Korea-Semiconductor on Korea Invest Insights

Anatomy of the Rebound: Foreign Buying Came in Two Bursts, and the Same Two Stocks Changed Hands

Sun, 16 Aug 2026 20:10:00 +0900

Connecting context: The Second Week of August in Review noted that foreign investors turned to four straight days of net buying starting Tuesday. That’s accurate, but it’s surface-level. This piece works from our own aggregation, breaking down eleven trading days of flow since the low by investor type, phase, and stock, to look at the structure underneath: who sold, who bought, and why buying at one moment differs in character from buying at another, even when both count as net buying.

TL;DR

The bottom line first. From August 10 to 14, the KOSPI rose from 6,258.77 to 6,977.94, up 11.49%, while retail investors net sold around 7 trillion won over the five sessions and foreign investors net bought in the mid-6 trillion won range. Institutions were modest net buyers. [Fact: exchange and media aggregation, totals by investor type vary by aggregation method] This rebound is a large-cap rally in which foreign investors absorbed the profit-taking supply that retail sold.
But split the full period since the low (July 30) into phases, and foreign buying was not continuous. Foreigners bought in the 7 trillion won range in a single day on July 31, absorbing capitulation supply, then sold back in the 8 trillion won range during the early-August stall (August 3 to 11), and bought again in the 7 trillion won range starting August 12 once the inflation rally began. [Fact: own aggregation] It’s an alternating structure: absorb the low’s supply, sell back during the stall, buy again on the news rally.
This structure carries a methodological lesson. Cut the window from August 1 to 14 instead, and you get the opposite conclusion: retail net buying, foreign net selling. That’s because July 31 alone falls outside the window. Reading flow data is half about how you set the window.
The rebound’s buying returned to the exact same spot as the crash’s selling. Since the low, the top foreign net-buying names have been Samsung Electronics and SK Hynix, together accounting for half of total net buying, and the two stocks also made up roughly 78% of the single day of buying on July 31. That is an exact mirror image of how foreign selling concentrated in these same two names during the July crash. [Fact: own aggregation] It is flow-side evidence for this series’ verdict that the decline was a positioning problem, not a fundamentals problem.
The direction of the handover in these two stocks is healthy. From August 10 to 14, in Samsung Electronics and SK Hynix, retail investors sold 4.6 trillion won, while foreign investors absorbed 2.15 trillion won and institutions absorbed 2.66 trillion won. Retail net selling in these two stocks accounted for about 65% of total retail net selling across the KOSPI. [Fact: own aggregation]
The caveats are just as clear. Foreign investors bought cash equities, but their cumulative August net-short position in futures was not fully unwound, and program trading was net selling. On the KOSDAQ, they were net sellers throughout the rebound. [Fact: own aggregation, public reporting confirms direction only] With cash strength, a futures hedge, and program-trading weakness all mixed together, it is too early to call this a full shift to risk-on.
The remaining checks are whether foreign net buying in cash equities holds up cumulatively, whether the cumulative futures short position shrinks, whether non-arbitrage program trading flips to net buying, and whether foreign ownership ratios in the two stocks rise. The last of these is still unmet. Despite the heavy buying, Samsung Electronics’ foreign ownership ratio is stuck at 46.70%, unchanged from the start of the month. [Fact: disclosure aggregation]
One more variable to flag before Monday’s open. Friday night, US retail sales fell 0.6% month over month, the largest miss versus consensus in 14 months, and the probability of a September hold rose into the 70% range. The US-Iran ceasefire extension remains at the agreement-in-principle stage past its expiration date, and Iran’s foreign minister said the two sides are not in negotiations. [Fact: reporting]

Key Framing

Even when the net-buying total looks the same, when the buying happened changes what it means. On July 31, foreign investors absorbed the supply that forced liquidation dumped into the market. In early August, they took profit on that position, and from August 12, they bought again in response to news on inflation and earnings. Read this alternation as one continuous buying streak and you overstate the strength of the flow. Cut the window starting from August 1 and you read the direction as the exact opposite. What is confirmed now is the symmetry: the same two stocks sold in the crash were bought back by the same hands in the rebound, and the handover ran from retail to foreign and institutional investors. What is not confirmed is whether this buying reflects conviction that has already shed its hedge. Futures and program trading will answer that question first.

1. The Surface of Five Days: Retail Sold, Foreigners Bought

Start with the numbers. From August 10 to 14, the KOSPI rose 11.49%, from 6,258.77 to 6,977.94. On Friday alone it gained 2.42% and touched an intraday high of 7,010.86, while Samsung Electronics closed at 274,500 won (up 2.43%) and SK Hynix closed at 1,645,000 won (up 3.26%). [Fact: exchange and media]

The five-day flow by investor type is where this piece starts.

Five days in mirror image. Retail selling and foreign buying face each other at nearly identical size every day; only Monday, when the rebound began, ran the other way. Own aggregation; the 10th, 12th and 14th match public reports, and for the 13th we use the value where most sources converge.

View as table

Date	KOSPI	Retail	Foreign	Institutional
Aug 10	+0.65%	+0.90tn won	-1.49tn won	+0.57tn won
Aug 11	+0.73%	-0.07tn won	+0.05tn won	+0.02tn won
Aug 12	+3.68%	-3.19tn won	+2.84tn won	+0.53tn won
Aug 13	+3.56%	-2.74tn won	+2.11tn won	+0.68tn won
Aug 14	+2.42%	-1.98tn won	+3.04tn won	-1.03tn won

The totals are approximately -7.1 trillion won for retail, +6.6 trillion won for foreigners, and +0.8 trillion won for institutions. Depending on aggregation method (whether off-exchange trades are included, how other-corporation trades are treated), weekly totals vary by outlet, from -7.4 trillion to -7.9 trillion won for retail and +6.6 trillion to +6.8 trillion won for foreigners, but the direction and magnitude agree. [Fact: media aggregation, cross-referenced] Notably, over the three days from the 12th to the 14th, foreigners net bought around 8 trillion won. This was not a one-day flash of buying. It was buying that built momentum, and the surface verdict, that this rebound is a large-cap rally in which foreign investors absorbed the profit-taking supply that retail sold, holds up this far.

2. Split Into Phases: Buying That Came in Two Bursts

But the picture changes once you start from the low. On a Samsung Electronics closing-price basis, the low came on July 30 (207,000 won), and the stock rose 32.6% over the eleven trading days from there to August 14. Splitting flow by investor type across this full period into phases looks like this.

Phase	Retail	Foreign	Detail
July 31 (single day)	-8.3 trillion won	+7.4 trillion won	The largest rebound day on record. After forced liquidation, foreign investors absorbed the capitulation supply. Samsung Electronics and SK Hynix made up about 78% of that day’s buying
August 3 to 11	+9.9 trillion won	-8.9 trillion won	Seven-session stall. Foreign investors took profit while retail bought the dip. Direction reversed
August 12 to 14	-7.7 trillion won	+7.7 trillion won	Three-day inflation rally. The reaction function shifted, and the investor types reversed again

[Fact: own aggregation, both-markets basis]

Buying split in two. Since the bottom, foreigners bought in two concentrated bursts, on July 31 and over the three days from August 12, and in the seven sessions between they sold back roughly what the first burst had taken in. Retail moved the opposite way in every phase. Own aggregation, both markets.

View as table

Phase	Retail	Foreign
July 31 alone	-8.3tn won	+7.4tn won
August 3 to 11	+9.9tn won	-8.9tn won
August 12 to 14	-7.7tn won	+7.7tn won

Foreign buying was not a continuous flow across the eleven trading days. It was two concentrated bursts of buying with one round of selling back in between. The first burst was the single day of July 31, and the second was the three days from August 12. In the seven trading days in between, foreigners sold back roughly as much as they had bought in the first burst, and retail investors took the other side. Absorb supply at the low, trim it during the stall, then come back in once direction is confirmed: this alternation is a textbook flow pattern that has repeated across first rallies off a bottom.

This breakdown carries a methodological lesson. Cut the window from August 1 to 14, and the numbers come out to roughly +2.2 trillion won in retail net buying and -1.2 trillion won in foreign net selling, the exact opposite of the surface verdict. That’s because it excludes the single day of July 31. With flow data, where you cut the calendar decides half the conclusion. One person could say “foreigners sold in August,” and another could say “foreigners drove the rebound,” and both could be arithmetically true. The window used for judgment should follow the structure of events (the low, the stall, the rally), not the calendar.

3. Buying That Returned to the Same Spot

What foreigners bought is the second finding. Since the low, the top foreign net-buying names have been Samsung Electronics (about 3.1 trillion won) and SK Hynix (about 2.6 trillion won), and the two stocks account for about 48% of total net buying. Broaden the semiconductor group to include SK Square and Samsung Electronics preferred shares, and the share tops half. That’s an exact mirror image of how foreign selling concentrated almost entirely in these same two stocks during the July crash. [Fact: own aggregation]

This symmetry tells us something. This series has judged the July decline as a positioning problem, not an information problem. If the same investor type sold and bought the same two stocks, it is more consistent to read this as a position being unwound and then restored than as a fundamental reassessment happening in between. Because the reason for selling was not fundamentals, the reason for buying didn’t wait for confirmation that fundamentals had improved either.

Looking at the handover in these two stocks by investor type over the five days makes the direction even clearer.

The handover in the same two stocks. Over five days retail sold 4.6 trillion won of Samsung Electronics and SK Hynix, and foreigners and institutions split the other side. Retail selling in these two names was about 65% of all KOSPI retail selling. Own aggregation.

View as table

Stock	Retail	Foreign	Institutional
Samsung Electronics	-2.70tn won	+0.95tn won	+1.91tn won
SK Hynix	-1.92tn won	+1.20tn won	+0.75tn won

In Samsung Electronics, while retail sold 2.70 trillion won, foreign investors absorbed 0.95 trillion won and institutions absorbed 1.91 trillion won. In SK Hynix, retail sold 1.92 trillion won, foreign investors absorbed 1.20 trillion won, and institutions absorbed 0.75 trillion won. Combined, that’s -4.62 trillion won for retail against +4.81 trillion won for foreign plus institutional investors, and retail net selling in these two stocks accounts for about 65% of total retail net selling across the KOSPI. [Fact: own aggregation] It’s a structure in which retail investors, who had built up positions through leverage and bargain buying during the sharp decline, took profit on the rebound, and that supply passed to foreign and institutional investors. As a direction of handover, this is healthy.

It should also be noted that handovers ran in the opposite direction too. Samsung Electro-Mechanics, which ranked among the top retail net-buying names (retail +0.38 trillion won against foreign -0.84 trillion won), and some small and mid-cap semiconductor stocks overlap with the top foreign net-selling names. And on the KOSDAQ, foreign investors were net sellers of about 1.7 trillion won throughout the rebound, with institutions and retail taking that supply. [Fact: own aggregation] Foreign investors’ August was a rotation out of the KOSDAQ and the periphery and into large-cap semiconductors, not a broader expansion of bets across the Korean market as a whole.

4. The Caveat: Cash Was Bought, the Hedge Remains

The reason it’s hard to call this buying a full shift to risk-on lies on the derivatives side.

By our own aggregation, foreign net buying in KOSPI 200 futures came to only 1,462 contracts from August 10 to 14, and the cumulative net-short position for the month, including the August 3 sell-off day, still stands at just over 14,000 contracts. Program trading over the same period was net selling of about 1.1 trillion won. [Fact: own aggregation, derivatives and program-trading data. Public reporting does not confirm the contract count] Media analysis points the same way: foreign investors are buying cash equities in size while combining it with futures selling to partially hedge market risk, and over the past month this pattern of flipping positions worth several trillion won on a single-day basis has repeated. [Fact: media analysis]

So the character of this buying doesn’t reduce to one thing. Foreign investors picked out and bought the large-cap names that had fallen the most, while offloading part of the index risk through futures, and in the process absorbed retail’s profit-taking and leverage-unwind supply. That’s different from a phase where passive money sweeps up the entire market.

Here is what needs to be confirmed before we can call this a full return to trend.

