← Back to Research
Compute Infrastructure

Compute's Bottleneck Moves to Memory: Micron Earns $37.7 Billion in One Quarter

FutureX Research · AI Lab · 2026.05.12 · 24 pp · preview 6 pp · Revised 2026.11.02

🎧

Listen · Audio Summary

5-8 min · AI narration in English · abstract + all key findings

Bottom Line

AI compute demand is still being written into contracts, while the bottleneck is moving from chips to memory, packaging and power.

Why now
On September 30, 2026, Micron reported quarterly net income of $37.7 billion, against $3.2 billion a year earlier.
What it means
For founders, compute costs must be re-priced around memory and bandwidth; for investors, risk is shifting from whether orders exist to whether buyers can keep financing the payments.

Abstract

AI compute demand is still being written into contracts, while the bottleneck is moving from chips to memory, packaging and power. Oracle's remaining performance obligations reached $664 billion (September 10), and Micron earned $37.7 billion in a single quarter (September 30). CoWoS lead times ran 52-78 weeks in July, and GlobalFoundries has signed on to supply interposers to TSMC (October 8). Risk is shifting to buyers: Anthropic's draft prospectus reportedly lists about $518 billion in compute commitments, while OpenAI's annualized revenue, on its own accounting basis, is approaching $50 billion, about $20 billion below what the market had assumed. (Data as of 2026-10-11)

Key Points

  • 01On September 30, 2026, Micron reported fiscal Q4 2026 net income of $37.7 billion, up from $3.2 billion a year earlier, as DRAM revenue rose 343% (Per company disclosure); memory suppliers are still collecting a shortage premium.
  • 02On October 8, 2026, TSMC reported Q3 revenue of about NT$1.49 trillion, above its July guidance; the same day GlobalFoundries signed a $2 billion deal to supply silicon interposers for TSMC's CoWoS (Verified).
  • 03On September 10, 2026, Oracle reported $664 billion of remaining performance obligations and 850 MW of new data-center capacity in one quarter (Per company disclosure); demand is on the order book, and on-time delivery is now the main variable.
  • 04On September 29, 2026, Reuters reported that Anthropic's draft prospectus lists about $518 billion in compute commitments, more than 110 times its 2025 revenue of nearly $4.6 billion (Reported).
  • 05On August 21, 2026, a16z published OpenRouter data showing that more than 85% of agentic tokens come from cached prompts (Verified); on that basis, this report sees the bottleneck moving from compute to memory.

In Late July Four Constraints Tightened at Once, and Capex Guidance Overtook Revenue Growth as the Main Driver of AI Stocks

In late July, demand, supply, power and capital all tightened at once across AI compute infrastructure. Demand has not slowed, supply remains constrained and power is now a first-order bottleneck, so all three of our earlier calls got stronger. The new variable is that public markets now price AI narratives with free-cash-flow discipline: capex guidance has replaced revenue growth as the main driver of stock prices.

Demand: On July 22, Alphabet raised full-year capex guidance to $195-205 billion. On July 27, Moonshot AI released Kimi K3's full weights, setting Hugging Face's fastest-growth record.

Capital: Microsoft and SK Hynix reported on the same day, July 29. Microsoft rose about 15.5% the next day on stronger-than-expected demand; SK Hynix's operating profit rose 557% year on year, yet the stock sank on a miss (Public market data). SK Hynix had listed on Nasdaq only on July 10. SpaceX raised $75 billion on June 12 in the largest US IPO ever, then broke below its $135 issue price in mid-July, shedding more than $1.2 trillion from its peak value (Reported). Anthropic, valued at $965 billion after its May round, began investor meetings on July 15 with October as the earliest listing target; it has since reportedly slipped to mid-November at the earliest. OpenAI reportedly slips to 2027.

Private markets: Moonshot plans to open talks in August on a final pre-IPO round targeting a $50 billion pre-money valuation, against $31.5 billion pre-money in its current round (Reported, per NBD). Zhipu, HK-listed since January, completed a placement of about HK$31.4 billion on July 13.

