The Half-Life of Frontier Models

A frontier AI lead now lasts months, but the compute deals behind it run for years. How China plays its hand, and what that means for the AI capex bet.

ByMichel JohannsenDate2026·10·19Read18 MINTopicFINANCEViews—
Share

In its quarterly report for the period ended 26 July 2026, Nvidia named three Chinese AI models in its risk factors. Any regulatory control that limits its ability to support third-party "applications built on foundation models originating in China such as DeepSeek, Qwen, or Kimi", the company wrote, "could have a material impact" on its business.1 The same filing notes that future demand may depend on the ability of "customers and partners, including those developing open models, to generate revenue and sustain investment in computing infrastructure".1

That is an unusual sentence for the world's most valuable chipmaker. The models it names are free to download. They are built by labs that own a small fraction of the world's AI compute, and, by Epoch AI's measure, Chinese models have trailed the best US models by about seven months on average since 2023.2 Yet the company whose chips power the American frontier now lists the ability to serve them as a business risk.

The final piece of my earlier series on AI infrastructure financing ended on one weak assumption: that per-token revenue holds up long enough to pay for hardware that is depreciated over five to six years.3 This piece starts there. It asks a narrower question than "is there a bubble". How long does a frontier lead stay worth money, and what happens to the companies that financed it if the answer is "a few months"?

This piece does not argue that anyone is planning to bleed anyone. It describes what can be observed: the cards China was dealt, how its labs and its state have played them, the game that results, and what that game means as a stress test for the US frontier labs and the hardware companies behind them. Where the evidence points in different directions, both directions are shown.

The Hand China Was Dealt

Start with the constraints, because they explain most of what follows.

Since October 2022 the United States has restricted the export of leading AI chips to China, tightening the rules in 2023 and adding high-bandwidth memory (HBM) in December 2024.4 In April 2025 Nvidia's China-specific H20 was put under a licence requirement "for the indefinite future". Nvidia booked a $4.5 billion charge and left about $2.5 billion of orders unshipped in one quarter, and expected roughly $8 billion of lost H20 revenue in the next.5 Since then the policy has loosened and tightened several times: H20 sales were allowed again in August 2025 against a 15% revenue share, H200 sales in December 2025 against 25%, and since January 2026 under a case-by-case rule that caps China-bound volume at half of what is shipped to US customers.6

The result shows up in the compute statistics. Epoch AI estimates that Chinese companies owned "just over 5%" of the cumulative compute of leading AI chips at the end of 2025, less than any single large US hyperscaler, and that the share is falling.7 Measured by the performance of large GPU clusters, China's share is around 14 to 15%.8 Huawei, China's main domestic alternative, will produce an estimated 1.5 million Ascend chips in 2026, against 5.9 million for Nvidia. In compute terms that is less than 4% of Nvidia's output, and the binding constraint is not chip design but HBM.9

The other half of the hand is stronger than the export-control debate usually admits.

The cards dealt (constraints)The cards held (strengths)
Compute~5% of leading AI-chip compute; HBM is the bottleneck; frontier training still largely on NvidiaDomestic chip stack scaling: Huawei at 28% of China's accelerator shipments in H1 2026, with units up from just over 300,000 to 572,000 year on year10
EnergyData-centre utilisation estimated at only 20 to 30%315 GW of solar and 119 GW of wind added in 2025, against 53 GW of total US capacity additions: about 8x11
CapitalLabs are loss-making; private funding smaller than in the USBig Fund III with RMB 344bn; a RMB 60bn national AI fund; compute vouchers; reported power discounts of up to 50% for data centres on domestic chips12
Market accessDistrust of Chinese APIs abroad; government restrictions or advisories on DeepSeek in at least eight jurisdictions13A domestic market large enough to sustain a price war among six or more serious labs
PolicyExposed to every change in US rulesState plans that name open source explicitly (below)

The policy texts are worth reading in the original. The State Council's "AI+" Opinions of August 2025 call for "a thriving open-source ecosystem" and for an open-source system "open to the world".14 The Party's recommendations for the 15th Five-Year Plan go further and put open source in the same sentence as self-reliance: China should "promote the cultivation of an open-source ecosystem" and "build self-reliant, controllable (自主可控), coordinated AI foundational software and hardware systems".15 The Global AI Governance Action Plan of July 2025 adds a qualifier that is easy to overlook: it supports "the open flow of non-sensitive technology resources".16

Read together, these documents do not describe open source as a gift. They describe it as a tool for independence, with a dial that can be turned down.

CHART 01 · Little compute, most of the open ecosystem% CHINESE SHARE
COMPUTEOPEN-MODEL USAGE0%20%40%60%80%Leading AI-chip compute ownedEND-2025~5%Global GPU-cluster performance2025/26~14%Hugging Face downloads12 MONTHS TO MAR 202641%New open-model fine-tunesFEB 202670%Open-model tokens on OpenRouterJAN 202672.7%
SOURCE · Epoch AI, Chip Owners Explorer (Apr 2026) and supercomputer dataset; Hugging Face, State of Open Source (Mar 2026); Lambert & Brand, ATOM Report (arXiv 2604.07190).

How the Hand Is Played

With little compute, abundant power and a state that wants independence, the observable moves follow a pattern.

