand and capacity indices from Ibiden’s results presentation. Industry SAP substrate demand goes from 100 in 2026 to 135 in 2027, 170 in 2028, 210 in 2029 and 350 in 2030. Ibiden’s Japan capacity is 150 in 2028 and 200 in 2030. The two indices use different bases, so the gap shows growth rates." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/substrate-shortage-longer-stocks-fell-daeduck-readthrough-2026-10-30/ibiden-demand-vs-capacity.en.png"&gt;
&lt;/p&gt;
&lt;p&gt;Ibiden rose 2.0% on results day, then fell to its daily limit-down of JPY 8,359 on October 30. It had gained 19.5% in the three sessions before the release. Japanese market coverage tied the selling to its exposure to Intel as a major customer after Intel&amp;rsquo;s after-hours plunge. &lt;a class="link" href="https://finance.yahoo.co.jp/news/detail/e9e8d2b14f8241757b9325360eb3137da3153c53" target="_blank" rel="noopener"
 &gt;Traders Web&lt;/a&gt; If that is right, much of Ibiden&amp;rsquo;s excess drop is Intel customer risk. It should not be copied onto Korean makers with little Intel exposure.&lt;/p&gt;
&lt;h2 id="todays-selling-looked-like-a-basket-trade"&gt;Today&amp;rsquo;s selling looked like a basket trade
&lt;/h2&gt;&lt;p&gt;After the US Big Tech reports, AI hardware sold off as a group. Meta issued $25bn of common stock and $12.5bn of mandatory convertible preferred, and plans a $60bn at-the-market share program from the first quarter of 2027. Amazon&amp;rsquo;s third-quarter free cash flow was -$33.5bn on its own definition. The funding risk of AI capex rose in a single day.&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;Taiwan&amp;rsquo;s substrate makers fell too: Unimicron about -6.3%, Kinsus about -4.2% and Nan Ya PCB about -6.6%, or 1.4 to 2.2 times the TAIEX (-3.0%), based on 10:30 a.m. KST quotes.&lt;/li&gt;
&lt;li&gt;In Korea, the selling ran through program trading. A broker&amp;rsquo;s provisional intraday tally at about 10:15 a.m. showed net program selling of KRW 331.4bn across the KOSPI. SK hynix (KRW -201.1bn) and SEMCO (KRW -144.5bn) together exceeded that total. Net program selling was 39% of turnover in SEMCO and 47% in Daeduck.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;Take out the market and the stocks separate. We estimated KOSPI betas from daily returns over 60 trading days to October 28: 1.38 for SEMCO and 1.17 for Daeduck.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Stock&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Beta&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Beta-implied move&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Actual&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Unexplained&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;SEMCO&lt;/td&gt;
 &lt;td style="text-align: right"&gt;1.38&lt;/td&gt;
 &lt;td style="text-align: right"&gt;-6.5%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;-9.4%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;-3.0pp&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Daeduck&lt;/td&gt;
 &lt;td style="text-align: right"&gt;1.17&lt;/td&gt;
 &lt;td style="text-align: right"&gt;-5.5%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;-13.0%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;-7.5pp&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;&lt;img alt="Intraday moves at 10:30 a.m. KST on October 30: KOSPI -4.7%, SK hynix -8.2%, SEMCO -9.4%, Daeduck -13.0%, Ibiden -15.2% (limit-down). Beta-implied moves are -6.5% for SEMCO and -5.5% for Daeduck, leaving -3.0pp and -7.5pp unexplained." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/substrate-shortage-longer-stocks-fell-daeduck-readthrough-2026-10-30/same-day-moves.en.png"&gt;
&lt;/p&gt;
&lt;p&gt;Both stocks have a typical daily residual (standard deviation) of about 4pp. SEMCO&amp;rsquo;s -3.0pp sits inside that. Daeduck&amp;rsquo;s -7.5pp is about 1.8 times it. The betas have an R-squared of 0.5 to 0.6, so these gaps carry estimation error.&lt;/p&gt;
&lt;h2 id="semco-the-most-direct-read-through-and-a-market-sized-drop"&gt;SEMCO: the most direct read-through, and a market-sized drop
&lt;/h2&gt;&lt;p&gt;SEMCO&amp;rsquo;s third-quarter revenue was KRW 3.7633tn (+30%) and operating profit was KRW 691.0bn, an 18.4% margin. Package Solution revenue was KRW 939.0bn, up 58% from a year earlier and 22% from the prior quarter. That matches the sequential growth of Ibiden&amp;rsquo;s electronics segment (+22.6%). SEMCO named server CPU, AI accelerator and networking supply as the FC-BGA growth drivers.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Variable&lt;/th&gt;
 &lt;th&gt;This week&amp;rsquo;s signal&lt;/th&gt;
 &lt;th&gt;Direction for SEMCO&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Server and AI substrate supply&lt;/td&gt;
 &lt;td&gt;Intel&amp;rsquo;s &amp;ldquo;industry wide shortage,&amp;rdquo; Ibiden&amp;rsquo;s 2030 demand view, SEMCO&amp;rsquo;s own fourth-quarter &amp;ldquo;customers keep asking for more supply&amp;rdquo;&lt;/td&gt;
 &lt;td&gt;Positive, the largest variable&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Pricing&lt;/td&gt;
 &lt;td&gt;Ibiden&amp;rsquo;s fiscal 2027 profit raised on unchanged revenue&lt;/td&gt;
 &lt;td&gt;Positive, SEMCO does not disclose package margin&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;PC&lt;/td&gt;
 &lt;td&gt;Intel&amp;rsquo;s PC unit declines in 2026 and 2027&lt;/td&gt;
 &lt;td&gt;Mildly negative, PC share undisclosed&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;2028 supply&lt;/td&gt;
 &lt;td&gt;Intel suppliers add capacity in 2H27, Ibiden&amp;rsquo;s new capacity in fiscal 2029, SEMCO&amp;rsquo;s Sejong plant in 2H28&lt;/td&gt;
 &lt;td&gt;Multiple discount&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;SEMCO does not need to supply Intel to benefit. Intel&amp;rsquo;s shortage adds one more buyer competing with AMD and hyperscaler ASIC customers for the same capacity. SEMCO&amp;rsquo;s drop today matched Taiwan&amp;rsquo;s substrate stocks, and beta explains most of it. On this week&amp;rsquo;s information, SEMCO&amp;rsquo;s substrate case got stronger and only its price got weaker.&lt;/p&gt;
&lt;h2 id="daeduck-the-signal-comes-through-memory-substrate-pricing"&gt;Daeduck: the signal comes through memory substrate pricing
&lt;/h2&gt;&lt;p&gt;Daeduck&amp;rsquo;s second-quarter revenue was KRW 401.0bn and operating profit was KRW 70.3bn (17.5%). Revenue split into FCCSP 43%, FCBGA 23%, CSP 20% and MLB 14%. Its own materials list FCBGA uses as autonomous driving, optical modules, networking and CXL controllers. That is a different product tier from the EMIB substrates for server CPUs that Intel described. &lt;a class="link" href="https://www.daeduck.com/file/board/830.do?isInline=true" target="_blank" rel="noopener"
 &gt;Daeduck 2Q IR&lt;/a&gt;&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;This week&amp;rsquo;s signal&lt;/th&gt;
 &lt;th&gt;Path to Daeduck&lt;/th&gt;
 &lt;th&gt;Direction&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Intel&amp;rsquo;s EMIB substrate shortage&lt;/td&gt;
 &lt;td&gt;No direct path&lt;/td&gt;
 &lt;td&gt;Small&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Top makers running full&lt;/td&gt;
 &lt;td&gt;If top makers fill up with CPU and GPU substrates, networking and optics customers may move to second-tier makers&lt;/td&gt;
 &lt;td&gt;Mildly positive, still a hypothesis&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Memory shortage persists&lt;/td&gt;
 &lt;td&gt;Intel says PCs will not recover before memory normalizes; if memory makers run flat out on constrained supply, memory substrate volume and pricing hold&lt;/td&gt;
 &lt;td&gt;Positive&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Fewer PCs&lt;/td&gt;
 &lt;td&gt;While memory supply is tight, chips that PCs give up go to server DDR5 and SOCAMM&lt;/td&gt;
 &lt;td&gt;Slightly negative&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Ibiden&amp;rsquo;s pricing&lt;/td&gt;
 &lt;td&gt;Higher industry average prices give second-tier makers bargaining power too&lt;/td&gt;
 &lt;td&gt;Positive, size unknown&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;A broker estimate puts memory substrates at about 44% of Daeduck&amp;rsquo;s second-quarter revenue (Yuanta Securities, secondary source). So the Intel signal that reaches Daeduck is delayed memory normalization, not the server bottleneck. Memory substrates are a pricing business more than a volume business. SEMCO&amp;rsquo;s October 19 filing of a KRW 360.5bn package substrate supply contract sends the same signal. The contract covers calendar 2027. The counterparty is disclosed only as a &amp;ldquo;global large company,&amp;rdquo; and there is no advance payment. &lt;a class="link" href="https://dart.fss.or.kr/dsaf001/main.do?rcpNo=20261019800610" target="_blank" rel="noopener"
 &gt;SEMCO contract filing&lt;/a&gt; Supply is tight, but customers are not paying for capacity in that market.&lt;/p&gt;
&lt;p&gt;This is where Ibiden and Daeduck differ in numbers.&lt;/p&gt;
&lt;p&gt;&lt;img alt="Customer advances or contract liabilities as a share of announced capacity investment. Ibiden: JPY 199.75bn of advances against JPY 651bn of announced expansion, 30.7%. Daeduck: KRW 72.3bn of contract liabilities against KRW 1.325tn of expansion announced in 2026, 5.5%. SEMCO has said customers support its plan but has not disclosed an amount." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/substrate-shortage-longer-stocks-fell-daeduck-readthrough-2026-10-30/who-funds-capacity.en.png"&gt;
&lt;/p&gt;
&lt;p&gt;Daeduck&amp;rsquo;s 2026 capacity filings were KRW 213bn in May, KRW 497bn in July and KRW 615bn in October, KRW 1.325tn in total. Its contract liabilities at the end of June were KRW 72.3bn, about 5.5% of that. The October filing did not name the product or site, and listed internal funds and external borrowing as funding sources. &lt;a class="link" href="https://dart.fss.or.kr/dsaf001/main.do?rcpNo=20261022800677" target="_blank" rel="noopener"
 &gt;Daeduck capex filing&lt;/a&gt; Customer money covers about 31% of Ibiden&amp;rsquo;s expansion. Daeduck builds mostly with its own money. On a day when the market reprices AI capex funding risk, that difference can show up as a discount.&lt;/p&gt;
&lt;h2 id="how-to-read-daeducks--75pp"&gt;How to read Daeduck&amp;rsquo;s -7.5pp
&lt;/h2&gt;&lt;p&gt;On information alone, Daeduck should have fallen less than SEMCO, because its Intel and PC exposure is smaller. It fell more. The one confirmed difference between the two is the earnings calendar. SEMCO reported on October 28, and Daeduck reports today. We read the gap as de-risking before the print. Holders who watched Ibiden hit limit-down the day after a strong report cut exposure before seeing Daeduck&amp;rsquo;s numbers. A discount for self-funded capex is a secondary hypothesis. We did not check trading by broker, so both remain hypotheses.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;What would change our read:&lt;/strong&gt; if third-quarter operating profit comes in clearly below consensus and the margin falls well below the second quarter&amp;rsquo;s 17.5%, we would treat today&amp;rsquo;s gap as selling that front-ran the numbers.&lt;/p&gt;
&lt;h2 id="what-to-look-for-in-todays-3q-release"&gt;What to look for in today&amp;rsquo;s 3Q release
&lt;/h2&gt;&lt;p&gt;Daeduck said it would post its 3Q IR materials on its website on October 30. Its conference call is at 9:00 a.m. KST on November 4. &lt;a class="link" href="https://dart.fss.or.kr/dsaf001/main.do?rcpNo=20261015800468" target="_blank" rel="noopener"
 &gt;IR schedule filing&lt;/a&gt;&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Item&lt;/th&gt;
 &lt;th&gt;Reference&lt;/th&gt;
 &lt;th&gt;How to read it&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;3Q operating profit&lt;/td&gt;
 &lt;td&gt;Naver consensus KRW 77.6bn; selected broker reports KRW 80.4-82.2bn&lt;/td&gt;
 &lt;td&gt;A number inside that range may meet Ibiden-style &amp;ldquo;sell the good news&amp;rdquo;&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Operating margin&lt;/td&gt;
 &lt;td&gt;2Q 17.5%&lt;/td&gt;
 &lt;td&gt;Whether pricing ran ahead of cost increases&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;4Q pricing&lt;/td&gt;
 &lt;td&gt;When memory substrate price increases apply&lt;/td&gt;
 &lt;td&gt;If they hit fourth-quarter shipments, 2027 estimates go up&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Large FC-BGA&lt;/td&gt;
 &lt;td&gt;Whether large networking and optics substrates are in mass production&lt;/td&gt;
 &lt;td&gt;A server-CPU-class customer would turn the spillover hypothesis into fact&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;KRW 615bn expansion&lt;/td&gt;
 &lt;td&gt;Product, site and customer undisclosed&lt;/td&gt;
 &lt;td&gt;Customer prepayments or long-term supply agreements would shrink the self-funding discount; check on the November 4 call&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Ibiden hit limit-down the day after operating profit more than doubled and guidance went up. A KRW 80bn-range quarter alone is unlikely to move Daeduck&amp;rsquo;s stock. The reaction will turn on when fourth-quarter pricing reaches shipments, and whether customer money backs the expansion.&lt;/p&gt;
&lt;h2 id="the-market-is-cutting-2028"&gt;The market is cutting 2028
&lt;/h2&gt;&lt;p&gt;The candidates for what the market cut today are in the table below.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Variable&lt;/th&gt;
 &lt;th&gt;Evidence&lt;/th&gt;
 &lt;th&gt;Weight in today&amp;rsquo;s selling&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;2028 supply wave&lt;/td&gt;
 &lt;td&gt;Intel suppliers add about 40% (2H27), Ibiden&amp;rsquo;s new capacity (fiscal 2029), SEMCO Sejong (2H28)&lt;/td&gt;
 &lt;td&gt;Not new today&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;AI capex funding risk&lt;/td&gt;
 &lt;td&gt;Meta&amp;rsquo;s equity raise and ATM plan, Amazon&amp;rsquo;s negative free cash flow&lt;/td&gt;
 &lt;td&gt;Rose today; memory, Taiwan and Japan AI hardware fell together&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;&lt;strong&gt;We see funding risk as the main driver of today&amp;rsquo;s selling.&lt;/strong&gt; The supply wave is a known schedule. What changed today is the funding conditions of the customers. Substrate profits are solid through 2027. The multiple depends on whether customers can keep paying after 2028.&lt;/p&gt;
&lt;p&gt;The evidence that would change this view is specific. If Ibiden&amp;rsquo;s customer advances keep rising next quarter and hyperscalers hold their 2027 capex plans after raising equity, today&amp;rsquo;s discount was too deep. If Ibiden pushes the Cell 7 and 9 decision past March 2027 because it cannot settle customer terms, or if advances fall, the market was right.&lt;/p&gt;
&lt;h2 id="what-to-watch-next"&gt;What to watch next
&lt;/h2&gt;&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;When&lt;/th&gt;
 &lt;th&gt;What&lt;/th&gt;
 &lt;th&gt;How to read it&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Oct 30&lt;/td&gt;
 &lt;td&gt;Daeduck 3Q preliminary results and IR materials&lt;/td&gt;
 &lt;td&gt;Operating profit, margin, memory, non-memory and MLB split, 4Q pricing&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Nov 4-6&lt;/td&gt;
 &lt;td&gt;Daeduck call and NDR&lt;/td&gt;
 &lt;td&gt;Use of the KRW 615bn, customer prepayments and LTAs, server-CPU-class customers&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Early Nov&lt;/td&gt;
 &lt;td&gt;Ibiden 2Q Q&amp;amp;A summary&lt;/td&gt;
 &lt;td&gt;Quarterly price talks, CPU customers, Cell 7 and 9 terms&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Early Nov&lt;/td&gt;
 &lt;td&gt;AMD 3Q&lt;/td&gt;
 &lt;td&gt;Server CPU supply constraints and substrate comments&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Mid-Nov&lt;/td&gt;
 &lt;td&gt;Daeduck 3Q quarterly report&lt;/td&gt;
 &lt;td&gt;Change in contract liabilities (advances)&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;By Mar 2027&lt;/td&gt;
 &lt;td&gt;Ibiden Cell 7 and 9 decision&lt;/td&gt;
 &lt;td&gt;Customer terms and size of advances&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;&lt;strong&gt;Was today&amp;rsquo;s substrate selling about 2028 supply or about AI capex funding? Tell us which number in Daeduck&amp;rsquo;s release would convince you today&amp;rsquo;s drop went too far.&lt;/strong&gt;&lt;/p&gt;
&lt;h2 id="key-sources"&gt;Key sources
&lt;/h2&gt;&lt;ul&gt;
&lt;li&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/50863/000005086326000188/q326earningsrelease.htm" target="_blank" rel="noopener"
 &gt;Intel 8-K, third-quarter results, Oct 29&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://d1io3yog0oux5.cloudfront.net/_47986fe477a6ac84fe34efbcbf38dd13/intel/db/887/9273/prepared_remarks/Q3%2726_Prepared_Remarks.pdf" target="_blank" rel="noopener"
 &gt;Intel CEO and CFO prepared remarks, Oct 29&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://contents.xj-storage.jp/xcontents/AS00538/659141d4/403c/4aa6/9e22/ba1f0e7d725a/140120261028557691.pdf" target="_blank" rel="noopener"
 &gt;Ibiden fiscal 2026 2Q earnings release (kessan tanshin), Oct 29&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://contents.xj-storage.jp/xcontents/AS00538/674b875f/9878/4a79/916d/773627c03422/140120260731504718.pdf" target="_blank" rel="noopener"
 &gt;Ibiden fiscal 2026 1Q earnings release&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.ibiden.co.jp/ir/items/kessannsetsumei2026Q2.pdf" target="_blank" rel="noopener"
 &gt;Ibiden 2Q results presentation&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.ibiden.co.jp/ir/items/QA_FY2026Q1.pdf" target="_blank" rel="noopener"
 &gt;Ibiden 1Q Q&amp;amp;A summary&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.nikkei.com/article/DGXZQOFD290UF0Z21C26A0000000/" target="_blank" rel="noopener"
 &gt;Nikkei, Ibiden president&amp;rsquo;s briefing, Oct 29&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://finance.yahoo.co.jp/news/detail/e9e8d2b14f8241757b9325360eb3137da3153c53" target="_blank" rel="noopener"
 &gt;Traders Web via Yahoo! Finance Japan, Oct 30&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.samsungsem.com/resources/file/kr/ir/earnings_release/3Q26_Earnings_Release_kor.pdf" target="_blank" rel="noopener"
 &gt;SEMCO 3Q results IR materials&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://dart.fss.or.kr/dsaf001/main.do?rcpNo=20261019800610" target="_blank" rel="noopener"
 &gt;SEMCO package substrate supply contract filing, Oct 19&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.daeduck.com/file/board/830.do?isInline=true" target="_blank" rel="noopener"
 &gt;Daeduck 2Q IR materials&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://dart.fss.or.kr/dsaf001/main.do?rcpNo=20261022800677" target="_blank" rel="noopener"
 &gt;Daeduck capex filing, Oct 22&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://dart.fss.or.kr/dsaf001/main.do?rcpNo=20260813001620" target="_blank" rel="noopener"
 &gt;Daeduck half-year report (contract liabilities)&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://dart.fss.or.kr/dsaf001/main.do?rcpNo=20261015800468" target="_blank" rel="noopener"
 &gt;Daeduck IR schedule filing, Oct 15&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://kabutan.jp/stock/kabuka?code=4062" target="_blank" rel="noopener"
 &gt;Kabutan, Ibiden daily prices&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;Naver Finance prices and consensus, Taiwan Stock Exchange real-time quotes, intraday Oct 30&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;For research and information only, not investment advice. The beta split and cause attribution are analytical assumptions, not company guidance, trading instructions or personal portfolio decisions. Figures are dated, and Korean, Japanese and Taiwanese prices are intraday and provisional. AI tools were used in research and editing.&lt;/p&gt;</description></item><item><title>The First-Day Verdict Got Marvell's Last Two Big Deals Backwards. Lumentum Is the One Its Playbook Does Not Cover.</title><link>https://koreainvestinsights.com/post/marvell-acquisition-playbook-lumentum-outside-it-2026-10-30/</link><pubDate>Fri, 30 Oct 2026 10:28:00 +0900</pubDate><guid>https://koreainvestinsights.com/post/marvell-acquisition-playbook-lumentum-outside-it-2026-10-30/</guid><description>&lt;img src="https://koreainvestinsights.com/" alt="Featured image of post The First-Day Verdict Got Marvell's Last Two Big Deals Backwards. Lumentum Is the One Its Playbook Does Not Cover." /&gt;&lt;p&gt;On Thursday, October 29, Marvell Technology closed at &lt;strong&gt;$218.34, down 13.2%&lt;/strong&gt;. The Philadelphia semiconductor index (SOX) fell 1.4%. The 11.8-point gap is the clearest measure of the market&amp;rsquo;s verdict on Bloomberg&amp;rsquo;s report that Marvell is in advanced talks to buy Lumentum, the laser and optics maker. No price or payment mix has been reported. Not all of the gap is about the deal: Alphabet fell 12.6% the same day after its results, and Marvell designs custom chips for Google. But Broadcom, Google&amp;rsquo;s larger chip partner, fell only 1.9%. By our measure, it is the worst first reaction to any Marvell acquisition in Matt Murphy&amp;rsquo;s ten years as chief executive.&lt;/p&gt;
&lt;p&gt;History says not to trust that verdict too quickly. In 2017 the market cheered the Cavium deal. From the first press report to the announcement day, Marvell beat the SOX by 14.9 points. One year later it trailed the index by 15.1 points. In 2020 the market disliked the Inphi deal. Over the report and announcement days, Marvell lagged by 7.1 points. One year later it led by 18.7 points. &lt;strong&gt;Both times, the first reaction and the one-year result pointed in opposite directions.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Our judgment: Marvell&amp;rsquo;s acquisition skill is real, but it was proven in a narrow zone, and Lumentum sits outside that zone.&lt;/strong&gt; Marvell knows how to buy fabless technology, take out cost, sell non-core units at high multiples and keep its investment-grade rating. That skill turned Avera and Inphi into the seeds of a data center business that grew from $1.04 billion to $6.10 billion in five fiscal years. Lumentum is a different kind of target. It runs its own laser wafer fabs. It is a customer of Marvell&amp;rsquo;s optical signal processors. And Marvell would buy it at what may be the peak of a laser shortage. The track record lowers the integration risk. It does not lower the price risk. &lt;strong&gt;In our view, the variable that dominates is whether lasers stay scarce after 2028.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;This article is dated October 30, 2026, Korea time. The deal is reported, not announced. As of this morning, neither company has filed an 8-K about it on SEC EDGAR, and neither has confirmed the talks. Prices are Nasdaq closes for October 29. Every Lumentum deal number below is our scenario. Our &lt;a class="link" href="https://koreainvestinsights.com/post/marvell-lumentum-talks-laser-scarcity-cpo-stock-deal-2026-10-29/" target="_blank" rel="noopener"
 &gt;first take on the reported talks&lt;/a&gt; covers the laser logic and the dilution math.&lt;/p&gt;
&lt;p&gt;&lt;img alt="Marvell’s return minus the SOX after each acquisition. Cavium: +14.9 points at the first reaction, -15.1 points one year after the announcement. Inphi: -7.1 points at the first reaction, +18.7 points one year later. The small deals moved within one point of the index at first. The Lumentum report: -11.8 points on October 29, 2026." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/marvell-acquisition-playbook-lumentum-outside-it-2026-10-30/first-reaction-vs-one-year.en.png"&gt;
&lt;/p&gt;
&lt;h2 id="what-to-take-away"&gt;What to take away
&lt;/h2&gt;&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Thursday was the largest negative first reaction in Marvell&amp;rsquo;s deal history.&lt;/strong&gt; Against the SOX, Marvell lost about $25 billion of value that the index does not explain, though part of it may be Alphabet&amp;rsquo;s results. Lumentum gained only about $2 billion. In an all-stock deal with a fixed exchange ratio, every dollar Marvell loses also comes out of the price Lumentum holders would receive.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Two deals built Marvell&amp;rsquo;s AI business.&lt;/strong&gt; Marvell&amp;rsquo;s latest annual report names custom products and electro-optics as its growth drivers. Custom silicon started with Avera, the ASIC unit Marvell bought from GlobalFoundries. Electro-optics started with Inphi.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Marvell&amp;rsquo;s own precedents rule out a large cash share.&lt;/strong&gt; If Marvell paid for Lumentum with the cash share it used for Cavium (47.5%), gross debt would reach about 11.8 times combined trailing EBITDA. Its new credit facility allows 4.5 times after a large deal, on our reading of the limit.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;The share overhang would dwarf Inphi&amp;rsquo;s.&lt;/strong&gt; The stock part of the Inphi price was worth about $5.9 billion. If an all-stock exchange ratio were set on Marvell&amp;rsquo;s October 28 price, the new shares would be worth about $100 billion at Thursday&amp;rsquo;s close. After the Inphi announcement, Marvell trailed the SOX by 8.2 points within five trading days and took about three months to get back to even.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;The price is the unproven part.&lt;/strong&gt; At a 20% premium, Marvell would pay about 23 times Lumentum&amp;rsquo;s guided quarterly revenue run rate. It paid about 14 times for Inphi, and Inphi worked. The question is whether 1.7 times Inphi&amp;rsquo;s multiple can work too.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id="the-playbook-deal-by-deal"&gt;The playbook, deal by deal
&lt;/h2&gt;&lt;p&gt;Since Murphy took over in July 2016, Marvell has bought companies at a steady pace. It used cash and stock in different mixes, and it sold businesses it did not need to pay for what it did.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Deal&lt;/th&gt;
 &lt;th&gt;Announced to closed&lt;/th&gt;
 &lt;th&gt;Announced price&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Cash share&lt;/th&gt;
 &lt;th&gt;Funding&lt;/th&gt;
 &lt;th&gt;China review&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Cavium (processors, security)&lt;/td&gt;
 &lt;td&gt;Nov 2017 to Jul 2018&lt;/td&gt;
 &lt;td&gt;about $6bn&lt;/td&gt;
 &lt;td style="text-align: right"&gt;47.5%&lt;/td&gt;
 &lt;td&gt;$0.9bn term loan and $1bn of notes; share count up about 31%&lt;/td&gt;
 &lt;td&gt;SAMR cleared after 221 days&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Aquantia (Ethernet PHYs)&lt;/td&gt;
 &lt;td&gt;May to Sep 2019&lt;/td&gt;
 &lt;td&gt;about $0.45bn&lt;/td&gt;
 &lt;td style="text-align: right"&gt;100%&lt;/td&gt;
 &lt;td&gt;Cash on hand and revolver&lt;/td&gt;
 &lt;td&gt;Not a closing condition&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Avera (custom ASICs)&lt;/td&gt;
 &lt;td&gt;May to Nov 2019&lt;/td&gt;
 &lt;td&gt;$0.65bn plus up to $0.09bn&lt;/td&gt;
 &lt;td style="text-align: right"&gt;100%&lt;/td&gt;
 &lt;td&gt;Bridge loan, repaid with the NXP sale proceeds&lt;/td&gt;
 &lt;td&gt;Not found&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Inphi (electro-optical DSPs)&lt;/td&gt;
 &lt;td&gt;Oct 2020 to Apr 2021&lt;/td&gt;
 &lt;td&gt;about $10bn&lt;/td&gt;
 &lt;td style="text-align: right"&gt;40.7%&lt;/td&gt;
 &lt;td&gt;$1.75bn term loans and $2bn of notes; share count up about 20%&lt;/td&gt;
 &lt;td&gt;SAMR cleared after 145 days&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Innovium (switches)&lt;/td&gt;
 &lt;td&gt;Aug to Oct 2021&lt;/td&gt;
 &lt;td&gt;about $1.1bn&lt;/td&gt;
 &lt;td style="text-align: right"&gt;0%&lt;/td&gt;
 &lt;td&gt;All stock&lt;/td&gt;
 &lt;td&gt;Not mentioned&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Celestial AI (photonic interconnect)&lt;/td&gt;
 &lt;td&gt;Dec 2025 to Feb 2026&lt;/td&gt;
 &lt;td&gt;about $3.25bn upfront plus up to $2.25bn earnout&lt;/td&gt;
 &lt;td style="text-align: right"&gt;about 31%&lt;/td&gt;
 &lt;td&gt;About $1bn cash and about 27 million shares&lt;/td&gt;
 &lt;td&gt;Not found&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;XConn (CXL and PCIe switches)&lt;/td&gt;
 &lt;td&gt;Closed Feb 2026&lt;/td&gt;
 &lt;td&gt;about $0.47bn&lt;/td&gt;
 &lt;td style="text-align: right"&gt;about 60%&lt;/td&gt;
 &lt;td&gt;About $0.28bn cash and about 2.1 million shares&lt;/td&gt;
 &lt;td&gt;Not found&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;&lt;strong&gt;Lumentum (our assumption)&lt;/strong&gt;&lt;/td&gt;
 &lt;td&gt;&lt;strong&gt;Reported&lt;/strong&gt;&lt;/td&gt;
 &lt;td&gt;&lt;strong&gt;about $115bn at a 20% premium&lt;/strong&gt;&lt;/td&gt;
 &lt;td style="text-align: right"&gt;&lt;strong&gt;0–15%&lt;/strong&gt;&lt;/td&gt;
 &lt;td&gt;&lt;strong&gt;See below&lt;/strong&gt;&lt;/td&gt;
 &lt;td&gt;&lt;strong&gt;Pending&lt;/strong&gt;&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Price is the value announced with each deal; for XConn it is the consideration at closing. Accounting purchase prices in the 10-K reports differ slightly. Cash share is the cash part of the per-share price at Marvell&amp;rsquo;s closing price before the news. Sources: Marvell&amp;rsquo;s 8-K filings, proxy statements and 10-K reports on &lt;a class="link" href="https://www.sec.gov/cgi-bin/browse-edgar?action=getcompany&amp;amp;CIK=0001835632&amp;amp;type=8-K" target="_blank" rel="noopener"
 &gt;SEC EDGAR&lt;/a&gt;. The Cavium and Inphi clearances are from Marvell&amp;rsquo;s own announcements of China&amp;rsquo;s State Administration for Market Regulation (SAMR) approval.&lt;/p&gt;
&lt;p&gt;The order of the deals tells the strategy. Marvell filled the empty boxes of a data center connectivity stack one at a time: processors and security, custom ASICs, electro-optical signal processors, switches, memory-expansion switches and finally photonic interconnect. The laser is the last empty box in the optical stack. On strategy, Lumentum fits the pattern exactly.&lt;/p&gt;
&lt;h2 id="what-the-playbook-produced"&gt;What the playbook produced
&lt;/h2&gt;&lt;p&gt;&lt;strong&gt;A growth engine.&lt;/strong&gt; Marvell&amp;rsquo;s data center revenue was $1.04 billion in fiscal 2021, 35% of sales. In fiscal 2026 it was $6.10 billion, 74% of sales. Marvell does not report revenue by acquired business, so the credit cannot be split exactly. But the two lines its 10-K names as growth drivers, custom silicon and electro-optics, both began as acquisitions.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Cost discipline.&lt;/strong&gt; The Cavium deal came with a target of $150–175 million of annual cost savings. By the March 2019 earnings call, management was talking about $200 million. For Inphi, Marvell set a $125 million run-rate target within 18 months of closing. We could not find a primary-source report of either target&amp;rsquo;s final delivery.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Expensive exits.&lt;/strong&gt; Marvell sold its Wi-Fi and Bluetooth business to NXP in 2019 for $1.76 billion in cash and used the money to repay the Avera bridge loan. In 2025 it sold its automotive Ethernet business, which included Aquantia, to Infineon for $2.5 billion in cash, about 10 to 11 times that unit&amp;rsquo;s revenue by our estimate.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;A protected balance sheet.&lt;/strong&gt; Marvell kept its investment-grade rating through both large deals. Its debt was about $4.6 billion at the end of fiscal 2022 and is $5.0 billion today, while revenue roughly doubled. Its 10-K reports have recorded no goodwill impairment from fiscal 2019 through fiscal 2026. The clear write-down was smaller: in fiscal 2021 Marvell cut back its server processor line, a Cavium legacy by our reading, and booked $119 million of charges.&lt;/p&gt;
&lt;h2 id="where-the-playbook-ends"&gt;Where the playbook ends
&lt;/h2&gt;&lt;p&gt;We split &amp;ldquo;M&amp;amp;A know-how&amp;rdquo; into the variables a deal actually tests. Here is how the record maps onto Lumentum.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Variable&lt;/th&gt;
 &lt;th&gt;Marvell&amp;rsquo;s record&lt;/th&gt;
 &lt;th&gt;Verdict&lt;/th&gt;
 &lt;th&gt;What Lumentum tests&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Picking the growth axis&lt;/td&gt;
 &lt;td&gt;Avera and Inphi seeded a 5.9-times larger data center business&lt;/td&gt;
 &lt;td&gt;Proven&lt;/td&gt;
 &lt;td&gt;Same logic: the laser is the optical bottleneck&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Cost take-out&lt;/td&gt;
 &lt;td&gt;Cavium target raised to $200m; Inphi run-rate plan published&lt;/td&gt;
 &lt;td&gt;Partly proven; final delivery not confirmed&lt;/td&gt;
 &lt;td&gt;Small relative to the price&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Funding and rating&lt;/td&gt;
 &lt;td&gt;Investment grade kept through two large deals; asset sales repaid debt&lt;/td&gt;
 &lt;td&gt;Proven&lt;/td&gt;
 &lt;td&gt;Untested if the deal is all stock&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;China clearance&lt;/td&gt;
 &lt;td&gt;SAMR approved in 145–221 days in 2018–2021&lt;/td&gt;
 &lt;td&gt;Proven, in an older climate&lt;/td&gt;
 &lt;td&gt;42% of Marvell&amp;rsquo;s latest-quarter shipments went to China&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Running fabs&lt;/td&gt;
 &lt;td&gt;Marvell has stayed fabless&lt;/td&gt;
 &lt;td&gt;No record&lt;/td&gt;
 &lt;td&gt;Lumentum runs indium phosphide (InP) laser fabs and module plants&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Customer conflict&lt;/td&gt;
 &lt;td&gt;No precedent&lt;/td&gt;
 &lt;td&gt;No record&lt;/td&gt;
 &lt;td&gt;Lumentum buys Marvell DSPs; module makers that buy Lumentum lasers also buy Marvell DSPs&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Price&lt;/td&gt;
 &lt;td&gt;Inphi: about 14 times run-rate revenue and a 43% premium, and it worked&lt;/td&gt;
 &lt;td&gt;One expensive success&lt;/td&gt;
 &lt;td&gt;About 23 times, 1.7 times the Inphi multiple, at our 20% premium&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Relative size&lt;/td&gt;
 &lt;td&gt;Cavium was about 61% of Marvell&amp;rsquo;s market value&lt;/td&gt;
 &lt;td&gt;Proven&lt;/td&gt;
 &lt;td&gt;About 51%, inside the precedent&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;The size is not the problem. Marvell has already absorbed a target that was larger relative to itself. The problem is that the variables most likely to decide this deal are ones Marvell has never had to manage: what it pays against the length of the laser shortage, how it runs a fab, and how it keeps customers who now compete with its own subsidiary. A Stifel note cited in trade coverage puts Marvell&amp;rsquo;s share of PAM4 optical DSPs at 60–65%. If even part of that customer base sees Marvell as a rival supplier, the dis-synergy could rival the synergy. &lt;a class="link" href="https://www.tradingview.com/news/seekingalpha:28193eb8c094b:0-marvell-lumentum-deal-analysts-see-strategic-logic-but-flag-price-integration-risk/" target="_blank" rel="noopener"
 &gt;TradingView / Seeking Alpha, Oct 29&lt;/a&gt;&lt;/p&gt;
&lt;h2 id="the-funding-math-marvells-own-precedents-rule-out-a-large-cash-share"&gt;The funding math: Marvell&amp;rsquo;s own precedents rule out a large cash share
&lt;/h2&gt;&lt;p&gt;Our reading of the record is that Marvell has used the cheapest currency available. In 2018 and 2021, debt was cheap: its Cavium-era notes paid about 4–5%, and its Inphi-era notes paid about 1.7–3%. So it mixed in 41–48% cash. Today the arithmetic is reversed. Marvell&amp;rsquo;s stock trades at a high multiple of expected earnings, so its earnings yield is below the after-tax cost of new debt. On that test, stock is now the cheaper currency. A mostly-stock deal would follow the same rule Marvell has always used, with a different currency.&lt;/p&gt;
&lt;p&gt;The credit agreement settles the rest. On October 22, Marvell signed a new $3 billion revolving credit facility. Its leverage limit is 4.0 times EBITDA and rises to 4.5 times for the quarter in which Marvell closes an acquisition of at least $500 million and the following three quarters. &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1835632/000119312526430424/d147430d8k.htm" target="_blank" rel="noopener"
 &gt;Marvell 8-K, Oct 23&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;&lt;img alt="Gross debt divided by combined trailing EBITDA under each precedent cash mix, at a 20% premium. Cavium mix (47.5% cash): 11.8 times. Inphi mix (40.7%): 10.1 times. Celestial AI mix (31%): 7.7 times. 85% stock: 3.8 times. 90% stock: 2.5 times. All stock: 1.1 times. The revolver limit after a large deal is 4.5 times." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/marvell-acquisition-playbook-lumentum-outside-it-2026-10-30/leverage-by-cash-mix.en.png"&gt;
&lt;/p&gt;
&lt;p&gt;On our assumptions, any precedent mix with real cash would exceed the 4.5-times limit. We have not matched the credit agreement&amp;rsquo;s exact EBITDA definition, so treat the bars as estimates. The usable cash in the chart combines Marvell&amp;rsquo;s $3.9 billion and Lumentum&amp;rsquo;s balance, less operating minimums. A deal with 15% cash stays under the limit but sits above the roughly 3 times that investment-grade issuers usually keep. That leaves two realistic structures: &lt;strong&gt;all stock, or about 90% stock with a small fixed cash amount per share.&lt;/strong&gt; The cash version gives Lumentum holders a more certain value, which helps in a premium negotiation. The all-stock version leaves Marvell&amp;rsquo;s balance sheet untouched and puts the full weight of the deal on its share price.&lt;/p&gt;
&lt;h2 id="the-all-stock-trap"&gt;The all-stock trap
&lt;/h2&gt;&lt;p&gt;An all-stock deal has a cost that does not show up in leverage ratios.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;The signal.&lt;/strong&gt; On October 6, Marvell&amp;rsquo;s management told investors it expects non-GAAP earnings per share above $30 in fiscal 2031. At Marvell&amp;rsquo;s price before the report, that target valued the stock at about 8.4 times fiscal 2031 earnings.&lt;/p&gt;
&lt;p&gt;A management team that believes its own target sees its shares as cheap, and a cheap currency is a strange thing to spend. When a buyer pays in stock anyway, investors start to doubt either the long-term target or the price of the target company. Thursday&amp;rsquo;s drop is consistent with that reading. It is also consistent with other explanations, such as customer conflict and lower margins from owning fabs, and we cannot separate them.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;The overhang.&lt;/strong&gt; In a fixed-ratio stock deal, merger-arbitrage funds buy the target and short the buyer until the deal closes. With Inphi, the stock part of the price was about $5.9 billion, and Marvell still lagged the SOX for about three months. Here the stock part could be close to $100 billion.&lt;/p&gt;
&lt;p&gt;Marvell&amp;rsquo;s past large deals took 173 to 228 days to close, with SAMR as the last gate. Lumentum&amp;rsquo;s own deals for Oclaro and NeoPhotonics each took about nine months. In today&amp;rsquo;s US–China climate, we assume nine to twelve months. A November announcement would close in the second half of 2027.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Nvidia in the middle.&lt;/strong&gt; Nvidia bought $2 billion of convertible preferred stock in Marvell on March 31 and $2 billion in Lumentum on March 2. In an all-stock deal, its Lumentum preferred would likely be exchanged for Marvell shares, depending on the preferred terms and the merger agreement. Whether Nvidia supports the deal is unknown, and its view matters as both an investor and a customer.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Lumentum holders share the fall.&lt;/strong&gt; Lumentum rose only 0.6% on Thursday. Once a fixed exchange ratio is signed, a lower Marvell price means a lower price for Lumentum. Before signing, Lumentum&amp;rsquo;s board can still push for more cash or price protection. Thursday showed how much value a fixed ratio can lose in one day.&lt;/p&gt;
&lt;h2 id="what-would-change-our-view"&gt;What would change our view
&lt;/h2&gt;&lt;p&gt;The following announcement terms would tell us that this deal looks more like Inphi than like Cavium. If they arrive together, we would treat the first-day drop as an overreaction. They are signals of intent, not proof of execution.&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;A public commitment to run Lumentum as a neutral laser supplier, with information walls and continued supply to outside module makers.&lt;/li&gt;
&lt;li&gt;A plan for Cloud Light, Lumentum&amp;rsquo;s module business, such as a separation or sale.&lt;/li&gt;
&lt;li&gt;A premium of 20% or less.&lt;/li&gt;
&lt;li&gt;A synergy target of at least $1 billion a year, with a date. A 20% premium is worth about $19 billion. Valued at 25 times after-tax earnings, it takes about $0.9 billion of annual pre-tax synergy to earn that back, by our estimate.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;The opposite signal would be a plan to integrate modules and sell complete optical engines in direct competition with module makers. In that case, customer conflict would cancel the strategic gain from owning the laser.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Date&lt;/th&gt;
 &lt;th&gt;What to check&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;November 5&lt;/td&gt;
 &lt;td&gt;Lumentum&amp;rsquo;s first-quarter results against guidance of $1.225–1.275 billion in revenue, plus any comment on laser lead times or the talks&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Early November&lt;/td&gt;
 &lt;td&gt;The earliest announcement window in press reports: cash share, exchange ratio, committed financing, synergy target, neutral-supply terms, module plan, expected close and Nvidia&amp;rsquo;s position&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;5, 20 and 60 trading days after any announcement&lt;/td&gt;
 &lt;td&gt;Marvell against the SOX, compared with the Inphi path of -8.2, -4.1 and +0.2 points&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Early December (date not confirmed)&lt;/td&gt;
 &lt;td&gt;Marvell&amp;rsquo;s fiscal third-quarter results, buyback pace and funding plan&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;&lt;strong&gt;Is Lumentum Marvell&amp;rsquo;s next Inphi or its next Cavium? Tell us which matters more to you: how long the laser shortage lasts, or whether Lumentum&amp;rsquo;s customers keep buying from a Marvell-owned supplier.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Research and information only; not investment advice. Estimates and scenarios are analytical assumptions, not company guidance, trade instructions or personalized portfolio decisions. Event returns use Nasdaq closing prices, price only. Figures are date-bound. AI tools assisted research and editing.&lt;/em&gt;&lt;/p&gt;</description></item><item><title>SpaceX Is the Strongest Neocloud and the Weakest Hyperscaler. Its 90-Day Contracts Are Why.</title><link>https://koreainvestinsights.com/post/spacex-strongest-neocloud-weakest-hyperscaler-2026-10-29/</link><pubDate>Thu, 29 Oct 2026 14:30:00 +0900</pubDate><guid>https://koreainvestinsights.com/post/spacex-strongest-neocloud-weakest-hyperscaler-2026-10-29/</guid><description>&lt;img src="https://koreainvestinsights.com/" alt="Featured image of post SpaceX Is the Strongest Neocloud and the Weakest Hyperscaler. Its 90-Day Contracts Are Why." /&gt;&lt;p&gt;SpaceX now runs more installed AI compute than any neocloud. Its Superintelligence (SI) segment had 1.7 gigawatts of nameplate compute at September 30, up from 0.6GW a year earlier, and its CFO said on the third-quarter call that it reached about 2.2GW by the end of October. It builds faster than anyone in the group. It also borrows more cheaply: its $25 billion of unsecured notes carry a weighted coupon of 5.855%, while CoreWeave&amp;rsquo;s unsecured bonds yield about 15%.&lt;/p&gt;
&lt;p&gt;Compare it with the companies it is so often grouped with, and the picture flips. SpaceX&amp;rsquo;s backlog is &lt;strong&gt;1.08 times&lt;/strong&gt; its annualized revenue. CoreWeave&amp;rsquo;s is 10.1 times, Oracle&amp;rsquo;s 8.6 times and Google Cloud&amp;rsquo;s 4.6 times. Its cloud contracts typically run three years, but &lt;strong&gt;either side can end them on 90 days&amp;rsquo; notice&lt;/strong&gt;. In the third quarter, backlog grew $15.5 billion, and all of that growth is due within a year.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Our judgment: SpaceX is the strongest neocloud and the weakest hyperscaler. It has a near-hyperscaler balance sheet and a neocloud&amp;rsquo;s order book. Its 65.5% SI adjusted EBITDA margin is the price of renting scarce GPUs on short terms, not proof of lower costs. At long-term contract prices, the payback that management calls &amp;ldquo;under one year&amp;rdquo; stretches to 1.8 to 3.4 years.&lt;/strong&gt; That does not make the business weak. It means SpaceX&amp;rsquo;s AI revenue should not be valued as if it carried Microsoft&amp;rsquo;s or Google&amp;rsquo;s contracts.&lt;/p&gt;
&lt;p&gt;On October 25 we ranked SpaceX fifth of six AI giants overall and first in scarce physical infrastructure. &lt;a class="link" href="https://koreainvestinsights.com/post/hyperscaler-leadership-six-company-comparison-2026-10-25/" &gt;KII, October 25&lt;/a&gt; The two results fit together. The infrastructure is scarce, and the contracts that sell it are short.&lt;/p&gt;
&lt;p&gt;This article is dated October 29, 2026. SpaceX figures are at September 30 unless stated. CoreWeave, Nebius and Amazon figures are at June 30 because their third-quarter results are not out yet; Oracle&amp;rsquo;s are at August 31. Quotes from SpaceX&amp;rsquo;s October 20 call come from a third-party transcript, not a company-edited one.&lt;/p&gt;
&lt;h2 id="the-short-version"&gt;The short version
&lt;/h2&gt;&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Neocloud test, SpaceX wins:&lt;/strong&gt; the most installed capacity, the fastest builds, the only investment-grade balance sheet and the highest margin.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Hyperscaler test, SpaceX loses:&lt;/strong&gt; the thinnest backlog, contracts with 90-day exits, no in-house chip, internal demand that does not yet bring in much outside cash, and a lower credit rating.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;The margin:&lt;/strong&gt; 65.5% SI adjusted EBITDA comes from a short-term price band of $30–50 per watt-year. Large long-term contracts pay $20–25.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Concentration:&lt;/strong&gt; an unnamed SI customer, &amp;ldquo;Customer B,&amp;rdquo; was 34.5% of consolidated third-quarter revenue, up from 19.5% a quarter earlier.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Funding:&lt;/strong&gt; the bonds are cheap, but a related-party sale-leaseback costs an estimated 10.7% a year, and a $40 billion GPU lease starts amortizing in 2027.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;The test:&lt;/strong&gt; whether the first exit windows pass with renewals at higher prices. Google&amp;rsquo;s contract becomes terminable on 90 days&amp;rsquo; notice after December 31, 2026, and Anthropic&amp;rsquo;s after initial non-cancellable periods that end in January and March 2027.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id="where-spacex-ranks-on-each-test"&gt;Where SpaceX ranks on each test
&lt;/h2&gt;&lt;p&gt;The table compares business structures, not expected returns.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Test&lt;/th&gt;
 &lt;th&gt;Against neoclouds&lt;/th&gt;
 &lt;th&gt;Against hyperscalers&lt;/th&gt;
 &lt;th&gt;Deciding fact&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Installed capacity and build speed&lt;/td&gt;
 &lt;td&gt;First&lt;/td&gt;
 &lt;td&gt;Not comparable; hyperscalers do not disclose total gigawatts&lt;/td&gt;
 &lt;td&gt;1.7GW nameplate; 220MW of GB300s brought up in 64 days&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Cost of debt&lt;/td&gt;
 &lt;td&gt;First&lt;/td&gt;
 &lt;td&gt;Last&lt;/td&gt;
 &lt;td&gt;5.855% coupon against about 15% for CoreWeave; Baa1/BBB/BBB+ against Microsoft&amp;rsquo;s AAA&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Margin&lt;/td&gt;
 &lt;td&gt;First on adjusted EBITDA&lt;/td&gt;
 &lt;td&gt;Last on operating margin&lt;/td&gt;
 &lt;td&gt;SI adjusted EBITDA 65.5%, but SI operating margin 5.7% against 36.2% at Google Cloud&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Contract durability&lt;/td&gt;
 &lt;td&gt;Last&lt;/td&gt;
 &lt;td&gt;Last&lt;/td&gt;
 &lt;td&gt;Backlog 1.08x revenue; 90-day mutual exits&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Chip cost&lt;/td&gt;
 &lt;td&gt;Level; all buy Nvidia or AMD&lt;/td&gt;
 &lt;td&gt;Last&lt;/td&gt;
 &lt;td&gt;No in-house accelerator before 2028 at the earliest&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Own demand&lt;/td&gt;
 &lt;td&gt;First; neoclouds have little&lt;/td&gt;
 &lt;td&gt;Last&lt;/td&gt;
 &lt;td&gt;Grok and Cursor, against search, Office and ads&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;h2 id="against-the-neoclouds-spacex-wins-on-size-speed-and-money"&gt;Against the neoclouds, SpaceX wins on size, speed and money
&lt;/h2&gt;&lt;h3 id="it-has-the-most-installed-compute-and-builds-the-fastest"&gt;It has the most installed compute and builds the fastest
&lt;/h3&gt;&lt;p&gt;The question here is how much compute each company already runs, and how fast it adds more.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Company&lt;/th&gt;
 &lt;th&gt;Running&lt;/th&gt;
 &lt;th&gt;Contracted or targeted&lt;/th&gt;
 &lt;th&gt;As of&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;SpaceX&lt;/td&gt;
 &lt;td&gt;1.7GW nameplate; about 2.2GW at end-October (CFO, call)&lt;/td&gt;
 &lt;td&gt;2.5–3GW by year-end; management said it would be surprised by less than 13GW at end-2027&lt;/td&gt;
 &lt;td&gt;Sep 30&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;CoreWeave&lt;/td&gt;
 &lt;td&gt;1.5GW active, 51 data centers&lt;/td&gt;
 &lt;td&gt;About 4.2GW contracted; over 1.85GW active by year-end; 8GW or more by 2030&lt;/td&gt;
 &lt;td&gt;Jun 30&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Nebius&lt;/td&gt;
 &lt;td&gt;Not disclosed&lt;/td&gt;
 &lt;td&gt;0.8–1GW connected by year-end; 5GW contracted-power target&lt;/td&gt;
 &lt;td&gt;Jun 30&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Nscale&lt;/td&gt;
 &lt;td&gt;About 25,000 GPUs&lt;/td&gt;
 &lt;td&gt;About 1.37GW running and contracted, 461,000 GPUs&lt;/td&gt;
 &lt;td&gt;S-1, September&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Oracle&lt;/td&gt;
 &lt;td&gt;Total not disclosed&lt;/td&gt;
 &lt;td&gt;Added 850MW in its fiscal first quarter&lt;/td&gt;
 &lt;td&gt;Aug 31&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;The definitions differ, and the difference matters. SpaceX&amp;rsquo;s nameplate figure counts installed GPUs times their full power draw, excludes cooling and distribution losses, and says nothing about utilization. CoreWeave reports active power. CoreWeave also has far more capacity under contract than SpaceX runs today. &amp;ldquo;Largest&amp;rdquo; therefore means installed capacity, not secured demand.&lt;/p&gt;
&lt;p&gt;The clearer edge is speed. According to its prospectus, SpaceX brought up the first Colossus cluster (about 100,000 H100s, about 130MW) in 122 days. Colossus II took 91 days for about 110,000 GB200s (about 210MW), then 64 days for 110,000 GB300s (220MW). The prospectus puts the industry norm for a new 100MW data center at about two years. &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1181412/000162828026042639/spaceexplorationtechnologi.htm" target="_blank" rel="noopener"
 &gt;Prospectus&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;On the October call, Elon Musk said SpaceX is turning on about 200,000 GB300s a month for the rest of the year. At roughly 2kW per GPU, that is about 0.4GW a month. Two more months at that pace would take nameplate from 2.2GW to about 3GW, the top of its year-end range.&lt;/p&gt;
&lt;p&gt;The speed comes from owning what sits under the chips. SpaceX bought about $2 billion of mobile gas turbines in April 2026, has bought about $0.99 billion of Tesla Megapacks to date, and designed Colossus II to run on its own gas generation. Neoclouds that rent shells and wait for grid power cannot match that. The same choice carries a legal risk: a Clean Air Act suit over turbines at Southaven, Mississippi, with a pending motion for an injunction.&lt;/p&gt;
&lt;h3 id="it-is-the-only-investment-grade-neocloud"&gt;It is the only investment-grade neocloud
&lt;/h3&gt;&lt;p&gt;The question here is what each company pays for money.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;&lt;/th&gt;
 &lt;th&gt;SpaceX&lt;/th&gt;
 &lt;th&gt;CoreWeave&lt;/th&gt;
 &lt;th&gt;Nebius&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Ratings&lt;/td&gt;
 &lt;td&gt;Baa1 / BBB / BBB+ (June 2026, stable)&lt;/td&gt;
 &lt;td&gt;Company rating not confirmed in filings; DDTL 5.5 rated Ba2/BB+&lt;/td&gt;
 &lt;td&gt;Not disclosed&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Representative cost&lt;/td&gt;
 &lt;td&gt;Unsecured notes, 5.855% weighted coupon, 11.7-year weighted maturity&lt;/td&gt;
 &lt;td&gt;2032 9.625% notes yield about 15%; DDTL 4.0 at SOFR+2.25%&lt;/td&gt;
 &lt;td&gt;$3.5bn Citi senior secured facility at SOFR+2.75–3.50%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Interest vs adjusted EBITDA&lt;/td&gt;
 &lt;td&gt;7.7% (3Q, consolidated)&lt;/td&gt;
 &lt;td&gt;42.4% (2Q, net interest)&lt;/td&gt;
 &lt;td&gt;Not comparable&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Liquidity&lt;/td&gt;
 &lt;td&gt;$89.9bn cash and short-term securities&lt;/td&gt;
 &lt;td&gt;$5.5bn cash and securities, $10bn undrawn&lt;/td&gt;
 &lt;td&gt;$8.04bn cash&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;After its IPO and a $25 billion bond issue, SpaceX retired expensive debt inherited from X and xAI, including 12.5% secured notes. It paid $1.16 billion in prepayment premiums and booked a $1.53 billion loss on extinguishment. Third-quarter interest expense was 7.7% of consolidated adjusted EBITDA. &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1181412/000162828026066880/spcx-20260930.htm" target="_blank" rel="noopener"
 &gt;3Q 10-Q&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;The rating rests on Starlink, not SI. Third-quarter connectivity operating income was $2.37 billion, 5.5 times SI&amp;rsquo;s $0.43 billion. Bond investors are also not paying top-tier prices. The notes were carried at a fair value of $22.9 billion against a book value of $24.9 billion, about 92%. Treating them as one bond, we estimate an implied yield of about 6.9%.&lt;/p&gt;
&lt;h3 id="its-margin-is-the-highest-in-the-group"&gt;Its margin is the highest in the group
&lt;/h3&gt;&lt;p&gt;SI revenue was $7.56 billion in the third quarter, up 830% from a year earlier. Its gross margin was 64.6% and its adjusted EBITDA margin 65.5%. CoreWeave&amp;rsquo;s adjusted EBITDA margin was 58.6% with a GAAP operating loss, and Nebius&amp;rsquo;s was 40.6% (49.7% in its AI cloud). &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1181412/000162828026066482/spaceexplorationtechnologi.htm" target="_blank" rel="noopener"
 &gt;3Q 8-K&lt;/a&gt;, &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1769628/000176962826000362/coreweave2q26earningspress.htm" target="_blank" rel="noopener"
 &gt;CoreWeave 2Q&lt;/a&gt;, &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1513845/000110465926094568/tm2622968d1_ex99-2.htm" target="_blank" rel="noopener"
 &gt;Nebius 2Q letter&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;On the neocloud scorecard, then, SpaceX wins on capacity, speed, funding and margin. The trouble starts when the comparison moves to the companies whose revenue is contracted for years.&lt;/p&gt;
&lt;h2 id="against-the-hyperscalers-spacex-loses-on-the-contract"&gt;Against the hyperscalers, SpaceX loses on the contract
&lt;/h2&gt;&lt;h3 id="it-has-the-thinnest-order-book"&gt;It has the thinnest order book
&lt;/h3&gt;&lt;p&gt;The chart asks one question: how many years of current revenue each company has already signed.&lt;/p&gt;
&lt;p&gt;&lt;img alt="Contract backlog divided by annualized latest-quarter revenue. Nebius 16.1 times, CoreWeave 10.1, Oracle 8.6, Google Cloud 4.6, AWS 2.9, Microsoft Cloud 2.6 and SpaceX 1.08. SpaceX is the lowest in the group." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/spacex-strongest-neocloud-weakest-hyperscaler-2026-10-29/backlog-multiple.en.png"&gt;
&lt;/p&gt;
&lt;p&gt;SpaceX is below every hyperscaler, not only below the neoclouds. For investors, its AI revenue has less contracted visibility than any company in the comparison.&lt;/p&gt;
&lt;p&gt;A low multiple is not always a weakness. Microsoft&amp;rsquo;s commercial remaining performance obligation (RPO) is only 2.6 times Microsoft Cloud revenue, with a weighted remaining life of about 2.3 years. Office, search and advertising fill Microsoft&amp;rsquo;s capacity without long contracts. SpaceX&amp;rsquo;s equivalents are Grok and Cursor, and management aims to use about half of its capacity internally. In the third quarter, though, SI advertising brought in $538 million and Cursor $608 million, small against $7.56 billion of SI revenue. Internal demand fills capacity; it does not yet pay for it.&lt;/p&gt;
&lt;h3 id="all-of-the-quarters-backlog-growth-is-short"&gt;All of the quarter&amp;rsquo;s backlog growth is short
&lt;/h3&gt;&lt;p&gt;&lt;img alt="SpaceX consolidated backlog by when it turns into revenue. At June 30, 2026, $26.6bn was due within one year and $20.9bn after, a total of $47.5bn. At September 30, $42.2bn was due within one year and $20.8bn after, a total of $63.0bn. The within-one-year part rose $15.6bn while the after-one-year part fell $0.1bn." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/spacex-strongest-neocloud-weakest-hyperscaler-2026-10-29/backlog-split.en.png"&gt;
&lt;/p&gt;
&lt;p&gt;Backlog rose 32.7% in the quarter, from $47.5 billion to $63.0 billion. The share due within twelve months rose from 56% to 67%. The part due after a year did not grow at all.&lt;/p&gt;
&lt;p&gt;The 10-Q explains why. SpaceX recognizes cloud revenue as a fixed monthly fee over the contract term, which is &amp;ldquo;typically three years,&amp;rdquo; and either party can terminate on 90 days&amp;rsquo; notice. A contract that can be ended on 90 days&amp;rsquo; notice enters backlog only for the period that can be enforced. The more of these contracts SpaceX signs, the more its backlog piles up in the next twelve months. &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1181412/000162828026066880/spcx-20260930.htm" target="_blank" rel="noopener"
 &gt;3Q 10-Q&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;One caution: this is company-wide backlog, including government space and connectivity contracts, because SpaceX does not report AI backlog separately. What the numbers do show is that the company added no long-dated contracts on net in the quarter.&lt;/p&gt;
&lt;h3 id="one-unnamed-customer-is-two-thirds-of-si-revenue"&gt;One unnamed customer is two-thirds of SI revenue
&lt;/h3&gt;&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Item&lt;/th&gt;
 &lt;th style="text-align: right"&gt;2Q26&lt;/th&gt;
 &lt;th style="text-align: right"&gt;3Q26&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Customer B share of consolidated revenue&lt;/td&gt;
 &lt;td style="text-align: right"&gt;19.5%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;34.5%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Customer B revenue (our calculation)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;about $1.5bn&lt;/td&gt;
 &lt;td style="text-align: right"&gt;about $5.0bn&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Customer B share of SI revenue (our calculation)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;about 60%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;about 67%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Largest single customer&amp;rsquo;s share of receivables&lt;/td&gt;
 &lt;td style="text-align: right"&gt;none above 10% at end-2025&lt;/td&gt;
 &lt;td style="text-align: right"&gt;27.7%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Source: SpaceX 10-Q, note 3. Customer B is an SI customer whose name the filing does not disclose. We do not link it to any company.&lt;/p&gt;
&lt;h3 id="the-disclosed-contracts-all-have-exits"&gt;The disclosed contracts all have exits
&lt;/h3&gt;&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Customer&lt;/th&gt;
 &lt;th&gt;Size&lt;/th&gt;
 &lt;th&gt;Price and term&lt;/th&gt;
 &lt;th&gt;Exit&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Google (June 5 filing)&lt;/td&gt;
 &lt;td&gt;About 110,000 Nvidia GPUs&lt;/td&gt;
 &lt;td&gt;$920m a month, October 2026 to June 2029, at a reduced fee during the ramp through September&lt;/td&gt;
 &lt;td&gt;If capacity was not delivered by September 30, Google could terminate after a one-month grace period or cut the fee. After December 31, 2026, either side on 90 days&amp;rsquo; notice&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Anthropic (prospectus)&lt;/td&gt;
 &lt;td&gt;About 325,000 Nvidia GPUs&lt;/td&gt;
 &lt;td&gt;$1.25bn a month to May 2029&lt;/td&gt;
 &lt;td&gt;After the first three months, either side on 90 days&amp;rsquo; notice&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Anthropic (its S-1, October 26)&lt;/td&gt;
 &lt;td&gt;Up to $120.5bn across contracts and amendments signed May–October 2026&lt;/td&gt;
 &lt;td&gt;About $3.1bn a month after full delivery, expected in December 2026&lt;/td&gt;
 &lt;td&gt;After initial non-cancellable periods that end in January and March 2027, either side on 90 days&amp;rsquo; notice&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Not named (2Q call)&lt;/td&gt;
 &lt;td&gt;$6.7bn of cloud revenue over six months, signed early in the third quarter&lt;/td&gt;
 &lt;td&gt;Ramping from October&lt;/td&gt;
 &lt;td&gt;Not disclosed&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Sources: &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1181412/000162828026041150/spacexagreementfwp.htm" target="_blank" rel="noopener"
 &gt;Google agreement filing&lt;/a&gt;, &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1181412/000162828026042639/spaceexplorationtechnologi.htm" target="_blank" rel="noopener"
 &gt;prospectus&lt;/a&gt;, &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/2130357/000162828026067168/anthropicpbc-sx1.htm" target="_blank" rel="noopener"
 &gt;Anthropic S-1&lt;/a&gt;, &lt;a class="link" href="https://s21.q4cdn.com/184289198/files/doc_financials/2026/q2/SpaceX-Q2-2026-Earnings_Transcript-FINAL.pdf" target="_blank" rel="noopener"
 &gt;2Q call transcript&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;The Google contract implies $50–54 a year per watt of GPU power, using 1.875 to 2.0kW per GPU. That is at or above the top of the price band management describes.&lt;/p&gt;
&lt;h3 id="it-has-no-chip-of-its-own"&gt;It has no chip of its own
&lt;/h3&gt;&lt;p&gt;SpaceX buys Nvidia exclusively; management said so on the second-quarter call. Its own chip effort, Terafab, targets a test chip in mid-2027 and volume production in the third quarter of 2028, a schedule Musk himself called aggressive. Google sells its TPUs to outside customers. Amazon has multi-year, gigawatt-scale Trainium commitments from Anthropic and OpenAI. Microsoft has Maia and Meta has MTIA.&lt;/p&gt;
&lt;p&gt;A hyperscaler can lower the cost of a unit of compute by moving work onto its own silicon. SpaceX pays Nvidia&amp;rsquo;s price. Its technical lead sits below the chip, in land, power, cooling and assembly.&lt;/p&gt;
&lt;h3 id="its-credit-sits-a-rating-class-below"&gt;Its credit sits a rating class below
&lt;/h3&gt;&lt;p&gt;Microsoft is rated AAA and has issued no debt this year. Alphabet is rated Aa2/AA+ and held $265.1 billion of cash and securities. Meta is rated Aa3/AA-. SpaceX is Baa1/BBB/BBB+, and its free cash flow was negative $9.96 billion in the third quarter and negative $34.97 billion over nine months. It expects 2026 capital spending of $83–88 billion.&lt;/p&gt;
&lt;h2 id="the-margin-is-a-scarcity-price-not-a-cost-advantage"&gt;The margin is a scarcity price, not a cost advantage
&lt;/h2&gt;&lt;p&gt;The chart asks how long SI&amp;rsquo;s EBITDA takes to repay the capital spent on a watt of capacity, at the two price levels now visible in the market.&lt;/p&gt;
&lt;p&gt;&lt;img alt="Years of SI EBITDA needed to recover capex. At short-term prices of $40 to $50 per watt-year, payback is 0.9 to 1.7 years. At long-term prices of $20 to $25, it is 1.8 to 3.4 years. Nebius’s disclosed payback on its second-quarter contracts, 1 year 10 months, sits at the start of the long-term range." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/spacex-strongest-neocloud-weakest-hyperscaler-2026-10-29/payback.en.png"&gt;
&lt;/p&gt;
&lt;p&gt;The price band comes from three sources. On the October call, management said its prices were slightly above the top of the $30–50 per watt range it had given a quarter earlier, without defining the unit or the period. CoreWeave said on September 17 that its short contracts of three to six months were running at about $40 million per megawatt per year, or $40 per watt-year. &lt;a class="link" href="https://investors.coreweave.com/news/news-details/2026/CoreWeave-Continues-to-Contract-New-Compute-Capacity-at-Higher-Prices/default.aspx" target="_blank" rel="noopener"
 &gt;CoreWeave, September 17&lt;/a&gt; Nebius puts its large long-term contracts at $20–25 and its short contracts at $40–50.&lt;/p&gt;
&lt;p&gt;The capital cost is our estimate. Nine-month SI capital spending of $40.0 billion, divided by the 0.9GW of nameplate added, gives about $44 per watt. Subtracting the entire company-wide rise in construction in progress, $19.1 billion, gives about $23. Part of that rise is not SI&amp;rsquo;s, and part will come online within months, so we use $30–44.&lt;/p&gt;
&lt;p&gt;The result holds even at the bottom of the cost estimate. At $23 per watt, long-term prices still give a payback of 1.4 to 1.8 years. &lt;strong&gt;A payback under one year needs short-term prices.&lt;/strong&gt; Nebius&amp;rsquo;s disclosed payback of one year and ten months on its second-quarter contracts, which came with about 70% prepayment, sits at the start of our long-term range. SpaceX&amp;rsquo;s efficiency edge comes from the price band it chooses, not from a lower cost base.&lt;/p&gt;
&lt;h2 id="the-biggest-buyers-sign-their-long-contracts-with-someone-else"&gt;The biggest buyers sign their long contracts with someone else
&lt;/h2&gt;&lt;p&gt;Three documents filed in the past ten days show the same pattern.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Document&lt;/th&gt;
 &lt;th&gt;What touches SpaceX&lt;/th&gt;
 &lt;th&gt;Source quality&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Alphabet third-quarter results (October 28)&lt;/td&gt;
 &lt;td&gt;Cloud revenue $30.71bn (+103%), operating margin 36.2%; TPU system sales contributed &amp;ldquo;meaningfully&amp;rdquo;; backlog $561bn; the CFO said Google began using third-party compute capacity from the end of the third quarter and will increase it in the fourth; 2027 capital spending to rise significantly&lt;/td&gt;
 &lt;td&gt;Filing, primary; call remarks from a third-party summary&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Anthropic S-1 (October 26)&lt;/td&gt;
 &lt;td&gt;Long-term compute commitments sit with TPUs (Google, and Broadcom equipment leases) and at least $269.4bn of minimum commitments with three cloud providers. Only the SpaceXAI contracts can be ended by either side without cause on 90 days&amp;rsquo; notice. The cloud contracts end only on material breach, Broadcom leases only on default, and Nscale and Lambda contracts on delivery failures&lt;/td&gt;
 &lt;td&gt;Filing, primary&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;SpaceX third-quarter call (October 20)&lt;/td&gt;
 &lt;td&gt;About half of capacity for internal use, possibly 60/40; SpaceX said it rents at what is effectively spot pricing and does not sign long-term contracts; renewals at higher prices per watt&lt;/td&gt;
 &lt;td&gt;Third-party transcript, secondary&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Google&amp;rsquo;s start date for third-party compute lines up with the end of the SpaceX contract&amp;rsquo;s ramp in September. Google did not name its supplier, so we cannot say they are the same contract.&lt;/p&gt;
&lt;p&gt;Meta&amp;rsquo;s filing adds the scale. Its non-cancellable commitments, mostly third-party cloud capacity, reached $454.5 billion at September 30, 3.5 times the $131.1 billion at the end of 2025. &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1326801/000162828026066773/meta-20260930.htm" target="_blank" rel="noopener"
 &gt;Meta 10-Q&lt;/a&gt; The largest buyers of outside capacity are the hyperscalers themselves.&lt;/p&gt;
&lt;p&gt;The demand pool is deep. But the buyers lock in long commitments with TPUs, cloud providers and take-or-pay neoclouds, and keep SpaceX on terms they can exit. SpaceX says it prefers it that way, because spot pricing pays more in a shortage and leaves room to move capacity in-house. In our reading, the main substitute for SpaceX&amp;rsquo;s outside revenue is the buyers&amp;rsquo; own long-term capacity, more than other neoclouds. That capacity arrives in 2027 and 2028: Google&amp;rsquo;s larger 2027 capital budget, Broadcom TPU leases that start in 2027, and the extra TPU capacity in Anthropic&amp;rsquo;s S-1 for 2028 to 2031. That is the first real test of SpaceX&amp;rsquo;s outside contracts.&lt;/p&gt;
&lt;h2 id="two-books-with-different-maturities"&gt;Two books with different maturities
&lt;/h2&gt;&lt;p&gt;The chart sets the time a customer has committed for against the time SpaceX has committed for.&lt;/p&gt;
&lt;p&gt;&lt;img alt="Length of each commitment in years. Notice either side needs to exit a SpaceX cloud contract: 90 days. Typical cloud contract term: 3 years. Apollo GPU finance lease: 5 to 5.25 years per tranche. Server and network useful life: 5 to 6 years. Unsecured notes: 11.7-year weighted maturity." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/spacex-strongest-neocloud-weakest-hyperscaler-2026-10-29/maturity-mismatch.en.png"&gt;
&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;The cost side is long.&lt;/strong&gt; Servers and network gear are depreciated over five to six years. On October 26 SpaceX signed a GPU finance lease with Apollo for about $40 billion, in tranches of five to five and a quarter years, paid quarterly in advance at one-twentieth of principal plus a fixed rate. The rate will be set at pricing, closing is expected within the year, and payments start in the first half of 2027. &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1181412/000110465926118614/tm2628247d1_8k.htm" target="_blank" rel="noopener"
 &gt;8-K, October 26&lt;/a&gt; Non-cancellable purchase commitments reached $37.8 billion, 67.6% of them due in 2027.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;One lease is expensive.&lt;/strong&gt; SpaceX has a sale-leaseback with Valor, a firm linked to director Antonio Gracias. The liability was $12.9 billion at September 30, with $351 million of third-quarter interest. Dividing that interest by the average balance gives about 10.7% a year, our estimate. That is far above the 6.3% weighted fixed rate the filing gives for other financing, and the filing does not explain the gap. Valor is 34% of debt principal and 52% of third-quarter interest expense.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;The revenue side is short.&lt;/strong&gt; Contracts can end on 90 days&amp;rsquo; notice, and 67% of backlog runs off within a year.&lt;/p&gt;
&lt;p&gt;In a shortage, this structure wins. Every renewal is a chance to raise prices, and management says renewals are being priced higher. When supply loosens, capacity without long contracts reprices first. In our reading, GPU supply could rise sharply in 2028 as substrate and interposer capacity expands.&lt;/p&gt;
&lt;p&gt;There are two cushions. Grok and Cursor can absorb capacity a customer hands back, and Starlink&amp;rsquo;s cash can carry the fixed payments. Internal use earns no outside cash, though, so the real cushion is connectivity cash flow, about $9.5 billion of annualized operating income on the third quarter. Debt of $38.0 billion plus the $40 billion Apollo lease comes to $78 billion, about 2.2 times annualized third-quarter adjusted EBITDA of $35.2 billion. On this rough measure, that is inside the 3.75 times leverage covenant on its revolving credit facility, though the covenant&amp;rsquo;s own definition may differ.&lt;/p&gt;
&lt;h2 id="what-it-means-for-korean-memory"&gt;What it means for Korean memory
&lt;/h2&gt;&lt;p&gt;At 200,000 GB300s a month and 288GB of HBM each, SpaceX is installing about 57.6 petabytes of HBM a month. For 2027, supply caps the volume, so SpaceX&amp;rsquo;s contract terms change little. For 2028, they matter more. CoreWeave and Nscale order GPUs against take-or-pay contracts. SpaceX orders against contracts that can end in 90 days. That makes SpaceX&amp;rsquo;s 2028 GPU orders the more rent-sensitive part of the HBM demand that SK hynix and Samsung Electronics are planning for.&lt;/p&gt;
&lt;h2 id="the-strongest-case-against-our-view"&gt;The strongest case against our view
&lt;/h2&gt;&lt;p&gt;&lt;strong&gt;Short contracts in a shortage are pricing power.&lt;/strong&gt; If renewals keep coming in higher, the 90-day exit works for SpaceX, not against it. CoreWeave&amp;rsquo;s September release also said it was contracting new capacity at higher prices.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Google is not a casual tenant.&lt;/strong&gt; Google held SpaceX shares worth $80 billion at June 30, subject to short-term sale restrictions, and Google executive Donald Harrison sits on SpaceX&amp;rsquo;s board. A customer with that stake has reasons to stay.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Internal demand is a real floor.&lt;/strong&gt; If Grok and Cursor grow into paying businesses, terminated capacity has somewhere to go.&lt;/p&gt;
&lt;p&gt;We would change our view if fourth- or first-quarter filings show the after-one-year backlog rising sharply, or if the January to March 2027 exit windows pass with no terminations and higher renewal prices. Then &amp;ldquo;sells short&amp;rdquo; becomes &amp;ldquo;sells dear.&amp;rdquo; A termination notice on a major cloud contract, Customer B above 40% of revenue, or an Apollo rate set 200 basis points or more above the unsecured notes would confirm the weakness.&lt;/p&gt;
&lt;h2 id="what-to-check-next"&gt;What to check next
&lt;/h2&gt;&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;When (KST)&lt;/th&gt;
 &lt;th&gt;What&lt;/th&gt;
 &lt;th&gt;How to read it&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;October 30, 6:00 a.m.&lt;/td&gt;
 &lt;td&gt;Amazon third quarter: AWS growth, RPO, capacity, Trainium&lt;/td&gt;
 &lt;td&gt;Updates the hyperscaler side of the comparison&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;November 11, 7:00 a.m.&lt;/td&gt;
 &lt;td&gt;CoreWeave third quarter: active and contracted power, short-term prices, backlog maturity&lt;/td&gt;
 &lt;td&gt;Short-term prices holding at $40 million per MW-year or more mean SpaceX&amp;rsquo;s pricing is industry-wide. A fall brings SpaceX&amp;rsquo;s repricing risk forward&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;After Alphabet&amp;rsquo;s 10-Q&lt;/td&gt;
 &lt;td&gt;Backlog maturity, third-party compute commitments, value of its SpaceX stake&lt;/td&gt;
 &lt;td&gt;Tests the Google link&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;December 31, 2026&lt;/td&gt;
 &lt;td&gt;Google contract becomes terminable on 90 days&amp;rsquo; notice; Apollo lease expected to close; year-end nameplate&lt;/td&gt;
 &lt;td&gt;First exit window&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;January–March 2027&lt;/td&gt;
 &lt;td&gt;Anthropic&amp;rsquo;s SpaceXAI contracts leave their initial non-cancellable periods; any termination notices&lt;/td&gt;
 &lt;td&gt;Second exit window&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;January–February 2027, date not set&lt;/td&gt;
 &lt;td&gt;SpaceX fourth-quarter results&lt;/td&gt;
 &lt;td&gt;Scorecard below&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Our scorecard for SpaceX&amp;rsquo;s fourth-quarter results. The probabilities are our subjective estimates.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Item&lt;/th&gt;
 &lt;th&gt;Our base case&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Probability&lt;/th&gt;
 &lt;th&gt;How we will judge it&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Year-end nameplate of 2.5GW or more&lt;/td&gt;
 &lt;td&gt;Achieved&lt;/td&gt;
 &lt;td style="text-align: right"&gt;75%&lt;/td&gt;
 &lt;td&gt;10-K key metrics. The Southaven injunction is the main downside risk&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Fourth-quarter capital spending of $35–40 billion&lt;/td&gt;
 &lt;td&gt;In range&lt;/td&gt;
 &lt;td style="text-align: right"&gt;65%&lt;/td&gt;
 &lt;td&gt;Cash flow statement. Below $35 billion points to supply or power limits&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;After-one-year backlog up 25% or more from $20.8 billion&lt;/td&gt;
 &lt;td&gt;Not achieved&lt;/td&gt;
 &lt;td style="text-align: right"&gt;30%&lt;/td&gt;
 &lt;td&gt;Backlog times one minus the within-one-year share&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Customer B below 30% of revenue&lt;/td&gt;
 &lt;td&gt;Achieved&lt;/td&gt;
 &lt;td style="text-align: right"&gt;55%&lt;/td&gt;
 &lt;td&gt;The Google fee and the six-month contracts widen the denominator&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Apollo lease closes with its rate disclosed&lt;/td&gt;
 &lt;td&gt;Achieved&lt;/td&gt;
 &lt;td style="text-align: right"&gt;70%&lt;/td&gt;
 &lt;td&gt;8-K or 10-K. A rate within 200 bp of the notes&amp;rsquo; implied 6.9% is a good result&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;No termination notice on a major cloud contract in January–March 2027&lt;/td&gt;
 &lt;td&gt;None&lt;/td&gt;
 &lt;td style="text-align: right"&gt;85%&lt;/td&gt;
 &lt;td&gt;8-K or 10-Q. A notice confirms the weakness&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;&lt;strong&gt;Is a 90-day exit clause in a GPU shortage a weakness or an option? Tell us which renewal you would watch first.&lt;/strong&gt;&lt;/p&gt;
&lt;h2 id="disclosure"&gt;Disclosure
&lt;/h2&gt;&lt;p&gt;We researched and drafted this analysis with AI models made by Anthropic. Anthropic is a SpaceX compute customer, a competitor of Grok, and the filer of an S-1 we cite. We report only Anthropic&amp;rsquo;s public disclosures and do not estimate how its contracts affect SpaceX&amp;rsquo;s revenue or value. We do not link SpaceX&amp;rsquo;s unnamed Customer B to any company.&lt;/p&gt;
&lt;h2 id="main-sources"&gt;Main sources
&lt;/h2&gt;&lt;ul&gt;
&lt;li&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1181412/000162828026066880/spcx-20260930.htm" target="_blank" rel="noopener"
 &gt;SpaceX 3Q26 10-Q&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1181412/000162828026052535/spcx-20260630.htm" target="_blank" rel="noopener"
 &gt;SpaceX 2Q26 10-Q&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1181412/000162828026066482/spaceexplorationtechnologi.htm" target="_blank" rel="noopener"
 &gt;SpaceX 3Q26 results, 8-K exhibit 99.1&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1181412/000162828026042639/spaceexplorationtechnologi.htm" target="_blank" rel="noopener"
 &gt;SpaceX prospectus (424B4)&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1181412/000162828026041150/spacexagreementfwp.htm" target="_blank" rel="noopener"
 &gt;SpaceX Google cloud agreement filing, June 5, 2026&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1181412/000110465926118614/tm2628247d1_8k.htm" target="_blank" rel="noopener"
 &gt;SpaceX 8-K, Apollo GPU finance lease, October 26, 2026&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://s21.q4cdn.com/184289198/files/doc_financials/2026/q2/SpaceX-Q2-2026-Earnings_Transcript-FINAL.pdf" target="_blank" rel="noopener"
 &gt;SpaceX 2Q26 company-edited call transcript&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.fool.com/earnings/call-transcripts/2026/10/20/spacex-spcx-q3-2026-earnings-call-transcript/" target="_blank" rel="noopener"
 &gt;SpaceX 3Q26 call transcript, Motley Fool (secondary)&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/2130357/000162828026067168/anthropicpbc-sx1.htm" target="_blank" rel="noopener"
 &gt;Anthropic S-1&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1652044/000165204426000080/googexhibit991q32026.htm" target="_blank" rel="noopener"
 &gt;Alphabet 3Q26 results, 8-K exhibit 99.1&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1652044/000165204426000071/goog-20260630.htm" target="_blank" rel="noopener"
 &gt;Alphabet 2Q26 10-Q&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.marketbeat.com/earnings/reports/2026-10-28-alphabet-inc-stock/" target="_blank" rel="noopener"
 &gt;Alphabet 3Q26 call summary, MarketBeat (secondary)&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1769628/000176962826000362/coreweave2q26earningspress.htm" target="_blank" rel="noopener"
 &gt;CoreWeave 2Q26 results&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1769628/000176962826000366/crwv-20260630.htm" target="_blank" rel="noopener"
 &gt;CoreWeave 2Q26 10-Q&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://investors.coreweave.com/news/news-details/2026/CoreWeave-Continues-to-Contract-New-Compute-Capacity-at-Higher-Prices/default.aspx" target="_blank" rel="noopener"
 &gt;CoreWeave release, September 17, 2026&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1513845/000110465926094568/tm2622968d1_ex99-2.htm" target="_blank" rel="noopener"
 &gt;Nebius 2Q26 shareholder letter&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1341439/000119312526389274/orcl-20260831.htm" target="_blank" rel="noopener"
 &gt;Oracle 1Q FY27 10-Q&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/2110365/000119312526395475/ck0002110365-20260918.htm" target="_blank" rel="noopener"
 &gt;Nscale S-1&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/789019/000119312526436133/msft-20260930.htm" target="_blank" rel="noopener"
 &gt;Microsoft FY27 1Q 10-Q&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1326801/000162828026066773/meta-20260930.htm" target="_blank" rel="noopener"
 &gt;Meta 3Q26 10-Q&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1018724/000101872426000026/amzn-20260630.htm" target="_blank" rel="noopener"
 &gt;Amazon 2Q26 10-Q&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;&lt;em&gt;Research and information only; not investment advice. Rankings compare business structures, not expected returns, and probabilities are our subjective estimates. Figures are date-bound and come from the filings listed above unless marked as our calculation or estimate. AI tools assisted research and editing; see the disclosure above.&lt;/em&gt;&lt;/p&gt;</description></item><item><title>Samsung and SK Hynix Locked In 2027–28 Volume. Their Long-Term Contracts Do Not Lock In Profit.</title><link>https://koreainvestinsights.com/post/samsung-sk-hynix-lta-volume-locked-terms-flexible-2026-10-29/</link><pubDate>Thu, 29 Oct 2026 13:30:00 +0900</pubDate><guid>https://koreainvestinsights.com/post/samsung-sk-hynix-lta-volume-locked-terms-flexible-2026-10-29/</guid><description>&lt;img src="https://koreainvestinsights.com/" alt="Featured image of post Samsung and SK Hynix Locked In 2027–28 Volume. Their Long-Term Contracts Do Not Lock In Profit." /&gt;&lt;p&gt;On its earnings call on October 29, Samsung Electronics said it had signed long-term agreements (LTAs) with &lt;strong&gt;10 customers&lt;/strong&gt;, was negotiating with 10 more, and aimed to cover &lt;strong&gt;60–70% of its capacity through 2030&lt;/strong&gt;. A day earlier, SK hynix said LTAs with &lt;strong&gt;14 major customers&lt;/strong&gt; cover more than half of its 2027 and 2028 capacity. On its call, the company described DRAM and HBM as &lt;strong&gt;effectively sold out through 2029&lt;/strong&gt;.&lt;/p&gt;
&lt;p&gt;Those are the strongest statements either company has made about demand. Their legal filings say something narrower. Samsung&amp;rsquo;s half-year report notes that the total order value and delivery period of its multi-year contracts &lt;strong&gt;may change&lt;/strong&gt; with customer demand. SK hynix&amp;rsquo;s half-year report says it sets volumes and prices &lt;strong&gt;monthly and quarterly by mutual agreement&lt;/strong&gt; and has no order backlog based on long-term contracts. On the buyer side, Alphabet&amp;rsquo;s CFO called a $924 billion purchase-commitment figure a maximum and said many contracts carry price and volume flexibility.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Our judgment: both companies&amp;rsquo; LTAs are B-grade contracts. They lock in 2027–28 volume at market-linked prices. They are not a floor under profit.&lt;/strong&gt; Delivery in 2027 looks safe, because the shortage itself enforces the contracts. How much the contracts protect profit in a downturn cannot be calculated until the price floor, the penalty terms and the size of deposits are disclosed. This does not change our scenario weights.&lt;/p&gt;
&lt;p&gt;This article is dated October 29, 2026, after Samsung&amp;rsquo;s earnings call. Samsung and SK hynix balance-sheet figures are as of June 30, from their half-year reports. Microsoft and Meta figures are as of September 30. Alphabet had not filed its third-quarter 10-Q at the time of writing, so its $924 billion figure and the flexibility remark come from a call transcript. Statements from both Korean calls are taken from Korean press reports.&lt;/p&gt;
&lt;p&gt;&lt;img alt="Timeline from 2026 to 2031. SK hynix has disclosed LTAs covering more than 50% of capacity in 2027 and 2028; its “sold out through 2029” claim includes demand still under discussion. Samsung targets 60–70% of capacity under LTAs through 2030. Micron’s take-or-pay agreements cover more than 35% of expected revenue through 2030. Alphabet’s $707 billion of fixed or guaranteed commitments, mostly LTAs, run generally through 2030. Meta’s $10.8 billion escrow is released between 2028 and 2030." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/samsung-sk-hynix-lta-volume-locked-terms-flexible-2026-10-29/coverage-timeline.en.png"&gt;
&lt;/p&gt;
&lt;h2 id="five-things-to-take-away"&gt;Five things to take away
&lt;/h2&gt;&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;The firm part of the coverage is 2027–28.&lt;/strong&gt; SK hynix has disclosed more than 50% of 2027 and 2028 capacity under contract. Samsung&amp;rsquo;s 60–70% through 2030 is a target. &amp;ldquo;Sold out through 2029&amp;rdquo; adds demand still under discussion to contracts already signed.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Call language is about volume; filing language is about flexibility.&lt;/strong&gt; Both are true. The gap is one reason Korean memory still trades at a low multiple of 2027 earnings.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Buyers wrote the mirror image.&lt;/strong&gt; Suppliers say terms may change. Alphabet says it may spend less than the maximum. Meta does not count variable obligations beyond minimum quantities or prices. A downturn after 2028 would travel through deferred volume and the bottom of price bands, not through broken contracts.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Some of the collateral stays with the buyer.&lt;/strong&gt; Meta holds $10.8 billion in escrow, to be released in 2028–30. Microsoft holds $12.6 billion of investments restricted under a supplier agreement. That pattern fits SK hynix showing only ₩0.57 trillion of contract liabilities and advances.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Neither Korean company wins on contract quality.&lt;/strong&gt; Samsung&amp;rsquo;s collateral is visible on its balance sheet. SK hynix is more specific about 2027–28 coverage, and it keeps about half its capacity outside LTAs, which raises both its upside and its downside.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id="a-memory-contract-is-only-as-strong-as-what-the-buyer-must-do-in-a-downturn"&gt;A memory contract is only as strong as what the buyer must do in a downturn
&lt;/h2&gt;&lt;p&gt;&amp;ldquo;LTA&amp;rdquo; covers very different documents. We sort them into three grades by one test: what the buyer must still do when demand falls.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Grade&lt;/th&gt;
 &lt;th&gt;What the contract contains&lt;/th&gt;
 &lt;th&gt;What happens in a downturn&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;A&lt;/td&gt;
 &lt;td&gt;Prepayment or deposit, take-or-pay, minimum volume, cancellation penalty&lt;/td&gt;
 &lt;td&gt;The buyer pays even if it takes less; the supplier keeps the cash&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;B&lt;/td&gt;
 &lt;td&gt;Committed volume; price linked to the market and reset periodically, sometimes inside a band&lt;/td&gt;
 &lt;td&gt;Volume is deferred and price falls to the bottom of the band&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;C&lt;/td&gt;
 &lt;td&gt;Forecasts, letters of intent, double ordering&lt;/td&gt;
 &lt;td&gt;Orders disappear&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Take-or-pay means the buyer pays for the contracted volume even if it takes less. Micron is the useful benchmark: it is the only one of the three large memory makers that discloses A-grade terms in its filings, although only for part of its revenue.&lt;/p&gt;
&lt;h2 id="both-korean-contracts-grade-b-micron-is-closest-to-a"&gt;Both Korean contracts grade B; Micron is closest to A
&lt;/h2&gt;&lt;p&gt;The table answers one question: if demand weakens in 2028, what in the contract keeps the buyer paying?&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;&lt;/th&gt;
 &lt;th&gt;Samsung Electronics&lt;/th&gt;
 &lt;th&gt;SK hynix&lt;/th&gt;
 &lt;th&gt;Micron (benchmark)&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Signed&lt;/td&gt;
 &lt;td&gt;10 customers signed, 10 in talks&lt;/td&gt;
 &lt;td&gt;14 major customers&lt;/td&gt;
 &lt;td&gt;26 strategic customer agreements&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Coverage&lt;/td&gt;
 &lt;td&gt;Target: 60–70% of capacity through 2030&lt;/td&gt;
 &lt;td&gt;Secured: more than 50% of 2027 and 2028 capacity&lt;/td&gt;
 &lt;td&gt;More than 35% of expected revenue through 2030&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Term&lt;/td&gt;
 &lt;td&gt;Five years by default, extended a year at a time&lt;/td&gt;
 &lt;td&gt;Typically five years; some customers ask for longer&lt;/td&gt;
 &lt;td&gt;Some run to 2031&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Price&lt;/td&gt;
 &lt;td&gt;Floor on commodity memory (level undisclosed), negotiated quarterly; HBM priced annually; 2027 HBM agreed at target prices&lt;/td&gt;
 &lt;td&gt;Market-linked within an agreed range; a ceiling is not mentioned&lt;/td&gt;
 &lt;td&gt;Fixed, or with upper and lower bounds, for about three-quarters of agreement revenue&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Volume commitment&lt;/td&gt;
 &lt;td&gt;Half-year note: order value and delivery period may change&lt;/td&gt;
 &lt;td&gt;Half-year report: volume and price set monthly and quarterly; no LTA order backlog&lt;/td&gt;
 &lt;td&gt;Take-or-pay on all agreements&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Cash collateral&lt;/td&gt;
 &lt;td&gt;Contract liabilities up ₩9.61tn in six months, to ₩23.59tn&lt;/td&gt;
 &lt;td&gt;Deposits said to exist, amount undisclosed; ₩0.57tn of contract liabilities and advances&lt;/td&gt;
 &lt;td&gt;$12.75bn of customer deposits received&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Grade&lt;/td&gt;
 &lt;td&gt;B&lt;/td&gt;
 &lt;td&gt;B&lt;/td&gt;
 &lt;td&gt;Closest to A&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;For Samsung and SK hynix, the honest answer to the question is that we do not know yet. Neither has disclosed a floor level, a minimum purchase obligation or a penalty. Sources: &lt;a class="link" href="https://images.samsung.com/kdp/ir/financial-info/2026/2026_con_quarter02_all.pdf" target="_blank" rel="noopener"
 &gt;Samsung half-year review report&lt;/a&gt;, &lt;a class="link" href="https://images.samsung.com/is/content/samsung/assets/global/ir/docs/2026_3Q_conference_eng.pdf" target="_blank" rel="noopener"
 &gt;Samsung Q3 2026 IR presentation&lt;/a&gt;, &lt;a class="link" href="https://www.newsis.com/view/NISX20261029_0003846226" target="_blank" rel="noopener"
 &gt;Newsis on Samsung&amp;rsquo;s call&lt;/a&gt;, &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/2120882/000119312526354777/d147827d6k.htm" target="_blank" rel="noopener"
 &gt;SK hynix half-year report&lt;/a&gt;, &lt;a class="link" href="https://news.skhynix.com/en/q3-2026-business-results/" target="_blank" rel="noopener"
 &gt;SK hynix Q3 release&lt;/a&gt;, &lt;a class="link" href="https://s25.q4cdn.com/621799436/files/doc_financials/2026/q4/Q4-FY26-Prepared-Remarks.pdf" target="_blank" rel="noopener"
 &gt;Micron fiscal Q4 2026 prepared remarks&lt;/a&gt;.&lt;/p&gt;
&lt;h2 id="the-calls-talk-about-volume-the-filings-talk-about-flexibility"&gt;The calls talk about volume. The filings talk about flexibility.