Power: On July 30, SpaceX (which absorbed xAI in February) said its Memphis Colossus site runs on 69 unpermitted gas turbines, to be removed by July 2027 as it moves to a permanent 1.2GW captive gas plant; its IPO filing discloses $2.8 billion of further turbine purchases over the next three years. The House Energy Committee wrote on July 29 demanding records by August 11. Musk-affiliated entities reportedly bought mobile-power firm APR Energy, and its fleet of more than 1GW, for over $1 billion (Axios).

Applications and disputes: SoftBank became Sierra's exclusive partner in Japan on July 14, and LINEMO's customer-service resolution rate rose to 97% (Reported). Apple sued OpenAI on July 10 over alleged theft of trade secrets; the claim is one-sided, and OpenAI denies it (Reported).

Demand Has Not Slowed: Open-Weight Kimi K3 Tops a Coding Leaderboard, and Big Tech's 2026 Capex Passes $700 Billion

The loudest demand signal came from open source. Kimi K3 is a 2.8-trillion-parameter MoE model with about 104 billion active parameters and a 1-million-token context window. Moonshot released its full weights on July 27 under a custom Kimi K3 license tiered by revenue, and, unusually, also open-sourced three pieces of training infrastructure: MoonEP, FlashKDA and AgentEnv. The model topped Hugging Face trending within 30 minutes.

On WebDev Arena, a blind front-end coding test, an open model took first place for the first time:

ModelWebDev Arena Elo
Kimi K3 (open weights)1679
Claude Fable 51631
GPT-5.6 Sol1618

K3 still trails both closed flagships on composite intelligence indices. Huawei Ascend 950, Alibaba Cloud's M890, Cursor and Cognition shipped Day-0 support. Chinese models reportedly account for about 60% of US token usage on OpenRouter.

The money tells the same story. Figures from the late-July earnings season:

CompanyMeasureFigure
AlphabetQ2 capex$44.9 billion, roughly double a year earlier
Alphabet2026 full-year capex guidance$195-205 billion
MicrosoftQuarterly capex$41 billion, +70%
Amazon2026 capexAbout $220 billion
MetaQ2 free cash flow$784 million (Reported)

Per Reuters, combined Big Tech spending exceeds $700 billion this year. Microsoft's Azure passed $100 billion in annual revenue. SoftBank's Masayoshi Son says AI will need $5 trillion of investment a year by 2040. The race has not cooled either: xAI is reportedly training a 2-trillion-parameter Grok 4.6 aimed at surpassing K3.

Supply Is Still Constrained: TSMC Lifts Capex to $60-64 Billion While CoWoS Lead Times Run 52-78 Weeks

TSMC posted its fifth straight record quarter on July 16: revenue of $40.2 billion (+36% year on year in NT dollars), net income up 77.4% and a 67.7% gross margin. HPC made up 66% of revenue, and 2nm began to ramp. Management raised two full-year targets:

Full-year guidanceBeforeAfter
USD revenue growthAbove 30%Above 40%
Capex$52-56 billion$60-64 billion

The supply side is voting for demand with capex. The constraint is still packaging: CoWoS is sold out with 52-78 week lead times, and Nvidia reportedly holds about 60% of capacity. On July 30, TSMC was reported to be launching an EMIB-like packaging route to add capacity, and its ADRs rose more than 7% (Public market data).

Memory is just as tight. SK Hynix posted Q2 operating profit of KRW 60.54 trillion, up 557% year on year, and has shipped HBM4 since Q2. Its CEO expects the worst memory shortage in history in 2027. The results still missed estimates and the stock sank (Public market data). When record results are not enough, the cycle is both hot and fragile.

85% of Agent Tokens Are Cached, Moving the Bottleneck From Compute to Memory

Verified (a16z, published August 21, 2026, data from openrouter.ai/rankings): agentic token consumption on OpenRouter reached 7.3 trillion on a seven-day average, about five times human usage. It has grown about 14x in the roughly six months since overtaking human usage on February 6, 2026.