Efficiency first. DeepSeek's V3 paper put the final training run at $5.6 million of GPU time and said explicitly that this excluded prior research and experiments. SemiAnalysis estimated the company's server capex at about $1.6 billion.17 The headline number was misleading, but the engineering was real: mixture-of-experts architectures, FP8 training and later sparse attention, which DeepSeek used to cut its API prices by half in September 2025.18 In August 2025 DeepSeek trained V3.1 with a number format, UE8M0 FP8, that it said was "designed for the next generation of domestically produced chips".19

Scale-out instead of better chips. Huawei's roadmap substitutes many weaker chips, optical interconnect and cheap power for per-chip performance: SuperPoDs of 8,192 chips in late 2026 and 15,488 in late 2027. Huawei's Eric Xu put it plainly: "Gaps still exist between Huawei and Nvidia in terms of individual chips and ecosystems. ... Our core focus is SuperPoDs and SuperClusters."20

Open weights as the export channel. Through 2025 and into 2026, Chinese labs released most of their flagships under MIT or Apache 2.0 licences, a pattern that has started to change (see below). By March 2026 Chinese models had overtaken US ones in cumulative Hugging Face downloads (1.15 billion vs 723 million), accounted for about 70% of new fine-tunes, and for 72.7% of the open-model tokens routed through OpenRouter.21 Alibaba now has more derivative models on Hugging Face than Google and Meta combined.22

Price. A domestic price war began in May 2024, when ByteDance priced its Doubao model "99% below" the industry average and Alibaba cut some Qwen prices by up to 97%.23 In January 2025, DeepSeek-R1's output tokens cost about 27 times less than OpenAI's o1.24

Rationing the competitor's hardware. Part of Nvidia's decline in China is a Chinese policy choice, not a US one. Chinese regulators questioned the H20 over alleged security risks, told ByteDance and Alibaba to cancel orders for Nvidia's RTX Pro 6000D, and issued a preliminary antitrust finding against Nvidia.125 Of more than 400,000 H200s that Beijing approved for ByteDance, Alibaba and Tencent in January 2026, roughly 5% had been delivered by August.26 In the first half of 2026, Nvidia's share of China's AI accelerator shipments fell to 34%, while Huawei reached 28%.10

Learning from the leader. In February 2026 Anthropic reported that around 24,000 fraudulent accounts linked to DeepSeek, Moonshot and MiniMax had generated more than 16 million exchanges with Claude, targeting reasoning, tool use and agentic coding.27 OpenAI made similar allegations about DeepSeek to the House China Committee.28 These are claims by interested parties. None has been tested in court, and critics point out that the US labs themselves trained on scraped data. By Anthropic's own figures, about 80% of the alleged volume came from MiniMax and less than 1% from DeepSeek, which is not where most of the public debate has focused.

Why do this? There are at least six readings in circulation, and each has a credible source behind it.

ReadingWhat it saysEvidence for it
NecessityForeign buyers do not trust Chinese APIs; downloadable weights need no trustMiniMax earns over 70% of revenue abroad29
Commoditise the complementThe model is cheap so the cloud, the app or the chip can earnAlibaba pitches Qwen as "the Android of the AI era"30
Domestic hypercompetitionSix or more labs in one market cannot all charge premium pricesThe 2024 price war; Baidu's turn to open models
Standard-settingWhoever supplies the default base model shapes the ecosystem and the hardware it runs onDay-0 support for DeepSeek V4 on eight Chinese chip platforms31
Soft powerOpen models travel to markets US firms do not serveDeepSeek usage estimated 2 to 4x higher in Africa than elsewhere32
Recruiting and respectOpen publication attracts talent and statusDeepSeek's Liang Wenfeng: "moats created by closed source are temporary"33

None of these readings excludes the others, and none requires a master plan. That matters for what comes next.

The Game

Economics has a long literature on what happens when a follower can copy a leader cheaply. In a survey of 129 lines of business published in 1987, Levin, Klevorick, Nelson and Winter found that imitating a major unpatented product typically took one to three years and cost 51 to 75% of the innovator's R&D. Lead time ranked far above patents as protection, second only to sales and service efforts.34

AI compresses both numbers. Model weights cannot be patented in any practical sense, outputs can be distilled, and the catch-up time is measured in months rather than years. Lead time, one of the strongest moats in the classic data, is exactly the moat that is shortest here.

Four frames help to read the resulting game. They are interpretations, not proofs, so each is tied to something observable.

FrameLogicWhat you would expect to seeWhat we see
Leader and follower with spilloversThe leader pays for the frontier; the follower pays a fraction to reach it a few months laterA stable lag, a large cost gapA 3 to 8 month lag (next section) on 1.5 to 2 orders of magnitude less training compute35
Commoditise your complement"Smart companies try to commoditize their products' complements" (Spolsky, 2002)36Firms that earn elsewhere release models; firms that sell models do notAlibaba, Nvidia and, until 2025, Meta release; pure model sellers mostly do not
War of attritionEach player accepts losses as long as falling behind looks worseSpending justified by the cost of being lateZuckerberg: "If we end up misspending a couple of hundred billion dollars, I think that that is going to be very unfortunate, obviously," with the risk of being late judged greater37
Red QueenRunning faster just to keep the same relative positionEfficiency gains spent on bigger models, not on marginsAmodei: efficiency gains "end up entirely devoted to training smarter models"38

The game has a second layer that is often missed. The pressure on US labs to spend comes first from each other. OpenAI, Anthropic, Google, Meta and xAI are in a classic prisoner's dilemma: each would be better off if all spent less, but none can afford to spend less alone. Chinese open models do not create that dilemma. They raise its stakes, because every month a US lab is not ahead is a month in which a free alternative is close.

