China's AI Compute Will Double on America's Own Chips
China's AI Compute Will Double on America's Own Chips
By Daniel Rao | October 24, 2026
Here is the paradox at the heart of America's AI race with China, stated as plainly as a new report from the Institute for Progress (IFP) allows: the United States is winning the chip war and losing the plot at the same time. China's access to AI computing power is set to roughly double in a single year — from about 4.6 million to 8.8 million H100-equivalents by the end of 2027, IFP estimates — and about 65% of that doubled capacity will run on US-designed chips. Washington spent four years building a fence around its best silicon. China is about to double its compute anyway, largely by renting the fence's own hardware, from clouds, in countries the fence was never built to cover. Export controls are not failing. They are leaking, through the cloud, at industrial scale — and that leak is now the single most important fact in the AI race.
We read the full IFP report, not just the South China Morning Post's summary of it, and cross-checked every figure below against its executive summary and methodology sections before writing this. That matters, because the report's numbers are estimates with wide uncertainty ranges, and its headline claim — that a package of policies using only existing executive-branch powers could roughly double America's AI lead by the end of 2027 — rests on a model the authors themselves call simplified. We report the model as the authors present it, with its ranges intact, and we label every IFP figure as IFP's. The report, "How to Increase America's AI Lead Over China," was published on October 19, 2026, updated October 21, and reported by the SCMP on October 24. Its authors are IFP president Tim Fist, technology fellow Georgia Adamson, and Konstantin Pilz of the Center for Technology & Statecraft.
The headline number, and why it is worse than it looks
Start with the number itself. IFP estimates China will have access to around 4.6 million H100-equivalents (H100-eq) of AI compute by the end of 2026, and 8.8 million by the end of 2027. An H100-equivalent is a yardstick, not a headcount: it measures computing capacity against Nvidia's H100 data-centre GPU, so a warehouse of weaker chips and a cluster of stronger ones can both be expressed in the same unit. IFP built its estimate from four channels: chips China makes itself, chips it imports legally, chips smuggled in, and — the channel that dominates the story — US chips in data centres outside China that Chinese entities access remotely.
The split is what turns a big number into a scandal. About 35% of China's expected year-end 2027 stock is already inside China, IFP says. Of the remaining 65%, about half comes from remote access — capacity that, in IFP's words, "can be almost entirely restricted today with new export controls" from the Bureau of Industry and Security (BIS). Read that again slowly. Roughly a third of China's future AI compute sits in foreign data centres, on American-designed silicon, available to rent, and closable by a regulator that has not yet closed it. The fence is not short. It has a gate, and the gate is a rental counter.
The remote-access figure itself is doing most of the doubling. IFP puts remote access at roughly 1 million H100-eq by the end of 2026, rising to around 2.8 million by the end of 2027 — with a stated range of about 680,000 to 1.46 million this year and 1.97 to 3.91 million next year, so treat the 2.8 million as a central estimate, not a fact carved in stone. Those data centres are heavily concentrated in Southeast Asia, IFP says — around 75% of the total. The report breaks the remote figure into two parts: clusters it can identify directly (about 280,000 H100-eq in 2026, rising to about 861,000 in 2027), including what it describes as ByteDance's Blackwell cluster in Norway and Tencent's 100,000-GPU lease with Oracle, and a larger inferred share from Chinese hyperscalers' overseas data-centre leasing (about 775,000 in 2026 and about 1.97 million in 2027), estimated from SemiAnalysis leasing data. The second part is inference, and IFP is honest about the assumptions baked into it — that leased capacity is operational, that its AI share matches regional patterns, and that Chinese hyperscalers will use it to access export-controlled GPUs. A sceptical reader should hold that inference at arm's length. But even the directly identified clusters alone are a large, embarrassing number, and the direction of travel is not in doubt.