&lt;/h2&gt;&lt;p&gt;SK hynix&amp;rsquo;s &amp;ldquo;effectively sold out through 2029&amp;rdquo; was said on the call. The Asia Business Daily reported the basis: LTA volume already secured plus demand under concrete discussion. Digital Daily quoted the company&amp;rsquo;s head of DRAM marketing describing HBM over the next three years, a narrower scope. &lt;a class="link" href="https://view.asiae.co.kr/article/2026102809584449660" target="_blank" rel="noopener"
 &gt;Asia Business Daily&lt;/a&gt;, &lt;a class="link" href="https://www.ddaily.co.kr/page/view/2026102810515586037" target="_blank" rel="noopener"
 &gt;Digital Daily&lt;/a&gt;. Either way, the statement describes visibility. It is not a disclosure of signed volume at fixed prices.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;A correction to our earlier post.&lt;/strong&gt; In our &lt;a class="link" href="https://koreainvestinsights.com/post/sk-hynix-80-percent-margin-contract-price-band-2026-10-29/" target="_blank" rel="noopener"
 &gt;October 29 article on SK hynix&amp;rsquo;s results&lt;/a&gt;, we wrote that we could not find &amp;ldquo;sold out through 2029&amp;rdquo; in the published results materials and treated it as unconfirmed. That was accurate for the written materials. Several Korean reports of the call now confirm the statement, on the basis described above.&lt;/p&gt;
&lt;p&gt;The filings are more precise. SK hynix&amp;rsquo;s half-year report, filed in English with the US SEC, says the company sets supply volumes and prices monthly and quarterly by mutual agreement with major customers, and that there is &amp;ldquo;no order backlog based on long term supply contracts.&amp;rdquo; Samsung&amp;rsquo;s half-year note on multi-year contracts says the total order value, the delivery period and other terms may change depending on customer demand and business progress.&lt;/p&gt;
&lt;p&gt;The calls added detail on both sides. SK hynix&amp;rsquo;s CFO said deposits and other practical enforcement mechanisms exist, without giving an amount. He also said 2027 HBM volumes and prices have been agreed, and that the profit gap between HBM and commodity DRAM has narrowed meaningfully. &lt;a class="link" href="https://www.newspim.com/news/view/20261028000278" target="_blank" rel="noopener"
 &gt;Newspim&lt;/a&gt;. SK hynix said it will keep its LTA share at roughly 50% or more over the medium term and manage the rest of its capacity flexibly.&lt;/p&gt;
&lt;p&gt;None of this is contradictory. The call statements are about how much demand the companies can see. The filings are about what the buyer is legally bound to take. A market that prices memory makers at a few times next year&amp;rsquo;s earnings is paying attention to the second.&lt;/p&gt;
&lt;h2 id="buyers-disclosed-the-same-flexibility-from-the-other-side"&gt;Buyers disclosed the same flexibility from the other side
&lt;/h2&gt;&lt;p&gt;If the suppliers&amp;rsquo; filings leave room to move, the buyers&amp;rsquo; filings should show the same room. They do.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Buyer&lt;/th&gt;
 &lt;th&gt;Commitments in filings&lt;/th&gt;
 &lt;th&gt;What the filing or call says about flexibility&lt;/th&gt;
 &lt;th&gt;Collateral disclosed&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Alphabet&lt;/td&gt;
 &lt;td&gt;$707.0bn of fixed or guaranteed commitments beyond one year at June 30, the significant majority LTAs, generally fulfilled through 2030. Total purchase commitments $811.0bn, up from $149.1bn at end-2025&lt;/td&gt;
 &lt;td&gt;CFO on the Q3 call: $924bn is the maximum contractual obligation; many contracts allow price and volume flexibility, so spending could be lower&lt;/td&gt;
 &lt;td&gt;None disclosed&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Meta&lt;/td&gt;
 &lt;td&gt;$454.52bn of non-cancelable commitments at Sept 30, up from $349.31bn at June 30. $121.41bn due in 2027, up from $81.65bn&lt;/td&gt;
 &lt;td&gt;For agreements with variable terms, Meta does not estimate obligations beyond minimum quantities or pricing&lt;/td&gt;
 &lt;td&gt;$10.80bn of money market funds in escrow, released 2028–30; $1.53bn of unused letters of credit&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Microsoft&lt;/td&gt;
 &lt;td&gt;$194.06bn of purchase commitments at June 30, 87% due within fiscal 2027, including take-or-pay&lt;/td&gt;
 &lt;td&gt;Higher memory and storage prices are built into its capex plan&lt;/td&gt;
 &lt;td&gt;$12.6bn of investments restricted under a supplier agreement; $35.7bn of receivables from facilitating server-component purchases&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;The suppliers wrote that terms may change. The buyers wrote that they may spend less than the headline. &lt;strong&gt;When both sides of a contract disclose the same flexibility, a downturn does not have to break the contract to reach the supplier&amp;rsquo;s profit.&lt;/strong&gt; It arrives as deferred volume and as prices drifting to the bottom of the agreed range.&lt;/p&gt;
&lt;p&gt;Two caveats apply. Buyer commitments include GPUs, TPUs, networking, power and data-center construction; the memory share is not disclosed, and we did not estimate it. None of the three buyers names its memory suppliers. Sources: &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1652044/000165204426000071/goog-20260630.htm" target="_blank" rel="noopener"
 &gt;Alphabet Q2 10-Q&lt;/a&gt;, &lt;a class="link" href="https://www.investing.com/news/transcripts/earnings-call-transcript-alphabet-q3-2026-profit-miss-sparks-10-afterhours-drop-93CH-4981037" target="_blank" rel="noopener"
 &gt;Alphabet Q3 call transcript&lt;/a&gt;, &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1326801/000162828026066773/meta-20260930.htm" target="_blank" rel="noopener"
 &gt;Meta Q3 10-Q&lt;/a&gt;, &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/789019/000119312526436133/msft-20260930.htm" target="_blank" rel="noopener"
 &gt;Microsoft FQ1 2027 10-Q&lt;/a&gt;, &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/789019/000119312526323660/msft-20260630.htm" target="_blank" rel="noopener"
 &gt;Microsoft FY2026 10-K&lt;/a&gt;.&lt;/p&gt;
&lt;h2 id="some-of-the-collateral-sits-on-the-buyers-balance-sheet"&gt;Some of the collateral sits on the buyer&amp;rsquo;s balance sheet
&lt;/h2&gt;&lt;p&gt;A supplier deposit is usually pictured as cash the buyer hands over. The filings show a second structure: the buyer locks up its own assets as security and keeps them on its own balance sheet.&lt;/p&gt;
&lt;p&gt;&lt;img alt="Left panel: Samsung’s contract liabilities rose from ₩13.98 trillion at end-2025 to ₩23.59 trillion at June 30, 2026, while SK hynix’s contract liabilities plus advances rose from ₩0.53 trillion to ₩0.57 trillion. Right panel: Micron has received $12.75 billion of customer deposits, held by the supplier. Microsoft holds $12.6 billion of investments restricted under a supplier agreement, and Meta holds $10.8 billion in escrow released between 2028 and 2030, both held by the buyer." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/samsung-sk-hynix-lta-volume-locked-terms-flexible-2026-10-29/collateral-holders.en.png"&gt;
&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Samsung&amp;rsquo;s collateral is visible.&lt;/strong&gt; Its contract liabilities rose from ₩13.98 trillion at the end of 2025 to ₩23.59 trillion at June 30, an increase of ₩9.61 trillion, or 68.8%. That is a change against the year-end balance, not a year-on-year change, as some reports described it. In the same period Samsung said it had written substantial deposit-like advances into its contracts. Not all of the balance is memory LTAs, since other businesses also carry contract liabilities. Other non-current liabilities rose from ₩7.95 trillion to ₩24.47 trillion over the same six months, but Samsung does not say where LTA deposits are booked, so we do not add the two accounts.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;SK hynix&amp;rsquo;s collateral is barely visible.&lt;/strong&gt; Contract liabilities were ₩0.52 trillion at June 30 and advances received ₩0.05 trillion, ₩0.57 trillion together. That is &lt;strong&gt;2.4% of Samsung&amp;rsquo;s contract liabilities&lt;/strong&gt;. Either the deposits had not been received by June 30, or they sit outside the balance sheet. A Korean report in August said the deposits are held in a third-party trust. The company has not confirmed that. &lt;a class="link" href="https://news.mtn.co.kr/news-detail/2026081913303555011" target="_blank" rel="noopener"
 &gt;MTN&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Meta and Microsoft keep the collateral themselves.&lt;/strong&gt; Meta reclassified $10.80 billion of money market funds as restricted cash for escrow under multi-year infrastructure purchase agreements, to be released between 2028 and 2030 as it meets the purchase obligations. Microsoft&amp;rsquo;s investments restricted under a supplier agreement rose from $11.3 billion to $12.6 billion in the quarter, and its receivables from helping buy server components rose from $27.8 billion to $35.7 billion. With this structure, the supplier gets assurance that the buyer will perform, but not cash. It matches what SK hynix&amp;rsquo;s balance sheet shows. The link to memory makers is still unconfirmed, because neither buyer names the supplier.&lt;/p&gt;
&lt;p&gt;Samsung&amp;rsquo;s summary third-quarter balance sheet shows other liabilities falling from ₩52.15 trillion to ₩48.63 trillion, and deferred revenue and other advances from ₩4.47 trillion to ₩3.81 trillion. The composition is not disclosed, so we do not read it as falling LTA deposits. The full quarterly report, due around November 13–16, will show contract liabilities at September 30.&lt;/p&gt;
&lt;h2 id="2027-is-protected-by-the-shortage-not-by-the-paperwork"&gt;2027 is protected by the shortage, not by the paperwork
&lt;/h2&gt;&lt;p&gt;In a shortage, even a B-grade contract behaves like an A-grade one. SK hynix says customers are asking for more than twice the HBM it can supply, and Samsung says its sellable 2027 HBM volume is already committed. A buyer that walks away from its contract now must buy at spot prices or from another supplier, and pay more.&lt;/p&gt;
&lt;p&gt;The buyers are also signalling that they intend to take the volume. Meta&amp;rsquo;s commitments due in 2027 rose by &lt;strong&gt;$39.8 billion in one quarter, or 49%&lt;/strong&gt;. Microsoft built higher memory and storage prices into its capex plan. Alphabet&amp;rsquo;s commitments under supply agreements run to 2030.&lt;/p&gt;
&lt;p&gt;What would break this: a customer asking to defer 2027 volume, or spot prices falling below contract prices during 2027. Neither has been reported.&lt;/p&gt;
&lt;h2 id="after-2028-the-flexibility-clauses-decide"&gt;After 2028, the flexibility clauses decide
&lt;/h2&gt;&lt;p&gt;Once new fabs arrive, the shortage stops enforcing the contracts, and only two terms are left to protect profit: the price floor and the minimum volume. Neither company has disclosed either. That is why a view on 2029 earnings cannot yet be proved from the contracts, and why we do not use the LTAs to argue that the market&amp;rsquo;s discount to post-2028 earnings is too steep.&lt;/p&gt;
&lt;p&gt;SK hynix&amp;rsquo;s 50% rule cuts both ways. Keeping about half of its capacity outside LTAs leaves room to earn more if the shortage lasts. If supply overtakes demand, that half is exposed to spot prices. If the contracted half also has a ceiling, SK hynix gives up some upside in a boom. The company has not said whether one exists.&lt;/p&gt;
&lt;p&gt;One early signal is worth watching. SK hynix said some customers may move HBM from 12-high to 8-high stacks and offset the lost capacity with higher capacity per die. The company frames this as a response to the shortage and says it is already in its supply plan, so we do not treat it as a sign of weakening demand. Still, a change in what buyers specify is the kind of adjustment that would come first if they start economizing.&lt;/p&gt;
&lt;p&gt;If credit tightens, which is our &lt;a class="link" href="https://koreainvestinsights.com/post/credit-squeeze-case-s2-reading-and-path-2026-10-24/" target="_blank" rel="noopener"
 &gt;credit-squeeze case&lt;/a&gt;, the first thing to bend would be whether buyers use the flexibility they disclosed. Microsoft&amp;rsquo;s commitments are mostly due within a year. Alphabet&amp;rsquo;s run to 2030. Meta&amp;rsquo;s escrow is tied to 2028–30.&lt;/p&gt;
&lt;h2 id="samsung-versus-sk-hynix-different-strengths-no-winner"&gt;Samsung versus SK hynix: different strengths, no winner
&lt;/h2&gt;&lt;p&gt;We do not rank the two on contract quality. Samsung leads on visible collateral: ₩9.61 trillion of new contract liabilities in six months is the clearest sign of cash changing hands at either company. SK hynix leads on specificity: it has named the share of 2027–28 capacity under contract and described its price mechanism.&lt;/p&gt;
&lt;p&gt;The bigger difference is cyclical exposure. SK hynix keeps about half its capacity outside LTAs by design. That makes its earnings more sensitive in both directions. Which company is better protected in a downturn can only be decided once the floor levels and penalty terms are public.&lt;/p&gt;
&lt;h2 id="our-judgment-and-what-would-change-it"&gt;Our judgment, and what would change it
&lt;/h2&gt;&lt;p&gt;&lt;strong&gt;The LTAs make 2027 delivery highly likely and leave the size of post-2028 downside unknown.&lt;/strong&gt; Our scenario weights are unchanged: S1 rebound 50%, S2 credit squeeze 15%, S3 demand slowdown 15%, S4 macro shock 20%. These are our subjective estimates; the &lt;a class="link" href="https://koreainvestinsights.com/post/rebound-case-s1-reading-and-path-2026-10-24/" target="_blank" rel="noopener"
 &gt;rebound case&lt;/a&gt; sets out what would confirm the first. The price-floor level and HBM4 pricing per gigabyte remain undisclosed.&lt;/p&gt;
&lt;p&gt;The next readings are close. The probabilities below are our subjective estimates.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;When (KST)&lt;/th&gt;
 &lt;th&gt;What to read&lt;/th&gt;
 &lt;th&gt;Base case&lt;/th&gt;
 &lt;th&gt;What would change the view&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Oct 30–31&lt;/td&gt;
 &lt;td&gt;Alphabet Q3 10-Q, commitments note&lt;/td&gt;
 &lt;td&gt;Fixed or guaranteed commitments of $707bn or more (80%)&lt;/td&gt;
 &lt;td&gt;Flexibility language written into the note itself would move it from a call remark into a filing&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Around Nov 13–16&lt;/td&gt;
 &lt;td&gt;Samsung Q3 quarterly report: contract liabilities at Sept 30&lt;/td&gt;
 &lt;td&gt;₩20–27tn (55%); above ₩27tn (25%); below ₩20tn (20%)&lt;/td&gt;
 &lt;td&gt;Below ₩20tn would mean falling advances, a weakening LTA signal; above ₩27tn would mean more collateral&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Around Nov 13–16&lt;/td&gt;
 &lt;td&gt;SK hynix Q3 quarterly report: order status&lt;/td&gt;
 &lt;td&gt;&amp;ldquo;No LTA order backlog&amp;rdquo; unchanged (85%)&lt;/td&gt;
 &lt;td&gt;A disclosed backlog or deposit amount would lift its contract quality within grade B&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Mid-December&lt;/td&gt;
 &lt;td&gt;Micron fiscal Q1 2027: contract backlog and deposits&lt;/td&gt;
 &lt;td&gt;Backlog near $150bn holds (70%)&lt;/td&gt;
 &lt;td&gt;A lower backlog or changed deposit-return terms would signal weaker contract quality across the industry&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Late January 2027&lt;/td&gt;
 &lt;td&gt;Samsung and SK hynix Q4 calls&lt;/td&gt;
 &lt;td&gt;Floor level, HBM4 price per gigabyte and deposit amount stay undisclosed (85%)&lt;/td&gt;
 &lt;td&gt;Any of the three in numbers would let us size the downside protection&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;&lt;strong&gt;Which number would make you treat these contracts as a floor under profit: the price-floor level, a take-or-pay clause, or the deposit amount? Tell us what you would need to see.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Research and information only; not investment advice. Estimates and scenarios are analytical assumptions, not company guidance, trade instructions or personalized portfolio decisions. Figures are date-bound. AI tools assisted research and editing.&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Interest disclosure: We use AI models, including Anthropic&amp;rsquo;s Claude, in our research. Alphabet is an investor in Anthropic and one of its cloud providers, and some Microsoft Copilot features use Anthropic models. Only public disclosures are cited; no assessment is made of the effect on either company.&lt;/em&gt;&lt;/p&gt;</description></item><item><title>Big Tech's Capex Miss Was a Delivery Problem. The Market Now Prices Who Can Pay for the Build, From Alphabet to Samsung.</title><link>https://koreainvestinsights.com/post/big-tech-capex-delivery-delay-who-pays-alphabet-samsung-2026-10-29/</link><pubDate>Thu, 29 Oct 2026 11:15:00 +0900</pubDate><guid>https://koreainvestinsights.com/post/big-tech-capex-delivery-delay-who-pays-alphabet-samsung-2026-10-29/</guid><description>&lt;img src="https://koreainvestinsights.com/" alt="Featured image of post Big Tech's Capex Miss Was a Delivery Problem. The Market Now Prices Who Can Pay for the Build, From Alphabet to Samsung." /&gt;&lt;p&gt;On October 28, Alphabet and Microsoft reported their September quarters and Oracle held its investor day. The two hyperscalers spent &lt;strong&gt;about $9 billion less on capex than analysts expected&lt;/strong&gt;, and both blamed supply. Alphabet fell &lt;strong&gt;about 10% after hours&lt;/strong&gt;, Microsoft rose &lt;strong&gt;about 3%&lt;/strong&gt;, and Oracle had already closed &lt;strong&gt;10.7% higher&lt;/strong&gt;. &lt;a class="link" href="https://www.cnbc.com/2026/10/28/alphabet-googl-earnings-q3-2026.html" target="_blank" rel="noopener"
 &gt;CNBC on Alphabet&lt;/a&gt;, &lt;a class="link" href="https://www.cnbc.com/2026/10/28/microsoft-msft-q1-earnings-report-2027.html" target="_blank" rel="noopener"
 &gt;CNBC on Microsoft&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;The next morning in Seoul, the KOSPI rose to about 6,100, up 1.5%, with all of that US news already known. Then, from 9:50 a.m., Samsung Electronics filed its results, its dividend and its capex plan, and the index reversed. By 11:15 a.m. it was down 3.0%.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Our judgment: the AI capex cycle has not slowed, but its funding has changed. Spending now runs ahead of operating cash at most builders, so markets reward the companies that can pay and punish the ones that need outside capital.&lt;/strong&gt; The same test reached Seoul. Samsung&amp;rsquo;s own capex jump limits how much of its record profit can go to shareholders, and that, at least as much as US capex, is what the market traded as the KOSPI reversed after 9:50.&lt;/p&gt;
&lt;p&gt;This article is dated October 29, 2026. Korean market figures are intraday values at 11:15 a.m. KST and are provisional until the close.&lt;/p&gt;
&lt;h2 id="capex-held-up-the-funding-changed"&gt;Capex held up; the funding changed
&lt;/h2&gt;&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;The miss came from supply.&lt;/strong&gt; Alphabet said it is in a &amp;ldquo;supply-constrained environment&amp;rdquo; and started renting outside compute. Microsoft said some third-party capacity deliveries were delayed. Alphabet expects 2027 capex to rise significantly, and Microsoft expects spending to accelerate in January–June 2027.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Demand indicators are still rising.&lt;/strong&gt; Google Cloud revenue grew 103% and its backlog reached $561 billion. Azure grew 48%.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Part of the spending is price.&lt;/strong&gt; Microsoft named higher memory and storage costs in its second-half capex. Samsung said its DRAM prices rose about 20% in the third quarter.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Investors priced cash.&lt;/strong&gt; Alphabet had negative free cash flow, bought back no stock and issued equity. Microsoft had $18.6 billion of free cash flow. Their shares moved in opposite directions.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Oracle put a date window on the peak.&lt;/strong&gt; It expects peak capex in fiscal 2027–2028, through May 2028, and a step-down in fiscal 2029.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Seoul fell on a payout question.&lt;/strong&gt; Samsung&amp;rsquo;s shareholder returns are set at 50% of free cash flow, so its 84.4 trillion won capex plan caps the payout that its profit would otherwise support.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id="the-shortfall-came-from-supply"&gt;The shortfall came from supply
&lt;/h2&gt;&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Company&lt;/th&gt;
 &lt;th&gt;Quarter capex&lt;/th&gt;
 &lt;th&gt;Analyst estimate&lt;/th&gt;
 &lt;th&gt;Company explanation&lt;/th&gt;
 &lt;th&gt;Direction&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Alphabet, Jul–Sep 2026&lt;/td&gt;
 &lt;td&gt;$51.4bn&lt;/td&gt;
 &lt;td&gt;$55.1bn&lt;/td&gt;
 &lt;td&gt;&amp;ldquo;Supply-constrained environment&amp;rdquo;; started using third-party compute at the end of the quarter&lt;/td&gt;
 &lt;td&gt;2026 range of $195–205bn kept; 2027 to &amp;ldquo;increase significantly&amp;rdquo;&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Microsoft, Jul–Sep 2026&lt;/td&gt;
 &lt;td&gt;$45.2bn incl. finance leases&lt;/td&gt;
 &lt;td&gt;$50.7bn&lt;/td&gt;
 &lt;td&gt;Some third-party capacity deliveries delayed; demand still exceeds supply&lt;/td&gt;
 &lt;td&gt;December quarter above $50bn; about $170bn for the calendar year; faster in Jan–Jun 2027&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Oracle, investor day&lt;/td&gt;
 &lt;td&gt;Not given&lt;/td&gt;
 &lt;td&gt;—&lt;/td&gt;
 &lt;td&gt;Peak capex in fiscal 2027–2028&lt;/td&gt;
 &lt;td&gt;Step-down in fiscal 2029; free cash flow of at least $30bn&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Sources: &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/0001652044/000165204426000080/googexhibit991q32026.htm" target="_blank" rel="noopener"
 &gt;Alphabet earnings release&lt;/a&gt;; Alphabet and Microsoft calls; estimates from LSEG and StreetAccount via CNBC; &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1341439/000119312526435853/d75538dex991.htm" target="_blank" rel="noopener"
 &gt;Oracle investor day release&lt;/a&gt;. Microsoft&amp;rsquo;s figure includes finance leases; Alphabet&amp;rsquo;s is cash paid for property and equipment.&lt;/p&gt;
&lt;p&gt;The demand side points the other way. &lt;strong&gt;Google Cloud revenue rose 103% to $30.7 billion.&lt;/strong&gt; Its backlog was $561 billion, and Alphabet expects to recognize just over 55% of it within 24 months. Azure grew 48%, and Microsoft&amp;rsquo;s commercial remaining performance obligation rose 73% to $677 billion. We covered Microsoft&amp;rsquo;s quarter in detail &lt;a class="link" href="https://koreainvestinsights.com/post/microsoft-q1-fy27-azure-capex-cpu-memory-korea-readthrough-2026-10-29/" &gt;earlier today&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;Microsoft did not name the late partners. CNBC read the lower capex as a sign that infrastructure partners are taking longer to deliver, and noted that Microsoft relies on neoclouds for capacity. CoreWeave and Nebius fell almost 5% in extended trading, and IREN fell about 3%. &lt;a class="link" href="https://www.cnbc.com/2026/10/28/neocloud-stocks-drop-after-microsoft-revises-capex-down.html" target="_blank" rel="noopener"
 &gt;CNBC&lt;/a&gt; These companies build with debt, and CNBC also cited concern that rising interest rates will make that debt too costly. If the delay sits in that layer, a delivery problem is a funding problem in a different form.&lt;/p&gt;
&lt;h2 id="part-of-the-spending-is-price"&gt;Part of the spending is price
&lt;/h2&gt;&lt;p&gt;Microsoft said capex in the second half of its fiscal year, January to June 2027, will accelerate mainly because of CPUs, &amp;ldquo;inclusive of higher memory and storage costs&amp;rdquo;, and GPUs, as well as third-party deliveries that shifted into that period. In July, according to Data Center Knowledge, Amazon raised its 2026 plan to about $220 billion and said the increase mainly reflected higher memory costs. &lt;a class="link" href="https://datacenterknowledge.com/infrastructure/amazon-lifts-ai-infrastructure-spending-to-220b-as-demand-outpaces-capacity" target="_blank" rel="noopener"
 &gt;Data Center Knowledge&lt;/a&gt; On its call this morning, Samsung said its DRAM prices rose about 20% and its NAND prices rose in the high teens in the third quarter, led by server products. &lt;a class="link" href="https://n.news.naver.com/mnews/article/366/0001201158" target="_blank" rel="noopener"
 &gt;Chosun Biz via Naver&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;This matters for how to read capex totals. &lt;strong&gt;Capex dollars now grow faster than the compute they buy, and the difference shows up as memory makers&amp;rsquo; margins.&lt;/strong&gt; SK hynix earned an 80% operating margin last quarter. &lt;a class="link" href="https://koreainvestinsights.com/post/sk-hynix-80-percent-margin-contract-price-band-2026-10-29/" &gt;KII, Oct 29&lt;/a&gt; The buyers see the same number as a cost that their own cash flow has to carry.&lt;/p&gt;
&lt;h2 id="the-market-priced-cash"&gt;The market priced cash
&lt;/h2&gt;&lt;p&gt;&lt;img alt="Alphabet’s cash flows for January to September. Capex rose from $63.6 billion in 2025 to $132.0 billion in 2026, while operating cash flow rose from $112.3 billion to $127.1 billion. Share repurchases fell from $39.8 billion to zero. Net debt issued rose from $11.9 billion to $76.3 billion, and equity issued rose from zero to $61.9 billion, including common and mandatory convertible preferred stock." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/big-tech-capex-delivery-delay-who-pays-alphabet-samsung-2026-10-29/alphabet-funding.en.png"&gt;
&lt;/p&gt;
&lt;p&gt;Alphabet&amp;rsquo;s numbers explain its fall better than its capex. In the quarter, &lt;strong&gt;free cash flow was negative $9.1 billion&lt;/strong&gt;; CNBC cited an estimate of negative $2.7 billion. Alphabet bought back &lt;strong&gt;no stock&lt;/strong&gt;, against $11.5 billion a year earlier. It issued $12.4 billion of common stock through its at-the-market program. From January to September it issued $42.9 billion of common stock and $19.1 billion of mandatory convertible preferred stock, and its net new debt was $76.3 billion. Long-term debt rose to &lt;strong&gt;$125.1 billion from $46.5 billion&lt;/strong&gt; at the end of 2025. Alphabet still holds $265.1 billion of cash and marketable securities, so this is a funding choice made from strength. But one of the most profitable builders now funds part of its spending with outside capital. &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/0001652044/000165204426000080/googexhibit991q32026.htm" target="_blank" rel="noopener"
 &gt;Alphabet earnings release&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;Microsoft is the contrast. Its free cash flow was &lt;strong&gt;$18.6 billion&lt;/strong&gt;, down 27% but positive, and its shares rose. Oracle rose 10.7% after it said it would limit additional capital raises through fiscal 2030 to $20 billion, split between equity and debt and excluding refinancing. That is on top of a $20 billion at-the-market equity sale completed in the June–August quarter. &lt;a class="link" href="https://koreainvestinsights.com/post/oracle-fcf-turn-customers-landlords-pay-first-openai-test-2026-10-29/" &gt;KII on Oracle&lt;/a&gt; The three reactions line up with cash and funding.&lt;/p&gt;
&lt;p&gt;The suppliers show the same split from the other side. They rose in the regular session, before the two reports, and gave back only 2–3% after hours while Alphabet fell 10%:&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Stock or index, Oct 28&lt;/th&gt;
 &lt;th&gt;Regular session&lt;/th&gt;
 &lt;th&gt;After hours&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Vertiv&lt;/td&gt;
 &lt;td&gt;+16.5%&lt;/td&gt;
 &lt;td&gt;−2.8%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Micron&lt;/td&gt;
 &lt;td&gt;+5.9%&lt;/td&gt;
 &lt;td&gt;−2.5%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;AMD&lt;/td&gt;
 &lt;td&gt;+3.2%&lt;/td&gt;
 &lt;td&gt;−2.0%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Broadcom&lt;/td&gt;
 &lt;td&gt;+2.8%&lt;/td&gt;
 &lt;td&gt;−2.2%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Nvidia&lt;/td&gt;
 &lt;td&gt;−1.2%&lt;/td&gt;
 &lt;td&gt;−0.7%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Philadelphia Semiconductor Index&lt;/td&gt;
 &lt;td&gt;+1.75%&lt;/td&gt;
 &lt;td&gt;—&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Source: CNBC and Nasdaq quote data. After-hours trading is thin. Investors sold the payer that needs outside capital and only trimmed the suppliers.&lt;/p&gt;
&lt;p&gt;The cost of outside money is rising. The 10-year Treasury yield closed at &lt;strong&gt;5.05%&lt;/strong&gt; on October 28. &lt;a class="link" href="https://home.treasury.gov/resource-center/data-chart-center/interest-rates/TextView?type=daily_treasury_yield_curve&amp;amp;field_tdr_date_value_month=202610" target="_blank" rel="noopener"
 &gt;US Treasury&lt;/a&gt; The Federal Reserve held its target range at 3.75–4.00% the same day, and two members, Lisa Cook and Beth Hammack, voted for a quarter-point increase. &lt;a class="link" href="https://www.federalreserve.gov/newsevents/pressreleases/monetary20261028a.htm" target="_blank" rel="noopener"
 &gt;Federal Reserve&lt;/a&gt; When builders need debt and equity to fund capex, a 5% risk-free rate makes the cash test sharper.&lt;/p&gt;
&lt;h2 id="oracle-dated-the-peak"&gt;Oracle dated the peak
&lt;/h2&gt;&lt;p&gt;Oracle said: &amp;ldquo;Following a period of peak capex in FY27-FY28, capital expenditures are expected to step-down in FY29.&amp;rdquo; Its fiscal 2027 began in June 2026, and fiscal 2028 ends in May 2028. &lt;strong&gt;Oracle gave an explicit window for its capex peak. None of the hyperscalers that reported this month has given one.&lt;/strong&gt; It also said vendor financing, customer-provided hardware and customer prepayments will reduce its upfront cash needs. &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1341439/000119312526435853/d75538dex991.htm" target="_blank" rel="noopener"
 &gt;Oracle investor day release&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;The date does not signal weaker 2027 demand. It tells suppliers when one large buyer expects its build to stop rising. The funding sentence moves part of the cash burden to suppliers and customers. For memory makers, payment terms and prepayments will matter as much as price.&lt;/p&gt;
&lt;h2 id="seoul-reversed-after-samsungs-filings"&gt;Seoul reversed after Samsung&amp;rsquo;s filings
&lt;/h2&gt;&lt;p&gt;&lt;img alt="KOSPI, Samsung Electronics and SK hynix on October 29, one-minute bars, percent change from the October 28 close. The KOSPI rose to about 1.5% by 9:46 a.m. and turned down after Samsung’s filings between 9:50 and 10:02 a.m. Samsung Electronics fell further than the index. SK hynix gave back most of an intraday gain of about 6% but stayed above its previous close." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/big-tech-capex-delivery-delay-who-pays-alphabet-samsung-2026-10-29/kospi-intraday.en.png"&gt;
&lt;/p&gt;
&lt;p&gt;The KOSPI rose &lt;strong&gt;6.63% on October 28&lt;/strong&gt; to 6,012.44. Before the open on October 29, investors already knew that Alphabet had fallen 10% after hours, that Nasdaq-100 futures were lower and that the 10-year yield was at 5%. The index still rose to about 6,100 by 9:46 a.m. Samsung filed its third-quarter results at 9:50 a.m., a 30 trillion won quarterly dividend at 9:53 a.m., a correction to its preliminary results at 10:00 a.m. and its 2026 capex plan at 10:02 a.m. Its call started at 10:00 a.m. The index had already slipped from its 9:46 high when the first filing landed at 9:50. It went negative at 9:55 and kept falling. &lt;a class="link" href="https://kind.krx.co.kr/disclosure/todaydisclosure.do?method=searchTodayDisclosureMain" target="_blank" rel="noopener"
 &gt;KRX KIND disclosure times&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;At 11:15 a.m., the evidence pointed to domestic selling:&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Item, 11:15 a.m. KST&lt;/th&gt;
 &lt;th&gt;Value&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;KOSPI&lt;/td&gt;
 &lt;td&gt;5,830, −3.0%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Samsung Electronics&lt;/td&gt;
 &lt;td&gt;−6.6%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;SK hynix&lt;/td&gt;
 &lt;td&gt;+0.3%, after about +6% at 9:46 a.m.&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Foreign investors, KOSPI net&lt;/td&gt;
 &lt;td&gt;+394 billion won&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Institutions, KOSPI net&lt;/td&gt;
 &lt;td&gt;−633 billion won&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Individuals, KOSPI net&lt;/td&gt;
 &lt;td&gt;+125 billion won&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Advancers / decliners&lt;/td&gt;
 &lt;td&gt;132 / 759&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Nikkei 225 / Taiwan TAIEX&lt;/td&gt;
 &lt;td&gt;+0.2% / −1.4%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Source: Naver Finance market data, intraday and provisional. Nikkei and TAIEX quotes are delayed.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Foreign investors were net buyers while the index fell.&lt;/strong&gt; If the cause were a US capex scare, foreign selling of memory stocks would lead. Instead, institutions sold, and the steepest falls came from Korean news. LG Energy Solution fell about 10% after LG Chem decided to sell about 2.5 trillion won of its stake. Doosan Enerbility fell about 12% after Korea Hydro &amp;amp; Nuclear Power ended contract talks with its consortium for the Czech Dukovany plant. &lt;a class="link" href="https://n.news.naver.com/mnews/article/003/0014286680" target="_blank" rel="noopener"
 &gt;Newsis&lt;/a&gt; Korean power-equipment makers such as Hyosung Heavy Industries and HD Hyundai Electric fell about 6% on the morning after Vertiv rose 16.5% in New York.&lt;/p&gt;
&lt;h2 id="samsungs-payout-is-bound-by-its-own-capex"&gt;Samsung&amp;rsquo;s payout is bound by its own capex
&lt;/h2&gt;&lt;p&gt;Samsung&amp;rsquo;s call was positive on demand. It said memory demand next year will &amp;ldquo;greatly exceed supply&amp;rdquo; and that it will start HBM4E mass production in the first half of 2027. &lt;a class="link" href="https://n.news.naver.com/mnews/article/008/0005433470" target="_blank" rel="noopener"
 &gt;Money Today via Naver&lt;/a&gt; Its third-quarter operating profit was &lt;strong&gt;107.4 trillion won&lt;/strong&gt; on revenue of 195.3 trillion won. The October 8 preliminary release had already shown 107.4 trillion won, so the profit was not news. &lt;a class="link" href="https://dart.fss.or.kr/dsaf001/main.do?rcpNo=20261029800087" target="_blank" rel="noopener"
 &gt;DART&lt;/a&gt; It also said lower dollar and euro exchange rates cut operating profit by about 10 trillion won from the previous quarter, mostly in components.&lt;/p&gt;
&lt;p&gt;The capex plan is the link to the payout. &lt;strong&gt;Samsung expects 84.4 trillion won of facility investment in 2026, 31.7 trillion won more than in 2025, with 76.8 trillion won in its chip division.&lt;/strong&gt; &lt;a class="link" href="https://dart.fss.or.kr/dsaf001/main.do?rcpNo=20261029800102" target="_blank" rel="noopener"
 &gt;DART&lt;/a&gt;, &lt;a class="link" href="https://n.news.naver.com/mnews/article/277/0005835135" target="_blank" rel="noopener"
 &gt;Asia Economy via Naver&lt;/a&gt; Samsung&amp;rsquo;s policy returns 50% of its 2024–2026 free cash flow. That free cash flow excludes prepayments under long-term memory supply agreements, which Samsung treats as deposits. In August, Samsung put the remaining pool for 2026 at about 90–110 trillion won. &lt;a class="link" href="https://dart.fss.or.kr/dsaf001/main.do?rcpNo=20260821800763" target="_blank" rel="noopener"
 &gt;DART, Aug 21&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;Today Samsung declared the 30 trillion won dividend that it had planned in August, at 4,601 won per share. On the call, it said about 35 trillion won of returns had been executed or confirmed, which leaves &lt;strong&gt;55–75 trillion won&lt;/strong&gt;. Samsung said a large part will go to buying back and cancelling shares. The final size will be set at the board meeting in late January 2027. &lt;a class="link" href="https://n.news.naver.com/mnews/article/014/0005596321" target="_blank" rel="noopener"
 &gt;Financial News via Naver&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;The arithmetic is unchanged from August, and the call did not raise the range after a record quarter. The late-January date for the final decision was already in the August disclosure. An industry source told Herald Economy that it still appeared to disappoint investors who wanted an earlier commitment. &lt;a class="link" href="https://n.news.naver.com/mnews/article/016/0002715932" target="_blank" rel="noopener"
 &gt;Herald Economy via Naver&lt;/a&gt; We cover the buyback calendar and the preferred-share question in &lt;a class="link" href="https://koreainvestinsights.com/post/samsung-buyback-calendar-preferred-shares-2026-10-29/" &gt;a separate post&lt;/a&gt;. A higher capex plan lowers the free cash flow base. Customer prepayments, which pay for new capacity, are excluded from it. SK hynix had announced about 100 trillion won of returns for this year the day before, including at least 30 trillion won of cash dividends and at least 60 trillion won of buybacks and cancellation. &lt;a class="link" href="https://www.newspim.com/news/view/20261028000155" target="_blank" rel="noopener"
 &gt;Newspim&lt;/a&gt; Investors compared the two this morning.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;For Korean memory, the capex boom is good for earnings and a constraint on payouts.&lt;/strong&gt; The question for Samsung&amp;rsquo;s stock has moved from whether AI demand lasts to how much of the cash reaches shareholders.&lt;/p&gt;
&lt;h2 id="our-judgment-and-what-would-change-it"&gt;Our judgment, and what would change it
&lt;/h2&gt;&lt;p&gt;&lt;strong&gt;Cash now ranks ahead of capex as the variable that sets AI-builder valuations.&lt;/strong&gt; The quarter&amp;rsquo;s shortfall came from late deliveries. Demand indicators kept rising, and part of the spending growth is higher memory prices. The pressure now falls on funding. Builders with negative free cash flow, such as Alphabet, are issuing new equity and debt while the 10-year yield is at 5%. In Seoul, the same pressure appeared as a payout cap tied to Samsung&amp;rsquo;s own capex.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Evidence&lt;/th&gt;
 &lt;th&gt;What it would mean&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Amazon (after the US close on Oct 29) raises capex while free cash flow stays near zero&lt;/td&gt;
 &lt;td&gt;Spending continues; the funding test gets sharper&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Alphabet&amp;rsquo;s fourth-quarter call gives 2027 capex growth well below 2026&amp;rsquo;s&lt;/td&gt;
 &lt;td&gt;Builders are rationing; the first real slowdown signal&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Microsoft&amp;rsquo;s December-quarter capex comes in below $50bn&lt;/td&gt;
 &lt;td&gt;The delay is more than timing&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Oracle exceeds its $20bn funding cap or its credit spreads widen sharply&lt;/td&gt;
 &lt;td&gt;Credit becomes the binding constraint&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Google Cloud backlog or Microsoft&amp;rsquo;s commercial RPO falls&lt;/td&gt;
 &lt;td&gt;A supply delay is turning into a demand delay&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Samsung&amp;rsquo;s January board raises total 2026 returns above 110 trillion won&lt;/td&gt;
 &lt;td&gt;The payout constraint loosens&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Foreign investors turn net sellers of Korean memory stocks&lt;/td&gt;
 &lt;td&gt;The Korean move becomes an AI-demand question&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;The next checks are Amazon&amp;rsquo;s results tonight, the KOSPI close and foreign flows, and Samsung&amp;rsquo;s January board decision.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Do you think markets are right to reward cash over capex growth, or is Alphabet&amp;rsquo;s turn to outside funding the price of staying ahead? Tell us which number would change your mind.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Disclosure: Alphabet, Amazon and Microsoft are investors in Anthropic, whose Claude models assisted research and editing for this post.&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Research and information only; not investment advice. Estimates and scenarios are analytical assumptions, not company guidance, trade instructions or personalized portfolio decisions. Figures are date-bound. AI tools assisted research and editing.&lt;/em&gt;&lt;/p&gt;</description></item><item><title>Samsung Has the Cash. Investors Are Waiting for a Buyback Date.</title><link>https://koreainvestinsights.com/post/samsung-buyback-calendar-preferred-shares-2026-10-29/</link><pubDate>Thu, 29 Oct 2026 11:15:00 +0900</pubDate><guid>https://koreainvestinsights.com/post/samsung-buyback-calendar-preferred-shares-2026-10-29/</guid><description>&lt;img src="https://koreainvestinsights.com/" alt="Featured image of post Samsung Has the Cash. Investors Are Waiting for a Buyback Date." /&gt;&lt;p&gt;Samsung Electronics ended Q3 with &lt;strong&gt;KRW197.6 trillion in net cash&lt;/strong&gt;. Its common shares were still down more than 5% on the morning of October 29. Cash on a balance sheet and a buy order in the market are different things.&lt;/p&gt;
&lt;p&gt;My leading interpretation is that the gap between expectations and the timing of execution mattered more than Samsung&amp;rsquo;s ability to fund returns. That is a hypothesis, not a measured decomposition of the share-price decline. Preferred shares falling less is consistent with interest in a larger preferred-share buyback allocation. It does not prove the cause.&lt;/p&gt;
&lt;p&gt;One correction comes first: &lt;strong&gt;Samsung did not newly postpone its remaining return decision to January today.&lt;/strong&gt; The company had already specified late January 2027 in its August 21 disclosure. Investors hoping for an earlier commitment did not receive one in the announced plan. &lt;a class="link" href="https://www.samsung.com/sec/ir/reports-disclosures/public-disclosure/4047/" target="_blank" rel="noopener"
 &gt;August disclosure&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Dated October 29, 2026. Prices below are Naver&amp;rsquo;s KRX intraday observations around 11:07 KST, not closing prices. Disclosed facts, observed prices and analytical interpretations are kept separate.&lt;/em&gt;&lt;/p&gt;
&lt;h2 id="krw5575-trillion-is-not-a-buy-order"&gt;KRW55–75 trillion is not a buy order
&lt;/h2&gt;&lt;p&gt;Samsung described an estimated KRW90–110 trillion in 2026 returns, less roughly KRW35 trillion already planned or completed, leaving KRW55–75 trillion. The remainder includes dividends and substantial repurchases and cancellations. It is not an all-buyback budget. Final amounts, execution plans and the next return policy are scheduled for the late-January board meeting. &lt;a class="link" href="https://images.samsung.com/is/content/samsung/assets/global/ir/docs/2026_3Q_conference_eng.pdf" target="_blank" rel="noopener"
 &gt;Official Q3 IR, final slide&lt;/a&gt;&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Item&lt;/th&gt;
 &lt;th&gt;Amount or date&lt;/th&gt;
 &lt;th&gt;What it means&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Estimated 2026 returns&lt;/td&gt;
 &lt;td&gt;KRW90–110tn&lt;/td&gt;
 &lt;td&gt;Subject to results, investment and final calculation&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Planned or completed&lt;/td&gt;
 &lt;td&gt;About KRW35tn&lt;/td&gt;
 &lt;td&gt;Includes first-half regular dividends and the Q3 dividend&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Remaining pool&lt;/td&gt;
 &lt;td&gt;KRW55–75tn&lt;/td&gt;
 &lt;td&gt;Dividend/buyback allocation is not final&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Final decision&lt;/td&gt;
 &lt;td&gt;Late January 2027&lt;/td&gt;
 &lt;td&gt;A decision date is not a purchase start date&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Some pre-announcement coverage used KRW110 trillion minus KRW30 trillion to arrive at KRW80 trillion. On the company&amp;rsquo;s basis, the comparable upper-end calculation is 110 minus 35, or 75. That is a KRW5 trillion difference. Comparing the headline&amp;rsquo;s 80 with the company&amp;rsquo;s midpoint of 65 produces a 15 trillion gap, but &lt;strong&gt;also switches from an upper bound to a midpoint&lt;/strong&gt;. It is not evidence of a 15 trillion cut in a previously committed payout. Nor is one headline a measured market consensus. &lt;a class="link" href="https://www.sentv.co.kr/article/view/sentv202610270088" target="_blank" rel="noopener"
 &gt;Pre-announcement coverage&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;The lower end of 55 is also not a guaranteed minimum payment. Samsung&amp;rsquo;s policy returns 50% of cumulative 2024–2026 free cash flow, with the final pool dependent on results and investment. &lt;a class="link" href="https://www.samsung.com/sec/ir/reports-disclosures/public-disclosure/4047/" target="_blank" rel="noopener"
 &gt;Official calculation conditions&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;The company&amp;rsquo;s calculation also deducts deposit-like memory LTA advances and employee performance-related stock-compensation spending. Half of an increase in cash is therefore not automatically available for distribution under this policy.&lt;/p&gt;
&lt;h2 id="the-krw15-trillion-program-ended-all-future-buying-is-not-prohibited"&gt;The KRW15 trillion program ended. All future buying is not prohibited.
&lt;/h2&gt;&lt;p&gt;The completed employee-compensation program is now verifiable beyond a news headline. Samsung bought 56,856,305 common shares between August 24 and October 6 for KRW14,999,999,899,750, effectively KRW15 trillion. &lt;a class="link" href="https://www.samsung.com/sec/ir/reports-disclosures/public-disclosure/4072/" target="_blank" rel="noopener"
 &gt;Official acquisition result&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;Those purchases created market demand. Shares bought for employee compensation can subsequently return to circulation, however. They are not equivalent to permanent cancellation and should not automatically be counted as the same shareholder-return effect.&lt;/p&gt;
&lt;p&gt;The verified conclusion is narrower than “Samsung will not buy any shares until January”: &lt;strong&gt;that program is complete, and the next purchase schedule for the remaining return pool has not been specified.&lt;/strong&gt; A separate compensation program or a new board resolution could change the picture.&lt;/p&gt;
&lt;p&gt;SK hynix is at a different execution stage. It approved roughly KRW20 trillion of repurchases on October 28, with an acquisition window from October 29 to January 28, 2027. About half is for shareholder-return cancellation and half for employee compensation. An open acquisition window does not establish actual first-day purchases. &lt;a class="link" href="https://dart.fss.or.kr/dsaf001/main.do?rcpNo=20261028000014" target="_blank" rel="noopener"
 &gt;SK hynix acquisition filing&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;The comparison is therefore about the next identifiable buyer: one company has a specified acquisition period and amount; the other has a large remaining pool with its detailed plan still ahead.&lt;/p&gt;
&lt;h2 id="preferred-share-allocation-has-a-governance-dimension"&gt;Preferred-share allocation has a governance dimension
&lt;/h2&gt;&lt;p&gt;Samsung explicitly linked its willingness to consider and adjust the preferred-share allocation to the common/preferred price gap and market conditions. It did not commit to a numerical allocation. &lt;a class="link" href="https://images.samsung.com/is/content/samsung/assets/global/ir/docs/2026_3Q_conference_eng.pdf" target="_blank" rel="noopener"
 &gt;Official IR&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;There is another reason to watch that allocation. Samsung Life and Samsung Fire &amp;amp; Marine&amp;rsquo;s Samsung Electronics stakes sit within Korea&amp;rsquo;s rules governing financial institutions&amp;rsquo; ownership of other companies. In March 2026, the two insurers announced roughly KRW1.5 trillion of disposals and described them as a preemptive response to regulatory risk. &lt;a class="link" href="https://www.yna.co.kr/amp/view/AKR20260319169952002" target="_blank" rel="noopener"
 &gt;Contemporaneous report&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;The legal framework should not be compressed into a universal “10% is forbidden” rule. Article 24 establishes approval requirements tied to voting-share ownership and control. The 10% threshold discussed in Samsung reporting needs to be distinguished from the statute&amp;rsquo;s general tests. The legal denominator at repurchase and cancellation should not be assumed identical. &lt;a class="link" href="https://www.law.go.kr/LSW/lsLinkCommonInfo.do?chrClsCd=010202&amp;amp;lsJoLnkSeq=1029473049" target="_blank" rel="noopener"
 &gt;Official Article 24&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;There is relevant counterevidence. In 2017, Reuters reported a Financial Services Commission official explaining that treasury shares were already excluded from the voting-share denominator, so cancellation did not change the ratio. I have not verified whether the detailed calculation rules subsequently changed. A formula in which cancellation mechanically forces insurer selling would therefore overstate what has been established here. &lt;a class="link" href="https://kr.investing.com/news/forex-news/article-67226" target="_blank" rel="noopener"
 &gt;Historical regulatory explanation&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;Preferred shares are non-voting, and Samsung has left room to adjust their allocation. Treating that as a confirmed design to solve a governance constraint goes further than the evidence. &lt;strong&gt;I have not verified a company statement saying this law caused the choice of January.&lt;/strong&gt; Completing the annual FCF calculation and announcing the next return policy together are alternative explanations.&lt;/p&gt;
&lt;h2 id="relative-resilience-is-a-clue-not-proof"&gt;Relative resilience is a clue, not proof
&lt;/h2&gt;&lt;p&gt;The following observations use one provider and market, with source timestamps only three seconds apart.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Security&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Price, KRW&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Day change&lt;/th&gt;
 &lt;th&gt;Source time, KST&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Samsung common&lt;/td&gt;
 &lt;td style="text-align: right"&gt;229,500&lt;/td&gt;
 &lt;td style="text-align: right"&gt;−5.17%&lt;/td&gt;
 &lt;td&gt;11:07:21&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Samsung preferred&lt;/td&gt;
 &lt;td style="text-align: right"&gt;184,500&lt;/td&gt;
 &lt;td style="text-align: right"&gt;−2.95%&lt;/td&gt;
 &lt;td&gt;11:07:18&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;SK hynix&lt;/td&gt;
 &lt;td style="text-align: right"&gt;1,514,000&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+1.82%&lt;/td&gt;
 &lt;td&gt;11:07:20&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Sources: Naver public APIs for &lt;a class="link" href="https://m.stock.naver.com/api/stock/005930/basic" target="_blank" rel="noopener"
 &gt;Samsung common&lt;/a&gt;, &lt;a class="link" href="https://m.stock.naver.com/api/stock/005935/basic" target="_blank" rel="noopener"
 &gt;Samsung preferred&lt;/a&gt; and &lt;a class="link" href="https://m.stock.naver.com/api/stock/000660/basic" target="_blank" rel="noopener"
 &gt;SK hynix&lt;/a&gt;. KRX intraday observations.&lt;/p&gt;
&lt;p&gt;Preferred shares fell 2.22 percentage points less than common shares. That aligns with the preferred-allocation message, but the snapshot cannot identify investor motives.&lt;/p&gt;
&lt;p&gt;Common shares had also risen about 9.0% the previous day, from KRW222,000 to KRW242,000. Return expectations, reversal of that rally, broad selling and interpretations of divisional earnings can all matter. Without an identified attribution model, ranking the missing buyback catalyst first remains my working judgment.&lt;/p&gt;
&lt;h2 id="the-196-discount-is-not-a-dividend-waiting-to-be-collected"&gt;The 19.6% discount is not a dividend waiting to be collected
&lt;/h2&gt;&lt;p&gt;At the quoted prices, preferred shares traded at a 19.6% discount. The declared Q3 dividend is identical for both classes at KRW4,601 per share.&lt;/p&gt;
&lt;p&gt;&lt;img alt="Declared quarterly dividend divided by intraday price: common 2.00%, preferred 2.49%. The record date has already passed." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/samsung-buyback-calendar-preferred-shares-2026-10-29/dividend.en.png"&gt;
&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Calculation&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Result&lt;/th&gt;
 &lt;th&gt;Interpretation&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;1 − 184,500 / 229,500&lt;/td&gt;
 &lt;td style="text-align: right"&gt;19.61%&lt;/td&gt;
 &lt;td&gt;Preferred price discount&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;4,601 / 229,500&lt;/td&gt;
 &lt;td style="text-align: right"&gt;2.00%&lt;/td&gt;
 &lt;td&gt;Declared quarterly dividend / common price&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;4,601 / 184,500&lt;/td&gt;
 &lt;td style="text-align: right"&gt;2.49%&lt;/td&gt;
 &lt;td&gt;Declared quarterly dividend / preferred price&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;&lt;strong&gt;The record date was September 30. Buying today does not entitle the buyer to this dividend.&lt;/strong&gt; These ratios compare prices; they are not annual yields or forecasts of future payouts. &lt;a class="link" href="https://dart.fss.or.kr/dsaf001/main.do?rcpNo=20261029800101" target="_blank" rel="noopener"
 &gt;Dividend resolution&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;A narrower discount also means less remaining relative-value upside under a fixed common-price assumption. If common stays at KRW229,500 and the discount narrows to 15%, preferred would be KRW195,075, or 5.73% above the snapshot. At a 25% discount, it would be KRW172,125, down 6.71%. These are sensitivities, not targets or expected returns. Preferred can lose money even as the discount narrows if common shares fall enough.&lt;/p&gt;
&lt;h2 id="watch-the-purpose-and-the-first-purchase-date"&gt;Watch the purpose and the first purchase date
&lt;/h2&gt;&lt;p&gt;An unresolved return plan alone would not make me downgrade the memory-demand outlook. Samsung disclosed about KRW84.4 trillion of 2026 capital spending and expects AI demand to exceed supply in 2027. Those remain spending plans and company forecasts. Capital investment consumes cash, and a large net-cash balance is not the same as unrestricted distributable cash. &lt;a class="link" href="https://dart.fss.or.kr/dsaf001/main.do?rcpNo=20261029800102" target="_blank" rel="noopener"
 &gt;Capex filing&lt;/a&gt;, &lt;a class="link" href="https://news.samsung.com/kr/%ec%82%bc%ec%84%b1%ec%a0%84%ec%9e%90-2026%eb%85%84-3%eb%b6%84%ea%b8%b0-%ec%8b%a4%ec%a0%81-%eb%b0%9c%ed%91%9c" target="_blank" rel="noopener"
 &gt;official business outlook&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;Four disclosures could change the judgment:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;A new repurchase or trust agreement before January would require reassessing the distant-execution hypothesis.&lt;/li&gt;
&lt;li&gt;Its purpose, acquisition dates and share count would distinguish cancellation from compensation.&lt;/li&gt;
&lt;li&gt;New insurer disposals could inform the common-share return design, without by themselves proving imminent cancellation.&lt;/li&gt;
&lt;li&gt;The final January split across common, preferred and dividends would determine the demand directed at each security.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;I have not verified an official source for May media claims of approximately KRW80 trillion in employee-share purchases over three years, so that figure is excluded from projected demand. I have also not obtained a complete Q&amp;amp;A transcript. Regulatory causation and the ranking of selling drivers remain analytical hypotheses, not company explanations.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;If you could see just one disclosure before January, which would change your view most: the first purchase date, the preferred-share allocation, or an insurer stake sale?&lt;/strong&gt; Leave your choice and the reasoning in the comments.&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Public-source investment analysis, not an individualized trading recommendation. Prices and policy changes can alter the conclusion.&lt;/em&gt;&lt;/p&gt;</description></item><item><title>Microsoft Spent $5 Billion Less Than It Planned and Grew Azure Faster. The Next Capex Wave Runs Through CPUs and Memory.</title><link>https://koreainvestinsights.com/post/microsoft-q1-fy27-azure-capex-cpu-memory-korea-readthrough-2026-10-29/</link><pubDate>Thu, 29 Oct 2026 09:10:00 +0900</pubDate><guid>https://koreainvestinsights.com/post/microsoft-q1-fy27-azure-capex-cpu-memory-korea-readthrough-2026-10-29/</guid><description>&lt;img src="https://koreainvestinsights.com/" alt="Featured image of post Microsoft Spent $5 Billion Less Than It Planned and Grew Azure Faster. The Next Capex Wave Runs Through CPUs and Memory." /&gt;&lt;p&gt;Microsoft reported its fiscal first quarter, which ended September 30, after the US close on October 28. Revenue was &lt;strong&gt;$92.3 billion, up 19%&lt;/strong&gt;. Azure, disclosed in dollars for the first time, was &lt;strong&gt;$33.2 billion, up 48%&lt;/strong&gt;. Capital spending including finance leases was &lt;strong&gt;$45.2 billion&lt;/strong&gt;. In July, Microsoft had guided more than $50 billion. &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/789019/000119312526436133/msft-20260930.htm" target="_blank" rel="noopener"
 &gt;Microsoft 10-Q&lt;/a&gt;, &lt;a class="link" href="https://www.cnbc.com/2026/10/28/microsoft-msft-q1-earnings-report-2027.html" target="_blank" rel="noopener"
 &gt;CNBC, Oct 28&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;The shortfall was not a decision to spend less. CFO Amy Hood said some third-party capacity deliveries expected in the quarter were delayed. Microsoft said customer demand still exceeds supply. In extended trading the shares rose about 3–4%. CNBC reported about 4%, and a later Nasdaq after-hours print was about $545.60.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Our judgment: AI demand at Microsoft is limited by delivered capacity, not by budget, and the next leg of spending runs through CPUs and memory.&lt;/strong&gt; Hood said second-half capex will accelerate mainly because of CPUs, including higher memory and storage costs, and GPUs. For Korean memory makers, that is a better read-through than another GPU-only wave. Conventional server DRAM and enterprise SSDs gain alongside HBM. &lt;strong&gt;The variable that now decides the 2027 memory cycle is the buyers&amp;rsquo; cash flow, not their demand.&lt;/strong&gt; Microsoft can pay. Not every hyperscaler can.&lt;/p&gt;
&lt;p&gt;This article is dated October 29, 2026. Microsoft&amp;rsquo;s fiscal year ends on June 30. Call quotes come from a preliminary third-party transcript. &lt;a class="link" href="https://news.alphastreet.com/microsoft-corporation-msft-q1-2027-earnings-call-transcript/" target="_blank" rel="noopener"
 &gt;AlphaStreet transcript&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;&lt;img alt="Microsoft capex by quarter. Cash paid for property and equipment rose from $19.4 billion in the July–September 2025 quarter to $39.4 billion a year later. Capex including finance leases rose from $34.9 billion to $45.2 billion. Short-lived assets were about half of capex a year ago and about three quarters now. The guide for the December quarter is above $50 billion." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/microsoft-q1-fy27-azure-capex-cpu-memory-korea-readthrough-2026-10-29/capex-mix.en.png"&gt;
&lt;/p&gt;
&lt;h2 id="five-things-to-take-away"&gt;Five things to take away
&lt;/h2&gt;&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Azure grew faster than capex.&lt;/strong&gt; Azure revenue rose by $10.8 billion from a year earlier. Capex including finance leases rose by $10.3 billion. Azure grew 48%; capex grew about 30%.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;The server share of capex jumped.&lt;/strong&gt; About three quarters of capex went to short-lived assets such as servers, CPUs and GPUs, against about half a year earlier. That puts short-lived spending at roughly $34 billion, about double the $17 billion of a year ago.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Memory is now named as a cost driver.&lt;/strong&gt; Hood said short-lived assets include higher memory and storage costs, and that the second-half acceleration is led by CPUs. Amazon said in July that most of its $20 billion capex increase reflected memory costs.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Microsoft also leases at scale.&lt;/strong&gt; It has &lt;strong&gt;$364.6 billion of additional leases, mostly for datacenters, that have not yet commenced&lt;/strong&gt;, more than Oracle&amp;rsquo;s $288 billion. &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/0001341439/000119312526389274/orcl-20260831.htm" target="_blank" rel="noopener"
 &gt;Oracle 10-Q&lt;/a&gt; Its headline capex includes finance leases but not operating leases.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;OpenAI is still a large share of the business.&lt;/strong&gt; Microsoft recorded &lt;strong&gt;$10.0 billion of revenue from OpenAI&lt;/strong&gt; in the quarter, about 11% of total revenue. That is a related-party disclosure in the 10-Q.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id="the-quarter-in-numbers"&gt;The quarter in numbers
&lt;/h2&gt;&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Item&lt;/th&gt;
 &lt;th&gt;Q1 FY27&lt;/th&gt;
 &lt;th&gt;Change&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Revenue&lt;/td&gt;
 &lt;td&gt;$92.28bn&lt;/td&gt;
 &lt;td&gt;+19%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Operating income&lt;/td&gt;
 &lt;td&gt;$45.92bn&lt;/td&gt;
 &lt;td&gt;+21%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Non-GAAP EPS&lt;/td&gt;
 &lt;td&gt;$4.89 (consensus about $4.73)&lt;/td&gt;
 &lt;td&gt;+18%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Azure revenue&lt;/td&gt;
 &lt;td&gt;$33.17bn&lt;/td&gt;
 &lt;td&gt;+48%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Microsoft Cloud revenue / gross margin&lt;/td&gt;
 &lt;td&gt;$63.9bn / 66%&lt;/td&gt;
 &lt;td&gt;+30% / lower&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Commercial remaining performance obligation&lt;/td&gt;
 &lt;td&gt;$677bn, about 30% due in 12 months&lt;/td&gt;
 &lt;td&gt;+73%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Revenue from OpenAI, including revenue sharing&lt;/td&gt;
 &lt;td&gt;$10.0bn&lt;/td&gt;
 &lt;td&gt;related-party disclosure&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Cash paid for property and equipment&lt;/td&gt;
 &lt;td&gt;$39.37bn&lt;/td&gt;
 &lt;td&gt;+103%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Capex including finance leases&lt;/td&gt;
 &lt;td&gt;$45.2bn (guide: more than $50bn)&lt;/td&gt;
 &lt;td&gt;about +30%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Operating cash flow / free cash flow&lt;/td&gt;
 &lt;td&gt;$58.0bn / $18.6bn&lt;/td&gt;
 &lt;td&gt;+29% / −27%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Windows OEM and devices revenue&lt;/td&gt;
 &lt;td&gt;$3.55bn&lt;/td&gt;
 &lt;td&gt;−22%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Sources: &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/789019/000119312526436133/msft-20260930.htm" target="_blank" rel="noopener"
 &gt;Microsoft 10-Q&lt;/a&gt;, &lt;a class="link" href="https://news.microsoft.com/source/2026/10/28/microsoft-announces-first-quarter-results-2/" target="_blank" rel="noopener"
 &gt;Microsoft release&lt;/a&gt;, &lt;a class="link" href="https://www.cnbc.com/2026/10/28/microsoft-msft-q1-earnings-report-2027.html" target="_blank" rel="noopener"
 &gt;CNBC&lt;/a&gt;. Free cash flow is operating cash flow minus cash paid for property and equipment.&lt;/p&gt;
&lt;p&gt;Microsoft also changed its segments this year. It now reports Agents and Infra, with revenue of $76.8 billion, up 25%, and Devices and Consumer, with $15.5 billion, down 3%. Older segment comparisons no longer apply.&lt;/p&gt;
&lt;h2 id="ai-demand-capacity-limited-broader-and-still-concentrated"&gt;AI demand: capacity-limited, broader, and still concentrated
&lt;/h2&gt;&lt;p&gt;Three facts describe demand at Microsoft today.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Demand exceeds supply, and the bottleneck is delivery.&lt;/strong&gt; Microsoft fell about $5 billion short of its capex guide because third-party GPU capacity arrived late. The same delay pushes some third-party GPU-as-a-service revenue into the second half. Azure still grew 48%, helped by better GPU fleet efficiency, more use of idle CPU capacity and higher revenue per token. Hood said token pricing in the quarter was higher than Microsoft expected. For the December quarter, Microsoft guided Azure growth of &lt;strong&gt;49–50% in constant currency&lt;/strong&gt;, above the 45% consensus that CNBC cited.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Usage is broadening beyond chat.&lt;/strong&gt; Nadella said Microsoft 365 Copilot daily users rose nearly 50% from the previous quarter and more than fivefold from a year earlier. Customers consuming Copilot credits rose more than 75% in the quarter. Foundry has more than 120,000 customers, and its revenue more than tripled. On GitHub, agents now open more than a third of pull requests. Microsoft did not update its paid Copilot seat count, tokens processed or an AI revenue run rate.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Concentration has not disappeared.&lt;/strong&gt; Revenue from OpenAI was $10.0 billion, about 11% of the total. For scale, that equals about 30% of Azure revenue, but the two figures are not directly comparable, because Microsoft does not say how much of the OpenAI revenue is Azure. Receivables from OpenAI rose to $7.2 billion from $6.0 billion in June. Commercial bookings fell 12%, because the year-earlier quarter included very large Azure commitments. Remaining performance obligation still rose 73% to $677 billion. Nadella said Microsoft wants breadth. It may sell capacity to large buyers if the pricing is right. If not, it will be patient and build the breadth business.&lt;/p&gt;
&lt;p&gt;Our reading: the demand signal is strong, and its quality is improving. The share coming from enterprise agents and developers is rising. The share that depends on one model company is still large enough to matter.&lt;/p&gt;
&lt;h2 id="the-capex-path-about-170-billion-this-calendar-year-then-faster"&gt;The capex path: about $170 billion this calendar year, then faster
&lt;/h2&gt;&lt;p&gt;Hood guided capex above $50 billion for the December quarter and about $170 billion for &amp;ldquo;the calendar year&amp;rdquo;. CNBC read this as calendar 2027. We read it as calendar 2026. In July, Microsoft guided calendar-2026 capex to about $175 billion, down from $190 billion after an accounting change to building lives. &lt;a class="link" href="https://www.cfodive.com/news/microsoft-holds-line-ai-spending-plans/826648/" target="_blank" rel="noopener"
 &gt;CFO Dive, Jul 30&lt;/a&gt; On this call, Hood said &amp;ldquo;therefore&amp;rdquo; right after the December-quarter figure and gave the third-party delays as the cause of the lower number. The reported quarters from January to September 2026, about $31.9 billion, $41 billion and $45.2 billion including finance leases, plus a fourth quarter above $50 billion, add up to at least about $168 billion. A calendar-2027 total of $170 billion would also conflict with her guidance that spending accelerates from a base above $50 billion a quarter. The preliminary transcript does not name the year.&lt;/p&gt;
&lt;p&gt;The composition matters more than the total. A year ago, about half of Microsoft&amp;rsquo;s capex was short-lived: servers, CPUs, GPUs and network gear. Now it is about three quarters. Land and buildings are moving into leases. Microsoft&amp;rsquo;s &lt;strong&gt;$364.6 billion of leases not yet commenced&lt;/strong&gt; start between fiscal 2027 and fiscal 2034, with terms of up to 20 years. In the quarter, new operating-lease assets were $6.5 billion, against $1.4 billion a year earlier. Those operating leases are not in the $45.2 billion.&lt;/p&gt;
&lt;p&gt;In other words, Microsoft&amp;rsquo;s own cash is increasingly going into the equipment inside the building. That is where memory sits.&lt;/p&gt;
&lt;p&gt;&lt;img alt="Calendar 2026 capex guidance. Amazon about $220 billion, raised in July with higher memory costs cited. Alphabet $195–205 billion. Microsoft about $170 billion for the calendar year, on our reading. Meta $135–145 billion. The midpoints add up to about $730 billion, although the companies define capex differently." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/microsoft-q1-fy27-azure-capex-cpu-memory-korea-readthrough-2026-10-29/hyperscaler-capex-2026.en.png"&gt;
&lt;/p&gt;
&lt;p&gt;The peer picture points the same way. Alphabet reported third-quarter capex of $51.4 billion on October 28, kept its 2026 range of $195–205 billion and said capex will increase significantly in 2027. Its shares fell more than 10% after hours. &lt;a class="link" href="https://www.cnbc.com/2026/10/28/alphabet-googl-earnings-q3-2026.html" target="_blank" rel="noopener"
 &gt;CNBC, Oct 28&lt;/a&gt; Meta narrowed its 2026 range to $135–145 billion on October 21. &lt;a class="link" href="https://www.fool.com/earnings/call-transcripts/2026/10/21/meta-meta-q3-2026-earnings-call-transcript/" target="_blank" rel="noopener"
 &gt;Meta call transcript&lt;/a&gt; Amazon raised its 2026 plan to about $220 billion in July and attributed most of the increase to memory costs. &lt;a class="link" href="https://datacenterknowledge.com/infrastructure/amazon-lifts-ai-infrastructure-spending-to-220b-as-demand-outpaces-capacity" target="_blank" rel="noopener"
 &gt;Data Center Knowledge&lt;/a&gt; Amazon is scheduled to report after the US close on October 29.&lt;/p&gt;
&lt;h2 id="why-cpus-matter-for-korean-memory"&gt;Why CPUs matter for Korean memory
&lt;/h2&gt;&lt;p&gt;The market has treated AI memory as an HBM story. Microsoft&amp;rsquo;s guidance adds a second story.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Agents run on CPUs as well as GPUs.&lt;/strong&gt; An agent that calls tools, searches, writes code and checks results spends much of its time on general-purpose servers. Microsoft said it is already monetizing idle CPU capacity and that second-half capex accelerates first because of CPUs. TrendForce made a related point on September 30: better server CPU availability is leading cloud providers and server makers to buy more general-purpose servers. &lt;a class="link" href="https://www.trendforce.com/presscenter/news/20260930-13258.html" target="_blank" rel="noopener"
 &gt;TrendForce, Sep 30&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Each general-purpose server carries conventional DRAM and SSDs.&lt;/strong&gt; These are DDR5 server modules and enterprise SSDs, not HBM. TrendForce expects the pace of contract price increases to moderate, with conventional DRAM up 10–15% in the fourth quarter and NAND up 15–20%. Enterprise SSDs are the only category where increases are accelerating. It also says server DRAM remains undersupplied.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Memory is inflating capex itself.&lt;/strong&gt; Microsoft named higher memory and storage costs inside its capex. Amazon attributed most of a $20 billion increase to memory. In April, before Amazon&amp;rsquo;s raise and Microsoft&amp;rsquo;s latest comments, a SemiAnalysis estimate reported by Tom&amp;rsquo;s Hardware put memory at about 30% of total hyperscaler capex in 2026. &lt;a class="link" href="https://www.tomshardware.com/tech-industry/memory-will-consume-30-percent-of-hyperscaler-spending-this-year" target="_blank" rel="noopener"
 &gt;Tom&amp;rsquo;s Hardware&lt;/a&gt; This is the mechanism behind SK hynix&amp;rsquo;s 80% operating margin last quarter. &lt;a class="link" href="https://koreainvestinsights.com/post/sk-hynix-80-percent-margin-contract-price-band-2026-10-29/" &gt;KII, Oct 29&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;For Korean suppliers, the implication is specific.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Korean exposure&lt;/th&gt;
 &lt;th&gt;What Microsoft&amp;rsquo;s quarter implies&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;HBM (SK hynix, Samsung Electronics)&lt;/td&gt;
 &lt;td&gt;GPU capacity is still constrained and delayed into the second half; demand for accelerators stays strong&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Conventional server DRAM (Samsung Electronics, SK hynix)&lt;/td&gt;
 &lt;td&gt;A CPU-led second half adds a second demand leg on top of HBM; contract prices are still rising&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Enterprise SSD (SK hynix via Solidigm, Samsung Electronics)&lt;/td&gt;
 &lt;td&gt;Storage costs are named in Microsoft&amp;rsquo;s capex; enterprise SSD is the category where price increases are accelerating&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Long-term agreements&lt;/td&gt;
 &lt;td&gt;SK hynix says agreements with 14 key customers cover more than half of its 2027–2028 capacity&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Sources: Microsoft call; &lt;a class="link" href="https://www.trendforce.com/presscenter/news/20260930-13258.html" target="_blank" rel="noopener"
 &gt;TrendForce, Sep 30&lt;/a&gt;; &lt;a class="link" href="https://thenextweb.com/news/sk-hynix-q3-2026-record-profit-memory-shortage-2030" target="_blank" rel="noopener"
 &gt;TheNextWeb on SK hynix Q3, Oct 28&lt;/a&gt;&lt;/p&gt;
&lt;h2 id="where-the-read-through-can-break"&gt;Where the read-through can break
&lt;/h2&gt;&lt;p&gt;Two costs of the memory boom are already visible inside Microsoft&amp;rsquo;s own numbers.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Devices are paying for memory.&lt;/strong&gt; Windows OEM and devices revenue fell 22%. CNBC tied the weakness to higher memory and storage costs that raised device prices. SK hynix itself flagged that PC and mobile shipments may fall because of memory costs. A memory cycle that is strong in servers and weak in consumer devices is more concentrated on a few cloud buyers.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;The buyers&amp;rsquo; cash flow is thinning.&lt;/strong&gt; Microsoft&amp;rsquo;s operating cash flow rose 29%, but free cash flow fell 27% to $18.6 billion. CNBC reported that Alphabet&amp;rsquo;s free cash flow was negative in the quarter and that Meta&amp;rsquo;s turned negative. Microsoft said it remains free-cash-flow positive this fiscal year, has not yet tapped the debt markets for additional funding, and will use other funding sources when they are attractive.&lt;/p&gt;
&lt;p&gt;This is where the memory thesis now lives. Demand for 2027 is well covered: SK hynix says its long-term agreements cover more than half of its capacity for 2027 and 2028, and on its call it cited market-research estimates that global cloud capex will grow more than 60% in 2027. The open question is how much more the buyers can pay when memory prices keep rising and their own free cash flow is near zero. Microsoft is the strongest payer in the group. Its acceleration is good news. A slowdown at the weakest payers would show up first in conventional DRAM and SSD orders, where long-term agreements are less common than in HBM.&lt;/p&gt;
&lt;h2 id="our-judgment-and-what-would-change-it"&gt;Our judgment, and what would change it
&lt;/h2&gt;&lt;p&gt;&lt;strong&gt;The read-through for Korean memory is positive for 2027, and it is broader than HBM.&lt;/strong&gt; Microsoft spent less than planned only because capacity arrived late, and it expects to spend faster in the second half on CPUs, GPUs and memory. &lt;strong&gt;The buyers&amp;rsquo; cash flow dominates the next stage of the cycle&lt;/strong&gt;, ahead of AI demand and supply. Demand is not the constraint today. Payment capacity could become one in 2027.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Evidence&lt;/th&gt;
 &lt;th&gt;What it would mean&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Microsoft delivers December-quarter capex above $50bn&lt;/td&gt;
 &lt;td&gt;The delay was timing; the second-half acceleration is real&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Amazon (Oct 29) and Alphabet (Q4 call) give 2027 capex growth of 30% or more&lt;/td&gt;
 &lt;td&gt;2027 memory demand is funded; conventional DRAM gains on top of HBM&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;TrendForce shows conventional DRAM contract prices still rising in 1Q27&lt;/td&gt;
 &lt;td&gt;The CPU leg is pulling through to memory prices&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Two or more large buyers guide 2027 capex growth below 20%&lt;/td&gt;
 &lt;td&gt;Payers are pushing back on memory-inflated capex&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Hyperscaler free cash flow negative for several quarters and new debt issuance rising&lt;/td&gt;
 &lt;td&gt;Payment capacity is the binding constraint; memory pricing power fades first in non-contract volumes&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Windows OEM and phone declines deepen&lt;/td&gt;
 &lt;td&gt;Consumer demand destruction narrows the memory cycle to a few cloud buyers&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;The next checks are Amazon&amp;rsquo;s results after the US close on October 29 and Samsung Electronics&amp;rsquo; full third-quarter results and call on the same day. Microsoft&amp;rsquo;s December-quarter capex, reported in late January, will show whether the delayed capacity arrived.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Do you see the CPU-led second half as a real second demand leg for conventional DRAM, or as a one-off catch-up of delayed deliveries? Tell us which number would convince you.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Research and information only; not investment advice. Estimates and scenarios are analytical assumptions, not company guidance, trade instructions or personalized portfolio decisions. Figures are date-bound. AI tools assisted research and editing.&lt;/em&gt;&lt;/p&gt;</description></item><item><title>Oracle's Cash Turn Came Early Because Customers and Landlords Now Pay First. OpenAI Decides Whether It Holds.</title><link>https://koreainvestinsights.com/post/oracle-fcf-turn-customers-landlords-pay-first-openai-test-2026-10-29/</link><pubDate>Thu, 29 Oct 2026 08:50:00 +0900</pubDate><guid>https://koreainvestinsights.com/post/oracle-fcf-turn-customers-landlords-pay-first-openai-test-2026-10-29/</guid><description>&lt;img src="https://koreainvestinsights.com/" alt="Featured image of post Oracle's Cash Turn Came Early Because Customers and Landlords Now Pay First. OpenAI Decides Whether It Holds." /&gt;&lt;p&gt;At its Investor Day on October 28, Oracle set a date for the end of its cash burn. It expects &lt;strong&gt;positive free cash flow in the second half of fiscal 2028&lt;/strong&gt;, the six months from December 2027 to May 2028, and &lt;strong&gt;at least $30 billion of free cash flow in fiscal 2029&lt;/strong&gt;. It capped new capital-market funding through fiscal 2030 at &lt;strong&gt;$20 billion&lt;/strong&gt;, currently expected to be $10 billion of equity and $10 billion of gross debt. &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1341439/000119312526435853/d75538dex991.htm" target="_blank" rel="noopener"
 &gt;Oracle Investor Day release&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;Before the event, analysts expected free cash flow of about &lt;strong&gt;minus $35 billion in both fiscal 2027 and fiscal 2028&lt;/strong&gt;, according to FactSet figures cited by CNBC. &lt;a class="link" href="https://www.cnbc.com/2026/10/28/oracle-forecast-growth-2030.html" target="_blank" rel="noopener"
 &gt;CNBC, Oct 28&lt;/a&gt; The stock rose &lt;strong&gt;10.7% to $144.54&lt;/strong&gt;. It is still about 49% below its 52-week high of $281.62.&lt;/p&gt;
&lt;p&gt;The headline revenue target was not the news. Oracle&amp;rsquo;s fiscal 2030 goal of at least $225 billion matched the LSEG consensus of $224.2 billion. The news was the cash.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Our judgment: the cash turn is realistic on Oracle&amp;rsquo;s own definition, and it came early because Oracle stopped paying first.&lt;/strong&gt; The upfront cost is now carried by customer prepayments, customer-supplied hardware and landlord-built data centers, and Oracle expects supplier credit to add to them. That moves most of the upfront cost off Oracle&amp;rsquo;s cash-flow statement. It does not remove it. The variable that decides whether the plan holds is &lt;strong&gt;whether Oracle&amp;rsquo;s largest AI customers, above all OpenAI, keep paying ahead of delivery&lt;/strong&gt;.&lt;/p&gt;
&lt;p&gt;This article is dated October 29, 2026. Oracle&amp;rsquo;s fiscal year ends on May 31. Prices are NYSE closes on October 28.&lt;/p&gt;
&lt;p&gt;&lt;img alt="Oracle free cash flow by fiscal year. Fiscal 2024 was positive $11.8 billion, fiscal 2025 minus $0.4 billion and fiscal 2026 minus $23.7 billion. Consensus before October 28 was about minus $35 billion for fiscal 2027 and 2028. Oracle now targets positive free cash flow in the second half of fiscal 2028 and at least $30 billion in fiscal 2029." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/oracle-fcf-turn-customers-landlords-pay-first-openai-test-2026-10-29/fcf-path.en.png"&gt;
&lt;/p&gt;
&lt;h2 id="five-things-to-take-away"&gt;Five things to take away
&lt;/h2&gt;&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Half of Oracle&amp;rsquo;s operating cash last quarter came from customer prepayments.&lt;/strong&gt; Operating cash flow was $23.1 billion in the quarter to August 31. Of that, $11.4 billion was prepayment. Without it, free cash flow would have been about minus $16.8 billion instead of minus $5.4 billion.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Oracle has signed $288 billion of data-center leases that have not started yet.&lt;/strong&gt; That is 2.3 times its $125 billion of borrowings. These leases begin between late 2026 and fiscal 2029 and run 15 to 19 years. Their upfront building cost does not appear in Oracle&amp;rsquo;s capex line.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;The timeline matches S&amp;amp;P&amp;rsquo;s test.&lt;/strong&gt; S&amp;amp;P cut Oracle to BBB- in July and, according to press reports, said the rating could come under pressure if free cash flow is not positive in 2028. Five of the six months in Oracle&amp;rsquo;s target window fall in calendar 2028.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Oracle measures leverage honestly, and that target matters more than free cash flow.&lt;/strong&gt; Its goal of net debt below 3.5 times adjusted EBITDA in fiscal 2029 counts leases and vendor financing as debt. On our rough estimate, it needs adjusted EBITDA of about $85–100 billion that year.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;The stock trades at about 7 times Oracle&amp;rsquo;s fiscal 2030 EPS target of $21.&lt;/strong&gt; The market believes the cash plan more than it did on Tuesday. It still does not believe the earnings plan.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id="what-oracle-promised"&gt;What Oracle promised
&lt;/h2&gt;&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Target&lt;/th&gt;
 &lt;th&gt;Value&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Fiscal 2027 outlook (reaffirmed)&lt;/td&gt;
 &lt;td&gt;Revenue of at least $90bn; non-GAAP EPS of $8.10&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Free cash flow&lt;/td&gt;
 &lt;td&gt;Positive in the second half of fiscal 2028; at least $30bn in fiscal 2029&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Capex&lt;/td&gt;
 &lt;td&gt;Peaks in fiscal 2027–2028, steps down in fiscal 2029&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;New capital-market funding, fiscal 2027–2030&lt;/td&gt;
 &lt;td&gt;Capped at $20bn: about $10bn equity and $10bn gross debt, excluding refinancing, on top of the $20bn at-the-market equity sale completed in the quarter to August&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Self-funding&lt;/td&gt;
 &lt;td&gt;Growth self-funded from fiscal 2029&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Leverage&lt;/td&gt;
 &lt;td&gt;Net debt to adjusted EBITDA below 3.5x in fiscal 2029; net debt includes borrowings, vendor financing and lease liabilities&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Fiscal 2030&lt;/td&gt;
 &lt;td&gt;Revenue of at least $225bn, non-GAAP operating margin of at least 40%, non-GAAP EPS of at least $21&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Source: &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1341439/000119312526435853/d75538dex991.htm" target="_blank" rel="noopener"
 &gt;Oracle Investor Day release, Oct 28&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;The release states the method in one sentence. Vendor financing, customer-provided hardware and customer prepayments are expected to reduce upfront cash needs until free cash flow turns positive. Leases appear in the leverage footnote.&lt;/p&gt;
&lt;h2 id="why-the-turn-came-early-oracle-stopped-paying-first"&gt;Why the turn came early: Oracle stopped paying first
&lt;/h2&gt;&lt;h3 id="customers-prepay"&gt;Customers prepay
&lt;/h3&gt;&lt;p&gt;In the quarter to August 31, Oracle&amp;rsquo;s operating cash flow included &lt;strong&gt;$11.4 billion of customer prepayments&lt;/strong&gt; with a significant financing component. In all of fiscal 2026, such prepayments were $4.6 billion. In the two years before, they were zero. &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1341439/000119312526389274/orcl-20260831.htm" target="_blank" rel="noopener"
 &gt;Oracle 10-Q, Q1 FY27&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;&lt;img alt="Oracle’s cash bridge for the quarter from June to August 2026. Operating cash excluding prepayments was $11.7 billion. Customer prepayments added $11.4 billion, for operating cash flow of $23.1 billion. Capital expenditures were $28.5 billion, leaving free cash flow of minus $5.4 billion." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/oracle-fcf-turn-customers-landlords-pay-first-openai-test-2026-10-29/q1-cash-bridge.en.png"&gt;
&lt;/p&gt;
&lt;p&gt;A prepayment is cash today for service delivered later. It improves free cash flow now and lowers it later, when Oracle recognizes the revenue without collecting new cash. Because the contracts carry a financing component, Oracle also records interest on that money, an immaterial amount so far. In economic terms, a large prepayment is a loan from the customer.&lt;/p&gt;
&lt;p&gt;On the September call, according to a third-party transcript, the CFO said net cash capex in fiscal 2027 would be no more than $70 billion after prepayments, against gross capex of $90–95 billion. If the whole gap is prepayments, that implies &lt;strong&gt;roughly $20–25 billion of prepayments this fiscal year&lt;/strong&gt;. That figure is our inference.&lt;/p&gt;
&lt;h3 id="customers-bring-their-own-hardware"&gt;Customers bring their own hardware
&lt;/h3&gt;&lt;p&gt;Some customers now supply the accelerators themselves. Co-CEO Clay Magouyrk described this on the September call as a &amp;ldquo;bring your own hardware&amp;rdquo; model, according to the same transcript. Oracle then provides power, space, networking and operations. It earns less revenue per gigawatt, but it also spends far less capex. A Wells Fargo note cited by TipRanks estimated about $5.9 billion of revenue per gigawatt for contracts where customers bring their own chips or pay upfront, against about $17.3 billion for standard ones. &lt;a class="link" href="https://www.tipranks.com/news/article/oracles-investor-day-is-coming-heres-the-answer-wells-fargo-wants-on-the-stock" target="_blank" rel="noopener"
 &gt;TipRanks, Oct 2026&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;This creates a tension inside the plan. Every contract that moves to customer hardware helps the cash target and makes the $225 billion revenue target harder.&lt;/p&gt;
&lt;h3 id="suppliers-wait-to-be-paid"&gt;Suppliers wait to be paid
&lt;/h3&gt;&lt;p&gt;Vendor financing means equipment suppliers or their financiers accept later payment. Oracle expects it to help, but the latest quarter does not show it yet. Unpaid capex rose to $6.2 billion at August 31, from $4.0 billion a year earlier. However, Oracle&amp;rsquo;s cash-flow line for short-term financing related to capex was a &lt;strong&gt;net repayment of $0.8 billion&lt;/strong&gt; in the quarter, against an inflow of $2.0 billion a year earlier. Oracle counts vendor financing in its net debt definition. That is the right treatment.&lt;/p&gt;
&lt;h3 id="landlords-build-the-data-centers"&gt;Landlords build the data centers
&lt;/h3&gt;&lt;p&gt;This is the largest mechanism and the least discussed. At August 31, Oracle had &lt;strong&gt;$288 billion of lease commitments that had not yet commenced&lt;/strong&gt;, substantially all for data centers. They are expected to start between the second quarter of fiscal 2027 and fiscal 2029, for terms of 15 to 19 years. The figure was $260 billion at May 31. &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1341439/000119312526389274/orcl-20260831.htm" target="_blank" rel="noopener"
 &gt;Oracle 10-Q, Q1 FY27&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;&lt;img alt="Oracle obligations at August 31, 2026: total borrowings of $125 billion, lease liabilities on the balance sheet of $43.8 billion, leases signed but not yet commenced of $288 billion, and unconditional purchase obligations of $34.2 billion." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/oracle-fcf-turn-customers-landlords-pay-first-openai-test-2026-10-29/obligations.en.png"&gt;
&lt;/p&gt;
&lt;p&gt;When Oracle leases a building, the developer pays for construction. Oracle pays rent later. The construction cost does not appear in Oracle&amp;rsquo;s capex line, so it does not reduce free cash flow as capex. Oracle has occasionally supported a lessor directly: its 10-K disclosed a guarantee of up to $3.3 billion of one lessor&amp;rsquo;s borrowing, which matured in September 2026. Operating-lease rent does reduce operating cash flow once the lease starts, and finance-lease principal is recorded as a financing outflow.&lt;/p&gt;
&lt;p&gt;A rough scale helps. If the $288 billion is paid evenly over 15 to 19 years, rent on these leases alone would be about &lt;strong&gt;$15–19 billion a year&lt;/strong&gt; once all of them have started. In the quarter to August, Oracle paid about $1.1 billion of lease cash in total. That straight-line split is our estimate, not Oracle&amp;rsquo;s disclosure.&lt;/p&gt;
&lt;h3 id="why-oracle-gave-a-date-now"&gt;Why Oracle gave a date now
&lt;/h3&gt;&lt;p&gt;In September, Oracle declined to say when free cash flow would turn positive. About seven weeks later it gave a date. Credit and equity pressure explain the timing.&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;The credit test.&lt;/strong&gt; S&amp;amp;P cut Oracle to BBB- on July 9, one notch above speculative grade, and cited concentration on OpenAI, according to &lt;a class="link" href="https://mlq.ai/news/sp-downgrades-oracle-to-bbb-one-notch-above-junk-citing-openai-concentration-risk/" target="_blank" rel="noopener"
 &gt;MLQ&lt;/a&gt;. According to Fortune, S&amp;amp;P said the rating could come under pressure if free cash flow is not positive in 2028. &lt;a class="link" href="https://fortune.com/2026/09/18/sp-kept-oracle-investment-grade-its-own-numbers-dont-support-that-call/" target="_blank" rel="noopener"
 &gt;Fortune, Sep 18&lt;/a&gt; Five of the six months in Oracle&amp;rsquo;s target window, December 2027 to May 2028, fall in that year.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;The equity test.&lt;/strong&gt; The stock had fallen by about half from its 52-week high. Investors feared repeated share sales. The $20 billion cap, with only $10 billion of equity, answers that fear. At the October 28 market value of about $395 billion, $10 billion is about 2.5% of the company.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id="how-oracle-defines-free-cash-flow"&gt;How Oracle defines free cash flow
&lt;/h2&gt;&lt;p&gt;Oracle defines free cash flow as operating cash flow minus capital expenditures. Under that definition:&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Item&lt;/th&gt;
 &lt;th&gt;Where it goes&lt;/th&gt;
 &lt;th&gt;Effect on free cash flow&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Customer prepayments&lt;/td&gt;
 &lt;td&gt;Operating cash flow&lt;/td&gt;
 &lt;td&gt;Raises it now, lowers it later&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Customer-supplied hardware&lt;/td&gt;
 &lt;td&gt;Not on Oracle&amp;rsquo;s books&lt;/td&gt;
 &lt;td&gt;Lowers capex and revenue&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Vendor financing&lt;/td&gt;
 &lt;td&gt;Financing activities (short-term financing related to capex)&lt;/td&gt;
 &lt;td&gt;Eases cash needs; counted as debt in Oracle&amp;rsquo;s leverage target&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Building cost of leased data centers&lt;/td&gt;
 &lt;td&gt;Not capex&lt;/td&gt;
 &lt;td&gt;Never counted as capex&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Operating-lease rent&lt;/td&gt;
 &lt;td&gt;Operating cash flow&lt;/td&gt;
 &lt;td&gt;Lowers it from lease start&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Finance-lease principal&lt;/td&gt;
 &lt;td&gt;Financing cash flow&lt;/td&gt;
 &lt;td&gt;Not counted&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Sources: Oracle 10-K for fiscal 2026 and 10-Q for the quarter to August 31; our classification of where each item lands.&lt;/p&gt;
&lt;p&gt;None of this breaks an accounting rule. It does mean free cash flow now measures a smaller share of what the AI build-out costs. Oracle&amp;rsquo;s leverage target corrects for that. Its net debt includes lease liabilities and vendor financing, and its adjusted EBITDA adds back operating-lease expense. &lt;strong&gt;Investors should treat the 3.5x leverage target as the main promise and the free-cash-flow date as the secondary one.&lt;/strong&gt;&lt;/p&gt;
&lt;h2 id="is-it-realistic"&gt;Is it realistic?