The leader's own responses fit the pattern. In January 2025 Sam Altman said: "We will produce better models, but we will maintain less of a lead than we did in previous years." When OpenAI released its open-weight gpt-oss models in August 2025, he gave the reason: "if we didn't do it, the world was gonna be mostly built on Chinese open-source models".39 That is defensive open-sourcing by the leader, made in response to the follower.

Azoulay, Krieger and Nagaraj describe the structural role in a 2024 NBER paper: an open ecosystem often depends on a "rogue" technology giant that uses openness as a weapon against other incumbents, while the real moats sit in complementary assets such as compute, distribution and data.40 Meta played that role in 2023 and 2024. In 2025 and 2026 Chinese labs played it collectively.

The central point of this section: the equilibrium does not need anyone to have planned it. A set of firms responding rationally to their own constraints produces the same outcome as a strategy designed to make the leader's capex less valuable. Whether that was intended is unknowable from the outside. That it is happening can be measured.

One formal result is worth keeping in mind for later. Qiu, Laufer, Kleinberg and Heidari model the incentives to open or close a model and find that, in equilibrium, weaker models tend to be open and the most capable models closed.41 If that holds, open weights are a phase, not a permanent position. The 2026 evidence, discussed below, points that way.

Measuring the Half-Life

A half-life needs a clock. Three of them can be read from public data.

Clock one: the catch-up time. Epoch AI estimates that open-weight models trailed the closed frontier by about three months in October 2025 and by about four months on average in 2026, slightly more than in autumn 2025.42 Chinese models specifically have trailed the US frontier by about seven months on average since 2023, within a range of four to fourteen.2 Using Epoch's raw capability-index data, each new open-weight record since mid-2024 arrived between one and eight months after a closed model first reached the same score. The narrowest gap was DeepSeek-R1 in January 2025, at 1.1 months. In 2026 it has run between 4.4 and 8 months at release.35

CHART 02 · The catch-up time is stable, not shrinkingMONTHS BEHIND AT RELEASE
3 TO 8 MONTH BAND02468102024 · 072025 · 012025 · 072026 · 012026 · 07DeepSeek-V3 · 3.4DeepSeek-R1 · 1.1Kimi K2 Thinking · 6.7GLM-5.1 · 8Kimi K3 · 4.4
SOURCE · Own calculation from Epoch AI Capabilities Index (ECI) data, snapshot 19 October 2026. Each new open-weight record vs the first closed model with an equal or higher score. Filled = Chinese lab, hollow = US lab.

That finding contradicts a popular story. The gap is not closing towards zero. It is holding in a band, while the frontier itself moves faster.

The yardstick matters. On chat-preference rankings the US-China gap is about 2.4%, which is where the "gap has closed" headlines come from.43 On held-out and agentic tests it is large. In NIST's evaluation of DeepSeek V4 Pro, the model scored 44% on a held-out software-engineering test against 78% for GPT-5.5, and 46% against 79% on the semi-private ARC-AGI-2 set. On public benchmarks the same pair was within three to seven points. NIST put the model about eight months behind the frontier; DeepSeek's own results implied about two.44 NIST is a government body with a policy interest, and it reproduced DeepSeek's reported GPQA score, which rules out simple configuration errors. The difference lies mainly in what is being measured: public tests that labs can train towards, or tests they have never seen.

CHART 03 · The gap depends on the yardstickPP, BEST US VS BEST CHINESE
PUBLIC BENCHMARKNEW BENCHMARKHELD-OUT / NON-PUBLIC010203040Chat preference (Arena, Elo gap in %)2.4SWE-bench Verified (Epoch runs)4.8GPQA (CAISI, V4 Pro vs GPT-5.5)6SWE-bench Verified (CAISI)7FrontierCode (new, Oct 2026)10.4ARC-AGI-2 semi-private (CAISI)33PortBench, held-out SWE (CAISI)34CTF-Archive-Diamond, cyber (CAISI)39
SOURCE · NIST CAISI, Evaluation of DeepSeek V4 Pro (1 May 2026); Epoch AI benchmark hub (Oct 2026); Arena leaderboard (Oct 2026), own calculation.

The commercial reading is that the follower is closest exactly where the volume is (chat, routine coding, public tasks) and furthest where the premium is (long agentic work, held-out problems, security).

Clock two: the price of a fixed capability. Epoch AI measured how fast the price of reaching a given benchmark score falls. Depending on the threshold, the decline ranged from 9x to 900x per year, and about 40x per year for GPT-4-level performance on a science benchmark.45 Andreessen Horowitz put the general rate at 10x per year.46 Translated into half-lives: at 10x per year, the price of a given capability halves every 3.6 months; at 40x, every 2.3 months.

