US export controls on advanced AI chips are binding and enforced, unlike Washington's voluntary framework for reviewing frontier models, but they restrain China's access to computing power rather than the pace of the frontier, and since 2025 parts of them have been relaxed. The controls date from October 2022 and were widened in 2023 and 2024. A global "AI diffusion" rule issued in January 2025 was dropped four months later, though its formal rescission had still not been published by mid-October 2026. Nvidia's H200 can now be licensed for China, but Nvidia says Beijing has restricted purchases and shipments came to under 1% of its data-centre revenue in its latest reported quarter. An Epoch AI analysis in January 2026 found that every frontier model since 2023 had been American, with Chinese models about seven months behind on average.
Four years of rules, then a partial reversal
The October 2022 rule required licences to export chips above set performance thresholds to China, along with some chipmaking equipment. Its stated reasons were national security and human rights: the Bureau of Industry and Security (BIS) cited weapons design, military decision-making and the surveillance of citizens. AI safety was not among them. A year later BIS said other chips could provide "nearly comparable AI model training capability", and the October 2023 rule tightened the thresholds and extended licensing to more countries and to Chinese-headquartered companies abroad. The December 2024 package added high-bandwidth memory and more manufacturing equipment.
The January 2025 diffusion framework went further, with worldwide licensing for advanced chips and a new control on the weights of closed models trained with at least 1026 operations. On 13 May 2025, two days before companies had to comply, BIS announced it would rescind the rule and stopped enforcing it. The formal rescission and a promised replacement were still listed as planned in the regulatory agenda published in August 2026. An older requirement survives: under guidance of 31 May 2026, advanced chips sold to companies headquartered in China or other arms-embargoed countries need a licence wherever the buyer is.
In April 2025 the government imposed a licence requirement on exports of Nvidia's H20 to China, then granted licences from August 2025, according to Nvidia's latest quarterly report. A January 2026 rule moved the H200 and equivalent chips from a presumption of denial to case-by-case review, on conditions that include capping China shipments at 50% of US shipments of the same product and independent testing in the US; more capable chips are still presumed denied. The same report, covering the period to 26 July 2026, says the Chinese government restricted sales of both chips and that H200 shipments were under 1% of Nvidia's data-centre revenue in the quarter.
Compute is the governable input, but its chokepoints span several countries
The case for chips as a lever is set out in a February 2024 paper on computing power and AI governance by researchers from OpenAI, the Centre for the Governance of AI, Cambridge, Mila and elsewhere. Compared with data or algorithms, they argue, compute is "detectable, excludable, and quantifiable", and its supply chain is extremely concentrated: in their account, TSMC makes the large majority of the most advanced AI chips, using extreme-ultraviolet lithography machines that only ASML builds. Concentration makes enforcement possible, but not unilateral. ASML is Dutch, and the Netherlands has required its own licences for advanced deep-ultraviolet lithography equipment since September 2023. Japan added advanced lithography and other chipmaking equipment to its own control list in July 2023, as the December 2024 US rule notes. Washington's July 2025 AI Action Plan still asks allies to follow US controls and "not backfill".
What a compute overhang is
A compute overhang, also called a hardware overhang, is a situation in which far more computing power is available than current AI software can use, so that when the missing software arrives, capability can jump rather than climb. Luke Muehlhauser and Anna Salamon described it in a 2012 chapter as the danger in a software bottleneck: when advanced AI is finally built, there may be "large amounts of inexpensive computing power" ready to run thousands of copies. AI Impacts defines it the same way. The overhang is the surplus hardware; an algorithmic advance is what releases it.
The idea has been used to argue for moving sooner. OpenAI's February 2023 post "Planning for AGI and beyond" said shorter timelines seemed more likely to bring a slower takeoff "due to less of a compute overhang". On that logic, hardware that piles up while development is held back fuels a sudden jump later, the fast-takeoff worry covered in our piece on hard and soft takeoff.
Algorithms also change what a fixed stock of chips can do. A study led by Epoch AI researchers found that the compute needed to reach a given language-model performance halved roughly every eight months between 2012 and 2023, although bigger training runs contributed even more to progress. The compute-governance paper concludes that any fixed capability gets cheaper to reach each year, so controls pegged to it erode, while, as long as scaling pays off, the frontier keeps demanding more. DeepSeek's V3 model, which the company reported as comparable to leading closed models, was trained on 2,048 Nvidia H800 chips for 2.788 million GPU hours, excluding earlier research runs. Anthropic chief executive Dario Amodei, whose company backs the controls, argued in January 2025 that this was "an expected point on an ongoing cost reduction curve" and did not show that the controls had failed.
What the controls have bought
Measured in hardware, a lot. Ege Erdil, then at Epoch AI, estimated in December 2024, before H200 sales to China were allowed, that the controls gave the US a lead of around four years in training price-performance but "essentially no lead" in serving models to users. Measured in capability, much less. Epoch's January 2026 analysis found that Chinese models trailed the American frontier by seven months on average (between four and 14 months), a gap it says resembles the one between closed and open-weight models generally. The figures measure different things, but together suggest a large hardware handicap became a much smaller capability gap. Algorithmic efficiency, stockpiled and smuggled chips, and distillation from American models are all candidate explanations.