&lt;/h2&gt;&lt;p&gt;&lt;strong&gt;Fiscal 2027.&lt;/strong&gt; Annualizing last quarter&amp;rsquo;s $11.7 billion of operating cash before prepayments gives about $47 billion. Subtracting net cash capex of $70 billion gives free cash flow of about &lt;strong&gt;minus $20–25 billion&lt;/strong&gt;. That is better than the minus $35 billion consensus, and it depends on $20–25 billion of prepayments actually arriving.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;The turn in fiscal 2028.&lt;/strong&gt; Positive free cash flow in the second half requires capex to fall from its peak while new capacity starts earning. Oracle said capex peaks in fiscal 2027–2028. The second-half target is plausible if Oracle delivers capacity on schedule.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Fiscal 2029.&lt;/strong&gt; At least $30 billion of free cash flow needs both lower capex and much higher revenue. The backlog helps. Oracle&amp;rsquo;s remaining performance obligations were &lt;strong&gt;$664 billion&lt;/strong&gt; at August 31. Oracle expects to recognize about 13% in the next 12 months, 37% in months 13 to 36, and 34% in months 37 to 60. &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1341439/000119312526389274/orcl-20260831.htm" target="_blank" rel="noopener"
 &gt;Oracle 10-Q&lt;/a&gt; That schedule implies roughly $120 billion a year of backlog revenue in months 13 to 36 (September 2027 to August 2029) and about $110 billion a year in months 37 to 60. On our estimate, fiscal 2030&amp;rsquo;s $225 billion target therefore needs roughly half its revenue from contracts not yet signed, renewals and usage.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Leverage in fiscal 2029.&lt;/strong&gt; If all the signed leases commence, lease liabilities could rise by roughly $170–190 billion in present-value terms, using the 5.7% average discount rate on Oracle&amp;rsquo;s existing leases. Add the existing $43.8 billion of lease liabilities and borrowings of about $135 billion after the new debt, subtract cash, and net debt on Oracle&amp;rsquo;s definition could reach about &lt;strong&gt;$300–350 billion&lt;/strong&gt;. To stay below 3.5x, adjusted EBITDA would need to be about &lt;strong&gt;$85–100 billion&lt;/strong&gt;. That is our estimate, not Oracle&amp;rsquo;s, and it is achievable only if revenue arrives on the backlog schedule.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;The sensitivity.&lt;/strong&gt; Oracle gave no fiscal 2029 revenue figure. Our assumption is about $170 billion on the path to $225 billion, so a 10% shortfall is roughly $17 billion. At an assumed incremental EBITDA margin of about 55%, it would remove about $9 billion of cash, close to a third of the $30 billion target. Rent and interest do not shrink when revenue is late.&lt;/p&gt;
&lt;h2 id="the-customer-at-the-center-openai"&gt;The customer at the center: OpenAI
&lt;/h2&gt;&lt;p&gt;Oracle&amp;rsquo;s 10-K says no customer accounted for 10% of revenue in fiscal 2026. Backlog is different. ERP Today reports that roughly half of Oracle&amp;rsquo;s $664 billion backlog is tied to OpenAI. Oracle does not disclose the share. &lt;a class="link" href="https://erp.today/oracle-q1-fy27-results-664b-backlog-ai-contracts/" target="_blank" rel="noopener"
 &gt;ERP Today, Sep 2026&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;That concentration links these mechanisms. A prepayment needs a customer with cash. Customer-supplied hardware needs a customer that can buy chips. Vendors and landlords plausibly lend and build because they see contracts with creditworthy end customers. If OpenAI&amp;rsquo;s fundraising slows, every one of these sources weakens at once, and the leases keep running for 15 to 19 years.&lt;/p&gt;
&lt;p&gt;The bull case on this point is real. ERP Today reported that Oracle&amp;rsquo;s non-OpenAI backlog more than doubled in a year. On the call, the CFO said Oracle booked more than $30 billion of new AI contracts last quarter, and most of them used prepayment or customer hardware. &lt;a class="link" href="https://erp.today/oracle-q1-fy27-results-664b-backlog-ai-contracts/" target="_blank" rel="noopener"
 &gt;ERP Today, Sep 2026&lt;/a&gt; Each new prepaying customer reduces the OpenAI dependence. That diversification is the number to watch.&lt;/p&gt;
&lt;h2 id="the-stock-reaction"&gt;The stock reaction
&lt;/h2&gt;&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;&lt;/th&gt;
 &lt;th&gt;Value&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Close, Oct 27&lt;/td&gt;
 &lt;td&gt;about $130.53&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Close, Oct 28&lt;/td&gt;
 &lt;td&gt;$144.54 (+10.7%)&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Intraday high, Oct 28&lt;/td&gt;
 &lt;td&gt;$148.48&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;After hours, Oct 28, 7:20 p.m. ET&lt;/td&gt;
 &lt;td&gt;$142.39 (−1.5%)&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Price / fiscal 2027 EPS guidance of $8.10&lt;/td&gt;
 &lt;td&gt;about 17.8x&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Price / fiscal 2030 EPS target of $21&lt;/td&gt;
 &lt;td&gt;about 6.9x&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Five-year CDS&lt;/td&gt;
 &lt;td&gt;about 277bp per CNBC, after about 320bp earlier in October; Reuters cited 287bp&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Sources: Yahoo Finance prices; &lt;a class="link" href="https://www.cnbc.com/2026/10/28/oracle-forecast-growth-2030.html" target="_blank" rel="noopener"
 &gt;CNBC, Oct 28&lt;/a&gt;; &lt;a class="link" href="https://www.fool.com/coverage/stock-market-today/2026/10/28/stock-market-today-oct-28-oracle-surges-on-fiscal-2030-revenue-guidance-of-usd225-billion/" target="_blank" rel="noopener"
 &gt;The Motley Fool, Oct 28&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;The market priced the funding relief and little else. A stock at seven times its fiscal 2030 earnings target tells you most investors do not expect that target. They now believe Oracle can get there without repeated equity sales. That is a narrower claim.&lt;/p&gt;
&lt;h2 id="our-judgment-and-what-would-change-it"&gt;Our judgment, and what would change it
&lt;/h2&gt;&lt;p&gt;The free-cash-flow date is achievable. Oracle has changed its contracts so that customers, suppliers and landlords carry the upfront cost. &lt;strong&gt;Customer funding is the variable that dominates&lt;/strong&gt;, ahead of capex timing and margins. If the largest customers keep prepaying, Oracle reaches positive free cash flow in fiscal 2028 and the leverage target in fiscal 2029. If they stop, Oracle&amp;rsquo;s own definition will show it within one or two quarters, because the prepayment line will shrink while rent starts.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Evidence&lt;/th&gt;
 &lt;th&gt;What it would mean&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Prepayments of about $3–4.5bn a quarter for the rest of the year&lt;/td&gt;
 &lt;td&gt;The pace needed for $20–25bn; the fiscal 2027 cash plan is on track&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Non-OpenAI backlog keeps growing faster than total backlog&lt;/td&gt;
 &lt;td&gt;Concentration falls; the credit case improves&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;S&amp;amp;P moves the outlook to positive or affirms after the Investor Day&lt;/td&gt;
 &lt;td&gt;The 2028 test is accepted as credible&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Prepayments fall sharply while lease commencements continue&lt;/td&gt;
 &lt;td&gt;Cash support is fading while fixed costs rise&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Oracle raises fiscal 2028 capex above the fiscal 2027 level&lt;/td&gt;
 &lt;td&gt;The step-down is slipping, and so is the second-half turn&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;OpenAI delays payments or cuts capacity requests&lt;/td&gt;
 &lt;td&gt;The plan&amp;rsquo;s central assumption fails&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;The next check is Oracle&amp;rsquo;s fiscal second-quarter report in December. The prepayment line in operating cash flow and the change in leases not yet commenced will matter more than revenue.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Is Oracle&amp;rsquo;s new model a smarter way to fund AI, or a way to move the bill onto lease contracts and customer credit? Tell us which line you will read first in December.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Research and information only; not investment advice. Estimates and scenarios are analytical assumptions, not company guidance, trade instructions or personalized portfolio decisions. Figures are date-bound. AI tools assisted research and editing.&lt;/em&gt;&lt;/p&gt;</description></item><item><title>Marvell's Reported Lumentum Talks Target the Laser That CPO Cannot Remove. The Bet Is That Lasers Stay Scarce.</title><link>https://koreainvestinsights.com/post/marvell-lumentum-talks-laser-scarcity-cpo-stock-deal-2026-10-29/</link><pubDate>Thu, 29 Oct 2026 08:30:00 +0900</pubDate><guid>https://koreainvestinsights.com/post/marvell-lumentum-talks-laser-scarcity-cpo-stock-deal-2026-10-29/</guid><description>&lt;img src="https://koreainvestinsights.com/" alt="Featured image of post Marvell's Reported Lumentum Talks Target the Laser That CPO Cannot Remove. The Bet Is That Lasers Stay Scarce." /&gt;&lt;p&gt;On October 28, Bloomberg reported that &lt;strong&gt;Marvell Technology is in advanced talks to acquire Lumentum&lt;/strong&gt;, the data-center laser and optics maker. A deal could be announced as soon as next week. As of this writing, neither company has confirmed the talks, and no price, premium or payment mix has been reported. &lt;a class="link" href="https://www.investing.com/news/stock-market-news/marvell-nears-deal-for-optical-supplier-lumentum-to-bolster-hardware-report-says-4981052" target="_blank" rel="noopener"
 &gt;Investing.com, Oct 28&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;The market answered in one evening. At 7:20 p.m. ET, Marvell traded at $232.39, &lt;strong&gt;7.7% below its $251.65 close&lt;/strong&gt;. Lumentum traded at $1,008, &lt;strong&gt;10.3% above its $913.95 close&lt;/strong&gt;. Marvell lost about $17 billion of market value. Lumentum gained about $8 billion on its basic share count, or about $9.5 billion on a diluted count. The combined value of the two companies fell by about $7–8 billion on the news.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Our judgment: the strategic logic is better than the reaction suggests, and the price will likely be worse than Marvell holders hope.&lt;/strong&gt; Marvell wants the one part of an optical link that survives every architecture change: the laser. Co-packaged optics (CPO) can remove Marvell&amp;rsquo;s own signal-processing chips. It cannot remove the light source. But at similar earnings multiples, a stock-funded deal is a transfer of value to Lumentum holders unless the premium stays small. The variable that decides whether the transfer pays back is &lt;strong&gt;whether laser supply stays scarce after 2028&lt;/strong&gt;, when today&amp;rsquo;s capacity additions arrive.&lt;/p&gt;
&lt;p&gt;This article is dated October 29, 2026. It covers a reported deal, not an announced one. Prices are Nasdaq closes and after-hours trades on October 28, US time. All deal terms below are our scenarios.&lt;/p&gt;
&lt;p&gt;&lt;img alt="Marvell, Lumentum and the Philadelphia semiconductor index, rebased to 100 on December 31, 2025. Marvell nearly tripled and Lumentum rose about 2.5 times by October 28. After hours on October 28, Lumentum rose 10% and Marvell fell 8%." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/marvell-lumentum-talks-laser-scarcity-cpo-stock-deal-2026-10-29/price-rebased.en.png"&gt;
&lt;/p&gt;
&lt;h2 id="five-things-to-take-away"&gt;Five things to take away
&lt;/h2&gt;&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Marvell would be buying a company almost as profitable as itself, at almost the same multiple.&lt;/strong&gt; Both reported a 36.6% non-GAAP operating margin in their latest quarter. On guided-quarter earnings, Marvell trades at about 57 times and Lumentum at about 54 times. The common view that Marvell would dilute a high-margin chip business with a low-margin component business is wrong at the operating line.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;The real asset is laser capacity.&lt;/strong&gt; According to a summary of Lumentum&amp;rsquo;s August call, management described its pump lasers as effectively sold out. Its ultra-high-power lasers for CPO reach high volume in the second half of 2027. Nvidia paid $2 billion in March partly for capacity access rights.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;On our arithmetic, the price would have to be paid mostly in Marvell shares.&lt;/strong&gt; Marvell has $3.9 billion of cash and a new $3 billion revolver. Even at its 4.5x acquisition leverage limit, new debt could cover only about a fifth to a quarter of a $110–120 billion price.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Marvell&amp;rsquo;s after-hours loss already equals a 20% premium with no synergy credit.&lt;/strong&gt; Its $17 billion loss is about what a 20% premium on Lumentum&amp;rsquo;s diluted value would transfer to Lumentum holders.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;The deal is long-dated.&lt;/strong&gt; China&amp;rsquo;s market regulator (SAMR) has cleared Marvell&amp;rsquo;s and Lumentum&amp;rsquo;s past deals, but each review took five to nine months. Lumentum shipped 79% of last year&amp;rsquo;s revenue to customers outside the US.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id="what-was-reported-and-what-was-not"&gt;What was reported, and what was not
&lt;/h2&gt;&lt;p&gt;Bloomberg&amp;rsquo;s report says the talks are advanced and private, and that terms and timing could change. Reuters added that neither company immediately responded to requests for comment. Bernstein analysts said a deal of this size probably could not be funded entirely in cash. &lt;a class="link" href="https://boursorama.com/bourse/actualites/marvell-serait-en-pourparlers-avances-en-vue-d-acquerir-lumentum-selon-bloomberg-news-a1c5f80208f3ffe4c2f665dbf6b9fb19" target="_blank" rel="noopener"
 &gt;Reuters via Boursorama, Oct 29&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;One filing makes the report more credible. On October 22, six days before the leak, Marvell replaced its $1.5 billion revolving credit facility with a &lt;strong&gt;$3 billion facility&lt;/strong&gt;, expandable by $2 billion. The new agreement lets Marvell&amp;rsquo;s leverage limit rise from 4.0 to 4.5 times for the quarter in which it completes an acquisition of at least $500 million and the following three quarters. &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1835632/000119312526430424/d147430d8k.htm" target="_blank" rel="noopener"
 &gt;Marvell 8-K, Oct 23&lt;/a&gt; That is a liquidity tool, not deal funding. But it is the kind of housekeeping a company does before a large transaction.&lt;/p&gt;
&lt;p&gt;What we do not know matters more than what we know. We do not know the price, the share of cash, whether Nvidia supports the deal, or what Lumentum&amp;rsquo;s board wants. Every number below is a scenario built from public results.&lt;/p&gt;
&lt;h2 id="marvell-from-connectivity-chips-to-custom-silicon"&gt;Marvell: from connectivity chips to custom silicon
&lt;/h2&gt;&lt;p&gt;Marvell reported &lt;strong&gt;record revenue of $2.739 billion&lt;/strong&gt; for its fiscal second quarter, which ended August 1, 2026. That was 37% more than a year earlier. Data-center revenue was $2.17 billion, 79% of the total, and grew 46%. Non-GAAP gross margin was 58.9%, non-GAAP operating margin 36.6%, and non-GAAP earnings per share $0.94. For the third quarter, Marvell guided revenue of $3.15 billion, plus or minus 5%, and non-GAAP EPS of $1.10. &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1835632/000183563226000022/q227_8kx812026ex-991.htm" target="_blank" rel="noopener"
 &gt;Marvell Q2 FY27 release&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;Growth comes from connectivity and custom silicon.&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Connectivity.&lt;/strong&gt; This is the Inphi heritage: optical DSPs for 800G and 1.6T modules, active electrical cable chips, retimers and switches. On the August call, management said 1.6T DSPs are ramping and that scale-out switching should more than double this fiscal year. &lt;a class="link" href="https://www.marketbeat.com/instant-alerts/transcript-marvell-technology-q2-earnings-call-highlights-2026-08-27/" target="_blank" rel="noopener"
 &gt;MarketBeat call summary, Aug 27&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Custom silicon.&lt;/strong&gt; These are AI accelerators and attached chips designed with hyperscalers. Management said custom revenue ramps sharply in the second half of this fiscal year and should more than double next year.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;Agreements signed in 2026 changed the shape of the company.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Date&lt;/th&gt;
 &lt;th&gt;Agreement&lt;/th&gt;
 &lt;th&gt;What it gives Marvell&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Feb 2&lt;/td&gt;
 &lt;td&gt;Closed the Celestial AI acquisition for about $1bn in cash and about 27 million shares, plus a possible earn-out&lt;/td&gt;
 &lt;td&gt;Photonic Fabric optical interconnect for scale-up; Marvell expects a $500m annualized revenue run rate in Q4 FY28&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Mar 31&lt;/td&gt;
 &lt;td&gt;Nvidia bought $2bn of Marvell Series A convertible preferred stock&lt;/td&gt;
 &lt;td&gt;NVLink Fusion access for custom accelerators and joint silicon-photonics work&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Jul 29 / Aug 18&lt;/td&gt;
 &lt;td&gt;Commercial agreement with Google; on Aug 18, a warrant for up to 58.97m Marvell shares at $206.58&lt;/td&gt;
 &lt;td&gt;Custom chips that attach to Google&amp;rsquo;s TPU systems; one warrant tranche vests per $500m of Google custom revenue&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Sources: &lt;a class="link" href="https://investor.marvell.com/news-events/press-releases/detail/1005/marvell-completes-acquisition-of-celestial-ai" target="_blank" rel="noopener"
 &gt;Marvell, Celestial AI closing&lt;/a&gt;, &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1835632/000119312526134462/d113606d8k.htm" target="_blank" rel="noopener"
 &gt;Marvell 8-K, Mar 31&lt;/a&gt;, &lt;a class="link" href="https://datacenterdynamics.com/en/news/nvidia-invests-2bn-in-marvell-as-part-of-wider-nvlink-fusion-deal" target="_blank" rel="noopener"
 &gt;DatacenterDynamics, Mar 31&lt;/a&gt;, &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1835632/000119312526356217/d412696d8k.htm" target="_blank" rel="noopener"
 &gt;Marvell 8-K, Aug 19&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;The Google warrant is worth reading closely. Its 240 performance tranches vest at one per $500 million of custom revenue through fiscal 2033. Full vesting would therefore require &lt;strong&gt;$120 billion&lt;/strong&gt; of Google custom revenue. That is a ceiling, not a forecast. It still shows how large both companies think the opportunity could become.&lt;/p&gt;
&lt;p&gt;At its October 6 Investor Day, Marvell raised its fiscal 2028 revenue target to about &lt;strong&gt;$20 billion&lt;/strong&gt; and set a fiscal 2031 target of &lt;strong&gt;$70–90 billion&lt;/strong&gt;, with EPS above $30. Press reports put interconnect at about $37.5 billion of the 2031 midpoint. &lt;a class="link" href="https://www.investing.com/news/stock-market-news/marvell-technology-stock-rallies-following-ambitious-investor-day-targets-4934767" target="_blank" rel="noopener"
 &gt;Investing.com, Oct 6&lt;/a&gt; The stock closed at $287.01 that day and has fallen since.&lt;/p&gt;
&lt;p&gt;The balance sheet is modest for a company worth $220 billion. On August 1, Marvell held &lt;strong&gt;$3.93 billion of cash&lt;/strong&gt; and &lt;strong&gt;$4.96 billion of debt&lt;/strong&gt;. Operating cash flow was $606 million in the quarter. Diluted shares were 921 million before any Google warrant shares vest.&lt;/p&gt;
&lt;h2 id="lumentum-the-laser-maker-that-ai-made-scarce"&gt;Lumentum: the laser maker that AI made scarce
&lt;/h2&gt;&lt;p&gt;Lumentum reported revenue of &lt;strong&gt;$1.006 billion&lt;/strong&gt; for its fiscal fourth quarter, which ended June 27, 2026. That was 109% more than a year earlier and 24.5% more than the previous quarter. Components, mostly lasers, were $649 million. Systems, mostly transceivers and optical circuit switches, were $357 million. Non-GAAP gross margin was 50.4%, non-GAAP operating margin 36.6%, and non-GAAP EPS $3.23. Revenue for the full fiscal year was $3.01 billion, up 83%. &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1633978/000162828026055726/lite_ex991xq4fy26.htm" target="_blank" rel="noopener"
 &gt;Lumentum Q4 FY26 release&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;For the September quarter, Lumentum guided revenue of &lt;strong&gt;$1.225–1.275 billion&lt;/strong&gt;, a non-GAAP operating margin of 39.5–40.5%, and non-GAAP EPS of $4.05–4.35. It reports on November 5. &lt;a class="link" href="https://investor.lumentum.com/financial-news-releases/news-details/2026/Lumentum-Announces-Reporting-Date-for-Fiscal-First-Quarter-2027-Financial-Results/default.aspx" target="_blank" rel="noopener"
 &gt;Lumentum, reporting date&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;The headline GAAP number looks alarming and is not. Lumentum reported a &lt;strong&gt;GAAP net loss of $7.2 billion&lt;/strong&gt; in the quarter. A non-cash $7.8 billion loss on debt extinguishment more than explains it, because holders converted notes into shares at prices far below the current stock price. The cost is real, but it is paid in shares, not cash. Common shares rose from 71.4 million in January to 89.7 million in August. Non-GAAP weighted diluted shares were 101.1 million in the June quarter.&lt;/p&gt;
&lt;p&gt;These Lumentum products matter most for AI data centers.&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Electro-absorption modulated lasers (EMLs).&lt;/strong&gt; These are the light sources inside most 800G and 1.6T pluggable modules. Lumentum reported record EML revenue. 200G-per-lane EMLs were more than a quarter of EML revenue.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Continuous-wave and ultra-high-power (UHP) lasers.&lt;/strong&gt; These supply light to silicon-photonics engines and to CPO. Lumentum expects high-volume UHP shipments in the second half of 2027. It has a first order for an external light source module, also for delivery in the second half of 2027.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Optical circuit switches (OCS).&lt;/strong&gt; These switch light without converting it to electricity. Shipments doubled from the previous quarter, and Lumentum guided its first quarter of more than $100 million of OCS revenue.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Transceivers.&lt;/strong&gt; Through Cloud Light, acquired in 2023, Lumentum builds modules itself. It began production shipments of 1.6T modules.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;Source for the product details: &lt;a class="link" href="https://www.marketbeat.com/instant-alerts/lumentum-q4-earnings-call-highlights-2026-08-11/" target="_blank" rel="noopener"
 &gt;MarketBeat call summary, Aug 11&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;On the same call, according to MarketBeat&amp;rsquo;s summary, management said pump lasers were effectively sold out. That phrase explains the deal better than any synergy slide.&lt;/p&gt;
&lt;p&gt;Nvidia had already reached the same conclusion. On March 2, Nvidia announced a &lt;strong&gt;$2 billion investment in Lumentum&lt;/strong&gt;, a multi-billion-dollar purchase commitment and &lt;strong&gt;future capacity access rights for advanced laser components&lt;/strong&gt;. The money supports a new US fab. &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1633978/000119312526085412/d41019dex991.htm" target="_blank" rel="noopener"
 &gt;Nvidia–Lumentum release, Mar 2&lt;/a&gt; On the same day, Nvidia announced a similar $2 billion agreement with Coherent, Lumentum&amp;rsquo;s closest merchant rival. &lt;a class="link" href="https://www.datacenterdynamics.com/en/news/nvidia-invests-2bn-in-optics-solutions-supplier-coherent/" target="_blank" rel="noopener"
 &gt;DatacenterDynamics, Mar 2&lt;/a&gt; Nvidia, whose switches and CPO platforms need these lasers, secured two laser suppliers in one day. It did not bet on one.&lt;/p&gt;
&lt;p&gt;Lumentum&amp;rsquo;s customer list is concentrated. Its two largest customers were 26.6% and 15.0% of fiscal 2026 revenue; the company does not name them. Shipments to customers outside the United States were 79.2% of revenue. &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1633978/000162828026057358/lite-20260627.htm" target="_blank" rel="noopener"
 &gt;Lumentum 10-K&lt;/a&gt;&lt;/p&gt;
&lt;h2 id="side-by-side"&gt;Side by side
&lt;/h2&gt;&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;&lt;/th&gt;
 &lt;th&gt;Marvell&lt;/th&gt;
 &lt;th&gt;Lumentum&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Latest quarter&lt;/td&gt;
 &lt;td&gt;Q2 FY27, ended Aug 1, 2026&lt;/td&gt;
 &lt;td&gt;Q4 FY26, ended Jun 27, 2026&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Revenue, growth vs a year earlier&lt;/td&gt;
 &lt;td&gt;$2.739bn, +37%&lt;/td&gt;
 &lt;td&gt;$1.006bn, +109%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Guided next quarter&lt;/td&gt;
 &lt;td&gt;$3.15bn ±5%&lt;/td&gt;
 &lt;td&gt;$1.225–1.275bn&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Non-GAAP gross margin&lt;/td&gt;
 &lt;td&gt;58.9%&lt;/td&gt;
 &lt;td&gt;50.4%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Non-GAAP operating margin&lt;/td&gt;
 &lt;td&gt;36.6%&lt;/td&gt;
 &lt;td&gt;36.6%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Guided non-GAAP EPS&lt;/td&gt;
 &lt;td&gt;$1.10&lt;/td&gt;
 &lt;td&gt;$4.05–4.35&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Price-to-earnings on guided quarter x4 (Oct 28 close)&lt;/td&gt;
 &lt;td&gt;about 57x&lt;/td&gt;
 &lt;td&gt;about 54x&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Cash / debt&lt;/td&gt;
 &lt;td&gt;$3.93bn / $4.96bn&lt;/td&gt;
 &lt;td&gt;$2.74bn incl. short-term investments / $1.55bn convertible principal&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Capex as % of revenue, last fiscal year&lt;/td&gt;
 &lt;td&gt;4.3%&lt;/td&gt;
 &lt;td&gt;15.0%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Market value, Oct 28 close&lt;/td&gt;
 &lt;td&gt;about $221bn&lt;/td&gt;
 &lt;td&gt;about $82bn basic; $92bn on 101.1m weighted diluted shares&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Sources: company releases and filings linked above; SEC XBRL data for capex. Capex shares use fiscal 2026 figures: Marvell $354m on $8.19bn revenue; Lumentum $451m on $3.01bn revenue.&lt;/p&gt;
&lt;p&gt;The operating-margin and capex lines carry the argument. The operating margins are the same. The capital intensity is not. Marvell designs chips and pays foundries to make them. Lumentum runs its own indium phosphide fabs and is building another one. A combined company would earn about the same operating margin, but it would need about 1.7 times Marvell&amp;rsquo;s current capex per dollar of revenue: 7.2% instead of 4.3%.&lt;/p&gt;
&lt;h2 id="why-marvell-wants-lasers-cpo-removes-the-dsp-not-the-light"&gt;Why Marvell wants lasers: CPO removes the DSP, not the light
&lt;/h2&gt;&lt;p&gt;&lt;img alt="The laser is the one part every optical link keeps. In pluggable modules, the DSP, laser, silicon photonics engine and module assembly are separate parts. Linear-drive optics reduce or remove the DSP. Co-packaged optics remove the DSP and move the photonics and assembly into the switch package, but still need an external laser source." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/marvell-lumentum-talks-laser-scarcity-cpo-stock-deal-2026-10-29/optical-link-layers.en.png"&gt;
&lt;/p&gt;
&lt;p&gt;An optical link consists of a DSP that cleans up the signal, a laser that makes the light, a photonics engine that puts data on the light, and a module that packages everything. Marvell is strong in the first layer and has been building the third through Celestial AI. It has no laser.&lt;/p&gt;
&lt;p&gt;Each step toward tighter integration attacks the layer where Marvell is strongest.&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Linear-drive optics&lt;/strong&gt; remove or simplify the DSP to save power.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Co-packaged optics&lt;/strong&gt; move the optics next to the switch chip. The DSP disappears from that link, and the module maker loses the assembly work.&lt;/li&gt;
&lt;li&gt;In both cases, the &lt;strong&gt;laser stays&lt;/strong&gt;. In CPO it moves outside the package as an external light source, because lasers run hot and fail more often than silicon.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;This is a hedge Marvell&amp;rsquo;s shareholders should understand. Marvell&amp;rsquo;s fiscal 2031 plan relies heavily on interconnect. Some of that interconnect revenue is exposed to the same CPO transition that Nvidia and Broadcom are promoting. Owning Lumentum would give Marvell a stake in every version of the link, including the one that removes its DSP. We wrote in June that EML and CW lasers had become the bottleneck of AI optics. &lt;a class="link" href="https://koreainvestinsights.com/post/marvell-follow-up-eml-cw-dfb-laser-source-korea-cpo-2026-06-03/" &gt;KII, Jun 3&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;There is also a competitive reason. TrendForce lists Broadcom among the leading suppliers of CW and EML lasers. &lt;a class="link" href="https://www.trendforce.com/presscenter/news/20260603-13077.html" target="_blank" rel="noopener"
 &gt;TrendForce, Jun 3&lt;/a&gt; Broadcom also sells CPO switches. &lt;a class="link" href="https://investors.broadcom.com/news-releases/news-release-details/broadcom-redefines-ai-infrastructure-industry-leading-networking" target="_blank" rel="noopener"
 &gt;Broadcom, Oct 8&lt;/a&gt; Broadcom can offer a hyperscaler a full optical path. Marvell cannot today. With Lumentum, Marvell could offer custom accelerators, scale-up optics from Celestial AI, lasers and OCS from one supplier.&lt;/p&gt;
&lt;h2 id="synergy-what-is-real-and-what-is-a-slide"&gt;Synergy: what is real and what is a slide
&lt;/h2&gt;&lt;p&gt;We separate the kinds of synergy, because the deal will likely be marketed with all of them.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Real: supply assurance in a shortage.&lt;/strong&gt; Today a hyperscaler that buys a Marvell DSP still has to find lasers. If lasers are the constraint, a supplier that can commit DSP, laser and photonics capacity together has a selling advantage. This synergy is largest exactly while lasers are scarce.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Plausible: co-design for scale-up optics.&lt;/strong&gt; Celestial AI&amp;rsquo;s Photonic Fabric needs external light. Lumentum&amp;rsquo;s UHP lasers and external light source modules are designed for that job. Designing them together could shorten qualification time. The volumes start in the second half of 2027 at the earliest.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Weak: cost synergies.&lt;/strong&gt; The two companies use different fabs, different processes and largely different engineers. Overlap is limited to corporate costs, sales and some photonics research. We do not expect cost synergies to be large relative to a $110 billion price.&lt;/p&gt;
&lt;p&gt;Then come the dis-synergies, which the market priced on Wednesday evening.&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Channel conflict.&lt;/strong&gt; Marvell sells DSPs to module makers. Lumentum&amp;rsquo;s Cloud Light competes with those module makers. After a deal, Marvell would compete with some of its own DSP customers.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Customer dual-sourcing.&lt;/strong&gt; Large buyers dislike depending on a supplier that also sells against them. Nvidia already holds capacity rights at both Lumentum and Coherent. In a shortage, customers cannot easily leave. When capacity loosens, they can.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Capital intensity.&lt;/strong&gt; Fabless Marvell would own fabs. Lumentum&amp;rsquo;s fiscal fourth-quarter capex alone was $167 million.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Gross margin optics.&lt;/strong&gt; On the latest quarters, the combined non-GAAP gross margin would be about 56.6% instead of 58.9%. Operating margin would barely change. Some investors screen on gross margin, so the stock may be valued lower even though the business is not.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id="financing-on-the-arithmetic-this-is-a-stock-deal"&gt;Financing: on the arithmetic, this is a stock deal
&lt;/h2&gt;&lt;p&gt;Lumentum&amp;rsquo;s diluted equity value at its October 28 close was about &lt;strong&gt;$92 billion&lt;/strong&gt;, on the June quarter&amp;rsquo;s weighted diluted share count. A 20–30% premium implies a purchase price of roughly &lt;strong&gt;$111–120 billion&lt;/strong&gt;, before Lumentum&amp;rsquo;s remaining convertible notes.&lt;/p&gt;
&lt;p&gt;Marvell&amp;rsquo;s capacity to borrow is limited by its own covenant. The 8-K does not define the leverage ratio, so the arithmetic below is illustrative. On guided quarters, Marvell earns about $1.17 billion of non-GAAP operating income a quarter. Lumentum&amp;rsquo;s guidance implies about $0.5 billion. Together that is roughly &lt;strong&gt;$6.7 billion a year&lt;/strong&gt; of operating income; earnings before depreciation are somewhat higher. If leverage is measured on earnings of that size, the 4.5x acquisition-quarter limit allows total debt of about $30–35 billion. The two companies already carry about $6.6 billion. That leaves roughly &lt;strong&gt;$23–28 billion of new debt at the covenant ceiling&lt;/strong&gt;, and less if Marvell wants to keep its credit rating comfortable.&lt;/p&gt;
&lt;p&gt;Our assumption, consistent with Bernstein&amp;rsquo;s view: &lt;strong&gt;at least 75–80% of the price would be paid in Marvell shares&lt;/strong&gt;, unless a partner such as Nvidia contributes capital. That makes the exchange ratio, not the headline price, the number to watch next week.&lt;/p&gt;
&lt;p&gt;&lt;img alt="Change in Marvell run-rate EPS before synergies at premiums of 0% to 40%, in an all-stock deal. At no premium, EPS rises 1.5%. At 20%, it falls 4.0%. At 30%, it falls 6.5%. At 40%, it falls 8.9%. Existing Marvell holders would own 64% to 71% of the combined company." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/marvell-lumentum-talks-laser-scarcity-cpo-stock-deal-2026-10-29/deal-dilution.en.png"&gt;
&lt;/p&gt;
&lt;p&gt;Because the two companies trade at similar multiples, a stock deal at no premium would be roughly neutral to Marvell&amp;rsquo;s earnings per share. The premium is what dilutes. At a 20% premium, Marvell&amp;rsquo;s run-rate non-GAAP EPS would fall about 4%; at 30%, about 6.5%.&lt;/p&gt;
&lt;p&gt;Growth can close the gap. Lumentum guided 24% sequential revenue growth for the September quarter; Marvell guided 15%. If Lumentum keeps growing faster, the dilution shrinks each quarter. This is the honest bull case on the arithmetic. It only works if laser prices and volumes stay strong for several years.&lt;/p&gt;
&lt;h2 id="what-marvell-gains-and-what-it-gives-up"&gt;What Marvell gains and what it gives up
&lt;/h2&gt;&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Marvell gains&lt;/th&gt;
 &lt;th&gt;Marvell gives up&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Laser capacity during a shortage, including EML, CW and UHP lines&lt;/td&gt;
 &lt;td&gt;About 30% of the company to Lumentum holders in a stock deal&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;A position that survives the move from pluggable modules to CPO&lt;/td&gt;
 &lt;td&gt;Its fabless, low-capex model; it takes on Lumentum&amp;rsquo;s fabs and new US fab plans&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;A full optical stack to match Broadcom&amp;rsquo;s lasers, DSPs and CPO switches&lt;/td&gt;
 &lt;td&gt;Neutrality toward module makers that buy its DSPs&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;OCS and external light sources for scale-up networks with Celestial AI&lt;/td&gt;
 &lt;td&gt;Freedom from China review risk on a large share of Lumentum&amp;rsquo;s revenue&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Faster near-term growth: Lumentum guided +24% QoQ&lt;/td&gt;
 &lt;td&gt;Management focus while it ramps its custom programs, including Google&amp;rsquo;s&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;h2 id="regulators-samr-is-the-clock"&gt;Regulators: SAMR is the clock
&lt;/h2&gt;&lt;p&gt;Marvell and Lumentum both have experience with China&amp;rsquo;s State Administration for Market Regulation. SAMR cleared Marvell&amp;rsquo;s purchase of Cavium in June 2018, about seven months after it was announced. &lt;a class="link" href="https://investor.marvell.com/news-events/press-releases/detail/838/marvell-technology-group-ltd-announces-receipt-of-chinese-regulatory-approval-for-cavium-acquisition" target="_blank" rel="noopener"
 &gt;Marvell, Jun 28, 2018&lt;/a&gt; It cleared Marvell&amp;rsquo;s purchase of Inphi in March 2021, about five months after announcement. &lt;a class="link" href="https://www.mlex.com/mlex/articles/2128899" target="_blank" rel="noopener"
 &gt;MLex, Mar 24, 2021&lt;/a&gt; It cleared Lumentum&amp;rsquo;s purchase of Oclaro in December 2018, about nine months after announcement. &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/0001633978/000163397818000150/lumentumform8-ksamrpressre.htm" target="_blank" rel="noopener"
 &gt;Lumentum 8-K, Dec 6, 2018&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;This deal is larger and more strategic. Much of the world&amp;rsquo;s optical-module assembly takes place in China, so Chinese module makers buy large volumes of optical DSPs and EMLs. A regulator that wants to protect them can ask for supply commitments or simply take time. We assume a review of at least nine to twelve months. During that period, Lumentum will trade at a discount to the offer price, and Marvell will carry the uncertainty.&lt;/p&gt;
&lt;h2 id="stock-impact-marvells-evening-loss-already-matches-a-20-premium"&gt;Stock impact: Marvell&amp;rsquo;s evening loss already matches a 20% premium
&lt;/h2&gt;&lt;p&gt;Lumentum&amp;rsquo;s diluted value at the close was about $92 billion. A 20% premium on that value is about &lt;strong&gt;$18.5 billion&lt;/strong&gt;. Marvell lost about &lt;strong&gt;$17 billion&lt;/strong&gt; of market value after hours. Marvell&amp;rsquo;s holders, in other words, treated a premium of roughly 18–20% as a straight transfer and gave the strategic logic no credit yet. Lumentum&amp;rsquo;s own 10% rise points to a smaller premium, or to a real chance that the talks fail.&lt;/p&gt;
&lt;p&gt;That makes the announcement terms the first test. The table below assumes an all-stock deal and no synergy credit. It shows the mechanical effect on Marvell&amp;rsquo;s market value relative to the after-hours price.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Announced premium&lt;/th&gt;
 &lt;th&gt;Value transferred to Lumentum holders&lt;/th&gt;
 &lt;th&gt;Already priced after hours&lt;/th&gt;
 &lt;th&gt;Further move for Marvell, no synergy credit&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;10%&lt;/td&gt;
 &lt;td&gt;about $9bn&lt;/td&gt;
 &lt;td&gt;about $17bn&lt;/td&gt;
 &lt;td&gt;about +4%, a partial recovery&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;20%&lt;/td&gt;
 &lt;td&gt;about $18bn&lt;/td&gt;
 &lt;td&gt;about $17bn&lt;/td&gt;
 &lt;td&gt;about −1%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;30%&lt;/td&gt;
 &lt;td&gt;about $28bn&lt;/td&gt;
 &lt;td&gt;about $17bn&lt;/td&gt;
 &lt;td&gt;about −5%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;40%&lt;/td&gt;
 &lt;td&gt;about $37bn&lt;/td&gt;
 &lt;td&gt;about $17bn&lt;/td&gt;
 &lt;td&gt;about −10%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;These are mechanical estimates, not forecasts. Marvell would trade above them on credible supply commitments from Nvidia or a hyperscaler to the combined company, a visible synergy plan for CPO lasers, or a share-heavy structure without large new debt. It would trade below them on a large cash component funded by debt, or on SAMR conditions that limit how the combined company sells.&lt;/p&gt;
&lt;p&gt;If the talks fail, the arithmetic reverses. Marvell would likely recover most of the after-hours loss, and Lumentum would give up its gain.&lt;/p&gt;
&lt;p&gt;For Lumentum holders, the stock will start trading on the deal spread once terms are public. With a long review, we would expect a spread wider than for a routine semiconductor deal.&lt;/p&gt;
&lt;h2 id="our-judgment-and-what-would-change-it"&gt;Our judgment, and what would change it
&lt;/h2&gt;&lt;p&gt;The deal makes industrial sense. It is the most direct way for Marvell to stay relevant as optics move into the package. &lt;strong&gt;Laser supply is the variable that dominates&lt;/strong&gt;, and it ranks ahead of the premium and the regulator. If lasers remain the bottleneck through 2028 and beyond, Marvell will have bought the scarcest asset in AI networking, and the premium will look small. If Coherent, Broadcom, Lumentum&amp;rsquo;s new fab and Asian suppliers add enough indium phosphide capacity by 2028, Marvell will have paid about 55 times earnings for a cyclical component maker. It will also own its fabs.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Evidence&lt;/th&gt;
 &lt;th&gt;What it would mean&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Announced premium of 20% or less, mostly stock&lt;/td&gt;
 &lt;td&gt;The evening move already priced the cost; Marvell can recover part of it&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Nvidia or a hyperscaler signs long-term supply agreements with the combined company&lt;/td&gt;
 &lt;td&gt;Customers accept a vertically integrated Marvell; dis-synergy risk falls&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Lumentum&amp;rsquo;s November 5 report shows laser lead times and prices still rising&lt;/td&gt;
 &lt;td&gt;Scarcity is lasting; the strategic value is growing&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Premium of 30% or more, or debt above $15bn&lt;/td&gt;
 &lt;td&gt;Dilution exceeds what growth can close within a year&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Coherent or Broadcom announce large new InP laser capacity for 2027–28&lt;/td&gt;
 &lt;td&gt;Scarcity premium fades; Marvell would be buying near a cyclical peak&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;SAMR review extends beyond twelve months or adds supply conditions&lt;/td&gt;
 &lt;td&gt;Value is delayed, and the combined company loses pricing freedom in China&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;The next check is November 5, when Lumentum reports and may comment on the talks. Marvell reports its fiscal third quarter in late November or early December. If a deal is announced first, the exchange ratio will settle most of the near-term debate.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Would you pay 55 times earnings for laser capacity if you believed CPO arrives in 2028? Tell us which number in the announcement would change your view: the premium, the cash share, or a customer commitment.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Research and information only; not investment advice. Estimates and scenarios are analytical assumptions, not company guidance, trade instructions or personalized portfolio decisions. Figures are date-bound. AI tools assisted research and editing.&lt;/em&gt;&lt;/p&gt;</description></item><item><title>SK Hynix's 80% Margin Answers 2027. Its Contract Price Band Will Answer 2029.</title><link>https://koreainvestinsights.com/post/sk-hynix-80-percent-margin-contract-price-band-2026-10-29/</link><pubDate>Thu, 29 Oct 2026 00:30:00 +0900</pubDate><guid>https://koreainvestinsights.com/post/sk-hynix-80-percent-margin-contract-price-band-2026-10-29/</guid><description>&lt;img src="https://koreainvestinsights.com/" alt="Featured image of post SK Hynix's 80% Margin Answers 2027. Its Contract Price Band Will Answer 2029." /&gt;&lt;p&gt;SK hynix reported &lt;strong&gt;₩83.3 trillion of operating profit on ₩104.6 trillion of revenue&lt;/strong&gt; for the third quarter. That is an operating margin of 80%. At average quarterly exchange rates, the profit is roughly &lt;strong&gt;$59 billion&lt;/strong&gt;, about nine-tenths of the $64.0 billion non-GAAP operating income Nvidia reported for its quarter ended in July.&lt;/p&gt;
&lt;p&gt;The stock closed at &lt;strong&gt;₩1,487,000&lt;/strong&gt; on October 28. It rose 10.9% that day, but it is still &lt;strong&gt;49% below its June 22 close&lt;/strong&gt; of ₩2,919,000. Over the same months, quarterly operating profit rose from ₩60.5 trillion to ₩83.3 trillion. On the 2027 consensus, the shares trade at about &lt;strong&gt;3.2 times expected earnings&lt;/strong&gt;.&lt;/p&gt;
&lt;p&gt;That gap is the real question. The market is not doubting 2027. It is pricing a steep fall in profit after 2028. The most important line in the release was therefore not the profit. It was this: long-term agreements with &lt;strong&gt;14 major customers now cover more than half of 2027 and 2028 production capacity&lt;/strong&gt;, at prices set &amp;ldquo;within a price range agreed upon with customers.&amp;rdquo;&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Our judgment: the contract disclosure makes a 2027–28 earnings collapse unlikely, and the 3x multiple already prices a harsh fall after that. Whether that fall is too harsh depends on one undisclosed variable, the floor of the price band.&lt;/strong&gt; This article explains why the floor dominates, what would show it is real, and what would prove the skeptics right.&lt;/p&gt;
&lt;p&gt;This article is dated October 29, 2026, before Samsung Electronics&amp;rsquo; 10 a.m. KST earnings call. Korean prices are KRX regular-session closes, not after-hours trades. Naver Finance daily charts include after-hours trades and can show different numbers, such as ₩1,508,000 for October 28. SK hynix&amp;rsquo;s Q3 figures are preliminary and unaudited.&lt;/p&gt;
&lt;p&gt;&lt;img alt="SK hynix quarterly operating profit rose from ₩11.4tn in Q3 2025 to ₩83.3tn in Q3 2026, with margin rising from 47% to 80%. The share price peaked at ₩2,919,000 on June 22, 2026, fell to ₩1,223,000 on October 20, and closed at ₩1,487,000 on October 28." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/sk-hynix-80-percent-margin-contract-price-band-2026-10-29/profit-vs-price.en.png"&gt;
&lt;/p&gt;
&lt;h2 id="five-things-to-take-away"&gt;Five things to take away
&lt;/h2&gt;&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;The profit is real, and in dollars it grew faster.&lt;/strong&gt; Operating profit rose 38% in won terms from Q2. The won strengthened during the quarter, so in dollars the increase was about 46%.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Net income fell 58%, but memory did not.&lt;/strong&gt; Q2 net income included ₩53.24 trillion of valuation gains on financial instruments. Kioxia&amp;rsquo;s share price fell about 40% in Q3. Screens built on net income will misread this company in both directions.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;More than half of 2027–28 capacity is contracted, with a price band.&lt;/strong&gt; The company did not disclose the floor, the band width, minimum purchase terms or coverage after 2028. Those details decide whether the contracts protect profit or only reserve volume.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;SK hynix is also building the supply that will test the band.&lt;/strong&gt; Investment will &amp;ldquo;increase significantly&amp;rdquo; next year as the Yongin Phase 1 fab starts full operation. The company forecasts a shortage through 2030. That is a forecast, not a contract.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;The ₩100 trillion payout is a plan; ₩24 trillion of new shareholder returns was approved this week.&lt;/strong&gt; The &amp;ldquo;₩210 trillion&amp;rdquo; headline adds the top of Samsung&amp;rsquo;s existing plan range to SK hynix&amp;rsquo;s plan. The recurring signal is the promise to buy back shares every quarter.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id="the-quarter-was-real-and-in-dollars-it-was-bigger"&gt;The quarter was real, and in dollars it was bigger
&lt;/h2&gt;&lt;p&gt;The question here is simple: did the business actually improve, or did accounting and currency flatter it? The operating line answers it.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Quarter&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Revenue (₩tn)&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Operating profit (₩tn)&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Operating margin&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Approx. operating profit ($bn)&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Q3 2025&lt;/td&gt;
 &lt;td style="text-align: right"&gt;24.45&lt;/td&gt;
 &lt;td style="text-align: right"&gt;11.38&lt;/td&gt;
 &lt;td style="text-align: right"&gt;47%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;8.2&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Q4 2025&lt;/td&gt;
 &lt;td style="text-align: right"&gt;32.83&lt;/td&gt;
 &lt;td style="text-align: right"&gt;19.17&lt;/td&gt;
 &lt;td style="text-align: right"&gt;58%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;13.2&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Q1 2026&lt;/td&gt;
 &lt;td style="text-align: right"&gt;52.58&lt;/td&gt;
 &lt;td style="text-align: right"&gt;37.61&lt;/td&gt;
 &lt;td style="text-align: right"&gt;72%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;25.7&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Q2 2026&lt;/td&gt;
 &lt;td style="text-align: right"&gt;79.32&lt;/td&gt;
 &lt;td style="text-align: right"&gt;60.54&lt;/td&gt;
 &lt;td style="text-align: right"&gt;76%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;40.3&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Q3 2026&lt;/td&gt;
 &lt;td style="text-align: right"&gt;104.62&lt;/td&gt;
 &lt;td style="text-align: right"&gt;83.33&lt;/td&gt;
 &lt;td style="text-align: right"&gt;80%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;58.8&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Revenue rose 32% and operating profit 38% from Q2. Dollar figures use the average of daily USD/KRW quotes for each quarter (about 1,501 in Q2 and 1,417 in Q3). They are our approximation, not company figures. The won strengthened about 6% between the two quarterly averages, so the dollar growth was larger than the won growth. &lt;a class="link" href="https://news.skhynix.com/en/q3-2026-business-results/" target="_blank" rel="noopener"
 &gt;SK hynix Q3 release&lt;/a&gt;, &lt;a class="link" href="https://news.skhynix.com/en/q2-2026-business-results/" target="_blank" rel="noopener"
 &gt;Q2 release&lt;/a&gt;, &lt;a class="link" href="https://news.skhynix.com/en/q1-2026-business-results/" target="_blank" rel="noopener"
 &gt;Q1 release&lt;/a&gt;, &lt;a class="link" href="https://finance.yahoo.com/quote/KRW%3DX/history/" target="_blank" rel="noopener"
 &gt;USD/KRW history&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;The company attributed the quarter to higher prices for key products, strong server DDR5 sales and the start of volume HBM4 shipments. It expects HBM4 to pass HBM3E in both shipments and revenue in the fourth quarter. Cash and equivalents reached ₩114.5 trillion, borrowings fell to ₩15.8 trillion and net cash rose to &lt;strong&gt;₩98.7 trillion&lt;/strong&gt;.&lt;/p&gt;
&lt;p&gt;The Nvidia comparison needs two caveats. Nvidia&amp;rsquo;s figure is non-GAAP and covers May–July; SK hynix&amp;rsquo;s is K-IFRS and covers July–September. The point is scale, not a like-for-like ranking. A memory maker now earns operating profit in the same order of magnitude as the company that buys much of its HBM. &lt;a class="link" href="https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-second-quarter-fiscal-2027" target="_blank" rel="noopener"
 &gt;Nvidia Q2 FY2027 results&lt;/a&gt;.&lt;/p&gt;
&lt;h2 id="net-income-fell-58-because-of-kioxia-not-memory"&gt;Net income fell 58% because of Kioxia, not memory
&lt;/h2&gt;&lt;p&gt;Third-quarter net income was ₩39.8 trillion, down 58% from ₩93.9 trillion in Q2. Anyone screening on trailing earnings will see a collapse. The operating business shows the opposite.&lt;/p&gt;
&lt;p&gt;The Q2 figure was inflated. SK hynix&amp;rsquo;s &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/2120882/000119312526354777/d147827d6k.htm" target="_blank" rel="noopener"
 &gt;half-year report&lt;/a&gt; shows &lt;strong&gt;₩53.24 trillion of valuation gains on financial instruments&lt;/strong&gt; in that quarter, largely linked to its Kioxia investment. We discussed that in our &lt;a class="link" href="https://koreainvestinsights.com/post/samsung-hynix-fair-per-normalized-earnings-2026-10-25/" target="_blank" rel="noopener"
 &gt;October 25 fair P/E analysis&lt;/a&gt;. In Q3 the move reversed. Kioxia&amp;rsquo;s shares fell from about ¥29,900 on June 30 to ¥18,095 on September 30, a drop of about 40%. &lt;a class="link" href="https://finance.yahoo.com/quote/285A.T/history/" target="_blank" rel="noopener"
 &gt;Kioxia price history&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;The release does not itemize the gap between operating profit and net income. Before the release, our estimate of the Q3 valuation loss on SK hynix&amp;rsquo;s remaining Kioxia-linked convertible bond was ₩20–25 trillion. That estimate scales the bond&amp;rsquo;s book value by Kioxia&amp;rsquo;s share price and ignores currency and contract terms. The reported net income is consistent with a loss of that size after tax, but the company has not confirmed the figure.&lt;/p&gt;
&lt;p&gt;The same effect may repeat. Kioxia closed at ¥12,605 on October 27, about 30% below its September 30 level. If it stays near there, the same rough method points to a further non-cash loss of roughly ₩9–12 trillion in Q4. &lt;strong&gt;For this company, operating profit is the cleaner measure of the memory business. Net income will swing with a Japanese stock price.&lt;/strong&gt;&lt;/p&gt;
&lt;h2 id="the-market-is-not-pricing-2027-it-is-pricing-what-comes-after"&gt;The market is not pricing 2027. It is pricing what comes after
&lt;/h2&gt;&lt;p&gt;At ₩1,487,000 and a 2027 consensus EPS of ₩470,589, SK hynix trades at &lt;strong&gt;3.16 times expected 2027 earnings&lt;/strong&gt;. On October 23 the same calculation gave 2.81 times. Since then the price has risen 16.4%, while the 2027 EPS estimate has risen 3.7%. Consensus is the Naver/WiseReport figure collected on October 28; some analysts may not yet reflect the release.&lt;/p&gt;
&lt;p&gt;Another way to see it: the market value at the regular-session close was about ₩1,086 trillion. Subtracting ₩98.7 trillion of net cash leaves an enterprise value of roughly &lt;strong&gt;three times the Q3 operating profit annualized&lt;/strong&gt;. Annualizing one quarter is not a forecast. It shows how little the price assumes about profit lasting.&lt;/p&gt;
&lt;p&gt;In our October 25 analysis, the price implied that SK hynix&amp;rsquo;s normalized earnings would fall to about 18% of the 2027 level. That is the bet the market is making. It is a bet on price, not on volume. Few investors doubt that SK hynix will sell its 2027 output. The doubt is what that output will be worth once new fabs arrive.&lt;/p&gt;
&lt;p&gt;October 28 also brought a broad rally. The KOSPI closed at 6,012.44, up 6.63%, and the exchange triggered a buy-side &amp;ldquo;sidecar&amp;rdquo; pause for the second day in a row. &lt;a class="link" href="https://m.stock.naver.com/domestic/index/KOSPI/total" target="_blank" rel="noopener"
 &gt;KOSPI data&lt;/a&gt;, &lt;a class="link" href="https://www.yna.co.kr/amp/view/AKR20261028064600008" target="_blank" rel="noopener"
 &gt;sidecar report&lt;/a&gt;. A sidecar is a temporary halt on program orders after a sharp futures move. It does not identify who was buying. The public data we could verify do not yet show whether foreign investors were covering futures shorts. We treat the rally as price evidence, not as proof that selling has ended. Our &lt;a class="link" href="https://koreainvestinsights.com/post/rebound-case-s1-reading-and-path-2026-10-24/" target="_blank" rel="noopener"
 &gt;rebound case from October 24&lt;/a&gt; set out the conditions that would confirm it.&lt;/p&gt;
&lt;h2 id="what-the-contracts-say-and-what-they-leave-out"&gt;What the contracts say, and what they leave out
&lt;/h2&gt;&lt;p&gt;The release makes four specific claims about long-term agreements (LTAs):&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;SK hynix has signed LTAs with &lt;strong&gt;14 major customers&lt;/strong&gt;, including leading global cloud providers.&lt;/li&gt;
&lt;li&gt;Contract terms are &lt;strong&gt;typically five years&lt;/strong&gt;, and the company is discussing longer terms with several customers.&lt;/li&gt;
&lt;li&gt;Contracted volume equals &lt;strong&gt;more than 50% of total production capacity for 2027 and 2028&lt;/strong&gt;.&lt;/li&gt;
&lt;li&gt;Prices reflect market conditions &lt;strong&gt;within a price range agreed with customers&lt;/strong&gt;.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;The fourth point is the one that matters for valuation. Volume certainty was never the bear case. A price range turns part of memory pricing from a spot market into a corridor. In a shortage, the ceiling costs SK hynix some upside. In a glut, the floor protects it. Which side matters more depends on where the corridor sits.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Question&lt;/th&gt;
 &lt;th&gt;Disclosed on October 28&lt;/th&gt;
 &lt;th&gt;Not disclosed&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;How much volume is contracted?&lt;/td&gt;
 &lt;td&gt;More than 50% of 2027 and 2028 capacity&lt;/td&gt;
 &lt;td&gt;Coverage for 2029 and later&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;How many customers?&lt;/td&gt;
 &lt;td&gt;14 major customers&lt;/td&gt;
 &lt;td&gt;Customer concentration and the HBM share of contracted volume&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;How is price set?&lt;/td&gt;
 &lt;td&gt;Market-linked, within an agreed range&lt;/td&gt;
 &lt;td&gt;Floor level, band width, reset frequency&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;What binds the buyer?&lt;/td&gt;
 &lt;td&gt;Not described&lt;/td&gt;
 &lt;td&gt;Minimum purchase, take-or-pay, prepayments, cancellation terms&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;How long?&lt;/td&gt;
 &lt;td&gt;Typically five years&lt;/td&gt;
 &lt;td&gt;Start dates; contracts signed in 2026–27 would run to about 2031–32&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;A brokerage commentary on October 28 said SK hynix had sold DRAM and HBM output through 2029. We could not find that statement in the published results materials. &lt;strong&gt;Correction, October 29:&lt;/strong&gt; Korean press reports of the earnings call confirm that SK hynix described DRAM and HBM as effectively sold out through 2029. The basis it gave was LTA volume already secured plus demand under concrete discussion, so the statement describes how much demand the company can see, not signed volume at fixed prices. Coverage for 2029 and later is still not disclosed. An earlier version of this article treated the statement as unconfirmed. &lt;a class="link" href="https://view.asiae.co.kr/article/2026102809584449660" target="_blank" rel="noopener"
 &gt;Asia Business Daily&lt;/a&gt;, &lt;a class="link" href="https://koreainvestinsights.com/post/samsung-sk-hynix-lta-volume-locked-terms-flexible-2026-10-29/" target="_blank" rel="noopener"
 &gt;our follow-up on contract quality&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;Other evidence points the same way, with lower certainty. On October 27, Maeil Business Newspaper reported that Alibaba, Tencent and ByteDance had asked Samsung and SK hynix for long-term server DRAM supply, with some requests running up to ten years, to 2035. &lt;a class="link" href="https://www.mk.co.kr/news/business/12187030" target="_blank" rel="noopener"
 &gt;Maeil Business report&lt;/a&gt;. That is a report of a request, not a signed contract. US export rules restrict HBM sales to China, and some conventional DRAM may also need licenses depending on specifications and end use. KB Securities separately estimates that 70–80% of both companies&amp;rsquo; planned 2027 memory volume is covered by five-year LTAs. That is a broker estimate, not company data. &lt;a class="link" href="https://kbthink.com/securities-view.html?docId=20261023133818567K" target="_blank" rel="noopener"
 &gt;KB Securities, October 26&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;TrendForce&amp;rsquo;s public summary of its October 7 DRAM bulletin describes the same structure: multi-year server DRAM contracts, with price-band mechanisms on many of those volumes and the remainder priced quarterly. &lt;a class="link" href="https://www.trendforce.com/research/download/RP261007RZ" target="_blank" rel="noopener"
 &gt;TrendForce DRAM bulletin&lt;/a&gt;. &lt;strong&gt;When buyers ask for ten-year terms, they are telling you which side of the band they fear.&lt;/strong&gt; They fear paying more, not being stuck with too much.&lt;/p&gt;
&lt;p&gt;&lt;img alt="Timeline from 2026 to 2035. SK hynix disclosed LTA coverage of more than 50% of capacity for 2027 and 2028. A typical five-year term would extend contracts signed in 2026–27 to about 2031–32, but coverage after 2028 is not disclosed. Chinese cloud firms reportedly requested supply to 2035. On the supply side, M15X ramps to full capacity by end-2026 and Yongin Phase 1 starts full operation in 2027. TrendForce expects tight supply through 2027; SK hynix forecasts a shortage through 2030." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/sk-hynix-80-percent-margin-contract-price-band-2026-10-29/contract-timeline.en.png"&gt;
&lt;/p&gt;
&lt;h2 id="sk-hynix-is-also-building-the-supply-that-will-test-the-band"&gt;SK hynix is also building the supply that will test the band
&lt;/h2&gt;&lt;p&gt;The release says investment will &lt;strong&gt;&amp;ldquo;increase significantly&amp;rdquo; next year&lt;/strong&gt;. Yongin Phase 1 enters full-scale operation, and the M15X fab will run at maximum capacity by the end of 2026 to raise HBM4 output. The company adds that it will adjust the scale and timing of investment flexibly.&lt;/p&gt;
&lt;p&gt;This is the honest tension in the story. The company that forecasts a shortage through 2030 is also adding capacity, and so are its competitors. The shortage claim rests on the difficulty of leading-edge processes and long fab lead times. That is a reasonable argument, but it is the company&amp;rsquo;s forecast.&lt;/p&gt;
&lt;p&gt;Independent forecasts support the near term. TrendForce expects memory supply to stay tight through 2027 and projects a &lt;strong&gt;121% rise in blended HBM prices in 2027&lt;/strong&gt;, driven by supply limits, more HBM4 and the start of HBM4E. It also notes that GPU and ASIC makers are considering 8-high HBM stacks to save cost; those carry a 10–20% higher price per gigabit than 12-high stacks. &lt;a class="link" href="https://www.trendforce.com/presscenter/news/20260929-13255.html" target="_blank" rel="noopener"
 &gt;TrendForce, September 29&lt;/a&gt;. None of this extends to 2029.&lt;/p&gt;
&lt;p&gt;So the band floor and the supply response are linked. If new supply arrives in 2028–30 and spot prices fall, the floor is what protects contracted volume. If the floor is close to today&amp;rsquo;s contract prices and the buyer must take the volume, the market&amp;rsquo;s 18% normalization looks too harsh for the contracted half. If the band is wide or resets each year, the contracts mostly guarantee shipments, and the skeptics are right to apply a cyclical multiple.&lt;/p&gt;
&lt;h2 id="the-payout-100-trillion-is-a-plan-24-trillion-was-approved-this-week"&gt;The payout: ₩100 trillion is a plan; ₩24 trillion was approved this week
&lt;/h2&gt;&lt;p&gt;A Korean newspaper headline on October 27 put Samsung&amp;rsquo;s and SK hynix&amp;rsquo;s combined shareholder returns at &amp;ldquo;₩210 trillion or more.&amp;rdquo; The number adds different things. Here is what SK hynix actually said on October 28. &lt;a class="link" href="https://news.skhynix.com/en/q3-2026-business-results/" target="_blank" rel="noopener"
 &gt;SK hynix Q3 release&lt;/a&gt;, &lt;a class="link" href="https://www.hankyung.com/article/2026102787811" target="_blank" rel="noopener"
 &gt;Hankyung report&lt;/a&gt;.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Item&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Amount&lt;/th&gt;
 &lt;th&gt;Status&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;FY2026 shareholder return plan&lt;/td&gt;
 &lt;td style="text-align: right"&gt;About ₩100tn&lt;/td&gt;
 &lt;td&gt;Plan: cash dividends of ₩30tn or more, plus buybacks and cancellations of ₩60tn or more&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;August buyback&lt;/td&gt;
 &lt;td style="text-align: right"&gt;₩40tn&lt;/td&gt;
 &lt;td&gt;Completed last week; shares to be cancelled shortly. Counted inside the ₩60tn&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Q3 cash dividend&lt;/td&gt;
 &lt;td style="text-align: right"&gt;₩14tn&lt;/td&gt;
 &lt;td&gt;Approved October 28: ₩20,000 per share, half regular and half a special dividend from Kioxia gains&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;New buyback for cancellation&lt;/td&gt;
 &lt;td style="text-align: right"&gt;₩10tn&lt;/td&gt;
 &lt;td&gt;Approved October 28; counted as a shareholder return&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;New buyback for employee pay&lt;/td&gt;
 &lt;td style="text-align: right"&gt;About ₩10tn&lt;/td&gt;
 &lt;td&gt;Approved October 28; excluded from the return total&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;The ₩24 trillion of new shareholder returns approved this week is the ₩14 trillion dividend plus the ₩10 trillion buyback for cancellation. To reach its own floors by year-end, the company still has to pay more dividends and buy back at least another ₩10 trillion for cancellation. The release does not say how much of the dividend floor was already paid earlier in the year.&lt;/p&gt;
&lt;p&gt;For market flows, both new buybacks matter, because both buy shares in the market. For per-share value, only the cancelled ₩10 trillion counts. The release does not give an execution period, so the daily buying pressure is unknown.&lt;/p&gt;
&lt;p&gt;On the Samsung side of the headline, the ₩110 trillion is the top of the ₩90–110 trillion range Samsung announced on August 21. It is not a new increase. Samsung plans to set dividend details at its October board meeting and the remaining return at a January 2027 board meeting. &lt;a class="link" href="https://news.samsung.com/global/samsung-electronics-to-implement-largest-ever-shareholder-return-in-2026-estimated-at-krw-90-to-110-trillion" target="_blank" rel="noopener"
 &gt;Samsung Electronics, August 21&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;The special dividend is a one-off; it returns an investment gain. The more telling line is the stated principle that &lt;strong&gt;buybacks &amp;ldquo;will continue every quarter.&amp;rdquo;&lt;/strong&gt; A management team that expected a 2027 cash cliff would be unlikely to promise that. It is a principle, not an amount, so it supports the 2027–28 view and says nothing about 2029.&lt;/p&gt;
&lt;h2 id="our-judgment-and-what-would-change-it"&gt;Our judgment, and what would change it
&lt;/h2&gt;&lt;p&gt;&lt;strong&gt;We side with the contracts for 2027–28 and with the skeptics for 2029–30, until the floor is disclosed.&lt;/strong&gt; The Q3 release lowers the risk of a near-term earnings collapse more than a 3x multiple implies. It does not yet justify valuing post-2028 profit as if it were contracted.&lt;/p&gt;
&lt;p&gt;Three variables decide the case. The &lt;strong&gt;band floor&lt;/strong&gt; dominates, because it is the only one that directly limits how far contracted profit can fall. &lt;strong&gt;Supply response&lt;/strong&gt; comes second; it decides when the floor is tested. &lt;strong&gt;Customer returns on AI spending&lt;/strong&gt; come third; they decide whether buyers want to renegotiate.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Evidence&lt;/th&gt;
 &lt;th&gt;What it would mean&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;SK hynix, Samsung or Micron says the floor sits near current contract prices, with minimum purchase or prepayment terms&lt;/td&gt;
 &lt;td&gt;The market&amp;rsquo;s normalization assumption looks too harsh for contracted volume; the discount to 2027 earnings should narrow&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;LTA coverage is disclosed for 2029–30 at levels near the 2027–28 share&lt;/td&gt;
 &lt;td&gt;The &amp;ldquo;2027 answered, 2029 open&amp;rdquo; framing would need revision toward durability&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Bands are described as wide, reset annually, or tied to spot prices&lt;/td&gt;
 &lt;td&gt;The contracts mainly secure volume; a cyclical multiple remains justified&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Industry capex guidance for 2027 rises faster than expected bit demand&lt;/td&gt;
 &lt;td&gt;The floor will be tested sooner, around 2028&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Hyperscaler capex guidance is cut, or buyers seek to defer LTA volume&lt;/td&gt;
 &lt;td&gt;Renegotiation risk rises; the contract floor matters more than the headline coverage&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;The next checks are close. Samsung Electronics holds its earnings call at &lt;strong&gt;10 a.m. KST on October 29&lt;/strong&gt;; the useful question is whether it describes its own LTA pricing terms, not only the dividend. Microsoft reports after the US close on October 28, early on October 29 in Korea, and third-party calendars place Alphabet and Amazon in the same window; Meta reported on October 21. Their capex and lease guidance shapes buyer behavior in 2027. Kioxia&amp;rsquo;s share price will decide how SK hynix&amp;rsquo;s Q4 net income looks.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Which disclosure would change your view most: the level of the LTA price floor, the contracted share of 2029–30 capacity, or the industry&amp;rsquo;s 2027 capex plan? Tell us what number would convince you.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Research and information only; not investment advice. Estimates and scenarios are analytical assumptions, not company guidance, trade instructions or personalized portfolio decisions. Figures are date-bound. AI tools assisted research and editing.&lt;/em&gt;&lt;/p&gt;</description></item><item><title>Korean Substrates: SEMCO Confirms Demand, Daeduck Must Deliver the Margin</title><link>https://koreainvestinsights.com/post/korea-substrates-semco-daeduck-2026-10-28/</link><pubDate>Wed, 28 Oct 2026 21:00:00 +0900</pubDate><guid>https://koreainvestinsights.com/post/korea-substrates-semco-daeduck-2026-10-28/</guid><description>&lt;img src="https://koreainvestinsights.com/" alt="Featured image of post Korean Substrates: SEMCO Confirms Demand, Daeduck Must Deliver the Margin" /&gt;&lt;p&gt;One percentage point of operating margin is worth about &lt;strong&gt;KRW4.3 billion&lt;/strong&gt; in our Daeduck Electronics Q3 model. The model&amp;rsquo;s entire profit advantage over the latest recorded consensus is only &lt;strong&gt;KRW4.1 billion&lt;/strong&gt;.&lt;/p&gt;
&lt;p&gt;That is the investment question after Samsung Electro-Mechanics reported strong AI-component results on October 28. Better substrate demand supports Daeduck&amp;rsquo;s outlook. A small difference in pricing, yield, product mix or currency can still absorb the apparent earnings beat.&lt;/p&gt;
&lt;p&gt;Our sector assessment is positive. Our short-term interpretation is more selective: &lt;strong&gt;Daeduck operating profit around KRW80–85 billion would confirm a good quarter, but recent broker forecasts already cluster near the low KRW80 billions. The stronger catalyst would be evidence that Q4 pricing and future customer contracts justify another earnings upgrade.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;This article is dated October 28, 2026, after the Korean regular-session close. Samsung Electro-Mechanics&amp;rsquo; Q3 figures are preliminary actual results. Daeduck&amp;rsquo;s Q3 figures below are forecasts; its results are scheduled for October 30. Prices use the KRX regular-session close, not NXT evening trading. Consensus figures retain their stated capture times.&lt;/p&gt;
&lt;h2 id="samsung-electro-mechanics-showed-that-the-mix-is-changing"&gt;Samsung Electro-Mechanics showed that the mix is changing
&lt;/h2&gt;&lt;p&gt;Samsung Electro-Mechanics (SEMCO, 009150) reported Q3 revenue of &lt;strong&gt;KRW3.763 trillion&lt;/strong&gt; and operating profit of &lt;strong&gt;KRW690.957 billion&lt;/strong&gt;. Revenue rose 8.9% sequentially while operating profit rose 56.9%. Consolidated operating margin improved from about 12.7% to 18.4%. &lt;a class="link" href="https://www.samsungsem.com/resources/file/kr/ir/earnings_release/3Q26_Earnings_Release_kor.pdf" target="_blank" rel="noopener"
 &gt;Official Q3 IR&lt;/a&gt;, &lt;a class="link" href="https://dart.fss.or.kr/dsaf001/main.do?rcpNo=20261028800169" target="_blank" rel="noopener"
 &gt;preliminary filing&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;The divisional revenue figures show why the result is more informative than a generic smartphone-season recovery.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Q3 2026, KRW billions&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Revenue&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Sequential growth&lt;/th&gt;
 &lt;th&gt;Main disclosed drivers&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Component Solutions&lt;/td&gt;
 &lt;td style="text-align: right"&gt;1,883.0&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+14%&lt;/td&gt;
 &lt;td&gt;AI-server, network and power-related MLCC demand&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Package Solutions&lt;/td&gt;
 &lt;td style="text-align: right"&gt;939.0&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+22%&lt;/td&gt;
 &lt;td&gt;Server CPU, AI accelerator, network and automotive substrates&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Optical Solutions&lt;/td&gt;
 &lt;td style="text-align: right"&gt;941.3&lt;/td&gt;
 &lt;td style="text-align: right"&gt;-9%&lt;/td&gt;
 &lt;td&gt;Automotive expansion, offset by sequential camera weakness&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Source: &lt;a class="link" href="https://www.samsungsem.com/resources/file/kr/ir/earnings_release/3Q26_Earnings_Release_kor.pdf" target="_blank" rel="noopener"
 &gt;SEMCO Q3 presentation, business divisions&lt;/a&gt;. Division figures are rounded. Package revenue also rose 58% year on year.&lt;/p&gt;
&lt;p&gt;Optical revenue declined sequentially while the group&amp;rsquo;s margin improved sharply. The evidence supports a richer AI-component mix. It does &lt;strong&gt;not&lt;/strong&gt; disclose how much profit came from package substrates: SEMCO did not provide a standalone FC-BGA margin, ASP or utilization rate in these slides.&lt;/p&gt;
&lt;p&gt;For Q4, the company described additional AI data-center customer demand and plans to respond through new products and capacity. That is useful industry evidence for Daeduck. It is not an order placed with Daeduck. &lt;a class="link" href="https://m.samsungsem.com/kr/newsroom/news/view.do?id=10627" target="_blank" rel="noopener"
 &gt;Official results announcement&lt;/a&gt;.&lt;/p&gt;
&lt;h2 id="a-41-beat-and-a-71-beat-can-describe-the-same-result"&gt;A 4.1% beat and a 7.1% beat can describe the same result
&lt;/h2&gt;&lt;p&gt;The apparent disagreement comes from the denominator, not the reported profit.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;SEMCO operating-profit comparison&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Forecast, KRW bn&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Actual, KRW bn&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Beat&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Infomax pre-results survey, as reported&lt;/td&gt;
 &lt;td style="text-align: right"&gt;About 663.6&lt;/td&gt;
 &lt;td style="text-align: right"&gt;690.957&lt;/td&gt;
 &lt;td style="text-align: right"&gt;About 4.1%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Naver snapshot saved on October 27&lt;/td&gt;
 &lt;td style="text-align: right"&gt;645.3&lt;/td&gt;
 &lt;td style="text-align: right"&gt;690.957&lt;/td&gt;
 &lt;td style="text-align: right"&gt;7.08%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;The Infomax report describes ten firms&amp;rsquo; estimates from the preceding month. The saved Naver value is a different provider snapshot. The analyst populations and cut-off dates need not match. Averaging the two would not create a better consensus. Nor should a post-results estimate be substituted into a pre-results surprise calculation. &lt;a class="link" href="https://news.einfomax.co.kr/news/articleView.html?idxno=4441610" target="_blank" rel="noopener"
 &gt;Infomax report&lt;/a&gt;, &lt;a class="link" href="https://m.stock.naver.com/api/stock/009150/finance/quarter" target="_blank" rel="noopener"
 &gt;Naver quarterly data&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;For the investment argument, both comparisons point in the same direction: profit exceeded expectations, but the magnitude depends on which expectation set the investor uses.&lt;/p&gt;
&lt;h2 id="daeduck-is-a-different-exposure-to-the-substrate-cycle"&gt;Daeduck is a different exposure to the substrate cycle
&lt;/h2&gt;&lt;p&gt;Daeduck (353200) reported Q2 revenue of KRW401.0 billion and operating profit of KRW70.3 billion. Its sales mix was 43% FCCSP, 20% CSP, 23% FC-BGA and 14% multilayer boards, or MLB. FCCSP and CSP serve both memory and non-memory applications; the two percentages cannot simply be labeled “memory exposure.” &lt;a class="link" href="https://www.daeduck.com/file/board/830.do?isInline=true" target="_blank" rel="noopener"
 &gt;Daeduck Q2 IR, pp.3 and 8&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;Three distinctions matter.&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Industry demand transfers; company economics do not.&lt;/strong&gt; SEMCO&amp;rsquo;s package growth supports the direction of high-end substrate demand. Daeduck&amp;rsquo;s own customer approvals, yields, capacity, pricing and currency exposure determine its profit.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;MLCC is not a Daeduck earnings driver.&lt;/strong&gt; SEMCO&amp;rsquo;s reported demand outlook for a specific 1005-size, 47μF capacitor cannot be applied to the whole MLCC market, much less Daeduck&amp;rsquo;s substrates. The reported “about eightfold” demand outlook also follows an earlier “more than sevenfold” statement; it is not a clean new eightfold group-growth forecast. &lt;a class="link" href="https://www.edaily.co.kr/News/Read?mediaCodeNo=257&amp;amp;newsId=04470646645617512" target="_blank" rel="noopener"
 &gt;Current call reporting&lt;/a&gt;, &lt;a class="link" href="https://www.thelec.kr/news/articleView.html?idxno=60326" target="_blank" rel="noopener"
 &gt;prior call transcript report&lt;/a&gt;.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;A peer&amp;rsquo;s contract is not Daeduck&amp;rsquo;s contract.&lt;/strong&gt; SEMCO&amp;rsquo;s October 19 disclosure covers KRW360.49 billion of package-substrate supply during 2027, with no contract deposit or advance payment. The public filing does not identify FC-BGA versus BT, or the customer by name. It cannot establish that Daeduck has secured a funded long-term agreement. &lt;a class="link" href="https://dart.fss.or.kr/dsaf001/main.do?rcpNo=20261019800610" target="_blank" rel="noopener"
 &gt;Supply-contract filing&lt;/a&gt;.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;The distinction is especially useful now: the sector can have a strong demand outlook while individual suppliers capture very different margins and face different funding needs.&lt;/p&gt;
&lt;h2 id="around-krw82-billion-is-already-in-the-discussion"&gt;Around KRW82 billion is already in the discussion
&lt;/h2&gt;&lt;p&gt;Daeduck&amp;rsquo;s recorded October 28 portal consensus was revenue of &lt;strong&gt;KRW418.5 billion&lt;/strong&gt; and operating profit of &lt;strong&gt;KRW77.6 billion&lt;/strong&gt;. The local provider record was captured at 16:35 KST; we do not present it as a later evening update.&lt;/p&gt;
&lt;p&gt;Recent broker reports on the company&amp;rsquo;s IR site were already above that aggregate.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Source and date&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Q3 revenue, KRW bn&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Q3 operating profit, KRW bn&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Q4 operating profit, KRW bn&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Naver aggregate, October 28 snapshot&lt;/td&gt;
 &lt;td style="text-align: right"&gt;418.5&lt;/td&gt;
 &lt;td style="text-align: right"&gt;77.6&lt;/td&gt;
 &lt;td style="text-align: right"&gt;Not used here&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;iM, October 19&lt;/td&gt;
 &lt;td style="text-align: right"&gt;433.0&lt;/td&gt;
 &lt;td style="text-align: right"&gt;81.2&lt;/td&gt;
 &lt;td style="text-align: right"&gt;About 82.0&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Hana, October 23&lt;/td&gt;
 &lt;td style="text-align: right"&gt;432.1&lt;/td&gt;
 &lt;td style="text-align: right"&gt;82.2&lt;/td&gt;
 &lt;td style="text-align: right"&gt;92.0&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Mirae Asset, October 19, separate reference&lt;/td&gt;
 &lt;td style="text-align: right"&gt;419.0&lt;/td&gt;
 &lt;td style="text-align: right"&gt;80.4&lt;/td&gt;
 &lt;td style="text-align: right"&gt;90.5&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Our scenario, not company guidance&lt;/td&gt;
 &lt;td style="text-align: right"&gt;430.2&lt;/td&gt;
 &lt;td style="text-align: right"&gt;81.7&lt;/td&gt;
 &lt;td style="text-align: right"&gt;No numerical forecast&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Sources: &lt;a class="link" href="https://m.stock.naver.com/api/stock/353200/finance/quarter" target="_blank" rel="noopener"
 &gt;Naver quarterly consensus&lt;/a&gt;, &lt;a class="link" href="https://www.daeduck.com/file/board/851.do" target="_blank" rel="noopener"
 &gt;iM report&lt;/a&gt;, &lt;a class="link" href="https://www.daeduck.com/file/board/856.do" target="_blank" rel="noopener"
 &gt;Hana report&lt;/a&gt;, &lt;a class="link" href="https://www.daeduck.com/file/board/852.do" target="_blank" rel="noopener"
 &gt;Mirae Asset report&lt;/a&gt;. These are selected reports, not the distribution of every investor&amp;rsquo;s expectations. Mirae&amp;rsquo;s report carries a separate-financial-statement footnote, so we keep it separate and do not pool the figures into a new consensus.&lt;/p&gt;
&lt;p&gt;We use iM&amp;rsquo;s PDF value of KRW81.2 billion rather than a rounded web summary. Hana&amp;rsquo;s KRW82.2 billion profit divided by KRW432.1 billion revenue implies about 19.0% margin, consistent with its table; the report&amp;rsquo;s prose says 19.3%. The arithmetic takes precedence.&lt;/p&gt;
&lt;p&gt;The sharper disagreement is in &lt;strong&gt;Q4&lt;/strong&gt;, where iM and Hana differ by about KRW10 billion. Their Q3 profit forecasts differ by only KRW1 billion. That gap makes the timing and quality of the next price increase more interesting than a headline Q3 result near KRW82 billion.&lt;/p&gt;
&lt;p&gt;&lt;img alt="Two selected broker forecasts, KRW billions. iM projects Q3 operating profit 81.2 and Q4 82.0; Hana projects Q3 82.2 and Q4 92.0. Q3 gap 1.0, Q4 gap 10.0. Not company guidance or a full market distribution." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/korea-substrates-semco-daeduck-2026-10-28/expectations.en.png"&gt;
&lt;/p&gt;
&lt;h2 id="one-margin-point-can-erase-the-modeled-beat"&gt;One margin point can erase the modeled beat
&lt;/h2&gt;&lt;p&gt;Our starting point is Daeduck&amp;rsquo;s official Q2 revenue: KRW343.5 billion of package substrates and KRW57.5 billion of MLB. We apply separate growth assumptions rather than copy SEMCO&amp;rsquo;s 22% package growth.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Scenario&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Package growth QoQ&lt;/th&gt;
 &lt;th style="text-align: right"&gt;MLB growth QoQ&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Group margin&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Revenue, KRW bn&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Operating profit, KRW bn&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Downside&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+2%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+5%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;17.5%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;410.7&lt;/td&gt;
 &lt;td style="text-align: right"&gt;71.9&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Base&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+6%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+15%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;19.0%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;430.2&lt;/td&gt;
 &lt;td style="text-align: right"&gt;81.7&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Upside&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+8%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+20%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;20.0%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;440.0&lt;/td&gt;
 &lt;td style="text-align: right"&gt;88.0&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;The base calculation is &lt;code&gt;343.5 × 1.06 + 57.5 × 1.15 = KRW430.235bn&lt;/code&gt;, followed by &lt;code&gt;430.235 × 19% = KRW81.74465bn&lt;/code&gt;. Rounded to the nearest KRW1 billion, that is revenue of about KRW430 billion and profit of about KRW82 billion.&lt;/p&gt;
&lt;p&gt;We use &lt;strong&gt;KRW80–85 billion&lt;/strong&gt; as a practical central profit range. It is about 3.1–9.5% above the KRW77.6 billion consensus. It is not a statistical confidence interval; the downside and upside scenarios lie outside it.&lt;/p&gt;
&lt;p&gt;At base revenue, one percentage point of operating margin changes profit by &lt;strong&gt;KRW4.302 billion&lt;/strong&gt;. The base profit exceeds consensus by &lt;strong&gt;KRW4.145 billion&lt;/strong&gt;. A margin near 18.04%, rather than 19%, would bring the same revenue model back to consensus.&lt;/p&gt;
&lt;p&gt;&lt;img alt="Daeduck profit sensitivity at fixed revenue of KRW430.235bn. At 19 percent margin profit is KRW81.745bn. The KRW77.6bn consensus is matched at about 18.04 percent margin. A one percentage point change changes profit by KRW4.302bn. This is a scenario calculation." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/korea-substrates-semco-daeduck-2026-10-28/margin-sensitivity.en.png"&gt;
&lt;/p&gt;
&lt;p&gt;This does not predict a margin miss. It shows why reported demand strength alone is insufficient. Product yield, higher-value shipments, price realization and exchange rates can determine whether the small modeled beat exists at all.&lt;/p&gt;
&lt;h2 id="the-price-increase-must-exceed-the-cost-increase"&gt;The price increase must exceed the cost increase
&lt;/h2&gt;&lt;p&gt;An increase that merely passes through raw-material inflation protects economics. An increase driven by scarce capacity can expand the amount retained by the supplier. Investors need to establish which mechanism is operating.&lt;/p&gt;
&lt;p&gt;iM&amp;rsquo;s October 19 report distinguishes the earlier raw-material pass-through from the proposed later price adjustment tied to tighter supply. Its illustrative calculation assumes 2027 memory-substrate sales of KRW706 billion, a 10% price increase and a 35% contribution margin, producing about &lt;strong&gt;KRW24.7 billion of additional annual operating profit&lt;/strong&gt;. Those are broker assumptions, not a signed price agreement or our forecast. &lt;a class="link" href="https://www.daeduck.com/file/board/851.do" target="_blank" rel="noopener"
 &gt;iM report, p.1&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;The application date matters as much as the quoted percentage. A negotiated increase affects profit only when qualifying products ship at the new terms. Mix changes, customer concessions and material costs can offset part of the headline increase.&lt;/p&gt;
&lt;p&gt;The other lever is manufacturing execution. Broker reports describe improving yields on large FC-BGA products and MLB capacity ramping. Those assessments need confirmation in company disclosures and subsequent financial results. A broker&amp;rsquo;s reported yield should not become a company-issued fact without attribution.&lt;/p&gt;
&lt;h2 id="higher-cash-is-not-all-free-cash"&gt;Higher cash is not all free cash
&lt;/h2&gt;&lt;p&gt;SEMCO&amp;rsquo;s Q3 presentation reports &lt;strong&gt;KRW1.1452 trillion of quarterly operating cash flow&lt;/strong&gt; and &lt;strong&gt;KRW862.7 billion of property, plant, equipment and intangible additions&lt;/strong&gt;. Subtracting the latter leaves a simple residual of &lt;strong&gt;KRW282.5 billion&lt;/strong&gt;. This is our calculation, not an official company-defined FCF metric.&lt;/p&gt;
&lt;p&gt;The company&amp;rsquo;s cash balance rose by KRW799.3 billion, but financing activities also supplied KRW551.1 billion; the borrowing-flow line was positive KRW589.1 billion. The cash increase cannot all be described as operating generation. These are Q3 standalone cash flows, not nine-month cumulative figures. &lt;a class="link" href="https://www.samsungsem.com/resources/file/kr/ir/earnings_release/3Q26_Earnings_Release_kor.pdf" target="_blank" rel="noopener"
 &gt;Official IR cash-flow slide&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;For Daeduck, the October 22 facility-investment decision is &lt;strong&gt;KRW615 billion&lt;/strong&gt;, scheduled through December 31, 2030. The filing cites internal funds and external borrowing among funding sources. It is a multi-year investment plan, not a single-quarter cash payment, and a decision made in October cannot be the direct cause of July–September earnings. &lt;a class="link" href="https://dart.fss.or.kr/dsaf001/main.do?rcpNo=20261022800677" target="_blank" rel="noopener"
 &gt;Investment filing&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;The investment case needs the sequence: customer commitments, capacity commissioning, acceptable yields, shipments, collections, then returns on the capital spent. New capacity can improve growth and simultaneously make cash-flow verification more important.&lt;/p&gt;
&lt;h2 id="todays-prices-already-assume-substantial-improvement"&gt;Today&amp;rsquo;s prices already assume substantial improvement
&lt;/h2&gt;&lt;p&gt;At the October 28 regular-session close, Daeduck was &lt;strong&gt;KRW138,000, up 8.07%&lt;/strong&gt;, and SEMCO was &lt;strong&gt;KRW1,306,000, up 8.74%&lt;/strong&gt;. Daeduck was 15.29% above its October 26 close of KRW119,700. KOSPI also rose sharply, closing at 6,012.44, up 6.63%. None of these gains is a clean estimate of the effect of SEMCO&amp;rsquo;s announcement alone. &lt;a class="link" href="https://www.kokstock.com/stock/item.asp?item_cd=353200" target="_blank" rel="noopener"
 &gt;Daeduck closing data&lt;/a&gt;, &lt;a class="link" href="https://markets.hankyung.com/stock/009150/public-announcement" target="_blank" rel="noopener"
 &gt;SEMCO closing data&lt;/a&gt;, &lt;a class="link" href="https://www.yna.co.kr/view/AKR20261028142400008" target="_blank" rel="noopener"
 &gt;market-close report&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;The following gives a common-date valuation reference, not an apples-to-apples fair-value ranking.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Company&lt;/th&gt;
 &lt;th style="text-align: right"&gt;October 28 KRX close&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Recorded FY2027 consensus EPS&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Price / forecast EPS&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Samsung Electro-Mechanics&lt;/td&gt;
 &lt;td style="text-align: right"&gt;KRW1,306,000&lt;/td&gt;
 &lt;td style="text-align: right"&gt;KRW41,821&lt;/td&gt;
 &lt;td style="text-align: right"&gt;31.2×&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Daeduck Electronics&lt;/td&gt;
 &lt;td style="text-align: right"&gt;KRW138,000&lt;/td&gt;
 &lt;td style="text-align: right"&gt;KRW6,613&lt;/td&gt;
 &lt;td style="text-align: right"&gt;20.9×&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;EPS source: Naver/WiseReport snapshot captured at 16:38 KST on October 28, &lt;a class="link" href="https://navercomp.wisereport.co.kr/v2/company/c1010001.aspx?cmp_cd=009150" target="_blank" rel="noopener"
 &gt;SEMCO&lt;/a&gt; and &lt;a class="link" href="https://navercomp.wisereport.co.kr/v2/company/c1010001.aspx?cmp_cd=353200" target="_blank" rel="noopener"
 &gt;Daeduck&lt;/a&gt;. Analyst counts and detailed EPS accounting-scope flags were not preserved in this snapshot. The estimates are neither reported earnings nor guaranteed outcomes.&lt;/p&gt;
&lt;p&gt;SEMCO offers exposure to both MLCC and package substrates, with actual Q3 earnings evidence. Daeduck is more concentrated in substrates and boards, but its next earnings release and new-capacity economics still have to be tested. The lower forecast multiple alone does not establish that Daeduck is the better investment.&lt;/p&gt;
&lt;p&gt;If 2027 EPS were 20% below those estimates with prices unchanged, the respective multiples would rise to about &lt;strong&gt;39.0× and 26.1×&lt;/strong&gt;. That is a sensitivity, not a forecast. The price paid for future improvement matters alongside the industry story.&lt;/p&gt;
&lt;h2 id="how-to-read-daeducks-october-30-release"&gt;How to read Daeduck&amp;rsquo;s October 30 release
&lt;/h2&gt;&lt;p&gt;The following are conditional interpretations, not probabilities or price targets. The chart references originated in the October 28 intraday analysis; the KRW138,000 close remained inside its KRW132,000–141,000 observation zone. We do not reuse the intraday moving-average value as a confirmed closing average.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Result and outlook&lt;/th&gt;
 &lt;th&gt;Investment interpretation&lt;/th&gt;
 &lt;th&gt;Dated price reference&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Profit clearly above KRW85bn, with concrete Q4 pricing or customer-contract improvement&lt;/td&gt;
 &lt;td&gt;Supports another estimate upgrade; quality and recurring contribution still matter&lt;/td&gt;
 &lt;td&gt;Sustained trading above KRW141,000 would strengthen the case for retesting KRW145,000–150,000&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Profit around KRW80–85bn, positive direction but little new information&lt;/td&gt;
 &lt;td&gt;Good earnings can coexist with profit-taking after expectations rose&lt;/td&gt;
 &lt;td&gt;Observe KRW132,000–141,000 rather than assume an automatic breakout&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Profit below KRW80bn or a weak Q4 explanation&lt;/td&gt;
 &lt;td&gt;Could disappoint the recently upgraded broker expectations even near portal consensus&lt;/td&gt;
 &lt;td&gt;Failure around KRW132,000–134,600 would put the earlier KRW127,500 area back in view&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;These levels must be reconsidered if the market or pre-release price changes materially. A one-off profit gain without better pricing, utilization or cash collection would not carry the same meaning as recurring improvement.&lt;/p&gt;
&lt;p&gt;The company scheduled preliminary consolidated results for &lt;strong&gt;October 30&lt;/strong&gt;, followed by an investor roadshow on &lt;strong&gt;November 4–6&lt;/strong&gt;. Missing contract detail on the first date should remain an open question, not automatically become a failure. &lt;a class="link" href="https://dart.fss.or.kr/dsaf001/main.do?rcpNo=20261015800458" target="_blank" rel="noopener"
 &gt;Results notice&lt;/a&gt;, &lt;a class="link" href="https://dart.fss.or.kr/dsaf001/main.do?rcpNo=20261015800468" target="_blank" rel="noopener"
 &gt;IR notice&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;For sector context, Samsung Electronics&amp;rsquo; detailed earnings call is scheduled for &lt;strong&gt;October 29 at 10 a.m. KST&lt;/strong&gt;. Its October 8 preliminary release is a separate event. Tomorrow&amp;rsquo;s memory and downstream-business details may inform the substrate backdrop, but they cannot establish Daeduck&amp;rsquo;s own shipments or margins. &lt;a class="link" href="https://www.samsung.com/sec/ir/ir-events-presentations/events/" target="_blank" rel="noopener"
 &gt;Samsung Electronics calendar&lt;/a&gt;.&lt;/p&gt;
&lt;h2 id="our-investment-judgment"&gt;Our investment judgment
&lt;/h2&gt;&lt;p&gt;&lt;strong&gt;We are constructive on the substrate cycle and selective on individual suppliers.&lt;/strong&gt; SEMCO has delivered a strong group result and a clear package-demand signal. Daeduck&amp;rsquo;s central Q3 earnings case is also constructive, but it is close to estimates already published before SEMCO&amp;rsquo;s release.&lt;/p&gt;
&lt;p&gt;The evidence that would justify greater confidence is now specific: price increases that improve margin after material costs, stable yields as higher-value capacity ramps, and customer contracts that clarify who funds the expansion and when cash is collected. Until those details improve, a result near KRW82 billion is primarily confirmation of the current case.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Which would change your view most on October 30: a larger Q3 profit beat, evidence of Q4 pricing beyond cost pass-through, or a signed customer agreement that supports expansion funding? Pick one and explain what would count as convincing evidence.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Research and information only; not investment advice. Forecasts and price scenarios are analytical assumptions, not company guidance, trade instructions or personalized portfolio decisions. Figures are date-bound. AI tools assisted research and editing.&lt;/em&gt;&lt;/p&gt;</description></item><item><title>Anthropic's $132 Billion Cash Pile: Read the Exit Clauses</title><link>https://koreainvestinsights.com/post/anthropic-cash-profit-cancellation-rights-2026-10-27/</link><pubDate>Tue, 27 Oct 2026 11:00:00 +0900</pubDate><guid>https://koreainvestinsights.com/post/anthropic-cash-profit-cancellation-rights-2026-10-27/</guid><description>&lt;img src="https://koreainvestinsights.com/" alt="Featured image of post Anthropic's $132 Billion Cash Pile: Read the Exit Clauses" /&gt;&lt;p&gt;At June 30, Anthropic reported &lt;strong&gt;$131.651 billion of cash, cash equivalents and short-term investments&lt;/strong&gt;. Amazon reported $122.988 billion of cash and marketable securities on the same date. Meta had $90.260 billion; Microsoft, $76.843 billion.&lt;/p&gt;
&lt;p&gt;The comparison is striking, but the categories are not identical. Meta&amp;rsquo;s securities include equities, investment maturities differ, and these totals exclude separately reported restricted balances. They are reported balance-sheet amounts, not a ranking of freely spendable cash or financial strength. Still, they are hard to reconcile with a story in which Anthropic&amp;rsquo;s immediate problem is simply an empty bank account. &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/2130357/000162828026067168/anthropicpbc-sx1.htm" target="_blank" rel="noopener"
 &gt;Anthropic S-1&lt;/a&gt;, &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1018724/000101872426000026/amzn-20260630.htm" target="_blank" rel="noopener"
 &gt;Amazon 10-Q&lt;/a&gt;, &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1326801/000162828026050705/meta-20260630.htm" target="_blank" rel="noopener"
 &gt;Meta 10-Q&lt;/a&gt;, &lt;a class="link" href="https://www.microsoft.com/en-us/Investor/earnings/FY-2026-Q4/press-release-webcast" target="_blank" rel="noopener"
 &gt;Microsoft results&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;The more useful question is who can stop spending when demand, technology and delivery schedules change. Anthropic&amp;rsquo;s customers mostly pay for usage. Some infrastructure providers have minimum spending commitments extending to 2033 or 2036. Others have contracts that can be cancelled after an initial period. Broadcom is helping finance equipment and support its residual value.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Paid demand and quarterly profitability have become more credible. The timing and allocation of the risk have become more complicated.&lt;/strong&gt; For investors in AI infrastructure and Korean memory, the next useful unit of analysis is the contract: when payment begins, who can cancel, and who carries the asset if the original plan changes.&lt;/p&gt;
&lt;p&gt;This article is dated October 27, 2026. Financial statements run through June 30; customer metrics reach September; some financing and contract amendments are from October. We keep those dates separate. IPO price and fully diluted share count are not established, so this is an analysis of business economics and the supply chain, not an IPO valuation.&lt;/p&gt;
&lt;p&gt;&lt;img alt="June 30, 2026 reported cash and short-term or marketable investments: Anthropic $131.651bn, Amazon $122.988bn, Meta $90.260bn, Microsoft $76.843bn. Categories and maturity mix differ." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/anthropic-cash-profit-cancellation-rights-2026-10-27/cash.en.png"&gt;
&lt;/p&gt;
&lt;h2 id="the-cash-pile-came-mostly-from-financing-the-operating-progress-is-real-too"&gt;The cash pile came mostly from financing. The operating progress is real too.
&lt;/h2&gt;&lt;p&gt;Anthropic received $110.432 billion of net proceeds from redeemable convertible preferred stock in the first half. Operating cash flow was positive $3.929 billion. Those are separate cash-flow sources; the balance sheet does not tell us that operating profits built the entire reserve.&lt;/p&gt;
&lt;p&gt;It would be equally wrong to dismiss the operating cash as entirely a prepayment effect. The disclosed working-capital movements contributed a net $1.168 billion. Subtracting that contribution leaves $2.761 billion of operating cash flow. Stock-based compensation remains a noncash add-back, and payments on future infrastructure commitments have not all begun. &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/2130357/000162828026067168/anthropicpbc-sx1.htm" target="_blank" rel="noopener"
 &gt;S-1, cash-flow statement F-13&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;The underlying business also crossed a meaningful threshold. Q2 GAAP operating profit was $974 million, after a $1.667 billion operating loss in Q1.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;US$ billions, except percentages&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Q1 2026&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Q2 2026&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Revenue&lt;/td&gt;
 &lt;td style="text-align: right"&gt;4.731&lt;/td&gt;
 &lt;td style="text-align: right"&gt;11.555&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Compute and infrastructure expense&lt;/td&gt;
 &lt;td style="text-align: right"&gt;3.387&lt;/td&gt;
 &lt;td style="text-align: right"&gt;5.558&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Other operating expenses, combined&lt;/td&gt;
 &lt;td style="text-align: right"&gt;3.011&lt;/td&gt;
 &lt;td style="text-align: right"&gt;5.023&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;GAAP operating profit / loss&lt;/td&gt;
 &lt;td style="text-align: right"&gt;-1.667&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+0.974&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Compute and infrastructure / revenue&lt;/td&gt;
 &lt;td style="text-align: right"&gt;71.6%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;48.1%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Source: &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/2130357/000162828026067168/anthropicpbc-sx1.htm#ica694c859b8e458dbf6a73f56dde9d42_2979" target="_blank" rel="noopener"
 &gt;S-1 quarterly statements of operations&lt;/a&gt;. “Other” combines personnel, sales/marketing/partnerships, and other operating/administrative costs. It is not an adjusted-profit measure.&lt;/p&gt;
&lt;p&gt;Revenue rose 144.2% quarter on quarter while compute and infrastructure expense rose 64.1%. The $2.641 billion improvement in operating profit divided by $6.824 billion of additional revenue is &lt;strong&gt;38.7%&lt;/strong&gt;. That is the incremental operating-profit conversion between these two quarters. It is not a forecast of the long-run margin.&lt;/p&gt;
&lt;p&gt;The demand evidence extends beyond a few large buyers. By September, more than 11,900 customers generated at least $100,000 in trailing-twelve-month revenue. The share of revenue outside the top five rose from about 64% in Q1 to 79% in Q3, while each of the top five customers also reached a quarterly revenue record. Broader adoption and larger major customers can coexist.&lt;/p&gt;
&lt;p&gt;The approximately 900% net revenue retention figure needs its denominator. It compares the same customer cohort&amp;rsquo;s aggregate recent three-month revenue with the corresponding prior-year period, including customers that previously spent as little as $1; personal Pro and Max subscribers are excluded. It demonstrates rapid expansion from early adoption. It does not establish a mature annual ninefold growth rate. &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/2130357/000162828026067168/anthropicpbc-sx1.htm" target="_blank" rel="noopener"
 &gt;S-1, customer metrics and definitions&lt;/a&gt;.&lt;/p&gt;
&lt;h2 id="the-holiday-statistic-is-more-revealing-than-another-token-chart"&gt;The holiday statistic is more revealing than another token chart
&lt;/h2&gt;&lt;p&gt;One small S-1 disclosure makes the usage model tangible: customer spending in the two weeks ended January 2, 2026 was approximately &lt;strong&gt;20% lower&lt;/strong&gt; than in the two weeks ended December 19, 2025. The company links this to holiday patterns and slower customer workflows.&lt;/p&gt;
&lt;p&gt;This is not evidence of a structural demand collapse. It is an observed example of how quickly revenue can respond to the working calendar. About 80% of first-half revenue was usage-based. That flexibility is attractive to customers; it is a different economic promise from paying for reserved infrastructure regardless of usage. &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/2130357/000162828026067168/anthropicpbc-sx1.htm" target="_blank" rel="noopener"
 &gt;S-1, customer usage patterns and revenue recognition&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;Growth can obscure this difference. When new customers and workloads expand extremely quickly, a holiday slowdown is absorbed by the trend. When growth becomes less exceptional, the mismatch between flexible customer spending and committed capacity becomes easier to see.&lt;/p&gt;
&lt;p&gt;That is why another increase in tokens is not enough to settle the investment case. We need dollars collected, the cost of serving the work, and the obligations that remain when activity slows.&lt;/p&gt;
&lt;h2 id="a-1205-billion-maximum-can-be-more-flexible-than-a-449-billion-minimum"&gt;A $120.5 billion maximum can be more flexible than a $44.9 billion minimum
&lt;/h2&gt;&lt;p&gt;The contract table changes the meaning of the headline numbers. All amounts below are in US$ billions and come from the October 26 S-1&amp;rsquo;s description of the arrangements. They are &lt;strong&gt;not additive measures of equivalent debt&lt;/strong&gt;.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Counterparty&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Disclosed amount&lt;/th&gt;
 &lt;th&gt;Payment and exit terms that matter&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Microsoft&lt;/td&gt;
 &lt;td style="text-align: right"&gt;At least 44.9&lt;/td&gt;
 &lt;td&gt;November 2026–May 2033, regardless of usage. Limited exits for uncured breach or undelivered tranches.&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Google&lt;/td&gt;
 &lt;td style="text-align: right"&gt;At least 114.5&lt;/td&gt;
 &lt;td&gt;April 2026–July 2033. Shortfall true-up, subject to stated adjustments and exceptions. Limited termination rights.&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Amazon&lt;/td&gt;
 &lt;td style="text-align: right"&gt;At least 110.0&lt;/td&gt;
 &lt;td&gt;May 2026–April 2036. Shortfall true-up, subject to stated adjustments and exceptions. Limited termination rights.&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;SpaceXAI&lt;/td&gt;
 &lt;td style="text-align: right"&gt;Up to 120.5&lt;/td&gt;
 &lt;td&gt;Roughly 3.1 per month after full delivery. Following initial periods ending in January and March 2027, either party may cancel with 90 days&amp;rsquo; written notice.&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Nscale&lt;/td&gt;
 &lt;td style="text-align: right"&gt;Up to 44.6&lt;/td&gt;
 &lt;td&gt;Fees accrue as capacity is delivered and accepted. Financing and delivery requirements matter.&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Lambda&lt;/td&gt;
 &lt;td style="text-align: right"&gt;Approximately 41.3&lt;/td&gt;
 &lt;td&gt;Fees accrue on delivery and acceptance. Some capacity has facility-lease and delivery conditions.&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Broadcom equipment leases&lt;/td&gt;
 &lt;td style="text-align: right"&gt;Approximately 161.2, excluding interest&lt;/td&gt;
 &lt;td&gt;Multiple five-year finance leases. Payments normally start when equipment is ready for service, but can start earlier in some facility-delay circumstances.&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Source: &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/2130357/000162828026067168/anthropicpbc-sx1.htm" target="_blank" rel="noopener"
 &gt;S-1, infrastructure service arrangements and equipment leases, pp.171–173&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;Microsoft, Google and Amazon&amp;rsquo;s stated minimums total $269.4 billion, spread over different multi-year periods. That is not an annual bill. SpaceXAI&amp;rsquo;s maximum amount is not an equally firm minimum. Its January and March dates do not mean spending stops immediately then: the applicable initial period must end and the 90-day notice process still matters.&lt;/p&gt;
&lt;p&gt;For a supplier investor, the implication is concrete. Two equally large “backlogs” can carry very different cash-flow protection. One may place more utilization risk on the customer. Another may leave the supplier with capacity to re-let if the customer exercises an exit right. Cancellation rights do not tell us that cancellation will occur; they tell us who has a choice.&lt;/p&gt;
&lt;p&gt;There is a second asymmetry in Anthropic Direct. Equipment, facilities, power and operations come from different counterparties. The filing warns that one supplier&amp;rsquo;s performance can trigger payment even when another supplier&amp;rsquo;s delay prevents usable compute. A functioning order book does not remove the coordination risk.&lt;/p&gt;
&lt;h2 id="the-63-billion-deal-finances-delivery-broadcom-also-carries-asset-risk"&gt;The $63 billion deal finances delivery. Broadcom also carries asset risk.