Clock three: the product cycle. In 2026 the US labs typically replaced their flagships every 60 to 90 days. Anthropic's Opus 5 was succeeded by Opus 5.5 after 60 days; OpenAI's GPT-6 Sol by GPT-6.1 Sol after seven.47

Set against these, the hardware runs on a different clock. Epoch AI finds that a chip stays in use for frontier training for a median of 3.9 years.48 Hyperscalers depreciate servers over five to six years. Meta moved to 5.5 years in 2025, adding $2.59 billion to net income; Amazon went the other way for part of its fleet, from six to five years, citing AI.49

CHART 04 · Two clocks that do not matchMONTHS, LOG SCALE
↑ MODEL CLOCK↓ HARDWARE CLOCK23612244896MONTHS (LOG SCALE)Price halving at fixed capability (10x–40x / yr)2.3–3.6 moFlagship replacement cadence, US labs 20262–3 moOpen-weight catch-up time3–8 moH100 rental price halving, 2024–2510.5 moMedian life of a chip in frontier training47 moServer depreciation schedule, hyperscalers60–72 mo
SOURCE · Epoch AI (LLM inference price trends, Mar 2025; GPU frontier lifespan); a16z (Nov 2024); Epoch ECI data; vendor changelogs; Silicon Data and SemiAnalysis; company 10-K filings. Own calculation.

Dario Amodei has described the economics with a useful image: "If you consider each model to be a company, the model that was trained in 2023 was profitable."50 The image holds, but the numbers above change what it implies. Each of those companies now has a commercial life measured in quarters, while the factory it runs in is written off over half a decade. The model is the shortest-lived asset in the system, and it does not appear on any balance sheet: as far as public disclosures show, training is expensed as R&D, not capitalised.51

A complication that cuts the other way. The price discount of Chinese models is shrinking at the top. In January 2025, R1 was 27 times cheaper than o1 on output. In October 2026, Moonshot's Kimi K3, the best open-weight model on Epoch's index, lists at $3 per million input tokens and $15 per million output, only 1.3 times below Anthropic's Claude Opus 5.5 at $4 and $20.52 Measured per completed task on the Artificial Analysis index, Alibaba's API-served Qwen3.8 Max costs about as much as Opus 5.5. The deep discounts now sit in the second tier: Xiaomi's MiMo-V2.6-Pro under an MIT licence is 23 times cheaper per output token than Opus 5.5.52 Meanwhile the US labs cut their own prices: the Opus line fell from $15 and $75 to $4 and $20 in under two years, and the cheapest US models now list below DeepSeek V4.1 Flash at peak hours.4753

The pressure on margins is therefore not only Chinese. Chinese open models set a public floor. Everyone, including the US labs themselves, is pushing prices towards it.

January 2025October 2026
Best Chinese model vs US flagship, output priceR1 vs o1: ~27x cheaperKimi K3 vs Opus 5.5: ~1.3x cheaper
Deepest second-tier discountn/aMiMo-V2.6-Pro vs Opus 5.5: ~23x cheaper
Licence of the best Chinese modelMITRestricted (commercial agreement above $20m revenue)
Catch-up time at release1.1 months (R1)4.4 months (Kimi K3)

SOURCE · Epoch AI; Artificial Analysis model pages (Oct 2026); own calculation

Stress Test I: The Labs

If frontier rents had a half-life of months, it should show up in the labs' numbers. So far, the top line says the opposite.

Anthropic's annualised revenue passed $65 billion at the end of July 2026, from about $9 billion at the end of 2025. Its revenue in the third quarter was $16.8 billion, and the FT reports a second consecutive quarter of adjusted operating profit.54 OpenAI's run-rate was about $40 billion at the end of July.54 Both figures need a caveat: Anthropic books spending through cloud partners gross and OpenAI books its share net, so they are not directly comparable.

Two observations explain why cheap Chinese models have not yet hit those numbers.

First, tokens and money have decoupled. Chinese open models dominate token counts on developer platforms. But Menlo Ventures found that Chinese models made up about 1% of enterprise LLM API usage at the end of 2025, and that the open-source share of enterprise usage fell from 19% to 11%.55 Epoch AI estimates that six leading Chinese AI companies together earn about 10% of what OpenAI and Anthropic earn.56 Volume migrates to the cheap tier; the revenue stays where trust, compliance and the hardest workloads are.

Second, the moat has moved from the model to the product. Claude Code, Codex and the agent harnesses built around them have become standard interfaces. DeepSeek's own documentation advertises compatibility with "Claude Code, and Codex".57 A model can be copied in months. A product embedded in a company's workflow, with security review and procurement behind it, has a longer half-life.

The stress is elsewhere: in the gap between what has been committed and what is earned.

CHART 05 · Long commitments, short half-livesUSD BN
COMPUTE COMMITMENTSANNUALISED REVENUE0100200300400500600700OpenAI600~$600BN COMPUTE SPEND TO 203040RUN-RATE, END JUL 2026Anthropic518≥$518BN, ~80% NON-CANCELLABLE67.2Q3 2026 REVENUE ×4
SOURCE · Anthropic draft S-1 via Reuters and Fortune (Sep 2026); FT via Yahoo Finance (Q3 2026); CNBC and Reuters (Feb 2026); Bloomberg via TechCrunch (Aug 2026).

