Rented capacity in Southeast Asian data centers is the fastest-growing source of compute for Chinese labs, and IFP's proposed remote-access controls would target it directly.
Compute available to Chinese companies will grow from about 4.6 million to about 8.8 million H100-equivalents between the end of 2026 and the end of 2027, the Washington think tank Institute for Progress (IFP) estimates in a report published October 19. An H100-equivalent measures processing power against one Nvidia H100, the US data center chip widely used to train AI models. The IFP authors, Tim Fist, Georgia Adamson and Konstantin Pilz, expect US chips to make up around 65% of that end-2027 stock. The South China Morning Post reported the findings on October 24.
The figure counts every route by which Chinese companies reach compute: chips made in China, chips imported legally, chips smuggled in, and capacity rented from data centers outside the country. Rented compute is the fastest-growing channel in the IFP estimate, rising from about 1 million H100-equivalents at the end of 2026 to about 2.8 million a year later. IFP says about 75% of those data centers are in Southeast Asia.
The report argues that the gap between US and Chinese AI depends on two levers Washington still controls: how much compute Chinese labs can reach, and how freely they can copy US models through distillation, which trains a model on another model’s answers. IFP puts the US lead at about 6.5 months even though US frontier labs hold roughly 16 times the training compute of the top five Chinese labs.
Domestic chips will supply about a third of China’s compute
For the end of 2026, IFP breaks the stock down as about 1.7 million H100-equivalents from domestic production, 846,000 from legal imports, 1 million smuggled and about 1 million reached remotely. By the end of 2027, it expects domestic chips to reach 3 million, legal imports 1.35 million and smuggled chips 1.6 million, alongside the 2.8 million rented abroad. Domestic chips would then make up about a third of the total, consistent with the 65% share IFP assigns to US chips.
China had about 743,000 H100-equivalents of domestic chips at the end of 2025, IFP estimates, and will produce about 986,000 in 2026 and 1.3 million in 2027. The binding constraint is high-bandwidth memory (HBM), the stacked memory that sits beside AI processors, which IFP says China cannot yet make at scale. Commercial Times, a Taipei business daily, cited an Epoch AI estimate from September that Huawei’s 2026 AI compute output will be less than 4% of Nvidia’s, and could fall to about 1% by 2028 without foreign HBM.
Huawei is trying to move faster. The company said in September that its Ascend 960DT, a chip designed for training AI models, is due in the first quarter of 2027, three quarters earlier than planned, Commercial Times reported. EastFrontier covered that timeline change at the time.
Legal imports and smuggling are the least certain parts of the estimate
Legal imports are mostly Nvidia H200s and gaming and workstation graphics cards, IFP says. Washington began reviewing H200 export applications case by case in January, Commercial Times noted, while Chinese regulators first banned H200 imports and later allowed small quantities, according to the report. IFP’s range for legal imports runs from 674,000 to 1.02 million H100-equivalents in 2026 and from 986,000 to 2.04 million in 2027.
The smuggling figures carry the widest uncertainty. IFP’s central estimate of 1 million smuggled H100-equivalents by the end of 2026 sits inside a range of 381,000 to 2.9 million, and its 2027 range runs from 468,000 to 6.7 million. The authors describe their forecasting model as a simplified tool for comparing the broad effect of policy choices.
The report also explains how Chinese models keep pace with less compute. IFP estimates that a 16-fold compute advantage would on its own give US labs a lead of about 13 months, and that distillation wins back about half of it. Commercial Times reported that Beijing has called US accusations of malicious distillation baseless. EastFrontier reported earlier this month on how Chinese labs stretch fixed compute under export controls.
IFP wants Washington to close the cloud route first
IFP proposes a package that it says uses existing executive-branch powers and could roughly double the US lead by the end of 2027, adding about 6.6 months. The measures include export controls on Chinese firms’ remote access to overseas training clusters, stronger due diligence to stop chip smuggling, tighter limits on the chips that can still be sold to China legally, such as H200-class parts and RTX cards, and action against smuggling of foreign HBM. It also asks the Department of Justice and the Federal Trade Commission for a narrow antitrust safe harbor so US AI companies can coordinate against distillation, which it estimates could cut the data Chinese labs harvest by 95%.
On cost to US chipmakers, the authors argue that demand outside China will exceed supply through at least 2027, so restrictions "would thus likely redirect sales, largely to US customers, rather than reduce them." China Times carried the report’s figures for Taiwanese readers on October 24.
The proposals now need a buyer in Washington. The clearest signal would be a US export rule covering remote access to cloud clusters, the channel IFP expects to grow almost threefold in 2027.