Skip to content
Developer312

Commerce Approved 879,000 AI Chips for China: The Number Behind Every Compute Model

House Foreign Affairs Chair Brian Mast released Commerce data showing BIS approved more than 879,000 Nvidia H200 and equivalent AMD AI chips for sale to China between December 2025 and May 2026. Named recipients include Alibaba, Tencent, ByteDance and Tesla Shanghai.

Published Report an error

House Foreign Affairs Chair Brian Mast released Commerce Department data on November 19 showing that the Bureau of Industry and Security approved more than 879,000 Nvidia H200 and equivalent AMD AI chips for sale to China between December 2025 and May 2026. That figure comes out of 3.4 million chips requested during the window, and named recipients include Alibaba, Tencent, ByteDance, Tesla Shanghai and Baidu. The release is billed as the most detailed public disclosure in the history of the Export Control Reform Act.

The number reframes a debate that has run on assumptions for years. Analysts modeling Chinese AI compute capacity have generally treated export controls as a binding constraint, while treating the leakage rate as guesswork. Now there is a measured quantity: 879,000 approved accelerators in roughly six months, with the approval rate, more than a quarter of the 3.4 million requested, documented rather than estimated. Whatever your model of Chinese training and inference capacity said last week, it needs a rerun with this figure in it.

Key Takeaways: What Matters Most?

  • BIS approved more than 879,000 Nvidia H200 and equivalent AMD AI chips for China between December 2025 and May 2026.
  • The approvals came out of 3.4 million chips requested, meaning roughly one in four requested export licenses was granted.
  • Named recipients include Alibaba, Tencent, ByteDance, Tesla Shanghai and Baidu, with two firms flagged as suspected smuggling fronts.
  • The firms Maginfra and New Matrix were flagged as suspected smuggling fronts in the House release.
  • The release is billed as the most detailed public disclosure in the history of the Export Control Reform Act.

Takeaways like these, every weekday morning. Get the briefing

What the approval record actually shows

The scale is the first thing to absorb. These are not gray-market anecdotes or smuggled pallets; these are approvals granted by the export-control apparatus itself, for H200-class hardware from Nvidia and equivalent AMD parts, to the largest AI buyers in China. Alibaba, Tencent, ByteDance and Baidu are the named recipients alongside Tesla Shanghai. When the customers are the most sophisticated cloud and AI operators in the country, the approved volume translates directly into training and inference capacity, not into warehouse inventory.

The second thing the record shows is that the process had a filter, just not a tight one. Of 3.4 million chips requested, more than 879,000 were approved, which means the review pipeline granted roughly one in four requests for top-tier AI silicon destined for China. A control regime that denies 74% of requests is doing something. A control regime that approves 26% of requests for H200-class hardware, to ByteDance and Alibaba, is also doing something else, and the release forces a naming of which thing is the policy.

The third finding is the darker one: the House release flags two firms, Maginfra and New Matrix, as suspected smuggling fronts. That means even within the approved pipeline, the government acknowledges units whose documented end users may not be the real ones. Approved-but-diverted is a category that did not exist in most public analyses, and it now has named examples attached to a documented approval volume.

Why the release exists, and who it pressures

Mast's release is a political act as much as a transparency one. A committee chair publishing Commerce data and billing it as the most detailed disclosure in the history of the Export Control Reform Act is inviting the public to compare the stated policy with the approval record. The recipients list does most of the argumentative work: it is one thing to defend approvals as supporting commercial relationships like Tesla Shanghai, and another to explain 879,000 approved units when the list includes the companies building China's frontier AI models. The release puts that comparison in the open, with numbers attached.

For the AI industry outside China, the disclosure cuts in an uncomfortable direction. Nvidia and AMD shareholders have treated China restrictions as a revenue ceiling; 879,000 approved units in six months suggests the ceiling was considerably higher than the rhetoric implied. For US-based AI labs competing with Chinese models, the data suggests the compute asymmetry narrative, that export controls were holding back rivals, was overstated for at least this window. And for anyone building compliance or market-entry plans around "China cannot get top-tier chips," the correct assumption for December 2025 through May 2026 was that China could get them by the hundred thousand, with paperwork.

The trade-off: controls that gate volume, not access

The honest cost of this disclosure is intellectual. If you have written anything, a memo, a model, a market forecast, that assumed effective denial of H200-class hardware to Chinese AI firms, the documented approval of 879,000 chips over six months invalidates the simplification. The regime as practiced gates volume and velocity, not access, and even the gating has documented gaps through suspected fronts like Maginfra and New Matrix. Rhetoric about chokepoints was always easier to state than to verify, and now the verification exists, and it shows the chokepoint was porous by design or by drift.

What To Do About It

  • Rerun your compute assumptions: Any competitive analysis of Chinese AI capacity written before November 19 should be updated with 879,000 approved H200-class units in the December 2025-May 2026 window.
  • Read the recipient list literally: Alibaba, Tencent, ByteDance, Tesla Shanghai and Baidu are named buyers, so model frontier-lab access to this silicon, not just enterprise demand.
  • Discount your smuggling estimates carefully: With Maginfra and New Matrix flagged as suspected fronts, treat a portion of the 879,000 approved units as untraceable in your end-user assumptions.
  • Track the policy response: Watch whether the Export Control Reform Act's most detailed-ever disclosure triggers a tightening of the approval pipeline, because the next six-month window is the real test.

The Bottom Line

Export control as practiced between December 2025 and May 2026 approved more than 879,000 top-tier AI accelerators for China out of 3.4 million requested, and the House release names the buyers and the suspected fronts. The regime gated volume, not access, and the gap between the policy's stated purpose and its approval record is now a documented number rather than a suspicion. Any serious model of Chinese AI compute, and any serious argument about the effectiveness of controls, has to be rebuilt with this figure in it.

Sources

  1. [1]CNBC — Congress data on Nvidia and AMD China chip approvals (opens in a new tab)
  2. [2]AI Weekly — 879,000 approved AI chip exports coverage (opens in a new tab)

Frequently Asked Questions

Get the next briefing

The lead story each weekday morning, with the key takeaways and a link to the full analysis.

One lead story each weekday morning: what happened, the key takeaways, and a link to the full analysis.

New posts daily, briefing weekday mornings. We use your email only to send the briefing. Unsubscribe anytime. Privacy policy.

Related Articles