Confirmation condition	Current status
Cumulative foreign net buying in cash equities holds up	In progress. Three straight days in the 2 trillion won range since the 12th
Cumulative monthly KOSPI 200 futures short shrinks	Not met. Cumulative net-short position remains
Non-arbitrage program trading flips to net buying	Not met. Net selling for the week
Foreign ownership ratio rises in Samsung Electronics and SK Hynix	Not met. Samsung Electronics at 46.70%, unchanged from the start of the month

[Fact: own aggregation and disclosures. Ownership ratio as of August 14]

The last line may be surprising. Foreign investors bought in the trillions of won over five days, yet Samsung Electronics’ foreign ownership ratio sits at 46.70%, unchanged from the start of the month. This could be a technical factor, such as an aggregation lag or a change in shares outstanding, but the cause is unverified. [Blocked: cause unverified] The point at which the ownership ratio is confirmed to actually rise will be the strongest evidence that this buying is a restoration of holdings, not short-term trading.

5. The Variables for the Next Trading Day

Here are the variables hanging over Monday’s open. Friday night, US retail sales for July fell 0.6% month over month, badly missing the consensus estimate of +0.1%. That’s the largest monthly decline in 14 months. That’s rate-friendly, and the probability of a September hold rose into the 70% range, but the caution this series flagged in early August, that once employment and consumption weaken too much, falling rates start to read as recession risk instead, was tested against real data for the first time. US equities saw only limited declines despite the retail sales shock. [Fact: reporting]

The 60-day US-Iran ceasefire has passed its expiration date and remains at the stage of an agreement in principle to extend, and Iran’s foreign minister said over the weekend that the two sides are not in negotiations with the US. The blockade of the Strait of Hormuz and attacks on merchant vessels are continuing. [Fact: reporting] The oil-price axis heads into Monday as a still-open variable.

The conclusion on flow is this. The first leg of the rebound was a genuine handover led by foreign buying of cash equities, and its direction matches this series’ verdict. But roughly half of that buying still wears a hedge. Confirming that the three follow-up indicators, futures, program trading, and the ownership ratio, shed that hedge is the task for reading the next phase.

Stocks mentioned in this piece are examples for analysis and do not constitute a recommendation to buy or sell any specific security. Responsibility for investment decisions and their outcomes rests with the investor. Flow data by phase and by stock is our own aggregation, combining Naver Finance’s investor-type trading trends, brokerage flow data by stock, and derivatives and program-trading data; it may differ slightly from the exchange’s official statistics, and the stock-level totals are meant for ranking and directional judgment only. Because investor-type figures for August 13 varied across outlets (foreign net buying ranged from 2.1 trillion to 2.5 trillion won), we used the value where most sources converged, and weekly totals should likewise be understood as a range depending on aggregation method. Futures contract counts and program-trading amounts are not confirmed by public reporting, so we note explicitly that these are our own aggregation. The low date is July 30 on a Samsung Electronics closing-price basis. Prices and flow data are as of the market close on August 14, 2026.

AI's Bottleneck Really Is HBM, Not the GPU: But Bottlenecks Invite Workarounds

Sun, 16 Aug 2026 14:30:00 +0900

Series context: over the course of August, this series has covered in turn the deceleration in memory prices, the gap between confirmed demand and the stock price, and the conditions for a rebound. This piece verifies the physical foundation beneath all of that: how far the claim that AI’s bottleneck has moved from the GPU to HBM actually holds, what profit that bottleneck guarantees suppliers and for how long, and what countervailing force the bottleneck itself is generating.

TL;DR

Start with where the claim and its numbers come from. Morgan Stanley’s June report offers the frame that “the bottleneck moves along the supply chain: first chips, then power, then memory, then networking, then cooling”, along with the judgment that “memory will be in shortage through the end of 2026.” [Fact: Morgan Stanley, June 9] But the roughly 50 billion Gb of 2027 HBM demand often cited alongside it is a figure absent from that document. The real sources are Kiwoom Securities (53.4 billion Gb) and Korean research aggregation (6.20 exabytes, about 49.6 billion Gb converted), and working backward from TrendForce’s bit-share forecast lands in the same range. Because three separate lines converge, the number is usable, but its attribution needs to be corrected.
The first basis for the bottleneck is the gap in growth rates between compute and bandwidth. A single GPU’s compute has grown three to four times per generation, but HBM capacity is the same 288GB for both GB300 and the next-generation Rubin. What did grow is bandwidth, from the 8 terabyte class to the 20 terabyte class, and even that falls short of the roughly 3.5x growth in compute. [Fact: NVIDIA specifications, SemiAnalysis] When memory cannot keep up, buying more chips just means compute sits and waits.
The second basis is a change in the nature of inference. When an agent works for an extended stretch while holding a long context, the KV cache, an intermediate memory, occupies memory for the whole session. Field data shows the processing cost of a cached token versus an uncached one differs by a factor of 10, and NVIDIA thought the problem large enough to launch a dedicated cache storage tier as a separate product line earlier this year. [Fact: industry data] That said, no published quantitative estimate yet exists for how many gigabytes of additional HBM demand agent workloads generate. [Blocked: no public estimate]
The third basis is the arithmetic of supply. Producing the same capacity in HBM uses roughly three times the wafers of ordinary DRAM, and even if a single die’s yield is 95%, cumulative yield falls to 66% at 8 layers of stacking and 44% at 16 layers. A fab takes three to five years from groundbreaking to volume production, and new capacity does not contribute meaningfully until 2028. [Fact: industry data, TrendForce] That is why 2027 tightness looks less like a forecast and more like physics. Micron said it cannot fill even half of data center demand and that “2027 will be tighter than 2026”, and SK Hynix CEO Kwak Noh-jung described 2027, from a supply standpoint, as the tightest year in the industry’s history. [Fact: company statements]
But the same arithmetic is moving buyers too. Once memory cost reached 29% of system cost, NVIDIA cut the CPU-side memory module on Vera Rubin in half, reports say Rubin Ultra’s HBM will be set at 192GB, below the prior generation’s 288GB, and the prefill-only chip (Rubin CPX) uses GDDR7 instead of HBM. The High Bandwidth Flash that SanDisk and SK Hynix standardized builds a cheaper capacity tier beneath HBM. [Fact: TrendForce, company announcements] The deeper the bottleneck, the bigger the reward for working around it.
Where the bottleneck’s profit ends up is a matter of contracts. Micron has locked in roughly half its revenue through contracts with 16 strategic customers running to around 2030, “several-fold” increases in 2027 HBM contract prices are being discussed, and custom base dies tailored to individual customers (custom HBM) lock supplier and customer to each other. [Fact: company data, TrendForce] The rent is being fixed in the form of multi-year contracts rather than a spike in spot prices.
The implication for Korea. Two of the three oligopoly players are Korean, and SK Hynix’s HBM share sits in the mid-to-high 50% range. Profit visibility through 2027 is backed by the physics of the bottleneck. What remains open is 2028. New capacity lands that year, and CXMT’s HBM also arrives around then, still carrying yield problems. That is also when the workaround technologies described above accumulate enough effect to matter. The bottleneck claim is correct, but it is a claim that holds through 2027, and beyond that it becomes a matter of contract structure and supply discipline.

Key Framing

The diagnosis that AI’s bottleneck has moved from the GPU to HBM passes verification. The gap between compute and bandwidth, the memory occupied by agent inference, three times the wafers for half the yield, and a three-year lead time are all measured facts. But the bottleneck is not static. Once memory reached three-tenths of system cost, the largest buyer began stripping memory out of its designs, and the owners of the bottleneck are themselves fixing prices through multi-year contracts to narrow the swing of the next downturn. The profit coming out of the bottleneck is real, but its size is set by contract terms rather than market price, and its lifespan runs only until new supply arrives in 2028. What investors should ask is not whether the bottleneck exists, but whose books the bottleneck’s rent lands on, and in what form.

1. Where the Claim and Its Numbers Come From

Morgan Stanley Investment Management’s June report, “Big Picture: Ten Truths About Investing in AI,” is the document cited as the origin of this claim. What the document actually argues is this. The bottleneck in AI infrastructure does not stay in one place; it moves along the supply chain. It was chips first, then power, now memory, and next comes networking and cooling. Memory will be in shortage through the end of 2026, and incremental memory demand in 2027 will run 75 to 100 exabytes, doubling again in 2028. Agentic AI requires roughly a million times the compute of the original conversational models, and token consumption grew more than tenfold in 2025 alone. [Fact: Morgan Stanley, June 9, 2026]

There is one thing to correct here. The figure widely cited alongside this report, “roughly 50 billion Gb in 2027 AI-related HBM demand,” does not appear in the document. The document contains no HBM-specific figures and does not name any of the three suppliers. Tracing this number to its actual source turns up two lines. Kiwoom Securities estimated 2027 HBM demand at 53.4 billion Gb, a 56% increase from the prior year, and Korean research aggregating TrendForce, McKinsey, and others put it at 6.20 exabytes, roughly 49.6 billion Gb when converted to gigabits. [Fact: Kiwoom Securities, VLSI Research Korea]

Three routes to 50 billion Gb. Kiwoom's estimate, a research synthesis converting 6.20 exabytes, and our own back-solve from a 13% bit share land in the same place. The figure does not appear in the Morgan Stanley document.

View as table

Route	2027 HBM demand	Method
Kiwoom Securities	53.4bn Gb	+56% year on year
Research synthesis	about 49.6bn Gb	6.20 exabytes converted
Own back-solve	about 50bn Gb	TrendForce 13% bit share

The number itself is solid. Two estimates using different methods converge near 50 billion Gb, and a third path verifies it as well. TrendForce expects HBM to reach about 13% of total DRAM bit supply in 2027; if 50 billion Gb is 13%, total DRAM works out to roughly 380 to 440 billion Gb, which matches the industry’s usual scale estimates. [Inference: our own back-calculation] The conclusion is this. The number is usable, but the source is Korean brokerages and research aggregation, not Morgan Stanley, and it should be cited that way.

2. Why HBM, Not the GPU: The Gap in Growth Rates

The first foundation of the bottleneck claim is the gap between how fast compute grows and how fast memory grows. Line up GPU specifications generation by generation and the gap is plain to see.

Compute jumps, capacity stalls. GB300 and Vera Rubin carry the same 288GB of HBM. The generational gain is bandwidth, and even that (about 2.75x) trails the compute gain (about 3.5x) over the same span.

View as table

GPU	HBM capacity	Bandwidth
H100	80GB	3.35TB/s
B200	192GB	about 8TB/s
GB300	288GB	about 8TB/s
Vera Rubin	288GB	22 targeted, about 20TB/s reported

The H100’s HBM was 80GB at 3.35 terabytes per second. B200 moved to 192GB at 8 terabytes, and GB300 raised capacity to 288GB. But the next-generation Vera Rubin’s HBM4 capacity stays at the same 288GB. What grew is bandwidth, targeted at 22 terabytes per second, with actual early shipments reported around 20 terabytes, roughly 2.75 times GB300. Over the same span, GPU compute performance grew about 3.5 times. [Fact: NVIDIA specifications, SemiAnalysis, reporting] The pattern of compute outpacing bandwidth keeps repeating generation after generation, and this is the substance of what is called the memory wall.

This gap shows up as real cost during inference. Inference splits into two stages. The prefill stage, which processes the entire prompt at once, is compute-intensive, so GPU compute performance is the bottleneck there. But the decode stage, which generates tokens one at a time, has to reread the model weights and the entire intermediate memory from memory at every single token, so the bottleneck becomes memory bandwidth rather than compute. During this stage the GPU sits idle for a good share of the time, waiting on memory. That is why buying more expensive chips does not raise throughput when HBM cannot keep up.

3. What Agents Change: Resident Memory

The second foundation is the change in workload. Unlike a chatbot that answers one question at a time, an agent searches the web, writes code, tests it, and fixes it, over stretches that run from tens of minutes to several hours. What the model maintains to remember everything up to that point is the KV cache, an intermediate memory, and this cache grows larger the longer the context, the longer the session runs, and the more agents run at once. A single agent’s task splits into multiple model calls, for routing, search, tool selection, verification, and each call holds its own cache, so the memory resident at any one time grows to a different order of magnitude than in the chatbot era.