The cloud loophole is legal — that is the problem
Why is any of this legal? Because American export law, as written over the last two decades, largely treats renting compute as something other than exporting a chip. The cloud provider is not shipping a GPU to China; it is making capacity available, often in a third country, to a customer who logs in from somewhere else. A companion report published on Tuesday, October 20, by the Institute for AI Policy and Strategy (IAPS) makes the same point from the other direction: US-designed compute remains legally available for rent to foreign entities that are otherwise barred from buying the hardware. IAPS urged BIS to restrict Chinese access to cloud capacity large enough to train frontier AI models or run them at scale, and to strengthen checks on customer identity and ownership. Its escalation framework runs in four levels, from reporting requirements, to banning remote access at or above frontier-scale thresholds, to banning all remote access with enhanced know-your-customer checks for flagged customers — and finally for all customers. IAPS separately estimates that rented US compute adds at least 60% to the compute China acquires in 2026, relative to what chip export controls alone would permit. Two think tanks, two methods, one conclusion: the rental market is now a first-order channel of the AI race, not a footnote to it.
IFP's own prescription is blunter than IAPS's ladder. Its authors write: "The most effective channel to limit is remote access, given its outsized impact on the US-China AI capabilities gap by the end of 2027." The report lays out three viable BIS pathways using existing authorities — a licence condition on chip exports, a new export control on frontier model weights, and a new remote-access end-use control — and models a world in which BIS restricts Chinese access to all overseas clusters large enough to support frontier training. In that modelled world, China effectively loses all of the remotely accessed compute that affects the US lead, while keeping access for other purposes such as running AI applications. Alongside the remote-access control, IFP's package includes a narrow antitrust safe harbour so US companies can coordinate against distillation, strengthened due-diligence rules to fight smuggling, updated controls on chips still legally exportable to China, and new controls and investigations aimed at foreign high-bandwidth memory (HBM) smuggling. The package, IFP estimates, would buy America 6.6 extra months of lead by the end of 2027 — roughly doubling today's lead. In an extreme hypothetical where China's compute froze at today's level and distillation stopped entirely, the gain would be 11.4 months. Models are not promises, and IFP says so. But the asymmetry is the point: the cheapest months of lead America can buy are the ones it is currently giving away through the rental market.
Distillation: the other half of the gap
Compute is only half the report. The other half is distillation — training a model on the outputs of a better model, so that a Chinese lab can acquire reasoning, coding and tool-use capabilities at a fraction of the cost America paid to create them. IFP's model is striking in its symmetry: America's roughly 16-fold advantage in AI training compute would, on compute alone, buy about a 13-month lead. Distillation claws back about half of it, leaving the observed gap of roughly 6.5 months between top American and top Chinese models, measured on the Epoch Capabilities Index and cross-checked against software-engineering "time horizon" measures (about 6.6 months). Put another way, IFP says distillation gives China the equivalent of having four times more physical compute than it actually has. And the crueller the logic, the worse it gets: the bigger America's compute advantage grows, the more powerful the models Chinese labs can distil from, and the more effective distillation becomes as a catch-up strategy. Building a bigger lead without plugging distillation, IFP argues, is like pouring water into a bucket with a hole in it — the bucket still fills, just slower than it should.
China, for its part, has previously rejected US allegations of "aggressive, malicious" AI distillation by its developers as groundless, the SCMP notes — and it is worth keeping the two claims distinct. Ordinary distillation is a standard, widely used machine-learning technique; what IFP and the US government allege is adversarial distillation at industrial scale, including what IFP describes as fraudulent accounts operated by Chinese entities interacting with US models through commercial APIs. IFP says US companies and the US government have publicly documented multiple instances of top Chinese AI companies doing this at scale. Its proposed fixes are correspondingly practical rather than rhetorical: a joint Justice Department and Federal Trade Commission letter confirming that sharing threat information and defensive techniques does not violate antitrust law, and sanctions — under existing cyber and trade-secret authorities — against a handful of key actors, to make distillation carry financial risk. IFP roughly estimates that success on the coordination front could cut harvested distillation data by 95%, which matters because each tenfold increase in distilled data volume adds roughly the same capability gain. Half-measures, in other words, barely move the needle; only near-total disruption does.