Verified (OpenRouter data, disclosed via a16z): more than 85% of agentic tokens come from cached prompts, and cached tokens account for nearly all of the category's relative growth. Agents run read-write-execute loops and keep context across steps, so the initial prompt that loads the task context is counted again and again.

What it means for infrastructure: this report had reduced the demand side to growth in inference calls and derived GPU and rack demand from that. The simplification needs correcting. Cache-heavy workloads keep the KV cache resident for long periods. Compute per token falls, while memory capacity and bandwidth use rise sharply. The marginal bottleneck is therefore more likely to sit in HBM supply, memory bandwidth and interconnect than in FLOPs.

How to test it: if this pattern holds, three things should follow. High-bandwidth memory gets tight earlier and harder than GPUs overall; memory-to-compute ratios rise in new platform generations; and inference providers price cached and non-cached tokens more differently.

Uncertainty: the data comes from a single platform. OpenRouter's workload mix need not represent the industry, and especially not hyperscalers' in-house inference clusters. Treat this section as directional only.

Outside Views: Patel and Baker Both See a Multi-Year Shortage, and the Falsifier Is Rents That Stop Repricing

Dylan Patel (founder, SemiAnalysis; August 17; Verified): the supply-demand gap is widening, so cost per token keeps rising. Memory is in a multi-year structural shortage with no cyclical peak in sight: capacity grows 20–30% a year while AI demand doubles. In late June he also argued that co-designing chips, kernels and models yields 100x, while optimizing any single layer yields 8x. That explains why frontier labs keep pushing custom silicon, and why a design partnership, the route Anthropic turned to in its MatX talks, is still worth having.

Gavin Baker (CIO, Atreides; August 23; Verified): his firm's internal AI spending was about 100x higher in August than in March and is still roughly doubling every month. He calls this a shortage that runs through 2028, not a bubble, and asks bears to name one quantitative metric in their own business that is getting worse. Reportedly, he also estimates that rents for clusters of several thousand B200s rose 50–60% in seven months. Repricing as old contracts expire is exactly the signal this report tracks.

Both men point the same way, and both sit on the buy side: one sells research, the other runs long positions. Read them with that in mind. The bears' hardest number is Meta's Q2 free cash flow, down 91% year on year. The shortage story is falsified if rent repricing stops.

FutureX Position · The Constraint Moves From Chips to Power, Memory and Data, and Debt Piles Up in the Infrastructure Layer (Xiamen keynote, 2026-09-03)

The Xiamen keynote reinforces this report's conclusion and supplies the mechanism behind shifting constraints. FutureX's view: the binding constraint keeps moving. Chips came first, then power, then memory; today the tightest one is data. Each move opens an investment window.

Infrastructure carries the most debt. On FutureX's five-layer map, circular financing, credit sliding down the stack and GPU depreciation all land in this layer. Our reading is a maturity mismatch:

MeasureYears
Financing bond tenorTypically 7 to 10
Book life of the asset5 to 6
Economic lifePossibly only 2 to 3

This report leaves the circular-financing question open. FutureX's position is to stay out of debt-financed compute rental.

Energy has not entered the narrative yet. FutureX's view: the ceiling on compute is written on the grid.

At the chip layer, the bottleneck is moving from raw compute to bandwidth and energy efficiency. Compute efficiency rose about 1,000x from 2012 to 2023. The sources break down as follows:

SourceContribution
Number formats (quantization)16x
Sparsity2x
Instruction sets12.5x
Process, 28nm to 5nm2.5x

The next 1,000x will come from silicon photonics, compute-in-memory and 3D integration, not from process nodes. This supports the report's conclusion: custom silicon is inevitable, but building it yourself is not. The gains sit in architecture.

Full deck: /reports/open-source-breakthrough (first 5 pages public).

