Key points:
- The US Commerce Department under the Trump administration allowed the import into China of more than 878,000 AI chips of the types Nvidia H200 and AMD MI325X from December 2025 to May 2026, including a significant amount to Alibaba, Tencent, Baidu, and BYD.
- Of the approved chips, over 30 percent were intended for businesses on the Pentagon’s list of Chinese military-affiliated entities, raising bipartisan doubts regarding the thoroughness of the national security analysis.
- According to Nvidia, shipments of H200s into China have ceased after that period, and licenses had conditions attached, but the exact amount of shipments is not known.
Following disclosure under the Export Control Reform Act by House Foreign Affairs Committee Chair Brian Mast on 19 th November, it has been revealed that the Bureau of Industry and Security approved export licenses for 879,868 advanced AI chips. These include H200 AI processors from Nvidia and MI325X processors from AMD. 150,000 of these chips will go to Tencent and affiliated companies, 75,000 to Alibaba, 43,000 to Baidu, and 4,000 to BYD.
All four companies are on a list of companies alleged to have military links from the Defense Department. This is the most granular information published to date and follows the reversal of prior Biden era restrictions in late 2025. Sales of H200 chips to permitted Chinese recipients are now allowed, but with a 25% tariff.
These advanced AI chips are dual use. They allow both commercial AI applications, or military applications including autonomous systems, analytics, and logistics. Over 250,000 advanced AI chips being exported to entities designated by the Pentagon has raised concerns that effective end-use and diversion controls did not occur. Some licences included recipients who were found to be potential shell companies.
In response, NVIDIA has stated that these exports were restricted, and are currently no longer shipping H200s to China. However, it is unclear how many chips have actually been delivered. Further confusion arises due to control mechanisms on imports within China and reports of storage of chips in third country warehouses. There appears to be a common dilemma in this area where export licencing can provide profit and market presence, but increases the AI capabilities of a strategic rival.
This decision represents a change in strategy. Following increasingly restrictive controls to maintain U.S. leadership within frontier AI applications, the Trump administration has adopted a managed export policy to allow revenue to be captured, but restricts quantities and applies conditions. Proponents argue complete prohibition is difficult to implement. A managed system allows more control than uncontrolled diversions and quick replacement from indigenous Chinese production.
However, opponents argue that even relatively ‘old’ high-end chips like the H200 represent significant processing power, relative to China’s native production. Even partial leakage can add up. The decision by the committee to publish this information demonstrates concerns within Congress. Members from both parties have expressed interest in implementing more restrictive legislative limits.
Depending on how many chips have been delivered to data centres within China, and how they are being used, the strategic significance of this information varies. If only a small amount of chips have been delivered, as suggested in some industry statements, the strategic impact may be limited. Conversely, if near the full quantity of licensed chips have been delivered, there would be considerable support to Chinese AI capabilities.
There is a clear requirement for clearer information on quantities of chips licensed, and subsequently delivered. Improving end use verification and creating cohesiveness across a long term policy to balance commercial benefits against gradual loss of technological advantage is important. The number of chips licensed (879,000) represents a key data point in the race to determine the pace of development in Artificial Intelligence.
Sources: Reuters