Why "this won't hurt US sales" is the argument to watch
The objection every export-control debate eventually meets is commercial: won't restricting sales hurt American chipmakers? IFP's answer is the report's most politically useful finding, and the one most likely to be quoted in Washington. In a supply-constrained market, it argues, a chip that cannot be sold to a Chinese customer will almost certainly be bought by someone else. All major suppliers of the advanced memory (HBM) that US AI chips depend on report capacity booked out through 2027, IFP says, with chip designers receiving only 60 to 70% of their requests, and Nvidia told investors in August that it expects supply constraints through at least fiscal year 2028. AWS has said it will still not have enough capacity to meet all its demand in 2026 and expects the same in 2027; Microsoft has described an "extreme moment" of customer demand; Google has said its cloud revenue growth would have been higher with more compute. If demand outside China outstrips supply through at least the end of 2027, then restrictions redirect sales — largely to US customers — rather than destroying them. That claim, if accepted, removes the industry's best argument against action. It is also, note carefully, a claim about 2027, not forever: it holds only while scarcity holds, and scarcity is a market condition, not a law of nature.
The smuggling channel, meanwhile, is where IFP's ranges get widest and its honesty is most visible. Smuggled chips account for an estimated 1 million H100-eq by the end of 2026, rising to 1.6 million by the end of 2027 — but the ranges run from 381,000 to 2.9 million this year and 468,000 to 6.7 million next year. That is not a measurement; it is a confession of how little anyone can see into the grey market. IFP notes one encouraging precedent: in July 2026, the Financial Times reported that Nvidia, under pressure from Washington, intensified due diligence in Singapore, Malaysia and Japan — visiting data centres, verifying contracts, interviewing end users — and excluded more than half of its previous customers in those markets, many of them neoclouds. Industry insiders, IFP says, credit that crackdown with contributing to a shortage of AI chips in China. But the screening is voluntary, applies only to Nvidia's direct customers, and is not required by regulation. IFP's fix is to make it mandatory and extend it to server makers and distributors. A voluntary crackdown that works is still a policy failure if it can be unwound by the next quarter's sales targets.
China is not standing still — and that is the other half of the race
None of this means China is helpless without American silicon, and any article that implied so would be misleading. The SCMP's same-day coverage adds the other side of the ledger: Huawei said last month that its Ascend 960DT chip, designed for AI training, will become available in the first quarter of 2027 — three quarters ahead of its original schedule — and that more than 1,000 of its Atlas 900 A3 SuperPoDs have already been deployed, with its Atlas 950 system in large-scale commercial use. Washington, meanwhile, moved in January to allow case-by-case export reviews for Nvidia's H200, AMD's MI325X and comparable processors under security conditions, including requirements that sales not divert supply from US customers and that Chinese buyers adopt formal export-compliance procedures and submit chips for independent testing in the US. And Epoch AI, in an updated report last month, identified HBM as the main constraint on Huawei's output, estimating that Huawei's chip output this year would deliver less than 4% of the computing power of Nvidia's global shipments. The picture is not "China is winning" or "China is losing." It is a race in which one side has a 16-fold compute advantage and a 6.5-month capability lead, and the other side is doubling its compute, building domestic alternatives on an accelerated schedule, and renting whatever it cannot buy or build. That is a race, not a rout — which is exactly why IFP's 6.6 months matter.
