U.S. posts official H200 and MI325X AI GPU export rules to China, but with plenty of caveats — a string of requirments greatly limits the total number of GPUs that can be shipped to China
The U.S. Department of Commerce has unveiled new export rules for shipments of advanced AI and HPC processors designed in America to China and Macau. However, while the new rules permit limited exports of specific accelerators — most notably AMD's Instinct MI325X, Nvidia's H200, and comparable lower-performance products — on a case-by-case basis, licenses are granted only if the products are readily available in the U.S. and if shipments to the People's Republic remain capped relative to U.S. volumes, which effectively rules out China-only SKUs.
When it comes to specifications, approved devices must feature a total processing performance (TPP) score below 21,000 points and total DRAM bandwidth under 6,500 GB/s, which means relaxation of performance-tied export rules. However, the key constraint is now U.S. supply priority: exporters must prove that domestic demand is fully met, that no U.S. orders are delayed, that advanced-node foundry capacity serving U.S. customers is not diverted, and that aggregate PRC-bound shipments do not exceed 50% of the same product shipped into the United States.
The U.S. DoC names AMD's Instinct MI325X and Nvidia's H200 GPUs as models of accelerators that may qualify. The MI325X delivers 1,300 TFLOPS of FP16 performance, which translates to a TPP score of 20,800, and pairs 256 GB of HBM3E memory with 6 TB/s of bandwidth. Nvidia’s H200 offers 989.5 TFLOPS of FP16 throughput, which converts to a TPP score of 15,832, alongside 141 GB of HBM and 4.8 TB/s of bandwidth. Both processors fall below the stated thresholds and are therefore eligible for shipments to China, provided a license is approved. Products that exceed the limits, or include re-exports that involve D:5 countries remains subject to denial.
However, while previous limitations focused solely on performance and specifications, the new compliance requirements extend well beyond specifications. Every shipment must undergo verification of TPP, memory bandwidth, interconnect bandwidth, and copackaged DRAM capacity in an independent U.S.-headquartered testing laboratory with no financial ties to the exporter or importer. Meanwhile, the U.S. DoC can revoke a lab's qualification at any time.
In addition, exporters must meet strict Know Your Customer and cloud-usage rules, disclose end users, prevent unauthorized remote access, and bar transfers of model weights or trained algorithms to restricted parties, such as Chinese military or secret service organizations.
To a large degree, the policy treats China as a market for AI processor leftovers, as China-bound volumes are limited by U.S. shipment history. For example, if a company had supplied 100,000 model A processors to American clients, it cannot ship more than 50,000 model A processors to Chinese customers at the same time. Meanwhile, Europe, Japan, and other regions remain normal destinations for AI and HPC hardware.
Also, while the rules give AMD and Nvidia a way to supply older accelerators to select customers in China or Macau, they make large-scale exports difficult as they need to ship to American clients and ensure no capacity is diverted from American clients, which reduces volumes that can be shipped to the PRC. In addition, the new technology particularly disadvantages smaller vendors that have limited access to advanced production and packaging capacities and can rarely offer performance comparable to that of AMD's or Nvidia's GPUs.
Stay On the Cutting Edge: Get the Tom's Hardware Newsletter
Get Tom's Hardware's best news and in-depth reviews, straight to your inbox.
In short, while the new export regulations allow tightly controlled sales of previous-generation (or at least not leading-edge) hardware to customers in China, which preserves the presence of American technologies in the Chinese AI sector, it does so in a way that greatly limits volumes that can be shipped to Chinese buyers.
Anton Shilov is a contributing writer at Tom’s Hardware. Over the past couple of decades, he has covered everything from CPUs and GPUs to supercomputers and from modern process technologies and latest fab tools to high-tech industry trends.