According to its draft prospectus, Anthropic has committed at least $518 billion to six compute partners over roughly ten years, about 80% of it non-cancellable or payable regardless of use. One contract with Google states: "If our actual spend falls short, we must pay Google the difference."58 OpenAI's commitments, announced as about $1.4 trillion over eight years in November 2025, were recast in February 2026 as about $600 billion of compute spending through 2030.59

These are long, fixed obligations. They are serviced by revenue from products whose underlying models lose pricing power on a clock measured in months. The labs can absorb that as long as revenue keeps growing faster than prices fall. The signals to watch are the ones that show whether it does. Anthropic's month-on-month run-rate growth slowed from about 33% in the first seven months of 2026 to between 3% and 7% in August and September, according to an Investing.com analysis of reported figures.60 Two customers accounted for nearly a quarter of its 2025 revenue, and many of its largest customers have no long-term contracts.58

There is one more twist. The FT reported that Anthropic told prospective IPO investors that distillation, rival labs training on Claude's outputs, accounted for 5 to 7% of its revenue, and that it had cut this back materially in the third quarter.61 If that is accurate, the leader was for a time being paid by its followers to help them follow. Stopping it costs revenue. Allowing it shortens the half-life.

Stress Test II: The Hardware

For the hardware layer, there is a natural experiment. On 27 January 2025, after DeepSeek-R1, Nvidia's share price fell about 17% and the company lost roughly $589 billion of market value, the largest one-day loss on record. Broadcom fell 17.4%, Oracle 13.8%.62 Within ten days, every large hyperscaler confirmed or raised its capex plan, and Satya Nadella publicly invoked the Jevons paradox: cheaper AI means more AI.63 Nvidia reached a $4 trillion market value in July 2025, and its data-centre revenue for the fiscal year to January 2026 rose 68% to $193.7 billion.6264

The volume data support that reading. Google processed 9.7 trillion tokens a month in May 2024 and more than 3.2 quadrillion in May 2026.65 GPU rental prices tell the same story from the other side. An H100 cost about $8 an hour in early 2024 and $1.70 at the trough in October 2025. Then, as models proliferated and inference demand grew, the price rose again, to $2.84 on 19 October 2026.66

CHART 06 · The hardware clock ran backwards in 2026USD / H100 GPU-HOUR
$0$3$6$92024 · 012024 · 072025 · 012025 · 072026 · 012026 · 07~$8 · EARLY 2024$1.70 · TROUGH OCT 2025$2.84 · 19 OCT 2026−79% to the trough, then +67% while model prices kept falling.
SOURCE · Silicon Data; SemiAnalysis 1-year contract index (2 Apr 2026); Silicon Data SDH100RT index (19 Oct 2026). Series differ; indicative path only.

This is the asymmetry at the core of the stress test. Model rents compress far faster than hardware rents. As long as compute is scarce, most of the efficiency gain goes to the buyers of tokens, not to the owners of GPUs. Chinese open weights are, for now, a source of demand for Nvidia outside China. The risk factor in the opening paragraph is consistent with that reading.

China itself is a different matter, and it is already priced in. Nvidia's China revenue fell from 26% of the total in fiscal 2022 to 9% in fiscal 2026, on a changed accounting basis. Shipments of data-centre Hopper chips to China were below 1% of data-centre revenue in the latest quarter, and Nvidia's guidance has assumed zero China data-centre revenue since early 2026.67

CHART 07 · Nvidia's China business, before and after% OF NVIDIA REVENUE
BILLING-LOCATION BASIS (OLD)CUSTOMER-HQ BASIS (SINCE Q3 FY26)0%10%20%30%FY22FY23FY24FY25FY26Q1 FY27Q2 FY2726.4%21.4%16.9%13.1%20.2%19.2%9.1%5.6%8.2%DATA-CENTRE HOPPER TO CHINA:<1% OF DATA-CENTRE REVENUE
SOURCE · Nvidia Form 10-K FY2024, FY2025, FY2026; Form 10-Q Q1 and Q2 FY2027. China incl. Hong Kong. Basis changed from billing location to customer headquarters in Q3 FY2026. Shares are own calculation.

So where would a short model half-life hurt the hardware? Not through token demand, but through credit. The demand for chips is increasingly financed by the people who sell them and by the labs whose rents are under pressure:

  • Nvidia held about $99 billion of equity stakes at the end of July 2026, against about $40 billion in January.1
  • It guarantees up to $105 billion of lease obligations for an OpenAI campus in Ohio. The guarantees end when OpenAI reaches a "satisfactory credit rating".1
  • It has committed $36 billion to buying back cloud capacity from AI clouds.1
  • Oracle, which carries $664 billion of contracted backlog, had negative free cash flow of $5.4 billion in its latest quarter, funded partly by $19.9 billion of new equity and $11.4 billion of customer prepayments.68

The transmission is indirect. If frontier rents compress faster than revenue grows, the labs' ability to raise money weakens. With it go the prepayments, the take-or-pay renewals and the leases backstopped by Nvidia. Hardware orders could then fall even while token volumes keep rising. No such case has been observed yet. But Nvidia's own risk language, and the credit-rating trigger in its OpenAI guarantee, show that the chain is understood by the people who wrote the contracts.

One more observation: in the second quarter of 2026 SK hynix reported an operating margin of 76%.69 In this cycle, the rent has moved to the bottleneck input, memory, which is also China's binding constraint.

The Follower's Own Half-Life

It would be easy to read all this as a story in which China wins and the US pays. The evidence does not support that either.