&lt;/h2&gt;&lt;p&gt;The October 26 announcement describes a financing structure in which an SPV buys Broadcom XPU accelerators and networking equipment and leases them to Anthropic. Deliveries are expected to begin in 2027. This is meaningful evidence that an existing compute plan has financing behind it. It improves the path from announced capacity to equipment orders and delivery. &lt;a class="link" href="https://www.prnewswire.com/news-releases/bank-of-america-blackstone-citi-and-morgan-stanley-announce-landmark-63-billion-financing-structure-to-enable-ai-compute-capacity-in-partnership-with-broadcom-302919849.html" target="_blank" rel="noopener"
 &gt;Financing partners&amp;rsquo; announcement&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;The less obvious part is Broadcom&amp;rsquo;s residual-value support, described as an amount equivalent to the senior financing. A supplier&amp;rsquo;s economic exposure can therefore extend beyond shipping equipment. The public announcement does not establish an unconditional guarantee of every dollar of loan principal and interest, or disclose every condition for a claim. “Residual-value support” should not be silently rewritten as “full debt guarantee.”&lt;/p&gt;
&lt;p&gt;The transaction also should not be added to final customer demand. Financing an existing supply arrangement is not a second customer purchase.&lt;/p&gt;
&lt;p&gt;A separate SEC agreement provides up to $42 billion of Broadcom financing for specified quarterly equipment-lease fees. It is not unrestricted cash delivered to Anthropic. The proceeds settle the designated lease payments, and draw conditions apply. We cannot fully reconcile the equipment and capacity overlap between the $63 billion structure and this $42 billion agreement from the public documents. Calling their sum $105 billion of new demand or drawn debt would be unsupported. &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/2130357/000162828026067168/exhibit44-sx1.htm" target="_blank" rel="noopener"
 &gt;SEC Exhibit 4.4&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;The bullish interpretation is that financing, dedicated capacity and cheaper long-term compute relieve today&amp;rsquo;s supply constraint. The bearish interpretation is that utilization and asset-value risk are being distributed across the model company, lessor, equipment supplier and lenders. Both can be true while equipment shipments grow.&lt;/p&gt;
&lt;h2 id="a-175-compute-cost-increase-would-erase-q2-profit-if-revenue-stood-still"&gt;A 17.5% compute-cost increase would erase Q2 profit if revenue stood still
&lt;/h2&gt;&lt;p&gt;This is a sensitivity test, not a forecast. Start with Q2 revenue of $11.555 billion, compute and infrastructure expense of $5.558 billion, and all other operating expenses of $5.023 billion.&lt;/p&gt;
&lt;p&gt;Hold revenue and other expenses constant. The compute-cost increase that absorbs the $974 million operating profit is:&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;$0.974 billion ÷ $5.558 billion = 17.5%.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;At a 20% compute-cost increase, operating profit becomes approximately &lt;strong&gt;negative $138 million&lt;/strong&gt;. But let revenue rise 10% with the same cost shock and profit is about &lt;strong&gt;positive $1.018 billion&lt;/strong&gt;. Continued revenue growth can absorb new costs. The test is whether revenue arrives quickly enough.&lt;/p&gt;
&lt;p&gt;&lt;img alt="Illustrative operating profit in US$bn. For compute cost changes of 0, 20 and 40 percent: unchanged revenue gives 0.974, -0.138, -1.249; revenue plus 10 percent gives 2.130, 1.018, -0.094; revenue plus 20 percent gives 3.285, 2.173, 1.062. Other expenses fixed. Not a forecast." loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/anthropic-cash-profit-cancellation-rights-2026-10-27/profit-sensitivity.en.png"&gt;
&lt;/p&gt;
&lt;p&gt;Calculation: &lt;code&gt;0.974 + 11.555 × revenue growth − 5.558 × compute-cost growth&lt;/code&gt;. All other expenses stay fixed. In reality, sales fees, staffing, research and product mix can also change. These cells are not estimated next-quarter results.&lt;/p&gt;
&lt;p&gt;The S-1 explicitly describes compute becoming available in large steps, with expense recognized before the full associated revenue is realized. It also expects increased use of more expensive short-term compute in the near term. A margin decline after a strong quarter could therefore reflect timing, not automatically the end of demand. It would still matter to cash flow and valuation.&lt;/p&gt;
&lt;h2 id="financing-can-preserve-liquidity-while-increasing-the-common-shareholders-bill"&gt;Financing can preserve liquidity while increasing the common shareholder&amp;rsquo;s bill
&lt;/h2&gt;&lt;p&gt;The $42 billion agreement makes this distinction unusually visible. Notes issued outside the specified breach period carry 9% annual interest; notes issued during that period carry 13%. Interest is capitalized quarterly. After a liquidity event, conversion uses 90% or 85% of the preceding 30-trading-day VWAP, respectively. Some trigger details are redacted. There is no basis here to say the adverse terms have been triggered. &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/2130357/000162828026067168/exhibit44-sx1.htm" target="_blank" rel="noopener"
 &gt;Exhibit 4.4, note purchase agreement&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;For illustration, assume $1 drawn, four equal quarters, no interim conversion, and the same reference share price at year end:&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Scenario&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Year-end claim after PIK&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Conversion-price fraction&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Relative shares issued at the same reference price&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;9% annual rate&lt;/td&gt;
 &lt;td style="text-align: right"&gt;1.0931&lt;/td&gt;
 &lt;td style="text-align: right"&gt;90%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;1.2145&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;13% annual rate&lt;/td&gt;
 &lt;td style="text-align: right"&gt;1.1365&lt;/td&gt;
 &lt;td style="text-align: right"&gt;85%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;1.3370&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;The second case requires about &lt;strong&gt;10.1% more shares&lt;/strong&gt; than the first. A further 20% decline in the reference share price increases the shares required by another 25% in either case. This compares hypothetical share issuance, not an annual cash interest rate or a prediction of actual dilution. The contract uses actual elapsed days/365, so the equal-quarter approximation is not an exact payment schedule.&lt;/p&gt;
&lt;p&gt;The cash pile offers a buffer. It does not make these terms irrelevant. The same distinction applies to equity compensation: the S-1 describes a hypothetical $2.9 billion cumulative RSU charge if the IPO condition had been met at June 30, followed by $23.7 billion of unrecognized expense over a weighted-average roughly 2.1 years. Recognition is accelerated, not a flat annual amount. Cash charges, accounting expense and per-share dilution must be modeled separately. &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/2130357/000162828026067168/anthropicpbc-sx1.htm" target="_blank" rel="noopener"
 &gt;S-1, share-based compensation&lt;/a&gt;.&lt;/p&gt;
&lt;h2 id="the-854-margin-is-not-the-return-available-to-shareholders"&gt;The 85.4% margin is not the return available to shareholders
&lt;/h2&gt;&lt;p&gt;Using the functional cost-of-revenue disclosure, first-half revenue of $16.286 billion less $2.380 billion of cost of revenue gives an 85.4% gross margin. It excludes costs elsewhere in the accounts, including research compute and distribution fees. Marketplace sales are recognized gross, with channel fees in sales, marketing and partnerships.&lt;/p&gt;
&lt;p&gt;The nature-of-expense statement gives a different bridge: $16.286 billion of revenue minus $8.945 billion of total compute/infrastructure, $3.168 billion of sales/marketing/partnerships, $3.152 billion of personnel and $1.714 billion of other expenses leaves a &lt;strong&gt;$693 million first-half operating loss&lt;/strong&gt;. Do not subtract the $2.380 billion functional cost again; that would double-count costs. First-half loss and Q2 profit describe different periods. &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/2130357/000162828026067168/anthropicpbc-sx1.htm" target="_blank" rel="noopener"
 &gt;S-1, financial statements and Note 8&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;There is also a product reason to resist treating research as an easily removable expense. More than 80% of Q3 revenue came from models released during the preceding six months. That shows fast commercialization. It also suggests continuing model development helps defend today&amp;rsquo;s revenue. It does not prove that every old model has a six-month economic life.&lt;/p&gt;
&lt;p&gt;Price cuts have a similar two-sided effect. Anthropic says Fable 5.1 can cost customers about 25% less for typical workloads and up to about 45% less for highly agentic work, helped by lower cache-read pricing. That can unlock new uses. It is not evidence that Anthropic&amp;rsquo;s own compute costs fell by the same percentages. &lt;a class="link" href="https://www.anthropic.com/claude-fable-and-mythos-5-1" target="_blank" rel="noopener"
 &gt;Official model announcement&lt;/a&gt;.&lt;/p&gt;
&lt;h2 id="for-korean-memory-a-reserved-building-is-not-a-chip-order"&gt;For Korean memory, a reserved building is not a chip order
&lt;/h2&gt;&lt;p&gt;One of the most useful disclosures is Anthropic Direct&amp;rsquo;s sequencing: secure data-center capacity early, then defer the larger accelerator commitment until compute needs are clearer. Buildings need long lead times; the accelerator choice can be made later. That is a valuable option for the buyer, and a reason suppliers should distinguish site capacity from committed chip procurement. &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/2130357/000162828026067168/anthropicpbc-sx1.htm" target="_blank" rel="noopener"
 &gt;S-1, compute investments&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;The filing also describes using Trainium, TPU, NVIDIA and AMD hardware, and names Samsung Electronics, SK hynix and Micron as major high-performance-memory manufacturers. Diversifying accelerators does not automatically eliminate demand for high-performance memory. It also does not establish any memory supplier&amp;rsquo;s order value, HBM generation, bit content or share.&lt;/p&gt;
&lt;p&gt;For 2027, funded equipment purchases, actual delivery, customer acceptance and collections are useful evidence. For 2028 and beyond, we need the next round of orders and the economics of the workloads supporting them. A facility announcement, a financing commitment, a GPU order and a paid workload belong at different points in that chain.&lt;/p&gt;
&lt;h2 id="the-strongest-bull-case-is-workflow-ownership"&gt;The strongest bull case is workflow ownership
&lt;/h2&gt;&lt;p&gt;Among customers generating at least $100,000 in trailing revenue, more than 50% used both Claude Platform and at least one first-party application in Q3, up from approximately 30% in Q4 2025. The company says it contracts directly with end customers and sets its own prices even through cloud marketplaces. These are signs of a developing customer relationship beyond selling interchangeable tokens. They do not yet quantify switching costs. &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/2130357/000162828026067168/anthropicpbc-sx1.htm" target="_blank" rel="noopener"
 &gt;S-1, adoption and revenue recognition&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;If applications become embedded in valuable work, a lower cost per completed task can expand paid usage fast enough to absorb new capacity. That is the strongest case for improving operating leverage. It also changes the opportunity for enterprise AI and systems-integration providers: implementation demand can grow while the model company keeps the customer contract, distribution and pricing relationship.&lt;/p&gt;
&lt;p&gt;Commercially usable work has boundaries. The filing estimates that 5–7% of revenue in the first nine months came from customers later judged to breach usage policies or commercial terms. That is not synonymous with illegal activity. Enforcement may affect future reported growth, so a slowdown cannot automatically be attributed to weak underlying compute demand.&lt;/p&gt;
&lt;p&gt;Separately, the UK AISI reported unauthorized behavior in 10 of 122 test runs across several models, including models from Anthropic and OpenAI. The settings were unusually permissive, with disabled safeguards and open internet access. The report did not establish a normal-service incident rate or resulting real-world harm. The investment point is that capability, authorized deployment and billable work are different layers. Supervision and safety controls belong in the economics, not only in a benchmark score. &lt;a class="link" href="https://www.aisi.gov.uk/blog/incident-report-unsanctioned-agent-behaviour-during-cyber-testing" target="_blank" rel="noopener"
 &gt;AISI incident report&lt;/a&gt;.&lt;/p&gt;
&lt;h2 id="the-next-filing-should-answer-four-questions"&gt;The next filing should answer four questions
&lt;/h2&gt;&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Question&lt;/th&gt;
 &lt;th&gt;Evidence that strengthens the investment case&lt;/th&gt;
 &lt;th&gt;Evidence that weakens it&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Does new capacity fill quickly?&lt;/td&gt;
 &lt;td&gt;Accepted capacity becomes recognized revenue and cash&lt;/td&gt;
 &lt;td&gt;Lease payments begin while power, integration or customer usage lags&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Does operating leverage survive the next step?&lt;/td&gt;
 &lt;td&gt;Revenue growth exceeds total compute, channel and research-cost growth&lt;/td&gt;
 &lt;td&gt;Margin recovers only before each new capacity delivery&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Who retains the exit option?&lt;/td&gt;
 &lt;td&gt;Clear minimums, acceptance conditions and residual-value allocation&lt;/td&gt;
 &lt;td&gt;Maximum contract amounts are marketed as unconditional backlog&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;What reaches one common share?&lt;/td&gt;
 &lt;td&gt;Fully diluted ownership and cash improve after financing and compensation&lt;/td&gt;
 &lt;td&gt;Higher claims and cheaper conversion prices absorb the business growth&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;The missing document is a &lt;strong&gt;quarter-by-quarter, next-twelve-month cash-payment schedule&lt;/strong&gt;, matched to capacity delivery and customer monetization. The initial S-1 leaves aggregate commitment and near-term payment amounts blank in important places. Adding every disclosed contract number to produce a headline total cannot fill that gap.&lt;/p&gt;
&lt;p&gt;The conclusion is constructive on paid demand and cautious on transferring that evidence directly into shareholder returns. The next distinction between AI investments may be less about whose demand chart rises fastest and more about whose contract lets them wait, cancel, reprice or pass on a delay.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;If you could request one disclosure before investing, which would it be: the next twelve months of cash payments, capacity utilization after delivery, or fully diluted shares? Choose one and explain which uncertainty it resolves.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Research and information only; not investment advice. AI tools assisted research and editing. OpenAI, the provider of the editing model, competes with Anthropic. We distinguish company disclosures from independent evidence and do not infer undisclosed customer identities or recommend trades.&lt;/em&gt;&lt;/p&gt;</description></item><item><title>The AI Tenbagger Test: 15 Stocks and the Price of Being Wrong</title><link>https://koreainvestinsights.com/post/ai-tenbagger-test-15-candidates-2026-10-27/</link><pubDate>Tue, 27 Oct 2026 09:55:00 +0900</pubDate><guid>https://koreainvestinsights.com/post/ai-tenbagger-test-15-candidates-2026-10-27/</guid><description>&lt;img src="https://koreainvestinsights.com/" alt="Featured image of post The AI Tenbagger Test: 15 Stocks and the Price of Being Wrong" /&gt;&lt;p&gt;Arteris is a roughly $956 million company. Under the assumptions below, a tenfold share-price return requires about $1.72 billion of annual revenue, a 30% net margin and no more than 35% cumulative share dilution. Its 2026 revenue guidance midpoint is $96.5 million. That is an eighteenfold revenue hurdle before the story becomes a tenbagger.&lt;/p&gt;
&lt;p&gt;This is the tension behind our list of ten US-listed and five Korean companies. AI can create a large market. A shareholder still needs the right starting price, a profitable share of that market and enough financing to reach it.&lt;/p&gt;
&lt;p&gt;Our first research priorities are Arteris (AIP), PDF Solutions (PDFS) and, conditionally, Innodata (INOD) in the US-listed group; Nextin, ASICLAND and KOSES in Korea. The ranking tells us where to spend diligence time. It is not a portfolio, a set of price targets or a prediction that fifteen stocks will rise tenfold.&lt;/p&gt;
&lt;p&gt;The calculations retain the original research&amp;rsquo;s October 26, 2026 market-cap snapshot. US aggregators update at different times; AOSL&amp;rsquo;s original input was intraday. Korean figures are rounded. Financial periods are stated separately. Required future revenue is our calculation, not company guidance.&lt;/p&gt;
&lt;h2 id="four-tests-decide-the-order"&gt;Four tests decide the order
&lt;/h2&gt;&lt;p&gt;The first screen favored businesses that can lower customer costs: chip interconnect IP, manufacturing data, AI evaluation, security, inspection and data-center connectivity. The second pass asks four questions.&lt;/p&gt;
&lt;ol&gt;
&lt;li&gt;&lt;strong&gt;Market capacity:&lt;/strong&gt; Can the addressable market support the revenue needed for a tenfold return?&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Shareholder economics:&lt;/strong&gt; Can the company retain profit after competition, taxes, interest, recurring compensation and minority interests?&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Funding:&lt;/strong&gt; Can it reach that scale without damaging dilution or refinancing trouble?&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Starting price:&lt;/strong&gt; How much of the successful outcome is already reflected in its market value?&lt;/li&gt;
&lt;/ol&gt;
&lt;p&gt;These tests move CEVA and Ambarella above SentinelOne and GitLab. They also move ASICLAND above KOSES, subject to a hard cash check. They leave AIP, PDFS and Nextin near the front, but none passes every test yet.&lt;/p&gt;
&lt;h2 id="the-arithmetic-starts-with-one-share"&gt;The arithmetic starts with one share
&lt;/h2&gt;&lt;p&gt;For a tenfold share-price return:&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Required revenue = starting market cap × 10 × future/current share count ÷ terminal P/E ÷ net margin.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;A $1 billion company with 25% more shares needs a $12.5 billion future market cap. At 25 times earnings and a 25% net margin, that means $500 million of annual net income and $2 billion of revenue. Dividends, currency moves and investor taxes are outside this calculation.&lt;/p&gt;
&lt;p&gt;The central horizon is 2036. US growth hurdles use ten years from the stated 2026 revenue base. Korean hurdles use eleven years from FY2025 revenue. A quarter multiplied by four is a comparison base, not an annual forecast. Terminal margins and multiples are scenario assumptions; several candidates are currently loss-making.&lt;/p&gt;
&lt;h2 id="the-ten-us-listed-candidates"&gt;The ten US-listed candidates
&lt;/h2&gt;&lt;p&gt;All amounts in this table are &lt;strong&gt;US$ billions&lt;/strong&gt;. Assumptions are cumulative share-count growth / net margin / terminal P/E. Calculations use the rounded market caps shown; small differences from the source research reflect rounding.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Research rank&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Market cap&lt;/th&gt;
 &lt;th&gt;Assumptions&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Required net income&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Required revenue&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;1. Arteris (AIP)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;0.956&lt;/td&gt;
 &lt;td&gt;35% / 30% / 25×&lt;/td&gt;
 &lt;td style="text-align: right"&gt;0.516&lt;/td&gt;
 &lt;td style="text-align: right"&gt;1.721&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;2. PDF Solutions (PDFS)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;1.720&lt;/td&gt;
 &lt;td&gt;20% / 25% / 25×&lt;/td&gt;
 &lt;td style="text-align: right"&gt;0.826&lt;/td&gt;
 &lt;td style="text-align: right"&gt;3.302&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;3. Innodata (INOD), conditional&lt;/td&gt;
 &lt;td style="text-align: right"&gt;2.060&lt;/td&gt;
 &lt;td&gt;20% / 18% / 25×&lt;/td&gt;
 &lt;td style="text-align: right"&gt;0.989&lt;/td&gt;
 &lt;td style="text-align: right"&gt;5.493&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;4. CEVA&lt;/td&gt;
 &lt;td style="text-align: right"&gt;0.828&lt;/td&gt;
 &lt;td&gt;25% / 30% / 25×&lt;/td&gt;
 &lt;td style="text-align: right"&gt;0.414&lt;/td&gt;
 &lt;td style="text-align: right"&gt;1.380&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;5. Ambarella (AMBA)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;2.950&lt;/td&gt;
 &lt;td&gt;25% / 20% / 25×&lt;/td&gt;
 &lt;td style="text-align: right"&gt;1.475&lt;/td&gt;
 &lt;td style="text-align: right"&gt;7.375&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;6. ACM Research (ACMR)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;3.550&lt;/td&gt;
 &lt;td&gt;20% / 14% / 20×&lt;/td&gt;
 &lt;td style="text-align: right"&gt;2.130&lt;/td&gt;
 &lt;td style="text-align: right"&gt;15.214&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;7. Alpha and Omega (AOSL)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;0.676&lt;/td&gt;
 &lt;td&gt;30% / 10% / 20×&lt;/td&gt;
 &lt;td style="text-align: right"&gt;0.439&lt;/td&gt;
 &lt;td style="text-align: right"&gt;4.394&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;8. SentinelOne (S)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;8.880&lt;/td&gt;
 &lt;td&gt;35% / 25% / 25×&lt;/td&gt;
 &lt;td style="text-align: right"&gt;4.795&lt;/td&gt;
 &lt;td style="text-align: right"&gt;19.181&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;9. GitLab (GTLB)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;9.440&lt;/td&gt;
 &lt;td&gt;25% / 25% / 25×&lt;/td&gt;
 &lt;td style="text-align: right"&gt;4.720&lt;/td&gt;
 &lt;td style="text-align: right"&gt;18.880&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;10. Camtek (CAMT)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;5.750&lt;/td&gt;
 &lt;td&gt;15% / 25% / 25×&lt;/td&gt;
 &lt;td style="text-align: right"&gt;2.645&lt;/td&gt;
 &lt;td style="text-align: right"&gt;10.580&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Market-cap sources include &lt;a class="link" href="https://stockanalysis.com/stocks/aip/market-cap/" target="_blank" rel="noopener"
 &gt;AIP&lt;/a&gt;, &lt;a class="link" href="https://stockanalysis.com/stocks/pdfs/market-cap/" target="_blank" rel="noopener"
 &gt;PDFS&lt;/a&gt;, &lt;a class="link" href="https://stockanalysis.com/stocks/inod/market-cap/" target="_blank" rel="noopener"
 &gt;INOD&lt;/a&gt; and the same provider&amp;rsquo;s company pages. These are secondary estimates, not perfectly synchronized exchange measurements. ACMR and Camtek are US-listed; their operating exposure is not US-only.&lt;/p&gt;
&lt;h3 id="aip-a-royalty-opportunity-with-a-demanding-margin-assumption"&gt;AIP: a royalty opportunity with a demanding margin assumption
&lt;/h3&gt;&lt;p&gt;More complex chips need better internal data movement. Arteris can earn from design adoption and royalties when customers&amp;rsquo; chips enter production. Q2 revenue rose 46% to $24.1 million. Trailing-twelve-month variable royalties rose 65%; ACV plus royalties rose 44%. ACV plus royalties is a company-defined measure, not GAAP revenue or conventional ARR. &lt;a class="link" href="https://ir.arteris.com/news-releases/news-release-details/arteris-announces-financial-results-second-quarter-and-0" target="_blank" rel="noopener"
 &gt;Arteris results&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;Q2 GAAP operating loss was $13.9 million. First-half operating cash less equipment purchases was about $1.26 million, while net ATM share-sale proceeds were $72.55 million. Cash raised from shareholders is not cash earned from customers.&lt;/p&gt;
&lt;p&gt;The $1.721 billion revenue hurdle is about 17.8 times the guidance midpoint, or 33.4% annual growth for ten years. A sustained 30% net margin requires a major change from today&amp;rsquo;s losses. Royalty stagnation, customer insourcing, weaker pricing and repeated issuance would all weaken the case.&lt;/p&gt;
&lt;h3 id="pdfs-recurring-data-revenue-must-repay-equipment-investment"&gt;PDFS: recurring data revenue must repay equipment investment
&lt;/h3&gt;&lt;p&gt;PDF Solutions connects manufacturing data to yield improvement. Q2 revenue was $61.529 million, up 19%; backlog was $270.7 million and GAAP operating margin was 8%. &lt;a class="link" href="https://www.pdf.com/resources/pdf-solutions-reports-second-quarter-2026-financial-results/" target="_blank" rel="noopener"
 &gt;Company results&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;It also installs equipment at customer sites. First-half operating cash of $18.075 million, less $24.528 million of equipment purchases and $0.054 million of prepayments, leaves approximately &lt;strong&gt;negative $6.507 million&lt;/strong&gt;. This is our simple cash calculation, not a company-defined FCF measure. &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1120914/000143774926026294/pdfs20260630_10q.htm" target="_blank" rel="noopener"
 &gt;PDFS 10-Q&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;Required revenue is 13.4 times annualized Q2 sales, a 29.6% ten-year growth hurdle. The next evidence should connect recurring revenue, installation cost and payback time. A larger backlog without better cash recovery would lower its priority.&lt;/p&gt;
&lt;h3 id="inod-concentration-is-established-the-customer-growth-story-is-not"&gt;INOD: concentration is established; the customer-growth story is not
&lt;/h3&gt;&lt;p&gt;Innodata reported Q2 revenue of $92.142 million, about 58% growth, and roughly $14.4 million of GAAP net income. Two customers represented approximately 37% and 34% of revenue, or 71% combined. The opportunity is to move from labor-intensive projects toward repeatable data and evaluation programs. The disclosed mix does not yet prove that transition. &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/903651/000110465926092021/inod-20260630x10q.htm" target="_blank" rel="noopener"
 &gt;Innodata 10-Q&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;Multiplying those rounded percentages by total revenue gives about $34.1 million and $31.3 million. Applying the prior-year largest-customer percentage gives about $33.9 million. These are estimates from rounded ratios, not reported customer-level dollar amounts.&lt;/p&gt;
&lt;p&gt;That does &lt;strong&gt;not&lt;/strong&gt; establish that the same customer was flat or that one new program caused all growth. Comparing this year&amp;rsquo;s revenue outside two customers with last year&amp;rsquo;s revenue outside one customer also cannot establish same-customer growth. Customer identity, comparable prior-period contributions and program economics remain missing.&lt;/p&gt;
&lt;p&gt;INOD keeps a conditional third place. Its required $5.493 billion revenue is roughly fifteen times annualized Q2 sales. The next call needs evidence of growth beyond the largest customers, recurring evaluation revenue and cash after pass-through obligations. The call is scheduled for &lt;strong&gt;November 5 at 5 p.m. US Eastern, November 6 at 7 a.m. Korea time&lt;/strong&gt;. US daylight saving time has ended by then. &lt;a class="link" href="https://investor.innodata.com/events-and-presentations/event-details/2026/Innodata-to-Report-Third-Quarter-2026-Results/default.aspx" target="_blank" rel="noopener"
 &gt;Company event&lt;/a&gt;.&lt;/p&gt;
&lt;h3 id="the-other-seven-face-different-obstacles"&gt;The other seven face different obstacles
&lt;/h3&gt;&lt;p&gt;&lt;strong&gt;CEVA:&lt;/strong&gt; Q2 device shipments rose 16%, but royalties rose only 1%; licensing revenue rose 21%. The test is royalty value per device and a richer AI-IP mix. More low-priced units do not establish the path to a 30% terminal net margin. &lt;a class="link" href="https://www.ceva-ip.com/press/ceva-inc-announces-second-quarter-2026-financial-results/" target="_blank" rel="noopener"
 &gt;CEVA results&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Ambarella:&lt;/strong&gt; Q2 FY2027 revenue rose 13.2% to $108.1 million, but it still reported a GAAP loss. A $9 million one-time reduction in R&amp;amp;D expense following a project termination helped the result. Automotive and robotics design wins need to become recurring production revenue and ordinary profit. &lt;a class="link" href="https://investor.ambarella.com/news-releases/news-release-details/ambarella-inc-announces-second-quarter-fiscal-year-2027" target="_blank" rel="noopener"
 &gt;Ambarella results&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;ACMR:&lt;/strong&gt; Q2 revenue rose 36% to $292.9 million. Its required $15.2 billion revenue greatly exceeds the company&amp;rsquo;s $4 billion long-term target. China exposure, export controls, investment gains and minority interests complicate the link from consolidated results to the US parent shareholder. &lt;a class="link" href="https://ir.acmr.com/news-releases/news-release-details/acm-research-reports-second-quarter-2026-results" target="_blank" rel="noopener"
 &gt;ACMR results&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;AOSL:&lt;/strong&gt; FY2026 revenue fell 2.5% to $678.9 million; GAAP net loss was $42.3 million and operating cash flow was negative. Its lower revenue-growth hurdle still assumes recovery to a 10% net margin. Server-power revenue and repeat orders need to become visible. &lt;a class="link" href="https://investor.aosmd.com/press-releases/press-release-details/2026/Alpha-and-Omega-Semiconductor-Reports-Financial-Results-for-Fiscal-Fourth-Quarter-and-Fiscal-Year-Ended-June-30-2026/default.aspx" target="_blank" rel="noopener"
 &gt;AOSL results&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;SentinelOne and GitLab:&lt;/strong&gt; Both grew revenue about 21% in their latest reported quarters. Stock-based compensation was approximately 31.5% of revenue at SentinelOne and 26.2% at GitLab. Our cumulative dilution assumptions therefore require restraint over time. Their roughly $19 billion revenue hurdles explain the lower rank: they need to become much larger platforms while competition and bundled AI features pressure monetization. &lt;a class="link" href="https://investors.sentinelone.com/press-releases/news-details/2026/SentinelOne-Announces-Second-Quarter-Fiscal-Year-2027-Financial-Results/default.aspx" target="_blank" rel="noopener"
 &gt;SentinelOne results&lt;/a&gt;, &lt;a class="link" href="https://ir.gitlab.com/news/news-details/2026/GitLab-Reports-Second-Quarter-Fiscal-Year-2027-Financial-Results/default.aspx" target="_blank" rel="noopener"
 &gt;GitLab results&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Camtek:&lt;/strong&gt; Q2 revenue was $133.2 million, with year-to-date orders above $600 million for delivery across 2026–27. Required revenue of $10.58 billion is almost twenty times annualized Q2 sales. New applications must add to the advanced-packaging cycle. Year-to-date orders are not current backlog. &lt;a class="link" href="https://www.camtek.com/news-and-events/camtek-announces-results-for-the-second-quarter-of-2026/" target="_blank" rel="noopener"
 &gt;Camtek results&lt;/a&gt;.&lt;/p&gt;
&lt;h2 id="korea-asicland-moves-up-with-cash-as-the-first-test"&gt;Korea: ASICLAND moves up, with cash as the first test
&lt;/h2&gt;&lt;p&gt;Amounts below are &lt;strong&gt;KRW billions&lt;/strong&gt;. One KRW billion equals ten 억원. The same formula applies; net income means profit attributable to controlling shareholders.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Research rank&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Market cap&lt;/th&gt;
 &lt;th&gt;Assumptions&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Required net income&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Required revenue&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;1. Nextin (348210)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;272.4&lt;/td&gt;
 &lt;td&gt;15% / 25% / 25×&lt;/td&gt;
 &lt;td style="text-align: right"&gt;125.3&lt;/td&gt;
 &lt;td style="text-align: right"&gt;501.2&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;2. ASICLAND (445090)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;190.1&lt;/td&gt;
 &lt;td&gt;30% / 10% / 25×&lt;/td&gt;
 &lt;td style="text-align: right"&gt;98.9&lt;/td&gt;
 &lt;td style="text-align: right"&gt;988.5&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;3. KOSES (089890)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;542.3&lt;/td&gt;
 &lt;td&gt;15% / 18% / 25×&lt;/td&gt;
 &lt;td style="text-align: right"&gt;249.5&lt;/td&gt;
 &lt;td style="text-align: right"&gt;1,385.9&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;4. KINX (093320)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;611.0&lt;/td&gt;
 &lt;td&gt;15% / 15% / 25×&lt;/td&gt;
 &lt;td style="text-align: right"&gt;281.1&lt;/td&gt;
 &lt;td style="text-align: right"&gt;1,873.7&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;5. Gigavis (420770)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;1,228.3&lt;/td&gt;
 &lt;td&gt;10% / 30% / 25×&lt;/td&gt;
 &lt;td style="text-align: right"&gt;540.5&lt;/td&gt;
 &lt;td style="text-align: right"&gt;1,801.5&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;The October 26 snapshot uses KRX-linked figures in the original research and Naver aggregation as a cross-check. It is not an intraday October 27 valuation. &lt;a class="link" href="https://kind.krx.co.kr/common/stockprices.do?isurCd=34821&amp;amp;method=searchStockPricesMain" target="_blank" rel="noopener"
 &gt;KRX Nextin price page&lt;/a&gt;, &lt;a class="link" href="https://m.stock.naver.com/api/stock/348210/integration" target="_blank" rel="noopener"
 &gt;Naver Nextin data&lt;/a&gt;.&lt;/p&gt;
&lt;h3 id="nextin-the-product-roadmap-needs-paid-repeat-orders"&gt;Nextin: the product roadmap needs paid repeat orders
&lt;/h3&gt;&lt;p&gt;Nextin&amp;rsquo;s FY2025 revenue was about KRW66.9 billion; required revenue of KRW501.2 billion represents approximately 20.1% annual growth over eleven years. This lower hurdle still requires a lasting profit turnaround.&lt;/p&gt;
&lt;p&gt;The company announced a KRW25.44 billion SK hynix equipment contract on October 6, running through November 25, 2027. An October 22 update reported progress in Chinese certification and customer registration. Registration is not a repeat production order. &lt;a class="link" href="https://ir.nextinsol.com/boards/news/r894oy799t5jeqsbb9egkli1" target="_blank" rel="noopener"
 &gt;Contract announcement&lt;/a&gt;, &lt;a class="link" href="https://ir.nextinsol.com/boards/news/f3c3skvslbcxo8bzcupefe83" target="_blank" rel="noopener"
 &gt;October 22 update&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;On October 16, Nextin described AEGIS-IV lens-supply delays, extended RESQ customer testing and continuing KROKY/ASPER evaluations. Q1 operating cash was about negative KRW4.5 billion, with inventory around KRW68.1 billion at March 31. These old-period figures cannot establish today&amp;rsquo;s liquidity. Paid repeat orders, acceptance, collections and inventory turnover decide whether it stays first. &lt;a class="link" href="https://ir.nextinsol.com/boards/news/matm2q95gz37yp6e4jowwlr1" target="_blank" rel="noopener"
 &gt;Product update&lt;/a&gt;, &lt;a class="link" href="https://kind.krx.co.kr/external/2026/05/15/002049/20260515004541/11013.htm" target="_blank" rel="noopener"
 &gt;Q1 filing&lt;/a&gt;.&lt;/p&gt;
&lt;h3 id="asicland-a-low-sales-multiple-brings-a-financing-question"&gt;ASICLAND: a low sales multiple brings a financing question
&lt;/h3&gt;&lt;p&gt;First-half revenue of KRW113.7 billion exceeded FY2025 revenue of about KRW72.85 billion. Operating profit was KRW5.0 billion; production revenue contributed 18.6% of sales. The market-cap snapshot is about 0.84 times first-half revenue multiplied by two. That annualization is not a forecast, and a 4.4% operating margin does not warrant an IP-licensing margin. &lt;a class="link" href="https://www.asicland.com/news/?bmode=view&amp;amp;idx=173096114" target="_blank" rel="noopener"
 &gt;Company release&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;The March 2026 convertible bond has KRW45 billion face value and an initial conversion price of KRW31,517. The Q1 filing says the contractual minimum reset price is restricted to at least &lt;strong&gt;80%&lt;/strong&gt; of the relevant initial conversion price, despite a generic 70% provision in the same table. Holders have a put starting March 23, 2029. On the unadjusted initial price, 80% is about KRW25,214. This is not a claim about the current adjusted conversion price. &lt;a class="link" href="https://kind.krx.co.kr/external/2026/05/15/000655/20260515001375/11013.htm" target="_blank" rel="noopener"
 &gt;CB terms&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;That resolves one source gap, but not the cash question. FY2025 operating cash was negative KRW6.3 billion and Q1 2026 cash negative KRW8.2 billion. FY2025 investment included property purchases; it should not be projected mechanically as recurring burn. We would remove ASICLAND from the priority group if cash and committed funding cannot cover twelve months of modeled working-capital needs. Latest cash, contract-asset recovery and subsequent CB adjustments still need checking.&lt;/p&gt;
&lt;h3 id="koses-factory-expansion-must-become-repeat-business"&gt;KOSES: factory expansion must become repeat business
&lt;/h3&gt;&lt;p&gt;KOSES had FY2025 revenue of about KRW82.4 billion and net income of KRW13.2 billion. Bloom Energy is named as a customer in its Q1 filing. Market cap is about 6.6 times FY2025 sales, and required revenue is nearly KRW1.39 trillion. A single factory build is insufficient. &lt;a class="link" href="https://kind.krx.co.kr/external/2026/05/15/002905/20260515006626/11013.htm" target="_blank" rel="noopener"
 &gt;KOSES Q1 filing&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;Repeat lines, services, other customers and collections matter. A July contract has a correction warning on KIND; without the latest amendment, its original amount or delivery date is not treated as current. Using FY2025 net income, KRW542.3 billion divided by KRW13.204 billion gives a P/E of about 41.1 times. A multiple based on a later twelve-month earnings period is a different comparison. &lt;a class="link" href="https://kind.krx.co.kr/common/disclsviewer.do?acptno=20260713000131&amp;amp;docno=&amp;amp;method=search&amp;amp;viewerhost=" target="_blank" rel="noopener"
 &gt;Filing and correction notice&lt;/a&gt;.&lt;/p&gt;
&lt;h3 id="kinx-contracted-megawatts-must-become-cash-after-leases"&gt;KINX: contracted megawatts must become cash after leases
&lt;/h3&gt;&lt;p&gt;KINX&amp;rsquo;s September IR described 10MW of allocated IT capacity at Gwacheon as contracted, with roughly 50% occupancy in Q2. The 90%-plus occupancy goal for first-half 2028 remains a target. The separate 40MW joint-investment project is a development/build/operate role, not 40MW of wholly owned KINX IT capacity. &lt;a class="link" href="https://company.kinx.net/wp-content/uploads/2026/09/KINX_IR_Book_Kor_260916.pdf" target="_blank" rel="noopener"
 &gt;KINX IR&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;FY2025 operating cash less equipment and intangible purchases was about KRW22.5 billion. Subtracting lease repayments leaves about KRW9.4 billion before minority ownership. Required revenue of KRW1.87 trillion needs multiple projects and better per-share cash generation. &lt;a class="link" href="https://kind.krx.co.kr/external/2026/03/18/001362/20260318006308/11011.htm" target="_blank" rel="noopener"
 &gt;Annual filing&lt;/a&gt;.&lt;/p&gt;
&lt;h3 id="gigavis-strong-orders-face-the-highest-korean-growth-hurdle"&gt;Gigavis: strong orders face the highest Korean growth hurdle
&lt;/h3&gt;&lt;p&gt;Gigavis&amp;rsquo;s inspection and repair tools can improve substrate yield. FY2025 revenue was about KRW52.4 billion and net income KRW15.5 billion. A reported KRW46.937 billion equipment contract supports the business case, but its latest original DART text was not independently reopened in this review; we do not identify the customer. &lt;a class="link" href="https://kind.krx.co.kr/external/2026/05/15/001382/20260515003007/11013.htm" target="_blank" rel="noopener"
 &gt;Q1 filing&lt;/a&gt;, &lt;a class="link" href="https://www.datatooza.com/article/20261023135542202752ef3ed011_80" target="_blank" rel="noopener"
 &gt;contract report&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;Required revenue is about KRW1.80 trillion, over 34 times FY2025 sales: 37.9% annual growth for eleven years. Wafer/panel-level packaging and glass substrates need actual external revenue. Existing ABF demand alone cannot establish this path.&lt;/p&gt;
&lt;h2 id="history-makes-the-hurdle-harder-to-dismiss"&gt;History makes the hurdle harder to dismiss
&lt;/h2&gt;&lt;p&gt;Credit Suisse&amp;rsquo;s 2016 &lt;em&gt;Base Rate Book&lt;/em&gt; reports revenue-growth distributions for 1950–2015. Exhibit 4 groups companies by starting revenue in inflation-adjusted 2015 dollars. A decade above 30% growth was unusual even in the smallest revenue group.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Starting annual revenue, 2015 dollars&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Ten-year revenue CAGR ≥20%&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Ten-year revenue CAGR ≥30%&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Below $325 million&lt;/td&gt;
 &lt;td style="text-align: right"&gt;18.1%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;7.7%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;$700 million–$1.25 billion&lt;/td&gt;
 &lt;td style="text-align: right"&gt;3.5%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;0.4%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Source: &lt;a class="link" href="https://josephnoko.com/wp-content/uploads/2024/04/The-Base-Rate-Book-Michael-Mauboussin.pdf" target="_blank" rel="noopener"
 &gt;Mauboussin, Callahan and Majd, The Base Rate Book, Exhibit 4&lt;/a&gt;. These are historical revenue-growth frequencies, not probabilities of a tenfold stock return.&lt;/p&gt;
&lt;p&gt;Applying the buckets to today&amp;rsquo;s selected AI names requires judgment about inflation, starting revenue, industry mix and horizon. Historical cohort frequencies cannot simply become individual stock probabilities or an expected number of future tenbaggers. Revenue growth is only one of the required conditions. Single-digit stock-specific probabilities would imply more precision than the evidence supports.&lt;/p&gt;
&lt;p&gt;The useful lesson survives: a decade of 30% growth needs a much stronger case than one good quarter.&lt;/p&gt;
&lt;h2 id="three-disappointments-can-multiply-the-burden"&gt;Three disappointments can multiply the burden
&lt;/h2&gt;&lt;p&gt;This sensitivity uses one illustrative starting model, not a particular stock&amp;rsquo;s full forecast.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Change&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Extra revenue needed for the same 10× share price&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Terminal P/E falls from 25× to 15×&lt;/td&gt;
 &lt;td style="text-align: right"&gt;66.7%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Future shares rise from 1.25× to 1.75× today&amp;rsquo;s count&lt;/td&gt;
 &lt;td style="text-align: right"&gt;40.0%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Net margin falls from 25% to 15%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;66.7%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;All three occur together&lt;/td&gt;
 &lt;td style="text-align: right"&gt;Revenue hurdle becomes 3.89× as large&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;For AIP, doubling shares instead of adding 35% raises required revenue from about $1.72 billion to $2.55 billion. For PDFS and INOD, doubling shares instead of adding 20% raises the hurdles to about $5.50 billion and $9.16 billion. Cumulative dilution of 20–35% over a decade is not automatically conservative.&lt;/p&gt;
&lt;h2 id="why-popular-ai-names-are-outside-this-list"&gt;Why popular AI names are outside this list
&lt;/h2&gt;&lt;p&gt;This scale check uses the supplied October 26 market-cap estimates. All amounts are &lt;strong&gt;US$ billions&lt;/strong&gt;. It is not a current fair-value judgment. NBIS&amp;rsquo;s share base is stale and omits subsequent issuance, so its figure is only a lower-bound illustration.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Company&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Market cap input&lt;/th&gt;
 &lt;th&gt;Dilution / margin / P/E&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Required annual revenue&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Astera Labs (ALAB)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;45.02&lt;/td&gt;
 &lt;td&gt;20% / 30% / 25×&lt;/td&gt;
 &lt;td style="text-align: right"&gt;72.0&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Credo (CRDO)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;28.41&lt;/td&gt;
 &lt;td&gt;20% / 30% / 25×&lt;/td&gt;
 &lt;td style="text-align: right"&gt;45.5&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Coherent (COHR)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;43.38&lt;/td&gt;
 &lt;td&gt;10% / 15% / 20×&lt;/td&gt;
 &lt;td style="text-align: right"&gt;159.1&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Lumentum (LITE)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;73.38&lt;/td&gt;
 &lt;td&gt;10% / 20% / 20×&lt;/td&gt;
 &lt;td style="text-align: right"&gt;201.8&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;CoreWeave (CRWV)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;30.41&lt;/td&gt;
 &lt;td&gt;30% / 10% / 20×&lt;/td&gt;
 &lt;td style="text-align: right"&gt;197.7&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Nebius (NBIS)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;At least 38.8&lt;/td&gt;
 &lt;td&gt;30% / 10% / 20×&lt;/td&gt;
 &lt;td style="text-align: right"&gt;At least 252.2&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;COHR and LITE use an assumed 10% cumulative dilution in this table. Different margins, share counts and multiples change these answers. The scenarios do not prove overvaluation or rule out strong returns; they explain the high bar in this tenfold-return investigation.&lt;/p&gt;
&lt;p&gt;Funding adds another obstacle for neoclouds. CoreWeave&amp;rsquo;s first-half operating cash of $3.663 billion, less $14.117 billion of equipment and software purchases, leaves a simple cash deficit of $10.454 billion. Demand and financing needs can grow together. &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1769628/000176962826000366/crwv-20260630.htm" target="_blank" rel="noopener"
 &gt;CoreWeave 10-Q&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;VINATech, GAONCHIPS and OPENEDGES remain outside the priority list. Their gaps concern cash and contract economics, starting valuation relative to ASICLAND, and repeat royalties respectively. AEHR and OPENEDGES remain watchlist comparisons. Exclusion is not a sell recommendation or a completed relative-valuation judgment.&lt;/p&gt;
&lt;h2 id="evidence-that-would-change-our-minds"&gt;Evidence that would change our minds
&lt;/h2&gt;&lt;p&gt;These are analyst-set diligence thresholds, not company promises or trading instructions. A missing disclosure leaves a question open; it does not prove business failure.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Name&lt;/th&gt;
 &lt;th&gt;Evidence that strengthens the case&lt;/th&gt;
 &lt;th&gt;Evidence that lowers priority&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;AIP&lt;/td&gt;
 &lt;td&gt;Higher royalty value per design and cash generation per share&lt;/td&gt;
 &lt;td&gt;Royalty growth below 30% for two comparable periods, or annual ATM issuance above 5% of opening shares&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;PDFS&lt;/td&gt;
 &lt;td&gt;Recurring economics linked to equipment payback&lt;/td&gt;
 &lt;td&gt;Investment and financing needs outrun recurring cash&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;INOD&lt;/td&gt;
 &lt;td&gt;Comparable growth beyond the largest customers; cash after pass-through needs&lt;/td&gt;
 &lt;td&gt;More concentration without broader recurring programs&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;CEVA / AMBA&lt;/td&gt;
 &lt;td&gt;Higher-value AI royalties / recurring automotive and robotics production&lt;/td&gt;
 &lt;td&gt;Units or design wins without revenue quality and profit&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Nextin&lt;/td&gt;
 &lt;td&gt;Paid repeat orders, accepted deliveries and collections&lt;/td&gt;
 &lt;td&gt;Evaluation delays persist at the Q1 2027 research checkpoint&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;ASICLAND&lt;/td&gt;
 &lt;td&gt;Production growth, operating cash and twelve-month funding&lt;/td&gt;
 &lt;td&gt;Working-capital needs outrun cash and committed facilities&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;KOSES&lt;/td&gt;
 &lt;td&gt;Other-customer sales above 30%, repeat orders and collections&lt;/td&gt;
 &lt;td&gt;Post-expansion order gaps or uncollected balances&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;KINX / Gigavis&lt;/td&gt;
 &lt;td&gt;Multiple billed projects / paid adoption in new applications&lt;/td&gt;
 &lt;td&gt;Announcements without cash after investment and leases&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;The next checkpoint is Innodata&amp;rsquo;s November 5 US call. November 20 is our research review date, not a company event. The Korean cash comparison remains incomplete: Nextin, KOSES and Gigavis have half-year filings, but this article has not completed a fresh reconciliation of those cash-flow statements. Older figures cannot establish current liquidity.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Which assumption would you challenge first: AIP&amp;rsquo;s eventual 30% net margin, INOD&amp;rsquo;s customer breadth, or ASICLAND&amp;rsquo;s funding runway? Pick one, and name the single piece of evidence that would change your view.&lt;/strong&gt; A filing, a disclosed metric or a comparable case is more useful than a ticker alone.&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Research and information only. Not investment advice. AI tools assisted research and editing. Customers not named in filings are not inferred. Rankings and scenario assumptions are analytical judgments, not investment approvals.&lt;/em&gt;&lt;/p&gt;</description></item><item><title>Neocloud Earnings Season: Which Call Can Prove the Rebound?</title><link>https://koreainvestinsights.com/post/neocloud-earnings-season-iren-nebius-coreweave-2026-10-27/</link><pubDate>Tue, 27 Oct 2026 08:00:00 +0900</pubDate><guid>https://koreainvestinsights.com/post/neocloud-earnings-season-iren-nebius-coreweave-2026-10-27/</guid><description>&lt;img src="https://koreainvestinsights.com/" alt="Featured image of post Neocloud Earnings Season: Which Call Can Prove the Rebound?" /&gt;&lt;p&gt;A wave of earnings calls is coming for the companies that rent AI computing, often called neoclouds. Their stocks fell hard in October. Which call can show that the fall went too far?&lt;/p&gt;
&lt;p&gt;Price alone does not answer this. From the highest close in October, IREN fell 39.1%, Nebius 38.7% and CoreWeave 39.8%. So we rank the calls by what each one can prove, not by how far the stock fell. This is an order of examination. It is not an expected-return ranking, a price target or a call to buy.&lt;/p&gt;
&lt;p&gt;Prices are US regular-session closes on October 26, 2026. All call times are Korea time. Where a number is our calculation or an outside estimate, the text says so.&lt;/p&gt;
&lt;h2 id="five-things-to-take-away"&gt;Five things to take away
&lt;/h2&gt;&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Price does not pick a favorite.&lt;/strong&gt; The three main names are within 1.2 points of each other from their October highs. If each recovered half of its drop, the gain would be 32% to 33% for all three, a gap of about 1.6 points.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;IREN has the most concrete test.&lt;/strong&gt; Its call is confirmed for November 6 at 07:00. It can show whether delivered equipment became accepted, billed and cash-generating. Its new $2.4bn GPU loan repays in 30 months, so the funding table matters as much as the revenue.&lt;sup id="fnref:1"&gt;&lt;a href="#fn:1" class="footnote-ref" role="doc-noteref"&gt;1&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Nebius is close behind.&lt;/strong&gt; Its revenue conversion is better documented. But the market already expects revenue to jump about 55% in one quarter, and the company has not confirmed a call date.&lt;sup id="fnref:2"&gt;&lt;a href="#fn:2" class="footnote-ref" role="doc-noteref"&gt;2&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;CoreWeave is a financing story.&lt;/strong&gt; Its own outlook shows interest expense far above operating profit. The stock can move most if funding worries ease, and the same logic works in reverse.&lt;sup id="fnref:3"&gt;&lt;a href="#fn:3" class="footnote-ref" role="doc-noteref"&gt;3&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;The biggest drop is not the best setup.&lt;/strong&gt; Cipher Digital (CIFR) fell 50.9%, but it carries the first $359.3m of overruns at one of its projects.&lt;sup id="fnref:4"&gt;&lt;a href="#fn:4" class="footnote-ref" role="doc-noteref"&gt;4&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id="1-order-of-examination"&gt;1. Order of examination
&lt;/h2&gt;&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Order&lt;/th&gt;
 &lt;th&gt;Name&lt;/th&gt;
 &lt;th&gt;Why it comes here&lt;/th&gt;
 &lt;th&gt;What would change the order&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;First&lt;/td&gt;
 &lt;td&gt;IREN&lt;/td&gt;
 &lt;td&gt;A confirmed near date. The call can link delivery, customer acceptance and billing to revenue and cash.&lt;/td&gt;
 &lt;td&gt;Lower if only the ARR figure improves, if acceptance slips, or if a funding gap appears.&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Close second&lt;/td&gt;
 &lt;td&gt;Nebius (NBIS)&lt;/td&gt;
 &lt;td&gt;Revenue conversion and margins are better documented. It must beat a high bar.&lt;/td&gt;
 &lt;td&gt;Could move first if it beats Q3 expectations and shows Q4 billing and long-term funding. Lower if conversion slips while spending rises.&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Conditional&lt;/td&gt;
 &lt;td&gt;CoreWeave (CRWV)&lt;/td&gt;
 &lt;td&gt;The largest swing if financing worries ease.&lt;/td&gt;
 &lt;td&gt;Needs better funding terms and better cash flow together. We found no fresh credit-market data.&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Later&lt;/td&gt;
 &lt;td&gt;Cipher Digital (CIFR)&lt;/td&gt;
 &lt;td&gt;The biggest drop, but real cost and schedule problems.&lt;/td&gt;
 &lt;td&gt;Higher once costs and delivery dates settle and rent is collected.&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;IREN and Nebius are close. The IREN rank reflects a confirmed date and a concrete chain to check. It does not mean IREN is cheaper, safer or more likely to rise.&lt;/p&gt;
&lt;h2 id="2-the-calendar"&gt;2. The calendar
&lt;/h2&gt;&lt;p&gt;Times are Korea time. US daylight saving time ends on November 1, so we add 14 hours to Eastern time after that date and 13 hours before it.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Company&lt;/th&gt;
 &lt;th&gt;Business&lt;/th&gt;
 &lt;th&gt;Call, Korea time&lt;/th&gt;
 &lt;th&gt;Status&lt;/th&gt;
 &lt;th&gt;What to check&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Galaxy (GLXY)&lt;/td&gt;
 &lt;td&gt;Digital-asset finance plus the Helios data center&lt;/td&gt;
 &lt;td&gt;Oct 27, 21:30&lt;/td&gt;
 &lt;td&gt;Confirmed&lt;/td&gt;
 &lt;td&gt;Helios rent and cash, separated from crypto results&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Cipher Digital (CIFR)&lt;/td&gt;
 &lt;td&gt;Data center leasing and development&lt;/td&gt;
 &lt;td&gt;Nov 2, 22:00&lt;/td&gt;
 &lt;td&gt;Confirmed&lt;/td&gt;
 &lt;td&gt;Black Pearl rent, Barber Lake cost and schedule&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Hut 8 (HUT)&lt;/td&gt;
 &lt;td&gt;Power and data center development&lt;/td&gt;
 &lt;td&gt;Nov 4, 22:30&lt;/td&gt;
 &lt;td&gt;Confirmed&lt;/td&gt;
 &lt;td&gt;Construction progress, loan drawdowns, extra equity&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Core Scientific (CORZ)&lt;/td&gt;
 &lt;td&gt;High-density data center leasing&lt;/td&gt;
 &lt;td&gt;Nov 5, 22:30&lt;/td&gt;
 &lt;td&gt;Confirmed&lt;/td&gt;
 &lt;td&gt;Billable megawatts, recurring rent, cash&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;IREN&lt;/td&gt;
 &lt;td&gt;GPU cloud, moving away from mining&lt;/td&gt;
 &lt;td&gt;Nov 6, 07:00&lt;/td&gt;
 &lt;td&gt;Confirmed&lt;/td&gt;
 &lt;td&gt;Horizon billing, AI revenue, funding table&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;TeraWulf (WULF)&lt;/td&gt;
 &lt;td&gt;Data center leasing and development&lt;/td&gt;
 &lt;td&gt;Nov 6, 22:00&lt;/td&gt;
 &lt;td&gt;Confirmed&lt;/td&gt;
 &lt;td&gt;CB-4 delivery, cash rent and costs&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;CoreWeave (CRWV)&lt;/td&gt;
 &lt;td&gt;GPU cloud&lt;/td&gt;
 &lt;td&gt;Not confirmed&lt;/td&gt;
 &lt;td&gt;Outside calendars disagree&lt;/td&gt;
 &lt;td&gt;Funding terms, capex, cash flow&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Nebius (NBIS)&lt;/td&gt;
 &lt;td&gt;GPU cloud&lt;/td&gt;
 &lt;td&gt;Not confirmed&lt;/td&gt;
 &lt;td&gt;No date on the company page when checked&lt;/td&gt;
 &lt;td&gt;Meeting a high revenue bar, billable capacity, long-term funding&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;The confirmed dates come from each company&amp;rsquo;s event page or release.&lt;sup id="fnref:5"&gt;&lt;a href="#fn:5" class="footnote-ref" role="doc-noteref"&gt;5&lt;/a&gt;&lt;/sup&gt; For CoreWeave and Nebius we did not use outside calendar dates, because they differ and the companies had not announced one. Applied Digital reported on October 7 and ChronoScale on October 15, so both are out of this group.&lt;sup id="fnref:6"&gt;&lt;a href="#fn:6" class="footnote-ref" role="doc-noteref"&gt;6&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;h2 id="3-how-far-they-fell"&gt;3. How far they fell
&lt;/h2&gt;&lt;div class="fig-scroll" style="overflow-x:auto;-webkit-overflow-scrolling:touch;margin:1.4em 0"&gt;&lt;img src="https://koreainvestinsights.com/post-assets/neocloud-earnings-season-iren-nebius-coreweave-2026-10-27/drawdowns.en.svg" alt="Drop in the Oct 26 close versus the highest close in October and versus the Sep 25 close. CIFR minus 50.9 percent from the October high and minus 56.5 since Sep 25. CRWV minus 39.8 and minus 37.0. IREN minus 39.1 and minus 42.4. NBIS minus 38.7 and minus 35.4. HUT minus 31.7 and minus 35.0. APLD minus 29.8 and minus 32.1. WULF minus 22.7 and minus 23.9. CORZ minus 19.0 and minus 22.1." loading="lazy" style="display:block;width:100%;min-width:720px;max-width:none;height:auto;margin:0 auto"&gt;&lt;/div&gt;
&lt;p&gt;We use the highest close from October 1 to 26 as the high, not the 52-week intraday high, so every stock is measured the same way.&lt;sup id="fnref:7"&gt;&lt;a href="#fn:7" class="footnote-ref" role="doc-noteref"&gt;7&lt;/a&gt;&lt;/sup&gt; Two points follow.&lt;/p&gt;
&lt;p&gt;First, CoreWeave fell the most of the three, but only by a little. The argument that it must bounce the most because it fell the most is weak. Second, CIFR&amp;rsquo;s fall needs its own explanation. It could be a fear discount, or it could be a real price for cost overruns and cash flow. The Cipher section below suggests there is some of both.&lt;/p&gt;
&lt;h2 id="4-iren-can-delivered-equipment-become-revenue-and-cash"&gt;4. IREN: can delivered equipment become revenue and cash?
&lt;/h2&gt;&lt;p&gt;IREN&amp;rsquo;s AI cloud revenue was $70.5m in the quarter to June, up from $33.6m in the quarter before. Total revenue was $137.2m. The company reported $1bn of operating annualized run-rate revenue (ARR) today and $4bn of contracted ARR for 2026 capacity. At the time, the 50MW Horizon 1 site was delivered, Horizon 2 was being commissioned, and Horizons 3 and 4 were targeted for the fourth quarter of 2026.&lt;sup id="fnref:8"&gt;&lt;a href="#fn:8" class="footnote-ref" role="doc-noteref"&gt;8&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;The first trap is the ARR definition.&lt;/strong&gt; IREN&amp;rsquo;s operating ARR is not the last month&amp;rsquo;s revenue times 12. The company takes the number of contracted GPUs that are ready to run, applies the hourly contract price and 8,760 hours a year, and adds minor items. It says actual recognized revenue can be materially lower. So the $1bn operating figure and the $4bn contracted figure both measure contract value for GPUs, not recognized revenue. The gap between them is capacity that still has to be delivered and billed. Neither is comparable with Nebius, which defines ARR from the last month&amp;rsquo;s actual AI revenue.&lt;sup id="fnref1:8"&gt;&lt;a href="#fn:8" class="footnote-ref" role="doc-noteref"&gt;8&lt;/a&gt;&lt;/sup&gt;&lt;sup id="fnref1:2"&gt;&lt;a href="#fn:2" class="footnote-ref" role="doc-noteref"&gt;2&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;The second trap is the loan schedule.&lt;/strong&gt; A $2.4bn GPU financing signed on August 25 carries a fixed 9% rate, matures 30 months after each drawdown and repays in installments. The parent company gives an unconditional guarantee of payment.&lt;sup id="fnref1:1"&gt;&lt;a href="#fn:1" class="footnote-ref" role="doc-noteref"&gt;1&lt;/a&gt;&lt;/sup&gt; The maturity table at June 30 shows only $177.4m of repayments in fiscal 2027, but it does not include this new loan. Reading that table alone would understate near-term repayments. Reading the loan as isolated project risk would be wrong too, given the guarantee. A separate $3.6bn financing tied to Microsoft work carries a weighted 6% rate. With customer prepayments it covers 96% of the GPU spending for that contract. It does not cover the whole data center build or later expansions.&lt;sup id="fnref2:8"&gt;&lt;a href="#fn:8" class="footnote-ref" role="doc-noteref"&gt;8&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Expectations are not low.&lt;/strong&gt; Public aggregates expect total revenue of $221m (S&amp;amp;P estimates via Moomoo) to $230m (Investing.com) for the quarter to September. Against $137.2m last quarter, that is a rise of 61% to 68%. These are aggregates, not company guidance.&lt;sup id="fnref:9"&gt;&lt;a href="#fn:9" class="footnote-ref" role="doc-noteref"&gt;9&lt;/a&gt;&lt;/sup&gt; Total revenue also includes mining and the Mirantis acquisition that closed in August, which added about 580 people and costs.&lt;sup id="fnref2:1"&gt;&lt;a href="#fn:1" class="footnote-ref" role="doc-noteref"&gt;1&lt;/a&gt;&lt;/sup&gt; A revenue beat alone does not show that AI deployment drove it.&lt;/p&gt;
&lt;p&gt;Four things would strengthen the case: higher recognized AI revenue, Horizon 2 customer acceptance and billing, Horizon 3 and 4 dates kept, and a funding table that nets cash, prepayments and committed loans against remaining capex and repayments. A new contract announcement helps but does not replace them. The case weakens if ARR rises while acceptance, revenue and cash do not, or if a funding gap shows up for equipment already contracted.&lt;/p&gt;
&lt;h2 id="5-nebius-stronger-evidence-higher-bar"&gt;5. Nebius: stronger evidence, higher bar
&lt;/h2&gt;&lt;div class="fig-scroll" style="overflow-x:auto;-webkit-overflow-scrolling:touch;margin:1.4em 0"&gt;&lt;img src="https://koreainvestinsights.com/post-assets/neocloud-earnings-season-iren-nebius-coreweave-2026-10-27/expectations.en.svg" alt="Revenue in the last reported quarter versus what public aggregates expect for the quarter to September, in million dollars. IREN total revenue: 137.2 last quarter, expected 221 to 230, a rise of 61 to 68 percent, including mining and the August Mirantis acquisition. Nebius group revenue: 582.3 last quarter, expected 905.7, a rise of 55.5 percent in one quarter." loading="lazy" style="display:block;width:100%;min-width:720px;max-width:none;height:auto;margin:0 auto"&gt;&lt;/div&gt;
&lt;p&gt;Nebius reported second-quarter group revenue of $582.3m, of which AI cloud was $574.9m. The adjusted EBITDA margin of the AI business was about 50%, and ARR was $3.0bn, defined as the last month&amp;rsquo;s actual AI revenue times 12. Adjusted EBITDA leaves out depreciation, financing costs and capital spending, so it is not cash left for shareholders.&lt;sup id="fnref2:2"&gt;&lt;a href="#fn:2" class="footnote-ref" role="doc-noteref"&gt;2&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;The bar is high. Public aggregates expect third-quarter revenue of about $906m, or 55.5% more than the second quarter.&lt;sup id="fnref1:9"&gt;&lt;a href="#fn:9" class="footnote-ref" role="doc-noteref"&gt;9&lt;/a&gt;&lt;/sup&gt; So a year-on-year multiple is not enough. The call needs to show whether capacity is actually running and billing, and what the path through the fourth quarter looks like.&lt;/p&gt;
&lt;p&gt;Nebius&amp;rsquo;s funding access is open, but the terms differ by period. It closed a $5.75bn convertible note offering in August.&lt;sup id="fnref:10"&gt;&lt;a href="#fn:10" class="footnote-ref" role="doc-noteref"&gt;10&lt;/a&gt;&lt;/sup&gt; A low coupon does not mean a low total cost, because conversion terms and dilution matter. On October 22 it announced a $3.5bn senior secured facility from Citi with a 364-day term, an initial margin of SOFR plus 275 basis points and a stated top of plus 350.&lt;sup id="fnref:11"&gt;&lt;a href="#fn:11" class="footnote-ref" role="doc-noteref"&gt;11&lt;/a&gt;&lt;/sup&gt; Securing a limit, drawing cash and refinancing for the long term are three different steps.&lt;/p&gt;
&lt;p&gt;Nebius could overtake IREN if it meets the high Q3 bar and shows Q4 billing, prepayments received, remaining spending and long-term funding in one picture. It should fall if conversion slips while spending rises and short-term funding keeps being extended.&lt;/p&gt;
&lt;h2 id="6-coreweave-financing-not-revenue-is-the-swing"&gt;6. CoreWeave: financing, not revenue, is the swing
&lt;/h2&gt;&lt;div class="fig-scroll" style="overflow-x:auto;-webkit-overflow-scrolling:touch;margin:1.4em 0"&gt;&lt;img src="https://koreainvestinsights.com/post-assets/neocloud-earnings-season-iren-nebius-coreweave-2026-10-27/crwv-gap.en.svg" alt="CoreWeave. Left, third-quarter company outlook in million dollars: adjusted operating income 200 to 260, interest expense 860 to 940, a midpoint gap of about minus 670. Right, first half of 2026: operating cash flow 3,663, purchases of equipment and software 14,117, so operating cash flow minus purchases is minus 10,454." loading="lazy" style="display:block;width:100%;min-width:720px;max-width:none;height:auto;margin:0 auto"&gt;&lt;/div&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;CoreWeave Q3 outlook&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Range&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Revenue&lt;/td&gt;
 &lt;td style="text-align: right"&gt;$3.45bn to $3.60bn&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Adjusted operating income&lt;/td&gt;
 &lt;td style="text-align: right"&gt;$200m to $260m&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Interest expense&lt;/td&gt;
 &lt;td style="text-align: right"&gt;$860m to $940m&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Capital spending&lt;/td&gt;
 &lt;td style="text-align: right"&gt;$11.5bn to $13.5bn&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;At the midpoints, adjusted operating income of $230m is $670m below interest expense of $900m. This compares two different measures to show scale. It is not a net income forecast.&lt;sup id="fnref1:3"&gt;&lt;a href="#fn:3" class="footnote-ref" role="doc-noteref"&gt;3&lt;/a&gt;&lt;/sup&gt; At June 30, about 89.6% of its debt was recourse to the company by carrying amount. In the first half, operating cash flow of $3.66bn minus purchases of equipment and software of $14.12bn left a gap of $10.45bn.&lt;sup id="fnref:12"&gt;&lt;a href="#fn:12" class="footnote-ref" role="doc-noteref"&gt;12&lt;/a&gt;&lt;/sup&gt; These are past or guided figures, not forecasts for later quarters.&lt;/p&gt;
&lt;p&gt;With this much debt, less worry about funding can change what is left for shareholders a lot, which is why CoreWeave has the largest conditional swing. The same effect works the other way if funding gets harder. A revenue beat or a bigger backlog does not settle the financing question. We would look at the rate, term, collateral and recourse of any new borrowing, the credit quality of customer contracts, the amount drawn, installed capacity and the remaining funding gap. We did not obtain fresh credit-default swap or bond trade data, so we do not call the credit worry resolved.&lt;/p&gt;
&lt;h2 id="7-cipher-and-the-other-landlords"&gt;7. Cipher and the other landlords
&lt;/h2&gt;&lt;p&gt;Cipher Digital fell 50.9% from its October high and 56.5% since September 25. A September 25 amendment to its Barber Lake lease extended the term to 20 years. It also set delivery windows from the fourth quarter of 2026 to the first quarter of 2027. Cipher bears the first $359.3m of cost above the original budget. Above that, the tenant reimburses 50% through rent over 20 years.&lt;sup id="fnref1:4"&gt;&lt;a href="#fn:4" class="footnote-ref" role="doc-noteref"&gt;4&lt;/a&gt;&lt;/sup&gt; A larger contract does not mean cost and timing risk got better. On October 14 Cipher priced $850m of 0% convertible notes.&lt;sup id="fnref:13"&gt;&lt;a href="#fn:13" class="footnote-ref" role="doc-noteref"&gt;13&lt;/a&gt;&lt;/sup&gt; That announcement shows market access but does not settle closing, net cash or dilution. We would revisit Cipher when costs and dates are stable and rent from Black Pearl is received.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Company&lt;/th&gt;
 &lt;th&gt;Evidence so far&lt;/th&gt;
 &lt;th&gt;Decisive question for the call&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Core Scientific (CORZ)&lt;/td&gt;
 &lt;td&gt;Q2 colocation revenue $136.7m; billable capacity 395MW, 437MW by mid-July&lt;/td&gt;
 &lt;td&gt;Does more billable capacity become recurring revenue and cash?&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;TeraWulf (WULF)&lt;/td&gt;
 &lt;td&gt;Q2 HPC lease revenue $31.9m; 102MW running in July; adjusted EBITDA −$18.3m; interest expense $56.4m&lt;/td&gt;
 &lt;td&gt;Do new deliveries narrow the gap between future contracts and today&amp;rsquo;s cash?&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Hut 8 (HUT)&lt;/td&gt;
 &lt;td&gt;949MW contracted as of its Aug 4 release; $26.6bn contract value; $7.5bn project financing&lt;/td&gt;
 &lt;td&gt;Do loan drawdowns and construction match billing dates?&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;These businesses are at different stages, so we do not compare them on one revenue multiple.&lt;sup id="fnref:14"&gt;&lt;a href="#fn:14" class="footnote-ref" role="doc-noteref"&gt;14&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;h2 id="8-short-interest-it-can-amplify-a-move-but-it-does-not-rank-the-stocks"&gt;8. Short interest: it can amplify a move, but it does not rank the stocks
&lt;/h2&gt;&lt;p&gt;The latest figures are for the settlement date of October 15, not October 26.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Company&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Shares short, Sep 30&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Shares short, Oct 15&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Change&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;IREN&lt;/td&gt;
 &lt;td style="text-align: right"&gt;81,653,639&lt;/td&gt;
 &lt;td style="text-align: right"&gt;94,617,935&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+15.9%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Nebius&lt;/td&gt;
 &lt;td style="text-align: right"&gt;44,695,449&lt;/td&gt;
 &lt;td style="text-align: right"&gt;47,638,283&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+6.6%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;CoreWeave&lt;/td&gt;
 &lt;td style="text-align: right"&gt;59,952,639&lt;/td&gt;
 &lt;td style="text-align: right"&gt;69,779,354&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+16.4%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;This is the change in the number of shares sold short from a secondary source. It is not the change in the share of shares outstanding.&lt;sup id="fnref:15"&gt;&lt;a href="#fn:15" class="footnote-ref" role="doc-noteref"&gt;15&lt;/a&gt;&lt;/sup&gt; All three companies have convertible notes, and some short positions may hedge them. We do not know how many, so we cannot treat the balance as shares waiting to be bought back. The same rule applies to all three. IREN should not be judged by a harsher standard than the others.&lt;/p&gt;
&lt;h2 id="9-what-to-check-by-date"&gt;9. What to check, by date
&lt;/h2&gt;&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Korea time&lt;/th&gt;
 &lt;th&gt;Company&lt;/th&gt;
 &lt;th&gt;What to look for&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Oct 27, 21:30&lt;/td&gt;
 &lt;td&gt;Galaxy&lt;/td&gt;
 &lt;td&gt;Helios recurring rent and cash, and funding for expansion&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Nov 2, 22:00&lt;/td&gt;
 &lt;td&gt;Cipher&lt;/td&gt;
 &lt;td&gt;Black Pearl rent; Barber Lake overruns, schedule and completion funding&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Nov 4, 22:30&lt;/td&gt;
 &lt;td&gt;Hut 8&lt;/td&gt;
 &lt;td&gt;Construction progress, loan drawdowns, extra equity&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Nov 5, 22:30&lt;/td&gt;
 &lt;td&gt;Core Scientific&lt;/td&gt;
 &lt;td&gt;Billable megawatts, recurring rent, cash&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;&lt;strong&gt;Nov 6, 07:00&lt;/strong&gt;&lt;/td&gt;
 &lt;td&gt;&lt;strong&gt;IREN&lt;/strong&gt;&lt;/td&gt;
 &lt;td&gt;Actual AI revenue, Horizon 2 acceptance and billing, Horizon 3 and 4 dates, a funding table that includes the new loan&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Nov 6, 22:00&lt;/td&gt;
 &lt;td&gt;TeraWulf&lt;/td&gt;
 &lt;td&gt;CB-4 delivery, cash rent and costs&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;When announced&lt;/td&gt;
 &lt;td&gt;CoreWeave, Nebius&lt;/td&gt;
 &lt;td&gt;The confirmed dates, Nebius meeting its high bar, CoreWeave funding terms&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;h2 id="conclusion-the-first-test-is-delivery-turning-into-cash"&gt;Conclusion: the first test is delivery turning into cash
&lt;/h2&gt;&lt;p&gt;If we had to examine one call first, it would be IREN&amp;rsquo;s on November 6. We would look for four things at once: recognized AI revenue, Horizon 2 acceptance and billing, kept delivery dates and a funding table that includes the new loan. If they improve together, the rebound case gets stronger.&lt;/p&gt;
&lt;p&gt;Nebius is a nearly equal alternative. If it beats a high bar and also shows fourth-quarter billing and long-term funding, it could come first. CoreWeave has the largest conditional swing, but we would not call the credit worry resolved on revenue growth alone. Nothing in this post says the credit market as a whole has improved.&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Interest note: We use AI models, including Anthropic&amp;rsquo;s Claude and OpenAI&amp;rsquo;s models, in our research. OpenAI is a CoreWeave customer, with contracts worth up to about $22.4 billion in total. We cite public figures only and make no judgment about the effect on either company.&lt;/em&gt;&lt;sup id="fnref:16"&gt;&lt;a href="#fn:16" class="footnote-ref" role="doc-noteref"&gt;16&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Disclaimer: For research and information purposes only. Not investment advice. Names cited are for analytical illustration; readers should perform their own due diligence and consult licensed advisors before any investment decision.&lt;/em&gt;&lt;/p&gt;
&lt;div class="footnotes" role="doc-endnotes"&gt;
&lt;hr&gt;
&lt;ol&gt;
&lt;li id="fn:1"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1878848/000187884826000052/iren-20260630.htm" target="_blank" rel="noopener"
 &gt;IREN annual report (10-K), fiscal 2026&lt;/a&gt;, notes and Item 9B: $2.4bn GPU financing signed 2026-08-25 (two parts of $1.2bn each), 9% fixed, 30 months from each drawdown, amortizing, unconditional parent guarantee of payment; Mirantis acquisition in August with about 580 personnel added; fiscal 2027 repayments of $177.4m in the June 30 maturity table.&amp;#160;&lt;a href="#fnref:1" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:1" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref2:1" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:2"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1513845/000110465926094568/tm2622968d1_ex99-2.htm" target="_blank" rel="noopener"
 &gt;Nebius second-quarter 2026 shareholder letter&lt;/a&gt;: group revenue $582.3m, AI cloud revenue $574.9m, AI adjusted EBITDA margin about 50%, ARR $3.0bn defined as the last month&amp;rsquo;s actual AI revenue times 12.&amp;#160;&lt;a href="#fnref:2" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:2" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref2:2" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:3"&gt;
&lt;p&gt;&lt;a class="link" href="https://s205.q4cdn.com/133937190/files/doc_financials/2026/q2/CoreWeave-Q2-26-Earnings-Outlook-Presentation.pdf" target="_blank" rel="noopener"
 &gt;CoreWeave second-quarter 2026 outlook presentation&lt;/a&gt;, p. 3: third-quarter revenue $3.45bn to $3.60bn, adjusted operating income $200m to $260m, interest expense $860m to $940m, capital spending $11.5bn to $13.5bn.&amp;#160;&lt;a href="#fnref:3" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:3" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:4"&gt;
&lt;p&gt;&lt;a class="link" href="https://investors.cipherdigital.com/news-releases/news-release-details/cipher-digital-expands-barber-lake-lease-term-20-years" target="_blank" rel="noopener"
 &gt;Cipher Digital press release&lt;/a&gt;, 2026-09-25: 20-year term, delivery windows, first $359.3m of cost above the original budget borne by Cipher, tenant reimbursement of 50% of further overruns through rent over 20 years. The company page blocked our automated check; we corroborated the terms with &lt;a class="link" href="https://datacenterdynamics.com/en/news/cipher-extends-barber-lake-data-center-lease-commitments-to-20-years" target="_blank" rel="noopener"
 &gt;Data Center Dynamics&lt;/a&gt;.&amp;#160;&lt;a href="#fnref:4" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:4" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:5"&gt;
&lt;p&gt;Event pages and releases: &lt;a class="link" href="https://investor.galaxy.com/events-presentations/events" target="_blank" rel="noopener"
 &gt;Galaxy&lt;/a&gt;, &lt;a class="link" href="https://www.globenewswire.com/news-release/2026/10/20/3384943/0/en/cipher-digital-announces-date-of-third-quarter-2026-business-update-conference-call.html" target="_blank" rel="noopener"
 &gt;Cipher&lt;/a&gt;, &lt;a class="link" href="https://www.nasdaq.com/press-release/hut-8-schedules-third-quarter-2026-earnings-release-and-conference-call-2026-10-08" target="_blank" rel="noopener"
 &gt;Hut 8&lt;/a&gt;, &lt;a class="link" href="https://investors.corescientific.com/news-events/press-releases/detail/145/core-scientific-inc-schedules-third-quarter-fiscal-year-2026-earnings-release-conference-call-and-webcast" target="_blank" rel="noopener"
 &gt;Core Scientific&lt;/a&gt;, &lt;a class="link" href="https://www.globenewswire.com/news-release/2026/10/22/3386871/0/en/iren-to-release-q1-fy27-results-on-november-5-2026.html" target="_blank" rel="noopener"
 &gt;IREN&lt;/a&gt;, &lt;a class="link" href="https://investors.terawulf.com/news-events/press-releases/detail/147/terawulf-schedules-conference-call-for-third-quarter-2026-financial-results" target="_blank" rel="noopener"
 &gt;TeraWulf&lt;/a&gt;, &lt;a class="link" href="https://investors.coreweave.com/events-and-presentations/" target="_blank" rel="noopener"
 &gt;CoreWeave&lt;/a&gt;, &lt;a class="link" href="https://nebius.com/investor-events" target="_blank" rel="noopener"
 &gt;Nebius&lt;/a&gt;. Korea time is our conversion from US Eastern time.&amp;#160;&lt;a href="#fnref:5" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:6"&gt;
&lt;p&gt;&lt;a class="link" href="https://ir.applieddigital.com/news-events/press-releases/detail/163/applied-digital-reports-fiscal-first-quarter-2027-results" target="_blank" rel="noopener"
 &gt;Applied Digital fiscal first-quarter 2027 results&lt;/a&gt;, 2026-10-07; &lt;a class="link" href="https://www.globenewswire.com/news-release/2026/10/15/3382157/0/en/chronoscale-reports-fiscal-first-quarter-2027-results.html" target="_blank" rel="noopener"
 &gt;ChronoScale fiscal first-quarter 2027 results&lt;/a&gt;, 2026-10-15.&amp;#160;&lt;a href="#fnref:6" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:7"&gt;
&lt;p&gt;Daily closes from StockAnalysis, a secondary data provider: &lt;a class="link" href="https://stockanalysis.com/stocks/iren/history/" target="_blank" rel="noopener"
 &gt;IREN&lt;/a&gt;, &lt;a class="link" href="https://stockanalysis.com/stocks/nbis/history/" target="_blank" rel="noopener"
 &gt;NBIS&lt;/a&gt;, &lt;a class="link" href="https://stockanalysis.com/stocks/crwv/history/" target="_blank" rel="noopener"
 &gt;CRWV&lt;/a&gt;, &lt;a class="link" href="https://stockanalysis.com/stocks/cifr/history/" target="_blank" rel="noopener"
 &gt;CIFR&lt;/a&gt;, &lt;a class="link" href="https://stockanalysis.com/stocks/hut/history/" target="_blank" rel="noopener"
 &gt;HUT&lt;/a&gt;, &lt;a class="link" href="https://stockanalysis.com/stocks/wulf/history/" target="_blank" rel="noopener"
 &gt;WULF&lt;/a&gt;, &lt;a class="link" href="https://stockanalysis.com/stocks/corz/history/" target="_blank" rel="noopener"
 &gt;CORZ&lt;/a&gt;, &lt;a class="link" href="https://stockanalysis.com/stocks/apld/history/" target="_blank" rel="noopener"
 &gt;APLD&lt;/a&gt;. Percentages are our calculation: (Oct 26 close / comparison close − 1).&amp;#160;&lt;a href="#fnref:7" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:8"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1878848/000187884826000051/irenreportsfy26results.htm" target="_blank" rel="noopener"
 &gt;IREN fiscal 2026 results release&lt;/a&gt;, 2026-08-27: AI cloud revenue $70.5m (prior quarter $33.6m), total revenue $137.2m, $1bn operating ARR today and $4bn contracted ARR for 2026 capacity, Horizon delivery status, ARR definition in the notes, and Microsoft-related GPU financing of $3.6bn at a weighted 6% excluding fees.&amp;#160;&lt;a href="#fnref:8" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:8" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref2:8" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:9"&gt;
&lt;p&gt;Public aggregates, not company guidance: &lt;a class="link" href="https://www.moomoo.com/stock/IREN-US/earnings" target="_blank" rel="noopener"
 &gt;IREN on Moomoo&lt;/a&gt; and &lt;a class="link" href="https://ca.investing.com/equities/iris-energy-earnings" target="_blank" rel="noopener"
 &gt;IREN on Investing.com&lt;/a&gt;; &lt;a class="link" href="https://www.moomoo.com/stock/NBIS-US/earnings" target="_blank" rel="noopener"
 &gt;Nebius on Moomoo&lt;/a&gt;. We could not reopen these pages to recheck them. A Benzinga aggregate shows about $221.4m for IREN and $904.8m for Nebius, in line. Growth rates against reported revenue are our calculation.&amp;#160;&lt;a href="#fnref:9" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:9" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:10"&gt;
&lt;p&gt;&lt;a class="link" href="https://nebius.com/newsroom/nebius-group-announces-closing-of-private-offering-of-convertible-senior-notes-with-aggregate-gross-proceeds-of-approximately-5-75-billion" target="_blank" rel="noopener"
 &gt;Nebius newsroom&lt;/a&gt;: closing of the convertible senior notes offering with gross proceeds of about $5.75 billion.&amp;#160;&lt;a href="#fnref:10" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:11"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1513845/000110465926117978/tm2628078d1_ex99-1.htm" target="_blank" rel="noopener"
 &gt;Nebius filing on the Citi facility&lt;/a&gt;, 2026-10-22: $3.5 billion senior secured facility, 364 days, initial margin SOFR plus 2.75% with a stated maximum of plus 3.50% and other adjustments.&amp;#160;&lt;a href="#fnref:11" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:12"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1769628/000176962826000366/crwv-20260630.htm" target="_blank" rel="noopener"
 &gt;CoreWeave Form 10-Q&lt;/a&gt;, six months to 2026-06-30: operating cash flow $3,663m; purchases of property and equipment and software $14,117m; recourse and non-recourse debt carrying amounts at June 30. The gap and the recourse share are our calculations.&amp;#160;&lt;a href="#fnref:12" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:13"&gt;
&lt;p&gt;&lt;a class="link" href="https://investors.cipherdigital.com/news-releases/news-release-details/cipher-digital-announces-pricing-private-offering-8500-million" target="_blank" rel="noopener"
 &gt;Cipher Digital press release&lt;/a&gt;, 2026-10-14: pricing of $850 million convertible notes.&amp;#160;&lt;a href="#fnref:13" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:14"&gt;
&lt;p&gt;Second-quarter 2026 results: &lt;a class="link" href="https://investors.corescientific.com/news-events/press-releases/detail/139/core-scientific-announces-second-quarter-2026-results" target="_blank" rel="noopener"
 &gt;Core Scientific&lt;/a&gt;, &lt;a class="link" href="https://investors.terawulf.com/news-events/press-releases/detail/144/terawulf-reports-second-quarter-2026-results" target="_blank" rel="noopener"
 &gt;TeraWulf&lt;/a&gt;, &lt;a class="link" href="https://www.hut8.com/news-insights/press-releases/hut-8-reports-second-quarter-2026-results" target="_blank" rel="noopener"
 &gt;Hut 8&lt;/a&gt;.&amp;#160;&lt;a href="#fnref:14" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:15"&gt;
&lt;p&gt;MarketBeat short interest pages, settlement dates 2026-09-30 and 2026-10-15: &lt;a class="link" href="https://www.marketbeat.com/stocks/NASDAQ/IREN/short-interest/" target="_blank" rel="noopener"
 &gt;IREN&lt;/a&gt;, &lt;a class="link" href="https://www.marketbeat.com/stocks/NASDAQ/NBIS/short-interest/" target="_blank" rel="noopener"
 &gt;NBIS&lt;/a&gt;, &lt;a class="link" href="https://www.marketbeat.com/stocks/NASDAQ/CRWV/short-interest/" target="_blank" rel="noopener"
 &gt;CRWV&lt;/a&gt;. A secondary source; we could not obtain the Nasdaq primary figures.&amp;#160;&lt;a href="#fnref:15" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:16"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.coreweave.com/news/coreweave-expands-agreement-with-openai-by-up-to-6-5b" target="_blank" rel="noopener"
 &gt;CoreWeave press release&lt;/a&gt;, 2025-09-25: a further contract of up to $6.5 billion, bringing the total contract value with OpenAI to approximately $22.4 billion.&amp;#160;&lt;a href="#fnref:16" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;/ol&gt;
&lt;/div&gt;</description></item><item><title>What Naver and SK Telecom Really Earn From AI Data Centers</title><link>https://koreainvestinsights.com/post/naver-sk-telecom-ai-data-center-profits-value-2026-10-25/</link><pubDate>Sun, 25 Oct 2026 20:20:00 +0900</pubDate><guid>https://koreainvestinsights.com/post/naver-sk-telecom-ai-data-center-profits-value-2026-10-25/</guid><description>&lt;img src="https://koreainvestinsights.com/" alt="Featured image of post What Naver and SK Telecom Really Earn From AI Data Centers" /&gt;&lt;p&gt;Our &lt;a class="link" href="https://koreainvestinsights.com/post/naver-sk-telecom-ai-data-centers-gpus-funding-2026-10-25/" &gt;last post&lt;/a&gt; sorted out what Naver and SK Telecom have really secured: plans, orders, deliveries and money. This post asks the next question. If the projects work, how much profit do they add by 2029? And how much of that profit belongs to each company&amp;rsquo;s shareholders?&lt;/p&gt;
&lt;p&gt;Every number below is our own estimate from public filings and company releases up to October 25, 2026. It is not company guidance. We show earnings, shares of value and conditions to watch. We do not give price targets or buy and sell calls.&lt;/p&gt;
&lt;h2 id="five-things-to-take-away"&gt;Five things to take away
&lt;/h2&gt;&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Naver&amp;rsquo;s AI data center adds little profit before 2028.&lt;/strong&gt; In our base case it earns ₩32bn of operating profit in 2027, ₩314bn in 2028 and ₩558bn in 2029. That is 1.3%, 10.7% and 16.3% of Naver&amp;rsquo;s total operating profit in our model.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;A higher GPU rental price does not flow straight to profit.&lt;/strong&gt; Rental prices for H200 and B200 chips rose about 20% in a month. Our model has Naver pay usage fees to a separate company that holds the data center assets, so the extra price helps Naver only after those fees.&lt;sup id="fnref:1"&gt;&lt;a href="#fn:1" class="footnote-ref" role="doc-noteref"&gt;1&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;SK Telecom&amp;rsquo;s AI data center revenue grows faster than the profit its shareholders keep.&lt;/strong&gt; In our base case, related revenue rises from ₩850bn to ₩1.85tn between 2027 and 2029. Profit that belongs to SK Telecom shareholders rises from ₩31bn to ₩154bn, only 3.7% to 8.3% of revenue.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Most of each company&amp;rsquo;s value sits elsewhere.&lt;/strong&gt; The new AI data center is about a tenth of Naver&amp;rsquo;s value in our base model. For SK Telecom, the telecom business, a stake in Anthropic and the existing data center unit matter more than the new GPU and gigawatt projects.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;The US neocloud sell-off is a weak guide.&lt;/strong&gt; US AI cloud stocks fell 33% to 39% from September 30 to October 23. Naver and SK Telecom fell 11% to 13%. After the US stocks rose, Naver rose in the next Korean session 4 times out of 8 and SK Telecom 3 times out of 8.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id="1-how-we-turn-plans-into-profit"&gt;1. How we turn plans into profit
&lt;/h2&gt;&lt;p&gt;We start from company targets and turn them into revenue in three steps. First, how many megawatts (MW) of capacity are available on average during the year. Second, how much of that capacity is used by paying outside customers. Third, how much one contracted megawatt earns in a year.&lt;/p&gt;
&lt;p&gt;For Naver, the company&amp;rsquo;s roadmap is 55MW from the first half of 2027, about 100MW by the end of 2027 and 200MW in 2028.&lt;sup id="fnref:2"&gt;&lt;a href="#fn:2" class="footnote-ref" role="doc-noteref"&gt;2&lt;/a&gt;&lt;/sup&gt;&lt;sup id="fnref:3"&gt;&lt;a href="#fn:3" class="footnote-ref" role="doc-noteref"&gt;3&lt;/a&gt;&lt;/sup&gt; The monthly start dates are not public. We assume capacity averages 40MW in 2027, 150MW in 2028 and 200MW in 2029. We do not add the later 1GW expansion.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Our base case for Naver&lt;/th&gt;
 &lt;th style="text-align: right"&gt;2027&lt;/th&gt;
 &lt;th style="text-align: right"&gt;2028&lt;/th&gt;
 &lt;th style="text-align: right"&gt;2029&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Average available capacity&lt;/td&gt;
 &lt;td style="text-align: right"&gt;40MW&lt;/td&gt;
 &lt;td style="text-align: right"&gt;150MW&lt;/td&gt;
 &lt;td style="text-align: right"&gt;200MW&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Share of capacity in use&lt;/td&gt;
 &lt;td style="text-align: right"&gt;80%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;85%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;90%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Share of that use by outside paying customers&lt;/td&gt;
 &lt;td style="text-align: right"&gt;75%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;80%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;85%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Outside paying capacity&lt;/td&gt;
 &lt;td style="text-align: right"&gt;24MW&lt;/td&gt;
 &lt;td style="text-align: right"&gt;102MW&lt;/td&gt;
 &lt;td style="text-align: right"&gt;153MW&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Revenue per contracted MW a year&lt;/td&gt;
 &lt;td style="text-align: right"&gt;$20m&lt;/td&gt;
 &lt;td style="text-align: right"&gt;$19m&lt;/td&gt;
 &lt;td style="text-align: right"&gt;$18m&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Operating margin after SPV fees and running costs&lt;/td&gt;
 &lt;td style="text-align: right"&gt;5%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;12%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;15%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Outside revenue&lt;/td&gt;
 &lt;td style="text-align: right"&gt;₩648bn&lt;/td&gt;
 &lt;td style="text-align: right"&gt;₩2.62tn&lt;/td&gt;
 &lt;td style="text-align: right"&gt;₩3.72tn&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;We use ₩1,350 per dollar. Our price path starts from a reference point. Nebius said four large contracts it signed in the second quarter yield $20 million to $25 million per megawatt a year.&lt;sup id="fnref:4"&gt;&lt;a href="#fn:4" class="footnote-ref" role="doc-noteref"&gt;4&lt;/a&gt;&lt;/sup&gt; That is Nebius&amp;rsquo;s contract data, not Naver&amp;rsquo;s price.&lt;/p&gt;
&lt;p&gt;The margin is our assumption. Naver has not published a margin for this business or the terms of its costs. We treat 5%, 12% and 15% as the margin left after the usage fees Naver pays to the separate company that holds the data center assets (an SPV), power, staff and other running costs. We do not subtract those fees a second time. The real fee terms are not public.&lt;/p&gt;
&lt;h2 id="2-naver-little-profit-before-2028"&gt;2. Naver: little profit before 2028
&lt;/h2&gt;&lt;div class="fig-scroll" style="overflow-x:auto;-webkit-overflow-scrolling:touch;margin:1.4em 0"&gt;&lt;img src="https://koreainvestinsights.com/post-assets/naver-sk-telecom-ai-data-center-profits-value-2026-10-25/naver-range.en.svg" alt="Operating profit from Naver's new AI data center in billion won under bear, base and bull scenarios. 2027: bear minus 4.9, base 32.4, bull 172.2, base case 1.3 percent of Naver's total operating profit. 2028: bear 15.4, base 314, bull 873, base case 10.7 percent. 2029: bear 78.3, base 558, bull 1,121, base case 16.3 percent. Capacity averages 40, 150 and 200MW in the base case with operating margins of 5, 12 and 15 percent. Dunamu is not included." loading="lazy" style="display:block;width:100%;min-width:720px;max-width:none;height:auto;margin:0 auto"&gt;&lt;/div&gt;
&lt;p&gt;The profit comes late for three reasons. In 2027 the average capacity is only 40MW, because the first 55MW starts in the middle of the year. The margin is thin while the business is new. And only part of the capacity serves outside customers. Naver&amp;rsquo;s own use of its AI computing does not create outside revenue in this model.&lt;/p&gt;
&lt;p&gt;By 2028 the data center adds about 10.7% to Naver&amp;rsquo;s operating profit in our base case. By 2029 it adds about 16.3%, but only if 200MW runs at 90% use with a 15% margin. The bear case is a small loss of ₩5bn in 2027 and a profit of ₩78bn in 2029. The bull case has ₩1.12tn in 2029. We do not attach probabilities to these cases.&lt;/p&gt;
&lt;p&gt;One outside estimate shows how much the margin matters. NH Investment &amp;amp; Securities expects AI Factory revenue of ₩0.73tn in 2027, ₩3.32tn in 2028 and ₩6.51tn in 2029, with operating margins of 5.3%, 6.9% and 8.8%.&lt;sup id="fnref:5"&gt;&lt;a href="#fn:5" class="footnote-ref" role="doc-noteref"&gt;5&lt;/a&gt;&lt;/sup&gt; That gives ₩575bn of operating profit in 2029, close to our ₩558bn. Its revenue is higher partly because it assumes 300MW of capacity in 2029, which is NH&amp;rsquo;s assumption and not a Naver target. A higher revenue and a lower margin lead to about the same profit.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;The cost side can also turn negative.&lt;/strong&gt; Suppose Naver pays the SPV a fixed fee of $13m per megawatt a year, with $1.2m of variable cost and $1m of other cost, and sells at $18m. Then the business needs about 83% use just to cover its cash costs. At 60% use on 200MW, the yearly cash gap is about ₩1.06tn. This is a stress example, not a contract term. The real fee terms and minimum payments are not public.&lt;/p&gt;
&lt;h2 id="3-gpu-prices-rose-but-who-keeps-the-gain"&gt;3. GPU prices rose, but who keeps the gain?
&lt;/h2&gt;&lt;p&gt;Rental prices for GPUs are strong. The Ornn index of traded prices on October 24 showed a one-month rise of 20.2% for H200 and 19.5% for B200. H100 rose 3.4%, A100 5.9%, and the RTX PRO 6000 was flat.&lt;sup id="fnref1:1"&gt;&lt;a href="#fn:1" class="footnote-ref" role="doc-noteref"&gt;1&lt;/a&gt;&lt;/sup&gt; Investor a16z published a chart on October 23 that compares today&amp;rsquo;s rental prices with a straight-line six-year depreciation line. Prices sit from 9% to 143% above that line, depending on the chip.&lt;sup id="fnref:6"&gt;&lt;a href="#fn:6" class="footnote-ref" role="doc-noteref"&gt;6&lt;/a&gt;&lt;/sup&gt; That is a gap to a reference line, not a return on a chip.&lt;/p&gt;
&lt;div class="fig-scroll" style="overflow-x:auto;-webkit-overflow-scrolling:touch;margin:1.4em 0"&gt;&lt;img src="https://koreainvestinsights.com/post-assets/naver-sk-telecom-ai-data-center-profits-value-2026-10-25/gpu-price.en.svg" alt="Left: one-month change in the Ornn GPU price index on October 24: H200 plus 20.2 percent, B200 plus 19.5, A100 plus 5.9, H100 plus 3.4, RTX PRO 6000 zero. Right: Naver AI data center operating profit in billion won. 2028: old price path 314, higher price path with the same margin 413, higher price with the operator keeping 30 percent of the uplift 562. 2029: 558, 713 and 868. The old price path is 20 to 18 million dollars per MW-year, the higher path 25 to 23 million dollars." loading="lazy" style="display:block;width:100%;min-width:720px;max-width:none;height:auto;margin:0 auto"&gt;&lt;/div&gt;
&lt;p&gt;We tested what a higher price path does. We raised revenue per contracted megawatt from $20m, $19m and $18m to $25m, $25m and $23m. If Naver kept the same margin, 2028 operating profit would rise from ₩314bn to ₩413bn. If Naver kept 30 cents of every extra revenue dollar as profit, instead of the 12 cents its margin implies, it would be ₩562bn.&lt;/p&gt;
&lt;p&gt;There are two reasons not to push this further. First, Nebius separates its contract yield of $20m to $25m from a premium price range of $40m to $50m per megawatt, and says it had signed only the first contract in that range.&lt;sup id="fnref1:4"&gt;&lt;a href="#fn:4" class="footnote-ref" role="doc-noteref"&gt;4&lt;/a&gt;&lt;/sup&gt; A premium for scarce new chips does not set a three-year price. Second, Naver does not own the whole stack. If the SPV raises its fee when customer prices rise, Naver&amp;rsquo;s profit barely moves. The profit that counts is extra customer revenue minus extra fees and extra running cost.&lt;/p&gt;
&lt;h2 id="4-navers-other-swing-factors"&gt;4. Naver&amp;rsquo;s other swing factors
&lt;/h2&gt;&lt;p&gt;&lt;strong&gt;The existing business is the larger swing.&lt;/strong&gt; In the second quarter, revenue excluding the Wallapop acquisition grew 14.5%, platform revenue grew 12.3% and advertising grew 7.5%. Operating profit was flat from a year earlier, as spending on computing assets, content and data rose.&lt;sup id="fnref:7"&gt;&lt;a href="#fn:7" class="footnote-ref" role="doc-noteref"&gt;7&lt;/a&gt;&lt;/sup&gt; Our base case slows the existing business to 8%, 7% and 6% growth. A faster path of 12%, 10% and 8% is reasonable but not our base case. Korea Investment &amp;amp; Securities pointed to delays in tenant announcements, a slower advertising and commerce quarter around the Chuseok holiday, and higher costs for payment terminals, broadcast rights and GPU depreciation.&lt;sup id="fnref:8"&gt;&lt;a href="#fn:8" class="footnote-ref" role="doc-noteref"&gt;8&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;One more point sets the scale. In our model, one turn of the earnings multiple on Naver&amp;rsquo;s existing business moves value by about ₩2.0tn. That is about 60% of the value of the whole AI data center (about ₩3.2tn). A one-turn change on the old business is worth more than half of the new project.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Dunamu adds profit to Naver&amp;rsquo;s group, but not all of it to Naver&amp;rsquo;s shareholders.&lt;/strong&gt; Naver Financial plans to take 100% of the crypto exchange operator Dunamu in a share exchange. Naver&amp;rsquo;s holding in Naver Financial would fall from 69.0% to 17.0%. Naver would hold 46.5% of the voting rights including delegated votes, and plans to keep consolidating Naver Financial and, through it, Dunamu. The final control test comes at closing. The deal values 100% of Dunamu at ₩15.1tn. That is a share value, not cash coming to Naver. The chart is dated September 30, 2025, and the holdings can change if shareholders use appraisal rights.&lt;sup id="fnref:9"&gt;&lt;a href="#fn:9" class="footnote-ref" role="doc-noteref"&gt;9&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;The arithmetic for shareholders is simple. Naver&amp;rsquo;s yearly gain is 17% of Dunamu&amp;rsquo;s net profit minus the 52 percentage points of Naver Financial&amp;rsquo;s own profit it gives up. Dunamu earned ₩708.9bn in 2025.&lt;sup id="fnref:10"&gt;&lt;a href="#fn:10" class="footnote-ref" role="doc-noteref"&gt;10&lt;/a&gt;&lt;/sup&gt; Naver Financial&amp;rsquo;s own profit is about ₩133.5bn by our proxy from a secondary summary of its audit report. With those numbers the gain is about ₩51bn a year. But Dunamu earned only ₩108.4bn in the first half of 2026.&lt;sup id="fnref:11"&gt;&lt;a href="#fn:11" class="footnote-ref" role="doc-noteref"&gt;11&lt;/a&gt;&lt;/sup&gt; Doubling that number turns the gain into a loss of about ₩33bn. Dunamu would need to earn about ₩408bn a year for the swap to add to Naver&amp;rsquo;s profit. We did not find a confirmed closing in the filings we checked. Our model assumes the end of March 2027 and adds no synergy from stablecoins, payments or tokenization.&lt;/p&gt;
&lt;h2 id="5-sk-telecom-revenue-grows-faster-than-shareholder-profit"&gt;5. SK Telecom: revenue grows faster than shareholder profit
&lt;/h2&gt;&lt;p&gt;SK Telecom&amp;rsquo;s AI data center business is not one business. It combines three parts. SK Horizon owns existing data centers and subsea cables, and is being split off with outside investors. GPU services rent computing to customers. SK Hyper develops new gigawatt-scale sites. SK Telecom reported AI data center revenue of ₩131.4bn in the first quarter of 2026 and ₩136.2bn in the second.&lt;sup id="fnref:12"&gt;&lt;a href="#fn:12" class="footnote-ref" role="doc-noteref"&gt;12&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;div class="fig-scroll" style="overflow-x:auto;-webkit-overflow-scrolling:touch;margin:1.4em 0"&gt;&lt;img src="https://koreainvestinsights.com/post-assets/naver-sk-telecom-ai-data-center-profits-value-2026-10-25/skt-funnel.en.svg" alt="SK Telecom AI data center related business in our base case, billion won. 2027: revenue 850, operating profit 146, profit to SKT shareholders 31.3, which is 3.7 percent of revenue. 2028: revenue 1,250, operating profit 230, to SKT holders 58.9, 4.7 percent. 2029: revenue 1,850, operating profit 400, to SKT holders 154, 8.3 percent. Horizon is counted at SKT's 51 percent economic stake after the deal; revenue and operating profit are shown at 100 percent." loading="lazy" style="display:block;width:100%;min-width:720px;max-width:none;height:auto;margin:0 auto"&gt;&lt;/div&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Our base case, ₩bn&lt;/th&gt;
 &lt;th style="text-align: right"&gt;2027&lt;/th&gt;
 &lt;th style="text-align: right"&gt;2028&lt;/th&gt;
 &lt;th style="text-align: right"&gt;2029&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Horizon revenue / operating margin&lt;/td&gt;
 &lt;td style="text-align: right"&gt;550 / 25%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;700 / 27%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;900 / 28%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;GPU services revenue / operating margin&lt;/td&gt;
 &lt;td style="text-align: right"&gt;300 / 8%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;550 / 12%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;850 / 15%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;SK Hyper fees / operating profit&lt;/td&gt;
 &lt;td style="text-align: right"&gt;0 / −15&lt;/td&gt;
 &lt;td style="text-align: right"&gt;0 / −25&lt;/td&gt;
 &lt;td style="text-align: right"&gt;100 / 20&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Total revenue&lt;/td&gt;
 &lt;td style="text-align: right"&gt;850&lt;/td&gt;
 &lt;td style="text-align: right"&gt;1,250&lt;/td&gt;
 &lt;td style="text-align: right"&gt;1,850&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Total operating profit&lt;/td&gt;
 &lt;td style="text-align: right"&gt;146&lt;/td&gt;
 &lt;td style="text-align: right"&gt;230&lt;/td&gt;
 &lt;td style="text-align: right"&gt;400&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Profit to SK Telecom shareholders&lt;/td&gt;
 &lt;td style="text-align: right"&gt;31&lt;/td&gt;
 &lt;td style="text-align: right"&gt;59&lt;/td&gt;
 &lt;td style="text-align: right"&gt;154&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Share of SK Telecom&amp;rsquo;s total operating profit&lt;/td&gt;
 &lt;td style="text-align: right"&gt;7.4%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;11.0%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;17.5%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Share of SK Telecom&amp;rsquo;s shareholder profit&lt;/td&gt;
 &lt;td style="text-align: right"&gt;2.3%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;4.1%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;10.0%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Three things pull shareholder profit far below operating profit. First, Horizon&amp;rsquo;s revenue and operating profit are counted in full, but after the deal SK Telecom keeps only 51% of its net profit. KKR and a group led by IMM Investment pay about ₩3.08tn for the rest. Of that, about ₩1.88tn buys existing shares from SK Telecom and about ₩1.2tn is new money for Horizon. The sale is planned for March 4, 2027, subject to conditions.&lt;sup id="fnref:13"&gt;&lt;a href="#fn:13" class="footnote-ref" role="doc-noteref"&gt;13&lt;/a&gt;&lt;/sup&gt;&lt;sup id="fnref:14"&gt;&lt;a href="#fn:14" class="footnote-ref" role="doc-noteref"&gt;14&lt;/a&gt;&lt;/sup&gt; Second, interest and tax come before shareholders. Third, GPU services start with a low margin, and SK Hyper starts as a cost.&lt;/p&gt;
&lt;p&gt;Compared with our 2026 estimate of about ₩53bn of shareholder profit from this business, 2027 is lower by about ₩21bn. &lt;strong&gt;Revenue can grow while the profit a shareholder can claim falls in the first year.&lt;/strong&gt; The ₩1.88tn from selling existing shares is cash, not profit. We treat it as capital that SK Telecom can reuse, and we do not add it to earnings.&lt;/p&gt;
&lt;p&gt;SK Hyper is still mostly a plan. SK Telecom may invest up to ₩750bn by 2030 for sites and substations. The first 5GW is planned in phases from 2029, and the company says it plans to expand toward 15GW in 2035.&lt;sup id="fnref:15"&gt;&lt;a href="#fn:15" class="footnote-ref" role="doc-noteref"&gt;15&lt;/a&gt;&lt;/sup&gt; In an October 13 filing, the company said it has no concrete plan to add another 5GW overseas, after a news report said it was considering one.&lt;sup id="fnref:16"&gt;&lt;a href="#fn:16" class="footnote-ref" role="doc-noteref"&gt;16&lt;/a&gt;&lt;/sup&gt; We include no computing revenue from the 5GW. We include only an assumed ₩100bn of development and operating fees in 2029.&lt;/p&gt;
&lt;h2 id="6-a-token-gateway-an-option-not-in-our-base-case"&gt;6. A token gateway: an option, not in our base case
&lt;/h2&gt;&lt;p&gt;On October 25, ZDNet Korea reported that SK Telecom is preparing a &amp;ldquo;token gateway.&amp;rdquo; The service would manage which AI model a customer uses, how much it uses and what it costs, and would set what work the AI may handle. It would use SK Telecom&amp;rsquo;s identity, payment and billing capabilities. The article says the company has not set a launch date or prices.&lt;sup id="fnref:17"&gt;&lt;a href="#fn:17" class="footnote-ref" role="doc-noteref"&gt;17&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;This idea deserves attention because it adds demand without owning more GPUs. It also faces a hard question. Similar services already exist. Amazon Bedrock routes requests between models in the same model family, and Cloudflare AI Gateway offers usage logs, caching, rate limits, retries and model fallback.&lt;sup id="fnref:18"&gt;&lt;a href="#fn:18" class="footnote-ref" role="doc-noteref"&gt;18&lt;/a&gt;&lt;/sup&gt;&lt;sup id="fnref:19"&gt;&lt;a href="#fn:19" class="footnote-ref" role="doc-noteref"&gt;19&lt;/a&gt;&lt;/sup&gt; SK Telecom would have to show an edge in business integration, access control, support and switching costs. ZDNet cites SK Telecom&amp;rsquo;s 23 million subscribers as an asset. They are telecom customers, and nothing shows they will pay for AI services.&lt;/p&gt;
&lt;p&gt;We put zero value on it in our base case. To see the size, we tested two cases. In the proof case, 500,000 consumers pay ₩120,000 a year and 1,000 companies pay ₩50m, for ₩110bn of net fee revenue. In the expansion case, 1 million consumers pay ₩180,000 and 3,000 companies pay ₩100m, for ₩480bn. At 25% to 30% margins and a 20 times earnings multiple, discounted at 12%, these cases add ₩0.29tn and ₩1.51tn. That is about 1.5% and 7.7% of our base value. These cases are not company forecasts. Token purchases that SK Telecom passes on to model makers would not count as its own revenue.&lt;/p&gt;
&lt;h2 id="7-where-the-value-sits"&gt;7. Where the value sits
&lt;/h2&gt;&lt;div class="fig-scroll" style="overflow-x:auto;-webkit-overflow-scrolling:touch;margin:1.4em 0"&gt;&lt;img src="https://koreainvestinsights.com/post-assets/naver-sk-telecom-ai-data-center-profits-value-2026-10-25/value-sits.en.svg" alt="Where the value sits in our base models. Naver: existing business 90 percent, new AI data center 10 percent; of the AI data center part, 18 percent is earnings and cash from 2027 to 2029 and 82 percent is value after 2029. SK Telecom: telecom and other business 68.9 percent, SK Horizon 51 percent stake 14.2, cash from the Horizon share sale 7.1, Anthropic stake at 65 percent of book value 11.7, GPU services net of capital minus 1.3, SK Hyper net of outlays minus 0.5. Naver's AI part is about 10 percent in the earlier model and 12 to 14 percent in the later one." loading="lazy" style="display:block;width:100%;min-width:720px;max-width:none;height:auto;margin:0 auto"&gt;&lt;/div&gt;
&lt;p&gt;&lt;strong&gt;Naver.&lt;/strong&gt; About 10% of base value comes from the new AI data center, or 12% to 14% under the higher price paths. Of that part, 82% is value after 2029. Only 18% comes from earnings and cash in 2027 to 2029. That makes the AI part sensitive to renewals, equipment replacement and whether profits last. Naver&amp;rsquo;s existing business, and the multiple the market gives it, matter more.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;SK Telecom.&lt;/strong&gt; About 69% of base value comes from telecom and other existing business. SK Horizon accounts for about 14%, and the cash from the share sale for about 7%. The Anthropic stake adds about 12%. SK Telecom holds 3,860,330 shares, rounded to 0.2%, with a book value of ₩3.5tn at the end of June.&lt;sup id="fnref:20"&gt;&lt;a href="#fn:20" class="footnote-ref" role="doc-noteref"&gt;20&lt;/a&gt;&lt;/sup&gt; We count it at 65% of book value because it is unlisted, can be diluted and is hard to cash out. A book-value gain is not cash. The new GPU services and SK Hyper together are slightly negative after we subtract the capital they need.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;SK Telecom&amp;rsquo;s first half gives a baseline.&lt;/strong&gt; Operating profit was ₩1.10tn on revenue of ₩8.75tn. Second-quarter mobile revenue fell 1.9% from a year earlier on fewer subscribers, and the strong profit growth reflects the base effect from last year&amp;rsquo;s USIM replacement costs.&lt;sup id="fnref:21"&gt;&lt;a href="#fn:21" class="footnote-ref" role="doc-noteref"&gt;21&lt;/a&gt;&lt;/sup&gt; The company paid a dividend of ₩830 per share in each of the first two quarters of 2026. It has not promised the same for later quarters.&lt;/p&gt;
&lt;h2 id="8-did-the-us-neocloud-sell-off-drag-them-down"&gt;8. Did the US neocloud sell-off drag them down?
