Key takeaways
- BIS issued an interim final rule on November 16 requiring a license to supply large amounts of remote compute to companies headquartered in China, Macau or other D:5 countries.
- Advanced chips going to Indonesia, Malaysia, Singapore, Thailand and Brazil now need a license, except for US-headquartered buyers that meet reporting rules.
- Data centers already serving D:5 companies have 90 days after the rule takes effect to get a license.
The US Commerce Department issued a new AI chip export rule on November 16 that reaches past the chips themselves to the data centers that rent them out. Under the interim final rule from the Bureau of Industry and Security (BIS), a data center anywhere in the world needs a license to keep supplying large amounts of computing power to companies headquartered in China, Macau or the other countries in the export group known as D:5.
Until now, the controls mostly followed the physical chip. A Chinese AI lab that couldn't buy Nvidia or AMD accelerators could still rent them from a cloud in Malaysia or Singapore. The rule text says its aim is to stop exactly that: companies in countries of concern using US chips through remote access to train frontier models.
What the AI chip export rule changes
The rule does four things.
- Remote compute needs a license. Exporting advanced chips to a data center that will give companies headquartered in D:5 or Macau access to more than 12,000,000 "aggregate TPP" now requires a license, reviewed case by case. TPP, total processing performance, is the measure the export rules use for a chip's computing power, and the cap counts linked data centers together, not only one building. Data centers, foreign chip factories and exporters must check who owns their customers, directly or indirectly.
- Five more countries need a license. Advanced chips going to Indonesia, Malaysia, Singapore, Thailand or Brazil now need one. Shipments to companies headquartered in the United States are exempt if the exporter meets new reporting rules.
- Chip factories get less room. Foundries and packagers making advanced-node logic chips for designers in licensed countries have fewer ways to overcome the presumption that they need a license.
- The AI Diffusion Rule is formally gone. The Biden-era rule from January 2025, which sorted countries into tiers and controlled model weights, was never enforced. This rule removes it from the regulations, including the control on model weights.
There is one carve-out. Computing power used only for commercial recommender systems, such as the ranking engines of online stores and social networks, is exempt.
Existing data centers have 90 days
A data center that already serves D:5 companies with significant compute can keep doing so for 90 days after the rule takes effect. After that it needs a license. The rule takes effect when it is published in the Federal Register, and comments are due 60 days after that. The version BIS posted on November 16 still has placeholders for both dates.
Why it matters
This closes the gap that let Chinese companies reach US chips without importing them. For cloud providers in Southeast Asia, it turns customer screening from good practice into a legal requirement, with license reviews for any big contract with a Chinese-owned customer. For US clouds and US-headquartered companies building in Malaysia, Singapore and the other three countries, the reporting exemption keeps the door open.
The rule also comes with a threat. Commerce says it will watch for countries that tolerate evasion, and lists possible responses: import restrictions, new license requirements, a lower export control status and Entity List designations.
Two details stand out. The press release says the rule was "written by and at the direction of the Department of Treasury and the National Security Council," not by Commerce alone. And the text uses "Super Intelligence" (SI) in place of "AI" throughout, citing Executive Order 14434.
What comes next
The rule lands after a week of pressure on Nvidia's China sales, covered in our story on Megaspeed and the $12 billion chip report, and a day before Nvidia reports earnings on November 17. Chip stocks were lower on Monday morning: as of 14:49 UTC, Broadcom was down 4.4%, Micron 5.5% and Arm 4.9%. Watch for the Federal Register publication, which starts the 90-day clock, and for what Nvidia's executives say about the rule on Tuesday's call.
Sources
- BIS Issues Rule Addressing Remote Access to American Super Intelligence TechnologyBureau of Industry and Security
- Addressing Remote Access to Advanced Semiconductors (RIN 0694-AK50), interim final ruleBureau of Industry and Security
- US Moves to Restrict Remote Access to Advanced AI ChipsBloomberg
