Server DRAM on Korea Invest Insights

Alphabet's Q2: Cloud +82% Ends the Demand Debate, Negative FCF Starts the Cash Debate

Thu, 23 Jul 2026 11:00:00 +0900

Context: Who Burns All Those Tokens? argued that final demand, the last weak link in the AI CAPEX debate, had started to get real numbers attached to it, and left the verdict to the earnings season around July 30. The first scorecard arrived earlier than scheduled. Alphabet reported Q2 results in the early hours of July 23 Korea time. This piece synthesizes two analysis notes that dissected the same print from different angles into one. The short version: the demand side of the debate is effectively over, and the debate has moved on to cash flow and return on capital.

TL;DR

Cloud revenue came in at $24.8bn, +82% (vs. consensus +64%), marking five straight quarters of accelerating growth: 32%, 34%, 48%, 63%, 82%. Backlog rose from $462bn to $514bn, a net increase of $52bn even as revenue recognition accelerated.
Demand quality matters more. Existing Cloud customers are consuming on average 50%+ above their initial commitments, and Gemini model API throughput hit 22bn tokens per minute, up 37.5% in a single quarter. With supply running short, Alphabet started, from June, paying SpaceX roughly $920M a month for AI compute. It is an unprecedented sight: a hyperscaler turning into a net buyer of compute.
On the flip side of the same print, quarterly FCF turned negative for the first time ever, at -$5.9bn. CAPEX of $44.9bn outran operating cash flow of $39.1bn, and 2026 CAPEX guidance was raised again from $180-190bn to $195-205bn, with a large further increase flagged for 2027.
The judgment splits three ways. The odds of a post-2028 AI demand cliff have fallen further. Alphabet’s FCF turnaround has been pushed out from 2027 into the 2028-2029 window. For memory, volume demand strength was reconfirmed, but the durability of pricing and peak margins remains a separate question.
We hold the 45/35/20 demand-scenario probabilities. But another strand of hard evidence has piled up in support of the 35% upside case, and the early-CAPEX-deceleration premise central to the 20% downside case was weakened by this print. The next scorecards are Microsoft (July 30 KST) and Amazon (July 31).

Key Framing

Cloud +82% and the first-ever negative quarterly FCF are two sides of the same print. Asked whether the demand is real, Alphabet answered with usage that exceeds commitments and with the act of paying a premium to rent someone else’s servers. What arrived sooner than expected, in exchange, was the bill for serving that demand. The market’s scoring criterion has now shifted from whether demand exists to the cash return after depreciation, and for memory investors this print is reinforcement for volume, not a warranty on margin.

1. The Numbers First: What Came Out

Item	Q2 2026 Actual	Comparison Basis	Verdict
Total revenue	$119.8bn, +24%	Consensus ~$116.9bn	Beat
Google Cloud revenue	$24.8bn, +82%	Consensus $22.4bn, +64%	Large beat
Cloud operating income	$8.8bn (35.6% margin)	Prior year $2.8bn (20.7% margin)	Improved
Search revenue	$63.3bn, +17%	Consensus $63.4bn	In line
Total operating income	$40.8bn, +30% (34.0% margin)	Margin below consensus	Mixed
Adjusted EPS	~$2.85	Consensus ~$2.89	Slight miss
Operating cash flow	$39.1bn	CAPEX $44.9bn	Reversed
Quarterly FCF	-$5.9bn	Prior quarter $10.1bn	First negative ever
Cloud backlog	$514bn	Prior quarter $462bn	+$52bn
2026 CAPEX guidance	$195-205bn	Prior $180-190bn	Raised again

[Fact: Alphabet disclosures and earnings call]

The GAAP numbers carry an optical illusion. Reported GAAP EPS was $9.11 and net income was $112.1bn, up 298% year over year, but that includes roughly $99bn (pre-tax; about $6.26 per share after tax) of unrealized gains on private stakes such as Anthropic and SpaceX. Strip out this operationally unrelated item and the underlying EPS was about $2.85, a slight miss versus market expectations. Revenue and Cloud won big, but the quality of earnings was not as strong as the headline number suggests. [Fact: recalculated from disclosures]

2. The Demand Debate: Why We Think It’s Effectively Over

The previous piece flagged final demand as the weak link and cited Codex’s three million in three days as the first hard data point. Alphabet’s print adds four more strands of hard evidence on top of that.

Usage is running ahead of contracts. Gemini model API throughput rose from 16bn to 22bn tokens per minute, up 37.5% in one quarter, and monthly developers passed 9M. Existing Cloud customers are consuming on average 50%+ above their initial commitments, and the overage widened from the prior quarter. New-customer acquisition is running at twice last year’s pace, and Marketplace transaction volume is up sevenfold. [Fact: CEO letter and earnings call] The hypothesis that demand rests solely on a handful of long-term frontier-lab contracts is not compatible with data showing post-contract actual consumption exceeding commitments.

The demand base has broadened. Alongside model-development demand such as large-scale Gemini 4 pretraining, the company laid out five channels at once: enterprise inference in finance, pharma, retail, telecom and manufacturing; BigQuery and security workloads; consumer services such as Search’s AI Mode and the Gemini app; and outright sales of TPU systems supplied directly into customer data centers. About 90% of the Fortune 100 now use Gemini Enterprise, more than 500 Cloud customers process over 1tn tokens a year, and over 2,000 exceed 100bn tokens. [Fact: CEO letter]

Supply is still capping growth. The CFO stated flatly that the company remains in a supply-constrained environment. To fill near-term shortfalls, Alphabet is leasing third-party data center capacity at elevated prices, and management said roughly six months of inefficiency is worth accepting to win large customers. That is the backdrop for the biggest news of the call: from June, Alphabet has been paying SpaceX roughly $920M a month, about $11bn annualized, to rent AI compute. [Fact: earnings call and press] That the most calculating company in the world is paying a premium to rent someone else’s servers is behavioral evidence that runs directly counter to the demand-overstatement hypothesis.

Backlog grew even while being recognized. Cloud revenue rose about 23.8% quarter over quarter, meaning large contracts are converting into revenue quickly, and yet the backlog balance still grew by another $52bn. The current backlog is about 5.2x annualized Q2 Cloud revenue, and the company said more than half will be recognized within 24 months. [Fact: disclosures and earnings call] As visibility through 2027, that is strong evidence. Customer concentration and the year-by-year recognition amounts beyond 2028, however, were not disclosed. [Blocked: backlog detail undisclosed]

The efficiency counter-argument also lost force this quarter. AI Mode’s cost per response has fallen to its lowest level since launch, yet users and query volume grew even faster. AI Mode has 1bn monthly users, and the Gemini app has 950M MAU. The relationship where usage growth outruns the decline in per-unit cost has now been confirmed at Alphabet’s scale too, running in the same direction as the Jevons pattern observed with Codex. [Inference: efficiency-usage relationship]

3. The Cash Debate: The Bill Arrived Sooner Than Expected

On the other side of the demand story, what this print confirmed is the scale and duration of spending.

Q2 CAPEX of $44.9bn, double the year-ago level, exceeded operating cash flow of $39.1bn, and quarterly capital intensity (CAPEX/OCF) came in at about 115%. The arithmetic is heavy too. With H1 CAPEX at roughly $80.6bn, hitting the top of the annual guidance range of $205bn requires an H2 quarterly average of $57.2-62.2bn, 27-38% above Q2. A meaningful FCF recovery in H2 2026 or in 2027 is unlikely. [Inference: own arithmetic]

There was a hint on 2027 as well. The CFO maintained guidance that 2027 CAPEX will rise sharply, and the market consensus sits at about $257bn. The path from $205bn to $257bn works out to roughly +25% growth. That places the company on a deceleration path, from about +76% this year to +25% next year, not yet a re-acceleration scenario toward the $300bn level. [Fact: earnings call and consensus] This distinction matters. On a deceleration path, the skeleton of the FCF-turnaround thesis stays intact; on a re-acceleration path, the skeleton itself collapses.

Running a 2028 sensitivity in advance sharpens the judgment criteria. The starting point is the current trailing-12-month operating cash flow of $185.7bn.