The Chinese labs are losing money heavily. Zhipu, listed in Hong Kong since January 2026, reported revenue of RMB 724 million for 2025 and a net loss of RMB 4.72 billion, about 6.5 times revenue. MiniMax lost an adjusted $251 million on $79 million of revenue.70 Open weights also let others capture the margin: Epoch estimates that DeepSeek earns only about 8% of the OpenRouter spending on its own models; the rest goes to third-party hosts.56

And the followers are starting to close up. Alibaba moved its Qwen Max models behind a paid API in spring 2026. When it later announced that it would open the weights of Qwen3.8-Max, Reuters reported that it would seek a revenue share from large commercial users. Moonshot's Kimi K3 requires a commercial agreement for companies above $20 million in revenue.71 This is what the Qiu et al. model predicts: as capability rises, release becomes more closed. In October 2026, according to reporting based on the South China Morning Post and AP, China's commerce ministry published a draft that would put models above 10^26 FLOP or two trillion parameters on its technology export list, with fast-track clearance for open-source releases. Consultation closes on 15 November.72

History offers a warning for both sides. Chinese solar manufacturers won about 80% of global module production after polysilicon prices fell 94% between 2008 and 2012, and a generation of Western producers went bankrupt. In 2025 the Chinese leaders themselves guided to combined losses of about $5.5 billion.73 The follower took the market and much of the margin with it.

The analogy breaks in one important place. A solar factory cannot produce a competitor's panel, but a data centre can run a competitor's model. AWS hosts DeepSeek and Qwen; Microsoft hosts DeepSeek. In August 2026 Mistral, Europe's "sovereign" AI champion, began hosting Zhipu's GLM-5.2 on its own platform, even though Zhipu is on the US Entity List.74 The capex does not become stranded when models commoditise. What is at risk is the margin earned on top of it, and that risk applies to every lab, Chinese ones included.

What to Watch

  • 9 and 10 November 2026: the suspension of the US "Affiliates Rule" and China's pause on rare-earth export controls both expire unless extended.75
  • 15 November 2026: consultation on China's draft export controls for AI models closes. Whether open-source releases are really fast-tracked will show whether open weights remain state policy.72
  • Anthropic's public S-1: the first chance to see how a frontier lab accounts for training costs, how concentrated its customers are, and how much revenue came from distillation.
  • Licence terms of the next Chinese flagships: more revenue-share clauses would mean the follower is turning open weights into a funnel.
  • Epoch AI's next lag estimate: a move below three months would mean the half-life is shortening; above eight, that the frontier is pulling away.
  • Tokens vs spend: if enterprise spend on US labs flattens while token volumes keep rising, the price clock has overtaken the volume clock.
  • H100 and B200 rental indices and the first exchange-distributed GPU forward curves, which CME began publishing on 15 October 2026 and which slope downwards.76

Exposure Card

WhoChannelIf frontier rents halveVisibility
US frontier labsPricing power of each model generationGrowth must outrun price declines to service $500bn+ commitmentsMedium (S-1, leaks)
NvidiaEquity stakes, lease guarantees, capacity buybacksCredit exposure to the labs grows; token demand may not fallHigh (10-Q)
Oracle, neocloudsBacklog, prepayments, take-or-pay contractsRenewal and funding riskMedium
Memory makersHBM pricingLeast affected while supply is tightHigh
Chinese labsIPO and venture fundingLosses widen; pressure to close flagshipsMedium (HKEX filings)
Index saversMarket-cap weight of the names aboveSubject of Part 2High

This was Part 1 of the series "Follow the Compute Money". Four more parts will follow over the coming weeks.

Personal View

China's open-model strategy is, in my view, the rational play given the hand it was dealt, and the open ecosystem it has built is good for most of the people who use AI. Cheaper, downloadable models put capable tools in the hands of companies and countries that would never sign a $500 billion compute contract. That is worth saying first.

But don't forget the other half. A lead that lasts five to eight months is a flow, not a stock. It has to be earned again every quarter, and it is being financed with obligations that run for five to ten years. The labs have so far grown fast enough to outrun the price clock. The question is not whether they are ahead today. It is whether being ahead is still worth what it costs to stay there.

The reframe I find more useful: the race being run is "who has the best model". The race that probably decides the economics is "who owns the layer that does not commoditise". For the US labs that looks like products, distribution and trust. For Nvidia it is increasingly credit. For China it may be the hardware stack and the standards that its open models carry with them. Chasing the leader on the leader's own terms is rarely the clever move. China seems to have understood that early. Whether the US labs have is, to me, the open question.

Footnotes

  1. NVIDIA Corporation, Form 10-Q for the quarter ended 26 July 2026 (Q2 FY2027), SEC EDGAR. sec.gov. Risk factors (models originating in China; customers developing open models), PRC regulatory actions, equity stakes, OpenAI lease guarantees, cloud service agreements. ↩ ↩2 ↩3 ↩4 ↩5 ↩6

  2. Epoch AI, "US vs China: ECI", 2 January 2026. epoch.ai ↩ ↩2

  3. Michel Johannsen, "The Hyperscaler Capex Stress Test", 11 May 2026. ↩

  4. US Bureau of Industry and Security, press release on semiconductor and HBM controls, 2 December 2024. media.bis.gov. Earlier rules of October 2022 and October 2023 as summarised there. ↩

  5. NVIDIA Corporation, Form 8-K, 9 April 2025, sec.gov ; NVIDIA Q1 FY2026 results release, sec.gov ↩

  6. Covington & Burling, "US Commerce Department revises license review policy for exports of certain advanced computing commodities to China and Macau", January 2026. cov.com ; NPR, 11 August 2025, npr.org ↩

  7. Epoch AI, "Introducing the AI Chip Owners Explorer", 6 April 2026. epoch.ai. Excludes smuggled chips and offshore cloud. ↩