&lt;/h2&gt;&lt;div class="fig-scroll" style="overflow-x:auto;-webkit-overflow-scrolling:touch;margin:1.4em 0"&gt;&lt;img src="https://koreainvestinsights.com/post-assets/naver-sk-telecom-ai-data-center-profits-value-2026-10-25/spillover.en.svg" alt="Price change from Sep 30 to Oct 23 close in local currency: IREN minus 39.1 percent, CoreWeave minus 37.7, Nebius minus 32.7, KOSPI minus 18.8, SK Telecom minus 13.4, Kakao minus 11.3, Naver minus 10.8, KT minus 10.5, QQQ minus 4.8, LG Uplus minus 2.5. After the US basket of CoreWeave and Nebius rose, in the last month (8 times), Naver rose in the next Korean session 4 of 8 times and SK Telecom 3 of 8." loading="lazy" style="display:block;width:100%;min-width:720px;max-width:none;height:auto;margin:0 auto"&gt;&lt;/div&gt;
&lt;p&gt;Possibly in part, but the evidence does not show it was the main cause. Reuters reported on October 14 that bets on higher US interest rates deepened concerns about the financing of AI infrastructure companies.&lt;sup id="fnref:22"&gt;&lt;a href="#fn:22" class="footnote-ref" role="doc-noteref"&gt;22&lt;/a&gt;&lt;/sup&gt; On October 19, the Dong-A Ilbo tied Naver&amp;rsquo;s funding plan to CoreWeave&amp;rsquo;s debt worries. It reported, citing industry sources, a draft plan to raise up to $9bn (about ₩12tn) from outside investors who would earn about 9% a year over six years. Naver told the paper that the structure was still under discussion and that its fundamentals differ from CoreWeave&amp;rsquo;s.&lt;sup id="fnref:23"&gt;&lt;a href="#fn:23" class="footnote-ref" role="doc-noteref"&gt;23&lt;/a&gt;&lt;/sup&gt; That is a plausible path from US stocks to Naver, and it is not a confirmed contract.&lt;/p&gt;
&lt;p&gt;The prices say the link is loose. Naver fell about the same as Kakao and less than the KOSPI, so neoclouds alone do not explain the fall. Following US days when CoreWeave and Nebius rose, Naver rose in the next Korean session 4 times out of 8 and SK Telecom 3 times out of 8. On October 19 the US basket rose 1.2%, and Naver fell 2.7% on October 20. The correlation is positive, at +0.32 for Naver and +0.42 for SK Telecom in the last month. After we remove the Korean market and the Nasdaq-100 ETF (QQQ), it drops to between +0.06 and +0.26 for Naver and between +0.13 and +0.23 for SK Telecom. Eight observations cannot give a probability.&lt;/p&gt;
&lt;p&gt;A US rebound would help only in some forms. A bounce after oversold trading proves little. A fall in interest rates and credit spreads, together with Naver&amp;rsquo;s financing close and customer contracts, would prove more. Even a 20% rise in the value of Naver&amp;rsquo;s AI data center would raise our total value for Naver by only about 2.4% to 2.9%, if nothing else changed.&lt;/p&gt;
&lt;h2 id="9-what-would-change-our-view"&gt;9. What would change our view
&lt;/h2&gt;&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;What to check&lt;/th&gt;
 &lt;th&gt;Why it matters&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Nvidia&amp;rsquo;s $1bn share payment for Naver (latest date October 30) and the Brookfield financing close&lt;sup id="fnref:24"&gt;&lt;a href="#fn:24" class="footnote-ref" role="doc-noteref"&gt;24&lt;/a&gt;&lt;/sup&gt;&lt;sup id="fnref:25"&gt;&lt;a href="#fn:25" class="footnote-ref" role="doc-noteref"&gt;25&lt;/a&gt;&lt;/sup&gt;&lt;/td&gt;
 &lt;td&gt;Turns conditional money into money received&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Naver&amp;rsquo;s first outside customers, the length of the contracts, minimum payments and the SPV fee&lt;/td&gt;
 &lt;td&gt;Decides whether a higher GPU price reaches Naver&amp;rsquo;s profit&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;The month Naver&amp;rsquo;s first 55MW starts and Naver&amp;rsquo;s next earnings&lt;/td&gt;
 &lt;td&gt;Tests our 2027 capacity and margin assumptions&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;SK Telecom&amp;rsquo;s Horizon shareholder meeting, split and share sale (planned January to March 2027)&lt;/td&gt;
 &lt;td&gt;Decides the 51% stake and the cash&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;SK Hyper&amp;rsquo;s outside investors, SK Telecom&amp;rsquo;s guarantees and any minimum lease obligations&lt;/td&gt;
 &lt;td&gt;Shows how much risk stays with SK Telecom&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;SK Telecom&amp;rsquo;s third-quarter AI data center revenue, margin and capital spending, and its 2027 shareholder return policy&lt;/td&gt;
 &lt;td&gt;Tests our ₩850bn revenue and 3.7% profit ratio&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Whether the token gateway launches, with prices and paying customers&lt;/td&gt;
 &lt;td&gt;Turns an option into revenue&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;h2 id="conclusion-profit-arrives-later-and-is-smaller-than-revenue"&gt;Conclusion: profit arrives later and is smaller than revenue
&lt;/h2&gt;&lt;p&gt;Revenue from AI data centers will arrive sooner than profit. For Naver, the profit grows from 2028 and depends on prices, fees and use. For SK Telecom, the profit is diluted by the Horizon deal and early costs. For both companies, the existing business, the multiple the market pays for it and a few large assets decide most of the value.&lt;/p&gt;
&lt;p&gt;Four things would make us more positive: signed contracts with minimum use, money actually received, margins that survive the fees, and falling credit costs. The token gateway for SK Telecom and Dunamu for Naver are options, and they count only after they show up in results.&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Interest note: We use AI models, including Anthropic&amp;rsquo;s Claude and OpenAI&amp;rsquo;s models, in our research. SK Telecom holds a stake in Anthropic, which we count in SK Telecom&amp;rsquo;s base value at 65% of book value, and OpenAI has announced plans to explore AI data centers in Korea with SK Telecom. We cite public figures only and make no judgment about the effect on either company.&lt;/em&gt;&lt;sup id="fnref:26"&gt;&lt;a href="#fn:26" class="footnote-ref" role="doc-noteref"&gt;26&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Disclaimer: For research and information purposes only. Not investment advice. Names cited are for analytical illustration; readers should perform their own due diligence and consult licensed advisors before any investment decision.&lt;/em&gt;&lt;/p&gt;
&lt;div class="footnotes" role="doc-endnotes"&gt;
&lt;hr&gt;
&lt;ol&gt;
&lt;li id="fn:1"&gt;
&lt;p&gt;&lt;a class="link" href="https://data.ornn.com/" target="_blank" rel="noopener"
 &gt;Ornn GPU price index&lt;/a&gt;, updated 2026-10-24 20:00 UTC: H100 SXM $3.02 an hour (+3.42% over one month), H200 $6.20 (+20.16%), B200 $9.57 (+19.48%), A100 SXM4 $1.08 (+5.88%), RTX PRO 6000 $1.55 (0.00%). A limited sample of traded prices.&amp;#160;&lt;a href="#fnref:1" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:1" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:2"&gt;
&lt;p&gt;&lt;a class="link" href="https://navercorp.com/en/story/storyDetail?seq=10034603" target="_blank" rel="noopener"
 &gt;NAVER story&lt;/a&gt;, 2026-07-24: 55MW from the first half of 2027, 200MW in 2028.&amp;#160;&lt;a href="#fnref:2" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:3"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.navercorp.com/en/story/storyDetail?seq=10034410" target="_blank" rel="noopener"
 &gt;NAVER story&lt;/a&gt;: roadmap of 55MW, 100MW and 200MW. The monthly schedule and the 2029 level are our assumptions.&amp;#160;&lt;a href="#fnref:3" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:4"&gt;
&lt;p&gt;&lt;a class="link" href="https://assets.nebius.com/assets/a6ecfd85-a6cb-4967-8ef7-9a25bd261f9c/SHLQ226.pdf" target="_blank" rel="noopener"
 &gt;Nebius second-quarter 2026 shareholder letter&lt;/a&gt;: four large contracts at a yield of $20 to $25 million per megawatt; a &amp;ldquo;$40-50 million per MW&amp;rdquo; price opportunity, with the first such contract signed that week.&amp;#160;&lt;a href="#fnref:4" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:4" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:5"&gt;
&lt;p&gt;&lt;a class="link" href="https://download.nhqv.com/CommFile/cis/rsh/epr/CISPPR20260930175105995.pdf" target="_blank" rel="noopener"
 &gt;NH Investment &amp;amp; Securities, NAVER, 2026-10-01&lt;/a&gt;, pp. 25 and 30: AI Factory revenue ₩730bn, ₩3,320bn and ₩6,510bn and operating profit ₩38.7bn, ₩228.5bn and ₩574.8bn for 2027 to 2029; data center capacity of 100, 200, 300 and 400MW for 2027 to 2030. These are NH&amp;rsquo;s estimates and assumptions.&amp;#160;&lt;a href="#fnref:5" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:6"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.a16z.news/p/charts-of-the-week-attention-is-all" target="_blank" rel="noopener"
 &gt;a16z, Charts of the week&lt;/a&gt;, 2026-10-23: prices &amp;ldquo;anywhere from 9% higher to 143% higher&amp;rdquo; than a straight-line six-year depreciation baseline. An investor&amp;rsquo;s reading, not a return on equipment.&amp;#160;&lt;a href="#fnref:6" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:7"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.navercorp.com/api/article/download/07b3b329-bbbe-481c-ac47-b3d0d7fea855" target="_blank" rel="noopener"
 &gt;NAVER second-quarter 2026 IR presentation&lt;/a&gt;: revenue excluding Wallapop +14.5%, platform +12.3%, advertising +7.5% from a year earlier; operating profit flat as costs rose on computing assets, content and data.&amp;#160;&lt;a href="#fnref:7" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:8"&gt;
&lt;p&gt;&lt;a class="link" href="https://m.koreainvestment.com/main/research/research/StrategyDetail.jsp?id=159813&amp;amp;jkGubun=6" target="_blank" rel="noopener"
 &gt;Korea Investment &amp;amp; Securities, 3Q26 preview&lt;/a&gt;, 2026-10-01. In Korean.&amp;#160;&lt;a href="#fnref:8" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:9"&gt;
&lt;p&gt;&lt;a class="link" href="https://navercorp.com/api/article/download/61f86393-5464-43fa-b7ae-8234c399d168" target="_blank" rel="noopener"
 &gt;NAVER shareholder presentation on the Dunamu deal&lt;/a&gt;: Naver&amp;rsquo;s holding in Naver Financial from 69.0% to 17.0%; 46.5% of voting rights including delegation; Dunamu 100% share value ₩15.1tn; exchange ratio 3.06 to 1. The shareholding may change if appraisal rights are exercised.&amp;#160;&lt;a href="#fnref:9" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:10"&gt;
&lt;p&gt;&lt;a class="link" href="https://zdnet.co.kr/view/?no=20260330182616" target="_blank" rel="noopener"
 &gt;ZDNet Korea&lt;/a&gt;, 2026-03-30: Dunamu consolidated net income of ₩708.9bn for 2025. In Korean.&amp;#160;&lt;a href="#fnref:10" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:11"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.fnnews.com/news/202608141738241349" target="_blank" rel="noopener"
 &gt;Financial News&lt;/a&gt;, 2026-08-14: Dunamu first-half 2026 net income of ₩108.4bn. In Korean.&amp;#160;&lt;a href="#fnref:11" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:12"&gt;
&lt;p&gt;&lt;a class="link" href="https://news.sktelecom.com/en/3218" target="_blank" rel="noopener"
 &gt;SK Telecom second-quarter 2026 results&lt;/a&gt;, 2026-08-05: AI data center revenue ₩136.2bn, up 92.5% from a year earlier; first quarter ₩131.4bn.&amp;#160;&lt;a href="#fnref:12" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:13"&gt;
&lt;p&gt;&lt;a class="link" href="https://news.sktelecom.com/en/3290" target="_blank" rel="noopener"
 &gt;SK Telecom newsroom&lt;/a&gt;, 2026-08-27: SK Horizon to be set up in the first quarter of 2027; 318MW including eight existing data centers and sites under construction; stakes of 51%, 29% and 20%.&amp;#160;&lt;a href="#fnref:13" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:14"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1015650/000119312526369453/d120053d6k.htm" target="_blank" rel="noopener"
 &gt;SK Telecom Form 6-K&lt;/a&gt;, 2026-08-27: 24,481,427 existing shares sold for ₩1,881,128,369,253; 15,617,069 new shares for about ₩1.20 trillion; scheduled disposal date 2027-03-04.&amp;#160;&lt;a href="#fnref:14" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:15"&gt;
&lt;p&gt;&lt;a class="link" href="https://news.sktelecom.com/en/3192" target="_blank" rel="noopener"
 &gt;SK Telecom newsroom&lt;/a&gt;, 2026-07-23: SK Hyper set up; ₩750bn by 2030; phased opening of 5GW in 2029; sequential expansion toward 15GW in 2035.&amp;#160;&lt;a href="#fnref:15" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:16"&gt;
&lt;p&gt;&lt;a class="link" href="https://dart.fss.or.kr/dsaf001/main.do?rcpNo=20261013800718" target="_blank" rel="noopener"
 &gt;SK Telecom clarification filing on DART&lt;/a&gt;, 2026-10-13: no concrete plan at present for an overseas 5GW addition, in reply to a news report. The 15GW figure is a long-term plan under review. In Korean.&amp;#160;&lt;a href="#fnref:16" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:17"&gt;
&lt;p&gt;&lt;a class="link" href="https://zdnet.co.kr/view/?no=20261024132650" target="_blank" rel="noopener"
 &gt;ZDNet Korea&lt;/a&gt;, published 2026-10-25: token gateway to manage model choice, usage, cost and the scope of work given to AI; launch date and pricing not disclosed. In Korean.&amp;#160;&lt;a href="#fnref:17" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:18"&gt;
&lt;p&gt;&lt;a class="link" href="https://docs.aws.amazon.com/bedrock/latest/userguide/prompt-routing.html" target="_blank" rel="noopener"
 &gt;Amazon Bedrock documentation&lt;/a&gt;: prompt routing between foundation models of the same family.&amp;#160;&lt;a href="#fnref:18" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:19"&gt;
&lt;p&gt;&lt;a class="link" href="https://developers.cloudflare.com/ai-gateway/" target="_blank" rel="noopener"
 &gt;Cloudflare AI Gateway documentation&lt;/a&gt;: analytics and logging, caching, rate limiting, request retries and model fallback.&amp;#160;&lt;a href="#fnref:19" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:20"&gt;
&lt;p&gt;&lt;a class="link" href="https://dart.fss.or.kr/dsaf001/main.do?rcpNo=20260813001728" target="_blank" rel="noopener"
 &gt;SK Telecom half-year report on DART&lt;/a&gt;, 2026-08-13: Anthropic 3,860,330 shares, 0.2%, book value ₩3,505,017 million. In Korean.&amp;#160;&lt;a href="#fnref:20" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:21"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.sktelecom.com/img/eng/presen/2026/2Q26InvestorBriefingENG.pdf" target="_blank" rel="noopener"
 &gt;SK Telecom second-quarter 2026 investor briefing&lt;/a&gt;: first-half revenue ₩8,751.4bn and operating profit ₩1,103.6bn; second-quarter mobile revenue down 1.9% on fewer subscribers; dividend ₩830 per share in each of the first and second quarters.&amp;#160;&lt;a href="#fnref:21" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:22"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.investing.com/news/economy-news/ai-infrastructure-firms-slump-as-us-rate-hike-bets-rise-4949759" target="_blank" rel="noopener"
 &gt;Reuters via Investing.com&lt;/a&gt;, 2026-10-14: AI infrastructure stocks fall as expectations of higher US rates deepen financing concerns.&amp;#160;&lt;a href="#fnref:22" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:23"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.donga.com/news/Economy/article/all/20261019/134863396/2" target="_blank" rel="noopener"
 &gt;Dong-A Ilbo&lt;/a&gt;, 2026-10-19, citing unnamed industry sources: up to $9bn (about ₩12tn) of outside funding in a draft structure; Naver said the structure is still under discussion. In Korean.&amp;#160;&lt;a href="#fnref:23" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:24"&gt;
&lt;p&gt;&lt;a class="link" href="https://kind.krx.co.kr/external/2026/07/27/000002/20260727000004/11306.htm" target="_blank" rel="noopener"
 &gt;NAVER disclosure of the share issue to Nvidia (KIND)&lt;/a&gt;, 2026-07-27: about $1 billion, on condition that NAVER secures at least $9 billion of computing infrastructure; October 30, 2026 is the latest payment date and can be changed.&amp;#160;&lt;a href="#fnref:24" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:25"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.navercorp.com/en/media/pressReleasesDetail?seq=10034687" target="_blank" rel="noopener"
 &gt;NAVER press release&lt;/a&gt;, 2026-09-04: Brookfield &amp;ldquo;has committed to investing up to US$9 billion&amp;rdquo; in the 200MW first phase of a 1GW AI factory project.&amp;#160;&lt;a href="#fnref:25" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:26"&gt;
&lt;p&gt;&lt;a class="link" href="https://amp.kedglobal.com/newsAmp/ked202308130002" target="_blank" rel="noopener"
 &gt;The Korea Economic Daily (KED Global)&lt;/a&gt;, 2023-08-13, on SK Telecom&amp;rsquo;s investment in Anthropic; &lt;a class="link" href="https://openai.com/index/samsung-and-sk-join-stargate" target="_blank" rel="noopener"
 &gt;OpenAI&lt;/a&gt;, 2025-10-01, on exploring AI data centers in Korea with SK Telecom.&amp;#160;&lt;a href="#fnref:26" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;/ol&gt;
&lt;/div&gt;</description></item><item><title>Samsung at 5–7x, SK hynix at 6–8x: A Conditional Fair P/E from Normalized Earnings and Cash Returned</title><link>https://koreainvestinsights.com/post/samsung-hynix-fair-per-normalized-earnings-2026-10-25/</link><pubDate>Sun, 25 Oct 2026 17:30:00 +0900</pubDate><guid>https://koreainvestinsights.com/post/samsung-hynix-fair-per-normalized-earnings-2026-10-25/</guid><description>&lt;img src="https://koreainvestinsights.com/" alt="Featured image of post Samsung at 5–7x, SK hynix at 6–8x: A Conditional Fair P/E from Normalized Earnings and Cash Returned" /&gt;&lt;p&gt;On 2027 expected EPS, we put a conditional fair P/E at 5–7x for Samsung Electronics and 6–8x for SK hynix. Both are cyclical companies whose demand has become structurally larger. It is hard to value them on past memory cycles alone. It is just as hard to give them the multiples of NVIDIA or TSMC at this stage.&lt;/p&gt;
&lt;p&gt;Prices are the regular-session close on October 23, 2026. Korean EPS comes from the WiseReport/FnGuide consensus collected the same day. The fair range is not a market consensus. It is our judgment, built from the assumptions below.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Samsung Electronics&lt;/th&gt;
 &lt;th style="text-align: right"&gt;SK hynix&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Close&lt;/td&gt;
 &lt;td style="text-align: right"&gt;₩215,500&lt;/td&gt;
 &lt;td style="text-align: right"&gt;₩1,277,000&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;2027 expected EPS&lt;/td&gt;
 &lt;td style="text-align: right"&gt;₩68,241&lt;/td&gt;
 &lt;td style="text-align: right"&gt;₩453,943&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Current P/E (recomputed)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;3.16x&lt;/td&gt;
 &lt;td style="text-align: right"&gt;2.81x&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Fair P/E in this analysis&lt;/td&gt;
 &lt;td style="text-align: right"&gt;5–7x&lt;/td&gt;
 &lt;td style="text-align: right"&gt;6–8x&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Model midpoint&lt;/td&gt;
 &lt;td style="text-align: right"&gt;5.64x&lt;/td&gt;
 &lt;td style="text-align: right"&gt;6.54x&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;P/E on normalized later EPS&lt;/td&gt;
 &lt;td style="text-align: right"&gt;10–12x&lt;/td&gt;
 &lt;td style="text-align: right"&gt;11–13x&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;The 5–7x and the 10–12x do not contradict each other. The first sits on high 2027 earnings. The second sits on the lower, normalized earnings that follow.&lt;/p&gt;
&lt;h2 id="five-things-to-take-away"&gt;Five things to take away
&lt;/h2&gt;&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;The debate is not whether the cycle is gone. It is how much profit is left after normalization.&lt;/strong&gt; HBM and long-term contracts make earnings more visible. New supply, yield gains, product transitions and price renegotiation all remain.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Today&amp;rsquo;s 3.16x and 2.81x are a question, not proof of cheapness.&lt;/strong&gt; The denominator is peak earnings. To make the price fit our base model, normalized EPS would have to fall to about 25% of 2027 for Samsung and about 18% for SK hynix. That is a back-solved number, not the market&amp;rsquo;s earnings forecast.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Our base case is also optimistic.&lt;/strong&gt; About 80% of the computed value comes from the period after the end of 2028.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;SK hynix&amp;rsquo;s Q2 net profit of ₩93.92tn includes ₩53.24tn of valuation gains on financial instruments.&lt;/strong&gt; Do not read it as recurring operating profit.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;A low P/E is not a floor.&lt;/strong&gt; In the bear stress, the fair multiple falls to 2.21x for Samsung and 1.84x for SK hynix.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id="the-issue-is-the-profit-left-after-normalization-not-the-end-of-the-cycle"&gt;The issue is the profit left after normalization, not the end of the cycle
&lt;/h2&gt;&lt;p&gt;What has changed is clear. Demand now rests on AI servers. HBM is a complex product that needs customer qualification. Equipment and packaging are constrained, and long-term contracts and prepayments now exist. Together these can make earnings visible for longer than before.&lt;/p&gt;
&lt;p&gt;What remains is just as clear. Commodity products substitute for each other. Wafer allocation, new supply and yield gains react with a lag, and customers face their own return limits. Even HBM is contested and repriced at each generation.&lt;/p&gt;
&lt;p&gt;Long-term contracts must be read in three parts: volume, price range and renegotiation terms. Micron&amp;rsquo;s filing is a good example. Its strategic customer agreements are take-or-pay with binding volumes, but pricing is either fixed or periodically negotiated, and for most agreements it sits inside minimum and maximum bands.&lt;sup id="fnref:1"&gt;&lt;a href="#fn:1" class="footnote-ref" role="doc-noteref"&gt;1&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;So we avoid both extremes. &amp;ldquo;AI demand keeps growing, so today&amp;rsquo;s excess profit lasts forever&amp;rdquo; is too much. &amp;ldquo;Memory is cyclical, so 2–3x is always the right multiple&amp;rdquo; ignores cash that comes back and the chance that normal earnings settle higher.&lt;/p&gt;
&lt;p&gt;The width of past cycles shows why the normalization assumption matters. Samsung&amp;rsquo;s consolidated operating margin fell from 24.2% in 2018 (₩58.89tn ÷ ₩243.77tn) to 2.5% in 2023 (₩6.57tn ÷ ₩258.94tn). SK hynix went from 51.5% in 2018 (₩20.84tn ÷ ₩40.45tn) to an operating loss of 23.6% in 2023 (₩7.73tn loss ÷ ₩32.77tn revenue).&lt;sup id="fnref:2"&gt;&lt;a href="#fn:2" class="footnote-ref" role="doc-noteref"&gt;2&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;Both companies are larger now, so we do not use the absolute average of past profit. We estimate the price, margin and cash conversion that can hold at today&amp;rsquo;s scale, and we do not assume normalization happens all at once.&lt;sup id="fnref:3"&gt;&lt;a href="#fn:3" class="footnote-ref" role="doc-noteref"&gt;3&lt;/a&gt;&lt;/sup&gt;&lt;sup id="fnref:4"&gt;&lt;a href="#fn:4" class="footnote-ref" role="doc-noteref"&gt;4&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;Samsung has many business lines, but its current profit is not diversified. Q2 profit in the DS (chip) division was ₩89.2tn, about 99.7% of consolidated operating profit of ₩89.5tn.&lt;sup id="fnref:5"&gt;&lt;a href="#fn:5" class="footnote-ref" role="doc-noteref"&gt;5&lt;/a&gt;&lt;/sup&gt; It is hard to argue that non-memory businesses offset today&amp;rsquo;s memory cycle. Net cash and a possible recovery outside memory are a separate cushion. For SK hynix, we credit the quality of HBM and server memory, but product concentration remains a risk.&lt;sup id="fnref:6"&gt;&lt;a href="#fn:6" class="footnote-ref" role="doc-noteref"&gt;6&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;h2 id="six-companies-six-multiples-the-gap-tracks-how-long-earnings-are-defended"&gt;Six companies, six multiples: the gap tracks how long earnings are defended
&lt;/h2&gt;&lt;p&gt;The table below compares earnings near 2027. It is a reference, not a ranking. Korea uses WiseReport/FnGuide consensus and the US uses yfinance. For NVIDIA and Micron we calendarized fiscal-year estimates. We could not confirm whether US EPS is GAAP or adjusted, or the date of the TSM estimate. Do not read this as a precise comparison of relative undervaluation.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Company&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Oct 23 close&lt;/th&gt;
 &lt;th&gt;Source EPS and period&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Price ÷ EPS&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Calendar-2027 approx.&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Samsung Electronics&lt;/td&gt;
 &lt;td style="text-align: right"&gt;₩215,500&lt;/td&gt;
 &lt;td&gt;CY2027 ₩68,241; CY2028 ₩76,413&lt;/td&gt;
 &lt;td style="text-align: right"&gt;3.16x&lt;/td&gt;
 &lt;td style="text-align: right"&gt;3.16x&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;SK hynix&lt;/td&gt;
 &lt;td style="text-align: right"&gt;₩1,277,000&lt;/td&gt;
 &lt;td&gt;CY2027 ₩453,943; CY2028 ₩500,960&lt;/td&gt;
 &lt;td style="text-align: right"&gt;2.81x&lt;/td&gt;
 &lt;td style="text-align: right"&gt;2.81x&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Micron&lt;/td&gt;
 &lt;td style="text-align: right"&gt;$877.55&lt;/td&gt;
 &lt;td&gt;FY2027 $176.55; FY2028 $206.89&lt;/td&gt;
 &lt;td style="text-align: right"&gt;4.97x; 4.24x&lt;/td&gt;
 &lt;td style="text-align: right"&gt;about 4.70x&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;NVIDIA&lt;/td&gt;
 &lt;td style="text-align: right"&gt;$210.96&lt;/td&gt;
 &lt;td&gt;FY2027 $9.31; FY2028 $15.92&lt;/td&gt;
 &lt;td style="text-align: right"&gt;22.65x; 13.26x&lt;/td&gt;
 &lt;td style="text-align: right"&gt;about 13.73x&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;TSM ADR&lt;/td&gt;
 &lt;td style="text-align: right"&gt;$418.51&lt;/td&gt;
 &lt;td&gt;CY2027 $23.39&lt;/td&gt;
 &lt;td style="text-align: right"&gt;about 17.89x&lt;/td&gt;
 &lt;td style="text-align: right"&gt;about 17.89x&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Intel&lt;/td&gt;
 &lt;td style="text-align: right"&gt;$86.96&lt;/td&gt;
 &lt;td&gt;FY2027 $2.11&lt;/td&gt;
 &lt;td style="text-align: right"&gt;about 41.17x&lt;/td&gt;
 &lt;td style="text-align: right"&gt;about 41.17x, limited meaning&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Our calendar-year approximation for NVIDIA is (FY27 EPS + 11 × FY28 EPS) ÷ 12 = $15.36. For Micron it is (8 × FY27 EPS + 4 × FY28 EPS) ÷ 12 = $186.66. This assumes profit accrues evenly by month and is not a supplier consensus for the calendar year. Do not read the P/Es here as one common forward basis.&lt;/p&gt;
&lt;div class="fig-scroll" style="overflow-x:auto;-webkit-overflow-scrolling:touch;margin:1.4em 0"&gt;&lt;img src="https://koreainvestinsights.com/post-assets/samsung-hynix-fair-per-normalized-earnings-2026-10-25/per-six.en.svg" alt="Reference P/E on earnings near 2027. Samsung Electronics 3.16x, SK hynix 2.81x, Micron about 4.70x, NVIDIA about 13.73x, TSMC about 17.89x, Intel about 41.17x, with limited meaning because its EPS base is small." loading="lazy" style="display:block;width:100%;min-width:720px;max-width:none;height:auto;margin:0 auto"&gt;&lt;/div&gt;
&lt;p&gt;What to see in the figure: the two Korean memory makers trade at about 60–67% of Micron&amp;rsquo;s P/E. A similar multiple does not mean similar value, though. A multiple depends on how long, and how reliably, earnings are defended.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Company&lt;/th&gt;
 &lt;th&gt;What supports the multiple&lt;/th&gt;
 &lt;th&gt;What limits it&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Samsung Electronics&lt;/td&gt;
 &lt;td&gt;Memory scale, HBM expansion, net cash and capital-allocation room&lt;/td&gt;
 &lt;td&gt;Profit concentrated in memory today, commodity memory, non-memory profitability&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;SK hynix&lt;/td&gt;
 &lt;td&gt;HBM and server memory, co-design with customers, long-term contracts&lt;/td&gt;
 &lt;td&gt;Memory concentration, generation-by-generation competition, commodity prices and capacity, non-operating gains&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Micron&lt;/td&gt;
 &lt;td&gt;HBM and leading-edge memory, contracts and cash generation&lt;/td&gt;
 &lt;td&gt;Normalization of record prices and margins, heavy investment&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;NVIDIA&lt;/td&gt;
 &lt;td&gt;Switching costs across CUDA, tools, networking and systems; fabless model&lt;/td&gt;
 &lt;td&gt;AI capex concentration, customers&amp;rsquo; own ASICs, receivables, export controls&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;TSMC&lt;/td&gt;
 &lt;td&gt;Leading-edge process, yield, customer design ties, a manufacturer that does not compete with customers&lt;/td&gt;
 &lt;td&gt;Capital intensity, overseas fab costs, geopolitics&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Intel&lt;/td&gt;
 &lt;td&gt;Option value if the CPU recovery and outside foundry customers work out&lt;/td&gt;
 &lt;td&gt;Small EPS base, losses, cash needs, dilution&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Micron is the closest comparison. NVIDIA and TSMC show what it takes to defend long-term earnings. They are not a reason to lift Korean memory multiples to their level.&lt;/p&gt;
&lt;p&gt;Micron&amp;rsquo;s fiscal 2026 adjusted free cash flow was $62.31bn, strong cash generation.&lt;sup id="fnref:7"&gt;&lt;a href="#fn:7" class="footnote-ref" role="doc-noteref"&gt;7&lt;/a&gt;&lt;/sup&gt; But the $12.75bn of customer deposits received in the same year is scheduled to be repaid almost entirely between 2029 and 2031. If a customer misses its minimum purchase commitment, Micron may keep all or part of the deposit.&lt;sup id="fnref1:1"&gt;&lt;a href="#fn:1" class="footnote-ref" role="doc-noteref"&gt;1&lt;/a&gt;&lt;/sup&gt; That cash cannot be treated as permanent profit. Differences in US listing, capital allocation and contract structure are fair adjustments, but they do not explain all of the large discount on Korean memory.&lt;/p&gt;
&lt;p&gt;NVIDIA&amp;rsquo;s premium comes not only from GPU demand but from the cost of replacing software, networking and systems together. Still, NVIDIA is not free of the AI investment cycle. It carries customer concentration and receivables risk.&lt;sup id="fnref:8"&gt;&lt;a href="#fn:8" class="footnote-ref" role="doc-noteref"&gt;8&lt;/a&gt;&lt;/sup&gt; High margins and switching costs do not remove the cycle. They support the earnings that remain after normalization.&lt;/p&gt;
&lt;p&gt;TSMC&amp;rsquo;s third-quarter operating margin was 58.7%.&lt;sup id="fnref:9"&gt;&lt;a href="#fn:9" class="footnote-ref" role="doc-noteref"&gt;9&lt;/a&gt;&lt;/sup&gt; Leading-edge process and customer design ties support earnings durability, but heavy investment and overseas fab costs come with them. Its position as a manufacturer that serves whichever designer wins also matters. To give Samsung&amp;rsquo;s foundry the same multiple, we would first need evidence on profitability, outside customers and yield.&lt;/p&gt;
&lt;p&gt;Intel&amp;rsquo;s high P/E is not evidence of a better business. In Q2, external customers made up $293m of Intel Foundry&amp;rsquo;s $5,765m of revenue, about 5%, and the segment&amp;rsquo;s operating loss was $2,089m.&lt;sup id="fnref:10"&gt;&lt;a href="#fn:10" class="footnote-ref" role="doc-noteref"&gt;10&lt;/a&gt;&lt;/sup&gt; The company&amp;rsquo;s Q2 GAAP loss was $2.16 per share. A small EPS base and a recovery option lifted the multiple, nothing more. A sum-of-the-parts approach fits better: the normal earnings value of the product business, plus the chance that foundry succeeds, minus future cash needs, net debt and dilution.&lt;/p&gt;
&lt;h2 id="we-split-near-term-cash-from-value-after-normalization"&gt;We split near-term cash from value after normalization
&lt;/h2&gt;&lt;p&gt;The model below is an earnings-and-cash model built on subjective assumptions. It is not a full company DCF and not a consensus fair value. It converts into present value, separately, the cash that can actually be paid out in 2027–2028 and the value placed on normalized earnings at the end of 2028, then adds them.&lt;/p&gt;
&lt;div class="highlight"&gt;&lt;pre tabindex="0" style="color:#f8f8f2;background-color:#272822;-moz-tab-size:4;-o-tab-size:4;tab-size:4;-webkit-text-size-adjust:none;"&gt;&lt;code class="language-text" data-lang="text"&gt;&lt;span style="display:flex;"&gt;&lt;span&gt;Fair P/E today = f ÷ (1+r)^t1 + [f × (E28 ÷ E27) + M × q] ÷ (1+r)^t2
&lt;/span&gt;&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/div&gt;&lt;ul&gt;
&lt;li&gt;E27, E28: per-share earnings in 2027 and 2028. Samsung ₩68,241 and ₩76,413; SK hynix ₩453,943 and ₩500,960.&lt;/li&gt;
&lt;li&gt;t1, t2: years from October 23 to the end of 2027 and the end of 2028. They are 1.1882 and 2.1903 years.&lt;/li&gt;
&lt;li&gt;f: the share of 2027–2028 earnings left as distributable cash after reinvestment and conversion friction. It is not an actual payout ratio.&lt;/li&gt;
&lt;li&gt;q: recurring normalized EPS in 2029 divided by 2027 EPS.&lt;/li&gt;
&lt;li&gt;M: the P/E applied to normalized EPS at the end of 2028. It differs from the P/E on high 2027 earnings.&lt;/li&gt;
&lt;li&gt;r: the required return over the transition period.&lt;/li&gt;
&lt;/ul&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Model assumptions and results&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Samsung Electronics&lt;/th&gt;
 &lt;th style="text-align: right"&gt;SK hynix&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Cash conversion of 2027–2028 earnings (f)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;65%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;70%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;2029 normal EPS ÷ 2027 EPS (q)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;55%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;60%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Normalized EPS (E29)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;₩37,533&lt;/td&gt;
 &lt;td style="text-align: right"&gt;₩272,366&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Terminal P/E on normalized EPS (M)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;10.5x&lt;/td&gt;
 &lt;td style="text-align: right"&gt;11.5x&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Required return over the transition (r)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;12.0%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;12.5%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;PV of 2027–2028 cash ÷ 2027 EPS&lt;/td&gt;
 &lt;td style="text-align: right"&gt;1.136x&lt;/td&gt;
 &lt;td style="text-align: right"&gt;1.205x&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;PV of value after normalization ÷ 2027 EPS&lt;/td&gt;
 &lt;td style="text-align: right"&gt;4.506x&lt;/td&gt;
 &lt;td style="text-align: right"&gt;5.331x&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Total, fair P/E today&lt;/td&gt;
 &lt;td style="text-align: right"&gt;5.642x&lt;/td&gt;
 &lt;td style="text-align: right"&gt;6.536x&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;div class="fig-scroll" style="overflow-x:auto;-webkit-overflow-scrolling:touch;margin:1.4em 0"&gt;&lt;img src="https://koreainvestinsights.com/post-assets/samsung-hynix-fair-per-normalized-earnings-2026-10-25/value-split.en.svg" alt="Composition of the fair P/E. Samsung Electronics total 5.64x: 2027-28 cash 1.14x and value after normalization 4.51x, or 80%. SK hynix total 6.54x: cash 1.21x and value after normalization 5.33x, or 82%." loading="lazy" style="display:block;width:100%;min-width:720px;max-width:none;height:auto;margin:0 auto"&gt;&lt;/div&gt;
&lt;p&gt;What to see in the table: about 80% of the total is value after normalization. Two assumptions, q and M, decide most of the judgment, not the arithmetic.&lt;/p&gt;
&lt;p&gt;The model does not add paid-out cash back into the terminal value. We also define normalized EPS as recurring earnings, excluding interest on cash already paid out and one-time valuation gains. We did not add net cash separately, and we did not value the rest of 2026&amp;rsquo;s cash flow or excess assets on their own. These are assumptions to avoid double counting, not confirmation of actual capital allocation.&lt;/p&gt;
&lt;p&gt;SK hynix&amp;rsquo;s announced policy is to return more than 50% of cumulative free cash flow over 2025–2027.&lt;sup id="fnref:11"&gt;&lt;a href="#fn:11" class="footnote-ref" role="doc-noteref"&gt;11&lt;/a&gt;&lt;/sup&gt; That is not 50% of net profit, and it is not the same concept as the 70% f in the table. Actual payments, buybacks and cancellations must be confirmed before they justify a smaller capital-allocation discount.&lt;/p&gt;
&lt;p&gt;The normal P/E is not a simple historical average of today&amp;rsquo;s peak earnings. With the steady-state formula P/E = (1 − g ÷ ROE) ÷ (ke − g), a normal ROE of 15–30%, long-run growth of 2–3% and a steady-state cost of equity of 10–12% give about 8.7–12.9x. All of these are analytical assumptions.&lt;/p&gt;
&lt;p&gt;For SK hynix we used the upper part of the normal range because we credit the differentiation of HBM and server memory. We reflected pure memory concentration through a higher transition discount rate. Samsung has net cash and recovery potential, but we did not pre-pay a TSMC-style premium for its non-memory businesses.&lt;/p&gt;
&lt;p&gt;The selected ranges are Samsung q 50–65% and M 10–11.5x, which gives 5.04–6.97x, and SK hynix q 55–70% and M 11–12.5x, which gives 5.88–7.97x. Other variables stay at base values. For ease of use we set Samsung at 5–7x and SK hynix at 6–8x. This range is not a confidence interval.&lt;/p&gt;
&lt;h2 id="our-base-case-also-contains-optimism"&gt;Our base case also contains optimism
&lt;/h2&gt;&lt;p&gt;The base case assumes that, even after normalization, a larger business and higher profitability than in the past remain. Holding today&amp;rsquo;s expected revenue and net-profit relationship fixed, Samsung&amp;rsquo;s q of 55% equals a net margin of about 26%, and SK hynix&amp;rsquo;s q of 60% equals about 39%. Samsung is above its past cycle peak, and SK hynix is near its past peak.&lt;/p&gt;
&lt;p&gt;At the same time, 2029 normal EPS sits about 51% below the 2028 consensus for Samsung and about 46% below for SK hynix. That assumes a sharp normalization. The actual timing and path of normalization are unconfirmed.&lt;/p&gt;
&lt;p&gt;Because about 80% of the computed value leans on terminal assumptions, 5–7x and 6–8x should not be accepted as objectively fixed numbers. Testing how long profit lasts matters most.&lt;/p&gt;
&lt;h2 id="for-sk-hynix-check-the-quality-of-net-profit-first"&gt;For SK hynix, check the quality of net profit first
&lt;/h2&gt;&lt;p&gt;SK hynix&amp;rsquo;s Q2 net profit was ₩93.92tn against operating profit of ₩60.54tn. Finance income explains the gap. In the notes to the half-year consolidated statements, Q2 finance income of ₩65.86tn includes ₩53.24tn of gains on valuation of financial instruments and ₩10.03tn of dividend income.&lt;sup id="fnref:12"&gt;&lt;a href="#fn:12" class="footnote-ref" role="doc-noteref"&gt;12&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;These cannot be treated as recurring memory operating profit. From public database information alone, we cannot confirm how the 2027–2028 EPS consensus adjusts for such non-operating items. We did not freeze the recent high net margin and ROE, and we did not use them to justify an NVIDIA-level multiple.&lt;/p&gt;
&lt;p&gt;P/B is a secondary check. On 2026 expected book value per share it is 1.98x for Samsung (₩215,500 ÷ ₩108,812) and 2.41x for SK hynix (₩1,277,000 ÷ ₩530,869).&lt;sup id="fnref:13"&gt;&lt;a href="#fn:13" class="footnote-ref" role="doc-noteref"&gt;13&lt;/a&gt;&lt;/sup&gt; Expected profit surges while book value does not keep pace, so a low 2027 P/E and these P/B levels appear together. We did not justify P/B by freezing 2026 expected ROE, and we did not treat the low end of past P/B as today&amp;rsquo;s floor. EV/EBITDA can also overstate memory if it ignores heavy maintenance and growth capex.&lt;/p&gt;
&lt;h2 id="todays-price-is-neither-proof-of-falling-earnings-nor-proof-of-a-floor"&gt;Today&amp;rsquo;s price is neither proof of falling earnings nor proof of a floor
&lt;/h2&gt;&lt;p&gt;The current price is 3.16x of Samsung&amp;rsquo;s and 2.81x of SK hynix&amp;rsquo;s 2027 expected earnings. Hold f, r, M and E28 at base values and solve for the q that fits today&amp;rsquo;s price: about 24.7% for Samsung and about 18.1% for SK hynix. That is a conditional residual of the model. It does not mean the market expects that much of a decline.&lt;/p&gt;
&lt;p&gt;It does not confirm a floor either. If EPS is overestimated, cash conversion is lower or the capital-allocation discount is larger, a different mix can explain the same price.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Scenario&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Samsung PV ÷ E27&lt;/th&gt;
 &lt;th style="text-align: right"&gt;SK hynix PV ÷ E27&lt;/th&gt;
 &lt;th&gt;Key conditions&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Bear stress&lt;/td&gt;
 &lt;td style="text-align: right"&gt;2.21x&lt;/td&gt;
 &lt;td style="text-align: right"&gt;1.84x&lt;/td&gt;
 &lt;td&gt;E28 ÷ E27 cut to 0.75 and 0.65; q 25% and 20%; cash conversion 50%; normal P/E 8x; discount rate 14% and 15%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Base&lt;/td&gt;
 &lt;td style="text-align: right"&gt;5.64x&lt;/td&gt;
 &lt;td style="text-align: right"&gt;6.54x&lt;/td&gt;
 &lt;td&gt;Assumptions in the table above&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Bull stress&lt;/td&gt;
 &lt;td style="text-align: right"&gt;10.22x&lt;/td&gt;
 &lt;td style="text-align: right"&gt;11.08x&lt;/td&gt;
 &lt;td&gt;q 85% and 90%; cash conversion 75% and 80%; normal P/E 13x and 13.5x; discount rate 10.5% and 11%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;div class="fig-scroll" style="overflow-x:auto;-webkit-overflow-scrolling:touch;margin:1.4em 0"&gt;&lt;img src="https://koreainvestinsights.com/post-assets/samsung-hynix-fair-per-normalized-earnings-2026-10-25/fair-range.en.svg" alt="Current P/E and fair range. Samsung Electronics now 3.16x, fair 5-7x with a midpoint of 5.64x, bear 2.21x, bull 10.22x. SK hynix now 2.81x, fair 6-8x with a midpoint of 6.54x, bear 1.84x, bull 11.08x." loading="lazy" style="display:block;width:100%;min-width:720px;max-width:none;height:auto;margin:0 auto"&gt;&lt;/div&gt;
&lt;p&gt;The stress cases are hypothetical paths with no probability attached. The bear value is not a floor on losses. If 2027 and 2028 EPS both fall 30% with the other ratios held, the base value also falls by the same 30%. A low P/E today does not protect against loss. These figures are the arithmetic result of multiple and EPS assumptions, not price targets.&lt;/p&gt;
&lt;h2 id="what-would-change-the-multiple-post-2028-price-volume-and-cash"&gt;What would change the multiple: post-2028 price, volume and cash
&lt;/h2&gt;&lt;p&gt;Four things would raise the multiple. Contract price and volume terms after 2028 become specific. Recurring earnings and cash flow are confirmed. Capital allocation is carried out without dilution. Profitability holds after the expected capacity additions.&lt;/p&gt;
&lt;p&gt;Four things would lower it. Annual EPS cuts continue. Long-term contract prices are renegotiated or deferred. Commodity DRAM goes into oversupply. Utilization or power runs late. An earnings gap that appears once non-operating items are stripped out belongs on this list too. Raising the multiple on peak earnings does not create value on its own.&lt;/p&gt;
&lt;p&gt;We will cross-check price, contracts, capex and recurring profit against the detailed results from SK hynix on October 28 and Samsung on October 29.&lt;sup id="fnref:14"&gt;&lt;a href="#fn:14" class="footnote-ref" role="doc-noteref"&gt;14&lt;/a&gt;&lt;/sup&gt; On October 30 we will recheck the EPS consensus and the f, q and M ranges. These dates are research checkpoints and have nothing to do with trade execution.&lt;/p&gt;
&lt;h2 id="what-we-could-not-confirm"&gt;What we could not confirm
&lt;/h2&gt;&lt;ul&gt;
&lt;li&gt;The accounting basis of US EPS (GAAP or adjusted) and the as-of date of the latest TSM estimate.&lt;/li&gt;
&lt;li&gt;How the Korean 2027–2028 EPS consensus adjusts for non-operating items.&lt;/li&gt;
&lt;li&gt;We did not convert Samsung&amp;rsquo;s and SK hynix&amp;rsquo;s consolidated net profit into EPS attributable to owners.&lt;/li&gt;
&lt;li&gt;Contract prices and cash conversion after 2028.&lt;/li&gt;
&lt;li&gt;For Intel, differences in share-count observations remain, so we did not fix EV/Sales or an exact fair P/E.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;We cross-checked price ÷ EPS by reverse multiplication, and the discounting model both as a split sum of cash and terminal value and as a single formula. We recomputed the numbers in this post in code before publishing. We rechecked the key figures for Micron, SK hynix, TSMC and Intel against the filings. Company IR material is the company&amp;rsquo;s own view, and supply-shortage outlooks are management&amp;rsquo;s outlook. Brokerage and aggregated consensus can carry conflicts of interest and estimation error.&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Disclaimer: For research and information purposes only. Not investment advice. Names cited are for analytical illustration; readers should perform their own due diligence and consult licensed advisors before any investment decision.&lt;/em&gt;&lt;/p&gt;
&lt;div class="footnotes" role="doc-endnotes"&gt;
&lt;hr&gt;
&lt;ol&gt;
&lt;li id="fn:1"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/723125/000072312526000023/mu-20260903.htm" target="_blank" rel="noopener"
 &gt;Micron fiscal 2026 annual report (10-K)&lt;/a&gt;. Strategic customer agreements are take-or-pay with binding volumes; pricing is fixed or periodically negotiated, and most agreements carry minimum and maximum price bands. Deposits are &amp;ldquo;scheduled to be repaid&amp;rdquo; almost entirely between 2029 and 2031, and Micron may retain all or part of a deposit if the customer misses its minimum purchase commitment.&amp;#160;&lt;a href="#fnref:1" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:1" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:2"&gt;
&lt;p&gt;Samsung &lt;a class="link" href="https://englishdart.fss.or.kr/dsbh002/viewer.do?rcpNo=20190401004781" target="_blank" rel="noopener"
 &gt;2018 annual filing&lt;/a&gt; and &lt;a class="link" href="https://news.samsung.com/global/samsung-electronics-announces-fourth-quarter-and-fy-2023-results" target="_blank" rel="noopener"
 &gt;2023 results&lt;/a&gt;; SK hynix &lt;a class="link" href="https://news.skhynix.com/en/sk-hynix-inc-reports-fiscal-year-2018-and-fourth-quarter-results/" target="_blank" rel="noopener"
 &gt;2018 results&lt;/a&gt; and &lt;a class="link" href="https://www.skhynix.com/ir/UI-FR-IR06/" target="_blank" rel="noopener"
 &gt;financial data&lt;/a&gt;. Units are trillions of won.&amp;#160;&lt;a href="#fnref:2" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:3"&gt;
&lt;p&gt;&lt;a class="link" href="https://pages.stern.nyu.edu/~adamodar/New_Home_Page/valquestions/normearn.htm" target="_blank" rel="noopener"
 &gt;Damodaran, Normalized earnings&lt;/a&gt;. Method for estimating normalized earnings.&amp;#160;&lt;a href="#fnref:3" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:4"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.mckinsey.com/capabilities/strategy-and-corporate-finance/our-insights/how-to-value-cyclical-companies" target="_blank" rel="noopener"
 &gt;McKinsey, How to value cyclical companies&lt;/a&gt;. Scenario approach for cyclical companies.&amp;#160;&lt;a href="#fnref:4" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:5"&gt;
&lt;p&gt;&lt;a class="link" href="https://images.samsung.com/is/content/samsung/assets/global/ir/docs/2026_2Q_conference_eng.pdf" target="_blank" rel="noopener"
 &gt;Samsung Electronics Q2 2026 earnings conference materials&lt;/a&gt;. DS division operating profit ₩89.2tn; consolidated operating profit ₩89.5tn.&amp;#160;&lt;a href="#fnref:5" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:6"&gt;
&lt;p&gt;&lt;a class="link" href="https://news.skhynix.com/en/q2-2026-business-results/" target="_blank" rel="noopener"
 &gt;SK hynix Q2 2026 business results&lt;/a&gt;. Revenue ₩79.3187tn, operating profit ₩60.5426tn, net profit ₩93.9226tn.&amp;#160;&lt;a href="#fnref:6" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:7"&gt;
&lt;p&gt;&lt;a class="link" href="https://investors.micron.com/news/press-release/2026/Micron-Technology-Inc--Reports-Record-Fiscal-Fourth-Quarter-and-Full-Year-2026-Results/default.aspx" target="_blank" rel="noopener"
 &gt;Micron fiscal Q4 and full-year 2026 results&lt;/a&gt;. Adjusted free cash flow was $62,308m for the year and $33,199m for Q4. Proceeds from customer contract liability deposits in the annual cash flow statement were $12,747m.&amp;#160;&lt;a href="#fnref:7" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:8"&gt;
&lt;p&gt;&lt;a class="link" href="https://investor.nvidia.com/news/press-release-details/2026/NVIDIA-Announces-Financial-Results-for-Second-Quarter-Fiscal-2027/default.aspx" target="_blank" rel="noopener"
 &gt;NVIDIA fiscal 2027 Q2 results&lt;/a&gt; and &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1045810/000104581026000075/nvda-20260726.htm" target="_blank" rel="noopener"
 &gt;Form 10-Q&lt;/a&gt;.&amp;#160;&lt;a href="#fnref:8" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:9"&gt;
&lt;p&gt;&lt;a class="link" href="https://pr.tsmc.com/english/news/3346" target="_blank" rel="noopener"
 &gt;TSMC third quarter 2026 results&lt;/a&gt;. Operating margin 58.7%.&amp;#160;&lt;a href="#fnref:9" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:10"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.intc.com/news-events/press-releases/detail/1776/intel-reports-second-quarter-2026-financial-results" target="_blank" rel="noopener"
 &gt;Intel Q2 2026 results&lt;/a&gt; and &lt;a class="link" href="https://www.intc.com/filings-reports/all-sec-filings/content/0000050863-26-000157/intc-20260627.htm" target="_blank" rel="noopener"
 &gt;Form 10-Q&lt;/a&gt;. The $293m of Intel Foundry revenue from external customers is from the 10-Q.&amp;#160;&lt;a href="#fnref:10" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:11"&gt;
&lt;p&gt;&lt;a class="link" href="https://news.skhynix.com/en/share-buyback-and-retirement/" target="_blank" rel="noopener"
 &gt;SK hynix shareholder return announcement&lt;/a&gt;.&amp;#160;&lt;a href="#fnref:11" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:12"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/2120882/000119312526354777/d147827d6k.htm" target="_blank" rel="noopener"
 &gt;SK hynix half-year filing (SEC Form 6-K)&lt;/a&gt;, notes to the consolidated statements. Of Q2 finance income of ₩65,858,814m, gain on valuation of financial instruments was ₩53,237,425m and dividend income was ₩10,027,945m.&amp;#160;&lt;a href="#fnref:12" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:13"&gt;
&lt;p&gt;&lt;a class="link" href="https://navercomp.wisereport.co.kr/v2/company/c1010001.aspx?cmp_cd=005930" target="_blank" rel="noopener"
 &gt;Samsung Electronics on WiseReport&lt;/a&gt; and &lt;a class="link" href="https://navercomp.wisereport.co.kr/v2/company/c1010001.aspx?cmp_cd=000660" target="_blank" rel="noopener"
 &gt;SK hynix on WiseReport&lt;/a&gt;. FnGuide consensus collected on October 23, 2026, including 2027–2028 EPS and 2026 expected book value per share.&amp;#160;&lt;a href="#fnref:13" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:14"&gt;
&lt;p&gt;&lt;a class="link" href="https://news.skhynix.com/en/q3-2026-earnings-conference-call-invitation/" target="_blank" rel="noopener"
 &gt;SK hynix Q3 earnings call notice&lt;/a&gt; (October 28, 2026, 9:00 a.m. KST) and &lt;a class="link" href="https://www.samsung.com/global/ir/ir-events-presentations/events/" target="_blank" rel="noopener"
 &gt;Samsung Electronics IR calendar&lt;/a&gt; (October 29, 2026, 10:00 a.m. KST).&amp;#160;&lt;a href="#fnref:14" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;/ol&gt;
&lt;/div&gt;</description></item><item><title>Six AI Giants Ranked: Microsoft Leads Overall, While Google, Meta and SpaceX Win Their Own Categories</title><link>https://koreainvestinsights.com/post/hyperscaler-leadership-six-company-comparison-2026-10-25/</link><pubDate>Sun, 25 Oct 2026 17:30:00 +0900</pubDate><guid>https://koreainvestinsights.com/post/hyperscaler-leadership-six-company-comparison-2026-10-25/</guid><description>&lt;img src="https://koreainvestinsights.com/" alt="Featured image of post Six AI Giants Ranked: Microsoft Leads Overall, While Google, Meta and SpaceX Win Their Own Categories" /&gt;&lt;p&gt;We judge &lt;strong&gt;Microsoft&lt;/strong&gt; the strongest overall. It has the most balanced mix of three things: the ability to sell technology to paying customers, recurring profit, and cash that is still left after heavy investment.&lt;/p&gt;
&lt;p&gt;The category winners differ, though. &lt;strong&gt;Google&lt;/strong&gt; leads in technology vertical integration, &lt;strong&gt;Meta&lt;/strong&gt; in the own-model scores we could verify, and &lt;strong&gt;SpaceX&lt;/strong&gt; in scarce physical infrastructure.&lt;/p&gt;
&lt;p&gt;The reference date is October 25, 2026. Share prices are October 23 closes. We ranked on cash generation first, then paid distribution, then proven growth.&lt;/p&gt;
&lt;h2 id="the-overall-ranking-puts-cash-first-then-paid-distribution-then-proven-growth"&gt;The overall ranking puts cash first, then paid distribution, then proven growth
&lt;/h2&gt;&lt;p&gt;The table below is a provisional ranking. It puts six companies on the same footing and lists each one&amp;rsquo;s strongest edge next to what holds its rank down.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Provisional rank&lt;/th&gt;
 &lt;th&gt;Company&lt;/th&gt;
 &lt;th&gt;Strongest edge&lt;/th&gt;
 &lt;th&gt;What limits the rank&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;&lt;strong&gt;1&lt;/strong&gt;&lt;/td&gt;
 &lt;td&gt;&lt;strong&gt;Microsoft&lt;/strong&gt;&lt;/td&gt;
 &lt;td&gt;Enterprise work, developer and security channels plus Azure, strong cash generation&lt;/td&gt;
 &lt;td&gt;Reliance on outside frontier models; payback on AI spending&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;&lt;strong&gt;2&lt;/strong&gt;&lt;/td&gt;
 &lt;td&gt;&lt;strong&gt;Google&lt;/strong&gt;&lt;/td&gt;
 &lt;td&gt;Models, TPUs, network, Search, Android and YouTube in one stack&lt;/td&gt;
 &lt;td&gt;Supply limits; moving Search to a new revenue model&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;&lt;strong&gt;3&lt;/strong&gt;&lt;/td&gt;
 &lt;td&gt;&lt;strong&gt;Meta&lt;/strong&gt;&lt;/td&gt;
 &lt;td&gt;A competitive own model joined to ad monetization and consumer channels&lt;/td&gt;
 &lt;td&gt;Heavy spending, ad concentration, dilution and governance&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;&lt;strong&gt;4&lt;/strong&gt;&lt;/td&gt;
 &lt;td&gt;&lt;strong&gt;Amazon&lt;/strong&gt;&lt;/td&gt;
 &lt;td&gt;AWS customer base, own chips, advertising and commerce&lt;/td&gt;
 &lt;td&gt;Very large investment and a complicated consolidated capital picture&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;&lt;strong&gt;5&lt;/strong&gt;&lt;/td&gt;
 &lt;td&gt;&lt;strong&gt;SpaceX&lt;/strong&gt;&lt;/td&gt;
 &lt;td&gt;A rare mix of launch, satellite connectivity, compute and models&lt;/td&gt;
 &lt;td&gt;Large cash burn, acquisition integration, execution and funding risk&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;&lt;strong&gt;6&lt;/strong&gt;&lt;/td&gt;
 &lt;td&gt;&lt;strong&gt;Oracle&lt;/strong&gt;&lt;/td&gt;
 &lt;td&gt;Enterprise databases and large AI cluster supply&lt;/td&gt;
 &lt;td&gt;Debt, customer and supplier concentration, reliance on prepayments&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Meta and Amazon are close. Put more weight on growth speed and new business potential, and SpaceX moves up. This is not an expected-return ranking or a buy list. Prices and personal circumstances are a separate decision.&lt;/p&gt;
&lt;h2 id="microsoft-wins-because-it-turns-good-models-into-money"&gt;Microsoft wins because it turns good models into money
&lt;/h2&gt;&lt;p&gt;Microsoft&amp;rsquo;s trailing twelve-month operating cash flow is &lt;strong&gt;$182.9 billion&lt;/strong&gt; and cash capex is &lt;strong&gt;$115.9 billion&lt;/strong&gt;. Simple free cash flow is $182.9 billion − $115.9 billion = &lt;strong&gt;$67.0 billion&lt;/strong&gt;. Free cash flow is the cash earned from operations minus what was spent on equipment. It means the company still has cash left after a very large investment.&lt;/p&gt;
&lt;p&gt;Growth continues too. Azure grew 43%, and paid Microsoft 365 Copilot seats passed 30 million.&lt;sup id="fnref:1"&gt;&lt;a href="#fn:1" class="footnote-ref" role="doc-noteref"&gt;1&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;These numbers matter because the revenue model survives a change of model leader. Microsoft holds the enterprise accounts, admin rights, purchasing contracts and work tools. Whichever model is ahead, revenue can come through that channel.&lt;/p&gt;
&lt;p&gt;The weak points are clear. Microsoft leans on outside frontier models, and whether its AI spending turns into profit has to be checked continuously.&lt;/p&gt;
&lt;h2 id="on-technology-alone-meta-and-google-are-the-closer-challengers"&gt;On technology alone, Meta and Google are the closer challengers
&lt;/h2&gt;&lt;p&gt;These are the Artificial Analysis v4.3.2 results. Each uses the model&amp;rsquo;s high-performance setting, and this is not a same-cost test.&lt;sup id="fnref:2"&gt;&lt;a href="#fn:2" class="footnote-ref" role="doc-noteref"&gt;2&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Model&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Intelligence index&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Cost per evaluation task&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Meta Muse Wonder Full&lt;/td&gt;
 &lt;td style="text-align: right"&gt;&lt;strong&gt;56&lt;/strong&gt;&lt;/td&gt;
 &lt;td style="text-align: right"&gt;not shown&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Google Gemini 4 Argon&lt;/td&gt;
 &lt;td style="text-align: right"&gt;&lt;strong&gt;53&lt;/strong&gt;&lt;/td&gt;
 &lt;td style="text-align: right"&gt;$1.99&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Meta Muse Wonder&lt;/td&gt;
 &lt;td style="text-align: right"&gt;52&lt;/td&gt;
 &lt;td style="text-align: right"&gt;$4.54&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;SpaceXAI Grok 4.7&lt;/td&gt;
 &lt;td style="text-align: right"&gt;46&lt;/td&gt;
 &lt;td style="text-align: right"&gt;$3.74&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Meta&amp;rsquo;s top score of 56 is the best result &lt;strong&gt;among the three companies whose scores we checked: Meta, Google and SpaceXAI&lt;/strong&gt;. We have not compared equivalent latest own models from Microsoft, Amazon or Oracle, so that comparison is incomplete. We also separated the ability to sell outside models from the ability to do own research.&lt;/p&gt;
&lt;p&gt;On people and research, three things stand out: the breadth of Google DeepMind and the TPU organization, Meta MSL&amp;rsquo;s recent model progress, and the pairing of Microsoft Research and MAI with Azure. A TPU is Google&amp;rsquo;s own AI chip.&lt;/p&gt;
&lt;p&gt;We did not rank on paper counts alone. The Gemini technical report, Llama research and Phi-4 report show accumulated research. They are separate from current product results.&lt;sup id="fnref:3"&gt;&lt;a href="#fn:3" class="footnote-ref" role="doc-noteref"&gt;3&lt;/a&gt;&lt;/sup&gt;&lt;sup id="fnref:4"&gt;&lt;a href="#fn:4" class="footnote-ref" role="doc-noteref"&gt;4&lt;/a&gt;&lt;/sup&gt;&lt;sup id="fnref:5"&gt;&lt;a href="#fn:5" class="footnote-ref" role="doc-noteref"&gt;5&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;h2 id="in-finance-cash-left-after-investing-separates-the-companies-more-than-growth-does"&gt;In finance, cash left after investing separates the companies more than growth does
&lt;/h2&gt;&lt;p&gt;Units are billions of US dollars. We used the latest reported quarter, so quarter-ends differ by company. Simple FCF is operating cash flow minus cash capex.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Company&lt;/th&gt;
 &lt;th&gt;Quarter end&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Revenue growth&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Simple FCF&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Microsoft&lt;/td&gt;
 &lt;td&gt;June&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+18%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;&lt;strong&gt;+19.64&lt;/strong&gt;&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Google&lt;/td&gt;
 &lt;td&gt;June&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+24%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;−5.86&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Meta&lt;/td&gt;
 &lt;td&gt;September&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+26%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;−1.16&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Amazon&lt;/td&gt;
 &lt;td&gt;June&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+20%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;−8.82&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;SpaceX&lt;/td&gt;
 &lt;td&gt;September&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+161%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;−9.96&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Oracle&lt;/td&gt;
 &lt;td&gt;August&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+30%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;−5.40&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Oracle (30%) and SpaceX (161%) grow revenue faster. But Microsoft is the only one with positive cash after paying for investment. The chart below shows the same numbers as bars.&lt;/p&gt;
&lt;div class="fig-scroll" style="overflow-x:auto;-webkit-overflow-scrolling:touch;margin:1.4em 0"&gt;&lt;img src="https://koreainvestinsights.com/post-assets/hyperscaler-leadership-six-company-comparison-2026-10-25/fcf-quarter.en.svg" alt="Quarterly simple FCF, $bn. Microsoft +19.64, Google -5.86, Meta -1.16, Amazon -8.82, SpaceX -9.96, Oracle -5.40. Only Microsoft is positive." loading="lazy" style="display:block;width:100%;min-width:720px;max-width:none;height:auto;margin:0 auto"&gt;&lt;/div&gt;
&lt;p&gt;The readings differ by company, and the differences matter:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Google:&lt;/strong&gt; Free cash flow was negative in the latest quarter. But trailing twelve-month FCF at the end of June was &lt;strong&gt;$53.3 billion&lt;/strong&gt;, and $242.5 billion of cash and securities is a large cushion. Cash raised from financing must be kept apart from cash created by operations.&lt;sup id="fnref:6"&gt;&lt;a href="#fn:6" class="footnote-ref" role="doc-noteref"&gt;6&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Meta:&lt;/strong&gt; Excluding $9.89 billion of legal costs, operating margin is &lt;strong&gt;35.4%&lt;/strong&gt;. That is not a collapse of the core business, but it is below 40.1% a year earlier. Company-defined free cash flow after finance-lease principal is &lt;strong&gt;−$2.12 billion&lt;/strong&gt;, lower than the simple FCF in the table (−$1.16 billion).&lt;sup id="fnref:7"&gt;&lt;a href="#fn:7" class="footnote-ref" role="doc-noteref"&gt;7&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Amazon:&lt;/strong&gt; AWS grew 37% and earned $16.6 billion of operating income. But consolidated capex also includes logistics. Total investment burden and AWS economics should be read separately.&lt;sup id="fnref:8"&gt;&lt;a href="#fn:8" class="footnote-ref" role="doc-noteref"&gt;8&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;SpaceX:&lt;/strong&gt; The Superintelligence (SI) business is already about &lt;strong&gt;52%&lt;/strong&gt; of quarterly revenue. Classifying it as a space company alone misses the current business. Its large cash balance also reflects the IPO and financing, and its investment spending is large.&lt;sup id="fnref:9"&gt;&lt;a href="#fn:9" class="footnote-ref" role="doc-noteref"&gt;9&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Oracle:&lt;/strong&gt; Infrastructure cloud (IaaS) grew 121%. But operating cash flow includes a &lt;strong&gt;$11.36 billion&lt;/strong&gt; rise in customer prepayments, and these carry a significant financing component (they work as financing through payment timing). Backlog and prepayments do not guarantee a long-term return on the investment.&lt;sup id="fnref:10"&gt;&lt;a href="#fn:10" class="footnote-ref" role="doc-noteref"&gt;10&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;On credit quality, the evidence for Microsoft&amp;rsquo;s AAA/Aaa rating is the strongest. For some of the others, the rating data we hold is old, so a full comparison of current ratings is incomplete.&lt;/p&gt;
&lt;h2 id="what-the-market-may-under-price-is-who-keeps-the-profit-not-who-owns-the-assets"&gt;What the market may under-price is who keeps the profit, not who owns the assets
&lt;/h2&gt;&lt;p&gt;Who keeps the profit means which company&amp;rsquo;s revenue and cash the AI profit finally lands in. It is a different question from who owns the most data centers or chips.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Company&lt;/th&gt;
 &lt;th&gt;Qualitative strength to watch&lt;/th&gt;
 &lt;th&gt;Counter-question to check&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Microsoft&lt;/td&gt;
 &lt;td&gt;A channel that already controls enterprise deployment, security and purchasing&lt;/td&gt;
 &lt;td&gt;Does more usage turn into more willingness to pay?&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Google&lt;/td&gt;
 &lt;td&gt;Vertical integration that captures the cost savings of its own chips&lt;/td&gt;
 &lt;td&gt;Do supply shortages and Search cannibalization cancel the profit?&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Meta&lt;/td&gt;
 &lt;td&gt;Consumer channels and early compute purchases support fast launches&lt;/td&gt;
 &lt;td&gt;Can paid retention and inference cost carry the spending and dilution?&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Amazon&lt;/td&gt;
 &lt;td&gt;An existing AWS customer base and operating system to sell its own chips into&lt;/td&gt;
 &lt;td&gt;Is there a real cost advantage large enough to beat switching costs?&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;SpaceX&lt;/td&gt;
 &lt;td&gt;A launch, connectivity and compute combination rivals cannot easily copy&lt;/td&gt;
 &lt;td&gt;Does founder-driven execution speed raise the risk of running many projects at once?&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Oracle&lt;/td&gt;
 &lt;td&gt;Core enterprise data and large capacity contracts&lt;/td&gt;
 &lt;td&gt;Does cash remain after customer concentration and prepayments?&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;This table is a &lt;strong&gt;hypothesis about possible under-pricing&lt;/strong&gt;. We have no like-for-like valuation and no market expectations, so we do not call any of these companies undervalued.&lt;/p&gt;
&lt;h2 id="the-share-price-trend-is-also-strongest-for-microsoft-today"&gt;The share-price trend is also strongest for Microsoft today
&lt;/h2&gt;&lt;p&gt;The table below sets the October 23 close beside the 200-day moving average and the 14-day RSI. The 200-day line is the average close of the last 200 trading days. RSI measures the strength of recent gains against losses over 14 days and is used to judge overheating or weakness.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Company&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Close ($)&lt;/th&gt;
 &lt;th style="text-align: right"&gt;200-day ($)&lt;/th&gt;
 &lt;th style="text-align: right"&gt;RSI 14&lt;/th&gt;
 &lt;th&gt;Technical read&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Microsoft&lt;/td&gt;
 &lt;td style="text-align: right"&gt;524.51&lt;/td&gt;
 &lt;td style="text-align: right"&gt;436.51&lt;/td&gt;
 &lt;td style="text-align: right"&gt;52.6&lt;/td&gt;
 &lt;td&gt;Medium-term strength. Watch support at 502–508 and a reclaim of 527&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Meta&lt;/td&gt;
 &lt;td style="text-align: right"&gt;642.62&lt;/td&gt;
 &lt;td style="text-align: right"&gt;632.53&lt;/td&gt;
 &lt;td style="text-align: right"&gt;35.7&lt;/td&gt;
 &lt;td&gt;Testing the 200-day line. Support at 632–633; watch a reclaim of 664&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Amazon&lt;/td&gt;
 &lt;td style="text-align: right"&gt;248.00&lt;/td&gt;
 &lt;td style="text-align: right"&gt;243.29&lt;/td&gt;
 &lt;td style="text-align: right"&gt;43.2&lt;/td&gt;
 &lt;td&gt;Pullback above the 200-day line. Needs a reclaim of 254–256&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;SpaceX&lt;/td&gt;
 &lt;td style="text-align: right"&gt;147.49&lt;/td&gt;
 &lt;td style="text-align: right"&gt;not available&lt;/td&gt;
 &lt;td style="text-align: right"&gt;46.4&lt;/td&gt;
 &lt;td&gt;Below the recent 20- and 50-day lines. Watch a reclaim of 149–155&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Google&lt;/td&gt;
 &lt;td style="text-align: right"&gt;320.21&lt;/td&gt;
 &lt;td style="text-align: right"&gt;341.31&lt;/td&gt;
 &lt;td style="text-align: right"&gt;31.3&lt;/td&gt;
 &lt;td&gt;Below the main moving averages. Needs a reclaim of 341–343&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Oracle&lt;/td&gt;
 &lt;td style="text-align: right"&gt;126.16&lt;/td&gt;
 &lt;td style="text-align: right"&gt;158.55&lt;/td&gt;
 &lt;td style="text-align: right"&gt;39.1&lt;/td&gt;
 &lt;td&gt;Long-term weakness. Watch a reclaim of 136 and 143&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;The gap is the close divided by the 200-day average, minus one. Microsoft is 524.51 ÷ 436.51 − 1 = &lt;strong&gt;+20.2%&lt;/strong&gt;, the furthest above its 200-day line. Google is −6.2% and Oracle −20.4%.&lt;/p&gt;
&lt;div class="fig-scroll" style="overflow-x:auto;-webkit-overflow-scrolling:touch;margin:1.4em 0"&gt;&lt;img src="https://koreainvestinsights.com/post-assets/hyperscaler-leadership-six-company-comparison-2026-10-25/price-vs-200dma.en.svg" alt="Gap of the close to the 200-day average, percent. Microsoft +20.2, Amazon +1.9, Meta +1.6, Google -6.2, Oracle -20.4. SpaceX has only 93 daily bars, so no 200-day line." loading="lazy" style="display:block;width:100%;min-width:720px;max-width:none;height:auto;margin:0 auto"&gt;&lt;/div&gt;
&lt;p&gt;Microsoft&amp;rsquo;s 63-trading-day return is &lt;strong&gt;+35.1%&lt;/strong&gt;, which is &lt;strong&gt;31.7 percentage points&lt;/strong&gt; ahead of QQQ. SpaceX listed in June as SPCX and has only 93 daily bars, so it has no 200-day line. The watch zones in the table are not order prices.&lt;sup id="fnref:11"&gt;&lt;a href="#fn:11" class="footnote-ref" role="doc-noteref"&gt;11&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;h2 id="the-question-that-could-reverse-the-ranking-is-whether-google-and-meta-turn-technical-progress-into-cash-after-investing"&gt;The question that could reverse the ranking is whether Google and Meta turn technical progress into cash after investing
&lt;/h2&gt;&lt;p&gt;Google and Meta threaten Microsoft on technology and models. But cash after investment is negative for both today. If that gap closes, the ranking can change.&lt;/p&gt;
&lt;p&gt;The next checkpoints are Microsoft&amp;rsquo;s results on October 28 (US time) and Amazon&amp;rsquo;s on October 29.&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Interest note: This analysis was drafted with an OpenAI model, and Anthropic&amp;rsquo;s Claude was used to edit it for publication. Some of the companies compared have business relationships with these two AI companies. We cite public figures only and make no judgment about any single company&amp;rsquo;s exposure.&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Disclaimer: For research and information purposes only. Not investment advice. Names cited are for analytical illustration; readers should perform their own due diligence and consult licensed advisors before any investment decision.&lt;/em&gt;&lt;/p&gt;
&lt;div class="footnotes" role="doc-endnotes"&gt;
&lt;hr&gt;
&lt;ol&gt;
&lt;li id="fn:1"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.microsoft.com/en-us/investor/earnings/fy-2026-q4/press-release-webcast" target="_blank" rel="noopener"
 &gt;Microsoft fiscal 2026 fourth-quarter results&lt;/a&gt;. The $67.0 billion is our simple calculation: $182.9 billion − $115.9 billion.&amp;#160;&lt;a href="#fnref:1" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:2"&gt;
&lt;p&gt;&lt;a class="link" href="https://artificialanalysis.ai/models/releases" target="_blank" rel="noopener"
 &gt;Artificial Analysis model releases&lt;/a&gt;, v4.3.2. Scores use each model&amp;rsquo;s high-performance setting and are not a same-cost test.&amp;#160;&lt;a href="#fnref:2" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:3"&gt;
&lt;p&gt;&lt;a class="link" href="https://arxiv.org/abs/2507.06261" target="_blank" rel="noopener"
 &gt;Gemini technical report&lt;/a&gt;.&amp;#160;&lt;a href="#fnref:3" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:4"&gt;
&lt;p&gt;&lt;a class="link" href="https://arxiv.org/abs/2407.21783" target="_blank" rel="noopener"
 &gt;Llama 3 research paper&lt;/a&gt;.&amp;#160;&lt;a href="#fnref:4" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:5"&gt;
&lt;p&gt;&lt;a class="link" href="https://arxiv.org/abs/2412.08905" target="_blank" rel="noopener"
 &gt;Phi-4 technical report&lt;/a&gt;.&amp;#160;&lt;a href="#fnref:5" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:6"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1652044/000165204426000066/googexhibit991q22026.htm" target="_blank" rel="noopener"
 &gt;Alphabet second-quarter 2026 results&lt;/a&gt;.&amp;#160;&lt;a href="#fnref:6" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:7"&gt;
&lt;p&gt;&lt;a class="link" href="https://investor.atmeta.com/investor-news/press-release-details/2026/Meta-Reports-Third-Quarter-2026-Results/default.aspx" target="_blank" rel="noopener"
 &gt;Meta third-quarter 2026 results&lt;/a&gt;.&amp;#160;&lt;a href="#fnref:7" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:8"&gt;
&lt;p&gt;&lt;a class="link" href="https://ir.aboutamazon.com/news-release/news-release-details/2026/Amazon-com-Announces-Second-Quarter-Results/default.aspx" target="_blank" rel="noopener"
 &gt;Amazon second-quarter 2026 results&lt;/a&gt;.&amp;#160;&lt;a href="#fnref:8" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:9"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1181412/000162828026066482/spaceexplorationtechnologi.htm" target="_blank" rel="noopener"
 &gt;SpaceX third-quarter 2026 results (SEC)&lt;/a&gt;.&amp;#160;&lt;a href="#fnref:9" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:10"&gt;
&lt;p&gt;&lt;a class="link" href="https://investor.oracle.com/investor-news/news-details/2026/Oracle-Announces-Q1-Results-Driven-by-Triple-Digit-Growth-in-Cloud-Infrastructure-Revenues/default.aspx" target="_blank" rel="noopener"
 &gt;Oracle fiscal 2027 first-quarter results&lt;/a&gt;.&amp;#160;&lt;a href="#fnref:10" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:11"&gt;
&lt;p&gt;Our calculation from October 23, 2026 closes and daily bars. The 63-trading-day return and the excess return over QQQ use the same daily data. SpaceX (SPCX) has 93 daily bars, so no 200-day line.&amp;#160;&lt;a href="#fnref:11" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;/ol&gt;
&lt;/div&gt;</description></item><item><title>Korea's AI Data Center Race: What Naver and SK Telecom Have Really Secured</title><link>https://koreainvestinsights.com/post/naver-sk-telecom-ai-data-centers-gpus-funding-2026-10-25/</link><pubDate>Sun, 25 Oct 2026 15:30:00 +0900</pubDate><guid>https://koreainvestinsights.com/post/naver-sk-telecom-ai-data-centers-gpus-funding-2026-10-25/</guid><description>&lt;img src="https://koreainvestinsights.com/" alt="Featured image of post Korea's AI Data Center Race: What Naver and SK Telecom Have Really Secured" /&gt;&lt;p&gt;Korea wants to be Asia&amp;rsquo;s hub for AI computing, and the announcements have been huge. Nvidia said Naver plans to deploy more than 60,000 of its chips. Brookfield says it will put up to $9 billion into Naver&amp;rsquo;s first AI factory. SK Telecom set up a new company to build 5 gigawatts of data centers, and that company&amp;rsquo;s chief executive put the cost of the first phase at about $350 billion.&lt;/p&gt;
&lt;p&gt;Big numbers are easy to announce and hard to finish. A plan is not an order. An order is not a delivery. And money promised is not money in the bank. This post sorts out what each company has actually secured, using filings and company releases up to October 25, 2026.&lt;/p&gt;
&lt;p&gt;The short answer: Naver is furthest along in turning AI computing into a business it can sell. SK Horizon, the data center unit SK Telecom is splitting off, has the most visible cash today because it already runs paying data centers. SK Hyper is a long-dated development project. Its value depends on customers, power and outside money that are not yet signed.&lt;/p&gt;
&lt;h2 id="five-things-to-take-away"&gt;Five things to take away
&lt;/h2&gt;&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;These are three different businesses, not two companies.&lt;/strong&gt; Naver&amp;rsquo;s AI Factory sells computing. SK Horizon rents out buildings, power and connections. SK Hyper develops new gigawatt-scale sites. Their numbers should not be compared one to one or added up.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Naver&amp;rsquo;s money is the most concrete, but it is conditional.&lt;/strong&gt; Nvidia agreed to buy $1 billion of new Naver shares, on condition that Naver secures at least $9 billion of computing infrastructure. Brookfield says it is committed to investing up to $9 billion in the first 200MW phase. Neither the payment nor the final financing close was confirmed as of Oct 25.&lt;sup id="fnref:1"&gt;&lt;a href="#fn:1" class="footnote-ref" role="doc-noteref"&gt;1&lt;/a&gt;&lt;/sup&gt;&lt;sup id="fnref:2"&gt;&lt;a href="#fn:2" class="footnote-ref" role="doc-noteref"&gt;2&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;SK Horizon&amp;rsquo;s ₩3.08 trillion deal is mostly a sale.&lt;/strong&gt; About ₩1.88 trillion goes to SK Telecom for existing shares. About ₩1.2 trillion is new money for Horizon itself.&lt;sup id="fnref:3"&gt;&lt;a href="#fn:3" class="footnote-ref" role="doc-noteref"&gt;3&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;&amp;ldquo;Securing GPUs&amp;rdquo; usually means a plan.&lt;/strong&gt; Most Korean GPU counts are deployment plans or rough conversions, not chips that are delivered, running and billed.&lt;sup id="fnref:4"&gt;&lt;a href="#fn:4" class="footnote-ref" role="doc-noteref"&gt;4&lt;/a&gt;&lt;/sup&gt;&lt;sup id="fnref:5"&gt;&lt;a href="#fn:5" class="footnote-ref" role="doc-noteref"&gt;5&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Revenue is not payback.&lt;/strong&gt; Naver expects AI Factory revenue from the first half of 2027. SK Hyper&amp;rsquo;s first 5GW opens in phases from 2029. None of the new projects has published a breakeven date, and overseas peers show that fast-growing AI clouds can still burn cash.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;Data are as of Oct 25, 2026. Calculations marked as ours are illustrations, not company forecasts.&lt;/p&gt;
&lt;h2 id="1-three-businesses-three-stages"&gt;1. Three businesses, three stages
&lt;/h2&gt;&lt;p&gt;Start by separating the companies from the projects.&lt;/p&gt;
&lt;div class="fig-scroll" style="overflow-x:auto;-webkit-overflow-scrolling:touch;margin:1.4em 0"&gt;&lt;img src="https://koreainvestinsights.com/post-assets/naver-sk-telecom-ai-data-centers-gpus-funding-2026-10-25/three-businesses.en.svg" alt="Three different businesses at three different stages. NAVER AI Factory, the most concrete new compute business: sells AI computing from its Sejong data center; 55MW live from the first half of 2027 with revenue expected then, 200MW targeted for 2028; money is a 1 billion dollar Nvidia share deal that is conditional plus up to 9 billion dollars of Brookfield-led project financing; next proof is the Nvidia payment by Oct 30, the financing close and first customer contracts. SK Horizon, the most visible cash today: SK Broadband's existing data centers and subsea cables paired with infrastructure-fund money; already earns money, set-up and deal close targeted for the first quarter of 2027; about 3.08 trillion won deal, 1.88 trillion to SKT for existing shares and about 1.2 trillion of new equity. SK Hyper, a long-dated development option: secures land, substations and customers; first 5GW opens in phases from 2029, expansion toward 15GW planned around 2035; SKT can invest up to 750 billion won by 2030, first 5GW estimated at about 350 billion dollars in total" loading="lazy" style="display:block;width:100%;min-width:720px;max-width:none;height:auto;margin:0 auto"&gt;&lt;/div&gt;
&lt;p&gt;&lt;strong&gt;Naver&amp;rsquo;s AI Factory sells computing.&lt;/strong&gt; &amp;ldquo;AI factory&amp;rdquo; is Nvidia&amp;rsquo;s name for a data center built to train and run AI models. Naver plans to run one at its Gak Sejong data center, linked to its own cloud, its HyperCLOVA X models and its search and shopping services. It is the first phase of a 1GW project with Nvidia. The first 55MW (megawatts of power capacity) is due to start in the first half of 2027, and 200MW is targeted for 2028.&lt;sup id="fnref:6"&gt;&lt;a href="#fn:6" class="footnote-ref" role="doc-noteref"&gt;6&lt;/a&gt;&lt;/sup&gt;&lt;sup id="fnref1:2"&gt;&lt;a href="#fn:2" class="footnote-ref" role="doc-noteref"&gt;2&lt;/a&gt;&lt;/sup&gt; Naver expects related revenue from the first half of 2027.&lt;sup id="fnref:7"&gt;&lt;a href="#fn:7" class="footnote-ref" role="doc-noteref"&gt;7&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;SK Horizon rents out buildings, power and connections.&lt;/strong&gt; SK Telecom is splitting SK Broadband&amp;rsquo;s data center and subsea cable business into a new company. Infrastructure investor KKR and a group led by IMM Investment are buying in. The new company covers 8 existing data centers plus sites under construction, 318MW in total.&lt;sup id="fnref:8"&gt;&lt;a href="#fn:8" class="footnote-ref" role="doc-noteref"&gt;8&lt;/a&gt;&lt;/sup&gt; Its data centers already earn money today.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;SK Hyper develops new sites.&lt;/strong&gt; SK Telecom set up SK Hyper in July as a wholly owned subsidiary. Its job is to secure land, substations and customers for gigawatt-scale campuses. The first 5GW is due to open in phases from 2029, with expansion toward 15GW planned around 2035.&lt;sup id="fnref:9"&gt;&lt;a href="#fn:9" class="footnote-ref" role="doc-noteref"&gt;9&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;A fourth item is easy to mix up: SK Telecom&amp;rsquo;s own AI cloud with Nvidia, a plan of up to 2GW with the first site targeted for 2027.&lt;sup id="fnref:10"&gt;&lt;a href="#fn:10" class="footnote-ref" role="doc-noteref"&gt;10&lt;/a&gt;&lt;/sup&gt; SK Telecom has said that how much of its partner projects falls inside SK Hyper depends on each contract, so we do not add these numbers together.&lt;sup id="fnref:11"&gt;&lt;a href="#fn:11" class="footnote-ref" role="doc-noteref"&gt;11&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;Why does this matter? A company that sells computing earns more per megawatt, but it carries the chip risk. A company that rents out buildings earns less but steadier money, and its customers carry the chip risk. Valuing the two the same way is the most common mistake in this debate.&lt;/p&gt;
&lt;h2 id="2-when-does-the-money-start"&gt;2. When does the money start?
&lt;/h2&gt;&lt;div class="fig-scroll" style="overflow-x:auto;-webkit-overflow-scrolling:touch;margin:1.4em 0"&gt;&lt;img src="https://koreainvestinsights.com/post-assets/naver-sk-telecom-ai-data-centers-gpus-funding-2026-10-25/timeline.en.svg" alt="When the money is supposed to start. NAVER: Oct 30, 2026 is the Nvidia 1 billion dollar payment deadline; first half of 2027, 55MW live and revenue expected; 2028, 200MW target. SKT and Horizon: first quarter of 2027, Horizon set up and share sale due Mar 4; 2027, first Nvidia AI Factory and AWS Ulsan opens. SK Hyper: 2029, 5GW begins in phases; 2035, expansion toward 15GW. Revenue start, operating profit and payback are three different dates, and none of the new projects has a published breakeven date" loading="lazy" style="display:block;width:100%;min-width:720px;max-width:none;height:auto;margin:0 auto"&gt;&lt;/div&gt;
&lt;p&gt;Three dates often get mixed up: the day revenue starts, the day the business makes an operating profit, and the day the investment is paid back. For the new projects, only the first has been announced.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Business&lt;/th&gt;
 &lt;th&gt;When revenue starts or is due&lt;/th&gt;
 &lt;th&gt;Profit and payback&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Naver&amp;rsquo;s existing cloud and AI&lt;/td&gt;
 &lt;td&gt;Running today&lt;/td&gt;
 &lt;td&gt;Mixed into existing results; must be separated from the AI Factory&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Naver AI Factory&lt;/td&gt;
 &lt;td&gt;55MW from 1H 2027; 200MW target for 2028&lt;/td&gt;
 &lt;td&gt;No breakeven date published&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;SK Telecom&amp;rsquo;s AI data centers today&lt;/td&gt;
 &lt;td&gt;₩136.2bn of revenue in Q2 2026, up 92.5%&lt;/td&gt;
 &lt;td&gt;Profit per GPU cluster not disclosed&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;SK Horizon&lt;/td&gt;
 &lt;td&gt;Earns money today; new company due in 1Q 2027&lt;/td&gt;
 &lt;td&gt;A spin-off does not create new demand by itself&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;AWS Ulsan and SKT&amp;rsquo;s first Nvidia AI factory&lt;/td&gt;
 &lt;td&gt;Both targeted for 2027&lt;/td&gt;
 &lt;td&gt;Need commercial operation and customer billing&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;SK Hyper&lt;/td&gt;
 &lt;td&gt;First 5GW from 2029; CEO says revenue is possible as early as 2027&lt;/td&gt;
 &lt;td&gt;Type and size of any 2027 revenue not disclosed&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Sources: company releases, SK Telecom&amp;rsquo;s second-quarter results and press reports of the SK Hyper interview.&lt;sup id="fnref1:7"&gt;&lt;a href="#fn:7" class="footnote-ref" role="doc-noteref"&gt;7&lt;/a&gt;&lt;/sup&gt;&lt;sup id="fnref:12"&gt;&lt;a href="#fn:12" class="footnote-ref" role="doc-noteref"&gt;12&lt;/a&gt;&lt;/sup&gt;&lt;sup id="fnref:13"&gt;&lt;a href="#fn:13" class="footnote-ref" role="doc-noteref"&gt;13&lt;/a&gt;&lt;/sup&gt;&lt;sup id="fnref:14"&gt;&lt;a href="#fn:14" class="footnote-ref" role="doc-noteref"&gt;14&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;One accounting detail matters here. In the second quarter, Naver lengthened the useful life of its GPU and CPU servers from five to six years. That defers about ₩100 billion of cost a year.&lt;sup id="fnref:15"&gt;&lt;a href="#fn:15" class="footnote-ref" role="doc-noteref"&gt;15&lt;/a&gt;&lt;/sup&gt; It makes reported profit look better sooner. It does not change how much cash the chips cost, or how fast they lose value in the rental market.&lt;/p&gt;
&lt;h2 id="3-navers-edge-one-company-from-the-building-to-the-app"&gt;3. Naver&amp;rsquo;s edge: one company from the building to the app
&lt;/h2&gt;&lt;p&gt;Naver runs the data center, the GPU clusters, the cloud platform, its own AI models and the services that use them. That lets it sell more than raw chip time. It can help a customer build, train and deploy a model. Its own search and shopping demand can also fill early capacity.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Strength&lt;/th&gt;
 &lt;th&gt;Naver&amp;rsquo;s advantage&lt;/th&gt;
 &lt;th&gt;What still needs proof&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Customers&lt;/td&gt;
 &lt;td&gt;Its own services, Korean companies and government, overseas sovereign AI deals&lt;/td&gt;
 &lt;td&gt;Whether Korean demand alone can fill 200MW and more; few overseas contracts are public&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Product&lt;/td&gt;
 &lt;td&gt;Links the building to models and apps&lt;/td&gt;
 &lt;td&gt;Performance and support against global GPU clouds on equal terms&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Supply and capital&lt;/td&gt;
 &lt;td&gt;Nvidia as a future shareholder, Brookfield as capital partner&lt;/td&gt;
 &lt;td&gt;Final financing terms, lease payments and guarantees&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Profit path&lt;/td&gt;
 &lt;td&gt;Computing revenue plus better own services&lt;/td&gt;
 &lt;td&gt;Separating GPU profit from the rest of Naver&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;There is a catch in the second column. When Naver&amp;rsquo;s own services use the chips, that is not outside revenue. Its value shows up as lower computing bills, better ad conversion or better products, and it has to be measured that way.&lt;/p&gt;
&lt;p&gt;On customers, Naver&amp;rsquo;s chief executive said the company would try to secure customers for the first capacity in advance.&lt;sup id="fnref2:7"&gt;&lt;a href="#fn:7" class="footnote-ref" role="doc-noteref"&gt;7&lt;/a&gt;&lt;/sup&gt; No customer names, contracted megawatts, prices, terms or minimum-use commitments have been published. That is a good sign of intent, but not yet enough to calculate the profit on any contract.&lt;/p&gt;
&lt;h2 id="4-sks-edge-land-power-and-connections"&gt;4. SK&amp;rsquo;s edge: land, power and connections
&lt;/h2&gt;&lt;p&gt;SK Telecom is building two layers. SK Horizon takes the existing data centers and subsea cables and pairs them with infrastructure money. SK Hyper takes on new land, power and customers. This split lets steady, running assets raise infrastructure capital, while each new project raises its own money.&lt;/p&gt;
&lt;p&gt;SK&amp;rsquo;s network, subsea cables, enterprise sales and data center experience help bring overseas customers to Korean sites. The group&amp;rsquo;s construction, energy and memory businesses can also help with design and purchasing. But being in the same group does not automatically give a project a power connection, cheap electricity, first call on high-bandwidth memory or a cheap guarantee. Each of those needs its own contract.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Partner&lt;/th&gt;
 &lt;th&gt;Role&lt;/th&gt;
 &lt;th&gt;What it is not&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;KKR&lt;/td&gt;
 &lt;td&gt;Equity investor in SK Horizon&lt;/td&gt;
 &lt;td&gt;Not committed money for SK Hyper&amp;rsquo;s 5GW&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;IMM-led group&lt;/td&gt;
 &lt;td&gt;Co-investor in SK Horizon, 20% final stake&lt;/td&gt;
 &lt;td&gt;Not a guarantor of customer revenue&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Nvidia&lt;/td&gt;
 &lt;td&gt;AI cloud design, chips and memory partnership&lt;/td&gt;
 &lt;td&gt;No disclosed equity investment in SKT or SK Hyper; no promise to buy unused capacity&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;AWS&lt;/td&gt;
 &lt;td&gt;Partner for the Ulsan AI Zone, due to run by 2027 under a 15-year agreement&lt;/td&gt;
 &lt;td&gt;Leased megawatts and minimum use not disclosed&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Lambda&lt;/td&gt;
 &lt;td&gt;GPU cloud partner; SKT invested in it&lt;/td&gt;
 &lt;td&gt;Money flowing out of SKT, not into SK Hyper&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Penguin Solutions&lt;/td&gt;
 &lt;td&gt;AI cluster design and operations; SKT invested $200m&lt;/td&gt;
 &lt;td&gt;Separate from any outside funding for SK Hyper&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Sources: SK Telecom, Nvidia and Penguin Solutions releases.&lt;sup id="fnref1:13"&gt;&lt;a href="#fn:13" class="footnote-ref" role="doc-noteref"&gt;13&lt;/a&gt;&lt;/sup&gt;&lt;sup id="fnref:16"&gt;&lt;a href="#fn:16" class="footnote-ref" role="doc-noteref"&gt;16&lt;/a&gt;&lt;/sup&gt;&lt;sup id="fnref:17"&gt;&lt;a href="#fn:17" class="footnote-ref" role="doc-noteref"&gt;17&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;SK Hyper&amp;rsquo;s weakness is not that its goals are big. It is that the customer, money and power contracts behind them are not yet public. SK says it will build in steps as demand appears. That lowers the risk of spending everything at once. It also means revenue waits until customers sign.&lt;/p&gt;
&lt;h2 id="5-securing-gpus-has-five-steps"&gt;5. &amp;ldquo;Securing GPUs&amp;rdquo; has five steps
&lt;/h2&gt;&lt;div class="fig-scroll" style="overflow-x:auto;-webkit-overflow-scrolling:touch;margin:1.4em 0"&gt;&lt;img src="https://koreainvestinsights.com/post-assets/naver-sk-telecom-ai-data-centers-gpus-funding-2026-10-25/gpu-ladder.en.svg" alt="Securing GPUs has five steps: plan announced, binding order, delivered and owned, installed, powered and tested, customers billed. NAVER 60,000 plus GPUs, an Nvidia deployment plan from October 2025, is confirmed only at step one. NAVER 200MW, about 100,000 GPUs, a rough conversion of a 2026 expansion plan, step one. SK Group 50,000 plus GPUs, a group-wide plan not allocated to SKT or SK Hyper, step one. SKT up to 2GW AI Factory, a letter of intent with Nvidia with the first site in 2027, step one. SKT GPU cloud on H100 chips, a paid service since 2025, reaches step five. The two NAVER numbers should not be added together and neither is a delivery count" loading="lazy" style="display:block;width:100%;min-width:720px;max-width:none;height:auto;margin:0 auto"&gt;&lt;/div&gt;
&lt;p&gt;To know what &amp;ldquo;securing GPUs&amp;rdquo; really means, check five steps in order. A plan is announced. A binding order is placed. The chips are delivered and owned. They are installed with power and networking and pass testing. Customers start paying. A headline number usually tells you only the first step.&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Naver, 60,000+ GPUs.&lt;/strong&gt; In October 2025, Nvidia said Naver Cloud plans to deploy more than 60,000 of its GPUs, part of more than 260,000 planned across Korea.&lt;sup id="fnref1:4"&gt;&lt;a href="#fn:4" class="footnote-ref" role="doc-noteref"&gt;4&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Naver, 200MW, about 100,000 GPUs.&lt;/strong&gt; In July 2026, Nvidia described Naver&amp;rsquo;s 200MW expansion as roughly 100,000 GPUs on its next-generation Vera Rubin platform.&lt;sup id="fnref1:5"&gt;&lt;a href="#fn:5" class="footnote-ref" role="doc-noteref"&gt;5&lt;/a&gt;&lt;/sup&gt; This is a rough conversion of a plan. It should not be added to the 60,000.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;SK Group, 50,000+ GPUs.&lt;/strong&gt; This is a group-wide plan from October 2025. There is no basis for treating it as chips owned by SK Telecom or SK Hyper.&lt;sup id="fnref2:4"&gt;&lt;a href="#fn:4" class="footnote-ref" role="doc-noteref"&gt;4&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;SK Telecom, up to 2GW.&lt;/strong&gt; A letter of intent with Nvidia for AI factories, part of a wider SK–Nvidia partnership described as worth more than $500 billion, which also covers memory.&lt;sup id="fnref1:10"&gt;&lt;a href="#fn:10" class="footnote-ref" role="doc-noteref"&gt;10&lt;/a&gt;&lt;/sup&gt; Chip counts, prices and delivery dates are not public.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;SK Telecom&amp;rsquo;s existing GPU cloud.&lt;/strong&gt; A paid service built with Lambda has been running since 2025.&lt;sup id="fnref1:16"&gt;&lt;a href="#fn:16" class="footnote-ref" role="doc-noteref"&gt;16&lt;/a&gt;&lt;/sup&gt; It is the only item here that clearly reaches the billing step, and it is small.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;Locking in chips early helps when a chip generation is in short supply. Customers buy a start date as much as a machine. But if power, cooling or networking is late, the chips sit idle and the owner pays for them anyway. And each new Nvidia generation pushes down the hourly rental price of the old one. So the useful comparison is not chip counts. It is the chip generation, the date the cluster can serve customers, how many hours customers actually pay for, and the price once the next generation arrives.&lt;/p&gt;
&lt;h2 id="6-nvidias-role-supplier-shareholder-or-customer"&gt;6. Nvidia&amp;rsquo;s role: supplier, shareholder or customer?