2028 CAPEX Assumption	OCF Needed for FCF = $0	Required OCF CAGR	OCF Needed for FCF = $50bn	Required Growth
$200bn	$200bn	~3.8%	$250bn	~16.0%
$230bn	$230bn	~11.3%	$280bn	~22.8%
$250bn	$250bn	~16.0%	$300bn	~27.1%

[Inference: own sensitivity, not company guidance]

If CAPEX stabilizes around $200bn in 2028, an FCF recovery is not difficult. If it keeps rising to $230-250bn, combined Cloud-and-Search operating cash flow needs to grow more than 20% a year to get back to prior FCF levels. That is not an impossible number if Cloud’s +82% growth and 35.6% margin hold up for a while, but depreciation, power costs, external lease payments, and a sales mix increasingly weighted toward lower-margin TPU hardware all push in the opposite direction.

A shift in the capital-allocation regime was also formalized this quarter. With $242.5bn in cash and marketable securities and $185.7bn in trailing-12-month operating cash flow, Alphabet’s ability to pay is not in question. But in Q2 it raised about $30.5bn in common equity and about $19.1bn in convertible preferred stock and added about $21.1bn in net debt, while buying back no stock at all. Debt rose from about $16bn to about $100bn in twelve months. [Fact: disclosures] A company that once covered both CAPEX and buybacks entirely from operating cash flow has become one that has paused buybacks for the sake of CAPEX and is now tapping debt and equity capital as well. This is not a liquidity-risk signal but a signal that the capital-allocation regime has changed, and the fact that the weight of the funding sits more with shareholders than with the credit market is the comforting part from a credit perspective.

4. Scoring the Three Questions

Scoring the three questions against this print comes out as follows.

Does AI demand hold up beyond 2028? The odds of a cliff have fallen further. Demand is broadening from training into inference, data, security and enterprise workflows; actual consumption is exceeding commitments; usage growth is outrunning efficiency gains; most TPU external-sales revenue is scheduled to be recognized in 2027; and supply is still short enough that CAPEX must rise sharply again in 2027. The more accurate way to read 2028 is not as the point where demand ends, but as the point where new supply capacity comes fully online and goes on trial.

Will Alphabet’s FCF turn around? Yes, but the timing has slipped. The quarterly FCF crossover arrived sooner than originally expected (early 2027), but even an optimistic 2027 operating-cash-flow estimate of $230-240bn is still FCF-negative for the full year against $257bn of CAPEX. A return to positive FCF requires 2028, and even then only if 2028 CAPEX growth slows to a single digit. A strong FCF re-acceleration is a conditional scenario confined to the 2028-2029 window.

Does this translate into sustained memory purchasing? Strongly positive for volume, neutral for price and margin. We unpack this in the next section.

5. Translating to Memory: Reinforcement for Volume, Not a Warranty on Margin

In Q2, about 60% of technical infrastructure CAPEX went to servers and 40% to data centers and networking. Server spend bundles together TPUs, GPUs, CPUs, HBM, server DRAM, SSDs and networking silicon. [Fact: earnings call]

The items supporting HBM demand are clear-cut: the new TPU generation and the rollout of NVIDIA Vera Rubin, large-scale Gemini 4 pretraining, growing enterprise inference volume, the next-generation network fabric tying together a 1M-accelerator network, and TPU system sales to outside customers, most of which will be recognized starting in 2027. What matters is that even where in-house TPUs partly displace NVIDIA GPUs, HBM demand does not disappear, it simply shifts purchasing channels from the GPU supply chain to the TPU-system supply chain. Add third-party compute leasing on top of that, and HBM and server DRAM demand now flows through three channels: Alphabet’s own data centers, TPU external sales, and leased compute. [Inference: demand-channel decomposition]

The case for server DRAM and eSSD also got thicker with this print. AI agents don’t run on accelerators alone; CPU servers handle preprocessing, state management, security and orchestration. Alphabet highlighted its Axion CPU alongside growing data and security workloads. The figure of 500 Cloud customers processing over 1tn tokens a year shows that storage demand flowing into logs, checkpoints, search indices and vector databases is not confined to training. [Fact: CEO letter] This confirms, via big-tech earnings, the demand-side backdrop for the server DRAM re-tightening covered in the previous piece (Q3 contract prices forecast +13-18%, lead times of 40 weeks).

Even so, this print alone is not grounds for raising memory companies’ 2028 earnings estimates. Just as equipment purchases are surging, memory supply capacity is also set to grow over 2027-2029. Broader in-house chip adoption weakens NVIDIA’s pricing power at the system level, inference cost per response keeps falling fast, and once Alphabet starts tightening CAPEX efficiency, downward price pressure on components will flow down the supply chain. The high prices of HBM4 and next-generation HBM embed early-generation scarcity and a yield premium that are not permanent values. Putting it together looks like this.

Item	Pre-Earnings View	Post-Earnings View
AI accelerator / HBM bit demand	High	Very high
Server DRAM / eSSD demand	Medium-high	High
Advanced packaging / networking	High	Very high
2028 demand cliff	Low probability	Lower probability
Memory ASP durability	Uncertain	Uncertain
Memory peak-margin durability	Low	Low
Early big-tech FCF recovery	Medium	Lower
Incremental return on CAPEX	Unproven	Partially proven, verification ongoing

[Inference: overall judgment]

The 2028 risk for semiconductors is not order cancellations but the normalization of ASP and margin. That is consistent with the 45/35/20 demand-scenario probabilities. This print added one more piece of hard evidence to the 35% upside case and weakened the early-big-tech-CAPEX-deceleration premise behind the 20% downside case. We will update the probabilities themselves once Microsoft and Amazon are also confirmed. From the perspective of Samsung Electronics and SK Hynix, this is a print that improves the demand backdrop heading into the 2027-vintage HBM pricing negotiations that begin in Q4.

6. The Stock Reaction, and Where the Two Notes Diverge

The market reaction summarizes the character of this moment. In after-hours trading, the stock opened down about 2%, briefly recovered to a modest gain, then reversed again once CAPEX guidance was raised on the call, with the decline widening to a 3-5% range. As of the morning of the 23rd Korea time, the share price sat around $342, about 1.5% below the prior close of roughly $347. [Fact: market data] In a market where beating estimates has become the baseline, the 2026 rule that CAPEX decides the reaction played out for a fifth time. Read the other way, the fact that the decline stayed this contained in the face of +82% growth is also a signal that the market will digest the spending as long as the growth keeps proving itself.

The two notes synthesized here read the same facts and reached different conclusions. One applied a structural-proof standard and chose to stay on hold. The accurate phrasing is not that return on capital has been proven but that it is now most likely to be provable, and it withheld a price target given negative FCF, the capital raises, and the flagged 2027 increase. The other maintained a buy on an entry-price basis. Taking a price already 15% off its peak as the starting point, it applied a 28-29x multiple to about $15 of ex-gains 2027E EPS to arrive at a 12-month target of $430, with a withdrawal condition attached: it will be recalculated if 2027 CAPEX is quantified near $300bn. [Fact: the two notes’ conclusions]

The difference between the two judgments is not a difference in facts observed but a difference in time horizon and standard. The structural view puts the burden of proof on the company until the cash return after depreciation is confirmed in the numbers; the price view calculates how much of that risk is already reflected in a stock that has already corrected. In keeping with this blog’s principle of not recommending a specific direction, we present both frames side by side, but the common denominator is clear. Either way, the next variable to be judged is not demand but cash.

7. Scorecard: The Next Two Weeks Will Decide

Here is a rundown of the boxes this print filled in, and the boxes it newly opened.

Alphabet’s share of the big-tech CAPEX commentary is now confirmed as a raise. What remains is Microsoft’s first FY27 guidance (call in the early hours of July 30 KST) and Amazon’s AWS growth rate (early hours of July 31). If Alphabet closed out the demand debate, these two will settle the margin and FCF debate.
Whether Cloud growth holds above 50% for the next 2-4 quarters and whether margin holds above 30% are the confirming indicators for the upside case. If growth falls below 30% once supply constraints ease, the possibility of demand overestimation needs to be recalculated.
Watch whether backlog keeps net-adding even after revenue recognition. A combination of a shrinking backlog, or one where only contract duration keeps extending, would signal deteriorating contract quality.
The quarter in which trailing-12-month operating-cash-flow growth overtakes CAPEX growth is the starting point of the FCF turnaround. Conversely, if 2028 CAPEX exceeds $250bn while operating cash flow falls short of that pace, we will call it structural FCF impairment.
Whether large additional equity or debt raises continue, whether SpaceX-style external leasing shrinks back in 2027-2028, and whether TPU external sales convert into recurring Cloud consumption are the checkpoints on the capital-efficiency side.
The discriminator on the memory side is unchanged: DRAM contract pricing, and confirmation of Q3 contract prices on the Samsung Electronics and SK Hynix calls around July 30.