  8. Epoch AI, "AI supercomputers: performance share by country", epoch.ai ; Al Jazeera, 24 September 2026, citing Epoch, aljazeera.com ↩

  9. Epoch AI, "Huawei's roadmap to 2031", 4 September 2026. epoch.ai ↩

  10. TechNode, citing IDC via Reuters, 14 October 2026. technode.com. Units, not compute. ↩ ↩2

  11. pv magazine, "China adds 315 GW of solar in 2025", 29 January 2026, pv-magazine-india.com ; US EIA, Today in Energy, eia.gov. Ratio is own calculation (434 / 53). Utilisation estimate via Al Jazeera, 28 May 2026, aljazeera.com ↩

  12. Global Trade Alert on Big Fund III, globaltradealert.org ; SCMP on the RMB 60bn AI fund, scmp.com ; Shanghai Municipal Government, compute vouchers, 29 July 2025, english.shanghai.gov.cn ; TechRadar citing FT on power discounts (not officially confirmed), 5 November 2025, techradar.com ↩

  13. Reuters factbox via Malay Mail, 29 June 2025. malaymail.com ; Inside Government Contracts on US federal and state restrictions, February 2025, insidegovernmentcontracts.com ↩

  14. State Council, Opinions on Deepening the Implementation of the "AI+" Action (Guofa [2025] No. 11), 26 August 2025. English translation by CSET: cset.georgetown.edu ↩

  15. Xinhua, text of the CPC Central Committee recommendations for the 15th Five-Year Plan as quoted, 6 December 2025. news.cn. Translation own. ↩

  16. Ministry of Foreign Affairs of the PRC, Global AI Governance Action Plan, 26 July 2025. mfa.gov.cn ↩

  17. DeepSeek-AI, "DeepSeek-V3 Technical Report", arXiv 2412.19437. arxiv.org ; SemiAnalysis, "DeepSeek Debates", 31 January 2025, newsletter.semianalysis.com (estimate). ↩

  18. VentureBeat on DeepSeek V3.2-Exp pricing, September 2025. venturebeat.com ↩

  19. The Register, "DeepSeek V3.1 hints at Chinese chips", 22 August 2025. theregister.com ↩

  20. Huawei, Eric Xu keynote at Huawei Connect 2025, September 2025. huawei.com ; RCR Wireless, 22 September 2025, rcrwireless.com ↩

  21. Nathan Lambert and Brand, "The ATOM Report", arXiv 2604.07190 (April 2026). arxiv.org ↩

  22. Hugging Face, "State of Open Source on Hugging Face: Spring 2026", 17 March 2026. huggingface.co ↩

  23. Fortune, "China's AI price war", 22 May 2024. fortune.com ↩

  24. Epoch AI, Gradient Updates, "What went into training DeepSeek-R1?", 31 January 2025. epoch.ai. Output prices $60 vs about $2.2 per million tokens. ↩

  25. Malay Mail citing the Financial Times, 17 September 2025. malaymail.com ↩

  26. Tom's Hardware, citing the FT, 19 August 2026. tomshardware.com. Share delivered is own calculation (about 20,000 of 400,000+). ↩

  27. Anthropic, "Detecting and preventing distillation attacks", 23 February 2026. anthropic.com. Shares by lab are own calculation from the stated exchange counts. ↩

  28. Rest of World, on OpenAI's memo to the House Select Committee on the CCP (12 February 2026). restofworld.org ↩

  29. MiniMax, "MiniMax announces full-year 2025 financial results". minimax.io ↩

  30. TechNode, "Alibaba pitches Qwen as Android of the AI era", 24 September 2025. technode.com ↩

  31. TrendForce, 29 April 2026. trendforce.com ↩

  32. Microsoft On the Issues, "Global AI adoption in 2025", 8 January 2026, blogs.microsoft.com ; Capacity, 13 January 2026, capacityglobal.com ↩

  33. ChinaTalk, translated interview with DeepSeek CEO Liang Wenfeng (July 2024). chinatalk.media ↩

  34. Levin, Klevorick, Nelson and Winter, "Appropriating the Returns from Industrial Research and Development", Brookings Papers on Economic Activity 1987(3). brookings.edu ↩

  35. Epoch AI, Capabilities and benchmarking data (benchmark_data.zip) and AI models dataset (all_ai_models.csv), snapshot 19 October 2026, CC-BY. epoch.ai. Lag series and compute ratios are own calculation. ↩ ↩2

  36. Gwern Branwen, "Laws of Tech: Commoditize Your Complement", quoting Joel Spolsky, "Strategy Letter V" (2002). gwern.net ↩

  37. Sherwood News, on Mark Zuckerberg's interview with Alex Heath (Access podcast), 19 September 2025. sherwood.news ↩

  38. Dario Amodei, "On DeepSeek and Export Controls", January 2025. darioamodei.com ↩

  39. VentureBeat on Sam Altman's Reddit AMA, 31 January 2025, venturebeat.com ; AOL, 19 August 2025, aol.com ↩

  40. Azoulay, Krieger and Nagaraj, "Old Moats for New Models", NBER Working Paper 32474 (2024). nber.org ↩

  41. Qiu, Laufer, Kleinberg and Heidari, "A Formal Model of the Economic Impacts of AI Openness Regulation", arXiv 2507.14193 (July 2025). arxiv.org ↩

