US Closes AI Chip Remote-Access Loophole With Licenses
Data centers anywhere in the world now need a US license to rent large amounts of American AI chip compute to Chinese-headquartered companies, under a Commerce Department rule released on November 16. The same rule adds license requirements for advanced chips shipped to Indonesia, Malaysia, Singapore, Thailand and Brazil, exempts US-headquartered buyers that report monthly, and formally scraps the Biden-era AI Diffusion Rule.
The US Commerce Department’s Bureau of Industry and Security (BIS) released an interim final rule on November 16 that targets the route China-linked companies have used to reach American AI chips without buying them: renting the compute from data centers in other countries. According to the press release, the rule was written by and at the direction of the Treasury Department and the National Security Council, and is meant to ensure that existing restrictions cannot be circumvented through remote computing arrangements in third countries. The rule text, titled “Addressing Remote Access to Advanced Semiconductors,” does four things at once.
A license for renting compute, not just shipping chips
The centrepiece is a new end-use control. Exporting, reexporting or transferring advanced chips now needs a license when the exporter knows they are headed for a data center, anywhere in the world, that will give companies headquartered in China, Macau or another Country Group D:5 destination access to more than 12,000,000 “aggregate TPP” of computing power. TPP, or total processing performance, is the export-control metric for chip throughput, and the threshold counts capacity at a single site or across “linked data centers,” which the rule defines to include digitally linked infrastructure, not just facilities wired together physically. By BitsMinds’ own arithmetic, using the roughly 15,800 TPP usually cited for an Nvidia H100, the line sits at around 750 H100-class chips.
The control also reaches intermediaries: it applies when a data center knowingly serves a non-D:5 customer that is training models for, or supplying models to, a D:5 company. A data center that only starts serving such customers after the chips arrive must get a license at that point, and operators already doing so before the rule was published have 90 days from its effective date to obtain one or face future export restrictions. Applications will be reviewed case by case. The one carve-out is for “commercial recommender systems,” the ranking engines run by online retailers and social networks, which the department says do not pose the same national security threat, subject to reporting.
Five countries added to the license list
Separately, a license is now required to send advanced chips to Indonesia, Malaysia, Singapore, Thailand and Brazil, which the rule describes as presenting a higher-than-normal risk of diversion; the Commerce Secretary can add more countries. Applications will be judged partly on how well each destination enforces export controls and cooperates with BIS. The requirement does not apply to buyers headquartered in the United States, or whose ultimate parent is, as long as the exporter files monthly reports with BIS on each shipment. Fortune reads that carve-out as covering cloud providers such as Microsoft, Oracle and Amazon, while noting that the remote-access limits still apply to them. The rule also warns that countries tolerating evasion could face import restrictions, new license requirements, a downgrade in their export-control status, and Entity List designations for the people involved.
The last two pieces are narrower. The rule tightens the presumptions that front-end chip fabricators and packaging houses face when making advanced-node logic chips, cutting the ways a designer in a license-required destination can qualify for production abroad. And it formally rescinds the regulations behind the AI Diffusion Rule published in January 2025, which BIS had already said in May 2025 it would scrap and which was never enforced. Throughout, the text says “Super Intelligence” where earlier rules said “artificial intelligence,” citing an executive order titled “Inaugurating the Era of Super Intelligence.”
Why Singapore and Malaysia, and why now
Neither the press release nor the rule mentions Nvidia. Fortune calls it the first update to the export-control regime since the scandal around former Nvidia chief executive Jensen Huang broke, after reports about his ties to a former executive of Megaspeed, a Singapore company suspected of funnelling controlled chips to China whose data centers were reportedly in Malaysia. Megaspeed has denied sending chips to China, and Huang resigned as Nvidia CEO on November 12. Fortune also notes that the House passed a Remote Access Security Act in January that is still waiting on the Senate; this rule gets much of the way there without new legislation.
The document posted by BIS still carries placeholders for its Federal Register date. It takes effect on publication there, and comments are due 60 days later through regulations.gov under RIN 0694-AK50. Because it is an interim final rule, the comments can reshape the details but do not hold up the license requirement, so the first test will come at the end of the 90-day window, when every overseas operator renting serious capacity to Chinese clients has to have applied or stopped.
Want AI news before everyone else?
The morning's most important AI stories, straight to your inbox. No fluff.