Names 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. Alphabet did not provide 2028 CAPEX or Cloud revenue guidance; the customer mix, year-by-year recognition amounts and cancellation terms behind the $514bn backlog were not disclosed; and purchase volumes and supplier mix for HBM, DRAM and NAND are also not disclosed items. The 2028 FCF sensitivity and the H2 CAPEX arithmetic are our own estimates starting from current disclosures, not company guidance. The pre-tax size of the unrealized gain on private-equity stakes and its after-tax per-share effect are recalculations based on disclosures and may differ slightly due to rounding. The $430 target is the calculation of one of the notes synthesized here, not this blog’s target price. Share prices and quotes are as of the morning of July 23, 2026 Korea time.

Who Burns All Those Tokens? NVIDIA's Customer Map, Sovereign AI and Codex at 9 Million Start Answering

Sun, 19 Jul 2026 19:00:00 +0900

Context This post builds on Will AI Memory Demand Exceed Expectations?, which probability-weighted demand at a base case of 45%, a beat case of 35% and a miss case of 20%, and on SK Hynix Chairman Chey Tae-won’s Two Months of Remarks, which read the words and moves of the supply side’s top executive. The weak link still left is final demand. With this much GPU and HBM being deployed, who exactly is going to burn all those tokens? This piece records the week in which numbers began attaching to that question.

TL;DR

In NVIDIA’s data center revenue, the hyperscale share has held around 50% for seven straight quarters. The share itself has not already collapsed. What has changed is the composition. In the February-April 2026 quarter, non-hyperscale (ACIE) revenue of $37bn reached near-parity with hyperscale’s $38bn, and the quarterly growth rate had already flipped, 31% QoQ versus 12% QoQ. The next disclosure could be the one where the share itself flips for the first time.
Over the same stretch, data center revenue’s year-over-year growth rate re-accelerated from +56% to +66%, +75% and +92%. The share holding steady while the total sped back up means incremental growth is being pulled by everything outside hyperscale. NVIDIA itself wrote in its disclosures that growth was led by the rest of its customer base.
Sovereign AI revenue topped $30bn for full-year FY2026, more than tripling from the prior year, and it grew more than 80% year-over-year in the most recent quarter as well. This is the buyer that the market missed while it kept watching only what the hyperscalers were saying.
Meritz Securities’ semiconductor team flagged, in a Sunday report, the start of full-scale sovereign purchasing and a renewed tightening in server DRAM supply since mid-July. TrendForce’s forecast of a +13-18% rise in third-quarter server DRAM contract prices, Taiwan’s Inventec testifying to lead times of over 40 weeks, and Micron’s statement that it can meet only 50-66% of key-customer demand all cross-verify this externally.
Numbers have appeared on the final-demand side too. OpenAI’s Codex went from 1 million in February to over 5 million by late May, then from 6 million on July 12, right after the GPT-5.6 launch, to about 9 million around July 15, adding 3 million users in three days. This is the stretch where the unit of demand shifts from subscriber counts to agent execution time.
The conclusion is not a premature upgrade of the probabilities. It is that, before litigating oversupply, this is the moment to re-examine whether demand was drawn too conservatively, and that the verdict will be delivered by the earnings season around July 30 and NVIDIA’s disclosure at the end of August.

Key Framing

The market has been doubting the sustainability of AI CAPEX while watching only what four Big Tech companies were saying. Meanwhile, on NVIDIA’s own books, non-hyperscale revenue reached parity with hyperscale, sovereign revenue tripled in a single year, and Codex added 3 million users in three days. While the stock price priced in the fear of oversupply first, the base of demand was quietly stacking up hard evidence pointing the other way. This asymmetry is the essence of the current correction.

1. What Kind of Week Was This: Between a Limit-Down and a Report

Let’s set the stage first. On Thursday, July 16, in the US market, the Philadelphia Semiconductor Index (SOX) plunged around 4% and entered bear-market territory, down more than 20% from its high, while the Nasdaq fell 1.47%. On Friday the 17th, the Nasdaq fell a further 1.40%, bringing the weekly decline to 2.90%. [Fact: market data] The same day in Tokyo, Kioxia hit limit-down right after the open, plunging 16.1%. The stock had fallen to less than half its June 22 peak, and roughly JPY 30tn in market cap had been erased. The direct trigger was a roughly $229 million patent-infringement damages verdict that a US jury awarded to Viasat, but the fact that TSMC fell 7.29%, SK Hynix’s ADR fell 13.69% and SanDisk fell 12.63% in the same session shows the real story was deleveraging across the entire AI rally. [Fact: Nikkei and Hankyung reports]

The Korean market sidestepped that Friday, because July 17 was a market holiday for Constitution Day, redesignated as a holiday for the first time in 18 years. But just before that, it had already gone through steep declines of its own: SK Hynix -15.37% and Samsung Electronics -10.70% on July 13, then SK Hynix -12.34% and Samsung Electronics -9.47% on July 16. [Fact: exchange data] The narrative the market is pricing in is clear. Supply is increasing, China is catching up, 2027 earnings are being cut, and nobody knows when Big Tech’s CAPEX will turn down.

Against this backdrop, on Sunday, Meritz Securities’ semiconductor team, analysts Kim Sun-woo, Yang Seung-su and Kim Dong-kwan, published a set of reports together. The timing was clearly conscious of Friday’s sell-off and the Kioxia episode. The thrust runs the other way. Memory demand right now is not just hyperscalers, sovereign-camp purchasing including the Middle East is moving into full swing, and server DRAM supply has been tightening again since mid-July. Their rebuttal is that the view treating 2028 shortage relief as a foregone conclusion has not priced in this demand. [Fact: Meritz Securities report summaries, 2026-07-19]

Rather than leave this as a war of words, we break the claim into three verifiable pieces and check each against hard evidence one at a time: NVIDIA’s customer mix, sovereign purchasing, and final usage.

Let’s start by precisely checking the claim that the hyperscaler share is shrinking. Laying NVIDIA’s quarterly disclosure language out as a time series looks like this.

Fiscal Quarter	Period	DC Revenue	YoY	Large CSP/Hyperscale Share (disclosure language)
FY25 Q3	Aug-Oct 2024	$30.77bn	+112%	~50%
FY25 Q4	Nov 2024-Jan 2025	$35.58bn	+93%	~50%
FY26 Q1	Feb-Apr 2025	$39.11bn	+73%	Just under 50%
FY26 Q2	May-Jul 2025	$41.10bn	+56%	~50%
FY26 Q3	Aug-Oct 2025	$51.22bn	+66%	Not disclosed
FY26 Q4	Nov 2025-Jan 2026	$62.31bn	+75%	Slightly over 50%, growth led by the rest of the customer base
FY27 Q1	Feb-Apr 2026	$75.2bn	+92%	Hyperscale $38bn (~50%) vs. ACIE $37bn

[Fact: NVIDIA CFO commentary and earnings calls]

Two things are visible at once. First, the share figure itself has not left the 50% band for seven straight quarters. The narrative that hyperscalers have already been pushed aside is not supported by the hard data. Second, even so, the direction of the mix has clearly shifted. The FY26 Q4 disclosure explicitly stated that growth was led by the rest of the customer base and that revenue had diversified, and starting in FY27 Q1, NVIDIA began disclosing data center revenue split between hyperscale and ACIE (AI cloud, industrial, enterprise). In that first quarter, ACIE came in at $37bn, nearly matching hyperscale’s $38bn, and the quarterly growth rate flipped to ACIE +31% QoQ versus hyperscale +12% QoQ. AI cloud revenue within ACIE was up more than threefold year over year. CEO Jensen Huang stated flatly on the call that over the long run, the second category would grow faster. [Fact: earnings call]

Overlay the growth-rate re-acceleration on this and the picture is complete. Data center revenue’s year-over-year growth rate bottomed at +56% in FY26 Q2 and climbed back up through +66%, +75% and +92%. The share holding steady while total growth sped up means, arithmetically, that more than half of the incremental growth is being generated outside hyperscale. [Inference: disclosure arithmetic] If this trend holds for even one more quarter, the FY27 Q2 disclosure at the end of August will print a number where ACIE overtakes hyperscale for the first time. From that moment on, the very frame of judging AI demand’s sustainability by watching only four Big Tech companies’ CAPEX guidance becomes outdated.