This shift already shows up in pricing. Field data puts the processing cost of a cached input token at one-tenth that of an uncached one, and operational data also shows that a single server’s DRAM ceiling (typically 1 to 2 terabytes) makes it hard to hold a cache for more than an hour. The fact that NVIDIA launched a dedicated KV cache storage tier (CMX) as a separate product line this past January is itself evidence that the problem has grown large enough to become a hardware product. [Fact: industry data, NVIDIA announcement] TrendForce says the spread of agents is shifting the CPU-to-GPU ratio in servers, with enterprise agents consuming up to four times the tokens of before, and projects the global memory market at $1.28 trillion in 2027. [Fact: TrendForce, May 29]

An honest gap is worth noting too. No published quantitative estimate yet exists for exactly how many exabytes agent workloads add to HBM demand. [Blocked: no public estimate] The direction is clear, but the magnitude still belongs to narrative rather than data.

4. Why Only Three Suppliers: The Arithmetic of Supply

The third foundation is a structure in which supply is hard to expand. Three numbers summarize that structure.

The arithmetic of stacking. A 95% die yield becomes 66% at 8-high and 44% at 16-high, and one bad via among more than 8,000 kills the stack. This arithmetic is why only three firms produce at scale and why HBM eats three times the wafers per bit.

View as table

Configuration	Cumulative yield	Calculation
Single die	95%	assumption
8-high stack	about 66%	0.95 to the 8th power
16-high stack	about 44%	0.95 to the 16th power

To produce the same capacity, HBM consumes roughly three times the wafers of ordinary server DRAM. This is because the dies are large, the through-silicon vias (TSVs) needed for stacking take up area, and yield losses compound. The arithmetic of yield is unforgiving. Even if a single die’s yield is 95%, cumulative yield falls to 66% at 8 layers of stacking and 44% at 16 layers. A single bad via, out of the more than 8,000 through-silicon vias on one die, is enough to kill the entire stack. [Fact: industry technical data] This barrier is what turned HBM into a three-way oligopoly, and Omdia concluded that HBM is far more complex than standard DRAM, that only three companies can produce it at volume, and that the bottleneck will last at least through 2027. [Fact: Omdia, July 30]

A time barrier sits on top of that. A fab takes three to five years from groundbreaking to volume production, and lithography equipment alone has a lead time of a year to a year and a half. Qualification eats time too. It took Samsung Electronics’s 12-layer HBM3E about 18 months to pass NVIDIA’s qualification, and the Blackwell cycle had already passed by the time it did. [Fact: industry data] That is why capacity now under construction (SK Hynix’s Cheongju M15X in mid-2027, Samsung’s Pyeongtaek P5 in 2028) cannot contribute meaningful volume before 2028 at the earliest. TrendForce’s supply-demand outlook points the same way. The DRAM fulfillment rate is -1% to -2% in 2026, the gap widens further in 2027, and meaningful new supply does not arrive until 2028. [Fact: TrendForce]

The result is what suppliers are saying right now. Micron’s chief business officer said on August 10 that the company cannot fill even half of data center customer demand and that “2027 will be tighter than 2026”, and SK Hynix CEO Kwak Noh-jung described 2027, from a supply standpoint, as the tightest year in the industry’s history. [Fact: TrendForce, press reporting] HBM’s share of DRAM wafer input rises from 18% (end of 2025) to 30% (end of 2027), but on a bit basis that is only 8% to 13%. The three-times-the-wafers structure shows up here as well. [Fact: TrendForce]

That covers the core of the bottleneck claim: the gap in growth rates, resident memory, and the arithmetic of supply. All three are measured facts, and the tightness through 2027 looks less like a forecast and more like physics.

5. The Workarounds the Bottleneck Invites

But this claim has another half. The deeper the bottleneck grows, the bigger the reward for designs that route around it. Over the past month, evidence of that workaround has been piling up.

The workarounds accumulate. Five confirmed this year: a dedicated cache tier, a prefill chip without HBM, module capacity halved, a flash-based capacity tier, and a flagship whose HBM may shrink. Design changes induced by price do not reverse when price falls.

View as table

When	Event	Nature
January	NVIDIA announces CMX	KV cache moved to a tier outside HBM
Roadmap	Rubin CPX adopts GDDR7	HBM removed from prefill
28 July	SOCAMM halving reported	CPU-side memory cut
3 August	High Bandwidth Flash spec	a capacity tier below HBM
Early August	Rubin Ultra at 192GB reported	flagship HBM under review

The most significant signal comes from NVIDIA itself. After memory cost climbed to 29% of Vera Rubin system cost, above the company’s preferred 20% ceiling, reports emerged that the CPU-side memory module (SOCAMM) would be cut in half, from 192GB to 96GB. CPU-side memory in a single rack falls from 55 terabytes to 28 terabytes. Supply allocation is part of the reason too: even combining all three suppliers, NVIDIA can only receive about 60% of estimated demand, so lowering the specification lets the company ship more systems from the same allocation. [Fact: TrendForce, July 28] Rubin Ultra went further. The specification originally announced as 1 terabyte of HBM4E has stepped down in stages, and reports now say the flagship configuration will be set at 192GB of HBM4, below even the current Rubin’s 288GB. Memory content falling as the generation rises is a direction never before seen in a flagship. [Fact: reporting, unconfirmed by NVIDIA]

The workaround is not limited to cutting specifications. Rubin CPX, designed solely for the prefill stage, skips HBM entirely and uses GDDR7 instead. The design splits prefill, a compute bottleneck, from decode, a memory bottleneck, at the hardware level, so it does not pay HBM’s price where HBM is not needed. High Bandwidth Flash (HBF), for which SanDisk and SK Hynix published a standard specification in early August, stacks NAND to deliver up to 512GB of capacity with bandwidth at the low end of the HBM class, and its consortium includes Google. It lays a cheaper capacity tier beneath HBM. On the software side, techniques that compress the KV cache itself have entered commercial models. The attention architecture used by the DeepSeek family cuts cache per token to one-third to one-fifth of the conventional approach. It is still concentrated in Chinese-origin models, but the direction is clear. [Fact: company announcements, technical literature]

The supplier’s own words confirm this list. SK Hynix Chairman Chey Tae-won, in a CNBC interview last week, likened demand to a war but also said prices had risen too fast. The reason the largest beneficiary of the bottleneck worries about how deep it runs is exactly the list above. Once a workaround triggered by price gets designed in, it does not reverse even if prices come back down.

6. Where the Bottleneck’s Rent Goes

If the bottleneck is real and the workarounds keep growing, the question left standing is who captures the bottleneck’s profit, and in what form. The answer is increasingly the contract.

On the demand side, multi-year contracts lock in volume. Micron has tied up roughly half its revenue through contracts with 16 strategic customers running to around 2030, and reports keep describing the three suppliers’ 2027 capacity as effectively fully allocated already. “Several-fold” increases in 2027 HBM contract prices are being discussed, though TrendForce’s own wording is a qualitative description that does not specify a multiple, so reading more precision into it than that would be an overreach. [Fact: company data, TrendForce] On the pricing side, customization is changing the structure. Starting with HBM4, the base die moves to TSMC logic processes (mainly 12nm, with premium 5nm and next-generation 3nm), and as NVIDIA and Google each demand their own custom specifications, HBM is turning from a general-purpose component into a part incompatible across customers. A custom base die makes it hard for a customer to switch suppliers, while also locking the supplier to that customer. [Fact: industry data]

What this structure means matches the conclusion this series has been building toward. The bottleneck’s rent is being fixed by the terms of multi-year contracts rather than a spike in spot prices. A fixed rent trims the upside in a boom in exchange for limiting the downside in a bust. The wide spread across institutions in HBM market size forecasts for 2027, from $116 billion to as much as 260 trillion won, is itself a feature of this transition, which makes the share of contract coverage a better thing to watch than any single figure. [Fact: various institutions, wide dispersion]

7. The Implication for Korea: 2027’s Physics, 2028’s Calendar

Here is what this verification means for Korean memory makers.

Ownership of the bottleneck is Korea’s position. Two of the three companies capable of volume production are Korean, and SK Hynix’s HBM share runs from the mid-50s to high-50s percent depending on the source. Some tallies put SK Hynix at roughly 70% of NVIDIA’s next-platform initial HBM4 volume. [Fact: reporting, share varies by source] Profit visibility through 2027 is backed by the physics laid out above. Because of lead times, 2027 supply is already determined at this very moment, and a substantial share of that volume is already sold under contract.

What remains open is 2028. New fabs begin contributing volume that year (Cheongju M15X, Pyeongtaek P5). CXMT’s HBM capacity, if it stays on plan, also climbs to a scale of 100,000 wafers that year, though still carrying a roughly 25% yield and a customer base limited to domestic Chinese buyers. That is also when the effects of the workaround technologies described above accumulate enough to matter. [Fact: company data, SemiAnalysis] The bottleneck claim is a matter of physics through 2027, but from 2028 on it reverts to a question of supply discipline and contract terms. The conclusion this series reached in early August, that the investment question is not whether prices rise further but whether they hold, and that holding is most solid when a contract guarantees it, survives intact even after passing through the physics of the bottleneck.

One last sentence, carried over directly from the Morgan Stanley report: “Infrastructure comes first. The applications that will justify that build-out are not yet visible.” Even the strongest supporter of the bottleneck claim attached this qualifier. A bottleneck produces rent only when demand is real. This earnings season showed demand measured in fact, and its persistence gets re-verified every quarter.

The stocks mentioned in this piece are illustrative examples for analysis and do not constitute a recommendation to buy or sell any particular stock. Responsibility for investment decisions and their outcomes rests with the investor. The figure of roughly 50 billion Gb in 2027 HBM demand is an estimate from Kiwoom Securities and Korean research aggregation, not a figure from the Morgan Stanley document, and the back-calculated share of total DRAM is our own computation. GPU specifications and bandwidth combine announced targets with reported figures and may differ from actual shipped specifications. Yield figures are example calculations drawn from industry technical data, and actual yields are not disclosed by company. The Rubin Ultra specification adjustment and the SOCAMM reduction come from research firms and press reporting and have not been confirmed by NVIDIA. Forecasts of HBM market size vary widely across institutions. The contribution of agent workloads to HBM demand has no published quantitative estimate, so only the direction is described. Prices and outlooks are as of mid-August 2026.

A Rebound Won't Come From a Single Trigger: A Verdict Calendar From Late August to Year-End

Sun, 09 Aug 2026 15:20:00 +0900

Related context: In What Hyperscaler Earnings Proved, and What They Didn’t: Three Explanations for the Memory Sell-Off, I wrote that demand had been confirmed, but pricing power and 2028 supply were still weighing on share prices. This piece takes up the next question. So what ends this phase? It scores the candidates across five axes for impact and timing, then lays out why no single one is enough and what to watch first.

