The Institute for Progress published a report in October 2026 estimating that total compute available to Chinese entities will rise from about 4.6 million H100-equivalents in 2026 to 8.8 million in 2027 — a 91 percent increase. Roughly 65 percent of that capacity will come from US-designed hardware.
The composition of that 65 percent matters as much as the number itself. IFP tallied four sources: domestically produced chips, legal imports, smuggled processors, and compute accessed remotely through overseas cloud networks. The fourth category is growing fastest. Remote access alone is projected to reach 2.8 million H100-equivalents by the end of 2027, up from roughly 1 million in 2026 — about a third of total projected capacity.
That growth is not happening through a channel the export control framework was designed to regulate.
The Four Sources
The IFP methodology is worth unpacking because it exposes a distinction the policy debate often collapses.
Domestic chips are the category Chinese policy has spent the most money and political capital on. Huawei's Ascend line, Cambricon, Hygon, and a long tail of smaller designers are all ramping. But the ramp is constrained by fabrication capacity, not design. SMIC's N+2 line, the only advanced-node production line capable of volume AI chip manufacturing in China, has a year-end 2026 target of 70,000 wafers per month — roughly 2.6 million AI chips annually, according to industry estimates. Estimated domestic demand for the same period is about 4.2 million. One in every two chips demanded cannot be manufactured.
Legal imports are constrained by export controls but not eliminated. In January 2026, Washington moved to allow case-by-case export reviews for Nvidia's H200, AMD's MI325X, and comparable processors under security conditions. That is a licensed channel, not a closed one.
Smuggled processors are the category that receives the most media attention and the least reliable data. IFP includes them in its tally, but the report does not disclose the estimated volume. Smuggling is a leakage the control regime has not closed, but the scale is uncertain.
Remote cloud access is the category that is growing fastest and is legally the most awkward. A Chinese entity that cannot buy an H100 can rent access to one hosted in a data center outside China. The transaction is a service, not a hardware sale. Export controls on chips do not apply to the rental of compute capacity. A separate report published the same week by the Institute for AI Policy and Strategy stated that US-designed compute power "remained legally available for rent to foreign entities otherwise barred from buying the hardware."
That is the channel IFP is pointing at when it urges US policymakers to impose stricter controls on overseas cloud access. It is not a loophole in the sense of an unintended drafting error. It is a structural feature of a control regime designed around the movement of physical goods, applied to an industry where compute can be delivered as a service across a border.
The Remote Access Case
The remote access channel is not hypothetical. In March 2026, reports indicated that ByteDance had secured access to Nvidia's Blackwell chips through a Malaysian cloud computing provider. According to Reuters and the Wall Street Journal, the arrangement involved roughly 500 Blackwell systems — about 36,000 B200 chips — deployed outside China for the company's AI research, at a cost exceeding $2.5 billion. The chips never entered China. The compute did. Nvidia's statement on the arrangement was direct: export rules allow cloud infrastructure to be built and operated outside controlled countries.
That structure is why the Institute for AI Policy and Strategy called on the Bureau of Industry and Security to restrict Chinese access to cloud capacity "large enough for training frontier AI models or running them at scale," while advocating for stronger checks on customer identity and ownership. The recommendation acknowledges the difficulty: cloud capacity is fungible, customers can be intermediated, and the identity of the ultimate user is not always visible to the provider.
Remote access is projected to account for about a third of China's total available compute by the end of 2027. If the policy goal is to constrain China's access to frontier training capacity, a channel that supplies a third of the total and is legally open is a material gap.
The Training Gap and the Model Gap
IFP's report contains a second pair of numbers that complicate the story further.
The United States maintains roughly a 16-fold advantage in AI training compute over China. That is a measure of raw capacity available for training frontier models. On that metric, the gap is large and, by IFP's estimate, not closing quickly.
But the performance gap between leading Chinese and American models is measured at 6.5 months. IFP attributes part of the narrowing to distillation — training systems on the outputs of superior models. China has rejected US allegations of "aggressive, malicious" distillation as groundless, but the IFP report treats it as a factor.
The two numbers are not contradictory. A 16-fold training compute advantage should produce a larger performance gap than 6.5 months. The fact that it does not suggests that compute is not the only variable, and that the relationship between raw capacity and model quality is mediated by architecture, data, and method. The policy implication is uncomfortable for a control regime built on the assumption that restricting compute restricts capability: the link between the two is real but not linear.
What the Domestic Ramp Can and Cannot Do
The 65 percent American-chip share is not a static baseline. It is a snapshot of a transition.
Chinese domestic AI chip output has been growing. One estimate put 2025 shipments at about 4 million units, rising to a projected 5 million in 2026, with domestic share moving from 40 percent to above 50 percent. Morgan Stanley projects China's GPU self-sufficiency rising from 27 percent in 2025 to 39 percent in 2026. Other estimates using broader definitions of "AI chip" — including ASICs and NPUs — put the 2026 domestic share closer to 46 percent by procurement budget, or even higher by shipment volume depending on the category.
The definitional disagreement is itself informative. A GPU-only measure produces a lower self-sufficiency figure than an all-accelerator measure, because non-GPU accelerators — ASICs, NPUs, custom silicon — are a growing share of Chinese deployments. IDC data for the first half of 2026 showed non-GPU accelerator shipments exceeding GPU shipments for the first time, with GPU volumes down 26 percent year over year and non-GPU volumes up nearly 80 percent.