Latest developments, newest first (data current as of 2026-10-11)

  • 2026-10-08 · TSMC's September revenue was NT$511.86 billion, down 0.6% month on month and up 54.6% year on year. Q3 revenue reached about NT$1.49 trillion (about US$46.75 billion), up 17% from Q2 and above its July guidance of US$44.6 billion to US$45.8 billion. January-September revenue was about NT$3.9 trillion, up 41.1%; full quarterly results were scheduled for its October 15 earnings call (Taipei Times, Electronics Weekly) · Verified
  • 2026-10-08 · GlobalFoundries signed a US$2 billion agreement with TSMC, with an initial five-year term, to supply silicon interposers for CoWoS advanced packaging from its fab in Malta, New York; volume ramp is planned for the first half of 2028 · Verified
  • 2026-10-08 · Samsung Electronics guided Q3 consolidated sales to about KRW 195 trillion and operating profit to about KRW 107.4 trillion, versus KRW 171.5 trillion and KRW 89.49 trillion in Q2. The figures are range midpoints, preliminary and unaudited · Per company disclosure
  • 2026-10-08 · OpenAI told investors its annualized revenue was approaching US$50 billion at the end of September, about US$20 billion below the roughly US$70 billion reported by media in late September. The gap is an accounting difference: Anthropic counts sales made through cloud partners such as AWS and Google Cloud, while OpenAI does not (Financial Times, Reuters) · Reported
  • 2026-10-08 · The Nasdaq Composite closed down 1.25%; Nvidia fell about 2.9%, Oracle 5.5% and CoreWeave nearly 8% (CNN, CNBC) · Public market data
  • 2026-09-30 · Micron's fiscal Q4 2026 revenue was US$54.23 billion (US$11.32 billion a year earlier) and net income US$37.7 billion (US$3.2 billion). DRAM revenue rose 343% to US$39.8 billion, 73% of sales. Next-quarter revenue guidance of about US$61.5 billion beat the US$57 billion analysts expected. Micron is investing US$250 billion in two new HBM campuses in New York and Idaho (CNBC) · Per company disclosure
  • 2026-09-29 · A non-public draft Anthropic prospectus seen by Reuters lists about US$518 billion of planned cloud, compute and infrastructure commitments over the coming years, most of them non-cancelable. Anthropic spent about US$7.33 billion on compute and infrastructure in 2025, roughly three times the prior year, on revenue of nearly US$4.6 billion · Reported
  • 2026-09-25 · Goldman Sachs expects hyperscaler AI capex of about US$800 billion in 2026, rising about 50% to roughly US$1.2 trillion in 2027 (Bloomberg, investingLive) · Reported
  • 2026-09-18 · Tencent News reported that DeepSeek plans to deploy at least 160,000 Huawei AI accelerators at a data center in Inner Mongolia · Reported
  • 2026-09-17 · At Huawei Connect in Shanghai, deputy chairman and rotating chairman Wang Tao unveiled the Ascend 960 SuperPoD, the first to use near-packaged optics (NPO): 4,096 cards per node, 8 EFLOPS of FP8 compute and 1 PB of HBM. 5,500 Hi-ONE units replace 48,000 800G optical modules, cutting power draw by more than 550 kW and lifting utilization 2.75x over a conventional cluster. The Ascend 960DT is due in Q1 2027, three quarters ahead of plan; the 970 and 980 follow in 2028 and 2029. Huawei says more than 1,000 Ascend SuperPoDs are deployed across more than 370 customers (Tencent News, September 18) · Verified
  • 2026-09-10 · Oracle's fiscal Q1 2027: remaining performance obligations up US$209 billion year on year to US$664 billion; more than US$30 billion of new AI cloud contracts in the quarter; total revenue US$19.3 billion (+30%), cloud revenue US$11.6 billion (+62%), of which cloud infrastructure US$7.4 billion (+121%); quarterly capex US$28.5 billion, about 3.4 times the US$8.5 billion a year earlier; 850 MW of new data-center capacity brought online in the quarter · Per company disclosure
  • 2026-09-09 · Nvidia announced work with eight Australian partners (Firmus, Sharon AI, IREN, ResetData, Megaport, CDC, NEXTDC and AirTrunk) to deliver up to 2 GW of AI compute by 2027. CDC says it operates more than 550 MW across Australia and New Zealand with a further 800 MW under construction; Sharon AI plans to deploy up to 68,000 Nvidia GPUs. No dollar figures were disclosed · Per company disclosure
  • 2026-09-03 · Nvidia announced it will acquire Hugging Face for US$12.93 billion. The platform serves 18 million developers and researchers, hosts 3 million models, 500,000 datasets and 1 million applications, and is used by more than 200,000 companies; Nvidia has published over 500 models and 250 open datasets there. Jensen Huang said developers will keep choosing their own models, frameworks, clouds and inference providers, and that building on or deploying through Hugging Face will not require Nvidia compute · Per company disclosure
  • 2026-09-03 · The Motley Fool compiled the two neoclouds' earlier public disclosures: CoreWeave carries US$35 billion of debt, guides 2026 capex to US$35-39 billion and paid US$640 million of interest last quarter, 2.4 times a year earlier; Nebius spent US$8 billion on capex in the first six months of 2026, plans more than US$20 billion for the year and has just raised US$5.75 billion through convertible notes · Reported
  • 2026-09-02 · Broadcom's fiscal Q3 2026 (quarter ended August 2): revenue US$29.6 billion (+86%), AI semiconductor revenue US$16.7 billion (+221%), adjusted EPS US$3.32 against a US$3.24 consensus; Q4 AI semiconductor guidance US$21.7 billion (+236%); on the call, a multi-year AI revenue outlook of about US$58 billion in fiscal 2026, US$115 billion in fiscal 2027 and US$230 billion in fiscal 2028 (CNBC and Seeking Alpha coverage of the release and call) · Reported
  • 2026-08-27 · Reuters reported exclusively (August 27–28) that Anthropic held talks to acquire AI chip startup MatX for about US$7 billion, then dropped the acquisition in favor of discussing a partnership. MatX, founded by former Google TPU engineers, builds chips for training large models and is reportedly raising a new round at about a US$4 billion valuation. Reuters could not determine why the talks ended; Anthropic declined to comment and MatX did not respond · Verified
  • 2026-08-10 · Per an Nvidia company statement, Nvidia signed memoranda of understanding with six asset managers (Apollo Global Management, Blackstone, BlackRock, Brookfield, Goldman Sachs and KKR) to build an independent financing platform that would mobilize more than US$500 billion of third-party capital over time for data-center construction and hardware purchases by cloud providers and AI companies. Jensen Huang answered circular-financing criticism, calling AI-factory compute an investable asset class, saying demand comes from real commercial workloads and that each project is independently diligenced. Nvidia closed down 2.86% that day · Verified