TL;DR

Start with the record of last week’s three days. On August 5 the KOSPI rose 3.76% (SK Hynix up 5.77%); on the 6th it fell 4.58% (SK Hynix down 10.37%); on the 7th it fell another 0.60%. [Fact: exchange, press] The trigger for the plunge on the 6th was not macro but guidance from AMD and SanDisk. Oil prices and interest rates were actually favorable that day.
These three days are the starting point for this piece. Even when macro lines up favorably, a single semiconductor-specific catalyst can still take a stock down more than 10% in a day. Macro is a condition that supports the floor, not one that resolves the narrative.
The macro axis has already fired once. On August 7, the US July jobs report came in well short of expectations, and the probability of a September rate hike collapsed while the probability of a hold flipped to about 60%. [Fact: press] Even so, the KOSPI still closed lower that day.
The nearest major event is NVIDIA’s earnings on August 26, but this will be the fourth straight quarter in which a beat has failed to lift the stock. NVIDIA beat estimates in August and November 2025 and February and May 2026, yet the reaction was flat or negative each time, and the 52-week high came before the May report, not after it. [Fact: market data] The beat is already priced in.
So what matters on August 26 is not this quarter’s numbers but next quarter’s guidance and customer mix. On the disclosed figures, top-customer concentration is actually declining. Four customers accounting for 63% of revenue in fiscal Q3 2026 became three customers accounting for 54% in fiscal Q1 2027. [Fact: compiled from quarterly filings]
Shareholder returns carry the greatest impact, but this week a condition got attached to that axis. On August 7, SK Hynix announced a new fab investment of roughly $38bn. [Fact: press] When a returns announcement arrives alongside a capacity expansion, its power to lift the valuation floor weakens. The two decisions are two sides of the same capital allocation choice.
Export data has little impact but the highest frequency. The Korea Customs Service publishes figures three times a month (the 11th, the 21st, and the 1st of the following month), which keeps score on the contract-price debate on an ongoing basis. July semiconductor exports came to $41.1bn, up 178.8% year on year but down 8.5% from the prior month. [Fact: Korea Customs Service]
The conclusion is stacking. On July 30, Samsung Electronics reported its highest-ever earnings and closed lower anyway. This is a phase in which single events get sold into, and what is needed is a three-layer structure in which returns build on top of a macro landing and an NVIDIA confirmation. That sequence is already laid out on the calendar from late August through year-end.
The earliest signal is not on the trigger list. Since July, good news has lasted one day and bad news three. The moment the market posts its first two consecutive up days on good news is the empirical marker that the reaction function has changed, and the two days following NVIDIA’s earnings are the observation window for it.

Key Framing

Right now the market prices in good news for one day and bad news for three. As long as this asymmetry holds, no single trigger can build a trend. NVIDIA beating estimates for four straight quarters without its stock rising, and Samsung Electronics closing lower on the day it posted record earnings, are the same phenomenon. What investors should count, then, is not the number of triggers but the order in which they arrive, and what to check first is not the news itself but whether the reaction to the news has changed.

1. What Last Week’s Three Days Showed

Start with the numbers. On August 5, the KOSPI rose 239.31 points to close at 6,598.26, up 3.76%, with Samsung Electronics up 2.50% and SK Hynix up 5.77%. Foreign investors were net buyers of 1.4513tn won. The backdrop was falling oil prices on progress in talks to reopen the Strait of Hormuz. The next day, the 6th, the KOSPI fell 301.88 points to close at 6,296.38, down 4.58%, and slid as much as 5.46% intraday, triggering a sell-side sidecar (trading curb). Samsung Electronics fell 6.30% to 230,500 won, and SK Hynix fell 10.37% to 1,495,000 won. Foreign investors were net sellers of 3.3285tn won on the KOSPI, of which 1.6936tn won was SK Hynix alone. On the 7th, the index fell another 0.60%. [Fact: compiled from exchange and press data]

Three sessions in early August. SK Hynix fell more than 10% on the 6th, a day when macro was favourable, and the index still fell on the 7th when September hike odds collapsed. The problem is the reaction, not the news.

View as table

Date	KOSPI	Detail
5 August	+3.76%	oil lower, SK Hynix +5.77%
6 August	-4.58%	AMD and SanDisk guidance, SK Hynix -10.37%
7 August	-0.60%	US jobs miss, September hike odds collapse

The trigger identified for the plunge on the 6th was that AMD and SanDisk guidance had missed market expectations, dragging down the US semiconductor index the night before. [Fact: press reports] There is a point worth noting here. Macro was not bad that day. Oil prices were lower than the prior week, the US 10-year yield was around 4.6%, and the won was trading around 1,420 to the dollar. Even on a day when macro was lining up favorably, a single semiconductor-specific catalyst was enough to take SK Hynix down more than 10% in one day.

These three days set the premise for this piece. Before listing rebound candidates, the question to ask is not what news is coming but how the market is responding to news.

2. Impact and Timing Across Five Axes

This section lays out five candidates that could produce a rebound, their impact, their timing, and what to watch for each.

Axis	Impact	Earliest timing	What to watch	Limitation
Macro	Medium	Already partly fired	September hike probability, long-term rates, oil prices	Supports the floor but does not resolve the semiconductor narrative
NVIDIA earnings	Medium-high	August 26	Next-quarter guidance, customer concentration	The beat is already priced in
Shareholder returns	High	Guided for Q3	Size, whether it includes share buyback and cancellation, relationship to capacity expansion	Weakens if it arrives alongside a capacity expansion
Export data	Low	August 11	Direction of unit prices, rebound from the prior month	Confirms current results but does not resolve 2028
Product qualification	High	Ad hoc	Samsung HBM4 customer qualification, HBM4 mix	Timing cannot be predicted

[Inference: the impact and timing assessments are this author’s own judgment]

What stands out in this table is that the axes with the greatest impact have the least certain timing, and the axes with the most certain timing have the least impact. The August 11 preliminary export figures have a fixed date, but that alone will not change the trend, while confirmation of Samsung’s HBM4 customer qualification would trigger the biggest re-rating, but there is no way to know when it will come.

3. Macro Has Already Fired Once

Macro easing, long flagged as a catalyst still to come, partly materialized last Friday. The US July jobs report, released on August 7, came in well short of expectations, and the probability of a September rate hike collapsed while the probability of a hold flipped to about 60%. [Fact: press] That is a reversal from as recently as late July, when the probability of a hike had stood above 80%.

Yet the KOSPI still closed lower that day. News that eased the discount-rate burden came out, and Korean semiconductor stocks did not respond to it. It is the same structure as August 6. Semiconductors do not rise even when macro is good, and they still fall on a good macro day if a semiconductor-specific negative shows up.

This week’s calendar is packed with macro data. US CPI comes on August 12, PPI on the 13th, and retail sales on the 14th, and Treasury auctions land on the same dates: $42bn of 10-year notes on the 12th and $25bn of 30-year bonds on the 13th. [Fact: Treasury Department] Because inflation data and long-bond auctions fall on the same days, rate volatility could pick up, and that cuts both ways.

Macro is a necessary condition. If it breaks down, a rebound is difficult even if every other axis is favorable, and if it improves on its own, that alone will not bring a rebound either.

4. NVIDIA Earnings: The Beat Is Already Priced In

NVIDIA reports fiscal 2027 second-quarter earnings on August 26. Company guidance points to revenue around $91bn, and consensus sits a little higher, around $91.8bn. [Fact: NVIDIA disclosures, market compilation]

It is natural to treat NVIDIA’s earnings as the nearest major trigger, but the record tells a different story.

Drawdown from recent highs. NVIDIA has barely come off its high while the two Korean memory names fell hard. This is where demand confirmation and share prices parted. Samsung's high rests on a single report and is approximate.

View as table

Name	From high	Date of high
NVIDIA	-5.3%	14 May
Samsung Electronics	-38.4%	19 June
SK Hynix	-50.0%	25 June

For four consecutive quarters, a beat has not translated into a higher stock price. The August and November 2025 reports and the February and May 2026 reports all came in above market expectations, yet the reaction was flat or negative each time. The May 20 report in particular beat on both revenue and profit and carried good guidance, but the stock drifted lower afterward, falling into the $200s by late July, and its 52-week high of $236.54 had already been set on May 14, before that earnings report. [Fact: market data]

So what matters on August 26 is not this quarter’s numbers. Next-quarter guidance comes first. Consensus currently expects next-quarter revenue growth of around 81%; only if the company guides toward the 90% range would this actually become a catalyst.

Second is customer mix. NVIDIA does not disclose revenue split between hyperscalers and other customers, so diversification cannot be confirmed along that line. Instead, its quarterly filings disclose how many customers account for more than 10% of revenue and what percentage each represents. On these figures, concentration is actually declining.

NVIDIA customer concentration. The combined share of customers above 10% of revenue fell from 63% to 54%. NVIDIA does not publish this as a metric; it is compiled from the customer concentration disclosure in the quarterly filings.

View as table

Quarter	Customers	Combined share
Fiscal 2026 Q3	4	63%
Fiscal 2026 Q4	4	61%
Fiscal 2027 Q1	3	54%

In fiscal Q3 2026, four customers accounted for 23%, 16%, 13%, and 11% of revenue, 63% combined; by fiscal Q1 2027, three customers accounted for 21%, 17%, and 16%, 54% combined. [Fact: compiled from quarterly filings] This is verifiable evidence that the demand base is broadening, and if this number falls further, it becomes a direct rebuttal to doubts about 2028 demand.

Third, days sales outstanding is worth watching alongside this. It is the metric most often cited as the accounting trace that vendor-financing concerns would leave behind, and the recent trend points toward normalization rather than deterioration. It went from 53 days in fiscal Q3 2026 to 51 days in Q4, then to 45 days in fiscal Q1 2027, close to the 2020 to 2024 average of 46 days. In that same quarter, receivables grew 84% year on year while revenue grew 85%, so receivables grew more slowly than revenue. [Fact: compiled from quarterly filings] If this metric worsens again, it would then become a trigger pointing the other way.

5. Shareholder Returns: High Impact, Now With a Condition Attached

Shareholder returns carry the greatest impact because they raise the valuation floor by contract. As discussed in an earlier piece, the market is currently pricing memory earnings with a low persistence rate, but a returns policy fixes the shareholder’s share regardless of that persistence rate. The greater the doubt about how long profits will last, the greater the relative value of a committed return.

Here is where the two companies stand. Samsung Electronics has a policy, running from 2024 through 2026, of returning 50% of free cash flow and paying a regular annual dividend of 9.8tn won, and it has disclosed a condition that it will consider additional returns if cumulative three-year return capacity exceeds the existing dividend. Net cash stood at 167.6tn won at the end of the second quarter. On July 31, the company said it was discussing a concrete plan that would include a special dividend. SK Hynix has a policy, running from 2025 through 2027, of a fixed dividend of 1,500 won per share plus roughly 50% of cumulative free cash flow as the source of returns, and at its July 29 earnings call it said additional returns were under review and would be announced within the year. On August 7, when it set a quarterly dividend of 375 won, it reconfirmed that it would finalize an additional returns plan during the third quarter. Net cash stood at 69.4tn won at the end of the second quarter. [Fact: both companies’ disclosures and earnings calls]

Yet on that same August 7, SK Hynix also announced a new fab investment of roughly $38bn. That covers 35.2tn won for DRAM at the Yongin P2 fab and 19.1tn won for NAND at the Cheongju M17 fab, both running through 2031. [Fact: press]

The fact that these two announcements landed on the same day changes the character of the shareholder-returns trigger. A large part of the discount the market applies to memory stocks reflects concern about 2028 supply, and when a returns announcement arrives alongside a capacity-expansion announcement, it becomes unclear whether capital allocation is leaning toward discipline or toward expansion. Returns do the most to lift the valuation floor when they read as a signal of supply discipline. Raising dividends and capacity at the same time lets the two messages cancel each other out.

So the amount is not the only thing to check in a returns announcement. It also needs to spell out whether the company will carry out a share buyback and cancellation, how long the policy period runs, and how it relates to capital expenditure plans over the same period. For reference, the 100tn won figure circulating in the market has not been confirmed by the company. [Blocked: unconfirmed by the company]

6. Export Data: Low Impact, Highest Frequency

The Korea Customs Service releases export data three times a month: preliminary figures for the 1st through the 10th around the 11th, figures through the 20th around the 21st, and the confirmed monthly total on the 1st of the following month. The next release is August 11. [Fact: Korea Customs Service]

Two readings of July semiconductor exports. The same $41.1bn is the second-highest month ever year on year and a decline month on month, because June set the record at $44.8bn.

View as table

Basis	Change	Note
Year on year	+178.8%	second highest ever
Month on month	-8.5%	June’s $44.8bn was the record

Here is where July stood. Total exports came to $98.89bn, up 62.8% year on year and the highest ever for a July, and semiconductor exports came to $41.1bn, up 178.8% and the second-highest on record. On a month-on-month basis, though, they fell 8.5%, because June had been the all-time high at $44.8bn. [Fact: Korea Customs Service, Ministry of Trade, Industry and Energy]

Any single release has little impact on its own. Exports are a lagging indicator confirming transactions that have already happened, so they cannot resolve the 2028 profitability debate. Instead, this data’s value lies in its frequency. It is the only public source that measures, three times a month, where contract prices are actually heading, and dividing export value by volume gives an approximation of unit-price direction. What to watch in the August data is not the year-on-year growth rate but whether the month-on-month decline seen in July gets reversed.