That shift explains part of the 65 percent figure. If a larger share of Chinese compute is coming from non-GPU accelerators, and a large share of those accelerators are domestic, then the American-chip share of total compute could decline even if the absolute volume of American chips in use continues to grow.
But the shift does not eliminate the dependency. The 65 percent figure is a measure of capacity in H100-equivalents, not of chip count. A domestic ASIC that delivers a fraction of an H100's throughput per unit contributes a fraction of an H100-equivalent. The metric weights by performance, not by unit volume. That is why the American share remains high even as domestic unit shipments rise.
What the Numbers Actually Constrain
The policy question that follows from the IFP data is not whether China can build AI without American chips. It can, and it is building that capacity. The question is what the American-chip share of China's compute implies about the effectiveness of a control regime designed to restrict access to frontier training capacity.
The answer depends on which channel carries the weight. Legal imports are licensed and visible. Smuggling is a leakage of uncertain scale. Domestic production is growing but constrained by fabrication and HBM. Remote access is legal, growing fastest, and supplies about a third of total capacity by 2027.
A control regime that restricts the first three channels but leaves the fourth open is restricting the channels that are visible and leaving open the one that is scaling. That is not a policy failure. It is a policy designed for a different problem — the movement of hardware — applied to an industry where the scarce input can be delivered as a service.
IFP's recommendation to tighten cloud access is an attempt to close that gap. Whether it can be closed depends on whether compute rental can be separated from other cloud services, whether customer identity can be verified at the point of access, and whether the providers — many of them American hyperscalers with global infrastructure — can be enlisted as enforcement agents.
Those are open questions. The 65 percent figure is not.
Sources: Institute for Progress report via South China Morning Post (October 24, 2026); Kantan.News summary of IFP report (October 24, 2026); China Times coverage of IFP findings (October 24-25, 2026); KuCoin summary of IFP report (October 24, 2026); Reuters report on ByteDance's Malaysian cloud arrangement (March 13, 2026); Wall Street Journal coverage of the ByteDance cloud deal (March 2026); Wccftech report on ByteDance Malaysian cloud access (March 13, 2026); 163.com analysis of domestic AI chip capacity and demand (July 23, 2026); Tencent Cloud Developer analysis of domestic AI server supply-demand gap (April 21, 2026); Sgpjbg report on China AI chip self-sufficiency metrics (October 20, 2026).
Disclaimer
The information provided in this article is for general informational and educational purposes only. It does not constitute legal, financial, or professional advice. The author and publisher are not responsible for any actions taken based on the content of this article. Readers should consult qualified professionals for advice specific to their situation. All trademarks and references to third-party products, services, or organizations are the property of their respective owners. The performance data and benchmarks discussed are based on specific research studies and may not generalize to all use cases or environments. As of the publication date, the AI landscape continues to evolve rapidly, and readers should verify current information independently.
Limitations
This analysis relies on the Institute for Progress report as summarized in public reporting; the underlying model and assumptions were not independently reproduced.
The reported volume of smuggled processors is not disclosed, so that channel is a residual rather than a measured quantity.
H100-equivalent conversions depend on assumptions about throughput per chip that differ across estimates.
Self-sufficiency figures vary widely by definition (GPU-only versus all accelerators) and are not directly comparable.
Reporting standards differ across jurisdictions; figures may not be directly comparable.
Market size and contract value figures come from third-party research and are estimates.
This sector is changing quickly; claims can become outdated soon.
Sources
- Institute for Progress report via South China Morning Post (October 24, 2026)
- Kantan.News summary of IFP report (October 24, 2026)
- China Times coverage of IFP findings (October 24-25, 2026)
- KuCoin summary of IFP report (October 24, 2026)
- Reuters report on ByteDance's Malaysian cloud arrangement (March 13, 2026)
- Wall Street Journal coverage of the ByteDance cloud deal (March 2026)
- Wccftech report on ByteDance Malaysian cloud access (March 13, 2026)
- 163.com analysis of domestic AI chip capacity and demand (July 23, 2026)
- Tencent Cloud Developer analysis of domestic AI server supply-demand gap (April 21, 2026)
- Sgpjbg report on China AI chip self-sufficiency metrics (October 20, 2026).
The information provided in this article is for general informational and educational purposes only. It does not constitute legal, financial, or professional advice. The author and publisher are not responsible for any actions taken based on the content of this article. Readers should consult qualified professionals for advice specific to their situation. All trademarks and references to third-party products, services, or organizations are the property of their respective owners. The performance data and benchmarks discussed are based on specific research studies and may not generalize to all use cases or environments. As of the publication date, the AI landscape continues to evolve rapidly, and readers should verify current information independently.
Limitations: This analysis relies on the Institute for Progress report as summarized in public reporting; the underlying model and assumptions were not independently reproduced.; The reported volume of smuggled processors is not disclosed, so that channel is a residual rather than a measured quantity.; H100-equivalent conversions depend on assumptions about throughput per chip that differ across estimates.; Self-sufficiency figures vary widely by definition (GPU-only versus all accelerators) and are not directly comparable.; Reporting standards differ across jurisdictions; figures may not be directly comparable.; Market size and contract value figures come from third-party research and are estimates.; This sector is changing quickly; claims can become outdated soon.