Huawei says it is constrained. Its rotating chairman, Eric Xu, repeated at the company's September 2025 conference that mainland China "will lag behind" in chip-manufacturing process nodes "for a relatively long time", and said Huawei would compensate by combining chips into larger systems. Jeffrey Kessler, the Commerce under secretary who runs BIS, told a House subcommittee in June 2025 that the government assessed Huawei's Ascend production capacity for 2025 at or below 200,000 chips, while warning that China was "catching up quickly".
The controls also leak. The 2023 rule itself cited Chinese companies buying through foreign subsidiaries and operating data centres outside China, and at the 2025 hearing Kessler called chip smuggling "a fact". His July 2026 testimony reported BIS penalties rising from $16 million in 2024 to $324 million in 2025, and cited an indictment over a multi-billion-dollar scheme to divert servers with advanced chips through Southeast Asia. The Remote Access Security Act, which would bring cloud access to controlled items under export controls, passed the House 369 to 22 in January 2026 and has sat in a Senate committee since. Capability can also travel without hardware. Anthropic said in February 2026 that three Chinese labs had distilled Claude at scale, arguing that this undermines the controls and, because it needs chips, also strengthens the case for them. We examined its later evidence in our piece on distillation and safeguards.
The safety case, for and against
The main safety argument for the controls is about margin. Leopold Aschenbrenner, a former OpenAI researcher, argued in his 2024 essay series Situational Awareness that a healthy lead lets the leader spend part of it on getting safety right, and that "if we have only a 2 month lead, we have no margin at all for safety." A March 2025 paper by Dan Hendrycks, Eric Schmidt and Alexandr Wang makes a non-proliferation case: treat advanced chips like enriched uranium and keep them away from rogue actors.
Against that, chip controls do not limit the compute American labs use, and American labs built every frontier model in Epoch's data, so as a brake on frontier progress they act on the follower; the pace is set by the race among US companies. Stephen Cave and Seán Ó hÉigeartaigh warned in a 2018 paper that race rhetoric and actual races risk incentivising corner-cutting on safety and raising the risk of conflict. The compute-governance paper notes that cutting China off strengthens its incentive to build a supply chain free of US involvement, which could weaken the lever itself. Nvidia, an interested party, says in its quarterly report that being shut out of China's data-centre market helped its competitors build larger ecosystems. CSIS analysts argued in April 2025 that without matching industrial investment, controls are "at best a short-term palliative". And by Aschenbrenner's yardstick, seven months is far closer to the two months he called no margin than to the two years he called healthy.
The overhang logic also runs in reverse. Where hardware rather than software is the bottleneck, methods refined under scarcity can sit waiting for chips, and a sudden loosening, such as the January 2026 H200 policy, could produce a jump of its own. That is our inference from the concept, not a documented event, and so far, with Beijing restricting purchases, H200 shipments have been small.
When an export control pulled a frontier model
In June 2026 an export control did act, briefly, as a brake on an American lab. On 12 June, three days after the models were released, the Commerce Department applied export controls to Anthropic's Claude Fable 5 and Claude Mythos 5, requiring a licence for any foreign national to use them. According to Anthropic, the order followed a report by Amazon researchers of a way around Fable 5's safeguards; because it took effect immediately and the company could not verify nationality in real time, it suspended both models for all users. Anthropic says the controls were lifted on 30 June, and Fable 5 returned the next day. The government's letter was not made public, and CSIS analysts questioned its legal basis. The episode, and the voluntary review framework ordered the same month, are covered in our piece on the "Pacing the Frontier" statement. Export controls are compulsory where that framework is not, but no published standard says when a model will be caught.
Two decisions are pending. BIS has yet to publish the rule replacing the diffusion framework. And a rule extending entity-list restrictions to companies at least 50% owned by listed firms, suspended for a year in November 2025 under a trade arrangement with Beijing, is due to return on 10 November 2026 unless extended. Treasury Secretary Scott Bessent said on 23 September that the arrangement would be extended to 10 January 2027, NBC News reported, but by mid-October BIS had published no matching extension of the suspension. Neither changes the basic limit. Chip controls decide who gets the chips; they do not govern what the labs that have them build, and they leave the alignment problem where it was.
Correction, 14 October 2026: An earlier version of this article defined a compute overhang as a store of undiscovered algorithmic improvements; the term means surplus computing power that such improvements could suddenly exploit. It said controlling the chip supply chain needs no international coordination, but the Netherlands licenses advanced lithography exports under its own rules. It called TSMC the only maker of the most advanced chips; it makes the large majority. It stated that GPT-4 was trained on about 25,000 A100 chips, a figure OpenAI has never disclosed. It attributed to Yann LeCun an argument that the controls are excessive, for which we found no source. It said Huawei's Ascend 910C was announced in late 2024 with benchmarks near Nvidia's H100; we found only press reports of Huawei's own claims to customers, not a formal announcement or a published benchmark. It also described the controls as the most effective current brake on AI capability development; they limit China's access to chips, not the pace of the American labs that built every frontier model in Epoch AI's data since 2023. Those passages have been corrected or removed.