7. Stacking and the Calendar

Putting the scoring together, the conclusion is that a single trigger is not enough. The evidence is what actually happened in July. On July 30, Samsung Electronics reported record quarterly operating profit of 89.5tn won, and its stock rose more than 4% intraday before reversing to close lower. NVIDIA beat estimates for four straight quarters without its stock rising. One good number does not end this phase.

NVIDIA’s 2023 episode is often cited, but even that case was not simple. After its May 2023 report, when it guided next-quarter revenue to 1.5 times consensus, the stock jumped around 25% in a single day, but it then traded sideways through the summer and fall, and when it beat by a wide margin again in November, the stock actually fell more than 2% the next day. The trend was only confirmed after that. [Fact: market data] It took a chain of confirmations, not one.

The calendar from August to October. The catalysts arrive in sequence: macro and demand confirmation first, then the items that move valuation.

View as table

When	Event
11 August	Korea export data, first ten days
12-14 August	US CPI, PPI, retail sales, Treasury auctions
26 August	NVIDIA fiscal Q2 2027 results
27-29 August	Jackson Hole
15-16 September	FOMC
During Q3	SK Hynix additional shareholder return
Late September	Micron fiscal Q4 results
Late October	Samsung and SK Hynix Q3 results

The calendar lays out the catalysts in sequence: preliminary export figures on August 11; US inflation data, retail sales, and Treasury auctions from the 12th through the 14th; NVIDIA earnings on the 26th; Jackson Hole from the 27th through the 29th; confirmed August export figures on September 1; the FOMC meeting on September 15 and 16; Micron earnings at the end of September; and then SK Hynix’s returns plan sometime in the third quarter along with third-quarter earnings at the end of October.

This sequence divides into three layers by function. A macro landing builds the floor. On top of that, NVIDIA earnings, export data, and product qualification confirm demand. Finally, shareholder returns and third-quarter earnings move the valuation. If only the first two layers fill in, the result is a technical rebound at best; a trend reversal requires all three.

8. The Earliest Signal Is Not a Trigger

One last signal, not on the list above. Since July, this market’s defining trait has been that good news doesn’t last a day while bad news lasts three. The record surge on July 31 was followed by the plunge on August 3, and the rise on August 5 was followed by the plunge on the 6th.

The moment the market posts its first two consecutive up days on good news is the empirical evidence that the reaction function has changed. This is likely to show up before any of the triggers do, and it is simple to check: watch the cumulative reaction over the two days following NVIDIA’s August 26 earnings. If the stock rises on day one and again on day two, the market has started digesting good news; if it rises on day one and gives that back on day two, it is still the same phase.

Here are the conditions that would overturn this piece’s verdict. If Samsung’s HBM4 customer qualification is confirmed, if third-quarter server DRAM contract prices come in well above forecasts, or if shareholder returns are announced alongside a scaled-back capacity plan, the impact ranking above would shift. Conversely, if August exports fall again month on month and NVIDIA guides next-quarter revenue below consensus, the three-layer structure this piece lays out would stall at the first stage.

The stocks mentioned in this piece are examples for analysis and are not a recommendation to buy or sell any specific security. Responsibility for investment decisions and their outcomes rests with the investor. The impact and timing assessments and the three-layer breakdown are judgment calls, not statistical estimates. NVIDIA’s customer concentration and days sales outstanding are figures compiled from its quarterly filings, not values the company discloses as separate metrics. The market’s figures for the size of shareholder returns have not been confirmed by either company. A month-on-month breakdown of July exports by product was not available in public data, so only the total figure is cited. Prices and indicators are as of the August 7, 2026 close, and the market is closed as of August 9.

What Hyperscaler Earnings Proved, and What They Didn't: Three Explanations for the Memory Sell-Off

Tue, 04 Aug 2026 18:50:00 +0900

Related context: In the August 3 Plunge and Contract Price Deceleration, I wrote that the pace of memory price increases had broken and the investment question had shifted from further price gains to whether prices would hold. This piece looks at a different axis. In the same week, Microsoft and Amazon produced evidence that AI investment is converting into real revenue, and at the very moment demand was confirmed, foreign investors concentrated their selling in the two memory makers. This is the point where the demand axis and the supply axis diverge, and it is what now explains Korean memory stock prices.

TL;DR

What Microsoft and Amazon changed is not the fact that capital spending is continuing, but the strength of the evidence that this spending is converting into revenue, profit, and backlog. Azure grew 43% for the quarter, and on a fiscal-year basis, annual revenue topped $100bn for the first time (41% annual growth). AWS grew 36.7%, with an operating margin of 39.4% and operating profit up 64%. [Fact: company statements and earnings calls]
The single strongest line came from Microsoft’s CFO. Commercial backlog rose 84% to $678bn, and the full $51bn net increase came from customers outside the frontier model companies. Backlog growth excluding OpenAI was still 25%. [Fact: earnings call] This is a concrete data point that directly cuts against the suspicion that backlog is mostly circular contracting with a handful of AI companies.
Backlog should not be read as if it were cash, though. The same call disclosed the structure behind it: a weighted average duration of 2.3 years, with about 30% due to be recognized within the next 12 months. The number that actually matters is that the 12-month portion rose 37% year over year. [Fact: earnings call]
Yet in the same week that demand was confirmed, Korean memory stocks sold off. Over the month of July, foreign investors net sold more than 11tn won of Samsung Electronics and SK Hynix combined, while net buying other Korean stocks over the same period. They bought semiconductor parts, materials and equipment names led by LG Innotek, along with autos. [Fact: aggregated from press reports; figures vary by source] This was not an exit from Korea. It was a reduction in memory specifically.
Domestic leverage unwinding compounded this. The margin loan balance fell from 38.6tn won at the end of June to 28.9tn won on July 31, with that single day’s decline the largest on record. Net assets in single-stock leveraged ETFs fell from about 16tn won at the end of June to about 5.7tn won on July 30. [Fact: Korea Financial Investment Association and press reports] The record 7.2tn won of foreign net buying on July 31 is interpreted as the trade that absorbed that forced selling.
We set up and scored three hypotheses for why memory sold off despite confirmed demand: a timing-lag hypothesis (how long confirmation takes to reach earnings), a pricing-regime hypothesis (pricing power shifting even as demand grows), and a supply-arrival hypothesis (2028 capacity additions overlapping). What governs price right now is not the first of these but the second and third, which is why good demand news alone does not resolve it.
The clearest evidence for the second hypothesis came this week. After memory cost rose to 29% of system cost, above NVIDIA’s preferred 20% ceiling, the company is cutting Vera Rubin’s module capacity in half. LPDDR5X per rack falls from about 55TB to about 28TB. [Fact: TrendForce] Even the strongest demand source redesigns around less memory once price rises far enough.
The KOSPI rose 1.62% today, and Samsung Electronics and SK Hynix both closed slightly higher. Both stocks, however, remain 36% and 47% below their June highs, respectively. [Fact: market data] Demand risk has fallen, and supply risk has not. That asymmetry is what today’s price reflects.

Key Framing

What this earnings season proved is that AI investment is returning as revenue. What it did not prove is how much of that revenue stays as margin for memory makers. Microsoft’s and Amazon’s income statements answered the first question. Contract structure and 2028 capacity additions will answer the second. Foreign investors selling only memory while buying parts and materials names is not an exit from Korea. It is pricing that separates these two questions. The logic of buying memory on confirmed demand alone is therefore only half right. The other half depends on where pricing power ends up.

1. What Was Actually Proven

Start with the numbers from both companies. Microsoft posted fiscal fourth-quarter revenue of $90bn and operating income of $40.6bn, with earnings per share up 32%. Azure grew 43% for the quarter, accelerating from 40% the prior quarter, and on a fiscal-year basis, Azure’s annual revenue topped $100bn for the first time, with annual growth of 41%. Amazon’s AWS revenue was $42.2bn, up 36.7%, the fastest growth rate in 18 quarters, and AWS operating income rose 64% to $16.6bn, putting the operating margin at 39.4%. [Fact: company statements and earnings calls]

Growth in Q2 2026. Azure and AWS both accelerated, and AWS lifted its operating margin to 39.4% while raising capital spending. Backlog growth is shown in a different colour because it measures something else.

View as table

Measure	Year on year
Azure revenue	+43%
AWS revenue	+36.7%
AWS operating income	+63.6%
Microsoft commercial RPO	+84%

What deserves more attention than the growth rate is that spending and returns moved together in the same quarter. AWS raised its operating margin from 32.9% a year earlier to 39.4% while increasing capital spending. That points to a phase where investment is generating margin, not eroding it.

The backlog figure is more decisive. Microsoft’s commercial backlog rose 84% to $678bn. What matters more than the headline number is its composition. Of the $51bn increase from $627bn the prior quarter, the CFO said the entire sequential increase came from commitments by customers outside the frontier model companies. Backlog growth excluding OpenAI was 25%, and bookings growth excluding OpenAI was 18%. [Fact: earnings call]

Why this line matters becomes clear if you recall last month’s doubts. If most of a large-looking backlog were commitments from a single customer, OpenAI, it would look more like circular dealing, and it would wobble if OpenAI’s finances wobbled. The fact that the entire net increase came from outside that customer means the base of demand has genuinely broadened.

2. How Far to Trust the Backlog

Backlog should not be read as equivalent to revenue, though. The structure the company disclosed on the same call forces that restraint. Microsoft’s backlog has a weighted average duration of 2.3 years, and about 30% is scheduled to be recognized as revenue within the next 12 months. [Fact: earnings call]

The term structure of the backlog. Only about 30% of the $678bn becomes revenue within 12 months. The number that matters is not the total but that this 12-month portion grew 37% year on year.

View as table

Portion	Amount	Note
Recognised within 12 months	about $203bn	up 37% year on year
Recognised later	about $475bn	weighted average duration 2.3 years
Total	$678bn	up 84% year on year

That works out to roughly $200bn of the $678bn becoming revenue within a year. The rest is recognized over the following two to three years. Reading the 84% backlog increase as an 84% increase in next year’s revenue would therefore be wrong.

The number that actually matters is different: the portion due for recognition within 12 months rose 37% year over year. That is the indicator of how fast backlog is converting into near-term revenue, and it lines up with Azure’s actual 43% growth rate and its 45% guidance for next quarter. Backlog and revenue are not moving independently. They are confirmed to be moving in the same direction.

It is also worth noting that the financial strain has not disappeared. Microsoft said it expects to remain free-cash-flow positive in the next fiscal year, but Amazon’s trailing-12-month free cash flow was negative $7.6bn, a reversal from positive $18.2bn in the same period a year earlier. That is the result of net property and equipment purchases rising $66.1bn in a single year. [Fact: company disclosures] Even within the same cycle, financial capacity diverges across companies.

3. So Why Did Memory Get Sold?

Korean memory stocks sold off in the very week that demand was confirmed. Who was selling can be identified down to the investor category.

Foreign net buying in July 2026. Foreign investors cut the two memory names specifically while buying the rest of the Korean market. The memory figure ranges from 11.3tn to 13.1tn won across sources, and the rest of KOSPI is back-solved.

View as table

Group	Net buying	Note
Samsung Electronics and SK Hynix	about -11tn won or more	11.3tn-13.1tn across sources
Rest of KOSPI	about +2.5tn won	back-solved

Over the month of July, foreign investors net sold more than 11tn won of Samsung Electronics and SK Hynix combined. Depending on the source, the total ranges from 11.3tn won to 13.1tn won. But foreign net selling across the entire KOSPI over the same period was smaller than that figure. Strip out the two stocks and foreign investors were net buyers of other Korean stocks. [Fact: aggregated from press reports; figures vary by source] The top foreign net-buy names in July were in fact led by LG Innotek, followed by semiconductor parts, materials and equipment names such as DB HiTek, Leeno Industrial, Hanmi Semiconductor and Samsung Electro-Mechanics, along with Hyundai Motor and Naver.