One point of confusion is worth clearing up. Direct-customer concentration in the 10-Q (Customer A at 22%, Customer B at 14%) has actually risen, but that is a distribution-stage metric that aggregates board partners, OEMs and large direct purchases, a different layer from the diversification of final demand. [Fact: 10-K] Distribution concentrating and final demand broadening are two things that hold true at the same time.

3. Sovereign AI: The Buyer the Market Missed While Watching Only the Talk

The largest chunk of non-hyperscale demand is sovereign. NVIDIA’s CFO stated that FY2026 sovereign AI revenue more than tripled from the prior year, topping $30bn. Canada, France, the Netherlands, Singapore and the UK led the way, sovereign revenue was up more than 80% year over year in FY27 Q1 as well, and NVIDIA infrastructure is now deployed across roughly 40 countries with combined GDP of about $50tn. [Fact: NVIDIA earnings call]

Physical substance is attached to this too. Saudi Arabia’s HUMAIN has finalized a plan to deploy up to 600,000 NVIDIA GPUs over three years and has received its initial shipment of 18,000 GB300 units. In the UAE, Stargate UAE is building Phase 1, a 200MW slice of a 1GW cluster, targeting activation in the third quarter of this year, with Phase 1 alone taking up to 35,000 GB300 units, and operator G42 received US export approval in mid-July. The EU has a EUR 20bn AI gigafactory program underway, India is targeting 100,000 public GPUs by year-end, and Japan’s SoftBank is targeting an October commercial launch of a sovereign GPU cloud. [Fact: company and government announcements]

How does this flow through to memory? Two channels are already public. OpenAI’s global Stargate program signed a letter of intent with Samsung Electronics and SK Hynix in October 2025 for up to 900,000 DRAM wafers per month, a volume equal to roughly 40% of global DRAM capacity. The limitation that this is a non-binding letter of intent clearly remains. [Fact: press reports, LOI] And in June of this year, SK Hynix formalized a multi-year HBM4 partnership with NVIDIA spanning the next-generation Vera Rubin platform. [Fact: NVIDIA announcement] On the sovereign-wealth-fund side, there has been a continuing stream of reports on cooperation between UAE capital, Mubadala, MGX, G42 and Khazna, and the SK Hynix camp.

There is a gap that also needs to be left in place honestly. No new disclosure confirming that a sovereign entity directly bought DRAM in bulk from Samsung Electronics or SK Hynix has surfaced in June or July. [Blocked: no direct contract disclosed] Sovereign memory demand flows in through server OEMs and system vendors, so it gets mixed in with hyperscalers inside the memory makers’ customer disclosures. [Inference: distribution structure] Not being visible is not the same as not existing. There is no scenario where 600,000 confirmed GPUs come without memory attached, so the question is not whether it exists but how visible its scale and timing are in disclosures.

4. Server DRAM Re-Tightening: Meritz’s Claim and External Hard Evidence

The re-tightening in server DRAM supply since mid-July that the Meritz team flagged is not a claim the house is making alone. Four pieces of external hard evidence point the same direction.

Evidence	Detail	Source/Date
Contract price outlook	Q3 server DRAM contract prices projected +13-18% QoQ. Cannot be raised on US CSPs locked into long-term agreements (LTA), so the increase concentrates on non-LTA customers	TrendForce, 7/9
Supply growth rate	RDIMM bit supply growth of 15-20% YoY falls short of CPU shipment growth. CSPs are stockpiling ahead of an expected 2027 shortfall. Modules are shifting down from 96/128GB to 32/64GB	TrendForce, 7/9
Lead time	Server DRAM lead times exceed 40 weeks, prices are up about 90% versus late 2025, quote validity has shrunk to 1-30 days	Inventec remarks reported, 7/16
Supplier testimony	Can meet only 50-66% of key customer demand, HBM is sold out through 2027, the supply shortfall persists beyond 2027	Micron earnings call, July

[Fact: sources as listed]

The absolute price level is already flashing an anomaly of its own. The fixed contract price for commodity PC DRAM (DDR4 8Gb) hit a record $21 in June, the highest since tracking began, and the spot price runs 72% above the contract price. There were also reports that Samsung Electronics notified Chinese customers it would raise Q3 DRAM prices by up to 20%. [Fact: industry reports] A combination of rising contract prices alongside a spot premium above 70% means demand willing to pay up for urgent volume actually exists outside the contract channel. The ones paying the high price are not the LTA-locked hyperscalers but everyone outside that circle, namely sovereigns, enterprises and second-tier clouds. TrendForce’s statement that the increase is concentrated on non-LTA customers and Meritz’s statement that sovereign purchasing is moving into full swing are two expressions of the same phenomenon. [Inference: cross-reading]

Analyst Kim Sun-woo goes a step further here. His view is that the market is misreading the situation, trapped inside a narrow frame of sacrificial long-term contracts and 2027 earnings downgrades, and that this misunderstanding will be dispelled quickly as events unfold, including early execution of shareholder returns and partnerships involving equity investment from Big Tech. [Fact: report claim] This is a house forecast, not a verified fact, so whether it is confirmed by actual events in this week’s earnings season is the grading standard. That said, the direction is consistent with SK Hynix Chairman Chey Tae-won’s Nasdaq listing proceeds and partnership moves covered in our earlier piece, and with SK Hynix’s multi-year contract with NVIDIA.

5. Codex at 9 Million: The Numbers on Final Demand Start to Appear

The point most attacked in the AI CAPEX debate was not the infrastructure but what sits at the end of it. Data centers keep getting built, but the evidence that final usage is growing to match has been weak. Notable numbers have started to appear at exactly this weak link.

Date	User Count	Note
February 2026	1 million	Codex-only active users
April 8	3 million	Weekly active; usage-limit reset promised at every 1 million milestone
April 21	4 million	Up 1 million in two weeks
Late May-early June	Over 5 million	6x versus February
July 9	GPT-5.6 general availability	Three variants, Sol, Terra and Luna; Codex merged into the ChatGPT desktop app
July 12	6 million	Metric from this point on is combined Codex + ChatGPT Work
Around July 15	9 million	Up 3 million in three days

[Fact: OpenAI executives’ public posts and press]

Two caveats need to be attached for fairness. The numbers from July 9 onward are not Codex alone but a combined metric that includes ChatGPT Work, so a strict continuous comparison with the earlier stretch is difficult. And an external extrapolation that simply extended the late-May growth pace put the 10-million mark in October, whereas the actual figure reached 9 million by mid-July. It should also be stated explicitly that this extrapolation line is not an official OpenAI projection. [Fact: extrapolation source verified] Even allowing for the change in metric definition, what remains is that an acceleration pulling the projected line forward by roughly three months was measured over the three days right after the GPT-5.6 launch.

The reason Codex matters is not the size of the number but the nature of the demand. Codex is a product that converts a person’s working hours into inference time. An agent can work longer without a person staying connected any longer, and when one person runs several tasks in parallel, the volume of work grows faster than the number of users. From this point on, the unit of demand is not monthly active users (MAU) but total agent execution time. This is exactly the same structure as rewriting the ceiling of the AI memory demand model, in our earlier piece, as the number of workloads times memory per workload times execution frequency.

The efficiency debate flips here too. OpenAI itself stated that GPT-5.6 Sol is 54% more token-efficient than the prior generation. The fact that users and usage surged right after the release of a model with lower cost per token is hard evidence for the Jevons hypothesis: that in the coding-agent domain, efficiency gains do not shrink compute demand but instead widen the range of work that can be handed to AI. [Inference: Jevons effect] What we should be watching is not the rate at which token prices are falling but the volume of work that the price decline has newly unlocked.

6. So How Does This Translate Into Memory

Restraint is needed too. A single number like Codex’s 9 million cannot explain the memory cycle on its own. For a user count to translate into memory bit demand, it has to pass through four conversion factors: concurrent execution volume, context length, memory capacity per accelerator, and the supply growth rate. [Inference: demand model] If any single one of these is weak, the headline-grabbing metric and actual demand come apart.