What This Means for Founders

Price compute around memory and bandwidth, not just FLOPs: more than 85% of agentic tokens on OpenRouter come from cached prompts, which makes HBM and interconnect the more likely bottleneck. Anthropic walked away from buying MatX for about $7 billion and moved to partnership talks, a sign that the barrier to custom silicon is not money; exits for AI chip teams are shifting from outright acquisition toward partnership plus a minority stake. Custom-silicon schedules depend first on a committed buyer: Huawei, with orders in hand, pulled the Ascend 960DT forward three quarters to Q1 2027. Signals to watch: whether inference providers price cached and non-cached tokens differently, and whether Nvidia keeps its platform-neutrality promise after buying Hugging Face.

What This Means for Investors

Demand has not slowed, and the profits are landing in memory: Micron's quarterly net margin is nearly 70%, and Samsung guided Q3 operating profit to about KRW 107.4 trillion, up about 20% from Q2. Risk is shifting to buyers' ability to pay: Anthropic's roughly $518 billion of compute commitments is more than 110 times its 2025 revenue. The layer that borrows to buy chips needs its own price: CoreWeave carries $35 billion of debt, and its quarterly interest bill is 2.4 times a year earlier. If Nvidia's $500 billion financing platform takes hold, the binding constraint shifts from capex capacity to asset-financing capacity; the independent diligence and lending terms of its first projects will test the circular-financing debate. This is not investment advice.