This narrows the interpretation. Foreign investors did not exit Korea, and they did not abandon semiconductors. They cut large-cap memory specifically and rotated into other parts of the same industry.

Domestic leverage unwinding compounded this. The margin loan balance fell from 38.6tn won at the end of June to 28.9tn won on July 31, with that single day’s drop the largest on record. Net assets in single-stock leveraged ETFs tracking Samsung Electronics and SK Hynix fell to roughly a third of their prior level, from about 16tn won at the end of June to about 5.7tn won on July 30. Average monthly returns on those funds were negative 47%. [Fact: Korea Financial Investment Association and press reports]

The July deleveraging. Margin loan balances and single-stock leveraged ETF assets both fell sharply within a month. The single-day drop in margin loans on 31 July was the largest on record.

View as table

Measure	End of June	End of July
Margin loan balance	38.6tn won	28.9tn won (31 Jul)
Single-stock leveraged ETF assets	about 16tn won	about 5.7tn won (30 Jul)

The record 7.2tn won of foreign net buying on July 31 should also be read in this context. Domestic brokerages interpret it as buying that came after a large share of the forced-selling volume had already been absorbed. [Inference: aggregated from brokerage commentary] After one more round of larger selling on August 3, the KOSPI rose 1.62% today, and Samsung Electronics and SK Hynix closed up 0.21% and 0.64% respectively. Both stocks remain 36% and 47% below their June highs, respectively. [Fact: market data]

4. Three Explanations, Scored

There are three hypotheses that explain why memory sold off despite confirmed demand. Here is a judgment on which one is governing price right now.

Hypothesis	Description	Implication	Current score
Timing lag	It takes two to three quarters for hyperscaler earnings visibility to pass through memory contracts and reach reported results. The market will not wait that long	If you wait, it resolves itself, so now is an opportunity	Partly true, but this alone does not explain buying parts and materials names while selling only memory
Pricing regime	Even as demand grows, pricing power shifting to a small number of large buyers lets volume rise while margin gets compressed	The discount is justified and will not be resolved by demand news	Dominant. Slowing contract price growth and the expansion of long-term contracts are evidence, and this week’s NVIDIA case is decisive
Supply arrival	Korea’s new fabs and CXMT’s capacity additions both land in 2028	Confirmed demand is a 2026-27 story and does not settle the 2028 debate	Dominant. The direct trigger for this morning’s weakness was also a report on CXMT reviewing a new Beijing plant

[Inference: hypothesis framing and scoring are our own judgment]

The third row has today’s real-world example. The direct trigger for this morning’s semiconductor weakness was a report that CXMT is reviewing construction of a new DRAM plant in Beijing. [Fact: press report] Three days after evidence of demand from U.S. big tech emerged, a single piece of Chinese capacity news pushed the sector back down. That contrast shows which axis the market is actually watching right now.

Domestic brokerages’ price target cuts point to the same axis. When several brokerages recently lowered their targets, the reasons cited were falling NAND contract prices, progress in China’s domestic lithography equipment, and CXMT’s planned listing. It is hard to find one that cited weak demand as the reason. [Fact: aggregated from brokerage reports] That said, most of the brokerages that cut targets kept their ratings unchanged and characterized the current pullback as excessive relative to fundamentals.

5. NVIDIA Is Cutting Back on Memory

The clearest evidence supporting the second hypothesis came this week. It is also the strongest counterargument to the bull case.

NVIDIA is moving to cut the capacity of the SOCAMM modules used in Vera Rubin systems in half, from 192GB to 96GB per module. The reason is that memory cost rose to 29% of the roughly $2.1 million system cost, above the company’s preferred 20% ceiling. If this change goes through, LPDDR5X per NVL72 rack falls from about 55TB to about 28TB, and the corresponding memory cost per rack falls from about $1.2 million to about $590,000. [Fact: TrendForce, July 28]

As price rose, the strongest demand source is redesigning its systems to use less memory. The same mechanism covered in yesterday’s post, the affordability ceiling on consumer devices, is now showing up inside AI servers. This does not mean demand is disappearing. It means price is limiting itself, and it shows where the math of multiplying volume growth by rising unit prices to project revenue eventually stops.

There is one more signal pointing the same way. What HBM4 adds is not capacity but bandwidth. The prior generation, GB300, already carried 288GB of HBM3E per GPU, and Vera Rubin’s HBM4 is also 288GB, the same capacity. What changed is bandwidth, up roughly 2.75 times from 8TB/s to as much as 22TB/s. [Fact: NVIDIA technical documentation] That runs against the common assumption that moving to HBM4 automatically means more memory content per system. Total volume still rises if GPU count per rack and unit shipments increase, but the generational transition itself does not automatically lift volume.

6. Rubin’s Timeline

Before sizing the benefit, the timing needs to be precise. Vera Rubin and Rubin Ultra are different products on different schedules.

The Rubin schedule. Vera Rubin is what can be put into supply-chain numbers today; Rubin Ultra belongs in the option set rather than the base case. The 26 August print is still Blackwell-led.

View as table

When	What
June 2026	Vera Rubin in full production, HBM4 qualified at three suppliers
Q3 2026	Partner shipments begin
26 August 2026	NVIDIA fiscal Q2 results
Q4 2026	Volume ramp; Rubin revenue from the November print
H2 2027	Rubin Ultra official schedule, with reported delays

Vera Rubin entered full production on June 1, with partner shipments in the third quarter of this year and volume ramp in the fourth quarter. All three HBM4 suppliers, Samsung Electronics, SK Hynix and Micron, were confirmed to have passed qualification and entered mass production as of early June. [Fact: NVIDIA and press reports] In other words, Vera Rubin is what can actually show up in supply chain results right now.

Rubin Ultra sits in the second half of 2027 on the official roadmap, but two separate delay signals have surfaced. Reports in early July said the Kyber rack is being pushed to 2028 because of substrate manufacturing problems, and reported HBM capacity per GPU has been revised down twice, from an initial 1TB to 384GB and then to around 192GB. NVIDIA has said its roadmap is intact, but that statement refers to the standard product shipping in the second half of this year, so both claims can be true at once. [Fact: press reports, unconfirmed by NVIDIA] Rubin Ultra and HBM4E should therefore stay in the outlook, but it is premature to treat them as the base case.

NVIDIA’s next earnings report is on August 26, and that report will still be centered on Blackwell. Rubin revenue will not actually show up in the numbers until the November report. [Fact: NVIDIA disclosures] What fills that gap in between is what to watch through August and September.

7. Verdict and Watch List

What this earnings season lowered is demand risk. What it did not raise is supply risk, and what it newly revealed is a path by which price limits demand on its own. The judgment on memory stock prices right now comes down to this. The logic of buying memory on the strength of hyperscaler earnings is only half right. The part about confirmed demand is correct. How much of that demand turns into memory makers’ margin is decided by a different axis.

Here is a record of what would need to be confirmed to rule each hypothesis in or out.

What to confirm	Timing	Which hypothesis it moves
Absolute level of contract liabilities in Samsung Electronics’ and SK Hynix’s quarterly reports	Mid-August	A large increase in customer prepayments weakens the pricing-regime hypothesis
NVIDIA’s fiscal Q2 results and commentary on Rubin	August 26	An earlier ramp schedule strengthens the timing-lag hypothesis
Third-quarter server DRAM contract price settlement	September	Coming in above the 13-18% forecast weakens the pricing-regime hypothesis
Whether CXMT’s Beijing plant actually breaks ground	Ongoing	Confirmation strengthens the supply-arrival hypothesis
Whether NVIDIA’s SOCAMM capacity cut is actually implemented	Ongoing	If implemented, price elasticity is confirmed and volume forecasts should be lowered
Whether foreign investors turn net buyers of memory again	Ongoing	The simplest indicator that the supply-demand overhang has been worked off

Finally, here are the conditions that would overturn this piece’s verdict. If third-quarter contract prices come in well above forecast, if NVIDIA reverses the SOCAMM capacity cut, or if CXMT’s capacity additions are confirmed to be delayed, the weight of the second and third hypotheses falls and the timing-lag hypothesis becomes dominant. At that point, the verdict would shift toward today’s discount having been an opportunity.

The stocks named in this piece are examples for analysis and do not constitute a recommendation to buy or sell any specific security. Responsibility for investment decisions and their outcomes rests with the investor. Microsoft’s and Amazon’s earnings figures are based on each company’s statements and earnings calls. The combined foreign net-selling figure for July for the two stocks ranges from 11.3tn won to 13.1tn won depending on the source, so this piece states it only as more than 11tn won; the estimate of net buying excluding memory is a derived figure and has limited precision. Margin loan and leveraged ETF figures are based on press reports, and some items show minor differences across sources. NVIDIA’s SOCAMM capacity cut, the Rubin Ultra specification changes, and the Kyber rack delay come from research firms and press reports and have not been confirmed by NVIDIA. The construction and scoring of the three hypotheses is judgment, not statistical estimation. Prices and indicators are as of the close on August 4, 2026.

Memory No Longer Needs Prices to Rise Further: Anatomy of the August 3 Plunge and the Contract Price Deceleration

Mon, 03 Aug 2026 18:40:00 +0900

Connected context: July Earnings Season Wrap: AI Demand Was Confirmed, and Memory Pricing Became an Industry-Wide Cost noted that memory prices had, for the first time, become a cost line item outside the memory companies themselves. Data showing that the pace of that price increase is slowing has become today’s market language. This piece dissects the August 3 plunge, sorts out what that data actually says and does not say, and lays out where this cycle’s investment question has moved.

TL;DR

On August 3, the KOSPI closed at 6,257.45, down 5.12%, while Samsung Electronics fell 8.76% and SK Hynix fell 8.79%, and the KOSDAQ actually rose 2.44%. [Fact: Exchange and media reports combined] Memory stocks gave back nearly twice what the index did, while small and mid caps rose. This asymmetry is the information of the day.
But the nature of this decline should not be misread. On the prior trading day, July 31, the KOSPI rose 17.91%, its largest single-day gain on record, and Samsung Electronics gained 26.81%, its own largest one-day gain ever. [Fact: Exchange] Compared with the July 29 close, Samsung Electronics is still 14.9% higher even after today’s plunge. Today’s 8.8% is not a new loss; it is a partial giveback of the surge from two days earlier.
The third-quarter DRAM contract price deceleration cited as the cause is not new data. It is a forecast that TrendForce published on July 3 and reaffirmed on July 9. [Fact: TrendForce] What is new today is not the data, but the market’s room to look at it.
Macro conditions created that room. Oil fell from $116 to $79 in late July, a joint yen-buying intervention that included the United States for the first time since 1998 was carried out on July 30 and 31, and the US 10-year yield came down to 4.68%. [Fact: Market data, Japan’s Ministry of Finance] Once the discount rate problem eased, the market immediately moved its question to the numerator, meaning how long the profits last.
Here is what the deceleration actually looks like. Quarterly contract price growth went from 90 to 95% in Q1, to 58 to 63% in Q2, and is now forecast at 13 to 18% for Q3. [Fact: TrendForce] Prices are still rising, but the pace of the rise has slowed for two straight quarters. The cause is not an increase in supply, but that PC and smartphone makers have reached the limit of what they can pay.
The contract structure has actually tilted toward suppliers. SK Hynix removed the price ceiling from its new long-term contracts, Samsung Electronics kept only a price floor and left 30 to 40% of volume for the spot market, and Micron set its price floor above its own highest margin from the previous cycle. [Fact: Reports and industry sources] The common assumption that buyers cap future price increases does not hold, at least for the two Korean suppliers.
The actual 2019 numbers define this cycle’s question. That year, DRAM bit demand grew about 19%, but as average selling prices fell 47.4%, revenue dropped 37.5%. [Fact: Gartner] Volume is not what generates memory profit; price is. That is why the fact that data centers are replacing consumer demand is not grounds for reassurance.
The conclusion is a change of thesis. The contest in this cycle has shifted to whether prices hold at current levels, not whether they rise further. If the market’s implied sustainability rate is 38 to 42%, merely holding steady leaves considerable room for re-rating. But margins in the 90% range have historically triggered capacity expansion, and holding steady will not happen on its own; it is an outcome that supply discipline has to produce.