Jensen Huang, Nvidia's chief executive, has been warning about this dynamic in unusually blunt language for months. In a podcast aired in April, ahead of the launch of DeepSeek's V4 foundation model, he said it would be "a horrible outcome" for the US if DeepSeek optimised its models to be trained and run on Huawei chips, claiming that China "will become superior to" the US as a result. SemiAnalysis, in its own recent analysis cited by the SCMP, went further: "Jensen's nightmare is starting to come true," its analysts wrote, as DeepSeek's testing on Huawei's Ascend hardware and software ecosystem narrows the gap between the two companies' tooling. Huang is not a neutral observer — Nvidia's China revenue is directly at stake in every control decision — but a warning is not less true because the warner profits from heeding it. When the CEO of the company whose chips define the yardstick says the ecosystem gap is closing, Washington should listen twice.
Our take: a fence with a rental counter is not a fence
Here is our judgment, stated plainly. IFP's report is not a counsel of despair, and it should not be read as one. Its central message is optimistic in a very specific, actionable way: America already holds the tools to roughly double its lead within a year, using authorities the executive branch already has, without new legislation and — on IFP's supply analysis — without costing American chipmakers their sales. The reason that message lands as an indictment rather than a reassurance is that the tools have not been used. The remote-access channel has been visible for years. The distillation campaigns have been publicly documented. The due-diligence crackdown that worked was voluntary. A policy establishment that can see all three facts and still leave the rental counter open is not being outmanoeuvred by Beijing; it is being out-waited by it, while every month of delay converts into Chinese training runs.
The honest caveats cut both ways, and we will not hide them. IFP's numbers are modelled estimates with wide ranges, especially on smuggling; its 6.6-month gain assumes policies are implemented "today" and work as modelled; and its claim that restrictions will not reduce US sales depends on scarcity lasting through 2027. Huawei's accelerated Ascend schedule is a reminder that export controls buy time, not victory — time that must be spent building domestic capacity, not congratulating oneself. But caveats are not an excuse for paralysis, and they are certainly not an excuse for the current arrangement, in which the world's most advanced AI hardware is simultaneously America's greatest strategic asset and, via a cloud login, China's fastest-growing compute source. Close the rental loophole, make due diligence mandatory, and give US labs the antitrust cover to fight distillation together. That is IFP's package, and on the evidence in its own report, it is the cheapest 6.6 months of national advantage Washington will ever buy. The alternative is to keep publishing reports about a doubling that everyone saw coming, one H100-equivalent at a time.
Frequently asked questions
What is an H100-equivalent?
A unit for comparing AI computing capacity against Nvidia's H100 data-centre GPU. IFP uses it so that different chips — domestic Chinese designs, legally imported US chips, smuggled chips and rented cloud capacity — can be added up in one total.
How much of China's AI compute is actually inside China?
About 35% of China's expected year-end 2027 stock, according to IFP. Of the remaining 65%, about half comes from remote access to data centres outside China — the channel IFP says BIS could almost entirely restrict with new export controls.
Is renting cloud compute to Chinese firms illegal?
Generally not under current US export rules, according to both IFP and the companion IAPS report — that is precisely the loophole both reports urge BIS to close, including through a new remote-access end-use control.
Does IFP say export controls will hurt Nvidia and other US chipmakers?
No — IFP argues the opposite for the near term: with demand outside China outstripping supply through at least 2027, restricted sales would largely be redirected to other buyers, including US customers, rather than lost.
For the hardware side of this story, see our posts on CXMT's memory leap that moved markets, Huang's warning on cutting off chip supply, and TSMC's $64B bet on AI chips.
Sources: Institute for Progress, "How to Increase America's AI Lead Over China" (published October 19, 2026, updated October 21, 2026; authors Tim Fist, Georgia Adamson, Konstantin Pilz); South China Morning Post (October 24, 2026); Institute for AI Policy and Strategy report on cloud compute (October 20, 2026) via SCMP and IAPS's own announcement; SemiAnalysis via SCMP (October 2026); Epoch AI (September 2026 update) via SCMP; Financial Times (July 2026) via IFP.
By Daniel Rao
Daniel Rao covers AI infrastructure and the business of compute for TechPulse USA — the chips, the clouds, and the controls deciding who gets to train the future.
Comments
Post a Comment