China’s AI computing capacity will nearly double from 4.6 million to 8.8 million H100-equivalents by the end of 2027, but US-designed chips will supply about 65% of it, according to a report by the Washington-based Institute for Progress covered by the South China Morning Post on October 24. Remote access to overseas cloud capacity alone will rise from about 1 million to 2.8 million H100-equivalents. The institute argues the US could roughly double its AI lead by end-2027 with tighter controls, while Chinese firms race to build domestic alternatives led by Huawei’s Ascend chips. China has previously rejected US allegations of aggressive, malicious distillation as groundless.
Where China’s Compute Comes From
The Institute for Progress tallied four channels of Chinese AI compute: domestically produced chips, legally imported US-designed chips, smuggled processors, and compute rented remotely through data centres outside China. Together they give Chinese entities about 4.6 million H100-equivalents by the end of 2026, rising to 8.8 million by the end of 2027. An H100-equivalent measures capacity against Nvidia’s data-centre graphics processors.
Remote access is the fastest-growing channel. The report puts it at roughly 1 million H100-equivalents this year and 2.8 million by the end of 2027. About 35% of China’s expected year-end 2027 chip stock is already inside China; of the rest, about half comes from remote access, which the institute says new Bureau of Industry and Security controls could almost entirely restrict today.
Nvidia’s GPUs and CUDA platform sit at the centre of the AI boom, and that centrality is exactly what the figures measure: even as China’s domestic industry grows, the benchmark for its capacity is still an American chip. Washington moved in January to allow case-by-case export reviews for Nvidia’s H200, AMD’s MI325X and comparable processors under security conditions, with exporters required to show no diversion from US customers and Chinese buyers required to adopt compliance procedures and independent US testing.
Distillation Halves The US Lead
The US holds a roughly 16-fold advantage in AI training compute over China, yet leading Chinese models lag their American rivals by only about 6.5 months. The report’s explanation is distillation: training systems on the outputs of superior models to capture their capabilities at a fraction of the original training cost.
By the institute’s estimates, distillation recovers about half of the capabilities gap. Without it, the 16-fold compute advantage would yield a lead of about 13 months; with it, the lead falls to 6.5 months, a gain for China equivalent to having four times more physical compute. The report notes that US companies and the US government have publicly documented multiple instances of top Chinese companies distilling at scale, while China has previously rejected US allegations of aggressive, malicious distillation as groundless.
DeepSeek’s efficiency shock showed how engineering can substitute for brute force, and the report treats that lesson as general: the larger the US compute advantage grows, the more effective distillation becomes as a catch-up strategy, because Chinese firms can distil ever more powerful models while needing relatively less compute of their own.
What IFP Wants Washington To Do
The institute’s central claim is that policies using existing executive-branch authorities could roughly double the current US lead by end-2027, adding about 6.6 months. Its list includes a narrow antitrust safe harbour so US companies can collaborate against adversarial distillation, new export controls on remotely accessed training clusters abroad, stronger due-diligence requirements against smuggling, updated controls on chips currently permitted for legal export, and controls on smuggled foreign high-bandwidth memory.
A separate report published on Tuesday by the Institute for AI Policy and Strategy points the same way. It found US-designed compute remains legally rentable to foreign entities otherwise barred from buying it, and called on the Bureau of Industry and Security to restrict Chinese access to cloud capacity large enough to train frontier models or run them at scale, alongside stronger checks on customer identity and ownership.
Beijing’s own draft controls on 23 AI technologies show the pressure runs both ways: China is moving to control exports of model weights and training methods even as Washington debates how to control the chips beneath them.
Why It Matters
The report reframes the US-China AI race as a contest over access, not just invention. China does not need to match American chip production; enough imported, smuggled or rented capacity, plus distillation, keeps it within months of the frontier.
For Chinese firms, the figures explain the urgency behind domestic hardware. Huawei said last month its Ascend 960DT training chip arrives in the first quarter of 2027, three quarters early, with more than 1,000 Atlas 900 A3 SuperPods already deployed. But scaling domestic production remains a challenge: research group Epoch AI identified high-bandwidth memory last month 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.
Background
The Institute for Progress published its report on October 19, and the South China Morning Post covered it on October 24. The report measures the gap with the Epoch Capabilities Index, which aggregates many benchmarks, and builds its compute figures from domestic production, legal imports, smuggling estimates and remote access. The debate it feeds — how far export controls can stretch an American lead, and how far distillation and domestic chips can shrink it — has run since Washington’s first advanced-chip controls in October 2022.
0 Comments