Still, lining up this week’s hard evidence against the upside variables set out as the case for the 35% beat scenario in our earlier demand-scenario piece shows an alignment.

The broadening base of accelerator demand now has hard evidence attached, in the form of the ACIE growth-rate reversal and the tripling of sovereign revenue.
Always-on agentic inference produced its first numbers, in Codex’s 3 million added in three days and the shift in the metric’s unit.
Spread beyond HBM has already shown up in prices, via the +13-18% forecast for server DRAM contract prices, 40-plus-week lead times, and a 72% spot premium.
The hypothesis that usage growth beats efficiency gains got a counterintuitive piece of hard evidence: a usage surge immediately following a 54% improvement in token efficiency.

This is also where the implication for the 2028 supply-relief narrative diverges. The relief thesis’s arithmetic generally treats the deceleration in hyperscaler CAPEX growth as the ceiling on the demand growth rate. But if half of the incremental demand is starting to come from outside that group, the ceiling assumption itself needs to be recalculated. This is exactly the point Meritz’s rebuttal is aimed at. We are not changing the probabilities right now. We hold the base 45%, beat 35%, miss 20% framework in place, while recording the fact that hard evidence has started attaching to the beat-side argument along three separate strands. The probability gets updated once the scorecard below fills in. [Inference: scenario management]

7. Checking the Counterarguments

The stronger an argument looks, the more important it is to set up its opposite.

The quality of non-hyperscale demand. ACIE includes neoclouds. Orders from second-tier clouds with fragile capital costs are the demand most likely to be cancelled first if interest rates and funding conditions tighten. A growth-rate reversal does not guarantee the durability of demand.
Sovereign political volatility. National projects sway with elections, budgets and export controls. The EU gigafactory program has already had its call for proposals pushed back twice, and Middle East volumes hinge on a single switch: US export approval.
The metric trap. Codex’s 9 million is a number that arrived right after the definition changed to a combined metric, and the criteria for what counts as active have not been disclosed. The direction of growth is hard evidence, but the magnitude has not yet been separated from marketing.
Price spikes are self-destructive. The server DRAM re-tightening is a bullish argument, but it also carries the risk, as seen in the IBM case, of eating into IT budgets and creating a demand vacuum after front-loaded buying. Chairman Chey himself has warned about demand destruction from excessive price surges.
Distinguishing forecast from hard evidence. Meritz’s event forecasts, early execution of shareholder returns, equity partnerships with Big Tech, are still forecasts. If they do not materialize, the narrow view turns out to have been right.

8. The Scorecard: What Will Settle the Answer

Whether, in the Samsung Electronics and SK Hynix earnings calls around July 30, the magnitude of the Q3 contract-price increase confirms TrendForce’s +13-18% range, and whether mentions of non-LTA customers and sovereign volumes appear.
Whether SK Hynix’s early execution of shareholder returns and equity-investment-style partnerships actually show up as disclosures. This is the direct grading item for Meritz’s forecast.
Whether ACIE overtakes hyperscale for the first time in NVIDIA’s FY27 Q2 disclosure at the end of August. If it does, the broadening demand base becomes a disclosed number rather than a narrative.
Whether a direct sovereign memory contract, or a large order routed through a server OEM, rises to the level of disclosure. Whether the Stargate letter of intent converts into a binding contract is the first candidate to watch.
Whether usage metrics for Codex and competing agents begin to be disclosed on an execution-time basis alongside crossing 10 million.

The arbiter does not change. If DRAM contract prices roll over, the market will throw out this entire demand narrative, and if they hold, the demand curve that had been drawn too conservatively gets redrawn.

The stocks mentioned in this piece are examples used for analysis and are not a recommendation to buy or sell any specific stock. Investment decisions and their outcomes are the sole responsibility of the investor. The Meritz Securities report is cited as a reconstructed summary of its main points, and responsibility for the original figures and forecasts rests with that house. Codex user counts changed definition after July 9 to a combined ChatGPT Work metric, making a strict comparison with the earlier stretch difficult, and the extrapolation line projecting 10 million in October is not an official OpenAI forecast. No direct sovereign purchase of memory has been confirmed by disclosure, and Stargate-related volumes remain at the stage of a non-binding letter of intent. Stock, index and price data are as of publicly available material through July 17, 2026, and do not reflect subsequent moves.

Kimi K3 Resets the AI Price Curve: From Kimi Linear to HBM and Big Tech Strategy

Fri, 17 Jul 2026 12:31:36 +0900

Kimi K3 launched on July 16, 2026 at $3 per million uncached input tokens and $15 per million output tokens. That is not the familiar ultra-low-price positioning of a Chinese challenger. It matches Claude Sonnet 5’s standard price and sits 40% below GPT-5.6 Sol on input and 50% below it on output. Moonshot AI is positioning K3 as a primary enterprise model, not a budget fallback.

The architecture is designed to support that claim. K3 has 2.8 trillion total parameters but effectively activates 16 of 896 experts per token. Kimi Delta Attention controls the cost of long context, Attention Residuals selectively retrieve earlier representations across depth, and quantization-aware training uses MXFP4 weights with MXFP8 activations. Moonshot recommends a supernode with at least 64 accelerators.

That combination creates a two-sided semiconductor outcome. Memory and compute per token can fall while the number of institutions able to deploy a frontier-class model, and the amount of work they run, can rise. K3 is both an efficiency technology and an infrastructure workload.

Related reading: AI token value and memory value capture / AI token futures and cost per token / US-China divergence in agentic inference infrastructure / Cross-checking the 2030 HBM shortage model

Executive Summary

K3 combines 2.8T parameters, native vision and a 1M-token context window. Its products and API are live, but as of July 17 the full weights, technical report and license have not been released. The open-weight thesis must be verified after the July 27 deadline.
Moonshot’s official GDPval-AA v2 score is 1,668, not 1,687. AA-Briefcase is 1,548. BrowseComp 91.2 uses context compaction; the no-compaction 1M-context result is 90.4. The results are strong, but mixed harnesses and company-run evaluations require independent replication.
Kimi Linear’s claims of up to 75% lower KV cache and up to 6.3x theoretical decoding throughput at 1M context come from a 48B-total, 3B-active research model. They should not be presented as measured K3 API performance.
K3 is not a low-cost model. It matches Sonnet 5’s standard price, is 50% more expensive than Sonnet’s temporary launch promotion, and is cheaper than GPT-5.6 Sol. Because only max reasoning is currently available and independent tests show high output-token use, cost per completed task may be less attractive than the rate card implies.
The semiconductor impact pits lower compute and KV cache per request against more self-hosted deployments and higher total workload. Server DRAM and enterprise SSDs are the cleanest second-order beneficiaries because long-lived agent context spills below HBM into disaggregated cache tiers.
The model and cloud layers experience opposite economics. OpenAI, Anthropic and Gemini API pricing face pressure, while Azure, AWS and Google Cloud can monetize K3 and other models through compute, storage and networking. Meta must defend US open-weight leadership.
The July 27 proof points are the license, full weights, external evaluation, throughput on NVIDIA and AMD, support on Chinese accelerators, actual output tokens per task, vLLM compatibility and cloud catalog adoption.

1. Correcting the Numbers Before Drawing Conclusions

Several numbers changed as the launch circulated through social media. The official values matter because some differences alter the investment interpretation.

Item	Official value	Investor interpretation
Total parameters	2.8T	Total model size, not per-token active compute
Expert routing	16 of 896	Extremely sparse MoE; routing and communication become first-order constraints
Context	1M tokens	Useful for repositories and research, but task cost depends on output length and cache hits
GDPval-AA v2	1,668	The official table does not show 1,687
AA-Briefcase	1,548	Above GPT-5.6 Sol at 1,495, below Claude Fable 5 at 1,583
BrowseComp	91.2	Uses compaction starting at 300K tokens
BrowseComp without compaction	90.4	The cleaner result for the native 1M-context claim
Product availability	Web, Work, Code and API live	Commercial testing can begin now
Weights	Promised by July 27	License and complete artifacts remain unverified as of July 17
Reasoning effort	Max only	Low-cost modes are not yet available

Moonshot explicitly states that K3 still trails Claude Fable 5 and GPT-5.6 Sol in overall user experience. At the same time, it reports frontier-level performance across coding, knowledge work and multimodal benchmarks. The credible interpretation is not that China has conclusively taken the lead. It is that a Chinese model has entered the performance band immediately below the strongest proprietary systems while promising a 1M context and downloadable weights.