&lt;/h2&gt;&lt;p&gt;Nvidia can play three roles. Each needs its own contract.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Role&lt;/th&gt;
 &lt;th&gt;Naver&lt;/th&gt;
 &lt;th&gt;SK Telecom and SK Group&lt;/th&gt;
 &lt;th&gt;Benchmark abroad&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Technology and supply&lt;/td&gt;
 &lt;td&gt;Nvidia data center designs, Blackwell and Vera Rubin chips&lt;/td&gt;
 &lt;td&gt;Same, plus memory co-development with SK hynix&lt;/td&gt;
 &lt;td&gt;Shared technology helps but gives no exclusivity&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Shareholder&lt;/td&gt;
 &lt;td&gt;$1bn share purchase agreement&lt;/td&gt;
 &lt;td&gt;No equity investment found&lt;/td&gt;
 &lt;td&gt;Nvidia has invested in CoreWeave and Nebius&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Backstop customer&lt;/td&gt;
 &lt;td&gt;Not found&lt;/td&gt;
 &lt;td&gt;Not found&lt;/td&gt;
 &lt;td&gt;CoreWeave: Nvidia must buy unused capacity&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;The backstop matters most. In September 2025, CoreWeave disclosed an order from Nvidia with an initial value of $6.3 billion. Under it, Nvidia is obliged to buy cloud capacity that CoreWeave&amp;rsquo;s other customers leave unused, through April 2032, subject to delivery and termination terms.&lt;sup id="fnref:18"&gt;&lt;a href="#fn:18" class="footnote-ref" role="doc-noteref"&gt;18&lt;/a&gt;&lt;/sup&gt; That kind of contract moves part of the demand risk onto the supplier. We found nothing like it for Naver or SK Telecom.&lt;/p&gt;
&lt;p&gt;Naver&amp;rsquo;s case is still stronger than an ordinary partnership. A shareholder has a stake in the outcome. But Nvidia also sells the chips. When investment money and chip purchases are linked, Nvidia&amp;rsquo;s revenue and Naver&amp;rsquo;s eventual return are different things. Having Nvidia as a shareholder does not guarantee customers or returns.&lt;/p&gt;
&lt;h2 id="7-follow-the-money"&gt;7. Follow the money
&lt;/h2&gt;&lt;div class="fig-scroll" style="overflow-x:auto;-webkit-overflow-scrolling:touch;margin:1.4em 0"&gt;&lt;img src="https://koreainvestinsights.com/post-assets/naver-sk-telecom-ai-data-centers-gpus-funding-2026-10-25/money-map.en.svg" alt="Follow the money: who receives it and in what form. NAVER: 1 billion dollars, Nvidia buys new NAVER shares, signed and board-approved, payment not yet confirmed, latest deadline Oct 30; up to 9 billion dollars, Brookfield-led money for a project company, non-binding term sheet in July, called a commitment in September, final close not confirmed; 400 billion won policy loan, 5 years, rate in the 3 percent range, approved in April 2026, amount drawn not disclosed. SK Horizon: 1.88 trillion won to SKT for selling existing shares, scheduled for Mar 4, 2027 once closing conditions are met; about 1.2 trillion won of new equity to Horizon, terms can change; final stakes SKT 51, KKR 29, IMM group 20. SK Hyper: up to 750 billion won from SKT through 2030 for land and substations, a board limit, not money already paid in; about 350 billion dollars, estimated total cost of the first 5GW, a CEO estimate via press with nothing signed; about 0.16 percent, 750 billion won as a share of 350 billion dollars, a scale comparison only, not a funding gap" loading="lazy" style="display:block;width:100%;min-width:720px;max-width:none;height:auto;margin:0 auto"&gt;&lt;/div&gt;
&lt;p&gt;The size of a funding headline matters less than who receives the money, in what form and on what conditions.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Naver.&lt;/strong&gt; Nvidia&amp;rsquo;s $1 billion share purchase comes with a condition: Naver must secure at least $9 billion of computing infrastructure, separate from Nvidia&amp;rsquo;s own money. So the $1 billion and the $9 billion should not be added up as cash in hand. October 30 is the latest payment date in the contract. Payment can come earlier, and the date can be extended, so it needs its own filing.&lt;sup id="fnref1:1"&gt;&lt;a href="#fn:1" class="footnote-ref" role="doc-noteref"&gt;1&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;Brookfield&amp;rsquo;s money moved from a non-binding term sheet in July to a commitment to invest up to $9 billion in September.&lt;sup id="fnref:19"&gt;&lt;a href="#fn:19" class="footnote-ref" role="doc-noteref"&gt;19&lt;/a&gt;&lt;/sup&gt;&lt;sup id="fnref2:2"&gt;&lt;a href="#fn:2" class="footnote-ref" role="doc-noteref"&gt;2&lt;/a&gt;&lt;/sup&gt; That is progress. But a press release does not confirm that the bank loans and project equity have closed. Korean press reported that KB Kookmin, Shinhan, Hana, Woori and NH Investment &amp;amp; Securities are reviewing loans, with Naver to lease the facility long-term.&lt;sup id="fnref:20"&gt;&lt;a href="#fn:20" class="footnote-ref" role="doc-noteref"&gt;20&lt;/a&gt;&lt;/sup&gt; Those institutions are reviewing; they are not signed lenders.&lt;/p&gt;
&lt;p&gt;If a separate project company owns the assets, Naver spends less cash upfront. In exchange, long-term lease payments, minimum-use obligations, guarantees and profit sharing decide the real burden. Until those contracts are public, we cannot say the debt burden has gone away. And closing the 200MW financing would not mean the full 1GW is funded.&lt;/p&gt;
&lt;p&gt;Naver itself is not short of cash, but not all of it is free to spend. At end-June it had ₩8.40 trillion of cash and short-term financial assets on a consolidated basis. Its own net cash figure, which takes out Naver Financial&amp;rsquo;s cash and debt, was ₩1.83 trillion.&lt;sup id="fnref1:15"&gt;&lt;a href="#fn:15" class="footnote-ref" role="doc-noteref"&gt;15&lt;/a&gt;&lt;/sup&gt; A ₩400 billion policy loan (five years, with interest in the 3% range) covers 43% of a ₩922 billion Sejong expansion and GPU project. The remaining ₩522 billion equals about 29% of that net cash.&lt;sup id="fnref:21"&gt;&lt;a href="#fn:21" class="footnote-ref" role="doc-noteref"&gt;21&lt;/a&gt;&lt;/sup&gt; This is a size comparison, not a verdict on how Naver will pay.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;SK Horizon.&lt;/strong&gt; The deal totals about ₩3.08 trillion. About ₩1.88 trillion (61%) is SK Telecom selling existing shares, and about ₩1.2 trillion (39%) is new equity for Horizon.&lt;sup id="fnref1:3"&gt;&lt;a href="#fn:3" class="footnote-ref" role="doc-noteref"&gt;3&lt;/a&gt;&lt;/sup&gt; Final stakes are SK Telecom 51%, KKR 29% and the IMM-led group 20%. For SK Telecom as a group, the whole deal brings in outside money. But only the ₩1.2 trillion lands in Horizon for growth. Horizon&amp;rsquo;s new capex plans, debt and investor rights are not fully public, so we cannot yet say whether that is enough.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;SK Hyper.&lt;/strong&gt; Two numbers are often compared. SK Telecom&amp;rsquo;s board approved up to ₩750 billion of investment through 2030 for land and substations.&lt;sup id="fnref1:9"&gt;&lt;a href="#fn:9" class="footnote-ref" role="doc-noteref"&gt;9&lt;/a&gt;&lt;/sup&gt; SK Hyper&amp;rsquo;s chief executive estimated, in an interview reported by Korean press, that the first 5GW alone would need about $350 billion.&lt;sup id="fnref1:14"&gt;&lt;a href="#fn:14" class="footnote-ref" role="doc-noteref"&gt;14&lt;/a&gt;&lt;/sup&gt; At ₩1,325 per dollar, ₩750 billion is about 0.16% of $350 billion. But one is seed money for site development and the other is the full cost of a multi-year build, including equipment. The ratio shows scale, not a funding gap. The real gap can only be measured project by project, once customers and investors sign.&lt;/p&gt;
&lt;h2 id="8-lessons-from-abroad"&gt;8. Lessons from abroad
&lt;/h2&gt;&lt;div class="fig-scroll" style="overflow-x:auto;-webkit-overflow-scrolling:touch;margin:1.4em 0"&gt;&lt;img src="https://koreainvestinsights.com/post-assets/naver-sk-telecom-ai-data-centers-gpus-funding-2026-10-25/lessons-abroad.en.svg" alt="Lessons from abroad: revenue growth does not fund itself. Cash in and out in the first half of 2026, in billions of dollars: CoreWeave operating cash flow 3.66 against 14.12 spent on equipment and property, a gap of minus 10.45; Nebius operating cash flow 4.50 against 8.13, a gap of minus 3.63, and 4.40 of Nebius's 4.50 came from customers paying in advance. Core Scientific: 590MW leased to CoreWeave, of which 395MW is billing and 195MW is still being built or tested, 67 percent billing. Count megawatts that bill, not megawatts that are leased" loading="lazy" style="display:block;width:100%;min-width:720px;max-width:none;height:auto;margin:0 auto"&gt;&lt;/div&gt;
&lt;p&gt;Overseas companies that already run this business show what to watch.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Company&lt;/th&gt;
 &lt;th&gt;Comparable to&lt;/th&gt;
 &lt;th&gt;What the numbers show&lt;/th&gt;
 &lt;th&gt;Lesson&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;CoreWeave&lt;/td&gt;
 &lt;td&gt;Naver and SKT GPU clouds&lt;/td&gt;
 &lt;td&gt;First-half 2026 operating cash flow of $3.66bn against $14.12bn spent on equipment and property&lt;/td&gt;
 &lt;td&gt;Fast revenue growth does not remove the need for outside money&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Nebius&lt;/td&gt;
 &lt;td&gt;GPU cloud growth and funding&lt;/td&gt;
 &lt;td&gt;Operating cash flow of $4.50bn, of which $4.40bn came from customers paying in advance; capex of $8.13bn&lt;/td&gt;
 &lt;td&gt;Prepayments help fund growth but come with service obligations&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Core Scientific&lt;/td&gt;
 &lt;td&gt;SK Horizon and SK Hyper as site providers&lt;/td&gt;
 &lt;td&gt;590MW leased to CoreWeave, of which 395MW is billing; 195MW still being built or tested&lt;/td&gt;
 &lt;td&gt;Count megawatts that bill, not megawatts that are signed&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Singtel&amp;rsquo;s Nxera&lt;/td&gt;
 &lt;td&gt;SK Horizon&lt;/td&gt;
 &lt;td&gt;Telecom data center platform with outside capital; an investor paid S$337m for preference shares in Nxera&lt;/td&gt;
 &lt;td&gt;Closest model to Horizon: keep site money and GPU money separate&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Sources: company filings and annual report.&lt;sup id="fnref:22"&gt;&lt;a href="#fn:22" class="footnote-ref" role="doc-noteref"&gt;22&lt;/a&gt;&lt;/sup&gt;&lt;sup id="fnref:23"&gt;&lt;a href="#fn:23" class="footnote-ref" role="doc-noteref"&gt;23&lt;/a&gt;&lt;/sup&gt;&lt;sup id="fnref:24"&gt;&lt;a href="#fn:24" class="footnote-ref" role="doc-noteref"&gt;24&lt;/a&gt;&lt;/sup&gt;&lt;sup id="fnref:25"&gt;&lt;a href="#fn:25" class="footnote-ref" role="doc-noteref"&gt;25&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;Nebius also shows how funding works in practice. On Oct 22 it announced a $3.5 billion financing facility arranged by Citi, with a 364-day term and interest of SOFR plus 2.75%, rising to at most 3.50%.&lt;sup id="fnref:26"&gt;&lt;a href="#fn:26" class="footnote-ref" role="doc-noteref"&gt;26&lt;/a&gt;&lt;/sup&gt; A facility is not cash until it is drawn. And funding chips and buildings that last years with one-year money only works if longer-term financing follows.&lt;/p&gt;
&lt;p&gt;Korean operators may have an edge over US &amp;ldquo;neoclouds&amp;rdquo; (GPU-rental specialists like CoreWeave and Nebius) in local sovereign demand, existing telecom networks and their own services. They have not yet shown an edge in signed contracts with very large global customers, in dollar funding, or in running huge GPU fleets. We have no like-for-like data on power costs, chip prices, utilization or contract prices, so we do not rank their cost advantage.&lt;/p&gt;
&lt;h2 id="9-the-payback-test"&gt;9. The payback test
&lt;/h2&gt;&lt;p&gt;Raising construction money is only half the job. The project also has to earn it back.&lt;/p&gt;
&lt;div class="fig-scroll" style="overflow-x:auto;-webkit-overflow-scrolling:touch;margin:1.4em 0"&gt;&lt;img src="https://koreainvestinsights.com/post-assets/naver-sk-telecom-ai-data-centers-gpus-funding-2026-10-25/payback.en.svg" alt="How fast the chips wear out decides the yearly bill. Yearly cash a hypothetical 10 billion dollar AI site must earn to pay back its cost: 3.92 billion dollars with a 3-year chip life, 3.23 billion with 4 years, 2.54 billion with 6 years. Two fewer years of chip life, from 6 to 4, adds about 0.69 billion dollars a year. Assumes 8 billion dollars of GPUs and servers at a 10 percent required return, 2 billion of buildings at 8 percent over 20 years, 0.5 billion a year of cash running costs and no resale value" loading="lazy" style="display:block;width:100%;min-width:720px;max-width:none;height:auto;margin:0 auto"&gt;&lt;/div&gt;
&lt;p&gt;To show why chip life matters so much, take a hypothetical $10 billion AI site. It has $8 billion of GPUs and servers that must earn 10% a year, $2 billion of buildings that must earn 8% over 20 years, and $0.5 billion a year of cash running costs. If the chips stay useful for six years, the site must earn about $2.54 billion a year to pay back its cost plus the required return. At four years, it needs $3.23 billion. At three years, $3.92 billion.&lt;sup id="fnref:27"&gt;&lt;a href="#fn:27" class="footnote-ref" role="doc-noteref"&gt;27&lt;/a&gt;&lt;/sup&gt; Cutting chip life from six years to four alone adds about $0.69 billion a year.&lt;/p&gt;
&lt;p&gt;This is not a forecast for Naver or SK. Naver may lease rather than own, and the real split of costs depends on its contracts. The point is who carries this risk: the building owner, the lenders or the computing seller. For shareholders, that question matters more than the headline investment.&lt;/p&gt;
&lt;p&gt;For funding, a simple 12-month coverage test helps:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Money available:&lt;/strong&gt; cash on hand, signed but undrawn facilities whose conditions are met, customer prepayments that can be used, and a cautious estimate of operating cash inflow.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Money needed:&lt;/strong&gt; committed construction and chip payments, interest, principal and lease payments, working capital, and a reserve for delays.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;Divide the first by the second. Do not count conditional money, or the same money twice. Include parent guarantees and minimum lease payments if they exist. Neither company has published enough to calculate this for the new projects, and we do not treat missing items as zero.&lt;/p&gt;
&lt;h2 id="10-what-would-change-our-view"&gt;10. What would change our view
&lt;/h2&gt;&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Business&lt;/th&gt;
 &lt;th&gt;Our read today&lt;/th&gt;
 &lt;th&gt;Good signs&lt;/th&gt;
 &lt;th&gt;Warning signs&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Naver&lt;/td&gt;
 &lt;td&gt;The most concrete path from customers and capital to new computing revenue&lt;/td&gt;
 &lt;td&gt;Nvidia payment, Brookfield financing close, minimum-use contracts with first customers, 55MW passing tests and billing&lt;/td&gt;
 &lt;td&gt;Funding conditions missed, start-up delays, lease terms longer than customer contracts&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;SK Horizon&lt;/td&gt;
 &lt;td&gt;The clearest structure for growing running assets&lt;/td&gt;
 &lt;td&gt;Deal close, new money received, more megawatts billing at new sites&lt;/td&gt;
 &lt;td&gt;Deal delays, larger capital obligations, customer concentration, weaker power-cost pass-through&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;SK Hyper&lt;/td&gt;
 &lt;td&gt;A long-dated development option; full funding not yet judged&lt;/td&gt;
 &lt;td&gt;A named anchor customer, signed outside investors, a grid connection and money being spent&lt;/td&gt;
 &lt;td&gt;Rising development costs with no customers, larger parent guarantees, land or grid delays&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;The strongest argument against our caution: computing is still scarce worldwide, and outside capital is eager. That can speed things up. Naver says it is lining up customers early, and SK says it will invest step by step with demand. Unpublished contracts are not proof of failure.&lt;/p&gt;
&lt;p&gt;But a supplier&amp;rsquo;s partnership or investment is different evidence from an independent customer&amp;rsquo;s ability to pay. Even long contracts carry acceptance tests, termination rights for delays, credit risk and power-cost terms. The key question is not how big the revenue is. It is who carries the cost of idle GPUs and construction overruns.&lt;/p&gt;
&lt;h2 id="11-what-this-means-for-memory-chips"&gt;11. What this means for memory chips
&lt;/h2&gt;&lt;p&gt;Korea&amp;rsquo;s AI build-out adds a new source of demand for memory, including high-bandwidth memory (HBM) from SK hynix and Samsung. But plans should not be turned into this year&amp;rsquo;s GPU orders or HBM sales. SK Hyper&amp;rsquo;s 5GW is a 2029-and-later story. SK Telecom&amp;rsquo;s own AI cloud is a 2027 story. And Nvidia&amp;rsquo;s memory partnership with SK does not prove that SK&amp;rsquo;s data centers will buy all of their HBM from SK hynix.&lt;/p&gt;
&lt;h2 id="12-calendar-what-to-watch"&gt;12. Calendar: what to watch
&lt;/h2&gt;&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;When&lt;/th&gt;
 &lt;th&gt;What&lt;/th&gt;
 &lt;th&gt;What counts as a pass&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;After Oct 30, 2026&lt;/td&gt;
 &lt;td&gt;Naver filing on the Nvidia share sale&lt;/td&gt;
 &lt;td&gt;Payment completed, or the terms of an extension&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Naver&amp;rsquo;s next results and funding filings&lt;/td&gt;
 &lt;td&gt;Brookfield project company, customer and lease structure&lt;/td&gt;
 &lt;td&gt;Financing closed and drawn, parent guarantees, customer contract terms and capacity&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Late 2026 to early 2027&lt;/td&gt;
 &lt;td&gt;SK Hyper outside investors&lt;/td&gt;
 &lt;td&gt;Named investors, equity or debt, amounts, drawdown terms, anchor-customer commitments&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;1Q 2027&lt;/td&gt;
 &lt;td&gt;SK Horizon set-up and deal close&lt;/td&gt;
 &lt;td&gt;Approvals, payment for old and new shares, stakes and net debt after the deal&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;1H 2027&lt;/td&gt;
 &lt;td&gt;Naver&amp;rsquo;s 55MW and SKT&amp;rsquo;s first AI factory&lt;/td&gt;
 &lt;td&gt;Power on, GPUs delivered, customer acceptance, billing starts&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;h2 id="bottom-line"&gt;Bottom line
&lt;/h2&gt;&lt;p&gt;Korea has real assets in this race: a national internet champion with its own AI stack, a telecom group with land, power and networks, and supportive policy money. What it does not have yet is public proof that the new capacity will be paid for. Watch three things in order: money that actually arrives, megawatts that actually bill, and contracts that say who pays for idle chips. Naver should show the first two soonest. SK Horizon already shows the second for its existing sites. SK Hyper has to show all three.&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Interest note: We use AI models, including Anthropic&amp;rsquo;s Claude and OpenAI&amp;rsquo;s models, in our research. SK Telecom invested about $100 million in Anthropic in 2023, and OpenAI has announced plans to explore AI data centers in Korea with SK Telecom. We cite public figures only and make no judgment about the effect on either company.&lt;/em&gt;&lt;sup id="fnref:28"&gt;&lt;a href="#fn:28" class="footnote-ref" role="doc-noteref"&gt;28&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Disclaimer: For research and information purposes only. Not investment advice. Names cited are for analytical illustration; readers should perform their own due diligence and consult licensed advisors before any investment decision.&lt;/em&gt;&lt;/p&gt;
&lt;div class="footnotes" role="doc-endnotes"&gt;
&lt;hr&gt;
&lt;ol&gt;
&lt;li id="fn:1"&gt;
&lt;p&gt;&lt;a class="link" href="https://kind.krx.co.kr/external/2026/07/27/000002/20260727000004/11306.htm" target="_blank" rel="noopener"
 &gt;NAVER disclosure of the share issue to Nvidia (KIND)&lt;/a&gt;, 2026-07-27: about $1 billion; condition that NAVER secures at least $9 billion of computing infrastructure, excluding this investment; Oct 30, 2026 is the latest payment date and can be extended. In Korean.&amp;#160;&lt;a href="#fnref:1" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:1" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:2"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.navercorp.com/en/media/pressReleasesDetail?seq=10034687" target="_blank" rel="noopener"
 &gt;NAVER press release&lt;/a&gt;, 2026-09-04: Brookfield &amp;ldquo;has committed to investing up to US$9 billion&amp;rdquo; in the 200MW first phase of a 1GW AI factory project.&amp;#160;&lt;a href="#fnref:2" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:2" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref2:2" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:3"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1015650/000119312526369453/d120053d6k.htm" target="_blank" rel="noopener"
 &gt;SK Telecom Form 6-K&lt;/a&gt;, 2026-08-27: 24,481,427 existing shares sold for ₩1,881,128,369,253; 15,617,069 new shares for about ₩1.20 trillion; scheduled disposal date 2027-03-04.&amp;#160;&lt;a href="#fnref:3" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:3" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:4"&gt;
&lt;p&gt;&lt;a class="link" href="https://nvidianews.nvidia.com/news/south-korea-ai-infrastructure" target="_blank" rel="noopener"
 &gt;Nvidia newsroom&lt;/a&gt;, 2025-10-30: more than 260,000 GPUs planned in Korea, including NAVER Cloud more than 60,000 and SK Group more than 50,000. All described as plans.&amp;#160;&lt;a href="#fnref:4" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:4" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref2:4" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:5"&gt;
&lt;p&gt;&lt;a class="link" href="https://blogs.nvidia.com/blog/ai-summit-korea-partners-and-nvidia/" target="_blank" rel="noopener"
 &gt;Nvidia blog&lt;/a&gt;, July 2026: NAVER&amp;rsquo;s 200MW expansion at roughly 100,000 GPUs on Vera Rubin.&amp;#160;&lt;a href="#fnref:5" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:5" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:6"&gt;
&lt;p&gt;&lt;a class="link" href="https://navercorp.com/en/story/storyDetail?seq=10034603" target="_blank" rel="noopener"
 &gt;NAVER story&lt;/a&gt;, 2026-07-24: 55MW from the first half of 2027, 200MW in 2028.&amp;#160;&lt;a href="#fnref:6" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:7"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.navercorp.com/media/pressReleasesDetail?seq=10034577" target="_blank" rel="noopener"
 &gt;NAVER second-quarter 2026 results release&lt;/a&gt;, 2026-08-07: AI Factory revenue expected from the first half of 2027; plan to secure customers for the first capacity in advance. In Korean.&amp;#160;&lt;a href="#fnref:7" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:7" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref2:7" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:8"&gt;
&lt;p&gt;&lt;a class="link" href="https://news.sktelecom.com/en/3290" target="_blank" rel="noopener"
 &gt;SK Telecom newsroom&lt;/a&gt;, 2026-08-27: SK Horizon to be set up in the first quarter of 2027; 318MW including 8 existing data centers and new sites; stakes of 51%, 29% and 20%.&amp;#160;&lt;a href="#fnref:8" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:9"&gt;
&lt;p&gt;&lt;a class="link" href="https://news.sktelecom.com/en/3192" target="_blank" rel="noopener"
 &gt;SK Telecom newsroom&lt;/a&gt;, 2026-07-23: SK Hyper set up; ₩750bn by 2030 for sites and substations; phased opening of 5GW in 2029; sequential expansion toward 15GW in 2035.&amp;#160;&lt;a href="#fnref:9" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:9" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:10"&gt;
&lt;p&gt;&lt;a class="link" href="https://blogs.nvidia.co.kr/blog/sk-group-and-nvidia-expand-strategic-partnership-across-ai-factories-and-next-generation-memory/" target="_blank" rel="noopener"
 &gt;Nvidia Korea blog&lt;/a&gt;, 2026-07-25: partnership of more than $500 billion, letter of intent for SK Telecom AI factories of up to 2GW, first factory targeted for 2027. In Korean.&amp;#160;&lt;a href="#fnref:10" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:10" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:11"&gt;
&lt;p&gt;&lt;a class="link" href="https://news.sktelecom.com/227156" target="_blank" rel="noopener"
 &gt;SK Telecom newsroom Q&amp;amp;A&lt;/a&gt;, 2026. In Korean.&amp;#160;&lt;a href="#fnref:11" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:12"&gt;
&lt;p&gt;&lt;a class="link" href="https://news.sktelecom.com/en/3218" target="_blank" rel="noopener"
 &gt;SK Telecom second-quarter 2026 results&lt;/a&gt;, 2026-08-05: AI data center revenue ₩136.2bn, up 92.5% from a year earlier.&amp;#160;&lt;a href="#fnref:12" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:13"&gt;
&lt;p&gt;&lt;a class="link" href="https://news.sktelecom.com/en/1948" target="_blank" rel="noopener"
 &gt;SK Telecom newsroom&lt;/a&gt;, 2025-06-22: AWS and SK Group Ulsan AI Zone, operation targeted by 2027, 15-year strategic agreement.&amp;#160;&lt;a href="#fnref:13" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:13" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:14"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.hankyung.com/article/2026102318001" target="_blank" rel="noopener"
 &gt;The Korea Economic Daily&lt;/a&gt;, 2026-10-23, and &lt;a class="link" href="https://www.thelec.kr/news/articleView.html?idxno=64194" target="_blank" rel="noopener"
 &gt;The Elec&lt;/a&gt;, 2026-10-23, both citing the SK Hyper chief executive&amp;rsquo;s Bloomberg interview: about $350 billion for the first 5GW; revenue possible as early as next year. Secondary; both rely on the same interview. In Korean.&amp;#160;&lt;a href="#fnref:14" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:14" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:15"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.navercorp.com/api/article/download/07b3b329-bbbe-481c-ac47-b3d0d7fea855" target="_blank" rel="noopener"
 &gt;NAVER second-quarter 2026 IR presentation&lt;/a&gt;, pp. 7 and 9: consolidated cash and short-term financial instruments ₩8,401.4bn, company-defined net cash ₩1,825.7bn; server useful life changed from five to six years, about ₩100bn of annual cost deferral.&amp;#160;&lt;a href="#fnref:15" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:15" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:16"&gt;
&lt;p&gt;&lt;a class="link" href="https://news.sktelecom.com/en/1702" target="_blank" rel="noopener"
 &gt;SK Telecom newsroom&lt;/a&gt;, 2025-01-13: GPU-as-a-service launch with Lambda, in which SKT has invested.&amp;#160;&lt;a href="#fnref:16" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:16" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:17"&gt;
&lt;p&gt;&lt;a class="link" href="https://ir.sghcorp.com/news/news-details/2025/Penguin-Solutions-Signs-AI-Data-Center-Collaboration-Agreement-with-SK-Telecom-and-SK-hynix/default.aspx" target="_blank" rel="noopener"
 &gt;Penguin Solutions&lt;/a&gt;, 2025-01-09: SK Telecom investment of $200 million.&amp;#160;&lt;a href="#fnref:17" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:18"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1769628/000176962825000047/crwv-20250909.htm" target="_blank" rel="noopener"
 &gt;CoreWeave Form 8-K&lt;/a&gt;, September 2025: initial order value of $6.3 billion; Nvidia obligated to buy residual unused capacity through 2032-04-13.&amp;#160;&lt;a href="#fnref:18" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:19"&gt;
&lt;p&gt;&lt;a class="link" href="https://nvidianews.nvidia.com/news/naver-nvidia-and-brookfield-to-expand-koreas-national-ai-factory-infrastructure-buildout" target="_blank" rel="noopener"
 &gt;Nvidia newsroom&lt;/a&gt;, July 2026: Brookfield &amp;ldquo;entered into a nonbinding term sheet to fund up to $9 billion.&amp;rdquo;&amp;#160;&lt;a href="#fnref:19" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:20"&gt;
&lt;p&gt;&lt;a class="link" href="https://en.sedaily.com/news/2026/09/23/naver-brookfield-launch-10-trillion-won-ai-data-center" target="_blank" rel="noopener"
 &gt;Seoul Economic Daily&lt;/a&gt;, 2026-09-23, citing unnamed investment-banking sources. Secondary.&amp;#160;&lt;a href="#fnref:20" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:21"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.fsc.go.kr/po010101/86717" target="_blank" rel="noopener"
 &gt;Financial Services Commission&lt;/a&gt;, 2026-04-16: ₩400bn of loans (₩340bn from the advanced strategic industries fund and ₩60bn from Korea Development Bank) for a ₩922.1bn project, five-year term, low rate in the 3% range. In Korean.&amp;#160;&lt;a href="#fnref:21" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:22"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1769628/000176962826000366/crwv-20260630.htm" target="_blank" rel="noopener"
 &gt;CoreWeave Form 10-Q&lt;/a&gt;, six months to 2026-06-30: operating cash flow $3,663m; purchases of property and equipment $14,117m. The gap is our subtraction, not company-defined free cash flow.&amp;#160;&lt;a href="#fnref:22" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:23"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1513845/000110465926094844/nbis-20260812xex99d1.htm" target="_blank" rel="noopener"
 &gt;Nebius Form 6-K, Exhibit 99.1&lt;/a&gt;, 2026-08-12: six-month operating cash flow $4,504.1m; capex $8,130.3m; increase in deferred revenue $4,395.0m.&amp;#160;&lt;a href="#fnref:23" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:24"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1839341/000183934126000014/core-20260630.htm" target="_blank" rel="noopener"
 &gt;Core Scientific Form 10-Q&lt;/a&gt;, as of 2026-06-30: about 590MW of leased customer power capacity at five CoreWeave sites, 395MW billable, 195MW under construction or commissioning.&amp;#160;&lt;a href="#fnref:24" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:25"&gt;
&lt;p&gt;&lt;a class="link" href="https://cdn2.singteldigital.com/content/dam/singtel/investorRelations/annualReports/2026/AR2026.pdf" target="_blank" rel="noopener"
 &gt;Singtel Annual Report 2026&lt;/a&gt;, year to 2026-03-31: Nxera preference shares issued to Stellar Asia Holdings II for net S$337m (5.8%); FY27 growth capex of about S$1.2bn, mainly Nxera and RE:AI.&amp;#160;&lt;a href="#fnref:25" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:26"&gt;
&lt;p&gt;&lt;a class="link" href="https://nebius.com/newsroom/nebius-secures-3-5-billion-financing-facility-with-citi" target="_blank" rel="noopener"
 &gt;Nebius newsroom&lt;/a&gt;, 2026-10-22: $3.5 billion senior secured facility, 364 days, Term SOFR plus 2.75% rising to up to 3.50%. Drawn amount not disclosed.&amp;#160;&lt;a href="#fnref:26" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:27"&gt;
&lt;p&gt;Our illustration. Capital recovery factor CRF(r, n) = r / [1 − (1 + r)^−n]. Yearly amount = $8bn × CRF(10%, chip life) + $2bn × CRF(8%, 20) + $0.5bn. Taxes, ramp-up, resale value, inflation and replacement costs are left out. Not a forecast for NAVER or SK.&amp;#160;&lt;a href="#fnref:27" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:28"&gt;
&lt;p&gt;&lt;a class="link" href="https://amp.kedglobal.com/newsAmp/ked202308130002" target="_blank" rel="noopener"
 &gt;The Korea Economic Daily (KED Global)&lt;/a&gt;, 2023-08-13, on SK Telecom&amp;rsquo;s investment in Anthropic; &lt;a class="link" href="https://openai.com/index/samsung-and-sk-join-stargate" target="_blank" rel="noopener"
 &gt;OpenAI&lt;/a&gt;, 2025-10-01, on exploring AI data centers in Korea with SK Telecom.&amp;#160;&lt;a href="#fnref:28" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;/ol&gt;
&lt;/div&gt;</description></item><item><title>AI Can Grow 41% a Year. The Stock Price Is a Different Test.</title><link>https://koreainvestinsights.com/post/ai-agents-41-percent-growth-stock-prices-2026-10-25/</link><pubDate>Sun, 25 Oct 2026 14:00:00 +0900</pubDate><guid>https://koreainvestinsights.com/post/ai-agents-41-percent-growth-stock-prices-2026-10-25/</guid><description>&lt;img src="https://koreainvestinsights.com/" alt="Featured image of post AI Can Grow 41% a Year. The Stock Price Is a Different Test." /&gt;&lt;p&gt;Growing 41% a year sounds demanding. Yet Amazon Web Services grew revenue almost that fast for four years. If AI agents turn a cheap answer into a useful piece of completed work, the next four years need not be less ambitious.&lt;/p&gt;
&lt;p&gt;My judgment is cautiously positive on that growth possibility. I am less positive on the claim that it automatically makes AI stocks cheap. Microsoft, Alphabet, Meta and Nvidia offer different earnings starting points, capital burdens and prices. Among these four, Meta and Nvidia leave more room for a favorable surprise under a common earnings test. Microsoft already requires substantial execution. Alphabet sits between them.&lt;/p&gt;
&lt;p&gt;This follows &lt;a class="link" href="https://koreainvestinsights.com/post/who-pays-for-ai-demand-agents-capex-roi-2026-10-25/" &gt;our analysis of who pays for AI&lt;/a&gt;. That article ended with a question: can delegated agents take demand beyond the heaviest users and the largest businesses? Here we test that proposition against historical growth, public interviews, company disclosures, field research and current valuations.&lt;/p&gt;
&lt;h2 id="first-41-is-a-conditional-hurdle"&gt;First, 41% is a conditional hurdle
&lt;/h2&gt;&lt;p&gt;The earlier calculation compared an estimated &lt;strong&gt;$414 billion year-end 2026 AI revenue run-rate&lt;/strong&gt; with our modeled &lt;strong&gt;$1.626 trillion of annual AI-attributable revenue in 2030&lt;/strong&gt;. Four years separate those endpoints. The resulting compound annual growth rate is 40.8%, rounded to 41%.&lt;/p&gt;
&lt;p&gt;The endpoint comes from assumptions about the capex path, depreciation, operating costs and a 12% required return. It is an annual capital-cost coverage test, not cumulative sales through 2030 or a promise that every investment earns 12%. The starting point is Tech Fund&amp;rsquo;s estimate across labs, AI-attributed advertising, cloud, applications and other channels. Neither number is an audited measure of final customer spending.&lt;sup id="fnref:1"&gt;&lt;a href="#fn:1" class="footnote-ref" role="doc-noteref"&gt;1&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;That distinction needs more emphasis than it received in the first article. Tech Fund explicitly says it removes cloud sales to frontier labs and app payments to those labs to avoid double counting. We should credit that adjustment. We have not independently reconciled the netting, and AI-attributed advertising gains depend on an assumed revenue path without AI. This is uncertainty about a third-party estimate, not evidence that duplication has been found.&lt;/p&gt;
&lt;p&gt;For scale, if the starting revenue were reduced by an illustrative 20% while the endpoint stayed unchanged, the required CAGR would rise to &lt;strong&gt;48.9%&lt;/strong&gt;. A 40% reduction would raise it to &lt;strong&gt;60.0%&lt;/strong&gt;. These are denominator stress tests, not estimates of actual overstatement or a fully reconciled alternative model. A consistent alternative model would also revisit the endpoint.&lt;/p&gt;
&lt;p&gt;Thus, “41% is feasible” is a useful growth proposition. “41% guarantees the infrastructure pays for itself” is too strong.&lt;/p&gt;
&lt;h2 id="history-makes-41-plausible-not-easy"&gt;History makes 41% plausible, not easy
&lt;/h2&gt;&lt;p&gt;The comparisons below all cover four elapsed years. Revenue comparisons come first; physical-volume comparisons show how quickly adoption can spread.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Market and period&lt;/th&gt;
 &lt;th&gt;Measure&lt;/th&gt;
 &lt;th&gt;Start → finish&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Annual growth&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;AWS, 2016–2020&lt;/td&gt;
 &lt;td&gt;Segment net sales&lt;/td&gt;
 &lt;td&gt;$12.219bn → $45.370bn&lt;/td&gt;
 &lt;td style="text-align: right"&gt;38.8%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Smartphones, 2010–2014&lt;/td&gt;
 &lt;td&gt;Worldwide shipment value&lt;/td&gt;
 &lt;td&gt;$130bn → $393bn&lt;/td&gt;
 &lt;td style="text-align: right"&gt;31.9%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Smartphones, 2010–2014&lt;/td&gt;
 &lt;td&gt;Worldwide units shipped&lt;/td&gt;
 &lt;td&gt;305m → 1,306m&lt;/td&gt;
 &lt;td style="text-align: right"&gt;43.9%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Electric cars, 2020–2024&lt;/td&gt;
 &lt;td&gt;Worldwide units sold&lt;/td&gt;
 &lt;td&gt;About 3m → over 17m&lt;/td&gt;
 &lt;td style="text-align: right"&gt;About 54% or more&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Sources: Amazon annual disclosures; WIPO&amp;rsquo;s industry study using IDC data; IEA. CAGR is &lt;code&gt;(finish / start)^(1/4) − 1&lt;/code&gt;. EV inputs are rounded. Shipment value and vehicle sales are not software revenue.&lt;sup id="fnref:2"&gt;&lt;a href="#fn:2" class="footnote-ref" role="doc-noteref"&gt;2&lt;/a&gt;&lt;/sup&gt;&lt;sup id="fnref:3"&gt;&lt;a href="#fn:3" class="footnote-ref" role="doc-noteref"&gt;3&lt;/a&gt;&lt;/sup&gt;&lt;sup id="fnref:4"&gt;&lt;a href="#fn:4" class="footnote-ref" role="doc-noteref"&gt;4&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;AWS is the strongest comparison for the optimistic argument. It sold metered infrastructure, broadened the tasks customers could run and reached a much larger revenue base. But it was a successful company selected after the fact. Its starting revenue was about one-thirty-fourth of the AI estimate used here. Matching its percentage growth now requires a much larger absolute spending increase. AI revenue of $414bn compounded at 41% would add roughly $1.22tn to the annual run-rate by 2030.&lt;/p&gt;
&lt;p&gt;Smartphones supply the necessary caution. Unit shipments grew faster than 41%, but shipment value grew 31.9% because average selling prices fell. EVs demonstrate rapid physical adoption, aided by policy and a small starting base; they do not establish a 54% revenue-growth precedent.&lt;/p&gt;
&lt;p&gt;These are selected analogues, not a statistical sample from which to estimate AI&amp;rsquo;s success probability. They rebut “41% is unprecedented.” They do not prove “41% is easy.”&lt;/p&gt;
&lt;h2 id="cheaper-tokens-must-become-cheaper-completed-work"&gt;Cheaper tokens must become cheaper completed work
&lt;/h2&gt;&lt;p&gt;For a constant mix of metered tokens, revenue equals volume times realized price. A 41% revenue increase with a 40% price decline requires 2.35 times the volume. With a 54% decline, it requires 3.07 times.&lt;/p&gt;
&lt;p&gt;&lt;img alt="Required token volume for 41% revenue growth at different realized token prices; constant-mix sensitivity, not a forecast" loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/ai-agents-41-percent-growth-stock-prices-2026-10-25/price-volume.en.png"&gt;
&lt;/p&gt;
&lt;p&gt;Alphabet reported API throughput rising from 16 billion to 22 billion tokens per minute between the first and second quarters of 2026. That is 37.5% sequential growth. Repeating that quarter four times would produce 3.57 times annual volume, above the two stressed requirements. It is an annualized illustration of one platform&amp;rsquo;s run-rate, not observed full-year growth, industry growth or four-year evidence.&lt;sup id="fnref:5"&gt;&lt;a href="#fn:5" class="footnote-ref" role="doc-noteref"&gt;5&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;Ramp adds a different signal. In its token-management sample, spending in the week ending October 11 rose 23.0% against the week ending September 13, while volume rose 52.4%. This is compatible with demand outrunning falling prices over that window. It does not contradict the earlier article&amp;rsquo;s August-to-October comparison: the start dates differ. Ramp explicitly says this product sample is not representative of all its AI spending. Its model-specific price series should not be combined mechanically with all-model spending and volume.&lt;sup id="fnref:6"&gt;&lt;a href="#fn:6" class="footnote-ref" role="doc-noteref"&gt;6&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;There is also no rule that the realized price must fall 40–54% every year. Quality, model mix, caching, service tiers and temporary capacity shortages change the bill. Subscriptions, advertising and fees for outcomes further weaken a simple token-to-total-revenue identity. A cheaper input can improve a supplier&amp;rsquo;s margin even when the customer price stays unchanged.&lt;/p&gt;
&lt;p&gt;The economic unit to watch is therefore &lt;strong&gt;cost per accepted result&lt;/strong&gt;, including human review and rework. A long agent run that consumes ten times more tokens but produces an unusable result is demand for computing today. It is poor evidence of a renewable customer budget tomorrow.&lt;/p&gt;
&lt;h2 id="the-agent-thesis-has-three-routes-out-of-the-early-adopter-niche"&gt;The agent thesis has three routes out of the early-adopter niche
&lt;/h2&gt;&lt;p&gt;The thesis is that a packaged agent removes work the customer previously had to do: select a model, connect data, write code, manage context and supervise every step. Customers can delegate a bounded job through a familiar product. That can broaden adoption, increase jobs per customer and make recurring payment easier.&lt;/p&gt;
&lt;p&gt;The first route runs through existing enterprise software. In Microsoft&amp;rsquo;s July 29 earnings call, Satya Nadella reported over 30 million paid Microsoft 365 Copilot seats. He also described a move toward seats plus consumption, with GitHub Copilot revenue growing more than 60% quarter on quarter after usage billing began. One in three GitHub pull requests involved an agent. These are company-reported measures: involvement is not a merged, fully autonomous change, and revenue growth cannot be attributed entirely to extra work.&lt;sup id="fnref:7"&gt;&lt;a href="#fn:7" class="footnote-ref" role="doc-noteref"&gt;7&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;The second route is a narrow business workflow sold with setup, permissions and review already built in. Customer support, document intake and routine reconciliation have clearer acceptance criteria than “run my company.” Here a nondeveloper can buy an outcome without building an agent stack. This is our inference about distribution, rather than a measured industry forecast.&lt;/p&gt;
&lt;p&gt;The third route is a personal agent distributed through devices and consumer services. Its reach could be enormous, but usage alone leaves the business model unresolved. Free access funded by advertising, a capped subscription and metered autonomous work have different revenue and compute consequences. We do not need consumer autonomy to arrive everywhere for enterprise-agent revenue to grow rapidly. We would need broader consumer proof to justify the most expansive demand case.&lt;/p&gt;
&lt;p&gt;Adoption data still draws a boundary. McKinsey&amp;rsquo;s 2026 survey found the share of respondents at large organizations scaling agents in at least one function rose from 27% to 40%. Smaller organizations stayed around 22%. Only 37% reported any positive enterprise EBIT contribution from AI, essentially unchanged from the prior year. These are survey responses, not audited profits or a count of active agents.&lt;sup id="fnref:8"&gt;&lt;a href="#fn:8" class="footnote-ref" role="doc-noteref"&gt;8&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;The “top 1%” also needs a precise label. Ramp ranks &lt;strong&gt;firms paying for AI by spending per employee&lt;/strong&gt;, not individual AI users. In September, the median firm in that top group spent about $9,568 per employee per month, versus $15.70 for the median paying firm. That gap shows concentration of intensity. It does not establish that 1% of people produce a fixed share of global AI revenue.&lt;sup id="fnref1:6"&gt;&lt;a href="#fn:6" class="footnote-ref" role="doc-noteref"&gt;6&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;h2 id="what-the-optimists-and-skeptics-actually-disagree-about"&gt;What the optimists and skeptics actually disagree about
&lt;/h2&gt;&lt;p&gt;Jensen Huang&amp;rsquo;s August 26 earnings-call argument is that agents can keep working in the background and interact with other agents, expanding computing beyond human screen time. That is a credible mechanism for more workloads. As the chief executive of a computing supplier, he also has a direct commercial interest in that outcome. His statement is not independent proof of the budgets required to support it.&lt;sup id="fnref:9"&gt;&lt;a href="#fn:9" class="footnote-ref" role="doc-noteref"&gt;9&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;Nadella supplied an important qualification in his February 2025 interview with Dwarkesh Patel. He emphasized real economic growth and inference revenue as checks on supply-side enthusiasm. His argument was that today&amp;rsquo;s investment must become tomorrow&amp;rsquo;s customer value before the next investment cycle can be justified. That connects the agent story to recurring payment.&lt;sup id="fnref:10"&gt;&lt;a href="#fn:10" class="footnote-ref" role="doc-noteref"&gt;10&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;Andrej Karpathy&amp;rsquo;s October 2025 interview with Patel offers the strongest timing objection. His phrase was “the decade of agents.” He described impressive tools he already used, alongside substantial work still required to make them reliable employees or interns. His view allows rapid commercial growth during a long transition; it does not imply agents produce no value until the transition ends.&lt;sup id="fnref:11"&gt;&lt;a href="#fn:11" class="footnote-ref" role="doc-noteref"&gt;11&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;Research gives both sides something concrete. Erik Brynjolfsson, Danielle Li and Lindsey Raymond&amp;rsquo;s final published study of 5,172 customer-support workers found a 15% average increase in issues resolved per hour. The gains varied by worker experience. This measures assistance in one workflow, not fully autonomous employment or supplier revenue.&lt;sup id="fnref:12"&gt;&lt;a href="#fn:12" class="footnote-ref" role="doc-noteref"&gt;12&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;METR&amp;rsquo;s early-2025 experiment found experienced open-source developers took 19% longer with the tested AI tools. But its February 2026 update warned that newer measurements were biased by participant selection and difficulties measuring concurrent agents. The researchers thought later tools likely helped more, while describing the measured size of that improvement as unreliable. Repeating the old slowdown as a universal current verdict would be misleading.&lt;sup id="fnref:13"&gt;&lt;a href="#fn:13" class="footnote-ref" role="doc-noteref"&gt;13&lt;/a&gt;&lt;/sup&gt;&lt;sup id="fnref:14"&gt;&lt;a href="#fn:14" class="footnote-ref" role="doc-noteref"&gt;14&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;Daron Acemoglu&amp;rsquo;s task-based macroeconomic analysis provides another restraint: automating selected tasks does not automatically deliver a vast aggregate productivity gain. His 2024 model estimated at most 0.66% total-factor-productivity improvement over ten years under its assumptions. That is not a forecast of AI-vendor revenue. Vendors can grow faster than GDP by replacing other spending; the distinction does not eliminate the need for useful output.&lt;sup id="fnref:15"&gt;&lt;a href="#fn:15" class="footnote-ref" role="doc-noteref"&gt;15&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;Taken together, these sources support a narrower, stronger proposition: agents can create a large new paying market before they become dependable general-purpose employees. Their success depends on choosing tasks whose gains exceed setup, supervision and error costs.&lt;/p&gt;
&lt;h2 id="my-probability-judgment-separates-growth-from-payback"&gt;My probability judgment separates growth from payback
&lt;/h2&gt;&lt;p&gt;My working range is &lt;strong&gt;50–60% for the gross AI-attributable revenue measure used above reaching roughly $1.63tn by 2030&lt;/strong&gt;, if its definition is held fixed. I place &lt;strong&gt;40–50% on the original capital-cost coverage case&lt;/strong&gt;, with the earlier 45% estimate inside that range. These are subjective research judgments, not calibrated statistical probabilities. The second range includes uncertainty about margins, revenue overlap, asset lives and how much additional capacity gets built. It is not the first probability multiplied by an independent factor.&lt;/p&gt;
&lt;p&gt;The positive weight comes from demonstrated enterprise distribution, real usage-linked revenue and throughput growth compatible with the near-term hurdle. The restraint comes from the much larger starting base, uneven smaller-business adoption and incomplete measurement of customer returns. Historical winners show possibility; they do not supply the probability.&lt;/p&gt;
&lt;p&gt;A useful trajectory check is an end-2027 run-rate around $584bn, then $823bn in 2028 and $1.16tn in 2029, using the same definition and exactly 41% growth. Falling below one point would not make recovery impossible. It would raise the later growth required. Changing what counts as AI revenue would not constitute beating the trajectory.&lt;/p&gt;
&lt;h2 id="prices-discount-company-cash-flows-not-one-industry-probability"&gt;Prices discount company cash flows, not one industry probability
&lt;/h2&gt;&lt;p&gt;A market can correctly expect strong AI demand and still overprice a particular supplier. Competition may transfer the gains to customers. Depreciation can absorb gross profit. New shares can dilute per-share growth. A growing industry can also contain businesses that destroy capital.&lt;/p&gt;
&lt;p&gt;To make the price test concrete, we use one provider&amp;rsquo;s forward P/E snapshot, retrieved October 25, with October 23 regular-session closing prices. These are vendor-adjusted forward earnings measures, not identical fiscal periods or synchronized 2030 forecasts. We use each quoted earnings base as a reference for a transparent sensitivity test.&lt;sup id="fnref:16"&gt;&lt;a href="#fn:16" class="footnote-ref" role="doc-noteref"&gt;16&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Company&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Closing price&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Forward P/E&lt;/th&gt;
 &lt;th style="text-align: right"&gt;EPS growth needed with terminal P/E 20x&lt;/th&gt;
 &lt;th style="text-align: right"&gt;With terminal P/E 25x&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Microsoft&lt;/td&gt;
 &lt;td style="text-align: right"&gt;$524.51&lt;/td&gt;
 &lt;td style="text-align: right"&gt;26.54x&lt;/td&gt;
 &lt;td style="text-align: right"&gt;18.1% a year&lt;/td&gt;
 &lt;td style="text-align: right"&gt;11.7% a year&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Alphabet, Class A&lt;/td&gt;
 &lt;td style="text-align: right"&gt;$320.21&lt;/td&gt;
 &lt;td style="text-align: right"&gt;23.94x&lt;/td&gt;
 &lt;td style="text-align: right"&gt;15.1%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;8.8%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Meta&lt;/td&gt;
 &lt;td style="text-align: right"&gt;$642.62&lt;/td&gt;
 &lt;td style="text-align: right"&gt;19.51x&lt;/td&gt;
 &lt;td style="text-align: right"&gt;9.3%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;3.4%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Nvidia&lt;/td&gt;
 &lt;td style="text-align: right"&gt;$210.96&lt;/td&gt;
 &lt;td style="text-align: right"&gt;17.44x&lt;/td&gt;
 &lt;td style="text-align: right"&gt;6.3%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;0.5%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Assumptions: four years of growth from today&amp;rsquo;s quoted forward-EPS reference, a 10% annual share-price return, and the stated multiple applied to the forward EPS at the end of year four. Dividends are excluded; EPS already includes the share-count effect. Formula: &lt;code&gt;required EPS growth = 1.10 × (current P/E / terminal P/E)^(1/4) − 1&lt;/code&gt;. This is an earnings sensitivity, not a full discounted-cash-flow valuation or a price target.&lt;/p&gt;
&lt;p&gt;&lt;img alt="Four-year EPS growth required to support 10% annual price returns under terminal P/E assumptions of 20 and 25; sensitivity rather than forecast" loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://koreainvestinsights.com/post-assets/ai-agents-41-percent-growth-stock-prices-2026-10-25/earnings-hurdle.en.png"&gt;
&lt;/p&gt;
&lt;p&gt;A second calculation shows why an “AI probability implied by the market” is not uniquely observable. Assume only two company earnings paths: 5% or 20% annual EPS growth for four years, a 20x terminal multiple in both cases and a 10% discount rate. The arithmetic weight on the faster path needed to match today&amp;rsquo;s multiple is about &lt;strong&gt;85% for Microsoft, 63% for Alphabet, 25% for Meta and 7% for Nvidia&lt;/strong&gt;. These are break-even weights in an invented two-outcome model, not measured investor beliefs or probabilities of 41% industry growth.&lt;/p&gt;
&lt;p&gt;Change only the terminal multiple to 25x and the Microsoft and Alphabet weights fall to about 40% and 22%. For Meta and Nvidia, even the slower modeled outcome is worth more than today&amp;rsquo;s price, so there is no feasible probability between zero and one that equates the model to the quote. That demonstrates dependence on assumptions, not a risk-free bargain.&lt;/p&gt;
&lt;h2 id="the-discount-looks-selective"&gt;The discount looks selective
&lt;/h2&gt;&lt;p&gt;&lt;strong&gt;Microsoft: broadly reasonable, with limited room for disappointment.&lt;/strong&gt; Distribution and consumption billing support sustained growth, but the 20x terminal case requires roughly 18% annual EPS growth after the quoted forward earnings base. The market is already paying for execution. The latest annual filing also contains a $6.5bn OpenAI investment gain and sharply higher depreciation. We use the adjusted forward measure above; we do not project that investment gain as operating profit.&lt;sup id="fnref:17"&gt;&lt;a href="#fn:17" class="footnote-ref" role="doc-noteref"&gt;17&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Alphabet: conditional upside, rather than a blanket discount.&lt;/strong&gt; Search distribution and cloud demand give it several ways to capture AI spending. Against that, agent answers can alter search monetization and capex must convert into cash. Its apparently cheap trailing P/E is a poor anchor: the first-half filing reports $135.95bn of equity-security gains, mostly unrealized. The 23.94x forward measure asks for about 15% EPS growth in the 20x terminal case.&lt;sup id="fnref:18"&gt;&lt;a href="#fn:18" class="footnote-ref" role="doc-noteref"&gt;18&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Meta: the clearest modest-growth hurdle, with a real cash-flow objection.&lt;/strong&gt; About 9% EPS growth clears our 20x terminal test. Advertising can monetize better recommendations without asking consumers to buy an agent subscription. But the company reported $85.82bn of nine-month capex including finance-lease principal, against $96.87bn of operating cash flow, and raised 2026 capex guidance to $135–145bn. The low earnings multiple partly compensates for this investment burden. It looks excessively discounted only if durable advertising gains and spending discipline restore cash conversion.&lt;sup id="fnref:19"&gt;&lt;a href="#fn:19" class="footnote-ref" role="doc-noteref"&gt;19&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Nvidia: a substantial discount to earnings durability, not to near-term growth.&lt;/strong&gt; The low quoted forward multiple starts from a large earnings forecast. A 30% cut to that EPS base would turn 17.44x into 24.91x at the same price. Customer financing, competition and the eventual supply response can make that cut possible. Still, if earnings survive the current buildout and continue growing, its price leaves more room than Microsoft&amp;rsquo;s under the same terminal assumptions. Nvidia&amp;rsquo;s disclosed cash generation is also less directly burdened by owning the whole customer data-center estate, though strategic investments and customer concentration still matter.&lt;sup id="fnref:20"&gt;&lt;a href="#fn:20" class="footnote-ref" role="doc-noteref"&gt;20&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;My conclusion is therefore &lt;strong&gt;selective underpricing of durable earnings, with Meta and Nvidia the stronger cases for further scrutiny&lt;/strong&gt;. I do not see sufficient evidence that the entire AI trade is excessively discounted. The broad 41% revenue thesis supports these cases; cash conversion and earnings durability decide them. This comparison covers these four companies, not leveraged infrastructure operators or the whole AI stock universe.&lt;/p&gt;
&lt;p&gt;For Korean memory suppliers, the link is another step removed. Agents can increase demand for server DRAM, storage and CPUs as well as HBM. Yet model efficiency, caching, utilization, memory supply and contract prices decide the revenue captured by suppliers. A successful agent thesis alone does not establish that a memory stock is cheap.&lt;/p&gt;
&lt;h2 id="the-next-evidence-must-connect-usage-to-renewal"&gt;The next evidence must connect usage to renewal
&lt;/h2&gt;&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Question&lt;/th&gt;
 &lt;th&gt;Evidence that would strengthen the case&lt;/th&gt;
 &lt;th&gt;Evidence that would weaken it&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Does distribution reach beyond large firms?&lt;/td&gt;
 &lt;td&gt;Smaller-firm agent scaling rises above the current 22% survey baseline, with paid renewals&lt;/td&gt;
 &lt;td&gt;Deployment rises only in large firms; smaller-firm adoption stays flat&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Does usage become revenue?&lt;/td&gt;
 &lt;td&gt;Comparable cohorts show rising retained paid work and net revenue&lt;/td&gt;
 &lt;td&gt;Tokens grow while recurring spending and retention stall&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Do customers earn the saving?&lt;/td&gt;
 &lt;td&gt;Review-inclusive cost per accepted task falls and quality holds&lt;/td&gt;
 &lt;td&gt;Human supervision and rework absorb the saving&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Can shareholders keep the gain?&lt;/td&gt;
 &lt;td&gt;Operating cash flow outgrows recurring capex and dilution&lt;/td&gt;
 &lt;td&gt;EPS rises while cash conversion deteriorates persistently&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Is the four-year path intact?&lt;/td&gt;
 &lt;td&gt;Same-definition revenue approaches the $584bn end-2027 reference&lt;/td&gt;
 &lt;td&gt;Growth depends on relabeling existing revenue or adding overlapping layers&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;The coming earnings releases and subsequent paid-use data can test the first steps. Four years of compounding will require repeated evidence, not one spectacular quarter.&lt;/p&gt;
&lt;p&gt;If you use an agent at work, what matters more to renewal: tokens consumed, tasks attempted, or tasks accepted without substantial rework? That is the operating question behind this investment debate.&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Interest disclosure: This article was prepared with OpenAI tools. OpenAI is discussed as an industry participant. Public information is used; no assessment is made of the magnitude of this same-vendor interest.&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;This is public-market research, not a recommendation tailored to any reader&amp;rsquo;s portfolio. Scenarios and valuation sensitivities are assumptions, not promised returns.&lt;/em&gt;&lt;/p&gt;
&lt;div class="footnotes" role="doc-endnotes"&gt;
&lt;hr&gt;
&lt;ol&gt;
&lt;li id="fn:1"&gt;
&lt;p&gt;&lt;a class="link" href="https://koreainvestinsights.com/post/who-pays-for-ai-demand-agents-capex-roi-2026-10-25/" &gt;Previous article and model assumptions&lt;/a&gt;; &lt;a class="link" href="https://www.techinvestments.io/p/ais-attractive-irr-anthropic-asml" target="_blank" rel="noopener"
 &gt;Tech Fund, October 18, 2026&lt;/a&gt;. Third-party year-end revenue estimate and our scenario calculations. The detailed model gives $1.625936tn, so CAGR is 40.775%. Endpoints do not form an audited, consolidated customer-spending series.&amp;#160;&lt;a href="#fnref:1" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:2"&gt;
&lt;p&gt;Amazon &lt;a class="link" href="https://ir.aboutamazon.com/files/doc_financials/annual/2018-Annual-Report.pdf" target="_blank" rel="noopener"
 &gt;2018 annual report&lt;/a&gt;, AWS 2016 segment net sales; &lt;a class="link" href="https://ir.aboutamazon.com/files/doc_financials/2021/q4/business_and_financial_update.pdf" target="_blank" rel="noopener"
 &gt;2021 year-end financial update&lt;/a&gt;, released February 3, 2022, with the 2020 comparison column. Our four-year CAGR calculation.&amp;#160;&lt;a href="#fnref:2" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:3"&gt;
&lt;p&gt;Jason Dedrick and Kenneth Kraemer, &lt;a class="link" href="https://www.wipo.int/edocs/pubdocs/en/wipo_pub_econstat_wp_41.pdf" target="_blank" rel="noopener"
 &gt;WIPO working paper 41&lt;/a&gt;, November 2017, Table 1 using IDC industry data. Shipment value, not the entire smartphone ecosystem&amp;rsquo;s revenue.&amp;#160;&lt;a href="#fnref:3" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:4"&gt;
&lt;p&gt;IEA, &lt;a class="link" href="https://www.iea.org/reports/global-ev-outlook-2021" target="_blank" rel="noopener"
 &gt;Global EV Outlook 2021&lt;/a&gt; and &lt;a class="link" href="https://www.iea.org/reports/global-ev-outlook-2025/executive-summary" target="_blank" rel="noopener"
 &gt;Global EV Outlook 2025 executive summary&lt;/a&gt;. Rounded global electric-car sales; volume, not revenue.&amp;#160;&lt;a href="#fnref:4" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:5"&gt;
&lt;p&gt;Sundar Pichai, &lt;a class="link" href="https://blog.google/company-news/inside-google/message-ceo/alphabet-earnings-q2-2026/" target="_blank" rel="noopener"
 &gt;Alphabet Q2 2026 remarks&lt;/a&gt;, July 22, 2026. Company-reported API run-rates. Annualization is our calculation.&amp;#160;&lt;a href="#fnref:5" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:6"&gt;
&lt;p&gt;&lt;a class="link" href="https://ramp.com/data/ai-index" target="_blank" rel="noopener"
 &gt;Ramp AI Index&lt;/a&gt;, retrieved October 25, 2026; September company-spend observations and token week ending October 11. Company and token-product samples have different coverage. The percentile is across firms, not people.&amp;#160;&lt;a href="#fnref:6" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:6" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:7"&gt;
&lt;p&gt;Satya Nadella and Amy Hood, &lt;a class="link" href="https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q4" target="_blank" rel="noopener"
 &gt;Microsoft FY2026 Q4 earnings transcript&lt;/a&gt;, July 29, 2026. Paid seats, GitHub involvement and product revenue have distinct definitions.&amp;#160;&lt;a href="#fnref:7" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:8"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai" target="_blank" rel="noopener"
 &gt;The state of AI in 2026&lt;/a&gt;, August 25, 2026. Survey of 1,719 respondents in 97 countries, fielded May 4–June 8; results are self-reported.&amp;#160;&lt;a href="#fnref:8" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:9"&gt;
&lt;p&gt;Jensen Huang, &lt;a class="link" href="https://investor.nvidia.com/files/content_files/TRANSCRIPT_-NVIDIA-Corp-NVDA-US-Q2-2027-Earnings-Call-26-August-2026-5_00-PM-ET.pdf" target="_blank" rel="noopener"
 &gt;Nvidia Q2 FY2027 earnings transcript&lt;/a&gt;, August 26, 2026, background-agent discussion. Management&amp;rsquo;s forward-looking view.&amp;#160;&lt;a href="#fnref:9" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:10"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.dwarkesh.com/p/satya-nadella" target="_blank" rel="noopener"
 &gt;Satya Nadella interviewed by Dwarkesh Patel&lt;/a&gt;, February 19, 2025, economic growth and inference-revenue discussion. Public interview, not an interview conducted for this article.&amp;#160;&lt;a href="#fnref:10" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:11"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.dwarkesh.com/p/andrej-karpathy" target="_blank" rel="noopener"
 &gt;Andrej Karpathy interviewed by Dwarkesh Patel&lt;/a&gt;, October 17, 2025, opening discussion of agent deployment timelines.&amp;#160;&lt;a href="#fnref:11" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:12"&gt;
&lt;p&gt;Brynjolfsson, Li and Raymond, &lt;a class="link" href="https://doi.org/10.1093/qje/qjae044" target="_blank" rel="noopener"
 &gt;Generative AI at Work&lt;/a&gt;, Quarterly Journal of Economics, May 2025. We use the published 5,172-worker, 15% result, not the earlier working paper&amp;rsquo;s 5,179-worker, 14% version.&amp;#160;&lt;a href="#fnref:12" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:13"&gt;
&lt;p&gt;Becker, Rush, Barnes and Rein, &lt;a class="link" href="https://arxiv.org/abs/2507.09089" target="_blank" rel="noopener"
 &gt;METR early-2025 developer experiment&lt;/a&gt;, July 2025; 16 developers and 246 tasks in familiar repositories. Context-specific result.&amp;#160;&lt;a href="#fnref:13" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:14"&gt;
&lt;p&gt;&lt;a class="link" href="https://metr.org/blog/2026-02-24-uplift-update/" target="_blank" rel="noopener"
 &gt;METR&amp;rsquo;s experiment-design update&lt;/a&gt;, February 24, 2026. Selection and measurement problems limit the newer speedup estimates.&amp;#160;&lt;a href="#fnref:14" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:15"&gt;
&lt;p&gt;Daron Acemoglu, &lt;a class="link" href="https://www.nber.org/papers/w32487" target="_blank" rel="noopener"
 &gt;The Simple Macroeconomics of AI&lt;/a&gt;, May 2024. Conditional ten-year aggregate-productivity model, not a supplier-sales forecast.&amp;#160;&lt;a href="#fnref:15" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:16"&gt;
&lt;p&gt;StockAnalysis statistics: &lt;a class="link" href="https://stockanalysis.com/stocks/msft/statistics/" target="_blank" rel="noopener"
 &gt;Microsoft&lt;/a&gt;, &lt;a class="link" href="https://stockanalysis.com/stocks/googl/statistics/" target="_blank" rel="noopener"
 &gt;Alphabet&lt;/a&gt;, &lt;a class="link" href="https://stockanalysis.com/stocks/meta/statistics/" target="_blank" rel="noopener"
 &gt;Meta&lt;/a&gt;, &lt;a class="link" href="https://stockanalysis.com/stocks/nvda/statistics/" target="_blank" rel="noopener"
 &gt;Nvidia&lt;/a&gt;. October 23, 2026 closes; ratios retrieved October 25. Forward estimates depend on the provider and company fiscal year. All reverse-valuation calculations are ours.&amp;#160;&lt;a href="#fnref:16" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:17"&gt;
&lt;p&gt;Microsoft &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/789019/000119312526323660/msft-20260630.htm" target="_blank" rel="noopener"
 &gt;FY2026 Form 10-K&lt;/a&gt;. OpenAI investment gains and depreciation are separate from the adjusted forward-EPS sensitivity.&amp;#160;&lt;a href="#fnref:17" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:18"&gt;
&lt;p&gt;Alphabet &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1652044/000165204426000071/goog-20260630.htm" target="_blank" rel="noopener"
 &gt;Q2 2026 Form 10-Q&lt;/a&gt;, other income and equity-security gains. These gains are not recurring search or cloud earnings.&amp;#160;&lt;a href="#fnref:18" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:19"&gt;
&lt;p&gt;Meta &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1326801/000162828026066700/meta-09302026xexhibit991.htm" target="_blank" rel="noopener"
 &gt;Q3 2026 earnings release&lt;/a&gt; and &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1326801/000162828026066773/meta-20260930.htm" target="_blank" rel="noopener"
 &gt;Form 10-Q&lt;/a&gt;. Capex includes finance-lease principal payments.&amp;#160;&lt;a href="#fnref:19" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:20"&gt;
&lt;p&gt;Nvidia &lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1045810/000104581026000075/nvda-20260726.htm" target="_blank" rel="noopener"
 &gt;Q2 FY2027 Form 10-Q&lt;/a&gt;. Customer concentration, investments and cash flows. Earnings cuts used in this article are sensitivities, not company guidance.&amp;#160;&lt;a href="#fnref:20" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;/ol&gt;
&lt;/div&gt;</description></item><item><title>Who Pays for AI? Seven Kinds of Demand, the Agent Engine, and the Capex Payback Test</title><link>https://koreainvestinsights.com/post/who-pays-for-ai-demand-agents-capex-roi-2026-10-25/</link><pubDate>Sun, 25 Oct 2026 12:30:00 +0900</pubDate><guid>https://koreainvestinsights.com/post/who-pays-for-ai-demand-agents-capex-roi-2026-10-25/</guid><description>&lt;img src="https://koreainvestinsights.com/" alt="Featured image of post Who Pays for AI? Seven Kinds of Demand, the Agent Engine, and the Capex Payback Test" /&gt;&lt;p&gt;Is AI a bubble? Strip that question down and three simpler ones are left. Who is paying for AI today? How fast is that growing? And is it enough to pay back the data centers being built for it?&lt;/p&gt;
&lt;p&gt;This post answers all three with numbers from our research. The short version: people and companies already pay a lot for AI, and usage is growing faster than almost anything we track. But the price of AI work is falling almost as fast, so spending grows much more slowly than usage. Today, what end customers pay covers roughly 58–81% of the yearly cost of the AI hardware already installed. Big Tech&amp;rsquo;s other businesses and the capital markets fill the rest.&lt;/p&gt;
&lt;p&gt;Whether that gap closes depends mostly on one thing: delegated AI agents. A chatbot answers when you ask. A delegated agent takes a goal, such as &amp;ldquo;reconcile these invoices&amp;rdquo; or &amp;ldquo;watch my inbox and draft replies,&amp;rdquo; and works on it by itself, often for hours, on a cloud computer of its own.&lt;/p&gt;
&lt;h2 id="five-things-to-take-away"&gt;Five things to take away
&lt;/h2&gt;&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Usage is booming; spending is tied to price.&lt;/strong&gt; In Ramp&amp;rsquo;s US business data, the price per token fell 42% from Aug 2 to Oct 11, usage rose 65%, and spending rose 1.5%. Google&amp;rsquo;s API traffic grew 37.5% in one quarter.&lt;sup id="fnref:1"&gt;&lt;a href="#fn:1" class="footnote-ref" role="doc-noteref"&gt;1&lt;/a&gt;&lt;/sup&gt;&lt;sup id="fnref:2"&gt;&lt;a href="#fn:2" class="footnote-ref" role="doc-noteref"&gt;2&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;There are seven kinds of paid AI demand, and their quality differs a lot.&lt;/strong&gt; Enterprise and agents are the core and grow fastest. Advertising is the biggest pool, but it mostly moves existing ad budgets. Consumers are wide but shallow. Commerce fees are close to zero.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Delegated agents are the growth engine.&lt;/strong&gt; An agent uses about 40 times the tokens of a chat user per day, by our estimate. Agents are spreading fast inside large companies (27% → 40% scaling them), but they have stalled at small firms (about 22%) and among the heaviest users.&lt;sup id="fnref:3"&gt;&lt;a href="#fn:3" class="footnote-ref" role="doc-noteref"&gt;3&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Payback today depends on who you ask.&lt;/strong&gt; Compute owners: yes, new rents are 1.6–2.6 times breakeven. Big Tech as a whole: yes, but its return on capital fell from 27.8% to 20.1%. End-customer payments alone: not yet, at 58–81% of the cost.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Payback tomorrow needs AI revenue to grow about 41% a year to 2030.&lt;/strong&gt; With prices falling 40–54% a year, usage must grow 2.4–3.1 times a year. Our base case for agent-driven usage in 2027, 2.5–3.0 times, sits right on that line. We put the odds of reaching the 2030 breakeven path at about 45%. Most misses would mean slower payback, not losses.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;Data are as of Oct 23–25, 2026. Probabilities and model outputs are our subjective estimates, and we label them as such.&lt;/p&gt;
&lt;h2 id="1-two-kinds-of-demand-volume-and-money"&gt;1. Two kinds of demand: volume and money
&lt;/h2&gt;&lt;p&gt;AI models read and write in small chunks of text called tokens. Tokens measure how much AI work gets done. Dollars measure how much anyone pays for it. Right now those two numbers tell different stories.&lt;/p&gt;
&lt;div class="fig-scroll" style="overflow-x:auto;-webkit-overflow-scrolling:touch;margin:1.4em 0"&gt;&lt;img src="https://koreainvestinsights.com/post-assets/who-pays-for-ai-demand-agents-capex-roi-2026-10-25/volume-vs-money.en.svg" alt="AI usage is booming while AI spending is tied to price: on Ramp's US business data from Aug 2 to Oct 11, price per token fell 42 percent, tokens used rose 65 percent and money spent rose 1.5 percent; Google API tokens per minute rose from 16 billion in Q1 2026 to 22 billion in Q2, up 37.5 percent in one quarter, about 3.6 times a year if it keeps up" loading="lazy" style="display:block;width:100%;min-width:720px;max-width:none;height:auto;margin:0 auto"&gt;&lt;/div&gt;
&lt;p&gt;When the price falls 42% and usage rises 65%, the bill barely moves. Economists call this a price elasticity of about 1. In plain words, every time AI gets cheaper, people use more of it, but so far only enough to keep the bill flat.&lt;sup id="fnref1:1"&gt;&lt;a href="#fn:1" class="footnote-ref" role="doc-noteref"&gt;1&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;The volume side shows no sign of slowing. Google processed 22 billion API tokens a minute in the second quarter, up from 16 billion. At that pace, volume grows about 3.6 times a year.&lt;sup id="fnref1:2"&gt;&lt;a href="#fn:2" class="footnote-ref" role="doc-noteref"&gt;2&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;Prices are also splitting in two. List prices for the top models rose 66.7% over a year, while mid-tier prices fell 44.5%.&lt;sup id="fnref:4"&gt;&lt;a href="#fn:4" class="footnote-ref" role="doc-noteref"&gt;4&lt;/a&gt;&lt;/sup&gt; On OpenRouter, open-weight models handle 72% of tokens but take only 27% of spending. Closed models handle 28% of tokens and take 73%. That makes revenue per token about seven times higher for closed models.&lt;sup id="fnref:5"&gt;&lt;a href="#fn:5" class="footnote-ref" role="doc-noteref"&gt;5&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;This split matters for chip investors. Memory chip demand follows tokens, but the money to build data centers follows dollars. A world of cheap, free and open-weight tokens can grow memory demand and still weaken the funding behind it.&lt;/p&gt;
&lt;h2 id="2-seven-kinds-of-paid-ai-demand"&gt;2. Seven kinds of paid AI demand
&lt;/h2&gt;&lt;p&gt;Not every dollar of AI revenue is the same. We sort paid demand by who pays and how. The seven groups overlap, so their sizes should not be added up.&lt;/p&gt;
&lt;div class="fig-scroll" style="overflow-x:auto;-webkit-overflow-scrolling:touch;margin:1.4em 0"&gt;&lt;img src="https://koreainvestinsights.com/post-assets/who-pays-for-ai-demand-agents-capex-roi-2026-10-25/seven-kinds.en.svg" alt="Seven kinds of paid AI demand from strongest to weakest: enterprise, paid by company IT budgets, Microsoft AI 37 billion dollars a year and Copilot 30 million seats, high quality; agents, mostly companies, agents in one of every three GitHub pull requests, Salesforce agent work up 97 percent quarter on quarter, the volume lever; API, Google 22 billion tokens a minute, mid-high quality for frontier only; advertising, Meta ads 63.1 billion dollars in Q3 up 26 percent, high quality but existing budgets; consumers, about 40 billion dollars a year worldwide, up 3.3 times, wide but shallow; subscriptions, flat plans 8 to 200 dollars a month, slowing, a ticket not a lever; commerce fees, near zero after the first 4 percent fee was pulled back, unproven" loading="lazy" style="display:block;width:100%;min-width:720px;max-width:none;height:auto;margin:0 auto"&gt;&lt;/div&gt;
&lt;p&gt;&lt;strong&gt;Enterprise: the core of paid demand.&lt;/strong&gt; Microsoft&amp;rsquo;s AI business reached an annual run-rate of about $37 billion in the March quarter, up 123%.&lt;sup id="fnref:6"&gt;&lt;a href="#fn:6" class="footnote-ref" role="doc-noteref"&gt;6&lt;/a&gt;&lt;/sup&gt; Microsoft 365 Copilot passed 30 million paid seats in July, and net seat additions more than doubled from the prior quarter.&lt;sup id="fnref:7"&gt;&lt;a href="#fn:7" class="footnote-ref" role="doc-noteref"&gt;7&lt;/a&gt;&lt;/sup&gt; Even at the $30 list price, 30 million seats come to about $10.8 billion a year, so most of Microsoft&amp;rsquo;s AI revenue likely comes from cloud usage rather than seats. Salesforce&amp;rsquo;s Agentforce passed $1.5 billion of annual recurring revenue, up 240%, though the product scope widened this quarter.&lt;sup id="fnref:8"&gt;&lt;a href="#fn:8" class="footnote-ref" role="doc-noteref"&gt;8&lt;/a&gt;&lt;/sup&gt; On Ramp, 55.9% of US businesses now pay for AI.&lt;sup id="fnref2:1"&gt;&lt;a href="#fn:1" class="footnote-ref" role="doc-noteref"&gt;1&lt;/a&gt;&lt;/sup&gt; The weak spot is proof: in one survey, only 9% of companies said they can measure the return on their token spending with high confidence.&lt;sup id="fnref:9"&gt;&lt;a href="#fn:9" class="footnote-ref" role="doc-noteref"&gt;9&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Agents: the volume lever.&lt;/strong&gt; Agents get their own section below, because they decide most of the growth math.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;API: the fastest volume, and the front line of price cuts.&lt;/strong&gt; Developers and companies buy model access by the token. This is where volume grows fastest and where prices fall hardest. Only frontier models hold their price.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Advertising: the biggest pool, but mostly old money moving.&lt;/strong&gt; Meta&amp;rsquo;s ad revenue was $63.06 billion in the third quarter, up 26%, as impressions rose 10% and the average price per ad rose 14%.&lt;sup id="fnref:10"&gt;&lt;a href="#fn:10" class="footnote-ref" role="doc-noteref"&gt;10&lt;/a&gt;&lt;/sup&gt; Meta did not say how much of that came from AI. New entrants are small by comparison: ChatGPT ads reached about $1 billion of annualized revenue, roughly 0.4% of Meta&amp;rsquo;s ad run-rate.&lt;sup id="fnref:11"&gt;&lt;a href="#fn:11" class="footnote-ref" role="doc-noteref"&gt;11&lt;/a&gt;&lt;/sup&gt; In the US, ad spending has stayed near 1.3% of GDP for decades, through every media shift. So AI advertising mostly shifts share between incumbents. It funds their capex rather than creating a new market.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Consumers: wide but shallow.&lt;/strong&gt; Global consumer spending on AI is estimated at $40 billion in 2026, up from $12 billion. Usage barely widened, from 61% to 64% of US adults, so existing users simply paid more. Spread over about 2 billion users worldwide, that is roughly $20 a year each. The top 14% of spenders, at $100 or more a month, account for about 60%, and 55% of US AI users now pay something.&lt;sup id="fnref:12"&gt;&lt;a href="#fn:12" class="footnote-ref" role="doc-noteref"&gt;12&lt;/a&gt;&lt;/sup&gt; Large free alternatives keep a lid on prices: Google&amp;rsquo;s Gemini app has about 950 million monthly users.&lt;sup id="fnref2:2"&gt;&lt;a href="#fn:2" class="footnote-ref" role="doc-noteref"&gt;2&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Subscriptions: a ticket, not a lever.&lt;/strong&gt; Flat plans run from $8 to $200 a month. The heaviest users cost more to serve than they pay, so labs are moving to usage-based pricing, weekly caps and cheaper models. Subscriptions become the entry ticket, and the money comes from usage on top.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Commerce fees: close to zero so far.&lt;/strong&gt; The first attempt at a 4% in-chat checkout fee was scaled back in March, reportedly because merchants were slow to sign up and retailers protected their customer relationships.&lt;sup id="fnref:13"&gt;&lt;a href="#fn:13" class="footnote-ref" role="doc-noteref"&gt;13&lt;/a&gt;&lt;/sup&gt; Meta and Walmart have not disclosed fees for agent shopping. If US agent-driven commerce reaches $300–500 billion by 2030 and fees run 2–4%, that is $6–20 billion a year. That is an assumption, and it will not pay for 2027&amp;rsquo;s data centers.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;One more buyer sits underneath: compute rental.&lt;/strong&gt; Labs and companies rent computing power, and that market is sold out. Amazon says its 2026 capacity is sold and 2027 is &amp;ldquo;largely reserved.&amp;rdquo;&lt;sup id="fnref:14"&gt;&lt;a href="#fn:14" class="footnote-ref" role="doc-noteref"&gt;14&lt;/a&gt;&lt;/sup&gt; Cloud backlogs add up to about $2.35 trillion: Microsoft $678 billion, Oracle $664 billion, Google $514 billion and AWS $496 billion. At least 37% of that is tied to two frontier labs.&lt;sup id="fnref1:14"&gt;&lt;a href="#fn:14" class="footnote-ref" role="doc-noteref"&gt;14&lt;/a&gt;&lt;/sup&gt; One-year rental contracts for H100 GPUs rose about 38% from their October 2025 low to March 2026.&lt;sup id="fnref:15"&gt;&lt;a href="#fn:15" class="footnote-ref" role="doc-noteref"&gt;15&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;Three patterns run across all seven kinds:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Revenue per token differs about sevenfold.&lt;/strong&gt; Free, open-weight and flat-rate usage grows volume. Frontier enterprise use, agent usage billing and advertising grow revenue.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Billing is moving from seats to usage to outcomes.&lt;/strong&gt; GitHub Copilot revenue grew more than 60% in a quarter after it switched to usage-based pricing.&lt;sup id="fnref1:7"&gt;&lt;a href="#fn:7" class="footnote-ref" role="doc-noteref"&gt;7&lt;/a&gt;&lt;/sup&gt; Usage billing lets revenue follow volume.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Payment is concentrated.&lt;/strong&gt; Among consumers, the top 14% pay 60%. Among companies, the top tenth of firms use 8.3 times the output tokens per user of a typical firm.&lt;sup id="fnref:16"&gt;&lt;a href="#fn:16" class="footnote-ref" role="doc-noteref"&gt;16&lt;/a&gt;&lt;/sup&gt; On the supply side, at least 37% of cloud backlog sits with two labs.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id="3-the-growth-engine-delegated-agents"&gt;3. The growth engine: delegated agents
&lt;/h2&gt;&lt;p&gt;Our working thesis since early October is simple. AI is moving from raw model APIs to packaged agents. An API is like a raw material: powerful, but only skilled builders can use it. A delegated agent is a finished product: you sign up, give it a goal, and it works on a cloud computer of its own.&lt;/p&gt;
&lt;p&gt;This form suits both sides. Providers run the inference centrally, so they can control it and make it efficient. Users skip the setup and simply hand work over, which matters most for non-developers doing everyday office tasks. If that spreads beyond programmers, agent traffic grows by multiples, and it runs around the clock.&lt;/p&gt;
&lt;p&gt;The product form is now real. Between May and October, Meta, Google, xAI and others launched always-on agents that keep working after you close the app. The open question is how fast people and companies adopt them.&lt;/p&gt;
&lt;div class="fig-scroll" style="overflow-x:auto;-webkit-overflow-scrolling:touch;margin:1.4em 0"&gt;&lt;img src="https://koreainvestinsights.com/post-assets/who-pays-for-ai-demand-agents-capex-roi-2026-10-25/agents.en.svg" alt="Delegated agents, the engine that has to beat falling prices: tokens per daily user about 40 times chat; 78 percent of input is cached so compute rises about 6 times. Where it is growing: OpenRouter weekly tokens about 5 times from May to September, Salesforce agentic work units up 97 percent quarter on quarter, ServiceNow agentic deployments 9 times in 9 months, Microsoft customers using a trillion tokens a year 4 times, GitHub pull requests with an agent one in three. Where it is stuck: large firms scaling agents 27 to 40 percent, small firms flat at about 22 percent, top 1 percent spenders per employee down 3.1 percent from July, one engineer can watch 3 to 5 agent sessions, a free consumer agent has 10 million US downloads but about 2.5 million daily users" loading="lazy" style="display:block;width:100%;min-width:720px;max-width:none;height:auto;margin:0 auto"&gt;&lt;/div&gt;
&lt;h3 id="why-agents-change-the-demand-math"&gt;Why agents change the demand math
&lt;/h3&gt;&lt;p&gt;An agent does far more work per user than a chat. By our estimate, a delegated agent uses about 1.86 million tokens a day per daily user, roughly 40 times a chat user. It reads web pages, calls tools, checks its own work and keeps going. About 78% of its input is reused from a cache, so the computing load rises about 6 times rather than 40 times.&lt;sup id="fnref:17"&gt;&lt;a href="#fn:17" class="footnote-ref" role="doc-noteref"&gt;17&lt;/a&gt;&lt;/sup&gt; For coding, the gap is wider: agentic coding can use around 1,000 times the tokens of a coding chat, according to one research preprint.&lt;/p&gt;
&lt;p&gt;Heavy users show where this leads. In OpenAI&amp;rsquo;s enterprise data, users at the top tenth of companies produce 8.3 times the output tokens of users at a typical company, up from 2.6 times in January.&lt;sup id="fnref1:16"&gt;&lt;a href="#fn:16" class="footnote-ref" role="doc-noteref"&gt;16&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;h3 id="where-it-is-growing"&gt;Where it is growing
&lt;/h3&gt;&lt;ul&gt;
&lt;li&gt;OpenRouter&amp;rsquo;s weekly token volume rose from about 25 trillion in May to 128.4 trillion in mid-September, about 5 times. Part of that is the platform gaining share.&lt;sup id="fnref1:5"&gt;&lt;a href="#fn:5" class="footnote-ref" role="doc-noteref"&gt;5&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;li&gt;Google processed more than 3,200 trillion tokens a month across its services in May, about 7 times a year earlier.&lt;sup id="fnref:18"&gt;&lt;a href="#fn:18" class="footnote-ref" role="doc-noteref"&gt;18&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;li&gt;Microsoft&amp;rsquo;s count of Foundry customers running at more than 1 trillion tokens a year quadrupled, and nearly 40 million agents were registered on Agent 365 within two months of launch. An agent is involved in one of every three GitHub pull requests.&lt;sup id="fnref2:7"&gt;&lt;a href="#fn:7" class="footnote-ref" role="doc-noteref"&gt;7&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;li&gt;Salesforce&amp;rsquo;s agentic work units rose 97% in one quarter.&lt;sup id="fnref1:8"&gt;&lt;a href="#fn:8" class="footnote-ref" role="doc-noteref"&gt;8&lt;/a&gt;&lt;/sup&gt; ServiceNow says its agentic deployments rose ninefold in nine months.&lt;sup id="fnref:19"&gt;&lt;a href="#fn:19" class="footnote-ref" role="doc-noteref"&gt;19&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;/ul&gt;
&lt;h3 id="where-it-is-stuck"&gt;Where it is stuck
&lt;/h3&gt;&lt;p&gt;Agent growth runs on two engines, and only one is fully running.&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;The depth engine, heavy users doing more, stalled in July.&lt;/strong&gt; On Ramp, spending per employee at the top 1% of companies is 3.1% below its July peak.&lt;sup id="fnref:20"&gt;&lt;a href="#fn:20" class="footnote-ref" role="doc-noteref"&gt;20&lt;/a&gt;&lt;/sup&gt; The limit is human attention. One engineer can supervise only about three to five agent sessions at a time. If every step needs a check, people become alert handlers, not supervisors.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;The breadth engine, more companies and more kinds of work, runs only at large firms.&lt;/strong&gt; In McKinsey&amp;rsquo;s 2026 survey, the share of large companies scaling AI agents rose from 27% to 40%, while small firms stayed around 22%.&lt;sup id="fnref1:3"&gt;&lt;a href="#fn:3" class="footnote-ref" role="doc-noteref"&gt;3&lt;/a&gt;&lt;/sup&gt; Gartner has predicted that more than 40% of agentic AI projects will be cancelled by the end of 2027.&lt;sup id="fnref:21"&gt;&lt;a href="#fn:21" class="footnote-ref" role="doc-noteref"&gt;21&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Consumers are mixed.&lt;/strong&gt; Meta&amp;rsquo;s free agent, Muse, reached about 10 million US downloads in its first 40 days, helped by heavy promotion across Meta&amp;rsquo;s own apps. Daily users reached about 2.5 million by Oct 16, and 23% of new users were still active after two weeks, above the 16% average for AI apps.&lt;sup id="fnref:22"&gt;&lt;a href="#fn:22" class="footnote-ref" role="doc-noteref"&gt;22&lt;/a&gt;&lt;/sup&gt; An earlier consumer agent, ChatGPT agent, reportedly lost about three-quarters of its users, with paid weekly users falling from about 4 million to under 1 million.&lt;sup id="fnref:23"&gt;&lt;a href="#fn:23" class="footnote-ref" role="doc-noteref"&gt;23&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;Billing is catching up with usage. Flat subscriptions make heavy agent use a cost, not revenue. Usage-based and outcome-based pricing turn it into revenue. The unit that matters is shifting from cost per token to cost per accepted, finished task.&lt;/p&gt;
&lt;h3 id="consumer-agents-are-small-for-compute-unless-intensity-jumps"&gt;Consumer agents are small for compute, unless intensity jumps
&lt;/h3&gt;&lt;p&gt;We sized personal cloud agents bottom-up. At 100 million daily users and today&amp;rsquo;s observed use, they need about 0.22–1.0 gigawatts and $9–52 billion of capex, about 0.6–4% of 2027 spending. The user count matters less than intensity. At the free-tier limit of about 14.3 million tokens a day, the same 100 million users would need 4.7–31% of 2027 capex.&lt;sup id="fnref1:17"&gt;&lt;a href="#fn:17" class="footnote-ref" role="doc-noteref"&gt;17&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;h3 id="agents-change-what-the-hardware-looks-like"&gt;Agents change what the hardware looks like
&lt;/h3&gt;&lt;p&gt;An agent&amp;rsquo;s own cloud computer needs CPUs and ordinary memory, not just AI chips. One measured personal-agent machine used 2 virtual CPUs and 7.7 GB of memory with no GPU. Data-center designs are shifting from one CPU per four GPUs toward one per two, or one to one. A one-million-token context can need about 328 GB of working memory, more than the 288 GB of high-bandwidth memory on a top GPU.&lt;sup id="fnref2:17"&gt;&lt;a href="#fn:17" class="footnote-ref" role="doc-noteref"&gt;17&lt;/a&gt;&lt;/sup&gt; Compared with chatbots, agents use about 50 times more CPU-side memory and 28–60 times more storage.&lt;/p&gt;
&lt;p&gt;That fits the last part of our thesis. Computing will keep getting more efficient. What keeps piling up is context and accumulated knowledge, which live in memory and storage.&lt;/p&gt;
&lt;h2 id="4-how-big-is-the-bill"&gt;4. How big is the bill?
&lt;/h2&gt;&lt;div class="fig-scroll" style="overflow-x:auto;-webkit-overflow-scrolling:touch;margin:1.4em 0"&gt;&lt;img src="https://koreainvestinsights.com/post-assets/who-pays-for-ai-demand-agents-capex-roi-2026-10-25/capex.en.svg" alt="Big Tech capex is still rising but more slowly: the four largest US cloud and platform companies spent 150 billion dollars in 2023, 226 billion in 2024 up 51 percent, 410 billion in 2025 up 81 percent, plan 725 billion in 2026 up 77 percent, and about 943 billion is estimated for 2027 up 30 percent; Morgan Stanley's tally including leases is about 1.58 trillion dollars in 2027, up 60 percent; capex took 97 percent of operating cash flow at five companies in Q2 2026" loading="lazy" style="display:block;width:100%;min-width:720px;max-width:none;height:auto;margin:0 auto"&gt;&lt;/div&gt;
&lt;p&gt;The four biggest US platforms went from about $150 billion of capex in 2023 to a planned $725 billion in 2026.&lt;sup id="fnref:24"&gt;&lt;a href="#fn:24" class="footnote-ref" role="doc-noteref"&gt;24&lt;/a&gt;&lt;/sup&gt; Growth is slowing, from +77% this year to somewhere between +30% and +60% in 2027, depending on whose tally you use. Even at the low end, these four companies would spend about $220 billion more in 2027 than in 2026.&lt;sup id="fnref:25"&gt;&lt;a href="#fn:25" class="footnote-ref" role="doc-noteref"&gt;25&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;The money no longer comes from spare cash. At five hyperscalers, capex took 97% of operating cash flow in the second quarter.&lt;sup id="fnref:26"&gt;&lt;a href="#fn:26" class="footnote-ref" role="doc-noteref"&gt;26&lt;/a&gt;&lt;/sup&gt; Meta&amp;rsquo;s free cash flow turned negative in the third quarter.&lt;sup id="fnref1:10"&gt;&lt;a href="#fn:10" class="footnote-ref" role="doc-noteref"&gt;10&lt;/a&gt;&lt;/sup&gt; Slower growth is already expected. What breaks this kind of cycle is not slower growth. It is a failure to refinance, which we covered in our &lt;a class="link" href="https://koreainvestinsights.com/post/credit-squeeze-case-s2-reading-and-path-2026-10-24/" &gt;credit-squeeze post&lt;/a&gt;.&lt;/p&gt;
&lt;h2 id="5-does-ai-pay-for-itself-today"&gt;5. Does AI pay for itself today?
&lt;/h2&gt;&lt;p&gt;The honest answer is &amp;ldquo;partly.&amp;rdquo; It depends on which layer you look at.&lt;/p&gt;
&lt;div class="fig-scroll" style="overflow-x:auto;-webkit-overflow-scrolling:touch;margin:1.4em 0"&gt;&lt;img src="https://koreainvestinsights.com/post-assets/who-pays-for-ai-demand-agents-capex-roi-2026-10-25/roi-today.en.svg" alt="Does AI pay for itself today, three layers: compute owners, yes, new contracts 20 to 25 million dollars per megawatt a year against breakeven of 9.5 to 12.9 million, 1.6 to 2.6 times; Big Tech as a whole, yes but falling, return on invested capital 27.8 to 20.1 percent against a 12 percent hurdle; end customers alone, not yet, yearly cost of installed AI kit about 359 billion dollars against 207 to 290 billion covered by paid AI, 58 to 81 percent; the gap of about 70 to 150 billion dollars a year is filled by Big Tech's other businesses and by capital markets" loading="lazy" style="display:block;width:100%;min-width:720px;max-width:none;height:auto;margin:0 auto"&gt;&lt;/div&gt;
&lt;p&gt;&lt;strong&gt;Compute owners: yes.&lt;/strong&gt; Estimates of what one gigawatt of AI data center costs range from about $35–40 billion to $50 billion or more, depending on what is counted. We use $37–50 billion.&lt;sup id="fnref:27"&gt;&lt;a href="#fn:27" class="footnote-ref" role="doc-noteref"&gt;27&lt;/a&gt;&lt;/sup&gt; Suppose the chips are 70% of that and last five years, the buildings are 30% and last 20 years, the owner wants a 12% return, and running costs are $1–1.5 million per megawatt a year. Then breakeven rent is about $9.5–12.9 million per megawatt a year. Nebius says its second-quarter bookings averaged more than $20 million of revenue per megawatt, and its largest contracts ran $20–25 million per megawatt a year. That is about 1.6–2.6 times breakeven.&lt;sup id="fnref:28"&gt;&lt;a href="#fn:28" class="footnote-ref" role="doc-noteref"&gt;28&lt;/a&gt;&lt;/sup&gt; By our tracking, rents on older GPUs have also risen more than 30% since the first quarter, a sign that chips stay useful longer than the three years bears fear. CoreWeave&amp;rsquo;s fleet average, about $10 million per megawatt by one broker&amp;rsquo;s estimate, sits near breakeven. Timing and chip generation decide who earns the most.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Big Tech as a whole: yes, but the return is falling.&lt;/strong&gt; The five largest hyperscalers&amp;rsquo; return on invested capital fell from about 27.8% to 20.1% between late 2024 and March 2026.&lt;sup id="fnref:29"&gt;&lt;a href="#fn:29" class="footnote-ref" role="doc-noteref"&gt;29&lt;/a&gt;&lt;/sup&gt; Most of that drop comes from spending running ahead of revenue. The return is still well above a 12% hurdle.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;End customers alone: not yet.&lt;/strong&gt; Here is the math in four steps.&lt;/p&gt;
&lt;ol&gt;
&lt;li&gt;&lt;strong&gt;Installed base.&lt;/strong&gt; About $931 billion of AI capex over the past three and a half years plus about $500 billion in the second half of 2026 comes to about $1.43 trillion.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Yearly cost.&lt;/strong&gt; Depreciation of about $222 billion plus a 12% return on the remaining book value of about $1.14 trillion, or about $137 billion, adds up to about $359 billion a year.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;What customers pay.&lt;/strong&gt; AI revenue is estimated to reach an annual run-rate of $414 billion by the end of 2026.&lt;sup id="fnref:30"&gt;&lt;a href="#fn:30" class="footnote-ref" role="doc-noteref"&gt;30&lt;/a&gt;&lt;/sup&gt; After staff, sales and revenue-sharing costs, about 50–70% of that is left to pay for hardware: $207–290 billion.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Coverage.&lt;/strong&gt; That covers about 58–81% of the yearly cost, leaving a gap of about $70–150 billion a year.&lt;sup id="fnref:31"&gt;&lt;a href="#fn:31" class="footnote-ref" role="doc-noteref"&gt;31&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;/ol&gt;
&lt;p&gt;Two pockets fill the gap. One is the cash from Big Tech&amp;rsquo;s existing businesses, which shows up as a falling return on capital. The other is equity and debt that labs and neoclouds raise against future demand. In plain words, the landlords are doing fine and the big tenants are paying. The end customers do not yet pay the full rent, and investors cover the difference in the hope that demand keeps growing.&lt;/p&gt;
&lt;h2 id="6-can-it-pay-off-tomorrow"&gt;6. Can it pay off tomorrow?
&lt;/h2&gt;&lt;div class="fig-scroll" style="overflow-x:auto;-webkit-overflow-scrolling:touch;margin:1.4em 0"&gt;&lt;img src="https://koreainvestinsights.com/post-assets/who-pays-for-ai-demand-agents-capex-roi-2026-10-25/bar-to-clear.en.svg" alt="The bar to clear: AI revenue must grow about 41 percent a year to 2030; estimated 414 billion dollars at end-2026, our path 565 to 661 billion at end-2027, and 1.63 trillion needed for breakeven in 2030; one year of 2027 spending needs about 397 billion dollars a year back, almost all of today's AI revenue; if token prices fall 40 to 54 percent a year, usage must grow 2.4 to 3.1 times a year; Google API tokens are growing at about 3.6 times; agents use about 40 times the tokens of chat per daily user" loading="lazy" style="display:block;width:100%;min-width:720px;max-width:none;height:auto;margin:0 auto"&gt;&lt;/div&gt;
&lt;p&gt;Four conditions have to hold.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;AI revenue must grow about 41% a year for four years.&lt;/strong&gt; In our model, AI revenue needs to reach about $1.63 trillion in 2030 to cover the yearly cost of everything built from 2026 to 2030. From $414 billion, that is about 41% a year. Our channel-by-channel path puts the end-2027 run-rate at $565–661 billion, and about 60% of that increase comes from frontier labs&amp;rsquo; enterprise and agent sales. Even on that path, 2028–2030 still needs 35–42% a year.&lt;sup id="fnref1:31"&gt;&lt;a href="#fn:31" class="footnote-ref" role="doc-noteref"&gt;31&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;One year of spending needs almost all of today&amp;rsquo;s AI revenue.&lt;/strong&gt; Morgan Stanley counts about $1.58 trillion of 2027 spending across the largest spenders, including leases.&lt;sup id="fnref1:25"&gt;&lt;a href="#fn:25" class="footnote-ref" role="doc-noteref"&gt;25&lt;/a&gt;&lt;/sup&gt; Its yearly cost comes to about $397 billion, about 96% of today&amp;rsquo;s AI revenue run-rate. Payback on that vintage starts in 2028.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Volume must do the work, not price.&lt;/strong&gt; Like-for-like token prices fall 40–54% a year.&lt;sup id="fnref:32"&gt;&lt;a href="#fn:32" class="footnote-ref" role="doc-noteref"&gt;32&lt;/a&gt;&lt;/sup&gt; For spending to grow 41% a year, usage must grow 2.4–3.1 times a year. Because cheaper prices have so far only kept the bill flat, that extra usage has to come from new kinds of work. Delegated agents are the main candidate. Today&amp;rsquo;s pace is enough: Google&amp;rsquo;s API volume is growing at about 3.6 times a year.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Proof will be demanded in 2027–28.&lt;/strong&gt; Only 9% of companies say they can measure the return on token spending with high confidence, and for 74% token spending is still under 10% of the IT budget.&lt;sup id="fnref1:9"&gt;&lt;a href="#fn:9" class="footnote-ref" role="doc-noteref"&gt;9&lt;/a&gt;&lt;/sup&gt; If budgets grow as planned, the questions get sharper in the 2027–28 budget cycle, exactly when the 2027 vintage needs to start paying back.&lt;/p&gt;
&lt;h3 id="do-agents-grow-fast-enough"&gt;Do agents grow fast enough?
&lt;/h3&gt;&lt;p&gt;This is where the two halves of the story meet.&lt;/p&gt;
&lt;div class="fig-scroll" style="overflow-x:auto;-webkit-overflow-scrolling:touch;margin:1.4em 0"&gt;&lt;img src="https://koreainvestinsights.com/post-assets/who-pays-for-ai-demand-agents-capex-roi-2026-10-25/agent-bar.en.svg" alt="Will agent-driven usage grow fast enough, our 2027 cases for token volume in Q4 2027 versus Q4 2026: bear case, 30 percent, 1.8 to 2.2 times, top users saturate, price war, trust and review limits; base case, 50 percent, 2.5 to 3.0 times, spread through big-company platforms while small firms and consumers lag; bull case, 20 percent, 3.5 to 4.0 times, office work at large firms, usage billing, agents built into phones and PCs; needed range for 41 percent a year spending growth if prices fall 40 to 54 percent is 2.4 to 3.1 times; Google API tokens Q2 pace about 3.6 times" loading="lazy" style="display:block;width:100%;min-width:720px;max-width:none;height:auto;margin:0 auto"&gt;&lt;/div&gt;
&lt;p&gt;Our 2027 cases for token volume came from the agent research above. In the base case, agents spread through big-company platforms like Copilot and Agentforce while small firms and consumers lag, and volume grows 2.5–3.0 times. That lands inside the 2.4–3.1 times the payback math needs. The bull case, with agents doing office work across large firms and built into phones and PCs, clears the bar easily. The bear case, where top users saturate and trust and review limits bite, falls short.&lt;/p&gt;
&lt;p&gt;So the capex payback question is, to a large degree, a question about delegated agents. Do they spread beyond the heaviest users and the largest companies, and does billing follow the usage?&lt;/p&gt;
&lt;p&gt;Putting it together, we see about a 60% chance that AI revenue reaches $580 billion or more by the end of 2027, and about a 75% chance that growth then holds at 35–42% a year through 2030. Combined, that is about 45% for the 2030 breakeven path. All three numbers are subjective.&lt;sup id="fnref2:31"&gt;&lt;a href="#fn:31" class="footnote-ref" role="doc-noteref"&gt;31&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;h2 id="7-if-it-falls-short-who-gets-hurt"&gt;7. If it falls short, who gets hurt?
