Did the U.S. Ease AI Chip Export Controls on China? How to Read the H200 and MI325X License Rule

Direct Answer

If a headline says Washington has reopened AI chip sales to China, the license rule deserves a closer read. The January 2026 Bureau of Industry and Security rule did not turn NVIDIA H200, AMD MI325X, or similar chips into automatically approved exports.

The practical change is narrower: certain advanced computing items moved from a presumption of denial to case-by-case license review, but only when the product, transaction path, end user, supply evidence, Know Your Customer controls, remote-access safeguards, and independent testing requirements line up.

The better working question is not whether the export control was lifted. It is whether a specific transaction can get through the rule’s gates before BIS even reaches a case-by-case licensing judgment.

What Actually Changed

BIS announced the policy revision on January 13, 2026, and the Federal Register final rule took effect on January 15, 2026. The BIS release describes the change as a revision to license review policy for certain semiconductors exported to China.

The rule focuses on advanced computing commodities below two performance thresholds: less than 21,000 Total Processing Performance, or TPP, and less than 6,500 GB/s total DRAM bandwidth. H200 and MI325X are named as examples of items in that range.

That example matters, but it is not a safe harbor by itself. A chip name does not answer the export-control question. The transaction still has to satisfy the product threshold, destination and end-user limits, supply conditions, certification requirements, and testing process.

The H200 and MI325X Rule in One Table

Checkpoint	What the rule is looking for	Why it matters in practice
Transaction path	Export from the United States to an end user in China or Macau	Reexports, exports from abroad, and transfers in-country may be treated differently.
Product performance	TPP below 21,000 and total DRAM bandwidth below 6,500 GB/s	The application needs technical data, not just a product name.
Example items	NVIDIA H200, AMD MI325X, and similar chips	Mention in the rule does not equal automatic approval.
U.S. supply	Enough supply so U.S. customer orders are not delayed	BIS is asking whether China-bound shipments crowd out U.S. demand.
Foundry capacity	No diversion of global foundry capacity needed for U.S. end users’ similar or more advanced integrated circuits	Production allocation becomes part of the licensing record.
China and Macau threshold	China/Macau aggregate TPP for the item must not exceed 50% of aggregate TPP for U.S. end users	The test is based on aggregate TPP, not a simple unit count.
End use and end user	No military, military-intelligence, WMD-related, or restricted-party use	Screening has to cover both the buyer and how the compute will be accessed.
KYC and physical security	Consignee KYC procedures and facility security must be described	BIS is looking beyond shipment paperwork to operational controls.
IaaS remote users	Certain remote users must be identified by location, headquarters, or ultimate parent	Cloud access can create a control problem even if the chips stay in one facility.
Third-party testing	A qualified independent U.S.-based tester verifies technical performance before export	The product specification becomes an externally checked licensing fact.

Start With the Product Threshold, Not the Brand Name

For H200, MI325X, or any similar accelerator, the first file to build is the technical specification file. The rule asks for more than a marketing name or a product family.

The application materials need to address TPP, total DRAM bandwidth, interconnect bandwidth, copackaged DRAM capacity, and peak power consumption at maximum TPP. If the model’s specifications changed after launch or after an earlier shipment, the applicant needs to explain that change.

That creates a simple but demanding first screen: the product must fit the stated performance window, and the application must describe that product in a way that can be checked.

The U.S. Supply Conditions Are a Real Gate

The rule also asks whether the proposed China or Macau shipment would affect U.S. supply. That is where a commercial export application turns into a capacity-allocation record.

Applicants need evidence that existing or new U.S. end-user orders will not be delayed. They also need to show that global foundry capacity used to supply similar or more advanced integrated circuits for U.S. end users is not being redirected to produce the China-bound item.

Then comes the 50% threshold. The applicant must show that cumulative aggregate TPP for the same AI item shipped to China and Macau does not exceed 50% of cumulative aggregate TPP for U.S. end users.

For an operator, that means the evidence set is broader than a sales contract. It may include U.S. shipment history, current U.S. order status, normal lead times, foundry capacity allocation, and the aggregate TPP calculation for China and Macau exposure.

End-User Control Now Includes Remote Access

AI chips are not only sold as boxes in a room. They can be used through infrastructure as a service, remote compute access, model training pipelines, and downstream access to trained model weights or algorithms. The rule reflects that.

The applicant has to address whether the item will support prohibited military, military-intelligence, nuclear, missile, chemical, or biological weapons-related end uses or end users. It also has to account for restricted parties and remote access.

If the consignee provides infrastructure as a service, the control problem becomes more detailed. The rule requires attention to IaaS end users, transfer of model weights trained on the covered AI item, and remote access to algorithms trained on that item.