The benchmark table is not a single controlled tournament. K3 runs at maximum reasoning effort. Depending on the task, models use Kimi Code, Claude Code or Codex, and some competitor scores are the best results across harnesses or are imported from external leaderboards. Independent reproduction remains necessary.

Still, Terminal-Bench 2.1 at 88.3, FrontierSWE at 81.2, SWE Marathon at 42.0, AutomationBench at 30.8, GPQA-Diamond at 93.5 and MMMU-Pro at 81.6 indicate that K3 is designed for long-horizon tool use and coding, not merely chat. The more important signal is the package: near-frontier capability, 1M context and promised open weights.

2. How a 2.8T Model Becomes Deployable

2.1 Total parameters are not active parameters

Computing all 2.8T parameters for every token would be prohibitively expensive. Stable LatentMoE effectively activates 16 of 896 experts, or about 1.8% of the expert pool. The technical report is needed to establish exact active parameter counts, but total parameters clearly do not equal per-token compute.

Sparse MoE shifts rather than eliminates bottlenecks.

What improves	What becomes harder
Active compute per token	Router quality and expert balance
Compute needed for a quality target	Communication across accelerators
Some inference costs	Avoiding idle accelerators and hot experts
Scaling model capacity	Keeping the full weight set accessible

This is why Moonshot recommends a supernode with 64 or more accelerators. Even when only a small subset of experts is active, the selected expert is not known in advance and the full model must remain accessible. At four bits, 2.8T weights imply a theoretical minimum of about 1.4 TB before scales, metadata, buffers, cache and replication. Sparse activation reduces arithmetic but does not remove model-residency memory or fabric requirements.

2.2 Kimi Linear addresses long-context cost

The Kimi Linear paper predates K3 and evaluates a 48B-total, 3B-active research model, not K3 itself. It combines Kimi Delta Attention with full Multi-head Latent Attention in a 3:1 ratio.

Full attention is strong at exact copying and fine-grained retrieval, but KV cache grows with context. Linear attention compresses history into a fixed-size state, reducing sequence-length dependence, but can lose exact detail. Kimi Linear uses three KDA layers followed by one full-attention layer to balance efficiency and expressiveness.

Component	Role	Trade-off
KDA	Compress long history into fixed-size state	Can weaken exact copying and fine retrieval
Full MLA	Restores precise token-to-token recall	KV cache still grows with context
3:1 hybrid	Balances efficiency and quality	Requires more complex kernels and serving software

The paper reports up to 75% lower KV cache and up to 6.3x theoretical decoding throughput at 1M context. A measured comparison in the complexity analysis reports 2.3x acceleration. These results establish direction, not guaranteed K3 API performance.

For investors, lower KV cache means more concurrent requests per accelerator. It also makes longer repositories, more documents and persistent agents economical. Memory per token can fall while total tokens rise.

2.3 Attention Residuals improve depth efficiency

Standard residual connections keep adding earlier layer outputs. In very deep networks, useful representations can be diluted. Attention Residuals allow a layer to select the earlier representations it needs. Block AttnRes groups layers and retains block-level representations, reducing memory overhead.

Moonshot’s research says roughly eight blocks recover most of the gains, and Block AttnRes can match a baseline trained with about 1.25x compute. This improves training capital efficiency, but it does not automatically reduce data-center demand. Better efficiency can be reinvested into larger models and longer tasks.

2.4 Quantization is a hardware-portability strategy

K3 applies quantization-aware training from supervised fine-tuning onward, using MXFP4 weights and MXFP8 activations. This should control accuracy loss better than aggressive post-training quantization and make deployment across multiple hardware platforms easier.

Moonshot’s kernel arena included NVIDIA H200 and a general-purpose GPU from an alternative vendor. The official post does not name a Chinese chip or claim H200-equivalent economics. What is verified is the strategic intent to optimize outside NVIDIA as well as on NVIDIA.

That matters for both China and global buyers. China needs frontier-level models that can survive constrained access to top NVIDIA systems. Other buyers want bargaining power across NVIDIA, AMD and custom accelerators.

3. What the Price Card Reveals

3.1 K3 is priced as a primary model

Model	Cached input per 1M	Standard input	Output	Current positioning
Kimi K3	$0.30	$3	$15	Frontier primary-model pricing
Claude Sonnet 5 launch promo	Varies	$2	$10	Temporary through August 31
Claude Sonnet 5 standard	Varies	$3	$15	Same headline price as K3
GPT-5.6 Sol	$0.50	$5	$30	67% higher input and 100% higher output than K3

K3 is more expensive than Sonnet 5’s current promotion. It is inaccurate to describe it as universally half-price. The clearer strategy is to match Sonnet’s standard tier while undercutting the most expensive frontier API.

The price is both a confidence signal and a monetization decision. Moonshot is no longer accepting a deep discount simply to acquire usage. It is claiming a position in the enterprise default-model tier.

3.2 Cost per completed task matters more than cost per token

Rate cards assume equal token use. Agents differ in planning length, tool calls, retries and verbosity. K3 currently exposes only max reasoning. Artificial Analysis reports that K3 used about 130 million output tokens in its Intelligence Index evaluation, more than twice the peer median of roughly 63 million. A cheaper output token can still lead to a costly completed task.

Task cost = input tokens × input rate + output tokens × output rate + tool and retry cost

Enterprise buyers will monitor task success, output tokens, time to first token, throughput, tool reliability, cache hit rate and session stability. Moonshot says Mooncake delivers more than 90% cache hits on coding workloads, making cached input one-tenth the uncached price. If that rate holds on enterprise traffic, K3’s effective economics improve materially.

3.3 Mooncake moves memory demand down the hierarchy

Mooncake separates prefill and decode clusters and disaggregates KV cache across CPU, DRAM and SSD rather than keeping everything in GPU HBM. Its paper reports up to 525% higher throughput in simulation and 75% more requests on production workloads.

This explains why AI memory demand extends beyond HBM.

Accelerator HBM → server DRAM → enterprise SSD → remote cache tier

KDA can reduce KV cache per request, yet persistent 1M-context agents increase total cache volume and retention time. Hot data remains in HBM and DRAM; colder context moves to SSD. Efficiency can soften an HBM-only thesis while strengthening the broader memory hierarchy.

4. Chinese Open Weights Move Up the Enterprise Stack

OpenRouter reports that Chinese models surpassed US models in token share on its platform in early June 2026, based on more than 450 trillion tokens from January through June 14. DeepSeek’s share rose from roughly 9% to 18% and it became the leading provider by mid-May.

The adoption path has three stages:

Low-cost open models take classification, translation and testing workloads.
Better coding and agent performance take repetitive enterprise workflows.
Near-frontier quality and million-token context compete for primary routing.

K3’s price is designed for the third stage. It is not following the most aggressive low-price strategy. It is charging a frontier-tier API price while promising weights that clouds, governments and enterprises can deploy themselves.

Proprietary APIs and open weights accumulate different strategic assets.

Proprietary frontier API	Near-frontier open weights
Controls the best UX and capability	Multiplies deployment routes and hardware options
Centralizes usage data	Lets enterprises retain data and operations
Changes pricing and policy centrally	Released files are difficult to withdraw
Model provider captures gross margin	Cloud, chip and application vendors share value

The strongest proprietary model can remain number one while losing paid token share. Enterprises can route repetitive work to K3-class models and reserve the most difficult legal, design and research tasks for the top closed model. Paid token mix and blended price matter more than leaderboard rank.

5. Semiconductor Demand: Efficiency and Diffusion Arrive Together

5.1 NVIDIA: near-term efficiency risk, medium-term deployment elasticity

Sparse MoE, KDA, quantization and autonomous kernel optimization reduce GPU time per task. Hardware portability also weakens the assumption that every frontier workload must stay on CUDA. Those are negative valuation signals for NVIDIA.

The positive side is equally material. Moonshot recommends at least 64 accelerators for self-hosted K3. Released weights could drive clouds, governments, laboratories and large enterprises to build their own clusters. Demand concentrated behind one API becomes hardware demand across many data centers.