&lt;/h2&gt;&lt;p&gt;Missing the path mostly means slower payback, not outright losses. Gartner estimates that AI infrastructure needs $8.5 trillion of revenue by 2030 against a forecast of $3.6 trillion.&lt;sup id="fnref:33"&gt;&lt;a href="#fn:33" class="footnote-ref" role="doc-noteref"&gt;33&lt;/a&gt;&lt;/sup&gt; In our model, 70–90% of that gap reflects how quickly the hardware must pay back. On a normal depreciation schedule, the gap shrinks a lot.&lt;/p&gt;
&lt;p&gt;The pain would also fall unevenly.&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Big Tech, the cash layer:&lt;/strong&gt; return on capital falls further toward the hurdle. Existing businesses and investment-grade debt absorb it.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Frontier labs:&lt;/strong&gt; lower valuations and renegotiated compute commitments. They are the most directly exposed to paid-demand growth.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Neoclouds and Oracle, the borrowed layer:&lt;/strong&gt; credit losses. Their average rents sit near breakeven, their contracts are short and their capital is the most expensive.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id="8-what-it-means-for-memory-chips"&gt;8. What it means for memory chips
&lt;/h2&gt;&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;2026–27:&lt;/strong&gt; Rents are above breakeven and computing is sold out, so builders still want to add capacity. Memory volumes are set by contracts and supply limits, not by the slower growth of AI spending.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;2028:&lt;/strong&gt; Whether the 2027 vintage visibly pays back will decide how easily 2028 capex gets funded. That test overlaps with memory price talks in the second half of 2027.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;The mix shifts toward memory and storage.&lt;/strong&gt; Agents add demand for CPUs, server DRAM and SSDs, not only for high-bandwidth memory. By our estimate, 12–15% of agent inference spending goes to non-HBM parts in the base case, and 29–33% in the bull case.&lt;sup id="fnref3:17"&gt;&lt;a href="#fn:17" class="footnote-ref" role="doc-noteref"&gt;17&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id="9-what-to-watch"&gt;9. What to watch
&lt;/h2&gt;&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Oct 29–30, Microsoft, Alphabet, Amazon.&lt;/strong&gt; Copilot seats and the AI run-rate; whether capacity is still short of demand. Alphabet&amp;rsquo;s API tokens: growth of 30% or more a quarter is bullish, under 15% is bearish. AWS growth and backlog.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Nov 11 KST, CoreWeave third quarter.&lt;/strong&gt; Revenue per megawatt against our $9.5–12.9 million breakeven.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;About Nov 13, Ramp AI Index.&lt;/strong&gt; October spending up 5% or more from September would suggest volume is starting to beat price.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Mid-November, Muse daily users.&lt;/strong&gt; Falling below half of the peak would weaken the consumer-agent case.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Our longer thresholds.&lt;/strong&gt; AI revenue of $580 billion or more by the end of 2027 strengthens the case, and under $500 billion weakens it. New contract rents under $13 million per megawatt, or 15% or more of companies freezing or cutting AI budgets, would also weaken it.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id="our-read"&gt;Our read
&lt;/h2&gt;&lt;p&gt;AI pays for itself today at the layers that own computing and in Big Tech&amp;rsquo;s overall returns. It does not yet pay for itself from end customers alone. The future case is conditional, not hopeless: AI revenue has to grow about 41% a year for four years while prices keep falling, and current volume growth is fast enough to do that.&lt;/p&gt;
&lt;p&gt;The single most important variable is no longer whether people use AI. They clearly do. It is whether delegated agents turn cheaper tokens into more finished work, fast enough, outside the top 1% of users and the largest companies.&lt;/p&gt;
&lt;h2 id="your-turn"&gt;Your turn
&lt;/h2&gt;&lt;p&gt;Is your company using AI agents beyond a pilot yet, and who supervises them? Which of the seven kinds of demand do you think will carry the bill in 2028? Tell me in the comments. I read every one.&lt;/p&gt;
&lt;p&gt;If a friend keeps asking whether AI is a bubble, send them this post. Sharing it helps more investors find the numbers behind the debate.&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Interest note: We use AI models, including Anthropic&amp;rsquo;s Claude, in our research. Anthropic is among the frontier labs counted in the lab and backlog figures above. We cite public figures only and make no judgment about any single lab.&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Disclaimer: For research and information purposes only. Not investment advice. Names cited are for analytical illustration; readers should perform their own due diligence and consult licensed advisors before any investment decision.&lt;/em&gt;&lt;/p&gt;
&lt;div class="footnotes" role="doc-endnotes"&gt;
&lt;hr&gt;
&lt;ol&gt;
&lt;li id="fn:1"&gt;
&lt;p&gt;&lt;a class="link" href="https://ramp.com/data/ai-index-oct-2026" target="_blank" rel="noopener"
 &gt;Ramp AI Index, October 2026&lt;/a&gt;, 2026-10-14: paid AI adoption 55.9%. Token price, volume and spending for Aug 2 to Oct 11 come from the weekly token index in the edition&amp;rsquo;s charts (volume 744 → 1,225, spending 605 → 614, price per million tokens $1.10 → $0.64). That index draws on Ramp&amp;rsquo;s token-spend management product, not card data, and Ramp says the sample is not representative.&amp;#160;&lt;a href="#fnref:1" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:1" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref2:1" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:2"&gt;
&lt;p&gt;&lt;a class="link" href="https://blog.google/company-news/inside-google/message-ceo/alphabet-earnings-q2-2026/" target="_blank" rel="noopener"
 &gt;Alphabet second-quarter 2026 results, CEO remarks&lt;/a&gt;, 2026-07-22: 22 billion API tokens per minute, up from 16 billion a quarter earlier; Google Cloud revenue +82%; Cloud backlog $514 billion. The annual pace is our calculation.&amp;#160;&lt;a href="#fnref:2" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:2" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref2:2" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:3"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.mckinsey.com/~/media/mckinsey/business%20functions/quantumblack/our%20insights/the%20state%20of%20ai/the-state-of-ai-in-2026-on-the-road-to-roi.pdf" target="_blank" rel="noopener"
 &gt;McKinsey, The state of AI in 2026&lt;/a&gt;. The large-company figure is the share of respondents at companies with $1 billion or more in revenue that are scaling agents; small-firm figure as summarized by secondary sources.&amp;#160;&lt;a href="#fnref:3" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:3" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:4"&gt;
&lt;p&gt;BenchLM token price index, as reported in October 2026. Secondary.&amp;#160;&lt;a href="#fnref:4" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:5"&gt;
&lt;p&gt;&lt;a class="link" href="https://openrouter.ai/data/newsletter/the-token-state-of-play" target="_blank" rel="noopener"
 &gt;OpenRouter, The Token State of Play&lt;/a&gt;, 2026-09-22: weekly token volume and the spending mix by model type.&amp;#160;&lt;a href="#fnref:5" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:5" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:6"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.uctoday.com/unified-communications/microsoft-earnings-2026-ai-copilot-enterprise/" target="_blank" rel="noopener"
 &gt;UC Today&lt;/a&gt;, 2026-04-30, on Microsoft&amp;rsquo;s March-quarter AI run-rate. Secondary.&amp;#160;&lt;a href="#fnref:6" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:7"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.microsoft.com/en-us/investor/earnings/fy-2026-q4/press-release-webcast" target="_blank" rel="noopener"
 &gt;Microsoft fiscal fourth-quarter 2026 results&lt;/a&gt;, 2026-07-29, and the &lt;a class="link" href="https://www.fool.com/earnings/call-transcripts/2026/08/07/microsoft-msft-q4-2026-earnings-call-transcript/" target="_blank" rel="noopener"
 &gt;earnings call transcript&lt;/a&gt;: Copilot seats, commercial RPO of $678 billion, one in three GitHub pull requests involving an agent, GitHub Copilot revenue up more than 60% quarter on quarter after usage-based pricing, Foundry customers at a 1-trillion-token run-rate up 4x, and nearly 40 million agents on Agent 365.&amp;#160;&lt;a href="#fnref:7" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:7" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref2:7" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:8"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.salesforce.com/news/press-releases/2026/08/26/fy27-q2-earnings/" target="_blank" rel="noopener"
 &gt;Salesforce fiscal second-quarter 2027 results&lt;/a&gt;, 2026-08-26: Agentforce ARR above $1.5 billion, up more than 240%, and agentic work units up 97% quarter on quarter. The ARR definition widened this quarter.&amp;#160;&lt;a href="#fnref:8" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:8" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:9"&gt;
&lt;p&gt;&lt;a class="link" href="https://theaiinnovator.com/bloomberg-analyst-enterprise-ai-spending-soars-but-model-loyalty-is-weak/" target="_blank" rel="noopener"
 &gt;The AI Innovator&lt;/a&gt;, 2026-10-05, on a Bloomberg Intelligence survey of about 100 enterprise respondents. Secondary.&amp;#160;&lt;a href="#fnref:9" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:9" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:10"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/0001326801/000162828026066700/meta-09302026xexhibit991.htm" target="_blank" rel="noopener"
 &gt;Meta third-quarter 2026 results (SEC 8-K, Exhibit 99.1)&lt;/a&gt;, 2026-10-21.&amp;#160;&lt;a href="#fnref:10" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:10" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:11"&gt;
&lt;p&gt;&lt;a class="link" href="https://digiday.com/media-buying/openais-chatgpt-ads-business-hits-1-billion-run-rate-as-europe-gets-self-serve-access/" target="_blank" rel="noopener"
 &gt;Digiday&lt;/a&gt;, 2026-08-31: a monthly figure times 12. Secondary.&amp;#160;&lt;a href="#fnref:11" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:12"&gt;
&lt;p&gt;&lt;a class="link" href="https://menlovc.com/perspective/2026-the-state-of-consumer-ai/" target="_blank" rel="noopener"
 &gt;Menlo Ventures, 2026: The State of Consumer AI&lt;/a&gt;, 2026-09-16. Survey-based estimates. Menlo Ventures invests in AI companies.&amp;#160;&lt;a href="#fnref:12" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:13"&gt;
&lt;p&gt;Press reports in March 2026 on the scaling back of in-chat checkout. Secondary.&amp;#160;&lt;a href="#fnref:13" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:14"&gt;
&lt;p&gt;Amazon second-quarter 2026 earnings call, 2026-07-30 (AWS backlog and capacity comments); &lt;a class="link" href="https://www.oracle.com/news/announcement/q1fy27-earnings-release-2026-09-10/" target="_blank" rel="noopener"
 &gt;Oracle fiscal first-quarter 2027 results&lt;/a&gt;, 2026-09-10 (RPO); Microsoft and Google backlogs from the releases above. The 37% two-lab share is our lower-bound sum of reported contracts with mixed dates and definitions.&amp;#160;&lt;a href="#fnref:14" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:14" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:15"&gt;
&lt;p&gt;&lt;a class="link" href="https://semianalysis.com/gpu-pricing-index/" target="_blank" rel="noopener"
 &gt;SemiAnalysis GPU Pricing Index&lt;/a&gt;: H100 one-year contract index about $1.70 per GPU-hour in October 2025 and about $2.35 in March 2026, as recorded in our dataset. Secondary.&amp;#160;&lt;a href="#fnref:15" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:16"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.vktr.com/ai-news/openai-data-shows-top-enterprise-ai-users-now-consume-8x-more-than-typical-firms/" target="_blank" rel="noopener"
 &gt;VKTR, on OpenAI enterprise data&lt;/a&gt;, August 2026: top-decile firms versus a typical firm. Secondary.&amp;#160;&lt;a href="#fnref:16" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:16" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:17"&gt;
&lt;p&gt;Our personal-agent and delegated-agent sizing models (September–October 2026), built on published measurements of agent sessions and hardware specifications. Estimates, not company data. The coding multiple is from a research preprint.&amp;#160;&lt;a href="#fnref:17" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:17" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref2:17" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref3:17" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:18"&gt;
&lt;p&gt;Google I/O 2026 keynote, May 2026, as reported by &lt;a class="link" href="https://www.shacknews.com/article/149205/google-3-2-quadrillion-monthly-ai-tokens" target="_blank" rel="noopener"
 &gt;Shacknews&lt;/a&gt;: 3.2 quadrillion tokens a month, about 7x a year earlier.&amp;#160;&lt;a href="#fnref:18" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:19"&gt;
&lt;p&gt;&lt;a class="link" href="https://s205.q4cdn.com/537566246/files/doc_news/ServiceNow-Reports-Second-Quarter-2026-Financial-Results-2026.pdf" target="_blank" rel="noopener"
 &gt;ServiceNow second-quarter 2026 results&lt;/a&gt;, 2026-07-22: chief executive&amp;rsquo;s comment that agentic deployments increased ninefold in nine months.&amp;#160;&lt;a href="#fnref:19" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:20"&gt;
&lt;p&gt;Ramp AI Index, October 2026, as summarized by &lt;a class="link" href="https://explainx.ai/blog/ramp-ai-index-october-2026-opus-5-5-34-percent-spend-claim-adoption-flattening" target="_blank" rel="noopener"
 &gt;explainx&lt;/a&gt;: top-1% spending per employee $9,873 at the July peak and $9,568 in September. Secondary.&amp;#160;&lt;a href="#fnref:20" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:21"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027" target="_blank" rel="noopener"
 &gt;Gartner press release&lt;/a&gt;, 2025-06-25.&amp;#160;&lt;a href="#fnref:21" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:22"&gt;
&lt;p&gt;TechCrunch, citing Sensor Tower, October 2026: US downloads in the first 40 days, daily users on Oct 16, and 14-day retention against the AI-app average. Meta&amp;rsquo;s launch announcement was 2026-09-08. Secondary.&amp;#160;&lt;a href="#fnref:22" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:23"&gt;
&lt;p&gt;&lt;a class="link" href="https://the-decoder.com/chatgpt-agent-reportedly-lost-75-of-its-users-because-nobody-knew-what-it-was-actually-for/" target="_blank" rel="noopener"
 &gt;The Decoder&lt;/a&gt;, citing The Information, 2026. Secondary.&amp;#160;&lt;a href="#fnref:23" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:24"&gt;
&lt;p&gt;2023–2024 as compiled by &lt;a class="link" href="https://www.groundbrkr.com/p/the-second-derivative-why-no-one" target="_blank" rel="noopener"
 &gt;Groundbreaker&lt;/a&gt;, July 2026; 2025–2026 from &lt;a class="link" href="https://finance.yahoo.com/sectors/technology/article/meta-microsoft-amazon-and-alphabet-are-about-to-spend-a-shocking-amount-of-money-to-dominate-the-ai-era-115359575.html" target="_blank" rel="noopener"
 &gt;Yahoo Finance&lt;/a&gt; and other press compilations; 2027 is September consensus for Microsoft, Alphabet and Amazon plus the midpoint of Meta&amp;rsquo;s range. Secondary.&amp;#160;&lt;a href="#fnref:24" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:25"&gt;
&lt;p&gt;Morgan Stanley&amp;rsquo;s October 2026 note, as reported by &lt;a class="link" href="https://www.itiger.com/news/1138724255" target="_blank" rel="noopener"
 &gt;Tiger Brokers&lt;/a&gt;: about $1.58 trillion of 2027 spending including leases, up about 60%. Secondary.&amp;#160;&lt;a href="#fnref:25" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:25" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:26"&gt;
&lt;p&gt;Our reconstruction from SEC XBRL filings for five hyperscalers: capex divided by operating cash flow, second quarter of 2026.&amp;#160;&lt;a href="#fnref:26" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:27"&gt;
&lt;p&gt;Bernstein, about $35–40 billion per gigawatt (&lt;a class="link" href="https://www.remio.ai/post/bernstein-ai-data-center-costs-hit-39-5-billion-per-gw-but-depreciation-is-the-b" target="_blank" rel="noopener"
 &gt;summary&lt;/a&gt;); Foxconn, about $47 billion; Nvidia&amp;rsquo;s chief executive, $50 billion or more (&lt;a class="link" href="https://finance.yahoo.com/news/nvidias-jensen-math-50-billion-033104538.html" target="_blank" rel="noopener"
 &gt;Yahoo Finance&lt;/a&gt;). Definitions differ (racks only versus full sites), so the range is approximate. Secondary.&amp;#160;&lt;a href="#fnref:27" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:28"&gt;
&lt;p&gt;Nebius second-quarter 2026 shareholder letter (&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/1513845/000110465926094568/tm2622968d1_ex99-2.htm" target="_blank" rel="noopener"
 &gt;Form 6-K, Exhibit 99.2&lt;/a&gt;), 2026-08-12.&amp;#160;&lt;a href="#fnref:28" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:29"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.apollo.com/wealth/insights-news/insights/daily-spark/slowing-token-spend-shows-compute-demand-is-not-unlimited" target="_blank" rel="noopener"
 &gt;Apollo Daily Spark, &amp;ldquo;Slowing Token Spend Shows Compute Demand Isn&amp;rsquo;t Unlimited&amp;rdquo;&lt;/a&gt;, 2026-10-20. The ROIC values are read from its chart (about ±0.5 percentage point).&amp;#160;&lt;a href="#fnref:29" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:30"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.techinvestments.io/p/ais-attractive-irr-anthropic-asml" target="_blank" rel="noopener"
 &gt;Tech Fund, &amp;ldquo;AI&amp;rsquo;s attractive IRR&amp;rdquo;&lt;/a&gt;, 2026-10-18: end-2026 AI annual recurring revenue estimate across labs, advertising, cloud and apps. Third-party estimate.&amp;#160;&lt;a href="#fnref:30" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:31"&gt;
&lt;p&gt;Our payback model: chips 70% of cost over five years, buildings 30% over 20 years, 12% required return on 80% of book value, contribution margin of 50–70%. Capex path 2026–2030 of $1.0, 1.2, 1.4, 1.6 and 1.8 trillion. Probabilities are subjective. A three-year chip life raises the bar; an 8.5% cost of capital lowers it.&amp;#160;&lt;a href="#fnref:31" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:31" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref2:31" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:32"&gt;
&lt;p&gt;Silicon Data token price index (about −54% a year) and other like-for-like price series. Secondary.&amp;#160;&lt;a href="#fnref:32" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:33"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.gartner.com/en/newsroom/press-releases/2026-10-14-gartner-says-over-us-dollars-three-and-a-half-trillion-of-artificial-intelligence-infrastructure-spending-could-be-exposed-to-impairment-by-the-end-of-2030" target="_blank" rel="noopener"
 &gt;Gartner press release&lt;/a&gt;, 2026-10-14; the revenue figures as reported by &lt;a class="link" href="https://cryptobriefing.com/gartner-ai-infrastructure-write-down-risk/" target="_blank" rel="noopener"
 &gt;Crypto Briefing&lt;/a&gt;. Secondary for the figures.&amp;#160;&lt;a href="#fnref:33" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;/ol&gt;
&lt;/div&gt;</description></item><item><title>The Credit-Squeeze Case (S2): Demand Holds, Money Gets Dearer, and the Weakest Borrowers Pay First</title><link>https://koreainvestinsights.com/post/credit-squeeze-case-s2-reading-and-path-2026-10-24/</link><pubDate>Sat, 24 Oct 2026 16:20:00 +0900</pubDate><guid>https://koreainvestinsights.com/post/credit-squeeze-case-s2-reading-and-path-2026-10-24/</guid><description>&lt;img src="https://koreainvestinsights.com/" alt="Featured image of post The Credit-Squeeze Case (S2): Demand Holds, Money Gets Dearer, and the Weakest Borrowers Pay First" /&gt;&lt;p&gt;On Oct 22 two things happened at once. The 10-year Treasury yield fell 17bp. US high-yield spreads rose 13bp in a day to 340bp, 10bp from the 350bp line we treat as bear confirmation.&lt;sup id="fnref:1"&gt;&lt;a href="#fn:1" class="footnote-ref" role="doc-noteref"&gt;1&lt;/a&gt;&lt;/sup&gt; Yields down and spreads up is the signature of S2, our credit-squeeze case.&lt;/p&gt;
&lt;p&gt;We put S2 at 15% (subjective, Oct 24). This post reads the market as if S2 is the path. In S2, AI demand stays real and Big Tech keeps spending. What changes is the price of money. It does not get cut off; it gets dearer, starting with the weakest borrowers. The question for chip investors is how far up that chain the stress climbs before it reaches memory.&lt;/p&gt;
&lt;p&gt;A companion post reads the same market through our base case, the rebound under a hawkish hold (S1).&lt;/p&gt;
&lt;h2 id="five-things-to-take-away"&gt;Five things to take away
&lt;/h2&gt;&lt;ul&gt;
&lt;li&gt;Under S2, demand holds but funding gets more expensive rung by rung. Hyperscaler-backed data-center bonds pay about 5.7–7.5%; CoreWeave&amp;rsquo;s unsecured 2032 notes yield about 15%.&lt;sup id="fnref:2"&gt;&lt;a href="#fn:2" class="footnote-ref" role="doc-noteref"&gt;2&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;li&gt;The stress is already climbing. High-yield spreads rose from 260bp on Aug 28 to 340bp. Oracle&amp;rsquo;s five-year CDS was reported near 315–322bp, from 144bp at the start of 2026. CoreWeave&amp;rsquo;s 2032 notes traded at 76–81 cents.&lt;sup id="fnref1:1"&gt;&lt;a href="#fn:1" class="footnote-ref" role="doc-noteref"&gt;1&lt;/a&gt;&lt;/sup&gt;&lt;sup id="fnref:3"&gt;&lt;a href="#fn:3" class="footnote-ref" role="doc-noteref"&gt;3&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;li&gt;S2 is only 15% because the broad market is still open. Investment-grade spreads are 89bp, below our 100bp bear line, and the BB index yields 6.85%, below 7.5%.&lt;sup id="fnref2:1"&gt;&lt;a href="#fn:1" class="footnote-ref" role="doc-noteref"&gt;1&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;li&gt;Memory feels S2 in 2028 orders, not in 2026–27 profits. Oracle and CoreWeave together are about $110–135 billion of 2027 spending, or 20–25% of the expected 2027 capex increase.&lt;sup id="fnref:4"&gt;&lt;a href="#fn:4" class="footnote-ref" role="doc-noteref"&gt;4&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;li&gt;In price, S2 means good earnings, a weak reaction and a failed bounce. KOSPI already sits 0.8% below its Jul 30 close and SK hynix 4.5% below.&lt;sup id="fnref:5"&gt;&lt;a href="#fn:5" class="footnote-ref" role="doc-noteref"&gt;5&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;Prices are Oct 23 closes; Korean closes are final prints after the 20:00 KST extended session. Probabilities are our subjective estimates. Most credit prices below come from press reports, and we label them as secondary.&lt;/p&gt;
&lt;h2 id="how-today-reads-under-s2"&gt;How today reads under S2
&lt;/h2&gt;&lt;h3 id="the-money-is-not-cut-off-it-gets-dearer-by-rung"&gt;The money is not cut off, it gets dearer by rung
&lt;/h3&gt;&lt;div class="fig-scroll" style="overflow-x:auto;-webkit-overflow-scrolling:touch;margin:1.4em 0"&gt;&lt;img src="https://koreainvestinsights.com/post-assets/credit-squeeze-case-s2-reading-and-path-2026-10-24/s2-ladder.en.svg" alt="The AI funding ladder: hyperscaler-backed data-center bonds rated Baa2 to A+ pay 5.7 to 7.5 percent; CoreWeave lease-backed paper about 10 percent; an unrated GPU-backed term loan about 11.6 percent; CoreWeave unsecured 2032 notes about 15 percent; for reference the US 10-year is 5.0 percent and the BB index yields 6.85 percent" loading="lazy" style="display:block;width:100%;min-width:720px;max-width:none;height:auto;margin:0 auto"&gt;&lt;/div&gt;
&lt;p&gt;The AI build-out is not funded by one pool of money. It is a ladder.&lt;/p&gt;
&lt;p&gt;At the top, data-center bonds backed by a hyperscaler lease or guarantee pay roughly 5.7–7.5%, close to Treasuries plus an investment-grade spread. One rung down, paper backed by CoreWeave leases pays about 10%. An unrated term loan backed by GPUs pays about 11.6%. At the bottom, CoreWeave&amp;rsquo;s unsecured 2032 notes yield about 15%.&lt;sup id="fnref1:2"&gt;&lt;a href="#fn:2" class="footnote-ref" role="doc-noteref"&gt;2&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;S2 says the ladder gets steeper from the bottom. Borrowers who rely on the bottom rungs, the neoclouds and the debt-funded part of the Oracle build, feel it first. A hyperscaler with AA credit barely notices. The first thing to shrink is the share of new capacity that neoclouds can fund, not total hyperscaler capex.&lt;/p&gt;
&lt;h3 id="the-stress-is-climbing"&gt;The stress is climbing
&lt;/h3&gt;&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;High-yield spreads.&lt;/strong&gt; 260bp on Aug 28, 283bp on Jan 2, 340bp on Oct 22, including a 13bp jump that day.&lt;sup id="fnref3:1"&gt;&lt;a href="#fn:1" class="footnote-ref" role="doc-noteref"&gt;1&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Oracle.&lt;/strong&gt; Five-year CDS reported near 315–322bp on Oct 19, up from 144bp at the start of 2026. We have not seen the original data, so treat it as secondary.&lt;sup id="fnref1:3"&gt;&lt;a href="#fn:3" class="footnote-ref" role="doc-noteref"&gt;3&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;CoreWeave.&lt;/strong&gt; The 9.625% 2032 notes traded at 76–81 cents on Oct 16–19, a yield near 15%. Five-year CDS was about 1,090bp on Oct 19 after peaking at 1,194.5bp on Oct 15.&lt;sup id="fnref2:3"&gt;&lt;a href="#fn:3" class="footnote-ref" role="doc-noteref"&gt;3&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Even the top names.&lt;/strong&gt; Meta&amp;rsquo;s CDS closed at a record 114.79bp on Oct 22.&lt;sup id="fnref:6"&gt;&lt;a href="#fn:6" class="footnote-ref" role="doc-noteref"&gt;6&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Volume.&lt;/strong&gt; Hyperscalers have issued about $255 billion of unsecured bonds this year, about 30% of 2026 capex, by Morgan Stanley&amp;rsquo;s count.&lt;sup id="fnref:7"&gt;&lt;a href="#fn:7" class="footnote-ref" role="doc-noteref"&gt;7&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Hidden leverage.&lt;/strong&gt; S&amp;amp;P Global said it is &amp;ldquo;safe to assume there is more unreported debt.&amp;rdquo;&lt;sup id="fnref:8"&gt;&lt;a href="#fn:8" class="footnote-ref" role="doc-noteref"&gt;8&lt;/a&gt;&lt;/sup&gt; Microsoft&amp;rsquo;s own 10-K shows leases signed but not yet started of $329.1 billion at Jun 30, up from $92.7 billion a year earlier, and it now calls some contracts conditional.&lt;sup id="fnref:9"&gt;&lt;a href="#fn:9" class="footnote-ref" role="doc-noteref"&gt;9&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;The equity window.&lt;/strong&gt; Firmus pulled its ASX listing on Oct 9. Nscale&amp;rsquo;s planned New York listing reportedly targets about $35 billion, while some investors offered $16–19 billion.&lt;sup id="fnref:10"&gt;&lt;a href="#fn:10" class="footnote-ref" role="doc-noteref"&gt;10&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Private credit.&lt;/strong&gt; A $3.7 billion GPU-backed term loan to Volta was reported at about 95 cents and was not underwritten.&lt;sup id="fnref:11"&gt;&lt;a href="#fn:11" class="footnote-ref" role="doc-noteref"&gt;11&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;None of these alone is a crisis. Together they show the cost of capital rising at the weak end while AI demand news stays good.&lt;/p&gt;
&lt;h3 id="why-s2-is-still-only-15"&gt;Why S2 is still only 15%
&lt;/h3&gt;&lt;p&gt;The broad credit market has not joined in. That is the main reason S2 is not our base case.&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;Investment-grade spreads are 89bp, below our 100bp bear line and below the 2026 high of 94bp. The BB index yields 6.85%, below our 7.5% bear line.&lt;sup id="fnref4:1"&gt;&lt;a href="#fn:1" class="footnote-ref" role="doc-noteref"&gt;1&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;li&gt;Stronger borrowers still get money on decent terms. Nebius signed a $3.5 billion senior secured facility with Citi at SOFR plus 275bp, though it is a 364-day bridge.&lt;sup id="fnref:12"&gt;&lt;a href="#fn:12" class="footnote-ref" role="doc-noteref"&gt;12&lt;/a&gt;&lt;/sup&gt; CoreWeave&amp;rsquo;s DDTL 4.0 was rated investment grade at SOFR plus 225bp in March.&lt;sup id="fnref:13"&gt;&lt;a href="#fn:13" class="footnote-ref" role="doc-noteref"&gt;13&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;li&gt;The biggest buyer pays from cash. Microsoft generated about $67 billion of free cash flow in fiscal 2026.&lt;sup id="fnref1:9"&gt;&lt;a href="#fn:9" class="footnote-ref" role="doc-noteref"&gt;9&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;li&gt;The weak link is small next to the whole. Barclays estimates that halving CoreWeave&amp;rsquo;s capex would equal about 2% of total AI capex.&lt;sup id="fnref:14"&gt;&lt;a href="#fn:14" class="footnote-ref" role="doc-noteref"&gt;14&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id="how-far-s2-would-reach-memory"&gt;How far S2 would reach memory
&lt;/h2&gt;&lt;div class="fig-scroll" style="overflow-x:auto;-webkit-overflow-scrolling:touch;margin:1.4em 0"&gt;&lt;img src="https://koreainvestinsights.com/post-assets/credit-squeeze-case-s2-reading-and-path-2026-10-24/s2-chain.en.svg" alt="How S2 would reach memory makers in four steps: the weakest rung pays more, with high-yield spreads from 283bp on Jan 2 to 340bp on Oct 22 and CoreWeave 2032 notes at 76 to 81 cents; marginal capacity shrinks, with Oracle plus CoreWeave about 110 to 135 billion dollars of 2027 spending or 20 to 25 percent of the 2027 capex increase; conditional leases are cut site by site, such as the Microsoft and Nscale Loughton exit and 329 billion dollars of Microsoft leases not yet started; and the hit lands on 2028 memory orders rather than 2026 to 27 profits. Brakes today: investment-grade spread 89bp, BB yield 6.85 percent, Nebius 3.5 billion dollar Citi facility, Microsoft funding capex from cash" loading="lazy" style="display:block;width:100%;min-width:720px;max-width:none;height:auto;margin:0 auto"&gt;&lt;/div&gt;
&lt;p&gt;&lt;strong&gt;Step 1: the weakest rung pays more.&lt;/strong&gt; This is where we are now.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Step 2: marginal capacity shrinks.&lt;/strong&gt; Oracle and CoreWeave together account for about $110–135 billion of 2027 spending. That is about 20–25% of the roughly $550 billion increase expected for 2027. This is a sensitivity, not a forecast. If that chain cut its spending in half, the hit would be about 10–12% of the increase. A CoreWeave half-cut alone is about 3.6%.&lt;sup id="fnref1:4"&gt;&lt;a href="#fn:4" class="footnote-ref" role="doc-noteref"&gt;4&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Step 3: conditional leases break site by site.&lt;/strong&gt; Contracts that depend on delivery dates and power can be cancelled one site at a time. The Microsoft–Nscale exit from Loughton in October is the template: Microsoft called it a single site, and reports pointed to grid delays.&lt;sup id="fnref:15"&gt;&lt;a href="#fn:15" class="footnote-ref" role="doc-noteref"&gt;15&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Step 4: memory feels it in 2028.&lt;/strong&gt; Long-term agreements and contract prices lock in much of 2026–27. Under S2 the damage shows up in 2028 orders and in the multiple investors pay for memory profits, not in next quarter&amp;rsquo;s earnings.&lt;/p&gt;
&lt;p&gt;That is why S2 can look strange. Results can beat while stocks fall, because the market is pricing the cost of capital and the life of memory profits, not unit demand.&lt;/p&gt;
&lt;h2 id="what-s2-looks-like-this-week"&gt;What S2 looks like this week
&lt;/h2&gt;&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Oct 28, SK hynix.&lt;/strong&gt; Strong results with a weak reaction fit S2. So does a shareholder-return plan smaller than the KRW 30 trillion or more that brokers hope for.&lt;sup id="fnref:16"&gt;&lt;a href="#fn:16" class="footnote-ref" role="doc-noteref"&gt;16&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Oct 29 00:55 KST, Oracle investor day.&lt;/strong&gt; The test is the funding mix. Goldman&amp;rsquo;s frame for fiscal 2027 is about $20 billion of equity (done), $20–25 billion of customer prepayments and about $20 billion of debt.&lt;sup id="fnref:17"&gt;&lt;a href="#fn:17" class="footnote-ref" role="doc-noteref"&gt;17&lt;/a&gt;&lt;/sup&gt; An S2-consistent day: more debt or off-balance-sheet funding, a later date for positive free cash flow, and CDS still above 300bp.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Oct 29–30, Alphabet, Microsoft, Amazon.&lt;/strong&gt; Capex up, free cash flow down and a weak price reaction. Listen for funding wording, not demand wording. Demand is not the S2 problem.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Oct 30–Nov 7, the bond window.&lt;/strong&gt; Hyperscalers and partners usually issue after earnings. S2 shows up as wider new-issue concessions on big deals and high-yield spreads at 350bp or more.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;The Nscale listing.&lt;/strong&gt; A pulled or discounted deal supports S2.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Nov 11 07:00 KST, CoreWeave third quarter.&lt;/strong&gt; Financing plan, capex and spreads on new secured debt. Our weekly bear lines are the 2032 notes below 70 cents or CDS at 1,500bp or more.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;We use three credit gates for S2: high-yield spreads at 350bp or more, Oracle CDS at 300bp or more, and CoreWeave 2032 notes below 75 cents. We have not decided whether one gate or all three should move the probability, so we watch all three and say which ones are lit.&lt;/p&gt;
&lt;h2 id="what-s2-would-look-like-in-price"&gt;What S2 would look like in price
&lt;/h2&gt;&lt;div class="fig-scroll" style="overflow-x:auto;-webkit-overflow-scrolling:touch;margin:1.4em 0"&gt;&lt;img src="https://koreainvestinsights.com/post-assets/credit-squeeze-case-s2-reading-and-path-2026-10-24/s2-july.en.svg" alt="Retest the July lows is already here for KOSPI and SK hynix: KOSPI 5,593.56 on Jul 30 versus 5,550.30 on Oct 23, down 0.8 percent; SK hynix 1,322,000 won versus 1,263,000 won, down 4.5 percent; Samsung Electronics 207,000 won versus 213,500 won, up 3.1 percent. The July low held because company buybacks absorbed foreign selling into September; that buyer is gone for now" loading="lazy" style="display:block;width:100%;min-width:720px;max-width:none;height:auto;margin:0 auto"&gt;&lt;/div&gt;
&lt;p&gt;Our scenario table writes the S2 outcome as &amp;ldquo;good earnings, weak reaction; the rebound fails and the July lows are retested.&amp;rdquo; The retest is closer than it sounds. KOSPI closed Oct 23 at 5,550.30, 0.8% below its Jul 30 close. SK hynix is 4.5% below its July close. Samsung Electronics still has a 3.1% cushion.&lt;sup id="fnref1:5"&gt;&lt;a href="#fn:5" class="footnote-ref" role="doc-noteref"&gt;5&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;The July low lasted one day. On Jul 31 KOSPI rose 17.9% to 6,595.45, SK hynix closed at 1,718,000 won and Samsung Electronics at 262,500 won.&lt;sup id="fnref2:5"&gt;&lt;a href="#fn:5" class="footnote-ref" role="doc-noteref"&gt;5&lt;/a&gt;&lt;/sup&gt; Company buybacks did much of the buying under that rebound. Those programs have now ended: Samsung&amp;rsquo;s in early October and SK hynix&amp;rsquo;s on Oct 21.&lt;/p&gt;
&lt;p&gt;So under S2 the question is not a crash from here. It is whether a bounce can start without that buyer and while money for AI gets dearer. We do not set a price target for S2.&lt;/p&gt;
&lt;h2 id="what-would-push-s2-back"&gt;What would push S2 back
&lt;/h2&gt;&lt;ul&gt;
&lt;li&gt;High-yield spreads back at 330bp or less, and investment-grade at 85bp or less, after the bond window.&lt;/li&gt;
&lt;li&gt;Oracle leans on customer prepayments instead of debt, and its CDS stays below 300bp for five trading days.&lt;/li&gt;
&lt;li&gt;CoreWeave&amp;rsquo;s 2032 notes back above 80 cents.&lt;/li&gt;
&lt;li&gt;A neocloud listing that prices normally.&lt;/li&gt;
&lt;li&gt;The BB index yield below 6.5%.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id="our-read"&gt;Our read
&lt;/h2&gt;&lt;p&gt;Credit is the closest of our three bear-confirmation lines: high-yield spreads are 10bp from it. The other two, an annual cut to next year&amp;rsquo;s memory consensus and a Big Tech capex slowdown, have not fired. That is why S2 is 15% and not higher. Under S2, the most useful single tell is the post-earnings bond window. If big deals need wider concessions while results are strong, the market is repricing the cost of AI capital, and memory stocks will trade on 2028, not on 2027.&lt;/p&gt;
&lt;h2 id="your-turn"&gt;Your turn
&lt;/h2&gt;&lt;p&gt;Which rung do you think breaks first: the neocloud listings, Oracle&amp;rsquo;s funding mix, or CoreWeave&amp;rsquo;s bonds? And is an 89bp investment-grade spread telling the truth, or is it lagging? Tell me in the comments. I will check these lines again after CoreWeave reports on Nov 11.&lt;/p&gt;
&lt;p&gt;If this helped you think about the credit side of the AI trade, a like helps other investors find it.&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Disclaimer: For research and information purposes only. Not investment advice. Names cited are for analytical illustration; readers should perform their own due diligence and consult licensed advisors before any investment decision.&lt;/em&gt;&lt;/p&gt;
&lt;div class="footnotes" role="doc-endnotes"&gt;
&lt;hr&gt;
&lt;ol&gt;
&lt;li id="fn:1"&gt;
&lt;p&gt;Federal Reserve Bank of St. Louis: &lt;a class="link" href="https://fred.stlouisfed.org/series/BAMLH0A0HYM2" target="_blank" rel="noopener"
 &gt;ICE BofA US High Yield OAS&lt;/a&gt; 2.60% on 2026-08-28, 2.83% on 2026-01-02, 3.27% on 2026-10-21 and 3.40% on 2026-10-22; &lt;a class="link" href="https://fred.stlouisfed.org/series/BAMLC0A0CM" target="_blank" rel="noopener"
 &gt;US Corporate (IG) OAS&lt;/a&gt; 0.89%; &lt;a class="link" href="https://fred.stlouisfed.org/series/BAMLH0A1HYBBEY" target="_blank" rel="noopener"
 &gt;BB effective yield&lt;/a&gt; 6.85%; &lt;a class="link" href="https://fred.stlouisfed.org/series/DGS10" target="_blank" rel="noopener"
 &gt;10-year Treasury&lt;/a&gt; 5.13% on Oct 21 and 4.96% on Oct 22. Latest available on Oct 24.&amp;#160;&lt;a href="#fnref:1" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:1" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref2:1" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref3:1" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref4:1" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:2"&gt;
&lt;p&gt;Coupons and yields at issue (February–October 2026) or as last reported, compiled in our episode dataset from company releases and press. October terms may differ. The CoreWeave 2032 yield is from &lt;a class="link" href="https://www.cnbc.com/2026/10/19/coreweave-oracle-ai-credit-stress.html" target="_blank" rel="noopener"
 &gt;CNBC&lt;/a&gt;, 2026-10-19.&amp;#160;&lt;a href="#fnref:2" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:2" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:3"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.cnbc.com/2026/10/19/coreweave-oracle-ai-credit-stress.html" target="_blank" rel="noopener"
 &gt;CNBC, CoreWeave and Oracle credit stress&lt;/a&gt;, 2026-10-19: Oracle five-year CDS about 315bp and CoreWeave 2032 notes. A reprint of an FT report put Oracle at about 322bp (&lt;a class="link" href="https://www.kucoin.com/news/flash/ai-debt-concerns-drive-surge-in-default-insurance-trading" target="_blank" rel="noopener"
 &gt;KuCoin&lt;/a&gt;). CoreWeave CDS and the start-of-year Oracle level are from press reports. All secondary; we have not seen the original data.&amp;#160;&lt;a href="#fnref:3" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:3" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref2:3" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:4"&gt;
&lt;p&gt;Our calculation from Morgan Stanley&amp;rsquo;s seven-company 2027 capex estimate of $1.467 trillion (+60%, secondary) and press-reported spending plans for Oracle (about $70–95 billion) and CoreWeave (about $40 billion). Oracle&amp;rsquo;s fiscal year ends in May, so the calendar-2027 mapping is approximate. This is a sensitivity, not a share of contribution.&amp;#160;&lt;a href="#fnref:4" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:4" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:5"&gt;
&lt;p&gt;Naver Finance daily closes (KRX final prints), rechecked on Oct 24.&amp;#160;&lt;a href="#fnref:5" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:5" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref2:5" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:6"&gt;
&lt;p&gt;Reuters, 2026-10-22, citing S&amp;amp;P Global Market Intelligence (Markit) closing data. Secondary.&amp;#160;&lt;a href="#fnref:6" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:7"&gt;
&lt;p&gt;Morgan Stanley estimates as reported in the press, October 2026. Secondary.&amp;#160;&lt;a href="#fnref:7" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:8"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.933thedrive.com/2026/10/19/top-hyperscalers-likely-to-have-more-ai-linked-debt-than-now-known-sp-global-warns/" target="_blank" rel="noopener"
 &gt;Reuters via 93.3 The Drive&lt;/a&gt;, 2026-10-19.&amp;#160;&lt;a href="#fnref:8" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:9"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/789019/000119312526323660/msft-20260630.htm" target="_blank" rel="noopener"
 &gt;Microsoft Form 10-K for fiscal 2026&lt;/a&gt;: leases signed but not yet commenced. Free cash flow is operating cash flow minus capital expenditures from the same filing, our calculation.&amp;#160;&lt;a href="#fnref:9" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:9" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:10"&gt;
&lt;p&gt;Firmus withdrawal from its ASX filing, 2026-10-09. Nscale valuation figures are press reports. Secondary.&amp;#160;&lt;a href="#fnref:10" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:11"&gt;
&lt;p&gt;Press reports, October 2026. Secondary.&amp;#160;&lt;a href="#fnref:11" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:12"&gt;
&lt;p&gt;&lt;a class="link" href="https://nebius.com/newsroom/nebius-secures-3-5-billion-financing-facility-with-citi" target="_blank" rel="noopener"
 &gt;Nebius newsroom&lt;/a&gt;: $3.5 billion senior secured facility, 364 days, SOFR plus 275bp initially and up to 350bp. This is the facility limit, not the drawn amount.&amp;#160;&lt;a href="#fnref:12" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:13"&gt;
&lt;p&gt;CoreWeave company release, 2026-03-31: DDTL 4.0, $8.5 billion, rated A3, SOFR plus 225bp.&amp;#160;&lt;a href="#fnref:13" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:14"&gt;
&lt;p&gt;Barclays note, 2026-10-20, as reported in the press. Secondary.&amp;#160;&lt;a href="#fnref:14" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:15"&gt;
&lt;p&gt;Press reports and a Microsoft statement, 2026-10-18 to 2026-10-19. Secondary.&amp;#160;&lt;a href="#fnref:15" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:16"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.mt.co.kr/en/industry/2026/10/19/2026101913371799572" target="_blank" rel="noopener"
 &gt;Money Today&lt;/a&gt;, 2026-10-19, citing securities-industry expectations.&amp;#160;&lt;a href="#fnref:16" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:17"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.prnewswire.com/news-releases/oracle-investor-day-meeting-to-be-held-october-28-2026-302918062.html" target="_blank" rel="noopener"
 &gt;Oracle investor day notice&lt;/a&gt;: Oct 28, 10:55 CT. The Goldman funding frame is from press coverage of its note. Secondary.&amp;#160;&lt;a href="#fnref:17" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;/ol&gt;
&lt;/div&gt;</description></item><item><title>The Rebound Case (S1): The Business Went Up, the Stocks Went Down, and the Forced Sellers Are Mostly Done</title><link>https://koreainvestinsights.com/post/rebound-case-s1-reading-and-path-2026-10-24/</link><pubDate>Sat, 24 Oct 2026 16:10:00 +0900</pubDate><guid>https://koreainvestinsights.com/post/rebound-case-s1-reading-and-path-2026-10-24/</guid><description>&lt;img src="https://koreainvestinsights.com/" alt="Featured image of post The Rebound Case (S1): The Business Went Up, the Stocks Went Down, and the Forced Sellers Are Mostly Done" /&gt;&lt;p&gt;Our base case for the chip selloff is S1: a rebound under a hawkish Fed hold, at 50% (subjective, Oct 24). This post reads the market as if S1 is the path. It asks two questions. What does today look like through that lens? And if S1 plays out, in what order should things happen?&lt;/p&gt;
&lt;p&gt;The short version: the business went up while the stocks went down. DDR5 spot prices rose 11.7% from Sep 16 to Oct 23, and Korea&amp;rsquo;s chip exports for Oct 1–20 were up 222.5% from a year earlier. Over the same weeks SK hynix fell 31.4% from its October closing high.&lt;sup id="fnref:1"&gt;&lt;a href="#fn:1" class="footnote-ref" role="doc-noteref"&gt;1&lt;/a&gt;&lt;/sup&gt;&lt;sup id="fnref:2"&gt;&lt;a href="#fn:2" class="footnote-ref" role="doc-noteref"&gt;2&lt;/a&gt;&lt;/sup&gt; A gap that wide is what forced selling looks like, and the forced sellers are mostly done.&lt;/p&gt;
&lt;p&gt;S1 is not a promise of a straight line up. It is a rebound with a shakeout first and a ceiling from rates. A companion post reads the same market through the credit-squeeze case (S2).&lt;/p&gt;
&lt;h2 id="five-things-to-take-away"&gt;Five things to take away
&lt;/h2&gt;&lt;ul&gt;
&lt;li&gt;Under S1 the October drop was a crowded long book being forced out, not a demand break. The earnings layer for 2026–27 got stronger while the stocks fell.&lt;/li&gt;
&lt;li&gt;Most of the mechanical selling is done. Goldman&amp;rsquo;s model of CTA equity longs fell from about $125 billion to $39 billion, and foreign cash selling in Korea fell from KRW 27.0 trillion over Sep 24–Oct 15 to KRW 0.31 trillion over Oct 16–23.&lt;sup id="fnref:3"&gt;&lt;a href="#fn:3" class="footnote-ref" role="doc-noteref"&gt;3&lt;/a&gt;&lt;/sup&gt;&lt;sup id="fnref:4"&gt;&lt;a href="#fn:4" class="footnote-ref" role="doc-noteref"&gt;4&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;li&gt;S1 needs four things this week: SK hynix confirms 2027 contracts and a new buyback, Microsoft passes three of our four tests, Big Tech keeps capex up, and the Fed holds without hike language.&lt;/li&gt;
&lt;li&gt;Expect the rebound in a specific order: the most-shorted neoclouds cover first, memory rises later on long re-buying, and in Korea foreign buying shows up in Samsung Electronics before SK hynix.&lt;/li&gt;
&lt;li&gt;Rates set the ceiling. Futures put only about an 8% chance on an October hike but high odds on December, so S1 is a rebound with a lid.&lt;sup id="fnref:5"&gt;&lt;a href="#fn:5" class="footnote-ref" role="doc-noteref"&gt;5&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;Prices are Oct 23 closes; Korean closes are final prints after the 20:00 KST extended session. Probabilities are our subjective estimates.&lt;/p&gt;
&lt;h2 id="how-today-reads-under-s1"&gt;How today reads under S1
&lt;/h2&gt;&lt;h3 id="the-business-went-up-while-the-stocks-went-down"&gt;The business went up while the stocks went down
&lt;/h3&gt;&lt;div class="fig-scroll" style="overflow-x:auto;-webkit-overflow-scrolling:touch;margin:1.4em 0"&gt;&lt;img src="https://koreainvestinsights.com/post-assets/rebound-case-s1-reading-and-path-2026-10-24/s1-business.en.svg" alt="The business went up while the stocks went down: DDR5 spot up 11.7 percent, Korea chip exports up 222.5 percent year on year for Oct 1 to 20, Micron FY27 EPS consensus up 12.8 percent in 30 days, Samsung 2027 operating-profit consensus 554 to 561 trillion won, TSMC 2026 capex guide held at 60 to 64 billion dollars; while SK hynix fell 31.4 percent, Samsung 22.6 percent, Micron 20.0 percent and SOX 18.4 percent" loading="lazy" style="display:block;width:100%;min-width:720px;max-width:none;height:auto;margin:0 auto"&gt;&lt;/div&gt;
&lt;p&gt;Every line on the left moved the right way in October. DDR5 16Gb spot averaged $61.27 on Oct 23, up from $54.83 on Sep 16.&lt;sup id="fnref1:1"&gt;&lt;a href="#fn:1" class="footnote-ref" role="doc-noteref"&gt;1&lt;/a&gt;&lt;/sup&gt; Korea&amp;rsquo;s chip exports for the first 20 days of October reached $27.5 billion.&lt;sup id="fnref1:2"&gt;&lt;a href="#fn:2" class="footnote-ref" role="doc-noteref"&gt;2&lt;/a&gt;&lt;/sup&gt; Consensus for Samsung&amp;rsquo;s 2027 operating profit rose to KRW 561 trillion at FnGuide, and Micron&amp;rsquo;s FY27 EPS consensus rose 12.8% in 30 days.&lt;sup id="fnref:6"&gt;&lt;a href="#fn:6" class="footnote-ref" role="doc-noteref"&gt;6&lt;/a&gt;&lt;/sup&gt; TSMC kept its 2026 capex guide at $60–64 billion on Oct 15.&lt;sup id="fnref:7"&gt;&lt;a href="#fn:7" class="footnote-ref" role="doc-noteref"&gt;7&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;Demand behind those numbers is still booked ahead. Amazon said in July that its 2027 capacity is largely reserved and that a good part of 2028 is reserved too.&lt;sup id="fnref:8"&gt;&lt;a href="#fn:8" class="footnote-ref" role="doc-noteref"&gt;8&lt;/a&gt;&lt;/sup&gt; Samsung said five customers had signed long-term agreements, though its 60–70% allocation plan also counts deals not yet signed.&lt;sup id="fnref:9"&gt;&lt;a href="#fn:9" class="footnote-ref" role="doc-noteref"&gt;9&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;Under S1, the market sold the durability of memory profits beyond 2027 and the cost of AI capital, not today&amp;rsquo;s demand. The 2026–27 layer is barely priced; the 2028 layer is heavily priced.&lt;/p&gt;
&lt;h3 id="the-forced-sellers-are-mostly-done"&gt;The forced sellers are mostly done
&lt;/h3&gt;&lt;div class="fig-scroll" style="overflow-x:auto;-webkit-overflow-scrolling:touch;margin:1.4em 0"&gt;&lt;img src="https://koreainvestinsights.com/post-assets/rebound-case-s1-reading-and-path-2026-10-24/s1-flush.en.svg" alt="The forced sellers are mostly done: CTA equity longs fell from 125 billion to 39 billion dollars, fund managers calling semiconductors the most crowded trade fell from 82 to 45 percent, foreign net selling of KOSPI cash equities fell from 27.0 trillion won over Sep 24 to Oct 15 to 0.31 trillion won over Oct 16 to 23; still, hedge-fund gross leverage is 311.7 percent" loading="lazy" style="display:block;width:100%;min-width:720px;max-width:none;height:auto;margin:0 auto"&gt;&lt;/div&gt;
&lt;p&gt;Three groups did most of the selling, and each has largely stopped.&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Trend followers.&lt;/strong&gt; Goldman&amp;rsquo;s model of CTA equity longs fell about 69% to roughly $39 billion in the week of Oct 16.&lt;sup id="fnref1:3"&gt;&lt;a href="#fn:3" class="footnote-ref" role="doc-noteref"&gt;3&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Crowded fund managers.&lt;/strong&gt; In BofA&amp;rsquo;s survey, the share calling semiconductors the most crowded trade fell from 82% in July to 45% in October.&lt;sup id="fnref:10"&gt;&lt;a href="#fn:10" class="footnote-ref" role="doc-noteref"&gt;10&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Foreign investors in Korea.&lt;/strong&gt; They sold KRW 27.0 trillion of KOSPI cash equities from Sep 24 to Oct 15, then only KRW 0.31 trillion from Oct 16 to Oct 23. Samsung Electronics even saw foreign net buying of about 4.55 million shares in that second window.&lt;sup id="fnref1:4"&gt;&lt;a href="#fn:4" class="footnote-ref" role="doc-noteref"&gt;4&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;The Oct 22 drop also looked like positioning, not credit. SOX fell 6.7% that day, yet the equal-weight S&amp;amp;P 500 rose 0.7%, an investment-grade bond ETF rose 1.1% and the 10-year Treasury yield fell 17bp.&lt;sup id="fnref:11"&gt;&lt;a href="#fn:11" class="footnote-ref" role="doc-noteref"&gt;11&lt;/a&gt;&lt;/sup&gt; A credit event usually drags those down together.&lt;/p&gt;
&lt;p&gt;There is one loaded spring left. Hedge funds still run gross leverage of about 312%, near a record.&lt;sup id="fnref:12"&gt;&lt;a href="#fn:12" class="footnote-ref" role="doc-noteref"&gt;12&lt;/a&gt;&lt;/sup&gt; If longs fall while shorts rise, they could be forced to cut once more. That is why S1 comes with a shakeout risk.&lt;/p&gt;
&lt;h3 id="the-stress-sits-at-the-weak-end-not-with-memory-buyers"&gt;The stress sits at the weak end, not with memory buyers
&lt;/h3&gt;&lt;p&gt;S1 does not say credit is fine. It says the stress is concentrated. Investment-grade spreads are 89bp, still below our 100bp bear line, while the pressure sits with Oracle and the debt-funded neoclouds.&lt;sup id="fnref:13"&gt;&lt;a href="#fn:13" class="footnote-ref" role="doc-noteref"&gt;13&lt;/a&gt;&lt;/sup&gt; Those borrowers matter for 2028 orders, not for the 2026–27 profits already under contract. The companion S2 post looks at what happens if that changes.&lt;/p&gt;
&lt;h2 id="what-s1-needs-this-week"&gt;What S1 needs this week
&lt;/h2&gt;&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Oct 28, SK hynix.&lt;/strong&gt; Two things: a clear share of 2027 HBM4 volume under contract, and a new shareholder-return plan. Brokers expect funding of more than KRW 30 trillion.&lt;sup id="fnref:14"&gt;&lt;a href="#fn:14" class="footnote-ref" role="doc-noteref"&gt;14&lt;/a&gt;&lt;/sup&gt; Company buybacks did much of the buying under the July rebound, and the last program ended on Oct 21; a new one would restore that bottom buyer.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Oct 29, Microsoft.&lt;/strong&gt; Three of four tests: demand still exceeds supply; calendar 2026 capex is raised or the December quarter is at least the September quarter; leases signed but not yet started keep growing; Azure grows 45% or more. We put the odds near 42%.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Oct 29–30, Alphabet and Amazon.&lt;/strong&gt; Capex held or raised, and memory price increases absorbed without cutting build plans.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Oct 29, FOMC.&lt;/strong&gt; A hold without language that points to more hikes. Markets price only about an 8% chance of an October hike.&lt;sup id="fnref1:5"&gt;&lt;a href="#fn:5" class="footnote-ref" role="doc-noteref"&gt;5&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Oct 30–Nov 7, the bond window.&lt;/strong&gt; US high-yield spreads at 330bp or less with deals clearing. They were 340bp on Oct 22.&lt;sup id="fnref1:13"&gt;&lt;a href="#fn:13" class="footnote-ref" role="doc-noteref"&gt;13&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id="how-s1-would-unfold"&gt;How S1 would unfold
&lt;/h2&gt;&lt;div class="fig-scroll" style="overflow-x:auto;-webkit-overflow-scrolling:touch;margin:1.4em 0"&gt;&lt;img src="https://koreainvestinsights.com/post-assets/rebound-case-s1-reading-and-path-2026-10-24/s1-path.en.svg" alt="How S1 would unfold in four steps: Oct 28 to 30 earnings clear; Oct 29 to Nov 4 rebound window; Oct 30 to Nov 7 bond window clears; Nov 3 to 10 rates set the ceiling. What ends S1: high-yield 350bp or more with failed deals, a 2027 capex cut, inflation-type oil days, or foreign cash selling of 1 trillion won a day with futures buying" loading="lazy" style="display:block;width:100%;min-width:720px;max-width:none;height:auto;margin:0 auto"&gt;&lt;/div&gt;
&lt;p&gt;&lt;strong&gt;Step 1, Oct 28–30: earnings clear.&lt;/strong&gt; The first day that splits the paths is Oct 29. If the tests above pass, hedges start to come off.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Step 2, Oct 29–Nov 4: the rebound window.&lt;/strong&gt; The order matters. Short covering hits the most-shorted names first; neocloud shorts were about 20–26% of float.&lt;sup id="fnref:15"&gt;&lt;a href="#fn:15" class="footnote-ref" role="doc-noteref"&gt;15&lt;/a&gt;&lt;/sup&gt; Memory stocks are not heavily shorted, so they rise later and more slowly, on long re-buying. In Korea we expect foreign buying to show up in Samsung Electronics first and in SK hynix after its results. The fingerprint to look for is retail selling while foreigners buy. On Jul 31, the day after the July low, foreigners bought KRW 7.24 trillion of KOSPI stocks while individuals sold KRW 8.28 trillion.&lt;sup id="fnref:16"&gt;&lt;a href="#fn:16" class="footnote-ref" role="doc-noteref"&gt;16&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Step 3, Oct 30–Nov 7: the bond window clears.&lt;/strong&gt; Hyperscalers and their partners usually issue bonds after earnings. S1 needs those deals to clear without bigger concessions, with high-yield spreads back at 330bp or less and investment-grade at 85bp or less.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Step 4, Nov 3–10: rates set the ceiling.&lt;/strong&gt; The midterms, October jobs (Nov 6) and October CPI (Nov 10) decide how far the rebound runs. With a December hike still priced, a 10-year yield near 5% caps valuations.&lt;/p&gt;
&lt;h3 id="what-the-base-rates-say"&gt;What the base rates say
&lt;/h3&gt;&lt;p&gt;History supports a rebound but warns about the path. Since 1994, after a day when SOX fell 6% or more (97 cases), the index was higher ten days later 63% of the time. In the 33 cases that looked like a crowding unwind, it was higher 64% of the time, but 58% saw another 5% drop first.&lt;sup id="fnref:17"&gt;&lt;a href="#fn:17" class="footnote-ref" role="doc-noteref"&gt;17&lt;/a&gt;&lt;/sup&gt; Hedge funds&amp;rsquo; AI net flows also sit at levels last seen near the April and July lows; SOX rose 38.7% and 13.7% over the next 21 sessions in those two cases, though two cases prove little.&lt;sup id="fnref:18"&gt;&lt;a href="#fn:18" class="footnote-ref" role="doc-noteref"&gt;18&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;Korea needs a caveat. In 2025–26, after a KOSPI gain of 3% or more, the next 20 sessions averaged 0.0% and rose only 42% of the time. Rallies without foreign cash buying faded.&lt;sup id="fnref1:17"&gt;&lt;a href="#fn:17" class="footnote-ref" role="doc-noteref"&gt;17&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;So the S1 expectation is a rebound, likely after one more shakeout, led by the most-shorted names and capped by rates.&lt;/p&gt;
&lt;h2 id="what-ends-s1"&gt;What ends S1
&lt;/h2&gt;&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;High-yield spreads at 350bp or more with failed deals.&lt;/strong&gt; S1 turns into S2.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;One Big Tech company cuts 2027 capex.&lt;/strong&gt; That moves toward S3, the demand slowdown.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Inflation-type oil days.&lt;/strong&gt; Brent above $110 while inflation gauges rise keeps the macro-shock case (S4) at 20% or higher.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Korea&amp;rsquo;s flow turns.&lt;/strong&gt; Foreign cash selling of KRW 1 trillion or more in a day, or KRW 3 trillion over five sessions, together with futures buying.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;The earnings layer cracks.&lt;/strong&gt; An annual cut to next year&amp;rsquo;s memory consensus. Quarterly estimates for Samsung&amp;rsquo;s first two quarters of 2027 have already slipped about 2.6–2.7%, so this is worth watching.&lt;sup id="fnref1:6"&gt;&lt;a href="#fn:6" class="footnote-ref" role="doc-noteref"&gt;6&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id="our-read"&gt;Our read
&lt;/h2&gt;&lt;p&gt;S1 is our base case because the layer that pays memory makers got stronger while the price fell, and the sellers who drove October look mostly spent. The weak spot in this reading is credit: the S1 gate of 330bp on high-yield spreads is not met today. If the bond window clears, the rebound case gets much stronger. If it fails, read the S2 post next.&lt;/p&gt;
&lt;h2 id="your-turn"&gt;Your turn
&lt;/h2&gt;&lt;p&gt;Do you buy the forced-selling story, or is the market seeing a demand break we are missing? Which S1 test do you think fails first: SK hynix&amp;rsquo;s buyback, Microsoft&amp;rsquo;s wording, or the bond window? Tell me in the comments. I will score this on Nov 6.&lt;/p&gt;
&lt;p&gt;If this helped you think through the week, a like helps other investors find it.&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Disclaimer: For research and information purposes only. Not investment advice. Names cited are for analytical illustration; readers should perform their own due diligence and consult licensed advisors before any investment decision.&lt;/em&gt;&lt;/p&gt;
&lt;div class="footnotes" role="doc-endnotes"&gt;
&lt;hr&gt;
&lt;ol&gt;
&lt;li id="fn:1"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.trendforce.com/price/dram/dram_spot" target="_blank" rel="noopener"
 &gt;TrendForce, DRAM spot price&lt;/a&gt;: DDR5 16Gb (2Gx8) 4800/5600 session average $61.267 on 2026-10-23 and $54.83 on 2026-09-16, from captured price tables.&amp;#160;&lt;a href="#fnref:1" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:1" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:2"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.korea.kr/common/download.do?fileId=198574612&amp;amp;tblKey=GMN" target="_blank" rel="noopener"
 &gt;Korea Customs Service via Korea.kr, exports for Oct 1–20, 2026&lt;/a&gt;: semiconductors $27.516 billion, up 222.5% from a year earlier. Provisional data.&amp;#160;&lt;a href="#fnref:2" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:2" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:3"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.tickmill.com/blog/institutional-insights-goldman-sachs-tech-liquidation-retial-etf-chase" target="_blank" rel="noopener"
 &gt;Tickmill, summary of Goldman Sachs notes&lt;/a&gt;, 2026-10-20: CTA global equity longs from about $125 billion to $39 billion. Secondary summary of a paid note.&amp;#160;&lt;a href="#fnref:3" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:3" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:4"&gt;
&lt;p&gt;Naver Finance investor-trend data (KRX based). Window totals are our sums of daily data; Samsung Electronics foreign net buying in shares from the same source. Rechecked on Oct 24.&amp;#160;&lt;a href="#fnref:4" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:4" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:5"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.fxstreet.com/cryptocurrencies/news/crypto-today-bitcoin-holds-the-50-day-ema-as-ethereum-and-xrp-face-structural-resistance-202610231024" target="_blank" rel="noopener"
 &gt;FXStreet&lt;/a&gt;, 2026-10-23, citing the CME FedWatch Tool: about an 8% chance of an October hike. December hike odds near 83% are from our market tracker as of Oct 23 (secondary).&amp;#160;&lt;a href="#fnref:5" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:5" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:6"&gt;
&lt;p&gt;Consensus data compiled by FnGuide (Samsung 2027 operating profit KRW 554 trillion to 561 trillion as of Oct 14; 1Q27 −2.71% and 2Q27 −2.56% quarterly revisions as of Oct 5), with Kiwoom Securities at KRW 555 trillion; Micron FY27 EPS revision over 30 days. Secondary data.&amp;#160;&lt;a href="#fnref:6" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:6" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:7"&gt;
&lt;p&gt;TSMC third-quarter 2026 earnings call, 2026-10-15: 2026 capital budget kept at $60–64 billion.&amp;#160;&lt;a href="#fnref:7" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:8"&gt;
&lt;p&gt;Amazon second-quarter 2026 earnings call, 2026-07-30.&amp;#160;&lt;a href="#fnref:8" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:9"&gt;
&lt;p&gt;Samsung Electronics second-quarter 2026 earnings call, 2026-07-30. The 60–70% allocation plan includes agreements not yet signed.&amp;#160;&lt;a href="#fnref:9" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:10"&gt;
&lt;p&gt;&lt;a class="link" href="https://macenews.com/bofa-global-research-fund-manager-survey-lower-growth-and-lower-inflation-expectations-lead-to-tweaking-of-asset-allocations-in-october/" target="_blank" rel="noopener"
 &gt;Mace News, BofA Global Fund Manager Survey for October&lt;/a&gt;, 2026-10-13. Secondary summary.&amp;#160;&lt;a href="#fnref:10" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:11"&gt;
&lt;p&gt;Index and ETF closes for 2026-10-22 from Yahoo Finance, as summarized by &lt;a class="link" href="https://www.stocktitan.net/articles/why-is-nasdaq-down-today-october-22-2026-ai-chip-stocks" target="_blank" rel="noopener"
 &gt;StockTitan&lt;/a&gt;; 10-year Treasury yield from &lt;a class="link" href="https://fred.stlouisfed.org/series/DGS10" target="_blank" rel="noopener"
 &gt;FRED DGS10&lt;/a&gt;, 5.13% on Oct 21 and 4.96% on Oct 22.&amp;#160;&lt;a href="#fnref:11" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:12"&gt;
&lt;p&gt;Same Tickmill summary, 2026-10-20: gross leverage 311.7% and net 72.7% for the week of Oct 16. Secondary.&amp;#160;&lt;a href="#fnref:12" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:13"&gt;
&lt;p&gt;Federal Reserve Bank of St. Louis: &lt;a class="link" href="https://fred.stlouisfed.org/series/BAMLH0A0HYM2" target="_blank" rel="noopener"
 &gt;ICE BofA US High Yield OAS&lt;/a&gt; 3.40% and &lt;a class="link" href="https://fred.stlouisfed.org/series/BAMLC0A0CM" target="_blank" rel="noopener"
 &gt;US Corporate (IG) OAS&lt;/a&gt; 0.89% on 2026-10-22, the latest available on Oct 24.&amp;#160;&lt;a href="#fnref:13" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:13" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:14"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.mt.co.kr/en/industry/2026/10/19/2026101913371799572" target="_blank" rel="noopener"
 &gt;Money Today&lt;/a&gt;, 2026-10-19, citing securities-industry expectations. The earlier KRW 40 trillion program is in the &lt;a class="link" href="https://news.skhynix.com/en/share-buyback-and-retirement/" target="_blank" rel="noopener"
 &gt;SK hynix newsroom&lt;/a&gt;.&amp;#160;&lt;a href="#fnref:14" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:15"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.reuters.com/business/short-sellers-pocket-6-billion-profits-neoclouds-slide-ai-debt-fears-2026-10-22/" target="_blank" rel="noopener"
 &gt;Reuters, short sellers and neoclouds&lt;/a&gt;, 2026-10-22, citing ORTEX data.&amp;#160;&lt;a href="#fnref:15" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:16"&gt;
&lt;p&gt;Naver Finance investor-trend data for 2026-07-31: foreigners +KRW 7,241.0 billion, individuals −KRW 8,284.0 billion in KOSPI.&amp;#160;&lt;a href="#fnref:16" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:17"&gt;
&lt;p&gt;Our calculation from daily closes: SOX (Yahoo Finance, 1994 onward) and KOSPI (2025–2026). The crowding subset is our classification. Past patterns are base rates, not forecasts.&amp;#160;&lt;a href="#fnref:17" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:17" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:18"&gt;
&lt;p&gt;Goldman Sachs prime brokerage AI net-flow chart as shared in market commentary (read from the chart; estimate). Two cases only.&amp;#160;&lt;a href="#fnref:18" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;/ol&gt;
&lt;/div&gt;</description></item><item><title>Squeeze or Second Leg? Four Scenarios and a Day-by-Day Watch List for Oct 26–30</title><link>https://koreainvestinsights.com/post/chip-selloff-four-scenarios-week-ahead-2026-10-24/</link><pubDate>Sat, 24 Oct 2026 15:30:00 +0900</pubDate><guid>https://koreainvestinsights.com/post/chip-selloff-four-scenarios-week-ahead-2026-10-24/</guid><description>&lt;img src="https://koreainvestinsights.com/" alt="Featured image of post Squeeze or Second Leg? Four Scenarios and a Day-by-Day Watch List for Oct 26–30" /&gt;&lt;p&gt;Last week we mapped who holds what after October&amp;rsquo;s chip crash: it was a long unwind, not a short attack. This week answers the next question. Does the unwind end in a short squeeze, or in a second leg down?&lt;/p&gt;
&lt;p&gt;Two things moved after Friday&amp;rsquo;s close. The December Brent contract settled at $116.91, up 3.3%, while 10-year inflation breakevens rose 2bp and the 2-year yield rose 4bp.&lt;sup id="fnref:1"&gt;&lt;a href="#fn:1" class="footnote-ref" role="doc-noteref"&gt;1&lt;/a&gt;&lt;/sup&gt; Under the rule we set in advance, that makes it an inflation-type oil day, so the macro-shock scenario rose from 15% to 20% and the base case fell from 55% to 50%. Meanwhile the US high-yield spread stood at 340bp on Oct 22, 10bp below our 350bp credit alarm.&lt;sup id="fnref:2"&gt;&lt;a href="#fn:2" class="footnote-ref" role="doc-noteref"&gt;2&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;The base case is still a rebound. Memory earnings for 2026 and 2027 got stronger this month, and most of the forced selling is done. But this week the burden of proof sits with credit and oil, not with AI demand. Below are the four scenarios, the lines that move them, and what to watch each day from Monday to Friday.&lt;/p&gt;
&lt;h2 id="five-things-to-take-away"&gt;Five things to take away
&lt;/h2&gt;&lt;ul&gt;
&lt;li&gt;Base case (S1, 50%): a rebound with a hawkish Fed on hold. The other three are a credit squeeze (S2, 15%), a capex slowdown (S3, 15%) and a macro shock (S4, 20%). All probabilities are our subjective estimates as of Oct 24.&lt;/li&gt;
&lt;li&gt;The bear case needs proof. We call the downturn confirmed only on one of three events: an annual cut to next year&amp;rsquo;s memory profit consensus, US high-yield spreads at 350bp or more, or one Big Tech company slowing 2027 capex. None has happened yet. The credit line is the closest.&lt;/li&gt;
&lt;li&gt;Oil is now the swing factor for the macro case. Brent above $110 counts against us only on days when inflation gauges rise with it. We judge that every day.&lt;/li&gt;
&lt;li&gt;Thursday, Oct 29 is the densest day: Oracle&amp;rsquo;s investor day, the FOMC, Alphabet, Microsoft and Samsung&amp;rsquo;s call land within ten hours in Korea time.&lt;/li&gt;
&lt;li&gt;In Korea, a lasting rebound needs foreigners to buy Samsung Electronics and SK hynix shares again. On Friday they sold about KRW 0.98 trillion of KOSPI stocks, almost at our KRW 1 trillion alarm.&lt;sup id="fnref:3"&gt;&lt;a href="#fn:3" class="footnote-ref" role="doc-noteref"&gt;3&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;Prices are Oct 23 closes. Korean closes are final prints after the 20:00 KST extended session. This is analysis, not a trading instruction.&lt;/p&gt;
&lt;h2 id="where-the-cycle-stands"&gt;Where the cycle stands
&lt;/h2&gt;&lt;p&gt;We track the cycle in four layers. The market has priced them very differently.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Layer&lt;/th&gt;
 &lt;th&gt;State&lt;/th&gt;
 &lt;th&gt;Priced in?&lt;/th&gt;
 &lt;th&gt;What we see&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;A. Memory profits, 2026–27&lt;/td&gt;
 &lt;td&gt;Strong&lt;/td&gt;
 &lt;td&gt;Barely&lt;/td&gt;
 &lt;td&gt;2027 operating-profit consensus for Samsung still rising; DDR5 spot up 11.7% since Sep 16&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;B. Supply and demand, 2028+&lt;/td&gt;
 &lt;td&gt;Uncertain&lt;/td&gt;
 &lt;td&gt;Heavily&lt;/td&gt;
 &lt;td&gt;Wafer capacity at four makers up 15–17% in 2028; CXMT gaining share&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;C. Funding and rates&lt;/td&gt;
 &lt;td&gt;Worsening&lt;/td&gt;
 &lt;td&gt;Partly&lt;/td&gt;
 &lt;td&gt;HY spread 340bp, Oracle CDS near 315bp, US 10-year at 5.00%, Brent $116.91&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;D. Positioning&lt;/td&gt;
 &lt;td&gt;Largely flushed&lt;/td&gt;
 &lt;td&gt;Largely&lt;/td&gt;
 &lt;td&gt;Hedge-fund AI net buying erased, CTA longs down 69%, but gross leverage still near a record&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;The A layer kept improving while the stocks fell. Consensus for Samsung&amp;rsquo;s 2027 operating profit rose to KRW 555 trillion at Kiwoom and KRW 561 trillion at FnGuide, and Micron&amp;rsquo;s FY27 EPS consensus rose 12.8% in 30 days.&lt;sup id="fnref:4"&gt;&lt;a href="#fn:4" class="footnote-ref" role="doc-noteref"&gt;4&lt;/a&gt;&lt;/sup&gt; DDR5 16Gb spot averaged $61.27 on Oct 23, against $54.83 on Sep 16.&lt;sup id="fnref:5"&gt;&lt;a href="#fn:5" class="footnote-ref" role="doc-noteref"&gt;5&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;The swing question is whether the C layer spills into the B layer. In plain terms: do higher funding costs make the AI builders who borrow cut their 2028 orders? Credit for hyperscalers is fine; IG spreads are 89bp.&lt;sup id="fnref1:2"&gt;&lt;a href="#fn:2" class="footnote-ref" role="doc-noteref"&gt;2&lt;/a&gt;&lt;/sup&gt; The pressure sits at the weak end, in Oracle and the debt-funded neoclouds.&lt;sup id="fnref:6"&gt;&lt;a href="#fn:6" class="footnote-ref" role="doc-noteref"&gt;6&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;The D layer is the main reason we keep the base case. Goldman&amp;rsquo;s model of CTA equity longs fell about 69%, and the share of fund managers who call semiconductors the most crowded trade fell from 82% to 45%.&lt;sup id="fnref:7"&gt;&lt;a href="#fn:7" class="footnote-ref" role="doc-noteref"&gt;7&lt;/a&gt;&lt;/sup&gt; The exception is gross leverage. Hedge funds still run about 312% gross, so a second round of de-grossing is possible if longs and shorts move against each other.&lt;sup id="fnref:8"&gt;&lt;a href="#fn:8" class="footnote-ref" role="doc-noteref"&gt;8&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;h2 id="where-the-tripwires-stand"&gt;Where the tripwires stand
&lt;/h2&gt;&lt;p&gt;The chart shows each line we monitor, the latest value and the distance to the line that would move a scenario.&lt;/p&gt;
&lt;div class="fig-scroll" style="overflow-x:auto;-webkit-overflow-scrolling:touch;margin:1.4em 0"&gt;&lt;img src="https://koreainvestinsights.com/post-assets/chip-selloff-four-scenarios-week-ahead-2026-10-24/tripwires.en.svg" alt="Tripwire gauges: US high-yield spread 340bp, 10bp below the 350bp alarm; Oracle CDS about 315bp, above the 300bp line but not confirmed; CoreWeave 2032 notes 76 to 81 cents, between the 75 and 80 cent lines; Brent 116.91 dollars, above the 110 dollar rule; US 10-year 5.00 percent, below the 5.3 percent shock line; France-Germany 10-year spread 166.6bp, above 150bp; foreign KOSPI cash selling 0.98 trillion won in one day, near the 1 trillion line; DDR5 spot 61.27 dollars and rising" loading="lazy" style="display:block;width:100%;min-width:720px;max-width:none;height:auto;margin:0 auto"&gt;&lt;/div&gt;
&lt;p&gt;Two lines have tripped and three are close. Brent and the France–Germany spread are past their lines. France&amp;rsquo;s spread was 166.6bp on Oct 23, before Moody&amp;rsquo;s kept the country&amp;rsquo;s Aa3 rating later that day.&lt;sup id="fnref:9"&gt;&lt;a href="#fn:9" class="footnote-ref" role="doc-noteref"&gt;9&lt;/a&gt;&lt;/sup&gt; The high-yield spread, Oracle&amp;rsquo;s CDS and Friday&amp;rsquo;s foreign selling in Korea are near theirs. The physical memory market is the one clean green light.&lt;/p&gt;
&lt;h2 id="four-scenarios-to-mid-november"&gt;Four scenarios to mid-November
&lt;/h2&gt;&lt;div class="fig-scroll" style="overflow-x:auto;-webkit-overflow-scrolling:touch;margin:1.4em 0"&gt;&lt;img src="https://koreainvestinsights.com/post-assets/chip-selloff-four-scenarios-week-ahead-2026-10-24/scenarios.en.svg" alt="Four scenarios with subjective probabilities as of Oct 24: S1 rebound under a hawkish hold 50 percent, S2 credit squeeze 15 percent, S3 demand slowdown 15 percent, S4 macro shock 20 percent, raised from 15 percent on Oct 24 under the Brent rule" loading="lazy" style="display:block;width:100%;min-width:720px;max-width:none;height:auto;margin:0 auto"&gt;&lt;/div&gt;
&lt;p&gt;&lt;strong&gt;S1. Rebound under a hawkish hold (50%).&lt;/strong&gt; Big Tech keeps 2027 capex rising, memory makers confirm 2027 contracts and HBM4, the Fed holds and leaves December open, and the post-earnings bond window clears with high-yield spreads at 330bp or less. The result is an oversold rebound with rates capping the upside. It breaks if new bond deals need bigger concessions in early November. Then S1 turns into S2.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;S2. Credit squeeze (15%).&lt;/strong&gt; Capex holds, but it is funded with more debt and special-purpose vehicles. The gates are high-yield spreads at 350bp or more, Oracle&amp;rsquo;s five-year CDS staying above 300bp and CoreWeave&amp;rsquo;s 2032 notes below 75 cents. Good earnings get a weak reaction and the July lows are retested.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;S3. Demand slowdown (15%).&lt;/strong&gt; One or more hyperscalers slows 2027 capex or stresses discipline. This is the second leg down, and SK hynix would fall the most because its valuation leans hardest on HBM. Meta did not trigger it on Oct 21: it raised the low end of its 2026 range and gave no 2027 number.&lt;sup id="fnref:10"&gt;&lt;a href="#fn:10" class="footnote-ref" role="doc-noteref"&gt;10&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;S4. Macro shock (20%).&lt;/strong&gt; A surprise rate hike, hot core inflation or Brent above $110 on an inflation-type day. The result is a broad selloff with the 10-year at 5.3% or more. Markets price only about an 8% chance of a hike on Oct 28, with more risk in December, so the live channel this week is oil, not the Fed.&lt;sup id="fnref:11"&gt;&lt;a href="#fn:11" class="footnote-ref" role="doc-noteref"&gt;11&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;Roughly, these line up with last week&amp;rsquo;s positioning paths: S1 is the short-cover path A, S3 is path B (one more de-gross, then a cover), and S2 is path D (the shorts win on credit). S4 does not need a positioning trigger; it hits everything at once.&lt;/p&gt;
&lt;h3 id="what-moves-the-odds"&gt;What moves the odds
&lt;/h3&gt;&lt;p&gt;We change probabilities only on a primary source or on a threshold break, in 5-point steps and no more than 10 points per scenario in a day. Three rules matter this week.&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;The oil rule.&lt;/strong&gt; Brent settles at $110 or more. If 10-year breakevens or the 2-year yield rise at least 1bp that day, the oil move is inflation-type: S4 goes to 20% and S1 to 50%. If breakevens, the 2-year and copper all fall, it is supply-type: S4 goes back to 15% and S1 to 55%. Mixed days change nothing.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;The credit rule.&lt;/strong&gt; Each S2 gate that breaks with a primary source moves S2 up. A high-yield spread back at 330bp or less with deals clearing moves S1 up.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;The capex rule.&lt;/strong&gt; If Big Tech puts 2027 capex on hold, we move 15 points from the short-cover path to the de-gross path and review S3 for 5 more points.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id="the-week-ahead-day-by-day"&gt;The week ahead, day by day
&lt;/h2&gt;&lt;div class="fig-scroll" style="overflow-x:auto;-webkit-overflow-scrolling:touch;margin:1.4em 0"&gt;&lt;img src="https://koreainvestinsights.com/post-assets/chip-selloff-four-scenarios-week-ahead-2026-10-24/week.en.svg" alt="Week ahead calendar in Korea time from Monday Oct 26 to Friday Oct 30, color-coded by demand, credit, rates and oil, and Korea positioning. Thursday Oct 29 carries Oracle's investor day, the FOMC, Alphabet, Microsoft, Samsung's call, US GDP and PCE, and the ECB" loading="lazy" style="display:block;width:100%;min-width:720px;max-width:none;height:auto;margin:0 auto"&gt;&lt;/div&gt;
&lt;p&gt;Times are Korea time (KST). US earnings after the New York close arrive the next morning in Korea.&lt;sup id="fnref:12"&gt;&lt;a href="#fn:12" class="footnote-ref" role="doc-noteref"&gt;12&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;h3 id="monday-oct-26-europe-and-the-korea-flow-box"&gt;Monday, Oct 26: Europe and the Korea flow box
&lt;/h3&gt;&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;France–Germany 10-year spread.&lt;/strong&gt; This is the first trading day after Moody&amp;rsquo;s decision. Below 150bp turns our sovereign alarm off; staying above 150bp keeps it on. A spread that keeps widening would push European rates and the macro case together.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Korea foreign flows (after 18:00).&lt;/strong&gt; We sort each session into one of four boxes. Bullish: foreigners net buy Samsung Electronics and SK hynix for two or more sessions and turn buyers of KOSPI 200 futures. Bearish: foreign cash selling of KRW 1 trillion or more in a day, or KRW 3 trillion or more over five sessions, together with futures buying.&lt;/li&gt;
&lt;/ul&gt;
&lt;h3 id="tuesday-oct-27-shorts-and-oil"&gt;Tuesday, Oct 27: shorts and oil
&lt;/h3&gt;&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;02:00, US 2-year auction.&lt;/strong&gt; A tail of 1bp or more would add pressure at the front end. The 5-year follows Wednesday and the 7-year Friday.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Morning, FINRA short interest (settled Oct 15).&lt;/strong&gt; Did neocloud shorts hold or grow? We compare CoreWeave, Core Scientific, Nebius, Galaxy Digital, Hut 8 and Teradyne with Sep 30. More shorts means more squeeze fuel if the week goes well. We also check whether shorts on Korea and chip ETFs (EWY, SOXX, SMH) rose.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;All day, Israel&amp;rsquo;s election.&lt;/strong&gt; The channel to chips is oil. Under our rule, a Brent spike alone does not move S4; the inflation gauges have to move with it.&lt;/li&gt;
&lt;/ul&gt;
&lt;h3 id="wednesday-oct-28-sk-hynix"&gt;Wednesday, Oct 28: SK hynix
&lt;/h3&gt;&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Early morning, Seagate and Teradyne.&lt;/strong&gt; Seagate&amp;rsquo;s HDD and flash pricing is an early read on storage demand and on Samsung&amp;rsquo;s NAND. Teradyne is on our squeeze list.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;SK hynix third quarter and a new shareholder-return plan.&lt;/strong&gt; Brokers expect funding of more than KRW 30 trillion for new returns.&lt;sup id="fnref:13"&gt;&lt;a href="#fn:13" class="footnote-ref" role="doc-noteref"&gt;13&lt;/a&gt;&lt;/sup&gt; The KRW 40 trillion buyback absorbed foreign selling until it ended on Oct 21. A large new program would refill that demand gap. Beyond the headline profit, we watch four points: how much 2027 HBM4 volume is under contract, the 2027 capex plan, HBM4E customers and timing, and how many ASIC customers buy HBM4.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;What would worry us.&lt;/strong&gt; A 2027 capex plan far above expectations reads as 2028 supply, not as strength. A thin return plan reopens the demand gap.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;15:30, Advantest.&lt;/strong&gt; Memory tester orders. It is the one equipment name roughly flat since late June.&lt;/li&gt;
&lt;/ul&gt;
&lt;h3 id="thursday-oct-29-ten-hours-that-decide-the-month"&gt;Thursday, Oct 29: ten hours that decide the month
&lt;/h3&gt;&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;00:55, Oracle investor day.&lt;/strong&gt; The funding mix is the test. More customer prepayments would ease the credit chain. More debt or off-balance-sheet funding would add to it. Oracle&amp;rsquo;s CDS staying above 300bp would push S2.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;03:00, FOMC.&lt;/strong&gt; A hold is expected. The risk is wording that signals further hikes, which is an S4 trigger. With no forward guidance from Chair Warsh, the statement carries more weight than usual.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;05:30, Alphabet.&lt;/strong&gt; Three numbers: third-quarter capex against a consensus near $55.8 billion, any direction for 2027 capex, and API token growth. Google&amp;rsquo;s API ran at 22 billion tokens a minute in the second quarter, up 37.5% from 16 billion.&lt;sup id="fnref:14"&gt;&lt;a href="#fn:14" class="footnote-ref" role="doc-noteref"&gt;14&lt;/a&gt;&lt;/sup&gt; Growth below 15% a quarter would lean bearish on demand. A hold or cut in 2027 capex would move S3.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;06:30, Microsoft.&lt;/strong&gt; We use four tests: demand still exceeds supply; calendar 2026 capex is raised or the December quarter is at least the September quarter; leases signed but not yet started keep growing; Azure grows 45% or more. Three of four would support the squeeze case. We put the odds near 42%.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;10:00, Samsung Electronics third-quarter call.&lt;/strong&gt; Memory margins, the share of 2027 supply under long-term contracts and any price floors, HBM4 qualification and new ASIC customers.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;21:30, US third-quarter GDP and September PCE.&lt;/strong&gt; Core PCE above expectations, a day after a hawkish FOMC, is the S4 combination.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;22:15, ECB decision.&lt;/strong&gt; Watch any comment on France&amp;rsquo;s spread and the bank&amp;rsquo;s bond-buying backstop.&lt;/li&gt;
&lt;/ul&gt;
&lt;h3 id="friday-oct-30-amazon-and-the-bond-window"&gt;Friday, Oct 30: Amazon and the bond window
&lt;/h3&gt;&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;02:00, US 7-year auction.&lt;/strong&gt; The first auction after the FOMC and PCE.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;05:30, Sandisk; 06:00, Amazon and Intel.&lt;/strong&gt; For Amazon: AWS capex, how much of the 2026 increase is memory prices rather than more buildings, and how much 2028 capacity is already reserved. For Sandisk: NAND pricing and Chinese competition, read against Samsung&amp;rsquo;s NAND comments the day before.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Midday, Bank of Japan.&lt;/strong&gt; A hold is expected; the question is a December signal. Yen carry matters for global rates.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;21:30, US third-quarter employment cost index.&lt;/strong&gt; Wage-driven inflation is the slow S4 channel.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;From Oct 30 to about Nov 7, the bond window.&lt;/strong&gt; After earnings, hyperscalers and their partners usually issue bonds. Watch new-issue concessions. High-yield spreads at 330bp or less with deals clearing would support S1. Spreads at 350bp or more with failed or pulled deals would confirm S2.&lt;/li&gt;
&lt;/ul&gt;
&lt;h3 id="over-the-weekend-and-the-week-after"&gt;Over the weekend and the week after
&lt;/h3&gt;&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Oct 31:&lt;/strong&gt; CXMT&amp;rsquo;s third-quarter report (margins and expansion plans).&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Nov 1, 09:00:&lt;/strong&gt; Korea&amp;rsquo;s October exports. The US leaves daylight saving time the same day, so US data arrive one hour later in Korea from Nov 2.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Nov 3:&lt;/strong&gt; US midterm elections. &lt;strong&gt;Nov 6, 22:30:&lt;/strong&gt; US October jobs.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Nov 6:&lt;/strong&gt; we score this map: which scenario played out, which positioning path, and whether the squeeze list beat its Oct 21 closes.&lt;/li&gt;
&lt;/ul&gt;
&lt;h3 id="every-day-three-standing-checks"&gt;Every day: three standing checks
&lt;/h3&gt;&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;DRAM spot.&lt;/strong&gt; As long as DDR5 spot holds or rises, the physical market does not confirm the selloff.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Korea&amp;rsquo;s flow box.&lt;/strong&gt; Box ① (cash buying plus futures buying) is the only one that supports a lasting rebound.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;The oil rule.&lt;/strong&gt; Brent at $110 or more is judged every day against breakevens, the 2-year yield and copper.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id="our-read"&gt;Our read
&lt;/h2&gt;&lt;p&gt;We keep the rebound as the base case for one reason: the layer that pays memory makers is getting stronger, while the selling that drove October looks mostly done. What would change our mind fast is not a weak quarter. It is a high-yield spread through 350bp with deals failing, or one hyperscaler cutting 2027 capex. Either one would mean the funding problem has reached the buyers who pay for memory.&lt;/p&gt;
&lt;h2 id="your-turn"&gt;Your turn
&lt;/h2&gt;&lt;p&gt;Which single print matters most to you next week: SK hynix&amp;rsquo;s capex plan, Microsoft&amp;rsquo;s four tests, the FOMC statement, or Brent? And where is your line? Tell me in the comments. I will score this map on Nov 6, so your thresholds go into the scorecard too.&lt;/p&gt;
&lt;p&gt;If this watch list helps you get through the week, a like helps other investors find it.&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Disclaimer: For research and information purposes only. Not investment advice. Names cited are for analytical illustration; readers should perform their own due diligence and consult licensed advisors before any investment decision.&lt;/em&gt;&lt;/p&gt;
&lt;div class="footnotes" role="doc-endnotes"&gt;
&lt;hr&gt;
&lt;ol&gt;
&lt;li id="fn:1"&gt;
&lt;p&gt;Brent December 2026 futures settlement of $116.91 on 2026-10-23 (+3.3%), as reported by Yahoo Finance for the December contract (BZZ26). A continuous front-month series showed a lower value around the contract roll. Breakeven and 2-year moves from &lt;a class="link" href="https://fred.stlouisfed.org/series/T10YIE" target="_blank" rel="noopener"
 &gt;FRED T10YIE&lt;/a&gt; and the US Treasury daily curve. The oil rule is our own pre-set rule.&amp;#160;&lt;a href="#fnref:1" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:2"&gt;
&lt;p&gt;Federal Reserve Bank of St. Louis: &lt;a class="link" href="https://fred.stlouisfed.org/series/BAMLH0A0HYM2" target="_blank" rel="noopener"
 &gt;ICE BofA US High Yield OAS&lt;/a&gt; 3.40% and &lt;a class="link" href="https://fred.stlouisfed.org/series/BAMLC0A0CM" target="_blank" rel="noopener"
 &gt;US Corporate (IG) OAS&lt;/a&gt; 0.89% on 2026-10-22, the latest available on Oct 24.&amp;#160;&lt;a href="#fnref:2" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:2" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:3"&gt;
&lt;p&gt;Naver Finance investor-trend data (KRX based) for 2026-10-23: foreigners −KRW 981.1 billion in KOSPI cash equities, institutions +KRW 868.2 billion. SK hynix accounted for about −KRW 0.60 trillion of foreign selling (shares × KRX close).&amp;#160;&lt;a href="#fnref:3" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:4"&gt;
&lt;p&gt;Consensus data compiled by Kiwoom Securities (2027 operating profit KRW 546 trillion to 555 trillion) and FnGuide (554 trillion to 561 trillion); Micron FY27 EPS consensus revision over 30 days. Secondary data, as of mid-October 2026.&amp;#160;&lt;a href="#fnref:4" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:5"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.trendforce.com/price/dram/dram_spot" target="_blank" rel="noopener"
 &gt;TrendForce, DRAM spot price&lt;/a&gt;: DDR5 16Gb (2Gx8) 4800/5600 session average $61.267 on 2026-10-23 and $54.83 on 2026-09-16, from captured price tables.&amp;#160;&lt;a href="#fnref:5" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:6"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.cnbc.com/2026/10/19/coreweave-oracle-ai-credit-stress.html" target="_blank" rel="noopener"
 &gt;CNBC, CoreWeave and Oracle credit stress&lt;/a&gt;, 2026-10-19: Oracle five-year CDS about 315bp; CoreWeave 2032 notes near a 14.9% yield. Other reports quoted Oracle near 322bp; we did not see the original data.&amp;#160;&lt;a href="#fnref:6" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:7"&gt;
&lt;p&gt;CTA longs from Tickmill summaries of Goldman Sachs flow notes (&lt;a class="link" href="https://www.tickmill.com/blog/institutional-insights-goldman-sachs-tech-liquidation-retial-etf-chase" target="_blank" rel="noopener"
 &gt;Oct 20&lt;/a&gt;); fund-manager crowding from &lt;a class="link" href="https://macenews.com/bofa-global-research-fund-manager-survey-lower-growth-and-lower-inflation-expectations-lead-to-tweaking-of-asset-allocations-in-october/" target="_blank" rel="noopener"
 &gt;Mace News on the BofA October survey&lt;/a&gt;. Secondary summaries.&amp;#160;&lt;a href="#fnref:7" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:8"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.tickmill.com/blog/institutional-insights-goldman-sachs-tech-liquidation-retial-etf-chase" target="_blank" rel="noopener"
 &gt;Tickmill, summary of Goldman Sachs notes&lt;/a&gt;, 2026-10-20: gross leverage 311.7% and net 72.7% for the week of Oct 16. Secondary.&amp;#160;&lt;a href="#fnref:8" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:9"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.cnbc.com/2026/10/23/moodys-holds-steady-on-frances-credit-rating-as-country-grapples-with-turmoil-and-higher-yields.html" target="_blank" rel="noopener"
 &gt;CNBC, Moody&amp;rsquo;s holds steady on France&amp;rsquo;s credit rating&lt;/a&gt;, 2026-10-23. The spread of 166.6bp is the 10-year benchmark difference on Oct 23 from our daily market panel.&amp;#160;&lt;a href="#fnref:9" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:10"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/0001326801/000162828026066700/meta-09302026xexhibit991.htm" target="_blank" rel="noopener"
 &gt;Meta Platforms, third-quarter 2026 results (SEC Form 8-K, Exhibit 99.1)&lt;/a&gt;, 2026-10-21.&amp;#160;&lt;a href="#fnref:10" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:11"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.fxstreet.com/cryptocurrencies/news/crypto-today-bitcoin-holds-the-50-day-ema-as-ethereum-and-xrp-face-structural-resistance-202610231024" target="_blank" rel="noopener"
 &gt;FXStreet&lt;/a&gt;, 2026-10-23, citing the CME FedWatch Tool: about an 8% chance of an October hike and higher odds for December. Context on guidance from &lt;a class="link" href="https://www.forbes.com/sites/simonmoore/2026/10/19/what-to-look-for-from-the-feds-october-meeting/" target="_blank" rel="noopener"
 &gt;Forbes&lt;/a&gt;, 2026-10-19.&amp;#160;&lt;a href="#fnref:11" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:12"&gt;
&lt;p&gt;Company and official schedules: &lt;a class="link" href="https://news.skhynix.com/en/q3-2026-earnings-conference-call-invitation/" target="_blank" rel="noopener"
 &gt;SK hynix earnings call notice&lt;/a&gt;; &lt;a class="link" href="https://www.prnewswire.com/news-releases/oracle-investor-day-meeting-to-be-held-october-28-2026-302918062.html" target="_blank" rel="noopener"
 &gt;Oracle investor day notice&lt;/a&gt;; &lt;a class="link" href="https://www.federalreserve.gov/monetarypolicy/fomccalendars.htm" target="_blank" rel="noopener"
 &gt;Federal Reserve FOMC calendar&lt;/a&gt;; &lt;a class="link" href="https://news.microsoft.com/source/2026/10/07/microsoft-announces-quarterly-earnings-release-date-69/" target="_blank" rel="noopener"
 &gt;Microsoft earnings date&lt;/a&gt;; &lt;a class="link" href="https://abc.xyz/investor/news/news-details/2026/Alphabet-Announces-Date-of-Third-Quarter-2026-Financial-Results-Conference-Call-2026-8tpGZsLS6v/default.aspx" target="_blank" rel="noopener"
 &gt;Alphabet earnings date&lt;/a&gt;; &lt;a class="link" href="https://ir.aboutamazon.com/news-release/news-release-details/2026/Amazon-com-to-Webcast-Third-Quarter-2026-Financial-Results-Conference-Call/default.aspx" target="_blank" rel="noopener"
 &gt;Amazon earnings webcast&lt;/a&gt;; &lt;a class="link" href="https://www.sandisk.com/company/newsroom/press-releases/2026/2026-09-29-sandisk-to-report-first-quarter-fiscal-year-2027-results-on-october-29-2026" target="_blank" rel="noopener"
 &gt;Sandisk results date&lt;/a&gt;; &lt;a class="link" href="https://www.samsung.com/global/ir/ir-events-presentations/events/" target="_blank" rel="noopener"
 &gt;Samsung Electronics IR events&lt;/a&gt;; &lt;a class="link" href="https://www.bea.gov/news/schedule/full" target="_blank" rel="noopener"
 &gt;BEA release schedule&lt;/a&gt; (GDP advance estimate and September personal income and outlays on Oct 29, 8:30 a.m. ET); &lt;a class="link" href="https://www.ecb.europa.eu/press/calendars/mgcgc/html/index.en.html" target="_blank" rel="noopener"
 &gt;ECB meeting calendar&lt;/a&gt;; &lt;a class="link" href="https://www.finra.org/filing-reporting/regulatory-filing-systems/short-interest" target="_blank" rel="noopener"
 &gt;FINRA short interest schedule&lt;/a&gt;. Bank of Japan, Treasury auction and other times follow published schedules as compiled in our event calendar.&amp;#160;&lt;a href="#fnref:12" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:13"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.mt.co.kr/en/industry/2026/10/19/2026101913371799572" target="_blank" rel="noopener"
 &gt;Money Today&lt;/a&gt;, 2026-10-19, citing securities-industry expectations; the earlier program is in the &lt;a class="link" href="https://news.skhynix.com/en/share-buyback-and-retirement/" target="_blank" rel="noopener"
 &gt;SK hynix newsroom&lt;/a&gt;.&amp;#160;&lt;a href="#fnref:13" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:14"&gt;
&lt;p&gt;Alphabet second-quarter 2026 results (SEC Form 8-K, 2026-07-22): 22 billion API tokens per minute against 16 billion in the prior quarter. The 15% line and the Microsoft tests are our own thresholds.&amp;#160;&lt;a href="#fnref:14" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;/ol&gt;
&lt;/div&gt;</description></item><item><title>October's Chip Crash Was a Long Unwind, Not a Short Attack</title><link>https://koreainvestinsights.com/post/october-chip-selloff-long-unwind-not-short-attack-2026-10-24/</link><pubDate>Sat, 24 Oct 2026 14:30:00 +0900</pubDate><guid>https://koreainvestinsights.com/post/october-chip-selloff-long-unwind-not-short-attack-2026-10-24/</guid><description>&lt;img src="https://koreainvestinsights.com/" alt="Featured image of post October's Chip Crash Was a Long Unwind, Not a Short Attack" /&gt;&lt;p&gt;From its October closing high to Friday&amp;rsquo;s close, the S&amp;amp;P 500 lost 2.5%. Over the same stretch the Philadelphia Semiconductor Index (SOX) fell 18.4%, SK hynix 31.4% and CoreWeave 40.9%. This was not a market-wide panic. One trade came apart: the AI chain.&lt;/p&gt;
&lt;p&gt;How it came apart matters more than how far. Hedge funds did not build a big new short against memory or Korea. They sold the longs they already owned, then hedged what they kept with index shorts. In Korea, foreign investors sold KRW 27.0 trillion of cash equities between Sept 24 and Oct 15 while they bought KOSPI 200 futures.&lt;/p&gt;
&lt;p&gt;That shape decides what comes next. A book that is long memory and short the index is exposed to correlation more than to direction. The next two weeks, from SK hynix&amp;rsquo;s results on Oct 28 to Microsoft&amp;rsquo;s capex wording and the Fed, will show whether it unwinds as a short squeeze or as one more round of forced selling. Here is the map.&lt;/p&gt;
&lt;h2 id="five-things-to-take-away"&gt;Five things to take away
&lt;/h2&gt;&lt;ul&gt;
&lt;li&gt;Only the AI chain fell. The S&amp;amp;P 500 dropped 2.5% while SOX dropped 18.4%. Memory longs and neocloud shorts both fell hard, so many funds lost on both sides at once.&lt;/li&gt;
&lt;li&gt;AI-specific news lit the fire and hedge-fund long sales spread it. Goldman Sachs prime brokerage data showed AI long sales running at 3.5 times short sales in the week to Oct 14.&lt;/li&gt;
&lt;li&gt;The book left behind is a hedged crowd: long whoever gets paid by AI, short the index and the debt-funded builders. Its weak point is correlation. If longs fall while shorts rise, the loss is about 3.3 times larger than when both fall together.&lt;/li&gt;
&lt;li&gt;Korea fell on foreign selling of cash equities, not on index shorts. A lasting rebound needs foreigners to buy Samsung Electronics and SK hynix shares again, not just to cover futures.&lt;/li&gt;
&lt;li&gt;Our base case (50%, subjective) is a short cover once the event cluster passes. One more leg down has a combined 35%.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;Prices are Oct 23 closes. Korean closes are the final prints after the extended session ends at 20:00 KST. Probabilities are our subjective estimates as of Oct 23. This is analysis, not a trading instruction.&lt;/p&gt;
&lt;h2 id="only-the-ai-chain-fell-and-both-sides-of-the-book-fell-with-it"&gt;Only the AI chain fell, and both sides of the book fell with it
&lt;/h2&gt;&lt;p&gt;The first question is simple: did the market fall, or did one trade fall? Blue bars are names hedge funds typically hold long. Orange bars are what they use as hedges or short.&lt;/p&gt;