Where We Could Be Wrong

  • 01If Micron's next-quarter revenue comes in clearly below its guidance of about $61.5 billion, the memory-shortage call should be downgraded.
  • 02If hyperscalers' full-year 2027 capex guidance, given in early 2027, sums to less than $1 trillion, this report's supply-tightness scenario should move back toward the base case. Goldman Sachs expects about $1.2 trillion; a cut to TSMC's fourth-quarter or 2027 outlook would trigger the same move.
  • 03If OpenAI or Anthropic discloses revenue growth clearly below the growth of its compute commitments, the supply-tightness scenario should move back toward the base case. If B200 cluster rents stop repricing as old contracts expire, the shortage story itself fails.

Key Questions

How much are Google, Microsoft and Amazon spending on AI capex in 2026?

Over $700 billion combined (Reuters). Alphabet's Q2 capex hit $44.9B with full-year guidance raised to $195-205B; Microsoft spent $41B in the quarter (+70%) as Azure passed $100B annual revenue; Amazon lifted 2026 capex to ~$220B. Markets now price free-cash-flow discipline — Meta's Q2 FCF reportedly fell to just $784M.

Is TSMC's CoWoS advanced packaging still sold out, and what are lead times?

Yes — CoWoS remains sold out with 52-78 week lead times. TSMC's July 16 Q2 showed $40.2B revenue (+36% YoY in NT$) and net income +77.4%, a fifth straight record; full-year capex was raised from $52-56B to $60-64B. On July 30 TSMC reportedly launched an 'EMIB-like' packaging route, lifting its ADRs over 7%. Advanced packaging is still the binding constraint.

What did Moonshot open-source with Kimi K3, and does it beat Claude and GPT?

On July 27 Moonshot released the full weights of the 2.8-trillion-parameter Kimi K3 under a bespoke MIT-style Kimi K3 License with revenue-tiered commercial terms, plus its MoonEP/FlashKDA/AgentEnv training infrastructure. It topped Hugging Face trending within 30 minutes — the platform's fastest launch ever — and topped WebDev Arena at 1679 Elo on its July 16 debut, above Claude Fable 5 (1631) and GPT-5.6 Sol (1618), with Day-0 support from Huawei Ascend 950 and Alibaba Cloud.

Watch & Listen

In China: search WeChat Channels for 「倩姐投AI」; full library → Qian on AI

Sourcing and standards

Compiled from public sources; data current as of 2026.10.11. The text separates four evidence grades: verified, reported, per company disclosure and public market data, and states the derivation behind every estimate. When we get something wrong, the correction is written into the report body with the original call left visible, and logged publicly.

Research standards & corrections →

📄 Full Report

Full report: 24 pages · provided to professional investors & partners only

This is the public preview. The full report includes the sections below. For compliance reasons it isn't posted publicly or offered as a free download. To request a copy, contact the FutureX team.

  • 🔒The Power Bottleneck: From APR Energy to a 1.2GW Captive Plant — the Cost of 'Bring Your Own Power' and the Regulatory Backlash
  • 🔒Capital Markets: SpaceX Below Issue, SK Hynix Sold Off on Records — Pricing Anthropic's IPO
  • 🔒Primary Market: Moonshot's $50B Pre-Money Push and Zhipu's HK$31.4B Placement — Falsifiable Premiums
  • 🔒Applications: SoftBank x Sierra Exclusivity and the Unit Economics of Agent Customer Service
  • 🔒Clash of Giants: Apple v. OpenAI and the Latest in the AI Hardware Talent War
  • 🔒Value-Chain Map, Risk Matrix and Scenario Analysis
Request the full report →

Where we stand on this

Questions people ask next

Building in this space, or want to discuss this report? Write to us. We usually reply within 48 hours →

Related Research

Industry research from FutureX Capital's AI Lab, compiled from public information; not investment advice; contains no fund performance, AUM, or offer to raise capital.