Total accelerator demand = lower GPU time per task × higher total workload × more self-hosting institutions

The first term is negative; the last two are positive. K3 alone is not enough to cut NVIDIA earnings estimates, but neither is it enough to assume that every efficiency gain creates more GPU demand. Workload elasticity is the deciding variable.

5.2 AMD: optionality from hardware choice

AMD benefits if MXFP4, MXFP8 and vLLM portability allow enterprises to separate model choice from accelerator choice. A near-frontier open model gives buyers a realistic workload on which to test NVIDIA alternatives.

The proof must come after the weights. ROCm kernels, expert-parallel communication, 1M-context throughput and 64-accelerator stability need to be measured. AMD’s upside grows only if K3 demonstrates superior cost per successful task on MI systems.

5.3 HBM: lower per-request cache, larger model residency

Demand layer	K3 efficiency impact	Direction
Weight residency	A 2.8T model must remain quickly accessible	More accelerators and high-bandwidth memory
Per-request KV cache	KDA lowers cache size and growth	Less HBM capacity per request
Concurrent users and agents	Lower cost and open deployment can expand workloads	More total HBM and DRAM

The headlines can be negative for SK hynix, Micron and Samsung because 75% lower KV cache and 6.3x throughput sound like lower memory intensity. Those numbers belong to a Kimi Linear research model, not measured K3 production. HBM also stores weights, activations, communication buffers and batches, not only KV cache.

The medium-term balance is neutral to positive if self-hosted clusters and agent workloads grow faster than efficiency. It turns negative if workload elasticity is weak and compression improves faster than usage.

5.4 Server DRAM and enterprise SSDs are the clearest second-order beneficiaries

Long context and cache reuse cannot remain entirely in HBM. Once serving separates prefill and decode and spills cache across CPU, DRAM and SSD, server DRAM and enterprise SSD become operating assets for AI inference.

Samsung can sell HBM, server DRAM, high-capacity SSDs, foundry and packaging.
SK hynix combines HBM and server DRAM with Solidigm enterprise SSDs.
Micron supplies HBM, data-center DRAM and SSDs.

K3 does not show that AI needs less memory. It shows that AI memory is becoming hierarchical. The hottest data stays in HBM, retained context sits in server DRAM, and colder cache moves to enterprise SSD. Vendors with a full memory stack have greater resilience than a single-product HBM thesis.

5.5 Networking and custom silicon are the hidden MoE bottlenecks

Spreading 896 experts across accelerators creates variable communication depending on token routing. This is why Moonshot emphasizes balanced expert-parallel training, static shapes and removal of host synchronization from the critical path. Efficient operation across 64 or more accelerators requires high-bandwidth scale-up and scale-out fabric.

That is structurally positive for Broadcom, Marvell, NVIDIA networking, optical interconnect and switch silicon. It also encourages inference ASICs optimized for open models. K3’s 48-hour small-chip design exercise is not a commercial product, but it illustrates faster model-hardware co-design.

6. US Big Tech Strategy and Stock Transmission

6.1 Microsoft: pressure on OpenAI economics, more Azure usage

Microsoft owns exposure to OpenAI IP and economics as well as Azure infrastructure. The April 2026 partnership update keeps Microsoft as a primary cloud partner and extends a non-exclusive IP license through 2032.

K3 pressures the model layer because repetitive work can move away from OpenAI APIs. It can support the cloud layer if Azure hosts K3 as managed or self-hosted infrastructure.

Microsoft layer	K3 impact
OpenAI revenue share	Negative through price and mix
Copilot	Positive if routing costs fall
Azure AI	Positive if multi-model usage rises
Custom silicon and data centers	Positive if open-weight optimization expands

The near-term stock impact is neutral. The medium-term question is whether Azure converts model price deflation into usage and Copilot margin.

6.2 Amazon: Anthropic and AWS have different economics

Amazon is a major Anthropic investor and the provider of AWS, Bedrock, Trainium and Inferentia. K3 can reduce Claude pricing power and the value of Amazon’s Anthropic stake. If enterprises run K3 on AWS, however, EC2, Bedrock, storage, networking and Trainium usage can rise.

Amazon’s optimal strategy is to keep the workload on AWS regardless of which model wins. If K3 arrives with a permissive license and efficient Trainium support, AWS can monetize a competitor’s diffusion. The stock impact is neutral to modestly positive, although inference price competition could lower cloud margins even as revenue grows.

6.3 Alphabet: Gemini pricing pressure, TPU and Vertex defense

Google controls a model, a chip, a deployment platform and final demand through Search, Ads and Workspace. Vertex Model Garden supports first-party, third-party and open models.

K3 pressures Gemini API pricing but can create TPU and Google Cloud demand. Lower model cost also reduces the expense of AI Overviews, ad generation and Workspace agents. The stock impact is neutral to positive because Google is not solely a model vendor. The risk is that Gemini loses both performance and price leadership, raising cloud customer-acquisition cost.

6.4 Meta: from open-weight beneficiary to defender

Meta benefited from open-weight diffusion by commoditizing competitors’ APIs and lowering its own recommendation, advertising and content costs. If Chinese labs release near-frontier capability and longer context first, Meta risks losing its position as the benchmark US open-weight ecosystem.

K3 forces two responses: the next Llama must compete on long context, agent tools and deployment cost, and Meta must provide a trusted US alternative for enterprises and governments unwilling to deploy Chinese weights. The near-term earnings impact is limited because advertising drives cash flow. The strategic risk is lower returns on AI infrastructure and developer ecosystem investment if Llama falls behind.

6.5 Existing market positioning matters more than one launch

From June 22 through July 16, before most of the K3 evidence could affect trading, the US AI complex had already diverged sharply.

Company	Return	Drawdown from period high	Positioning signal
Meta	+17.9%	-3.1%	Strong ad cash flow and AI utilization expectations
Microsoft	+9.2%	-1.2%	Cloud and software resilience
Amazon	+7.3%	-3.2%	AWS and consumer recovery expectations
Alphabet	+1.4%	-5.5%	Mixed search and cloud positioning
NVIDIA	-0.6%	-3.1%	Relatively resilient
Broadcom	-4.5%	-9.7%	Custom AI optimism meets valuation pressure
AMD	-9.2%	-14.3%	Alternative accelerator optionality with high volatility
Micron	-29.6%	-32.0%	Correction after elevated memory expectations
Oracle	-29.1%	-32.7%	Tension between AI infrastructure growth and financing
Marvell	-38.8%	-40.1%	Severe derating in networking and custom silicon

These are not K3-caused returns. They show the positions into which K3 arrived. A relatively resilient NVIDIA may be more sensitive to efficiency headlines, while already-corrected Micron and Marvell could react more strongly if the released weights create verifiable infrastructure demand.

7. Company Impact Map

Company or layer	Near-term signal	Medium-term path	What to do now
NVIDIA	Valuation pressure from efficiency and portability	Self-hosted 64+ accelerator clusters offset efficiency	Watch workload elasticity, do not change estimates on the launch alone
AMD	Alternative deployment optionality	Share upside if ROCm economics are proven	Wait for post-July 27 throughput
Broadcom and Marvell	Networking complex already derating	MoE expert parallelism raises fabric demand	Verify orders and actual K3 cluster topology
SK hynix	Sensitive to KV-cache efficiency headlines	HBM, server DRAM and Solidigm SSD hierarchy exposure	Value the full memory stack, not HBM alone
Samsung Electronics	HBM efficiency risk plus catch-up position	HBM, server DRAM, SSD and foundry optionality	Broadest portfolio, execution still required
Micron	High expectations already corrected	Integrated US AI memory and storage exposure	Watch deployment volume versus pricing
Microsoft	OpenAI price and mix pressure	Azure and Copilot cost leverage	Cloud usage matters more than model margin
Amazon	Pressure on Anthropic asset value	AWS, Bedrock and Trainium benefit from model choice	The cleanest internal offset
Alphabet	Gemini price pressure	TPU, Vertex and Search cost benefits	Application-layer defense remains strong
Meta	Pressure on US open-weight leadership	Llama acceleration or broader open ecosystem	Strategic impact exceeds near-term earnings impact

There is not enough evidence for a new buy or sell call from this launch alone. Weights, license and external hardware throughput remain unavailable. The order of proof matters more than the direction of the narrative.