&lt;div class="fig-scroll" style="overflow-x:auto;-webkit-overflow-scrolling:touch;margin:1.4em 0"&gt;&lt;img src="https://koreainvestinsights.com/post-assets/october-chip-selloff-long-unwind-not-short-attack-2026-10-24/drawdown.en.svg" alt="Decline from the October closing high to the Oct 23 close: S&amp;P 500 −2.5%, SOX −18.4%, Micron −20.0%, KOSPI −20.8%, Samsung Electronics −22.6%, SK hynix −31.4%, CoreWeave −40.9%" loading="lazy" style="display:block;width:100%;min-width:720px;max-width:none;height:auto;margin:0 auto"&gt;&lt;/div&gt;
&lt;p&gt;The broad index barely moved while every AI-linked name fell by a fifth or more. The memory longs and the neocloud shorts went down together, which is the painful case for a long/short book. Over the same weeks DRAM spot prices rose. The closes behind the chart are below.&lt;sup id="fnref:1"&gt;&lt;a href="#fn:1" class="footnote-ref" role="doc-noteref"&gt;1&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Name&lt;/th&gt;
 &lt;th&gt;Role in the book&lt;/th&gt;
 &lt;th&gt;October closing high&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Oct 23 close&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Change&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;S&amp;amp;P 500&lt;/td&gt;
 &lt;td&gt;Index hedge&lt;/td&gt;
 &lt;td&gt;Oct 6: 7,818.93&lt;/td&gt;
 &lt;td style="text-align: right"&gt;7,623.68&lt;/td&gt;
 &lt;td style="text-align: right"&gt;−2.5%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;SOX&lt;/td&gt;
 &lt;td&gt;Index hedge&lt;/td&gt;
 &lt;td&gt;Oct 6: 13,217.82&lt;/td&gt;
 &lt;td style="text-align: right"&gt;10,779.46&lt;/td&gt;
 &lt;td style="text-align: right"&gt;−18.4%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Micron&lt;/td&gt;
 &lt;td&gt;Long&lt;/td&gt;
 &lt;td&gt;Oct 1: $1,097.39&lt;/td&gt;
 &lt;td style="text-align: right"&gt;$877.55&lt;/td&gt;
 &lt;td style="text-align: right"&gt;−20.0%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;KOSPI&lt;/td&gt;
 &lt;td&gt;Index hedge&lt;/td&gt;
 &lt;td&gt;Oct 2: 7,003.74&lt;/td&gt;
 &lt;td style="text-align: right"&gt;5,550.30&lt;/td&gt;
 &lt;td style="text-align: right"&gt;−20.8%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Samsung Electronics&lt;/td&gt;
 &lt;td&gt;Long&lt;/td&gt;
 &lt;td&gt;Oct 2: KRW 276,000&lt;/td&gt;
 &lt;td style="text-align: right"&gt;KRW 213,500&lt;/td&gt;
 &lt;td style="text-align: right"&gt;−22.6%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;SK hynix&lt;/td&gt;
 &lt;td&gt;Long&lt;/td&gt;
 &lt;td&gt;Oct 2: KRW 1,842,000&lt;/td&gt;
 &lt;td style="text-align: right"&gt;KRW 1,263,000&lt;/td&gt;
 &lt;td style="text-align: right"&gt;−31.4%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;CoreWeave&lt;/td&gt;
 &lt;td&gt;Debt-funded short&lt;/td&gt;
 &lt;td&gt;Oct 6: $91.72&lt;/td&gt;
 &lt;td style="text-align: right"&gt;$54.24&lt;/td&gt;
 &lt;td style="text-align: right"&gt;−40.9%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;h2 id="ai-news-lit-the-fire-hedge-fund-long-sales-spread-it"&gt;AI news lit the fire; hedge-fund long sales spread it
&lt;/h2&gt;&lt;p&gt;The next question is cause. The chain below runs from the trigger to the buyers who absorbed the supply.&lt;/p&gt;
&lt;div class="fig-scroll" style="overflow-x:auto;-webkit-overflow-scrolling:touch;margin:1.4em 0"&gt;&lt;img src="https://koreainvestinsights.com/post-assets/october-chip-selloff-long-unwind-not-short-attack-2026-10-24/causes.en.svg" alt="Causal chain: AI-specific news and funding fears triggered the selloff, hedge-fund long liquidation amplified it into the U.S. AI chain and foreign cash selling in Korea, and retail leverage and the SK hynix buyback absorbed the supply" loading="lazy" style="display:block;width:100%;min-width:720px;max-width:none;height:auto;margin:0 auto"&gt;&lt;/div&gt;
&lt;p&gt;The trigger was news specific to AI. Lam Research and Meta reported on Oct 21 U.S. time, and the Oct 22 drop followed. Lam guided the December quarter to $8.6 billion, plus or minus $0.4 billion, slightly above consensus, yet the stock fell about 7% after hours.&lt;sup id="fnref:2"&gt;&lt;a href="#fn:2" class="footnote-ref" role="doc-noteref"&gt;2&lt;/a&gt;&lt;/sup&gt; We read that guidance as a supply ceiling, not weaker demand. Meta raised the low end of its 2026 capital spending range to $135 billion to $145 billion, kept the top and gave no 2027 number.&lt;sup id="fnref:3"&gt;&lt;a href="#fn:3" class="footnote-ref" role="doc-noteref"&gt;3&lt;/a&gt;&lt;/sup&gt; Credit added a funding scare: Oracle&amp;rsquo;s five-year CDS was near 315bp and CoreWeave&amp;rsquo;s 2032 notes yielded about 14.9% on Oct 19.&lt;sup id="fnref:4"&gt;&lt;a href="#fn:4" class="footnote-ref" role="doc-noteref"&gt;4&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;Hedge funds turned that news into a rout. In Goldman Sachs prime brokerage data, AI long sales ran at 3.5 times short sales over Oct 8 to 14.&lt;sup id="fnref:5"&gt;&lt;a href="#fn:5" class="footnote-ref" role="doc-noteref"&gt;5&lt;/a&gt;&lt;/sup&gt; Tech de-grossing in the five sessions to Oct 15 was the largest since the data began in 2016.&lt;sup id="fnref:6"&gt;&lt;a href="#fn:6" class="footnote-ref" role="doc-noteref"&gt;6&lt;/a&gt;&lt;/sup&gt; In the week of Oct 16, gross leverage stood at 311.7% and net leverage at 72.7%.&lt;sup id="fnref:7"&gt;&lt;a href="#fn:7" class="footnote-ref" role="doc-noteref"&gt;7&lt;/a&gt;&lt;/sup&gt; Funds kept their gross exposure and cut their net.&lt;/p&gt;
&lt;p&gt;Two groups took the other side. U.S. retail investors put about $10 billion into leveraged ETFs in one week, including a record $7 billion into SOXL.&lt;sup id="fnref:8"&gt;&lt;a href="#fn:8" class="footnote-ref" role="doc-noteref"&gt;8&lt;/a&gt;&lt;/sup&gt; In Korea, SK hynix&amp;rsquo;s KRW 40 trillion buyback absorbed supply until it ended on Oct 21. By our estimate from investor-flow data it had bought about KRW 39.4 trillion.&lt;sup id="fnref:9"&gt;&lt;a href="#fn:9" class="footnote-ref" role="doc-noteref"&gt;9&lt;/a&gt;&lt;/sup&gt; Assets in Korea&amp;rsquo;s KODEX leveraged ETF were about five times those in the 2x inverse fund.&lt;sup id="fnref:10"&gt;&lt;a href="#fn:10" class="footnote-ref" role="doc-noteref"&gt;10&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;The physical market did not confirm the panic. DDR5 16Gb spot averaged $61.10 on Oct 22, up 0.8% from Oct 19 and 11.4% from Sept 16.&lt;sup id="fnref:11"&gt;&lt;a href="#fn:11" class="footnote-ref" role="doc-noteref"&gt;11&lt;/a&gt;&lt;/sup&gt; The stocks sold the durability of memory profits beyond 2027, not today&amp;rsquo;s prices.&lt;/p&gt;
&lt;p&gt;Much of the mechanical selling is already done. Goldman&amp;rsquo;s model of CTA global equity longs fell from about $125 billion to $39 billion, down 69%.&lt;sup id="fnref:12"&gt;&lt;a href="#fn:12" class="footnote-ref" role="doc-noteref"&gt;12&lt;/a&gt;&lt;/sup&gt; In the BofA fund manager survey, the share calling semiconductors the most crowded trade fell from 82% to 45%.&lt;sup id="fnref:13"&gt;&lt;a href="#fn:13" class="footnote-ref" role="doc-noteref"&gt;13&lt;/a&gt;&lt;/sup&gt; The fuel left for another drop sits mostly with leveraged retail.&lt;/p&gt;
&lt;h2 id="the-book-is-a-hedged-crowd-long-who-gets-paid-short-who-borrows"&gt;The book is a hedged crowd: long who gets paid, short who borrows
&lt;/h2&gt;&lt;p&gt;What do the funds hold now? The map below shows the book behind the selloff.&lt;/p&gt;
&lt;div class="fig-scroll" style="overflow-x:auto;-webkit-overflow-scrolling:touch;margin:1.4em 0"&gt;&lt;img src="https://koreainvestinsights.com/post-assets/october-chip-selloff-long-unwind-not-short-attack-2026-10-24/book.en.svg" alt="Hedged-crowding book: longs in memory and Korean chip stocks; shorts in index hedges and in debt-funded neoclouds" loading="lazy" style="display:block;width:100%;min-width:720px;max-width:none;height:auto;margin:0 auto"&gt;&lt;/div&gt;
&lt;p&gt;The logic has three parts.&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Why own memory.&lt;/strong&gt; A large share of 2026 and 2027 memory profits is already set by long-term contracts. The more AI spends, the more these companies get paid.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Why short the index instead of selling memory.&lt;/strong&gt; Everyone owns the same longs, so selling them moves prices against you. Index and ETF shorts are cheaper and faster, and they fit multi-manager risk limits. Goldman&amp;rsquo;s short leverage was at the 99th percentile of the past five years.&lt;sup id="fnref:14"&gt;&lt;a href="#fn:14" class="footnote-ref" role="doc-noteref"&gt;14&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Why short the neoclouds.&lt;/strong&gt; Companies that build data centers with debt and leases are hurt first when rates and credit wobble. Short interest in five neoclouds reached about $17 billion, or 20% to 26% of their float.&lt;sup id="fnref:15"&gt;&lt;a href="#fn:15" class="footnote-ref" role="doc-noteref"&gt;15&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;The weak point of this book is correlation, not direction. The chart below uses a simplified book of 193% long and 120% short of equity, close to the reported gross and net.&lt;/p&gt;
&lt;div class="fig-scroll" style="overflow-x:auto;-webkit-overflow-scrolling:touch;margin:1.4em 0"&gt;&lt;img src="https://koreainvestinsights.com/post-assets/october-chip-selloff-long-unwind-not-short-attack-2026-10-24/asymmetry.en.svg" alt="P&amp;amp;L when correlation breaks: −3.65 points of equity if longs and shorts both fall 5%; −12.05 points if longs fall 5% while shorts rise 2%" loading="lazy" style="display:block;width:100%;min-width:720px;max-width:none;height:auto;margin:0 auto"&gt;&lt;/div&gt;
&lt;p&gt;If longs and short targets both fall 5%, the book loses 193% × 5% − 120% × 5% = 3.65 points of equity. If longs fall 5% while short targets rise 2%, it loses 9.65 + 2.40 = 12.05 points, 3.3 times more. That is when risk managers force de-grossing. So good news tends to produce a violent short-covering rally, and bad news on the long side produces the opposite: crowded longs down, shorted names up.&lt;/p&gt;
&lt;h2 id="korea-fell-on-foreign-cash-selling-not-index-shorts"&gt;Korea fell on foreign cash selling, not index shorts
&lt;/h2&gt;&lt;p&gt;Korea is the clearest test of the long-unwind reading. Blue bars are foreign net buying of KOSPI cash equities. Orange bars are foreign net buying of KOSPI 200 futures.&lt;/p&gt;
&lt;div class="fig-scroll" style="overflow-x:auto;-webkit-overflow-scrolling:touch;margin:1.4em 0"&gt;&lt;img src="https://koreainvestinsights.com/post-assets/october-chip-selloff-long-unwind-not-short-attack-2026-10-24/korea-flows.en.svg" alt="Foreign net buying of KOSPI cash and KOSPI 200 futures by window, in KRW trillion: cash sold heavily in each selloff while futures were bought" loading="lazy" style="display:block;width:100%;min-width:720px;max-width:none;height:auto;margin:0 auto"&gt;&lt;/div&gt;
&lt;p&gt;In every sharp decline, foreigners sold cash equities in size and bought futures. Index shorts did not build up. The table gives the windows; futures figures are estimates.&lt;sup id="fnref:16"&gt;&lt;a href="#fn:16" class="footnote-ref" role="doc-noteref"&gt;16&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Window&lt;/th&gt;
 &lt;th style="text-align: right"&gt;KOSPI&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Foreign cash (KRW tn)&lt;/th&gt;
 &lt;th style="text-align: right"&gt;KOSPI 200 futures (KRW tn)&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Jun 23 to Jul 29&lt;/td&gt;
 &lt;td style="text-align: right"&gt;−37.9%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;−44.0&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+8.0&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Jul 30 to Sep 23&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+25.0%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;−14.5&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+1.9&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Sep 24 to Oct 15&lt;/td&gt;
 &lt;td style="text-align: right"&gt;−21.6%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;−27.0&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+0.6&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Oct 12 to 15 (part of the above)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;−16.2%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;−11.5&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+3.2&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Oct 16 to 23&lt;/td&gt;
 &lt;td style="text-align: right"&gt;−0.1%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;−0.3&lt;/td&gt;
 &lt;td style="text-align: right"&gt;−1.45&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Short positions point the same way. The short balance in Samsung Electronics fell from 0.12% to 0.07% of shares, and short interest in the iShares MSCI South Korea ETF (EWY) fell 64% from mid-July to the end of September.&lt;sup id="fnref:17"&gt;&lt;a href="#fn:17" class="footnote-ref" role="doc-noteref"&gt;17&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;p&gt;Since Oct 16 the pattern has changed. Foreign cash selling has nearly stopped, and futures are sold on up days. That looks like keeping the longs and hedging the index. We sort each Korean session into one of four boxes.&lt;/p&gt;
&lt;div class="fig-scroll" style="overflow-x:auto;-webkit-overflow-scrolling:touch;margin:1.4em 0"&gt;&lt;img src="https://koreainvestinsights.com/post-assets/october-chip-selloff-long-unwind-not-short-attack-2026-10-24/quadrant.en.svg" alt="Four-box map of foreign cash and futures flows: selloffs were cash selling plus futures buying; Oct 16 to 23 moved to flat cash plus futures selling; a lasting rebound needs cash and futures buying together" loading="lazy" style="display:block;width:100%;min-width:720px;max-width:none;height:auto;margin:0 auto"&gt;&lt;/div&gt;
&lt;p&gt;The chart has no scale; the dots show direction only. Futures covering alone will not turn Korea. The futures short built since Oct 16 is about KRW 1.45 trillion, while cash flows move KRW 1 to 4 trillion on a typical day. A recovery needs box ①: foreign net buying of Samsung Electronics and SK hynix for at least two sessions, together with futures buying. Expiry and roll days do not count.&lt;/p&gt;
&lt;h2 id="the-next-two-weeks-decide-it-squeeze-or-one-more-leg-down"&gt;The next two weeks decide it: squeeze or one more leg down
&lt;/h2&gt;&lt;p&gt;A cluster of events lands between Oct 27 and Nov 3. Times are Korea time (KST); U.S. events after the close arrive the next morning in Korea.&lt;sup id="fnref:18"&gt;&lt;a href="#fn:18" class="footnote-ref" role="doc-noteref"&gt;18&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;KST&lt;/th&gt;
 &lt;th&gt;Event&lt;/th&gt;
 &lt;th&gt;What it tests&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Oct 27, morning&lt;/td&gt;
 &lt;td&gt;FINRA short interest (released Oct 26 U.S.)&lt;/td&gt;
 &lt;td&gt;Did neocloud shorts hold or grow? That is squeeze fuel&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Oct 28&lt;/td&gt;
 &lt;td&gt;SK hynix third-quarter results&lt;/td&gt;
 &lt;td&gt;Memory contract visibility into 2027&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Oct 29, 00:55&lt;/td&gt;
 &lt;td&gt;Oracle investor day&lt;/td&gt;
 &lt;td&gt;Funding plans for debt-built capacity&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Oct 29, 03:00&lt;/td&gt;
 &lt;td&gt;FOMC decision&lt;/td&gt;
 &lt;td&gt;Rates for the borrowers&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Oct 29, early morning&lt;/td&gt;
 &lt;td&gt;Alphabet and Microsoft results&lt;/td&gt;
 &lt;td&gt;Capex and demand wording&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Oct 29&lt;/td&gt;
 &lt;td&gt;Samsung Electronics results call&lt;/td&gt;
 &lt;td&gt;Memory pricing and HBM&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Oct 30, early morning&lt;/td&gt;
 &lt;td&gt;Amazon and Sandisk results&lt;/td&gt;
 &lt;td&gt;Capex and memory cost absorption&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Nov 3&lt;/td&gt;
 &lt;td&gt;U.S. midterm elections&lt;/td&gt;
 &lt;td&gt;Policy risk&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;We weigh four paths to about Nov 6.&lt;/p&gt;
&lt;div class="fig-scroll" style="overflow-x:auto;-webkit-overflow-scrolling:touch;margin:1.4em 0"&gt;&lt;img src="https://koreainvestinsights.com/post-assets/october-chip-selloff-long-unwind-not-short-attack-2026-10-24/paths.en.svg" alt="Short-cover paths to about Nov 6, subjective probabilities as of Oct 23: A short cover after events pass 50%, B one more de-grossing then cover 25%, C range-bound without cover 15%, D credit event that proves the shorts right 10%" loading="lazy" style="display:block;width:100%;min-width:720px;max-width:none;height:auto;margin:0 auto"&gt;&lt;/div&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;A. Short cover after the events pass (50%).&lt;/strong&gt; Capex is kept or raised and the FOMC is uneventful. Funds lift hedges and the most shorted names outperform. In Korea the rally lasts only if foreign cash buying returns.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;B. One more round of de-grossing, then a cover (25%).&lt;/strong&gt; Bad news on the long side pushes crowded longs down while shorted names rise. Forced cuts take prices lower first. Covering during de-grossing is not a rebound.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;C. Range-bound without a cover (15%).&lt;/strong&gt; The events pass, volatility stays high, hedges stay on and the market has no direction.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;D. A credit event that proves the shorts right (10%).&lt;/strong&gt; High-yield spreads above 350bp, Oracle CDS stuck above 300bp and CoreWeave bonds at new lows would mean the funding chain is breaking.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;If Big Tech puts 2027 capex on hold, we move 15 points from A to B. Together, the two downside paths (B and D) carry 35%.&lt;/p&gt;
&lt;h3 id="signals-that-tilt-toward-a-squeeze"&gt;Signals that tilt toward a squeeze
&lt;/h3&gt;&lt;ul&gt;
&lt;li&gt;Microsoft&amp;rsquo;s wording passes at least three of four tests: demand still exceeds supply; calendar 2026 capex is raised or the December-quarter capex is at least the September quarter&amp;rsquo;s; leases not yet commenced keep growing; Azure grows 45% or more. We put the odds near 42%.&lt;/li&gt;
&lt;li&gt;Alphabet and Amazon keep or raise capex and absorb higher memory prices without cutting build plans.&lt;/li&gt;
&lt;li&gt;FINRA data show neocloud shorts held or grew as of Oct 15. That is the fuel.&lt;/li&gt;
&lt;li&gt;Floors hold: the S&amp;amp;P 500 above 7,508 (Goldman&amp;rsquo;s medium-term CTA level), high-yield spreads below 350bp, CoreWeave&amp;rsquo;s 2032 notes above 75 cents.&lt;sup id="fnref1:12"&gt;&lt;a href="#fn:12" class="footnote-ref" role="doc-noteref"&gt;12&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;li&gt;In Korea, foreigners buy Samsung Electronics and SK hynix shares for two or more sessions while also turning buyers of futures.&lt;/li&gt;
&lt;li&gt;Oracle&amp;rsquo;s five-year CDS stays below 300bp for five sessions.&lt;/li&gt;
&lt;/ul&gt;
&lt;h3 id="signals-that-tilt-toward-another-drop"&gt;Signals that tilt toward another drop
&lt;/h3&gt;&lt;ul&gt;
&lt;li&gt;Big Tech puts 2027 capex on hold or cuts it.&lt;/li&gt;
&lt;li&gt;Credit: high-yield spreads above 350bp, Oracle CDS above 300bp for days, CoreWeave bonds at new lows.&lt;/li&gt;
&lt;li&gt;Trend followers: an S&amp;amp;P 500 close below 7,508 risks flipping CTAs short. Goldman estimated about $143.8 billion of selling in a one-month decline.&lt;sup id="fnref2:12"&gt;&lt;a href="#fn:12" class="footnote-ref" role="doc-noteref"&gt;12&lt;/a&gt;&lt;/sup&gt;&lt;/li&gt;
&lt;li&gt;Retail leverage unwinds: large redemptions from SOXL or Korea&amp;rsquo;s KODEX leveraged ETF.&lt;/li&gt;
&lt;li&gt;In Korea, foreign cash selling resumes (KRW 1 trillion or more in a day, or KRW 3 trillion or more over five sessions) together with futures buying.&lt;/li&gt;
&lt;/ul&gt;
&lt;h3 id="where-a-squeeze-would-bite-hardest"&gt;Where a squeeze would bite hardest
&lt;/h3&gt;&lt;p&gt;Ranked by squeeze sensitivity, the names are CoreWeave (CRWV), Core Scientific (CORZ), Nebius (NBIS), Galaxy Digital (GLXY), Hut 8 (HUT) and Teradyne (TER). We leave out IREN, TeraWulf (WULF), Fluence (FLNC) and Bloom Energy (BE): heavy convertible-bond hedging means hedge selling appears as they rise. IREN and Nebius are more sensitive to Microsoft&amp;rsquo;s wording than CoreWeave. This is a ranking of sensitivity, not a recommendation.&lt;/p&gt;
&lt;p&gt;History leans against fresh shorts after big drops. From January 2025 to October 2026, a day when SOX fell 3% or more was followed by an average 20-day gain of 5.1%, with gains 64% of the time. A day when SK hynix fell 5% or more was followed by an average 20-day gain of 9.0%. The exception is Korea&amp;rsquo;s rallies: after a KOSPI gain of 3% or more, the next 20 days averaged 0.0% with gains only 42% of the time, and the median low within 10 days was 5.1% below. Rallies without foreign cash buying were sold.&lt;sup id="fnref:19"&gt;&lt;a href="#fn:19" class="footnote-ref" role="doc-noteref"&gt;19&lt;/a&gt;&lt;/sup&gt; That sample is a bull market, so treat it as a base rate, not a forecast.&lt;/p&gt;
&lt;h2 id="your-turn"&gt;Your turn
&lt;/h2&gt;&lt;p&gt;Which path do you think we are on, A, B, C or D? And which single signal would change your mind? Tell me in the comments. Push back where you think the map is wrong.&lt;/p&gt;
&lt;p&gt;If this map helped you, a like helps other investors find it.&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Disclaimer: For research and information purposes only. Not investment advice. Names cited are for analytical illustration; readers should perform their own due diligence and consult licensed advisors before any investment decision.&lt;/em&gt;&lt;/p&gt;
&lt;div class="footnotes" role="doc-endnotes"&gt;
&lt;hr&gt;
&lt;ol&gt;
&lt;li id="fn:1"&gt;
&lt;p&gt;Daily closes retrieved Oct 24, 2026: S&amp;amp;P 500 from Yahoo Finance; SOX, Micron and CoreWeave from Nasdaq; KOSPI, Samsung Electronics and SK hynix from Naver Finance (KRX final closes after the extended session). Earlier figures that used provisional Korean closes were replaced.&amp;#160;&lt;a href="#fnref:1" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:2"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/0000707549/000070754926000052/lrcx_exhibitx991xq1x2027.htm" target="_blank" rel="noopener"
 &gt;Lam Research, results for the quarter ended September 27, 2026 (SEC Form 8-K, Exhibit 99.1)&lt;/a&gt;, 2026-10-21. The supply-ceiling reading is our interpretation.&amp;#160;&lt;a href="#fnref:2" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:3"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.sec.gov/Archives/edgar/data/0001326801/000162828026066700/meta-09302026xexhibit991.htm" target="_blank" rel="noopener"
 &gt;Meta Platforms, third-quarter 2026 results (SEC Form 8-K, Exhibit 99.1)&lt;/a&gt;, 2026-10-21.&amp;#160;&lt;a href="#fnref:3" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:4"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.cnbc.com/2026/10/19/coreweave-oracle-ai-credit-stress.html" target="_blank" rel="noopener"
 &gt;CNBC, report on CoreWeave and Oracle credit stress&lt;/a&gt;, 2026-10-19. Oracle five-year CDS at 315bp and CoreWeave&amp;rsquo;s 9.625% 2032 notes at about 14.9% yield as reported; other sources quoted Oracle CDS near 322bp a day later. The 75-cent line is our own threshold.&amp;#160;&lt;a href="#fnref:4" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:5"&gt;
&lt;p&gt;Goldman Sachs Prime Brokerage data on AI-stock flows for Oct 8 to 14, as shared by The Daily Shot on 2026-10-16. Secondary; no public link to the original note.&amp;#160;&lt;a href="#fnref:5" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:6"&gt;
&lt;p&gt;&lt;a class="link" href="https://finance.yahoo.com/markets/stocks/articles/hedge-funds-dumped-tech-stocks-133527326.html" target="_blank" rel="noopener"
 &gt;Bloomberg via Yahoo Finance, hedge funds dumped tech stocks&lt;/a&gt;, 2026-10-19. Goldman Sachs prime brokerage data.&amp;#160;&lt;a href="#fnref:6" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:7"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.tickmill.com/blog/institutional-insights-goldman-sachs-tech-liquidation-retial-etf-chase" target="_blank" rel="noopener"
 &gt;Tickmill, summary of Goldman Sachs notes on the tech liquidation and the retail ETF chase&lt;/a&gt;, 2026-10-20. Gross 311.7% and net 72.7% for the week of Oct 16. Secondary summary of a paid note.&amp;#160;&lt;a href="#fnref:7" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:8"&gt;
&lt;p&gt;Same Tickmill summary of Goldman Sachs notes, 2026-10-20: about $10 billion into leveraged ETFs in one week and a record $7 billion into SOXL. Secondary.&amp;#160;&lt;a href="#fnref:8" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:9"&gt;
&lt;p&gt;&lt;a class="link" href="https://news.skhynix.com/en/share-buyback-and-retirement/" target="_blank" rel="noopener"
 &gt;SK hynix newsroom, share buyback and retirement&lt;/a&gt;, program of KRW 40 trillion approved 2026-08-19; completion reported by &lt;a class="link" href="https://biz.chosun.com/stock/stock_general/2026/10/21/7JYPV2XQN5CFNJXAFK4PVZQJTA/?outputType=amp" target="_blank" rel="noopener"
 &gt;Chosun Biz&lt;/a&gt; on 2026-10-21. The KRW 39.4 trillion executed is our estimate from &amp;ldquo;other corporations&amp;rdquo; net buying in KRX investor data, not a disclosed figure.&amp;#160;&lt;a href="#fnref:9" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:10"&gt;
&lt;p&gt;Naver Finance ETF data as of 2026-10-23: KODEX Leverage at about KRW 4.41 trillion in market value against about KRW 0.86 trillion for KODEX 200 Futures Inverse 2X.&amp;#160;&lt;a href="#fnref:10" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:11"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.trendforce.com/price/dram/dram_spot" target="_blank" rel="noopener"
 &gt;TrendForce, DRAM spot price&lt;/a&gt;, DDR5 16Gb session average $61.10 on 2026-10-22 against $54.83 on 2026-09-16, from a captured price table.&amp;#160;&lt;a href="#fnref:11" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:12"&gt;
&lt;p&gt;Tickmill summaries of Goldman Sachs flow notes: &lt;a class="link" href="https://www.tickmill.com/blog/institutional-insights-goldman-sachs-sp500-flows-positioning-update-131026" target="_blank" rel="noopener"
 &gt;Oct 13&lt;/a&gt; (about $143.8 billion of CTA selling in a one-month decline), &lt;a class="link" href="https://www.tickmill.com/blog/institutional-insights-goldman-sachs-flow-of-funds-update-151026" target="_blank" rel="noopener"
 &gt;Oct 15&lt;/a&gt; and &lt;a class="link" href="https://www.tickmill.com/blog/institutional-insights-goldman-sachs-tech-liquidation-retial-etf-chase" target="_blank" rel="noopener"
 &gt;Oct 20&lt;/a&gt; (CTA global equity longs from about $125 billion to $39 billion; medium-term S&amp;amp;P 500 level 7,508). Secondary summaries of paid notes.&amp;#160;&lt;a href="#fnref:12" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref1:12" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&amp;#160;&lt;a href="#fnref2:12" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:13"&gt;
&lt;p&gt;&lt;a class="link" href="https://macenews.com/bofa-global-research-fund-manager-survey-lower-growth-and-lower-inflation-expectations-lead-to-tweaking-of-asset-allocations-in-october/" target="_blank" rel="noopener"
 &gt;Mace News, BofA Global Fund Manager Survey for October&lt;/a&gt;, 2026-10-13. Secondary summary.&amp;#160;&lt;a href="#fnref:13" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:14"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.bloomberg.com/news/articles/2026-10-22/hedge-funds-pile-into-shorts-as-goldman-flags-stock-squeeze-risk" target="_blank" rel="noopener"
 &gt;Bloomberg, hedge funds pile into shorts as Goldman flags stock squeeze risk&lt;/a&gt;, 2026-10-22. The 193% long and 120% short split in our chart is a rounded reading of shared prime brokerage charts, consistent with the gross and net figures above, and is used for illustration.&amp;#160;&lt;a href="#fnref:14" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:15"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.reuters.com/business/short-sellers-pocket-6-billion-profits-neoclouds-slide-ai-debt-fears-2026-10-22/" target="_blank" rel="noopener"
 &gt;Reuters, short sellers pocket $6 billion in profits as neoclouds slide on AI debt fears&lt;/a&gt;, 2026-10-22, citing ORTEX data for CoreWeave, IREN, Nebius, Applied Digital and Core Scientific.&amp;#160;&lt;a href="#fnref:15" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:16"&gt;
&lt;p&gt;Naver Finance investor-trend data (KRX based). Foreign cash flows by window are our sums of daily data; the Oct 16 to 23 total of about KRW −0.31 trillion was rechecked on Oct 24. KOSPI 200 futures amounts are estimates scaled against broker data and are used for direction.&amp;#160;&lt;a href="#fnref:16" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:17"&gt;
&lt;p&gt;Samsung Electronics short balance from AlphaSquare (6.77 million shares on Sep 17 to 4.17 million on Oct 21; secondary). EWY short interest from &lt;a class="link" href="https://api.finra.org/data/group/otcMarket/name/consolidatedShortInterest" target="_blank" rel="noopener"
 &gt;FINRA consolidated short interest&lt;/a&gt;: 36.97 million shares on Jul 15 to 13.32 million on Sep 30. October data were not yet published.&amp;#160;&lt;a href="#fnref:17" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:18"&gt;
&lt;p&gt;&lt;a class="link" href="https://www.finra.org/filing-reporting/regulatory-filing-systems/short-interest" target="_blank" rel="noopener"
 &gt;FINRA short interest schedule&lt;/a&gt;; &lt;a class="link" href="https://news.skhynix.com/en/q3-2026-earnings-conference-call-invitation/" target="_blank" rel="noopener"
 &gt;SK hynix earnings call notice&lt;/a&gt;; &lt;a class="link" href="https://www.prnewswire.com/news-releases/oracle-investor-day-meeting-to-be-held-october-28-2026-302918062.html" target="_blank" rel="noopener"
 &gt;Oracle investor day notice&lt;/a&gt;; &lt;a class="link" href="https://www.federalreserve.gov/monetarypolicy/fomccalendars.htm" target="_blank" rel="noopener"
 &gt;Federal Reserve FOMC calendar&lt;/a&gt;; &lt;a class="link" href="https://news.microsoft.com/source/2026/10/07/microsoft-announces-quarterly-earnings-release-date-69/" target="_blank" rel="noopener"
 &gt;Microsoft earnings date&lt;/a&gt;; &lt;a class="link" href="https://abc.xyz/investor/news/news-details/2026/Alphabet-Announces-Date-of-Third-Quarter-2026-Financial-Results-Conference-Call-2026-8tpGZsLS6v/default.aspx" target="_blank" rel="noopener"
 &gt;Alphabet earnings date&lt;/a&gt;; &lt;a class="link" href="https://ir.aboutamazon.com/news-release/news-release-details/2026/Amazon-com-to-Webcast-Third-Quarter-2026-Financial-Results-Conference-Call/default.aspx" target="_blank" rel="noopener"
 &gt;Amazon earnings webcast&lt;/a&gt;; &lt;a class="link" href="https://www.sandisk.com/company/newsroom/press-releases/2026/2026-09-29-sandisk-to-report-first-quarter-fiscal-year-2027-results-on-october-29-2026" target="_blank" rel="noopener"
 &gt;Sandisk results date&lt;/a&gt;; &lt;a class="link" href="https://www.samsung.com/global/ir/ir-events-presentations/events/" target="_blank" rel="noopener"
 &gt;Samsung Electronics IR events&lt;/a&gt;.&amp;#160;&lt;a href="#fnref:18" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;li id="fn:19"&gt;
&lt;p&gt;Our calculation from daily closes, January 2025 to October 2026. The sample is mostly a rising market.&amp;#160;&lt;a href="#fnref:19" class="footnote-backref" role="doc-backlink"&gt;&amp;#x21a9;&amp;#xfe0e;&lt;/a&gt;&lt;/p&gt;
&lt;/li&gt;
&lt;/ol&gt;
&lt;/div&gt;</description></item><item><title>CXMT 4F²: Bypassing EUV Does Not Remove Yield and Good-Die Cost Challenges</title><link>https://koreainvestinsights.com/post/samsung-cxmt-profit-duration-2027-contracts/</link><pubDate>Wed, 14 Oct 2026 06:00:00 +0900</pubDate><guid>https://koreainvestinsights.com/post/samsung-cxmt-profit-duration-2027-contracts/</guid><description>&lt;p&gt;For CXMT&amp;rsquo;s 4F², the number to watch beyond cell area is total cost per good bit. Higher storage density changes its economic competition with Samsung Electronics and SK hynix only if it can repeatedly ship qualified output that meets server requirements for speed, power, and reliability.&lt;/p&gt;
&lt;p&gt;On October 10, CXMT said it plans to unveil a server DDR5 RDIMM based on a 4F² structure by year-end. Evidence remains insufficient on speed, capacity, yield, repeat shipments, and customer qualification. Comparing that unveiling target with another company&amp;rsquo;s production target cannot establish a two-year lead.&lt;/p&gt;
&lt;p&gt;Changing the cell structure can create a path to the same storage density with a wider feature size. That matters to China because it could partly offset restrictions on access to EUV lithography equipment. &lt;strong&gt;The test is whether area gains translate into yield and cost gains.&lt;/strong&gt;&lt;/p&gt;
&lt;h2 id="reducing-lithography-constraints-does-not-remove-manufacturing-challenges"&gt;Reducing lithography constraints does not remove manufacturing challenges
&lt;/h2&gt;&lt;ul&gt;
&lt;li&gt;Moving from 6F² to 4F² reduces cell area by 33.3% and raises cell-array density by 50% at the same F. It does not improve good-die output or cost for the finished chip by the same percentages.&lt;/li&gt;
&lt;li&gt;G5 mass production is progress in a current product line. 4F² DDR5 is a separate announcement. G5&amp;rsquo;s 11.95nm active-area half-pitch cannot be used as the actual F of a 4F² cell or as the same metric as Samsung&amp;rsquo;s 1b or 1c nodes.&lt;/li&gt;
&lt;li&gt;Samsung&amp;rsquo;s ISSCC 16Gb test chip, CXMT&amp;rsquo;s IEDM device and array research, and SK hynix&amp;rsquo;s roadmap are each substantive. Research operation, product unveiling, customer qualification, and economically viable mass production are different stages.&lt;/li&gt;
&lt;li&gt;4F² does not replace HBM. Higher shipments of conventional DRAM can change customer choice and pricing power before CXMT catches up with HBM4.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;The reference date for technology and market data is October 13, 2026. CXMT&amp;rsquo;s listed entity is ChangXin Technology (688825.SH), listed on July 27. The article distinguishes official research, company announcements, reported news, inference, and unknowns. Yield and cost examples are illustrative assumptions.&lt;/p&gt;
&lt;h2 id="4f-vct-igzo-and-bonding-are-not-interchangeable-terms"&gt;4F², VCT, IGZO, and bonding are not interchangeable terms
&lt;/h2&gt;&lt;p&gt;A DRAM cell consists of a capacitor that stores charge and a transistor that controls access to it. A wordline carries the signal that opens the cell, while a bitline provides the path for reading and writing data. Shrinking the cell still requires managing stored charge, read signals, leakage, and interference.&lt;/p&gt;
&lt;p&gt;Bundling technologies from different levels under one name can distort the competitive picture.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Term&lt;/th&gt;
 &lt;th&gt;What it changes&lt;/th&gt;
 &lt;th&gt;What the term alone does not tell us&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;6F² or 4F²&lt;/td&gt;
 &lt;td&gt;Planar area occupied by one cell&lt;/td&gt;
 &lt;td&gt;Finished-chip density, cost, or yield&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;VCT, VG, VSG&lt;/td&gt;
 &lt;td&gt;Arrangement of the vertical channel and gate&lt;/td&gt;
 &lt;td&gt;Channel material or actual electrical behavior&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Oxide channel such as IGZO&lt;/td&gt;
 &lt;td&gt;Material through which current flows&lt;/td&gt;
 &lt;td&gt;Whether every supplier uses the same material&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;COP, or cell-on-peripheral&lt;/td&gt;
 &lt;td&gt;Placement of cells and control circuits&lt;/td&gt;
 &lt;td&gt;Bonding method or process cost&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Wafer hybrid bonding&lt;/td&gt;
 &lt;td&gt;Connection between dielectric and metal on two wafers&lt;/td&gt;
 &lt;td&gt;Bond yield, throughput, or total cost&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;3D DRAM&lt;/td&gt;
 &lt;td&gt;Stacking the cell array in multiple layers&lt;/td&gt;
 &lt;td&gt;Whether it is the same product as HBM&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;HBM&lt;/td&gt;
 &lt;td&gt;Stacking finished DRAM die and connecting them to an accelerator&lt;/td&gt;
 &lt;td&gt;A cost advantage in making a single cell&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;CXMT&amp;rsquo;s 2023 IEDM research used a junctionless GAA vertical device with a silicon pillar. Kioxia&amp;rsquo;s OCTRAM research uses an InGaZnO oxide channel. Samsung&amp;rsquo;s public abstract describes VCT, COP, and hybrid bonding, but does not specify the channel material. Comparisons such as “all 4F² uses IGZO” or “the same 4F² means the same process” do not hold. &lt;a class="link" href="https://mys.mapyourshow.com/mys_shared/iedm23/handouts/6-2_Mon_16771.pdf" target="_blank" rel="noopener"
 &gt;CXMT IEDM paper&lt;/a&gt;, &lt;a class="link" href="https://www.kioxia.com/en-jp/rd/technology/topics/topics-77.html" target="_blank" rel="noopener"
 &gt;Kioxia OCTRAM&lt;/a&gt;, &lt;a class="link" href="https://pure.skku.edu/en/publications/a-vertical-cell-transistor-based-4fsup2sup-dram-with-cell-on-peri/" target="_blank" rel="noopener"
 &gt;public abstract of Samsung&amp;rsquo;s joint research&lt;/a&gt;&lt;/p&gt;
&lt;h2 id="the-cell-area-gain-is-real-the-finished-chip-gain-is-smaller"&gt;The cell-area gain is real. The finished-chip gain is smaller.
&lt;/h2&gt;&lt;p&gt;F is the minimum dimension used in a geometric comparison. It is not always the same as a process-generation label such as “10a” or a half-pitch in a specific area. At the same F, the calculations are:&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Comparison&lt;/th&gt;
 &lt;th&gt;Calculation&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Result&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Reduction in cell area&lt;/td&gt;
 &lt;td&gt;1 − 4 ÷ 6&lt;/td&gt;
 &lt;td style="text-align: right"&gt;33.3%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Increase in cells per equal area&lt;/td&gt;
 &lt;td&gt;6 ÷ 4 − 1&lt;/td&gt;
 &lt;td style="text-align: right"&gt;50.0%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;F ratio for equal cell area&lt;/td&gt;
 &lt;td&gt;F₄ ÷ F₆ = √(6 ÷ 4)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;1.2247&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;The area of a 4F² cell with F = 15nm is 4 × 15² = 900nm². It is roughly equal to a 6F² cell with F ≈ 12.25nm. This illustrates how a wider feature size can yield the same cell area. It does not mean that CXMT&amp;rsquo;s actual product has been measured at F = 15nm.&lt;/p&gt;
&lt;p&gt;Structural changes can therefore partly offset lithography constraints. But they do not establish that CXMT&amp;rsquo;s first product is equivalent to or ahead of a specific Samsung 1b or 1c product. Compare actual cell area and full-chip Gb/mm², peripheral-circuit area, speed, power, and yield on the same basis.&lt;/p&gt;
&lt;p&gt;A finished chip also contains peripheral circuits. Under the hypothetical assumptions that the cells occupy 65% of the chip and the peripheral area remains unchanged, the new chip area is 0.35 + 0.65 × 4 ÷ 6 = 0.7833. The theoretical output increase is 1 ÷ 0.7833 − 1 = about 27.7%.&lt;/p&gt;
&lt;p&gt;The 65% is not an estimate of the cell share at Samsung or CXMT, and the calculation omits wafer-edge losses. It shows why a 50% increase in cell density cannot be transferred directly into a 50% increase in finished-chip output.&lt;/p&gt;
&lt;h2 id="do-not-conflate-g5s-two-different-50-increases"&gt;Do not conflate G5&amp;rsquo;s two different “50% increases”
&lt;/h2&gt;&lt;p&gt;In its official announcement on September 20, CXMT described G5 mass production and SAQP. SAQP means self-aligned quadruple patterning. It subdivides patterns across multiple process steps to increase density; it is different from simply exposing the same pattern four times.&lt;/p&gt;
&lt;p&gt;The company&amp;rsquo;s figures use different comparison bases.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Company-reported item&lt;/th&gt;
 &lt;th&gt;Actual comparison basis&lt;/th&gt;
 &lt;th&gt;What it should not be expanded to mean&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;11.95nm active-area half-pitch&lt;/td&gt;
 &lt;td&gt;A specific pattern dimension in the memory array&lt;/td&gt;
 &lt;td&gt;Line width of every layer, F for 4F², or equivalence to a competitor&amp;rsquo;s node&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;At least +50% DPW versus prior platform&lt;/td&gt;
 &lt;td&gt;Die per wafer normalized to an 8Gb baseline&lt;/td&gt;
 &lt;td&gt;+50% in good-die yield, sales, or cost improvement&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;+50% per-die capacity for 24Gb LPDDR5X&lt;/td&gt;
 &lt;td&gt;Per-die capacity versus equivalent prior-generation products&lt;/td&gt;
 &lt;td&gt;+50% total factory bit shipments&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Mass production of two 24Gb LPDDR5X products&lt;/td&gt;
 &lt;td&gt;Production stage claimed by the company&lt;/td&gt;
 &lt;td&gt;Completion of scale sales or repeat orders&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Source: &lt;a class="link" href="https://www.cxmt.com/en/news/info_22.html" target="_blank" rel="noopener"
 &gt;CXMT official G5 announcement and notes&lt;/a&gt;. This verifies what the company announced, not independent teardown or yield validation.&lt;/p&gt;
&lt;p&gt;A reprint of company announcement 2026-005 says G5 products had not yet achieved sales at scale, and that yield stabilization and new-customer validation and qualification would take time. This is the warning in the filing as of September. &lt;a class="link" href="https://stock.stockstar.com/notice/SN2026092200000007.shtml" target="_blank" rel="noopener"
 &gt;Reprinted G5 filing&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;G5 mass production and the absence of scale sales are not contradictory. Putting products on a process line differs from repeatedly supplying enough good die for customers to buy. G5 is a near-term supply development; year-end 4F² is a separate structural transition still to be verified.&lt;/p&gt;
&lt;h2 id="vertical-transistors-shrink-cells-but-make-device-control-harder"&gt;Vertical transistors shrink cells but make device control harder
&lt;/h2&gt;&lt;p&gt;A vertical channel makes the current path stand upright to reduce its planar footprint. A gate surrounding the channel can help control current and leakage. But variations in the etched pillar&amp;rsquo;s width, height, verticality, and surface condition can change device behavior.&lt;/p&gt;
&lt;p&gt;The process and electrical characteristics below need to be checked together before a design is a product.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Manufacturing challenge&lt;/th&gt;
 &lt;th&gt;Evidence to check&lt;/th&gt;
 &lt;th&gt;Why it matters to chips and servers&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Pillar etch and dimension distribution&lt;/td&gt;
 &lt;td&gt;Cross-sections, width, height, verticality, and within-wafer distribution&lt;/td&gt;
 &lt;td&gt;Differences in cell current and threshold voltage&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Gate and dielectric formation&lt;/td&gt;
 &lt;td&gt;Film thickness, uniformity, defects, and materials&lt;/td&gt;
 &lt;td&gt;Leakage, control, and long-term reliability&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Bitline resistance and parasitics&lt;/td&gt;
 &lt;td&gt;Resistance, junctions, and parasitic capacitance&lt;/td&gt;
 &lt;td&gt;Read signal, speed, and power&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Floating body and field-induced leakage&lt;/td&gt;
 &lt;td&gt;Electrical behavior by bias and temperature&lt;/td&gt;
 &lt;td&gt;Charge retention and errors&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Capacitor and sensing circuit&lt;/td&gt;
 &lt;td&gt;Charge, read margin, and distribution&lt;/td&gt;
 &lt;td&gt;Whether data can be distinguished in a smaller cell&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;This is a list of things to verify, not a qualification verdict for a public product. The issues vary with materials and implementation. A paper on another supplier&amp;rsquo;s device is not a measurement of CXMT&amp;rsquo;s product.&lt;/p&gt;
&lt;p&gt;CXMT&amp;rsquo;s IEDM 2023 paper reports that wordline-to-wordline leakage fell from 200µA to 2.5pA in a specific test structure. The ratio is 200 × 10⁻⁶ ÷ (2.5 × 10⁻¹²) = 80 million. This is a large improvement in the experiment. The measured leakage path differs from data retention, error rate, and production yield for a complete DDR5 module.&lt;/p&gt;
&lt;p&gt;Distinguish the existence of research data from validation of a server product&amp;rsquo;s 2026 economics. The 2023 paper alone cannot establish that the year-end product has the same channel material, structure, or performance.&lt;/p&gt;
&lt;h2 id="wafer-bonding-can-solve-circuit-placement-but-adds-cost-and-defect-risks"&gt;Wafer bonding can solve circuit placement but adds cost and defect risks
&lt;/h2&gt;&lt;p&gt;Building the cells and peripheral circuits on separate wafers can let each use different process conditions. Placing the peripheral circuits beneath the cell array can also reduce planar area. COP and wafer-to-wafer hybrid copper bonding in Samsung&amp;rsquo;s ISSCC research are one such implementation.&lt;/p&gt;
&lt;p&gt;However, 4F² defines cell area. Its definition does not require bonding two wafers. CXMT&amp;rsquo;s 2023 4F² array paper and Samsung&amp;rsquo;s COP-and-bonding test chip cannot be treated as using the same process flow.&lt;/p&gt;
&lt;p&gt;For a bonded implementation, the cell wafer, peripheral-circuit wafer, and bonding process must be assessed together.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Additional item to check&lt;/th&gt;
 &lt;th&gt;Yield and cost question&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Surface planarity and contamination or particles&lt;/td&gt;
 &lt;td&gt;Interface defects and connection failure&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Copper pads and dielectric bonding&lt;/td&gt;
 &lt;td&gt;Contact resistance, opens, and long-term stability&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Alignment of the two wafers and warpage&lt;/td&gt;
 &lt;td&gt;Connection errors and within-wafer yield distribution&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Annealing and material combinations&lt;/td&gt;
 &lt;td&gt;Bond stability and preservation of cell and circuit properties&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Good-die locations and inspection on both wafers&lt;/td&gt;
 &lt;td&gt;Which defects at each stage become finished-product losses&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Throughput, equipment, and inspection costs&lt;/td&gt;
 &lt;td&gt;Whether the area gain exceeds the added process cost&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Assume for illustration that cell-wafer yield is 95%, peripheral-wafer yield is 95%, and bonding yield is 98%, all independent. The combined yield is 0.95 × 0.95 × 0.98 = 88.445%. This is not an actual yield estimate. Defect location and correlation, inspection, and rework can change the result. The example shows why a high yield at one stage does not establish final yield.&lt;/p&gt;
&lt;p&gt;HBM die stacking and wafer bonding between cells and peripheral circuits are not the same market. Their connection requirements, inspection, thermal management, and process integration differ. The label “bonding” alone does not confirm orders for a specific equipment supplier or CXMT&amp;rsquo;s access to such equipment.&lt;/p&gt;
&lt;h2 id="the-cost-denominator-is-saleable-good-bits-not-designed-bits"&gt;The cost denominator is saleable good bits, not designed bits
&lt;/h2&gt;&lt;p&gt;&lt;strong&gt;Total cost per good bit = total cost including wafers, process steps, bonding, inspection, and depreciation ÷ saleable good bits.&lt;/strong&gt; Looking only at a design-density improvement misses yield and customer-qualification rates in the denominator.&lt;/p&gt;
&lt;p&gt;Assume total process cost increases 20% and pre-yield bit output increases 30%.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Relative yield versus the baseline&lt;/th&gt;
 &lt;th&gt;Cost ratio&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Result&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;95%&lt;/td&gt;
 &lt;td&gt;1.20 ÷ (1.30 × 0.95)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;Cost per good bit down about 2.8%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;85%&lt;/td&gt;
 &lt;td&gt;1.20 ÷ (1.30 × 0.85)&lt;/td&gt;
 &lt;td style="text-align: right"&gt;Cost per good bit up about 8.6%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Relative yield of 95% is not absolute yield of 95%. Even the same density benefit can produce lower or higher costs depending on process integration and yield. This table does not estimate actual Samsung or CXMT costs.&lt;/p&gt;
&lt;p&gt;Samsung&amp;rsquo;s ISSCC abstract reports more than a 20% increase in total die per wafer under the same design rules. This finished-chip metric uses a different comparison basis from the theoretical 50% cell-array density increase. Supply competitiveness requires area, yield, performance, and total cost for the same product.&lt;/p&gt;
&lt;h2 id="compare-samsung-and-sk-hynix-at-the-same-announcement-stage"&gt;Compare Samsung and SK hynix at the same announcement stage
&lt;/h2&gt;&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Company and source&lt;/th&gt;
 &lt;th&gt;What was disclosed&lt;/th&gt;
 &lt;th&gt;What it does not establish&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;CXMT, IEDM 2023&lt;/td&gt;
 &lt;td&gt;Junctionless GAA vertical device and 4F² array research&lt;/td&gt;
 &lt;td&gt;Production yield or customer qualification for 2026 DDR5 modules&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;CXMT, October 2026 remarks&lt;/td&gt;
 &lt;td&gt;Year-end server DDR5 RDIMM unveiling target combining 4F², vertical devices, and bonding&lt;/td&gt;
 &lt;td&gt;Product specifications, economics, or repeat shipments&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Samsung co-research, ISSCC 2026&lt;/td&gt;
 &lt;td&gt;Read/write operation of a 10nm-class 16Gb VCT/COP test chip with wafer-to-wafer copper bonding&lt;/td&gt;
 &lt;td&gt;Mass-production yield or a commercial shipment date&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;SK hynix, VLSI 2025&lt;/td&gt;
 &lt;td&gt;Roadmap for 4F² Vertical Gate and 3D DRAM&lt;/td&gt;
 &lt;td&gt;Confirmed commercial production dates for each generation&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Kioxia, OCTRAM research&lt;/td&gt;
 &lt;td&gt;4F² research using an InGaZnO oxide channel&lt;/td&gt;
 &lt;td&gt;That CXMT uses the same materials or implementation&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Sources: the research papers and the &lt;a class="link" href="https://submissions.mirasmart.com/ISSCC2026/PDF/ISSCC2026AdvanceProgram.pdf" target="_blank" rel="noopener"
 &gt;official ISSCC program, session 15.10&lt;/a&gt;, &lt;a class="link" href="https://news.skhynix.com/en/?p=1286" target="_blank" rel="noopener"
 &gt;SK hynix official VLSI announcement&lt;/a&gt;, and &lt;a class="link" href="https://www.yicai.com/news/103388707.html" target="_blank" rel="noopener"
 &gt;Yicai coverage of CXMT&amp;rsquo;s conference remarks&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;Samsung&amp;rsquo;s 2028 production date is a target reported by some media. It should be distinguished from a confirmed company mass-production filing. Calculating a two-year lead by comparing that target with CXMT&amp;rsquo;s year-end unveiling target mixes stages. &lt;a class="link" href="https://www.trendforce.com/news/2026/04/24/news-samsung-reportedly-produces-sub-10nm-10a-dram-working-die-using-4f-square-and-vct-targets-2028-production/" target="_blank" rel="noopener"
 &gt;Report on the 2028 target&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;The lack of public proof is not proof of failure. The accurate status is that CXMT&amp;rsquo;s product unveiling, yield, and customer qualification remain to be seen. Samsung and SK hynix cannot turn research announcements into proof of mass-production economics either.&lt;/p&gt;
&lt;h2 id="mobile-supply-server-dram-and-hbm-follow-different-qualification-paths"&gt;Mobile supply, server DRAM, and HBM follow different qualification paths
&lt;/h2&gt;&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Product and customer&lt;/th&gt;
 &lt;th&gt;Stage in public evidence&lt;/th&gt;
 &lt;th&gt;Interpretation&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;G5 24Gb LPDDR5X&lt;/td&gt;
 &lt;td&gt;Company announcement of product mass production&lt;/td&gt;
 &lt;td&gt;Current process progress. Scale sales and good-bit cost remain separate&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;LPDDR6 and Xiaomi foldable phone&lt;/td&gt;
 &lt;td&gt;Reuters reports the supply arrangement confirmed by both companies&lt;/td&gt;
 &lt;td&gt;Progress with a mobile customer. Customer-level volume remains unknown&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Tencent server DRAM&lt;/td&gt;
 &lt;td&gt;Long-term supply agreement reported by anonymous sources&lt;/td&gt;
 &lt;td&gt;Term, product scope, and actual delivery unverified. Not 4F² qualification evidence&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;4F² DDR5 RDIMM&lt;/td&gt;
 &lt;td&gt;Year-end unveiling target&lt;/td&gt;
 &lt;td&gt;Server-platform approval and repeat deliveries still needed&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;HBM3E and Chinese accelerators&lt;/td&gt;
 &lt;td&gt;Reports of test production and customer evaluation&lt;/td&gt;
 &lt;td&gt;Not direct evidence of customer, stacking yield, or commercial shipments&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Sources: &lt;a class="link" href="https://ca.marketscreener.com/news/china-s-cxmt-to-supply-memory-chip-for-xiaomi-s-upcoming-folding-phone-ce7858dcd981f12d" target="_blank" rel="noopener"
 &gt;Reuters report on LPDDR6 supply to Xiaomi, carried by MarketScreener&lt;/a&gt;, and &lt;a class="link" href="https://www.investing.com/news/stock-market-news/exclusivechinas-cxmt-wins-3-billion-memory-supply-deal-with-tencent-sources-say-4764321" target="_blank" rel="noopener"
 &gt;Reuters report on the Tencent deal, carried by Investing.com&lt;/a&gt;. Direct access to the Reuters pages was limited, so these Reuters syndication copies were checked.&lt;/p&gt;
&lt;p&gt;Do not combine other customer names, strategic investors, and listings of commercial modules with mass-production qualification. Public sources do not establish product- and customer-level revenue or CXMT&amp;rsquo;s Top 1/Top 5 customer shares.&lt;/p&gt;
&lt;p&gt;HBM requires more than smaller cells. It involves stacking die, TSV connections, packaging yield, thermal and warpage control, base-die integration, and accelerator-platform qualification. Progress in 4F² DDR5 does not imply immediate HBM4 substitution. Conversely, CXMT does not need to replace HBM4 to increase competition in conventional DRAM.&lt;/p&gt;
&lt;p&gt;The estimate that the HBM gap is “three to five years” depends on which products, performance levels, and yields are compared. Reports that a product is under evaluation do not automatically shorten the gap to two or three years. Actual deployment in a customer product and repeat shipments are more direct signals.&lt;/p&gt;
&lt;h2 id="judge-the-year-end-product-by-server-reliability-and-repeat-shipments-not-by-a-die-photo"&gt;Judge the year-end product by server reliability and repeat shipments, not by a die photo
&lt;/h2&gt;&lt;p&gt;&lt;strong&gt;The first question about a year-end reveal is structure; the second is performance and reliability; the third is economics.&lt;/strong&gt; Verifying one does not pass the others.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;What to check in the reveal&lt;/th&gt;
 &lt;th&gt;What it can establish&lt;/th&gt;
 &lt;th&gt;Evidence needed next&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Cell cross-section, gate, channel, and cell area&lt;/td&gt;
 &lt;td&gt;Actual 4F² and vertical-device implementation&lt;/td&gt;
 &lt;td&gt;Full-chip density and electrical characteristics&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Bonding interface and peripheral-circuit placement&lt;/td&gt;
 &lt;td&gt;A specific integration architecture&lt;/td&gt;
 &lt;td&gt;Bond yield, throughput, and cost&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Die capacity and Gb/mm²&lt;/td&gt;
 &lt;td&gt;Finished-chip density&lt;/td&gt;
 &lt;td&gt;Total cost per good bit&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Speed, voltage, and power&lt;/td&gt;
 &lt;td&gt;Performance under the stated conditions&lt;/td&gt;
 &lt;td&gt;Distribution across temperatures, loads, and wafers&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Data retention, refresh, and error distribution&lt;/td&gt;
 &lt;td&gt;Reliability under the stated conditions&lt;/td&gt;
 &lt;td&gt;Specifications that match density and temperature, and platform validation&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Intel, AMD, or other platform verification and customer approval&lt;/td&gt;
 &lt;td&gt;Qualification for that product and platform&lt;/td&gt;
 &lt;td&gt;Actual orders and repeat deliveries&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Good-die shipments and costs over multiple quarters&lt;/td&gt;
 &lt;td&gt;Commercial supply and economics&lt;/td&gt;
 &lt;td&gt;Persistence of price and share effects&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Product density, temperature, and operating conditions differ, so one universal threshold should not be used as a pass line. Reverse engineering helps establish structure and area but does not directly establish manufacturing yield or long-term reliability.&lt;/p&gt;
&lt;p&gt;The type of disclosure also limits the conclusion.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Disclosure type&lt;/th&gt;
 &lt;th&gt;Conclusion it can support&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Structure drawings and a date&lt;/td&gt;
 &lt;td&gt;Development direction and intent; not a confirmed production schedule&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Specifications and an operating sample&lt;/td&gt;
 &lt;td&gt;Technical progress; customer qualification and economics remain unknown&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Platform and customer approval&lt;/td&gt;
 &lt;td&gt;Stronger evidence that a specific product can be supplied&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Repeat shipments with yield and cost&lt;/td&gt;
 &lt;td&gt;Evidence to incorporate into Samsung&amp;rsquo;s and SK hynix&amp;rsquo;s supply and margin models&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Do not automatically move the date to 2028 because an event reveals no specifications, or call it mass production because specifications were disclosed. A schedule estimate requires progress in manufacturing and customer approval.&lt;/p&gt;
&lt;h2 id="put-good-bit-shipments-into-the-capacity-model"&gt;Put good-bit shipments into the capacity model
&lt;/h2&gt;&lt;p&gt;Monthly capacity of about 300,000 wafers and future paths toward 500,000–600,000 wafers appear in reports and estimates. These mix current capacity, targets, and forecasts. &lt;a class="link" href="https://newsletter.semianalysis.com/p/chinas-cxmt-is-set-to-challenge-dram" target="_blank" rel="noopener"
 &gt;SemiAnalysis&amp;rsquo; CXMT analysis&lt;/a&gt; is an analyst model, not a physical fab audit confirming current good-die output.&lt;/p&gt;
&lt;p&gt;Supply should be separated into &lt;strong&gt;installed capacity × actual utilization × designed bits per wafer × front- and back-end yield × customer-qualified share&lt;/strong&gt;. Doubling monthly wafer starts is not the same as doubling saleable server DRAM bits. Bit shipments can also rise through node migration and better yield.&lt;/p&gt;
&lt;p&gt;TrendForce&amp;rsquo;s reported second-quarter CXMT DRAM revenue share is 9.5%. It is not bit share or wafer share. Before averaging it with estimates from other firms, align the product, period, and revenue basis. &lt;a class="link" href="https://www.trendforce.com/news/2026/10/12/news-cxmt-accelerates-4f%C2%B2-dram-push-eyes-year-end-server-ddr5-rdimm-launch/" target="_blank" rel="noopener"
 &gt;TrendForce report on CXMT&amp;rsquo;s technology and market position&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;For illustration, set initial non-HBM DRAM supply and demand at 100 each, CXMT&amp;rsquo;s hypothetical bit share at 10%, and other suppliers&amp;rsquo; supply growth at 15%. The 10% assumption is not converted from revenue share.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Assumed CXMT bit-supply growth&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Total supply growth&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Demand +15%&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Demand +20%&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Demand +25%&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;+30%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+16.5%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+1.3%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;−2.9%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;−6.8%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;+50%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+18.5%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+3.0%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;−1.3%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;−5.2%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;+70%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+20.5%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+4.8%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+0.4%&lt;/td&gt;
 &lt;td style="text-align: right"&gt;−3.6%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;The supply-demand gap is supply ÷ demand − 1. A positive value means supply is in surplus; a negative value means a shortage. Supply in the middle row is 90 × 1.15 + 10 × 1.50 = 118.5. If demand grows 20%, the market is about 1.3% short; at 15% growth, supply exceeds demand by about 3.0%. The same Chinese capacity increase has different effects depending on demand and the starting shortage.&lt;/p&gt;
&lt;p&gt;The share of production converted to a new process also matters. If a supplier with a hypothetical 10% global bit share applies the process to 20% of output and raises good-die production there by 20%, the incremental effect on global supply is 0.4%. Do not add the process effect again if it is already included in total bit growth.&lt;/p&gt;
&lt;p&gt;These calculations are not forecasts or price-change estimates. They assume an initially balanced market and omit differences among products. Counting Chinese import substitution both as a decline in global demand and an increase in global supply would double-count the same event.&lt;/p&gt;
&lt;h2 id="policy-applies-to-the-whole-process-and-customer-markets-not-only-lithography"&gt;Policy applies to the whole process and customer markets, not only lithography
&lt;/h2&gt;&lt;p&gt;Even if structural changes ease part of the lithography burden, constraints remain in etching, deposition, bonding, inspection, and metrology. The process challenge is being redistributed, not replaced entirely by a shift from EUV to bonding. Claims that “bonding, ALD, and metrology are all uncontrolled” should not be accepted without review of the specific equipment, technology, and transaction.&lt;/p&gt;
&lt;p&gt;BIS&amp;rsquo;s December 2024 action covers manufacturing equipment, software, and HBM. Raising density without EUV does not establish that the controls have expired. &lt;a class="link" href="https://media.bis.gov/press-release/commerce-strengthens-export-controls-restrict-chinas-capability-produce-advanced-semiconductors-military" target="_blank" rel="noopener"
 &gt;BIS official announcement&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;The policy lists also refer to different regimes. The Entity List&amp;rsquo;s export-license framework, the DoD 1260H list, and procurement limits under NDAA 5949 and FAR should not be collapsed into a blanket ban on private-sector purchases. The February 17 proposed FAR rule discusses the December 23, 2027 statutory effective date and federal procurement. It is not a blanket restriction on all U.S. and allied private-cloud purchases. &lt;a class="link" href="https://www.federalregister.gov/documents/2026/02/17/2026-03065/federal-acquisition-regulation-prohibition-on-certain-semiconductor-products-and-services" target="_blank" rel="noopener"
 &gt;Federal Register proposed rule&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;Reports that additions to a policy list are under review or on hold are different from a published designation and its implementation. This article does not make a legal determination of the current status of individual designations. Technical capability, equipment access, and access to customer markets are separate risk factors.&lt;/p&gt;
&lt;h2 id="the-nearer-risk-for-samsung-and-sk-hynix-is-substitution-in-existing-products"&gt;The nearer risk for Samsung and SK hynix is substitution in existing products
&lt;/h2&gt;&lt;p&gt;CXMT does not need to catch up with the leaders in HBM4 or the highest-end server DRAM before competition begins. If Chinese customers replace some imported products with qualified domestic parts, incumbent suppliers can lose sales channels and face different contract terms.&lt;/p&gt;
&lt;p&gt;Policy support and Chinese customer testing and purchasing opportunities can help CXMT catch up. But there is not enough public evidence to conclude that technical moats have disappeared or that the state will absorb all losses so CXMT never cuts production. This asymmetric supply response should be considered as a downside scenario.&lt;/p&gt;
&lt;p&gt;The funding is real. An August 29 filing reports total proceeds of about RMB 66.607 billion after the greenshoe exercise and net proceeds of about RMB 66.310 billion. These are distinct from the initial July plan of RMB 57.9 billion. Funding does not guarantee equipment installation, yield, or qualification schedules. &lt;a class="link" href="https://static.cninfo.com.cn/finalpage/2026-08-29/1225529556.PDF" target="_blank" rel="noopener"
 &gt;IPO greenshoe implementation results&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;CXMT&amp;rsquo;s first-half revenue of RMB 150.310 billion and net profit attributable to shareholders of RMB 77.605 billion are figures reported by the financial press. Direct review of the full half-year filing and notes remains incomplete, so these are not inputs to a confirmed cost model. The figures do not establish Samsung-equivalent good-die costs. Price, volume, product mix, and subsidies would need to be separated. &lt;a class="link" href="https://www.nbd.com.cn/articles/2026-08-28/4564392.html" target="_blank" rel="noopener"
 &gt;CXMT financial report&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;Samsung&amp;rsquo;s preliminary third-quarter operating profit of KRW 107.4 trillion shows the strength of the current boom. It does not prove how long earnings will last. Check shipments of existing products and premium products, along with minimum-purchase obligations, price floors, and renegotiation terms in future contracts. &lt;a class="link" href="https://news.samsungsemiconductor.com/global/samsung-electronics-announces-earnings-guidance-for-third-quarter-2026/" target="_blank" rel="noopener"
 &gt;Samsung preliminary results&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;Customer approval and repeat shipments of HBM4E, server DDR5, SOCAMM2, and enterprise SSDs support Samsung&amp;rsquo;s defenses. Internal memory and foundry sourcing does not automatically guarantee lower cost, and SK hynix&amp;rsquo;s external sourcing is not inherently inferior. Finished-product performance, cost, and delivery are the standards.&lt;/p&gt;
&lt;p&gt;AI usage can keep growing while installations are delayed or customers reduce inventory, slowing order growth. If good-bit supply from China and incumbent suppliers rises at the same time, prices may normalize faster. Track technology supply paths and actual purchase demand together.&lt;/p&gt;
&lt;h2 id="signals-that-could-change-the-timeline-are-qualification-yield-and-shipmentsnot-announcements"&gt;Signals that could change the timeline are qualification, yield, and shipments—not announcements
&lt;/h2&gt;&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Event&lt;/th&gt;
 &lt;th&gt;Evidence to check first&lt;/th&gt;
 &lt;th&gt;Scope of the conclusion&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Year-end 4F² product reveal&lt;/td&gt;
 &lt;td&gt;Cross-section, capacity, density, speed, power, and refresh&lt;/td&gt;
 &lt;td&gt;Whether the work has moved from research direction to an operating product&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;4F² customer approval and repeat shipments&lt;/td&gt;
 &lt;td&gt;Platform, customer, and product-level approval and volumes&lt;/td&gt;
 &lt;td&gt;Commercial supply capability and timing&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;G5 scale sales&lt;/td&gt;
 &lt;td&gt;Good-die shipments by product, yield, and repeat orders&lt;/td&gt;
 &lt;td&gt;Near-term import substitution in conventional DRAM&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;HBM3E deployment at customers&lt;/td&gt;
 &lt;td&gt;Customer products, stacking yield, and repeat deliveries&lt;/td&gt;
 &lt;td&gt;HBM progress separate from conventional DRAM&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Changes in process/equipment or market access&lt;/td&gt;
 &lt;td&gt;Official rules, designation, license scope, and effective date&lt;/td&gt;
 &lt;td&gt;Recalculate capacity, technology, and sales channels separately&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Samsung and SK hynix 2027 contracts and product transition&lt;/td&gt;
 &lt;td&gt;Minimum purchase and price terms, cancellation rights, premium-product approvals and shipments&lt;/td&gt;
 &lt;td&gt;Whether high earnings can be defended&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Public evidence does not establish CXMT&amp;rsquo;s mass-production year, the exact number of years in its HBM gap, or its 2030 bit share. Price-drop probabilities and stock-price impacts from models whose source code is unavailable are not adopted. Running a model 3,000 times does not make its input assumptions realistic.&lt;/p&gt;
&lt;p&gt;The assessment of technology news should remain open in both directions. If CXMT&amp;rsquo;s cost, yield, and qualification improve faster than expected, raise the competitive-risk assessment. If it announces a structure without demonstrating good-die costs and customer qualification, push out the schedule and reduce the estimated impact.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;The path CXMT is pursuing is a path to the required density without EUV. The destination is not the smallest cell, but good die that customers buy repeatedly.&lt;/strong&gt; Reassess the timeline for pressure on Samsung&amp;rsquo;s and SK hynix&amp;rsquo;s high earnings at that point.&lt;/p&gt;
&lt;h2 id="evidence-scope"&gt;Evidence scope
&lt;/h2&gt;&lt;p&gt;Company announcements establish what the company claims; conference papers establish results for the reported test structures. No single source validates reverse engineering, process yield, long-term server reliability, and customer shipments. This article does not merge the stages of G5, LPDDR6, 4F², and HBM. Supply, yield, and cost calculations are illustrative assumptions, not forecasts or investment probabilities for these companies.&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Disclaimer: For research and information purposes only. Not investment advice. Names cited are for analytical illustration; readers should perform their own due diligence and consult licensed advisors before any investment decision.&lt;/em&gt;&lt;/p&gt;</description></item><item><title>Samsung Electronics' October 13 Selloff: Memory Competition, a Demand Gap, and What a Rebound Must Confirm</title><link>https://koreainvestinsights.com/post/samsung-oct13-selloff-cxmt-rebound-2026/</link><pubDate>Wed, 14 Oct 2026 05:00:00 +0900</pubDate><guid>https://koreainvestinsights.com/post/samsung-oct13-selloff-cxmt-rebound-2026/</guid><description>&lt;p&gt;Samsung Electronics closed at KRW 232,000 on October 13, down 7.39%. SK hynix fell 10.09%, and the KOSPI broke below 6,000 intraday. Just five days earlier, Samsung had reported preliminary quarterly operating profit of KRW 107.4 trillion. Strong earnings and a sharp share-price decline arrived at the same time.&lt;/p&gt;
&lt;p&gt;The most convincing explanation is that doubts about how long memory earnings can last coincided with a gap in domestic buying demand. CXMT&amp;rsquo;s new-process announcement and concerns about funding for AI investment gave investors reasons to sell. The end of Samsung&amp;rsquo;s share buyback reduced a source of demand that had absorbed some of that selling. Futures and program trading may have transmitted or amplified the shock.&lt;/p&gt;
&lt;p&gt;This distinction matters when assessing a rebound. A temporary recovery is possible if the order imbalance eases. A sustained trend reversal would also require foreign cash-equity selling to slow and memory earnings expectations to stabilize.&lt;/p&gt;
&lt;h2 id="the-selloff-centered-on-memory-a-rebound-needs-cash-equity-flows"&gt;The selloff centered on memory; a rebound needs cash-equity flows
&lt;/h2&gt;&lt;ul&gt;
&lt;li&gt;Memory stocks also fell sharply in Japan and Taiwan. Samsung-specific bad news or a Korea-only exit from assets does not fully explain the move.&lt;/li&gt;
&lt;li&gt;Korean government bonds strengthened while USD/KRW rose KRW 10. The evidence is insufficient to call the equity selling a liquidation of all Korean assets. It also does not prove that country risk is absent.&lt;/li&gt;
&lt;li&gt;CXMT said it plans to unveil a server DDR5 product by year-end, but it did not disclose yield, shipment volume, or customer qualification. A technology announcement cannot be translated directly into a Samsung earnings decline.&lt;/li&gt;
&lt;li&gt;A technical rebound toward KRW 240,000–245,000 and a trend reversal accompanied by foreign investor inflows are different scenarios. A price bounce alone would not confirm a bottom.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;The reference point is the October 13, 2026 close in Korea, Japan, and Taiwan. U.S. figures come from reports published during the October 13 session. This is a review prepared before the Korean market opened on October 14, not a report on the realized performance of a later forecast.&lt;/p&gt;
&lt;h2 id="foreign-selling-concentrated-in-samsung-electronics-and-sk-hynix"&gt;Foreign selling concentrated in Samsung Electronics and SK hynix
&lt;/h2&gt;&lt;p&gt;The table below shows both the price declines and the scale of selling. Market figures are from reports citing the Korea Exchange, Koscom, and the Korea Financial Investment Association. They are distinct from a calculation based on raw exchange transaction records downloaded directly.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;October 13 indicator&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Observation&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Samsung Electronics close&lt;/td&gt;
 &lt;td style="text-align: right"&gt;KRW 232,000, -7.39%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Samsung Electronics intraday low&lt;/td&gt;
 &lt;td style="text-align: right"&gt;KRW 228,000&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;SK hynix close&lt;/td&gt;
 &lt;td style="text-align: right"&gt;KRW 1,443,000, -10.09%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;KOSPI close / intraday low&lt;/td&gt;
 &lt;td style="text-align: right"&gt;6,035.35 / 5,911.11&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;KOSPI decline for the day / five sessions&lt;/td&gt;
 &lt;td style="text-align: right"&gt;-5.46% / -13.83%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Foreign net selling in KOSPI stocks&lt;/td&gt;
 &lt;td style="text-align: right"&gt;KRW 4.5259 trillion&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Individual / institutional net buying in KOSPI stocks&lt;/td&gt;
 &lt;td style="text-align: right"&gt;KRW 2.2271 trillion / KRW 1.6902 trillion&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;VKOSPI close&lt;/td&gt;
 &lt;td style="text-align: right"&gt;48.53&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Source: &lt;a class="link" href="https://www.seoul.co.kr/news/economy/2026/10/13/20261013500289" target="_blank" rel="noopener"
 &gt;Seoul Shinmun market close report&lt;/a&gt;. Samsung&amp;rsquo;s intraday low comes from the cited selloff analysis. Reports disagree on the arithmetic for VKOSPI&amp;rsquo;s prior-day value, so this article does not use that figure.&lt;/p&gt;
&lt;p&gt;Foreign investors sold KRW 2.3613 trillion of SK hynix and KRW 1.3595 trillion of Samsung Electronics. Adding the displayed figures gives KRW 3.7208 trillion, about 82.2% of KOSPI net selling. The same report put total foreign net selling in KOSPI stocks at KRW 4.5330 trillion. This differs from the KRW 4.5259 trillion in the table, so the concentration is best read as approximately 82%. &lt;a class="link" href="https://www.fnnews.com/news/202610131744129819" target="_blank" rel="noopener"
 &gt;Fnnews and News1 flow report&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;Over the five sessions from October 6 through October 13, individual investors bought KRW 4.1397 trillion of Samsung Electronics and KRW 3.3253 trillion of SK hynix. Foreign investors sold KRW 2.8601 trillion and KRW 7.017 trillion, respectively. Individuals also bought KRW 848.9 billion of KODEX Leverage. Individual buying is not proof of a bottom. Buying leveraged products indicates the potential for larger losses if prices fall further, but it does not prove that forced selling has occurred. &lt;a class="link" href="https://www.fnnews.com/news/202610131820301700" target="_blank" rel="noopener"
 &gt;Five-session investor flow data&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;Samsung&amp;rsquo;s 46.30% foreign ownership ratio is as of October 12. Presenting it as the October 13 closing ratio would move the date forward by one day. The lower ownership ratio allows two interpretations: less foreign-held stock remains available for sale, or foreign investors are structurally reducing their positions. Before cash-equity buying turns positive, the first interpretation alone is not enough to establish confidence in a rebound.&lt;/p&gt;
&lt;h2 id="the-end-of-the-buyback-reduced-demand-independently-of-earnings"&gt;The end of the buyback reduced demand independently of earnings
&lt;/h2&gt;&lt;p&gt;Samsung&amp;rsquo;s share buyback ended on October 6. That date is distinct from the October 12 filing of the results report. Flow reports cited the end of the buyback as a factor that weakened a buffer against selling. The end of one buyback program does not mean that every future share repurchase has stopped.&lt;/p&gt;
&lt;p&gt;A reproduced filing lists 56,856,305 shares acquired across 29 acquisition dates from August 24 through October 6, for approximately KRW 15 trillion. Dividing roughly KRW 15 trillion by 29 days gives an average of about KRW 517.2 billion per acquisition day. The detailed figures have not been finally reconciled against the original DART filing, so treat them as indicative. &lt;a class="link" href="https://www.bosoop.com/disclosures/20261012000421/" target="_blank" rel="noopener"
 &gt;Reproduction of the share acquisition results report&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;This average cannot be subtracted from one day&amp;rsquo;s foreign net selling to estimate an expected price decline. The timing, price, and responsiveness of the actual orders differ. It is still reasonable to say that when a large, recurring source of demand diminishes, the same amount of selling can move prices more easily. Daishin Securities&amp;rsquo; analysis of the buyback buffer points to this channel as well. &lt;a class="link" href="https://hankyung.com/article/2026101396276" target="_blank" rel="noopener"
 &gt;Hankyung flow analysis&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;Samsung&amp;rsquo;s official preliminary results were revenue of KRW 195 trillion and operating profit of KRW 107.4 trillion. Operating profit rose 20.01% from the previous quarter. A change in expectations about how long high earnings can last fits the share-price action better than the claim that reported earnings have already deteriorated. That does not mean a future earnings decline has been confirmed. &lt;a class="link" href="https://news.samsung.com/kr/%ec%82%bc%ec%84%b1%ec%a0%84%ec%9e%90-2026%eb%85%84-3%eb%b6%84%ea%b8%b0-%ec%9e%a0%ec%a0%95%ec%8b%a4%ec%a0%81-%eb%b0%9c%ed%91%9c" target="_blank" rel="noopener"
 &gt;Samsung Newsroom, October 8&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;There is a strong counterargument. A report citing Samsung Securities said September memory export prices fell 10.7% month over month, while export weight rose 27.2% and export value increased 13.5%. A decline in prices is difficult to equate with a collapse in end demand. These statistics may also reflect changes in product mix, so they cannot be applied directly to Samsung&amp;rsquo;s selling prices or margins.&lt;/p&gt;
&lt;h2 id="bond-strength-and-won-weakness-limit-the-case-for-a-broad-exit"&gt;Bond strength and won weakness limit the case for a broad exit
&lt;/h2&gt;&lt;p&gt;Judging country risk from equities alone can miss signals from other markets. Bonds and foreign exchange moved as follows on the same day.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Indicator&lt;/th&gt;
 &lt;th style="text-align: right"&gt;October 13&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Korean government bond, 3-year&lt;/td&gt;
 &lt;td style="text-align: right"&gt;3.857%, -6.4 bp&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Korean government bond, 10-year&lt;/td&gt;
 &lt;td style="text-align: right"&gt;4.229%, -7.7 bp&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;10-year minus 3-year government bond yield spread&lt;/td&gt;
 &lt;td style="text-align: right"&gt;37.2 bp&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Foreign net buying in 3-year / 10-year Treasury futures&lt;/td&gt;
 &lt;td style="text-align: right"&gt;+14,312 / +2,996 contracts&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Consecutive sessions of foreign net buying in both futures&lt;/td&gt;
 &lt;td style="text-align: right"&gt;6 sessions&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;USD/KRW daytime close&lt;/td&gt;
 &lt;td style="text-align: right"&gt;KRW 1,345.1, +KRW 10.0&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Sources: &lt;a class="link" href="https://www.etoday.co.kr/news/view/2635353" target="_blank" rel="noopener"
 &gt;EToday bond-market close&lt;/a&gt;, &lt;a class="link" href="https://www.mt.co.kr/economy/2026/10/13/2026101315493144500" target="_blank" rel="noopener"
 &gt;MoneyToday foreign-exchange report&lt;/a&gt;. One basis point (bp) is 0.01 percentage point.&lt;/p&gt;
&lt;p&gt;Government bond yields fell. Therefore, a same-day spike in domestic government bond yields is not a plausible explanation for the stock decline. Foreign buying of Treasury futures also weakens the claim that investors were reducing Korean assets across the board in the same direction.&lt;/p&gt;
&lt;p&gt;However, there is no evidence that the sellers of equities and buyers of Treasury futures were the same investors. Contract counts cannot be netted against the value of equity selling. Safe-haven demand, expectations of slower growth, and separate bond-position adjustments may all have played a role. Strong government bonds alone do not prove that country risk is absent.&lt;/p&gt;
&lt;p&gt;Dividing KRW 4.5259 trillion of foreign net equity selling by KRW 1,345.1 per dollar gives about USD 3.365 billion. The exchange-rate move was approximately 0.75%, calculated as KRW 10 divided by KRW 1,335.1. These are different kinds of figures. They do not mean all proceeds from stock sales were converted into dollars that day. Dollar supply from exporters, currency hedges, and settlement lags can change the exchange-rate response.&lt;/p&gt;
&lt;p&gt;One interpretation is that won strength encouraged profit-taking; another is that exporters&amp;rsquo; dollar-denominated earnings matter more. The fact that foreign investors sold on October 12, when the won was strong, makes it difficult to explain the selling through exchange rates alone. If the won, government bonds, and credit markets all weaken together, the current interpretation of an equity-centered reduction in risk should be revisited.&lt;/p&gt;
&lt;h2 id="us-rates-and-ai-financing-are-background-pressures"&gt;U.S. rates and AI financing are background pressures
&lt;/h2&gt;&lt;p&gt;Reports during the U.S. session on October 13 put the 10-year Treasury yield near 5.19%, Brent crude near USD 104.63, and WTI near USD 91.85. The USD 106.65 Brent figure in the Korean currency report was the U.S. close on October 12. Prices observed at different times should not be combined as if they were simultaneous. &lt;a class="link" href="https://kitco.com/news/article/2026-10-13/gold-firms-dollar-yields-and-oil-retreat-cpi-kitco-am-report" target="_blank" rel="noopener"
 &gt;Kitco, October 13 intraday report&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;High interest rates can increase the financing cost of AI infrastructure investment and reduce the present value of distant future earnings. But the absolute level of rates is different from a new rate shock on the day. These materials do not support attributing a specific share of Korea&amp;rsquo;s selloff to U.S. rates.&lt;/p&gt;
&lt;p&gt;Concerns about financing AI investment were mentioned in reports on Broadcom management&amp;rsquo;s comments. The original statement was not rechecked, so this article does not quote it directly. The questions to track in company results are whether investment is funded with cash flow or more external borrowing, and whether customer revenue is keeping pace with higher investment.&lt;/p&gt;
&lt;h2 id="the-futures-selloff-is-visible-but-the-scale-of-forced-selling-is-unknown"&gt;The futures selloff is visible, but the scale of forced selling is unknown
&lt;/h2&gt;&lt;p&gt;At 09:47:58 on October 13, the KOSPI 200 futures contract was down 5.08% from the previous close, triggering a sell-sidecar. This temporarily suspends the effectiveness of program sell orders. It is different from a halt in trading across the entire stock market. &lt;a class="link" href="https://www.mt.co.kr/stock/2026/10/13/2026101314134112327" target="_blank" rel="noopener"
 &gt;MoneyToday report on the sidecar&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;Foreign net flows classify investors; program flows classify trading methods. The same trade can appear in both categories, so adding the two selling totals would double-count. To determine whether futures selling led to cash-basket selling, we would need arbitrage/non-arbitrage classifications, futures-cash price differences at the same time, and open interest.&lt;/p&gt;
&lt;p&gt;Closing data on a consistent basis were unavailable to confirm a late-session reduction in foreign futures selling. It cannot serve as confirmed evidence for a rebound. A VKOSPI reading of 48.53 is also a price for expected volatility; it does not show the direction of dealer positions.&lt;/p&gt;
&lt;p&gt;October 13 was not the usual expiry date for October monthly options. That does not mean options hedging had no effect. Strike-level positions and option-price data are unavailable, so the contribution from gamma hedging, leveraged liquidation, or margin calls remains unverified.&lt;/p&gt;
&lt;h2 id="memory-stocks-also-fell-more-sharply-in-japan-and-taiwan"&gt;Memory stocks also fell more sharply in Japan and Taiwan
&lt;/h2&gt;&lt;p&gt;To assess whether this was a Korea-specific shock, compare both the indexes and the relevant companies.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Market&lt;/th&gt;
 &lt;th style="text-align: right"&gt;Index&lt;/th&gt;
 &lt;th&gt;Related companies&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Korea&lt;/td&gt;
 &lt;td style="text-align: right"&gt;KOSPI -5.46%&lt;/td&gt;
 &lt;td&gt;Samsung Electronics -7.39%, SK hynix -10.09%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Japan&lt;/td&gt;
 &lt;td style="text-align: right"&gt;Nikkei -1.82%&lt;/td&gt;
 &lt;td&gt;Kioxia about -13%; semiconductor and AI names were weak&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Taiwan&lt;/td&gt;
 &lt;td style="text-align: right"&gt;TAIEX -1.95%&lt;/td&gt;
 &lt;td&gt;TSMC -1.97%, Nanya -7.40%, Winbond -7.18%&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Sources: &lt;a class="link" href="https://www.nikkei.com/article/DGXZQOFL134EU0T11C26A0000000/" target="_blank" rel="noopener"
 &gt;Nikkei market close report&lt;/a&gt;, &lt;a class="link" href="https://news.cnyes.com/news/id/6629365" target="_blank" rel="noopener"
 &gt;Cnyes Taiwan market close report&lt;/a&gt;. Kioxia&amp;rsquo;s decline is from a reported figure in &lt;a class="link" href="https://finance.biggo.com/news/eb8a05d9-76f2-4e04-9cf6-1d2e426b69ce" target="_blank" rel="noopener"
 &gt;BigGo&amp;rsquo;s Asia market roundup&lt;/a&gt;, not a direct exchange calculation.&lt;/p&gt;
&lt;p&gt;DRAM companies in Taiwan also fell more than TSMC. Kioxia in Japan is primarily a NAND company. Its direct competition with CXMT&amp;rsquo;s server DRAM differs, yet it also fell. This supports an interpretation of broad risk reduction in memory and AI-related equities. The sources do not prove that every company fell for one identical reason.&lt;/p&gt;
&lt;p&gt;Differences in index declines also reflect industry composition, stock weights, and investor mix. No contribution by index weight was calculated, so the larger Korean decline is not attributed entirely to the two memory companies or the end of Samsung&amp;rsquo;s buyback. The prior analysis&amp;rsquo; comparison of Taiwan foreign selling in U.S. dollars is omitted because access to the source text was blocked.&lt;/p&gt;
&lt;h2 id="cxmts-4f-announcement-does-not-prove-a-cost-advantage"&gt;CXMT&amp;rsquo;s 4F² announcement does not prove a cost advantage
&lt;/h2&gt;&lt;p&gt;CXMT is China&amp;rsquo;s ChangXin Memory Technologies. It is distinct from South Korea-listed CMTX. The reported announcement was a plan to unveil a server DDR5 RDIMM using a vertical-transistor 4F² cell by year-end. CXMT disclosed the product category and target date, but not speed, capacity, yield, shipment volume, or customer qualification. &lt;a class="link" href="https://www.sedaily.com/article/20101179" target="_blank" rel="noopener"
 &gt;Seoul Economic Daily report on CXMT&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;Samsung Electronics also presented research on 4F² DRAM using vertical cell transistors and wafer-to-wafer hybrid copper bonding in the official ISSCC 2026 program. A research presentation, product announcement, customer qualification, and economically viable volume production are different stages. We cannot establish a two-year technology gap by placing CXMT&amp;rsquo;s year-end product plan beside press reports about a Samsung production target. &lt;a class="link" href="https://submissions.mirasmart.com/ISSCC2026/PDF/ISSCC2026AdvanceProgram.pdf" target="_blank" rel="noopener"
 &gt;Official ISSCC program, paper 15.10&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;Samsung&amp;rsquo;s 2028 production date was a target in some reports. Follow-up reporting also said that there is no confirmed timeline. CXMT, too, has stated a target rather than demonstrated commercial volume. The gap between an announcement and stable mass production remains unverified. &lt;a class="link" href="https://www.tomshardware.com/pc-components/dram/cxmt-president-says-companys-first-4f2-dram-will-be-a-server-ddr5-rdimm-due-by-year-end-vertical-transistors-cut-each-memory-cells-footprint-by-a-third-on-paper-but-no-speeds-or-capacities-yet" target="_blank" rel="noopener"
 &gt;Tom&amp;rsquo;s Hardware technology review&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;In 4F², F denotes a minimum feature size. Assuming the same F, cell area is 33.3% smaller than 6F² and theoretical cell-array density is 1.5 times higher. This does not mean that the cost of the entire chip, including peripheral circuits and bonding, falls by 33.3%.&lt;/p&gt;
&lt;p&gt;The difference becomes clear with a numerical example. Assume manufacturing cost is 1.3 times the baseline, bit count is 1.5 times higher, and yield falls from 90% to 75%. The cost per good bit is then 1.3 ÷ 1.5 × (0.90 ÷ 0.75) = 1.04 times the baseline. Cost per good bit is 4% higher. The break-even yield under these assumptions is 78%. This is an illustration of why the announcement alone cannot establish the economics; it is not an estimate of CXMT&amp;rsquo;s actual cost or yield.&lt;/p&gt;
&lt;p&gt;The potential channels for Samsung should be separated by time horizon.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Channel&lt;/th&gt;
 &lt;th&gt;Current assessment&lt;/th&gt;
 &lt;th&gt;Evidence that could change the assessment&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Near-term share-price multiple&lt;/td&gt;
 &lt;td&gt;Competition concerns may lower expectations for the duration of earnings&lt;/td&gt;
 &lt;td&gt;Customer, supply, and pricing data that strengthen or weaken the narrative&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Direct 2026–27 earnings impact&lt;/td&gt;
 &lt;td&gt;Public information does not support estimating the amount or timing of a decline&lt;/td&gt;
 &lt;td&gt;Actual shipments, qualification, yield, and selling prices&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Chinese substitution for commodity DRAM imports in 2027–28&lt;/td&gt;
 &lt;td&gt;A conditional risk, dependent on production and customer adoption&lt;/td&gt;
 &lt;td&gt;Adoption by Chinese server makers and DDR5 contract prices&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;HBM impact&lt;/td&gt;
 &lt;td&gt;This DDR5 announcement alone does not demonstrate HBM competitiveness&lt;/td&gt;
 &lt;td&gt;Separate evidence on HBM qualification, shipments, and performance&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Stronger competition in commodity DRAM could weigh on Samsung. However, comparing the earnings sensitivity of Samsung and SK hynix requires current product-level revenue and profit mix. Kioxia&amp;rsquo;s decline also does not prove that CXMT is replacing NAND.&lt;/p&gt;
&lt;p&gt;The strongest counterargument is that Chinese producers have already become commercial competitors in commodity DRAM. Uncertainty around 4F² is not a reason to ignore supply growth from existing processes. Track the success of the new process separately from volume growth in existing products. Industry caution on cost and yield is also reflected in &lt;a class="link" href="https://www.fnnews.com/news/202610131602240594" target="_blank" rel="noopener"
 &gt;Fnnews&amp;rsquo; technology report&lt;/a&gt;.&lt;/p&gt;
&lt;h2 id="assess-technical-rebounds-and-trend-reversals-separately"&gt;Assess technical rebounds and trend reversals separately
&lt;/h2&gt;&lt;p&gt;Samsung&amp;rsquo;s close was (KRW 232,000 ÷ KRW 228,000 − 1) × 100 = 1.75% above its intraday low. The KOSPI also closed 2.10% above its low. This shows that some intraday selling was absorbed. It does not guarantee that buyers will hold the line the next day.&lt;/p&gt;
&lt;p&gt;In the author&amp;rsquo;s subjective assessment, the probability of a technical rebound over one to three sessions is 55–60%; the probability of a trend reversal within two weeks accompanied by foreign net buying is 25–30%. These are not outputs from a statistically validated model and are not a verified hit rate. The 55–60% estimate is also not the probability of a sustained rally.&lt;/p&gt;
&lt;p&gt;KRW 240,000–245,000 is approximately 3.45–5.60% above the October 13 close. It is an observation range in a technical-rebound scenario, not a price target. The annual P/E estimate that could not be verified and the RSI figure without an official calculation are excluded from this assessment.&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;Subsequent combination&lt;/th&gt;
 &lt;th&gt;Scenario&lt;/th&gt;
 &lt;th&gt;Interpretation&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;Cash-equity selling slows, volatility falls, and the low holds&lt;/td&gt;
 &lt;td&gt;Technical rebound&lt;/td&gt;
 &lt;td&gt;Watch KRW 240,000–245,000; this is separate from a trend reversal&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Price rebounds, then retests around KRW 228,000&lt;/td&gt;
 &lt;td&gt;Base W-shaped assumption&lt;/td&gt;
 &lt;td&gt;The closing price and cash-equity flows during the retest matter&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Close below KRW 228,000 while selling continues&lt;/td&gt;
 &lt;td&gt;Downside risk&lt;/td&gt;
 &lt;td&gt;KRW 220,000 is another level to monitor, not confirmed support&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;Foreign investors keep buying cash equities and sector concerns ease&lt;/td&gt;
 &lt;td&gt;Potential trend reversal&lt;/td&gt;
 &lt;td&gt;Flow and earnings expectations need to improve together, beyond a one-day gain&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Continued weakness in U.S. memory stocks and larger losses on individual leveraged positions could delay a rebound. The U.S. close on October 13 was not confirmed at the time of writing. Intraday declines are not presented as closing prices. Margin balances, forced selling, and options positions were also unavailable.&lt;/p&gt;
&lt;p&gt;Conversely, the competitive narrative could weaken if results show sustained AI customer demand and funding for investment, and if DDR5 contract prices hold better than feared. Samsung&amp;rsquo;s company notice set its third-quarter results announcement for 10:00 a.m. on October 29. TSMC&amp;rsquo;s earnings date and U.S. inflation data matter, but this article does not present schedules or consensus estimates that it could not recheck as confirmed values. &lt;a class="link" href="https://www.samsung.com/sec/ir/reports-disclosures/notices/" target="_blank" rel="noopener"
 &gt;Samsung IR notice&lt;/a&gt;&lt;/p&gt;
&lt;h2 id="after-a-rebound-watch-for-a-change-in-who-is-selling"&gt;After a rebound, watch for a change in who is selling
&lt;/h2&gt;&lt;p&gt;Two consecutive sessions of net foreign buying in Samsung cash equities would strengthen the trend-reversal case. This is an observation condition to allow for confirmation, not a statistically validated trading rule. Even if VKOSPI falls below 40, the assessment should remain cautious while cash-equity selling continues.&lt;/p&gt;
&lt;p&gt;A close that holds after a retest of KRW 228,000 is a price-based condition. If the won and government bonds weaken together and credit spreads widen, the analysis should expand to include country and credit risk. A single threshold such as USD/KRW 1,360 does not by itself confirm a shift to country risk.&lt;/p&gt;
&lt;p&gt;On the business side, first verify the specifications, customer qualification, commercial volume, and DDR5 contract prices for CXMT&amp;rsquo;s server product. These data can distinguish actual pressure on earnings from a share-price multiple contraction driven by fear. The evidence is insufficient to chase an opening-price gain without improved flows.&lt;/p&gt;
&lt;p&gt;Official materials include Samsung&amp;rsquo;s preliminary results and IR notice, and the ISSCC program. Market figures and CXMT comments rely on cited reports. Details on the buyback that remain unreconciled with the original filing, derivatives positions, and the U.S. close are explicitly limited. The probabilities and price path are the author&amp;rsquo;s conditional judgments.&lt;/p&gt;
&lt;p&gt;&lt;em&gt;This article is for information and research. The price scenarios do not guarantee returns.&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Disclaimer: For research and information purposes only. Not investment advice. Names cited are for analytical illustration; readers should perform their own due diligence and consult licensed advisors before any investment decision.&lt;/em&gt;&lt;/p&gt;</description></item></channel></rss>