Institute for Progress researcher: US export controls on H20 chips were overdue and enforcement gaps remain critical

Summary

The H20 chip was Nvidia's workaround for China: designed in early 2024 to stay within export control limits, it turned out to be exceptionally good at AI inference — running models rather than training them. As test-time compute and techniques like synthetic data generation made inference the key strategic input to AI development, over a million H20s shipped to China in 2024. The chip that was supposed to be compliant became a significant channel for AI compute into a rival state.

Washington was slow to respond. A reported Mar-a-Lago dinner between Trump and Nvidia CEO Jensen Huang preceded a White House reversal on plans to restrict the H20. Then, last week, Chinese companies placed new orders for over 1.3 million additional units. The Bureau of Industry and Security has since moved to block those orders — but the researcher from the Institute for Progress framing this as a short-term fix, not a structural solution.

The enforcement gap is the real story

Illicit chip flows are no longer a suitcase-smuggling problem. Estimates suggest around 100,000 controlled GPUs entered China last year through black-market channels, with individual cases involving over 20,000 GPUs at once — shipments worth hundreds of millions to potentially billions of dollars. The pattern fits China's established playbook: it used similar methods to route sanctioned goods to Russia during the Ukraine war.

The controls that do exist have repeatedly been outrun. When HBM memory restrictions were rumored, Huawei reportedly stockpiled chips equivalent to around 2 million units before the rules landed. SMIC's 7-nanometer breakthrough in 2023 relied on secondhand DUV machines that had already been acquired before export controls covered them. Deepseek trained its models on US chips sold before restrictions tightened — a fact that caught policymakers off guard.

China's chip gap, and where it isn't closing

On raw chip performance, Chinese domestic GPUs are roughly three to four years behind Nvidia's best. That gap had been narrowing, but export controls across the full stack — chips, tooling, and fab equipment — are now constraining China's ability to scale production. The quantity problem may matter more than the quality gap: producing high-end AI chips by the millions requires manufacturing infrastructure China doesn't yet have, and restricted access to TSMC-level tooling keeps that ceiling in place.

On funding, China is reportedly spending the rough equivalent of the US CHIPS Act's $52 billion not as a one-time package but annually, sustaining a long-term push to close the gap domestically.

Huawei's next move

With US-sourced data center infrastructure becoming more expensive under tariffs, Huawei is positioned to run the same playbook it used in 5G — aggressively courting global markets with lower-end data center offerings where it does have competitive products. Alibaba Cloud is doing the same. Neither has comparable AI chip technology at the high end, but they don't need it to win infrastructure contracts in markets where US alternatives are pricing themselves out.

What the administration should do

The consistent failure mode is acting after the window has closed. The Institute for Progress is pushing for a dedicated, technically competent team inside government with the mandate to anticipate which AI inputs will matter in six months — and to act before stockpiling and workarounds become the response. The H20 situation, the HBM stockpiling, the SMIC DUV machines: each was a case where the US identified the problem after leverage had already been lost.