8. Three Scenarios

Base case: strong model, limited re-rating

Subjective probability: 55%. Full weights and a reasonably permissive license arrive by July 27. External results reproduce 85% to 95% of official performance. K3 runs on NVIDIA and AMD, but 64+ accelerators, max reasoning and high output-token use keep deployment expensive.

Proprietary APIs face more price pressure on repetitive work.
Clouds gain K3 hosting and self-deployment demand.
Per-token efficiency offsets higher workload.
HBM is neutral to modestly positive.
Server DRAM, enterprise SSD and networking are positive.

Bull case: open weights enter primary enterprise routing

Subjective probability: 25%. The license is close to MIT, vLLM and SGLang support is stable, and NVIDIA, AMD and Chinese accelerators show strong throughput. External evaluation confirms performance immediately below the top proprietary models. Lower reasoning modes reduce task cost. Major clouds and enterprise gateways add K3 to default routing.

Proprietary model blended price and token share fall.
Hyperscaler multi-model usage rises.
Self-serving clusters increase accelerator demand.
HBM, networking, server DRAM and enterprise SSD benefit from deployment volume.
Meta and US open-model programs accelerate investment.

Bear case: weights reveal a cost and quality gap

Subjective probability: 20%. Weights are delayed or restricted, external scores fall below the official table, preserved-thinking-history requirements and excessive proactiveness create enterprise failures, and task cost exceeds Sonnet 5 because of max reasoning and verbosity. The 64-accelerator requirement limits self-hosting.

Moonshot may have to retreat from frontier pricing.
Proprietary APIs retain a UX and reliability premium.
Incremental semiconductor demand remains limited.
Existing earnings and capex regain control of stock prices.

9. The July 27 Verification Checklist

Proof point	Strong signal	Weak signal
Weights and license	Full release on time with commercial modification rights	Delay, restrictions or missing components
External evaluation	Official scores broadly reproduced under one harness	Large drop versus company table
Cost per task	Lower reasoning modes and fewer output tokens	Max-only reasoning and persistent verbosity
NVIDIA throughput	Stable expert parallelism on 64-accelerator nodes	Fabric bottlenecks and low utilization
AMD throughput	Cost advantage under ROCm and vLLM	Immature kernels or accuracy loss
Chinese accelerators	Named hardware and measured throughput	Only “alternative GPU” language
Caching	Near-90% hits on enterprise traffic	Sharp decline outside coding
Cloud adoption	AWS, Azure, Google Cloud or Oracle listings	Limited to a few Chinese platforms
Competitive pricing	Standard price cuts or wider cache discounts	Temporary promotions only
Memory orders	Upward revisions to HBM, server DRAM and SSD volumes	Efficiency rises while volumes stagnate

10. The Strongest Counterargument

The strongest bear case is straightforward: semiconductor demand falls if efficiency improves faster than usage. Model compression, linear attention, quantization, cache reuse and inference ASICs could process the same number of useful tasks with far fewer GPUs and less HBM. Open-weight competition can reduce API prices without producing enough incremental paid work to justify AI capex.

Evidence for that outcome would include slower inference revenue despite falling prices, higher accelerator utilization but fewer new clusters, HBM bit growth below model-efficiency gains, weak conversion of cloud AI backlog into revenue, and enterprises using open models only to cut costs rather than create new workflows.

The bull case requires price declines to unlock new work: repository-scale coding, research automation, video editing, chip design and workflows that were previously uneconomic. Workload elasticity relative to model efficiency is the common proof point for both accelerator and HBM investing.

11. Conclusion

Kimi K3 does not prove that China has surpassed the strongest US proprietary models. Moonshot acknowledges the remaining UX gap. As of July 17, the full weights and technical report are not available, and a mixed-harness benchmark table cannot declare a winner.

It does change market structure. A 2.8T, 1M-context, near-frontier model is live at Sonnet’s standard price, with weights promised within ten days. Chinese open weights are moving from cheap second-tier substitutes toward primary enterprise workloads.

For semiconductors, the event is more demand reallocation than demand destruction. KDA and quantization reduce HBM and GPU time per request. Sparse MoE and 64-accelerator supernodes increase model-residency memory and fabric. Mooncake pushes cache into server DRAM and enterprise SSDs. The HBM-only story becomes less simple, while the full memory and data-center hierarchy remains exposed to wider deployment.

US Big Tech faces the same split. Model APIs absorb price pressure; clouds and applications absorb lower model cost. Microsoft, Amazon and Alphabet can host K3 even if their preferred models lose share. Meta must defend its strategic role as the trusted US open-weight standard.

On July 27, the important evidence is not the existence of a weight file. It is a permissive license, reproducible evaluation, multi-hardware throughput and competitive cost per completed task. If those conditions hold, K3 becomes an event that changes both the AI price curve and the semiconductor demand path. If they do not, it remains an impressive product launch rather than an industry reset.

Sources and Limitations

Primary materials: Kimi K3 launch, Kimi K3 API documentation, Kimi Linear paper, Kimi Linear GitHub, Attention Residuals paper, Mooncake paper, Claude Sonnet 5 pricing, GPT-5.6 Sol pricing, OpenRouter model adoption analysis, Microsoft-OpenAI partnership, Google Vertex Model Garden, and Amazon-Anthropic partnership.

The semiconductor and stock transmission analysis is an inference from public architecture, serving and market data. Exact active parameters, weight size, license terms, AMD and Chinese-accelerator throughput, and enterprise cost per completed task remain blocked as of July 17. This article is for research and information purposes only and is not investment advice.

Server DRAM on Korea Invest Insights

Alphabet's Q2: Cloud +82% Ends the Demand Debate, Negative FCF Starts the Cash Debate

TL;DR

1. The Numbers First: What Came Out

2. The Demand Debate: Why We Think It’s Effectively Over

3. The Cash Debate: The Bill Arrived Sooner Than Expected

4. Scoring the Three Questions

5. Translating to Memory: Reinforcement for Volume, Not a Warranty on Margin

6. The Stock Reaction, and Where the Two Notes Diverge

7. Scorecard: The Next Two Weeks Will Decide

Related Posts

Who Burns All Those Tokens? NVIDIA's Customer Map, Sovereign AI and Codex at 9 Million Start Answering

TL;DR

1. What Kind of Week Was This: Between a Limit-Down and a Report

2. NVIDIA’s Customer Map: The Share Sits at 50%, the Growth Sits Outside

3. Sovereign AI: The Buyer the Market Missed While Watching Only the Talk

4. Server DRAM Re-Tightening: Meritz’s Claim and External Hard Evidence

5. Codex at 9 Million: The Numbers on Final Demand Start to Appear

6. So How Does This Translate Into Memory

7. Checking the Counterarguments

8. The Scorecard: What Will Settle the Answer

Related Posts

Kimi K3 Resets the AI Price Curve: From Kimi Linear to HBM and Big Tech Strategy

Executive Summary

1. Correcting the Numbers Before Drawing Conclusions

2. How a 2.8T Model Becomes Deployable

2.1 Total parameters are not active parameters

2.2 Kimi Linear addresses long-context cost

2.3 Attention Residuals improve depth efficiency

2.4 Quantization is a hardware-portability strategy

3. What the Price Card Reveals

3.1 K3 is priced as a primary model

3.2 Cost per completed task matters more than cost per token

3.3 Mooncake moves memory demand down the hierarchy

4. Chinese Open Weights Move Up the Enterprise Stack

5. Semiconductor Demand: Efficiency and Diffusion Arrive Together

5.1 NVIDIA: near-term efficiency risk, medium-term deployment elasticity

5.2 AMD: optionality from hardware choice

5.3 HBM: lower per-request cache, larger model residency

5.4 Server DRAM and enterprise SSDs are the clearest second-order beneficiaries

5.5 Networking and custom silicon are the hidden MoE bottlenecks

6. US Big Tech Strategy and Stock Transmission

6.1 Microsoft: pressure on OpenAI economics, more Azure usage

6.2 Amazon: Anthropic and AWS have different economics

6.3 Alphabet: Gemini pricing pressure, TPU and Vertex defense

6.4 Meta: from open-weight beneficiary to defender

6.5 Existing market positioning matters more than one launch

7. Company Impact Map

8. Three Scenarios

Base case: strong model, limited re-rating

Bull case: open weights enter primary enterprise routing

Bear case: weights reveal a cost and quality gap

9. The July 27 Verification Checklist

10. The Strongest Counterargument

11. Conclusion

Sources and Limitations