Key Framing

Prices are still rising, the contract structure has shifted in suppliers’ favor, and inventory is not yet at down-cycle levels. Even so, the reason the market gave back twice as much in memory stocks is not that it doubts the size of the profits, but that it has started recalculating how long those profits will last. The fact that the growth rate has slowed for two consecutive quarters moved the cycle’s clock forward by one notch. But deceleration and decline are different things, and this cycle has a contract structure that did not exist before. The question investors need to answer is not whether prices rise further, but how long they hold at this level.

1. What Actually Happened on August 3

Start with the numbers. The KOSPI fell 338.00 points, or 5.12%, to close at 6,257.45, after slipping as far as 6,223.29 intraday, down 5.64% at the low. Samsung Electronics fell 23,000 won to 239,500 won, down 8.76%, and SK Hynix fell 151,000 won to 1,567,000 won, down 8.79%. On the flow side, retail investors net bought 4.6527tn won, while foreign investors net sold 2.8427tn won and institutions net sold 1.9475tn won. [Fact: Exchange and media reports combined]

Moves on 3 August 2026. The index fell 5.12% while the two memory makers dropped 8.8% and the KOSDAQ rose. The selling concentrated in large-cap memory rather than the whole market.

View as table

Name	Change
KOSDAQ	+2.44%
KOSPI	-5.12%
SOX index	-3.24%
Samsung Electronics	-8.76%
SK Hynix	-8.79%

What deserves attention here is the KOSDAQ. On the same day, the KOSDAQ rose 2.44% to 737.35, its second consecutive daily gain. [Fact: Exchange] That means this was not a day when the whole index collapsed; it was a day when large-cap semiconductor stocks alone were sold heavily. The contrast in the United States points the same way. On July 31, Micron closed down 5.90% at $823.03, while Nvidia rose 2.93% to $200.75 the same day. [Fact: Market data] With one day’s lag, the same pattern showed up in both Korea and the United States. It was memory that sold off, not AI broadly.

But there is a fact that must be considered alongside this decline. On the prior trading day, July 31, the KOSPI rose 17.91% to close at 6,595.45. That is the largest single-day percentage gain on record. Samsung Electronics rose 26.81% to 262,500 won, also its largest one-day gain ever, surpassing the 15% gain of December 2001, and SK Hynix rose to its upper price limit. Foreign investors net bought 7.2tn won that day, also a record. [Fact: Exchange and media reports combined]

Samsung across four sessions. It gave back its gains on results day, posted its largest ever single-day rise the next session, then returned part of it. The 3 August close is still 14.9% above 29 July.

View as table

Date	Close (KRW)	Note
29 July	208,500	before the surge
30 July	about 207,000	results day, about -0.7%
31 July	262,500	+26.81%, its largest ever
3 August	239,500	-8.76%

The picture changes once you line up the sequence. On July 30, Samsung Electronics reported record earnings, yet gave back all of an intraday 4.08% gain and closed down about 0.7%. Media outlets explicitly called this a sell-the-news reaction. [Fact: Media reports] The following trading day it surged 26.81%, and the day after that it gave back 8.76%. Comparing the July 29 close of 208,500 won with today’s close of 239,500 won, the stock is still 14.9% higher across the three days. Today’s plunge is not a new loss; it is a retracement of the largest surge on record. The reasons Korean media cited included profit-taking after the prior day’s surge, a string of executive share-disposal disclosures, and foreign selling in night futures. [Fact: Media reports]

If that is so, what needs explaining in today’s decline is not the retracement itself, but the excess: the fact that memory gave back twice as much as the index.

2. The Deceleration Data Is Not New

What is widely cited as the cause of today’s decline is the deceleration in the third-quarter DRAM contract price outlook. Let’s get the content right first. By TrendForce’s numbers, quarterly contract price growth for commodity DRAM was 90 to 95% in Q1 and 58 to 63% in Q2, with a Q3 forecast of 13 to 18%. Server DRAM (RDIMM) is forecast at the same 13 to 18%, and NAND at 10 to 15%. Only PC DRAM was revised upward, from 8 to 13% to 15 to 20%, on June 30. [Fact: TrendForce] Counterpoint independently forecasts Q3 DRAM at 10 to 20%, a range that overlaps. [Fact: Counterpoint, July 7]

Conventional DRAM contract price growth. Prices rise in all four quarters, but the pace of increase has broken twice. Q3 and Q4 are forecasts, and Q4 is the PC DRAM figure.

View as table

Quarter	QoQ growth	Status
Q1 2026	+90-95%	actual
Q2 2026	+58-63%	actual
Q3 2026	+13-18%	forecast
Q4 2026	+3-8% (PC DRAM)	forecast

But the publication date of this forecast matters. TrendForce put out these numbers on July 3 and reaffirmed them on July 9. [Fact: TrendForce] As of today, that data is a month old. And the deceleration did not start this quarter, either. The drop from 90 to 95% down to 58 to 63% was already one deceleration; the Q3 forecast is the second.

So what is new today is not the data. What is new is the market’s room to look at that data. Throughout July, the variable that dominated the market was the discount rate. The US-Iran conflict pushed oil above $116, the yen was pushed down to the 163 range, its weakest in 40 years, and the US 30-year yield climbed to its highest level since 2007. All three unwound at once in the last week of July. Oil fell for three straight days after the July 27 report of a halt in strikes, coming down to around $79, and on July 30 and 31 a joint yen-buying intervention that included the United States for the first time since 1998 was carried out, while the US 10-year yield settled at 4.68%. [Fact: Japan’s Ministry of Finance, market data]

Once the discount rate problem eases, what remains is the numerator. The market immediately shifted its attention to how long the profits will last, and the material it happened to have on hand was the month-old contract price forecast. That is the structure behind why memory alone fell twice as much today. It is not that new bad news arrived; it is that there was suddenly room to look at old data. [Inference: Own assessment]

This distinction matters in practice. If it were new bad news, the next piece of bad news would be what to wait for. But if it is a change in the question being asked, then every piece of data that comes out from here will be graded against that same question.

3. Why the Deceleration Happened: A Demand-Side Event, Not a Supply-Side One

The fact that the deceleration was not caused by oversupply defines the character of this cycle. TrendForce’s own language is clear. Record-high contract prices mean that customers in consumer markets like PCs and smartphones have hit the limit of what they can pay, and with contract prices already at record highs, slowing consumer demand has pushed price tolerance to its limit. The same report describes supply as still extremely tight. [Fact: TrendForce, July 3]

Traces of that pass-through are already visible in set prices. Dell raised PC prices 15 to 20%, and Lenovo has also flagged a January price increase. Gartner expects DRAM and SSD prices to rise about 130% in 2026, pushing up PC prices by 17% and smartphone prices by 13%. [Fact: Gartner, February 26] Budget Android makers have chosen to downgrade camera and display specifications to protect margins.

Here is one data point that runs counter to conventional wisdom. Third-quarter contract price growth for PC and mobile DRAM, at 15 to 20%, is actually higher than the 13 to 18% for server DRAM. That is the result of tighter consumer-side supply as suppliers shift capacity toward servers and HBM. [Fact: TrendForce] In other words, what has cooled on the consumer side is not the realized price, but the capacity to absorb it. The two-tier structure, where AI demand is strong and consumer demand is cooling, is real, but that structure is showing up in capacity allocation right now, not on the price tag.

One counterexample is worth recording. The chairman of Taiwanese module maker ADATA said in early July that suppliers had notified the company of Q3 price increases of 20 to 30% for DRAM and 35 to 40% for NAND. [Fact: Reports, July 8] The 13 to 18% figure applies to large contracts; smaller buyers with no bargaining power face much steeper increases. The deceleration is an industry-average story, not a story that applies to every transaction.

4. The Contract Structure Has Actually Tilted Toward Suppliers

A widespread interpretation holds that hyperscalers locked in volume through long-term contracts and capped future price increases in the process. Checking this against the record shows the picture differs by supplier, and for the two Korean suppliers, it is the opposite of what is assumed.

Supplier	Price Ceiling	Price Floor	Notes
SK Hynix	Removed in new contracts	Not clearly confirmed	About 10 customers, up to 5 years. During shortages, spot price gains feed into contract prices
Samsung Electronics	None	Capped at a maximum 5% quarterly decline	Only 60 to 70% of capacity allocated to long-term contracts; 30 to 40% held for the spot market
Micron	Retained	Retained	16 customers, cumulative $100bn guaranteed through 2030. Floor set above its own highest margin from the prior cycle

[Fact: TrendForce and media reports combined]

The fact that the three companies made different choices is itself informative. Micron chose to lock in both sides of the price to reduce volatility, while the two Korean suppliers chose to keep the upside open. In particular, Samsung Electronics keeping 30 to 40% of volume for the spot market is a choice not to give up excess profit during the upswing.

That said, buyer bargaining power has not disappeared entirely. TrendForce expects that starting in Q3, the main source of server DRAM price increases will shift to customers without long-term contracts and volume outside existing contracts. [Fact: TrendForce, July 9] Some US cloud providers’ contracts still constrain how much suppliers can raise prices. Combined with the ADATA example above, the picture becomes clear: this is a two-tier structure in which large buyers are protected and small buyers absorb the increases.

Historical context is worth noting too. In the past, the standard for long-term memory contracts was to commit to volume while resetting price every quarter. [Fact: Industry sources] The three-to-five-year contracts of 2026 break that standard, and some even include a structure where 10 to 30% of the contract value is paid upfront. HBM has always been priced on an annual basis set in advance, so it moves independently of commodity prices. As a result, 2026 even saw a reversal in which DDR4 spot prices exceeded HBM3E contract prices. [Fact: Reports]

5. Volume Is Not the Issue; Price Is What Makes the Profit

A common objection needs to be addressed here: even if consumer demand falls, isn’t it fine because data centers are more than making up for it? On volume alone, this objection is correct. Server DRAM already overtook mobile as the largest demand source by bit shipments in 2023 (37.6% versus 36.8%), and HBM consumed 23% of total DRAM wafer input as of April 2026. [Fact: TrendForce] One AI server contains as much DRAM as hundreds of smartphones.

But in memory, what drives the swing in profit is not volume; it is price. 2019 shows this.

Volume and price in 2019. Bit shipments grew by nearly 19%, yet a 47% fall in price cut revenue by 37.5%. Bit shipments are back-solved from the revenue and price changes.

View as table

Measure	2019, year on year
Bit shipments	about +18.8% (back-solved)
Average selling price	-47.4%
DRAM revenue	-37.5%

In 2019, the average DRAM selling price fell 47.4% and revenue fell 37.5%. Working backward from those two figures, bit shipments grew about 19%. [Fact: Gartner, January 2020] Volume grew by nearly 20% while more than a third of revenue evaporated. The fact that Samsung Electronics and SK Hynix both guided to 20 to 21% shipment growth that year supports this back-calculation.

Bit volume grows even in down cycles. The one real exception is 2016, when shipments fell 3% and price fell 16%, shrinking the market 19%. [Fact: Industry compilations] In every other downturn, bits kept growing and profit kept following price. The idea that rising volume makes things fine has never held up in memory industry history.

So what does this shift in demand mix actually mean? Two effects pull in opposite directions. On the side that pushes prices up, hundreds of dispersed consumer buyers have played the role of bidding up marginal prices, and that pressure disappears once they hit their payment limit. What remains is a small number of large buyers, and they sit at the negotiating table. On the side that supports prices, the larger the volume locked into contracts, the narrower the path by which a spot price crash spreads into earnings, and because data center demand keeps replacing and expanding equipment even in down cycles, a 2019-style demand cliff becomes shallower.