  42. Epoch AI, "Open-weights vs closed-weights models" (30 Oct 2025) and "The gap between open and closed models" (29 May 2026). epoch.ai ; epoch.ai ↩

  43. Stanford HAI, AI Index Report 2026. hai.stanford.edu ; Arena leaderboard, October 2026, arena.ai (own calculation). ↩

  44. NIST Center for AI Standards and Innovation, "CAISI Evaluation of DeepSeek V4 Pro", 1 May 2026. nist.gov ↩

  45. Epoch AI, "LLM inference prices have fallen rapidly but unequally across tasks", 12 March 2025. epoch.ai. Half-lives are own calculation (12 x ln2 / ln(rate)). ↩

  46. Guido Appenzeller, a16z, "Welcome to LLMflation", 12 November 2024. a16z.com ↩

  47. Digital Applied, AI model API pricing tracker Q3 2026. digitalapplied.com ; Anthropic pricing documentation, platform.claude.com ; OpenAI API pricing, openai.com ↩ ↩2

  48. Epoch AI, "How long are GPUs used for frontier training?" epoch.ai ↩

  49. Company 10-K filings as summarised by GPU Insights, "Server useful life and GPU depreciation 2026". gpuinsights.net ↩

  50. Cheeky Pint podcast (Stripe), John Collison with Dario Amodei, released 6 August 2025, at about 16:56. youtu.be ↩

  51. Epoch AI, "OpenAI compute spend", epoch.ai ; Quartz on OpenAI's 2025 audited figures (FT-verified), qz.com. No filing reviewed shows capitalised model weights. ↩

  52. Artificial Analysis model pages and leaderboard, accessed 19 October 2026: artificialanalysis.ai ; artificialanalysis.ai ; artificialanalysis.ai ; artificialanalysis.ai. Ratios are own calculation. ↩ ↩2

  53. DeepSeek API documentation, Models and Pricing (accessed 19 October 2026). api-docs.deepseek.com ↩

  54. TechCrunch, "Anthropic's annualized revenue surges to $65B", 17 August 2026, techcrunch.com ; Yahoo Finance citing the FT on Q3 2026 revenue, finance.yahoo.com ↩ ↩2

  55. Menlo Ventures, "2025: The State of Generative AI in the Enterprise", 9 December 2025. menlovc.com ↩

  56. Epoch AI, "How do Chinese AI companies make money?", 6 October 2026. epoch.ai ↩ ↩2

  57. DeepSeek API documentation (accessed 19 October 2026). api-docs.deepseek.com ↩

  58. Fortune on Anthropic's draft IPO prospectus, 29 September 2026, fortune.com ; The Next Web, thenextweb.com ↩ ↩2

  59. Yahoo Finance citing CNBC and Reuters, 20 February 2026. finance.yahoo.com ; TechCrunch, 6 November 2025, techcrunch.com ↩

  60. Investing.com analysis, "Anthropic Q3: spectacular numbers, harder questions". investing.com (secondary). ↩

  61. Simon Willison, summarising the Financial Times, 15 October 2026. simonwillison.net (secondary; FT report based on four sources). ↩

  62. Bloomberg, 27 January 2025, bloomberg.com ; Reuters via The Daily Herald, thedailyherald.sx ; Bloomberg Law on the $4 trillion milestone, news.bloomberglaw.com ↩ ↩2

  63. TechCrunch, 6 February 2025. techcrunch.com ↩

  64. NVIDIA Q4 FY2026 results, Form 8-K. sec.gov ↩

  65. Google, Sundar Pichai at I/O 2026. blog.google ↩

  66. Silicon Data, "H100 price spike", silicondata.com ; SemiAnalysis, 2 April 2026, newsletter.semianalysis.com ; Silicon Data H100 index, silicondata.com. Series differ; indicative path. ↩

  67. NVIDIA Form 10-K FY2025, sec.gov ; Form 10-K FY2026, sec.gov ; Form 10-Q Q2 FY2027 (as in note 1). Shares are own calculation. ↩

  68. Oracle Corporation, Form 10-Q for the quarter ended 31 August 2026. sec.gov ↩

  69. Pulse 2.0, "SK hynix reports record Q2 2026 revenue", pulse2.com. Margin is own calculation (60.5 / 79.3). ↩

  70. The Standard (Hong Kong) on Zhipu FY2025 results, thestandard.com.hk ; MiniMax FY2025 results (as above). ↩

  71. Implicator, on Alibaba's closed releases, implicator.ai ; ProPakistani citing Reuters, 8 August 2026, propakistani.pk ; Hugging Face, "State of Open Models: Summer 2026", huggingface.co ↩

  72. Business Standard, 19 October 2026, citing SCMP and AP. business-standard.com. Primary MOFCOM text not yet reviewed. ↩ ↩2

  73. IEA, Solar PV Global Supply Chains, iea.org ; ORF on polysilicon prices, orfonline.org ; Techloy citing Nikkei, 21 January 2026, techloy.com ↩

  74. Techzine, "Mistral to host external models, Chinese Z.ai to be first", techzine.eu ; Mistral docs, docs.mistral.ai ; AWS, aws.amazon.com ↩

  75. Jones Day, "BIS suspends Affiliates Rule; China pauses rare earth export controls", November 2025. jonesday.com ↩

  76. Crypto Briefing, 15 October 2026. cryptobriefing.com (secondary; not yet confirmed on cmegroup.com). ↩

END · 07