The application also has to identify prospective IaaS remote end users located in, headquartered in, or ultimately parented from Belarus, China, Cuba, Iran, Macau, North Korea, Russia, or Venezuela.

Third-Party Testing Is Not a Paper Formality

The case-by-case pathway includes a pre-export third-party testing step. The rule does not require every individual chip to be tested one by one; it allows review of a representative sample from a batch selected by the testing organization. But the tester and the test location are constrained.

The testing organization must be headquartered in the United States, and the testing must occur inside the U.S. customs territory. The tester cannot be controlled by a company headquartered or ultimately parented in Macau or Country Group D:5.

Independence also matters. The testing organization cannot have an ownership or financial interest in the consignee, exporter, or other transaction parties. It must also have the expertise to verify performance and functionality against the application’s technical description, including TPP, total DRAM bandwidth, interconnect bandwidth, and copackaged DRAM capacity.

BIS can revoke a tester’s qualification. If that happens, exporters using that tester would need to identify a new qualified testing organization before relying on this pathway.

Where the Presumption of Denial Can Still Apply

The narrowest mistake is to read case-by-case review as a general relaxation. The rule keeps several categories outside the easier lane.

Watch for these risk points:

• The transaction is a reexport, export from abroad, or transfer in-country rather than a direct U.S. export to a China or Macau end user.
• The destination issue arises in a reexport, export from abroad, transfer in-country, or another application that sits outside the specific U.S.-export-to-China/Macau case-by-case lane.
• Certain entity headquarters or ultimate-parent structures involving Macau or Country Group D:5 can keep the application under a presumption of denial, especially outside the specific U.S.-export-to-China/Macau case-by-case pathway.
• The license application lacks the required certification or a pre-export commitment to submit certification.
• The applicant cannot screen out military, military-intelligence, WMD-related, or restricted-party concerns.
• The consignee cannot explain KYC procedures, physical security, IaaS user controls, model-weight transfer limits, or remote-access restrictions.

In other words, the rule is less a broad reopening than a filter. It tells applicants which transactions may receive case-by-case review and which facts BIS expects to see before that review can be meaningful.

A Practical Pre-Filing Checklist

Before treating an H200, MI325X, or similar chip transaction as eligible for the new review policy, work through the gates in order.

1. Confirm the transaction path: U.S. export, reexport, export from abroad, or transfer in-country.
2. Map the end user’s location, headquarters, and ultimate parent company.
3. Prepare the product specification record: TPP, total DRAM bandwidth, interconnect bandwidth, copackaged DRAM capacity, and peak power consumption at maximum TPP.
4. Document any product specification changes since launch or prior shipment.
5. Collect U.S. customer supply evidence, including order status, shipment history, lead times, and any expected delays.
6. Document foundry capacity allocation for similar or more advanced integrated circuits supplied to U.S. end users.
7. Calculate whether China and Macau cumulative aggregate TPP for the item stays at or below 50% of cumulative aggregate TPP for U.S. end users.
8. Screen for military, military-intelligence, WMD-related, and restricted-party risks.
9. Review the consignee’s KYC procedures, physical security, and access controls.
10. If IaaS is involved, prepare the remote end-user list and the model-weight and algorithm access controls.
11. Select a qualified independent third-party testing organization and prepare the pre-export certification process.
12. Submit the required certification through SNAP-R when the filing path requires it.

How to Read the Market Signal

For supply-chain and market-risk readers, the rule does not say that China-bound AI chip revenue is simply back on the table. It creates three bottlenecks.

The first is the product bottleneck. The chip has to fit the performance thresholds, and the specs have to survive third-party verification.

The second is the supply bottleneck. The applicant has to show that China or Macau shipments do not delay U.S. customers and do not redirect foundry capacity needed for U.S. end users.

The third is the access-control bottleneck. When compute can be rented or accessed remotely, the relevant question is not only who buys the chip. It is who can use the compute, receive model weights, or access algorithms trained on the controlled item.

That is the cleaner frame for the rule: some transactions may now have a path to case-by-case review, but the path runs through product specs, U.S. supply evidence, end-user controls, remote-access controls, and independent testing.

What to Watch Next

The most useful updates will probably come from the rule’s operating details, not from broad claims about whether controls are easing or tightening.

Track three items first:

• BIS notices or guidance, especially on qualified third-party testing organizations and any revocations.
• Changes to license application requirements, certifications, KYC materials, IaaS remote-user lists, or technical data submissions.
• Product revisions and shipment records, because the same product name may not answer the performance or aggregate TPP questions.

For any specific H200, MI325X, or similar transaction, the next checkpoint is concrete: identify the transaction path, confirm the performance thresholds, gather U.S. supply evidence, map the end user and remote users, and verify that an eligible third-party testing route exists.