Combine the two forces and they point in one direction: peak margins get lower and trough margins get higher. That is the judgment on where this mix shift ultimately lands, a third answer that is neither collapse nor permanence. [Inference: Own assessment] The memory subscription service concept that SK Hynix Chairman Chey Tae-won raised at the Nasdaq listing event on July 10 targets the same problem. The idea is to remove the cycle from the business structure rather than try to predict it. He himself stated it is still just an idea at this stage. [Fact: Reports, July 10]

6. What Is the Same and What Is Different From 2017-18

The precedent invoked most often for a deceleration phase is the previous major cycle. Contract price growth started slowing after the fourth quarter of 2017. In the first quarter of 2018, the month-over-month increase in PC DRAM slowed to 5%, and by the third quarter it had come down to 1 to 2%, marking the end of nine consecutive quarters of increases. Prices did not actually start falling until the fourth quarter of 2018, and within that quarter, contract prices were even cut twice. [Fact: TrendForce]

Stock prices moved earlier than that. Samsung Electronics hit its first peak on November 1, 2017, and SK Hynix on October 11, 2017; SK Hynix then corrected 27%, rebounded, and set its final peak on May 23, 2018. Both companies’ operating profit peaked in the third quarter of 2018. [Fact: Reports and disclosures] The commonly cited summary that “stock prices led by six months” is not accurate. Measured from the first peak, the lead was 9 to 11 months; measured from the final peak, it was 2 to 4 months. Stock prices at the time formed a double top, not a single peak. Micron shows the scale of the decline: it fell 56%, from $64 in May 2018 to $28 in December, consistent with the typical 50 to 60% peak-to-trough range for memory cycles. [Fact: Market data]

But this cycle has three things that did not exist back then. One is the contract structure discussed above: more than half of capacity is locked into multi-year contracts, with upfront payments involved. Another is the starting valuation. In 2018, the downturn began from peak multiples, but the sustainability rate the market is currently implying is 38 to 42%, a level that already prices in only a replacement-demand floor. [Inference: Own back-calculation] The last is the absolute level of margins. SK Hynix’s DRAM gross margin is estimated at 90.9%, and Micron’s contract floor is set above its own highest margin from the prior cycle. [Fact: Bernstein estimate, industry sources]

But that last item is both a shield and a risk. Margins in the 90% range have historically been the level that invites capacity expansion.

7. Holding Steady Will Not Happen on Its Own

Translating the discussion above into an investment judgment, the conclusion is a change of thesis. The contest in this cycle is not whether prices rise further, but whether they hold at current levels. An implied sustainability rate of 38 to 42% means the current stock price already assumes that only around 40% of peak profit remains. If that is the case, then even if prices simply hold without rising further, that assumption turns out to be significantly too conservative, opening room for re-rating. Holding steady is not a passive scenario; it is a bullish scenario relative to the current stock price. [Inference: Own back-calculation, sensitive to the sustainability rate assumption]

Two caveats attach to this, though. Commodity prices dislike sitting on a plateau and tend to trend instead. And current margins are at a level that invites the strongest possible supply response. There is precedent. In the PC cycle from 1993 to 1996, gross margins above 50% drew roughly 50 new fab construction announcements in 1995 and 1996 alone, and that glut ended the cycle. [Fact: Industry analysis] The expansions announced so far are Samsung’s Pyeongtaek P5 (construction complete first half of 2027, production in the second half of 2028), SK Hynix’s Cheongju M15X (mid-2027) and Yongin phase 1 (first quarter of 2027), Micron’s Idaho fab (second half of 2027) and New York fab (after 2029), and CXMT’s roadmap for 500,000-plus wafers per month by 2028. [Fact: Company disclosures, reports] Since a fab takes 3 to 5 years from groundbreaking to production, this capacity will arrive all at once around 2028.

The inventory asymmetry. Customer inventory has reached the warning line, but supplier inventory is still lean and far from the levels that led past downturns. Bars show the midpoint of each range.

View as table

Measure	Level	Reading
Supplier inventory	3-5 weeks	flat since February, still tight
Customer inventory	7-9 weeks	up from 2-4 weeks at the start of the year
Usual warning line	8 weeks	industry convention
Pre-downturn level	10-15 weeks	not reached

So “holding steady would be enough” is true, but “it will hold steady” is a separate claim. The former is arithmetic; the latter is an outcome that supply discipline has to produce. And whether contracts can substitute for that discipline is this cycle’s experiment. If a contract guarantees a price floor, holding steady becomes a condition rather than a hope; if it only guarantees volume, price remains a hostage to supply and demand. As the table above showed, Samsung Electronics and Micron currently have floors, while SK Hynix’s floor is unconfirmed.

8. The Second Force on the Same Day: The Yen

Beyond memory-specific factors, there was also a currency backdrop to the August 3 decline. But the dates need to be precise here too.

The joint intervention carried out on July 30 and 31 was unusual in both scale and character. Japan’s Ministry of Finance officially confirmed on August 3 that it had bought yen jointly with the US Treasury, marking the first direct US participation in such an intervention in 28 years, since 1998. Reuters reported that Japan and Korea carried out the first publicly confirmed joint intervention. The estimated scale was about $53 to 59bn on July 30 and about $34bn on July 31. [Fact: Japan’s Ministry of Finance, Reuters, Bloomberg] The US Treasury Secretary said he would not hesitate to pursue further joint intervention.

As a result, USD/JPY fell from 163.75 on July 27 to 159.54 on July 30, a 2.36% move in a single day, and reached 156.80 on August 3. In other words, the sharp single-day yen appreciation of over 2% happened on July 30, and the move on August 3 was only 0.50%. [Fact: Market data] Explaining today’s decline by a yen surge does not match the timeline. That said, the pace of appreciation itself, about 4.2% from 163.75 to 156.80 over the week following the intervention, is a condition that increases pressure toward carry trade unwinding.

The Korean won moved in the same direction. It fell from 1,465.07 on July 27 to 1,423.92 on July 30, and closed at 1,428.73 on August 3. [Fact: Market data] Won strength is a necessary condition for foreign investor buying. Currency stabilization lies behind the shift from foreign investors net selling more than 12tn won between July 1 and 16 to net buying 7.2tn won in a single day on July 31.

Many observers compare this to August 5, 2024. Back then, the yen appreciated 6.15% in a week, the Nikkei fell more than 12% in a single day, and the KOSPI dropped more than 8% as well. Recovery took one to two weeks. [Fact: Market data] Current positioning looks lighter than it was then. Net short yen positions in Chicago futures markets stood at 152,000 contracts as of July 24, below the roughly 180,000 contracts seen just before August 2024. [Fact: Chicago Mercantile Exchange, media reports] Korean brokerages have also suggested that, unlike two years ago, the carry funds available to unwind have already largely dried up. That said, there are dissenting views, and the actual peak in positioning right before the intervention cannot be confirmed because of reporting lags.

One thing needs correcting here. The common belief that joint intervention goes hand in hand with a recovery in risk assets is not historically well established. The intervention that included the United States in June 1998 did not produce an immediate effect, and the yen kept weakening through that August. [Fact: Market historical records] It is true that the KOSPI posted its largest-ever gain right after this intervention, but easing tension from Iran and Microsoft’s earnings coincided in the same period, so the effect cannot be isolated to the intervention alone.

9. What to Watch, and What to Stop Watching

Contract price growth rates now carry less informational value. The deceleration has already become an established fact, and the forecast for the next quarter is already out. Here is a prioritized list of what to watch instead.

Indicator	Current Value	Reading
Customer-side channel inventory	7 to 9 weeks (2 to 4 weeks at start of year)	Approaching the 8-week warning line. Top-priority condition for a down-cycle transition
Supplier-side inventory	3 to 5 weeks, flat since February	Far from the 10 to 15 week level seen ahead of past downturns. This asymmetry is the key point
DRAM spot price	New high of $41.10 on July 21	Has not gone flat or fallen for two straight weeks since April. Has not turned yet
Q4 contract price outlook	PC DRAM at 3 to 8% (revised up from 0 to 5% at end of June)	Now in single digits. Server and NAND figures not yet available
Contract price terms	Samsung has a floor; SK removed its ceiling	Whether a floor exists is the contractual basis for the holding-steady scenario
Upfront payments (contract liabilities)	Absolute amounts undisclosed by either company	Mid-August quarterly reports will be the first confirmation window

[Fact: TrendForce and industry compilations, Blocked: Quarterly reports not yet filed]

It is true that customer-side inventory is approaching the warning line, but supplier inventory remains thin and spot prices keep hitting new highs. For deceleration to turn into a decline, all three of these need to point in the same direction, and right now only one of them does. Slowing and declining are different things, and the moment that distinction breaks down is the moment the verdict changes.

Finally, here are the conditions that would overturn this piece’s verdict. If supplier inventory rises into double-digit weeks, if spot prices fall for two or more consecutive weeks, or if the Q4 server DRAM contract price outlook comes in flat or lower, then this piece’s judgment that this is “deceleration, not decline” is void as of that point.

Stocks mentioned in this piece are illustrative examples for analysis and are not a recommendation to buy or sell any particular security. Responsibility for investment decisions and their outcomes rests with the investor. Contract price figures are forecasts from TrendForce and Counterpoint and may differ from finalized monthly figures, and the third- and fourth-quarter items are estimates that have not yet materialized. The implied sustainability rate of 38 to 42% is an in-house back-calculation, is sensitive to its assumptions, and does not allow for a definitive statement about the size of any re-rating room. Channel inventory and spot price figures are industry compilations and vary by data provider. SK Hynix’s DRAM gross margin of 90.9% is a brokerage estimate, not a disclosed figure. The pricing terms of long-term contracts are drawn from a combination of reports and industry sources; since the original contracts are not public, the details may differ by supplier, and neither company has disclosed the absolute amount of its contract liabilities. The attribution of causes behind today’s plunge and the yen’s contribution to it are qualitative judgments, not a quantitative decomposition. Prices and indicators are as of the closing prices on August 3, 2026.

Korea-Semiconductor on Korea Invest Insights

Anatomy of the Rebound: Foreign Buying Came in Two Bursts, and the Same Two Stocks Changed Hands

TL;DR

1. The Surface of Five Days: Retail Sold, Foreigners Bought

2. Split Into Phases: Buying That Came in Two Bursts

3. Buying That Returned to the Same Spot

4. The Caveat: Cash Was Bought, the Hedge Remains

5. The Variables for the Next Trading Day

Related posts

AI's Bottleneck Really Is HBM, Not the GPU: But Bottlenecks Invite Workarounds

TL;DR

1. Where the Claim and Its Numbers Come From

2. Why HBM, Not the GPU: The Gap in Growth Rates

3. What Agents Change: Resident Memory

4. Why Only Three Suppliers: The Arithmetic of Supply

5. The Workarounds the Bottleneck Invites

6. Where the Bottleneck’s Rent Goes

7. The Implication for Korea: 2027’s Physics, 2028’s Calendar

Related Posts

A Rebound Won't Come From a Single Trigger: A Verdict Calendar From Late August to Year-End

TL;DR

1. What Last Week’s Three Days Showed

2. Impact and Timing Across Five Axes

3. Macro Has Already Fired Once

4. NVIDIA Earnings: The Beat Is Already Priced In

5. Shareholder Returns: High Impact, Now With a Condition Attached

6. Export Data: Low Impact, Highest Frequency

7. Stacking and the Calendar

8. The Earliest Signal Is Not a Trigger

Related Posts

What Hyperscaler Earnings Proved, and What They Didn't: Three Explanations for the Memory Sell-Off

TL;DR

1. What Was Actually Proven

2. How Far to Trust the Backlog

3. So Why Did Memory Get Sold?

4. Three Explanations, Scored

5. NVIDIA Is Cutting Back on Memory

6. Rubin’s Timeline

7. Verdict and Watch List

Related Posts

Memory No Longer Needs Prices to Rise Further: Anatomy of the August 3 Plunge and the Contract Price Deceleration

TL;DR

1. What Actually Happened on August 3

2. The Deceleration Data Is Not New

3. Why the Deceleration Happened: A Demand-Side Event, Not a Supply-Side One

4. The Contract Structure Has Actually Tilted Toward Suppliers

5. Volume Is Not the Issue; Price Is What Makes the Profit

6. What Is the Same and What Is Different From 2017-18

7. Holding Steady Will Not Happen on Its Own

8. The Second Force on the Same Day: The Yen

9. What to Watch, and What to Stop Watching

Related Posts