Federal Register 2026-00789, effective 15 January 2026, moves a narrow band of U.S. H200-class exports to China and Macau onto case-by-case review. It is a licensing-policy change. It creates no general license. An electronic application is not authorization.
The news
BIS issued final rule 2026-00789, docket 260112-0028, RIN 0694-AK43, published in the Federal Register 15 January 2026 and effective that day (91 FR 1684). The summary says BIS is revising license review for certain semiconductors to China and Macau from presumption of denial to case-by-case.
Covered items are advanced computing commodities with total processing performance (TPP) less than 21,000 and total DRAM bandwidth less than 6,500 GB per second, "such as the NVIDIA H200 or AMD MI325X," the rule says, and only if the product is commercially available in the United States at publication. A Commerce news release dated 13 January 2026 names the same two SKUs.
Case-by-case applies to exports from the United States to end-users in China or Macau when the applicant files the certifications in Supplement No. 2 to Part 748, paragraph (dd). The rule keeps presumption of denial for reexports, exports from abroad, and in-country transfers to Macau or Country Group D:5, and for entities headquartered, or with a parent headquartered, in Macau or D:5, even if the end-user sits elsewhere.
BIS estimates about 100 extra license applications a year. That is an administrative forecast. It does not count approved shipments.
Who is bound
BIS wrote the rule under the Export Administration Regulations. Exporters of qualifying H200, MI325X, and lesser-equivalent SKUs from the United States are bound if they want case-by-case treatment. Ultimate consignees in China or Macau must supply Know Your Customer procedures, physical-security descriptions, and lists of intended infrastructure-as-a-service remote end users in named destinations.
U.S. buyers of the same SKUs are bound in practice. The exporter must certify that filling a China license would not delay existing or new U.S. orders for advanced-node integrated circuits, taking normal lead times into account, and that global foundry capacity would not be diverted from similar or more advanced parts for U.S. end users. Cloud operators running training or inference on imported units inherit the remote-access and model-weight transfer limits in the certification text.
What's new
Before this rule, H200-class exports to China and Macau sat under presumption of denial. The new path is discretionary review against a checklist: U.S. units already shipped; performance specs; a 50 percent aggregate-TPP cap versus U.S. commercial end-use shipments of the same product; prohibited end-use attestations; KYC against unauthorized remote access; IaaS limits on weight transfer; and a qualified U.S. third-party testing lab confirming specs on a representative sample before each export.
Under Secretary Jeffrey Kessler said in the 13 January news release: "Export controls should evolve with changes in technology, while protecting national security. Permitting the sale of the H200 to China under controlled conditions will strengthen the American technology ecosystem." That is a policy rationale. It grants no license.
What it does not settle
Dates: news release 13 January 2026; rule published and effective 15 January 2026. Hardware: TPP under 21,000 and DRAM bandwidth under 6,500 GB per second; H200 and MI325X are named examples. More capable SKUs stay outside this band. Region: U.S. exports to China or Macau only; reexports and D:5-headquartered entities stay presumption of denial. Volume: aggregate TPP to China and Macau no more than 50 percent of aggregate TPP shipped to U.S. customers for U.S. end use of the same commodities, from first U.S. commercial shipment to the application date. Testing: labs must be U.S.-headquartered, test in the U.S. customs territory, hold no ownership in the parties, and can be disqualified by BIS at any time. Price, approved license count, and Chinese offtake are not disclosed.
What to do now
Export-compliance teams at NVIDIA, AMD, and board partners should map which SKUs sit under the TPP and DRAM-bandwidth ceilings and which stay on presumption of denial. U.S. cluster buyers should put in writing that China-bound lots of the same SKU cannot jump their queue.
A 29 September indictment alleges a smuggling route for GPU servers to China through Malaysia and Singapore. See US v. Lui: Earthmade Computer owner charged with routing H100 and RTX 5090 servers to China through Malaysia and Singapore.