Can NVIDIA sell its H200 AI chips in China? The short answer: there's a legal path, but it's narrow, and NVIDIA isn't counting on it. Here's how the rule works.
What changed
On Jan 13, 2026, the Commerce Department's Bureau of Industry and Security (BIS) announced a revised licensing policy. The final rule took effect Jan 15, 2026.
- Before: applications to export these chips to China or Macau were reviewed under a presumption of denial (default: no).
- After: BIS reviews them case-by-case, if conditions are met.
BIS said it followed President Trump's Dec 8, 2025 announcement that the U.S. would allow the H200 and similar products to ship to approved customers.
Which chips
The rule names the NVIDIA H200 and AMD MI325X as examples. It applies to chips with:
- TPP under 21,000 (total processing performance, a standard export-control measure of compute), and
- Total DRAM bandwidth under 6,500 GB/s (memory speed, including stacked HBM).
They also have to be commercially available in the U.S. when the rule was published. Newer, faster chips aren't covered. Background on why memory bandwidth is a control lever: HBM explained.
The conditions
To get case-by-case review, the exporter must certify that:
- There's sufficient U.S. supply of the product.
- Making it for China won't divert foundry capacity from U.S. customers for similar or more advanced chips.
- The Chinese buyer has export compliance procedures, including customer screening.
- Every shipment is checked by an independent, third-party testing lab in the U.S. to verify its specs. BIS can revoke a lab's qualification "at any time and for any reason."
Applications also have to list the chip's TPP, DRAM bandwidth, interconnect bandwidth, copackaged memory capacity and peak power.
What still isn't allowed
The rule keeps a presumption of denial for re-exports and in-country transfers to China and Macau, and for entities headquartered (or with a parent headquartered) in China, Macau or other Country Group D:5 destinations, even if they're located elsewhere.
Why NVIDIA still guides to zero
A license path isn't revenue. In its Aug 26, 2026 CFO commentary, NVIDIA's guidance for its October quarter assumed no data-center compute revenue from China, and China Hopper shipments were under 1% of data-center revenue in the July quarter. More in our NVIDIA Q3 preview.
China is also building its own supply chain, including memory: see CXMT, China's DRAM maker.
