Commerce Requires Licences for AI Chips Rented Out to Chinese Firms Abroad and Adds Five Countries to Its Licence List
The Bureau of Industry and Security’s 16 November interim rule targets overseas data centres serving China-headquartered customers and rescinds the Biden-era AI Diffusion Rule.

Key facts
- The Commerce Department’s Bureau of Industry and Security released the interim final rule “Addressing Remote Access to Advanced Semiconductors” on 16 November 2026.[1][3]
- Katten and IBTimes report that the rule requires a licence for chips sent to data centres providing more than 12,000,000 aggregate TPP of compute to companies headquartered in restricted countries.[1][4]
- Indonesia, Malaysia, Singapore, Thailand and Brazil now need a licence for covered advanced chips, which BIS says present a “higher-than-normal risk” of diversion.[3][2]
- Data centres already serving restricted-country companies at covered scale have 90 days after the effective date to obtain a licence.[1][4]
- The rule formally rescinds the Biden-era AI Diffusion Rule, including its country tiers.[1][3]
The U.S. Commerce Department's Bureau of Industry and Security (BIS) on 16 November 2026 issued an interim final rule, "Addressing Remote Access to Advanced Semiconductors," that requires export licences for advanced AI chips headed to overseas data centres serving companies headquartered in China and other restricted jurisdictions, and adds Indonesia, Malaysia, Singapore, Thailand and Brazil to the list of destinations that need a licence for such chips. The rule also formally rescinds the Biden-era AI Diffusion Rule.
The practical target is rented compute. Rather than smuggle chips into China, Chinese developers can lease clusters in third countries and train models there. According to Katten's advisory on the rule, the measure is meant to stop companies in restricted countries from renting large amounts of U.S.-based AI compute from data centres elsewhere, and it was directed by the Treasury Department and the National Security Council. The rule does not ban U.S. chips outright, Katten notes.
What does the rule require?
The core provision is a licence requirement. Under the rule as summarised by Katten, a licence is needed to export, reexport or transfer covered items, generally those under export classification numbers 3A090 and 4A090, when the supplier knows they will go to a data centre that provides, or will provide, more than 12,000,000 aggregate TPP (total processing performance, the metric BIS uses to size AI chips) to or for a company headquartered in a restricted country. IBTimes reported the same threshold and said capacity spread across several linked data centres counts toward it.
Other outlets describe the trigger less precisely. Manufacturing Dive wrote only of data centres that supply "significant" computing capacity remotely and gave no numeric threshold, and STR Trade likewise referred to a "significant volume of computing power" without defining it. The 12 million TPP figure therefore rests on two of the four accounts read for this article, and readers should check the rule text.
The restricted jurisdictions are Macau plus the countries in Country Group D:5, which includes China, Russia and Iran. The rule looks through corporate structure. According to Katten, it also reaches affiliates elsewhere whose ultimate parent is headquartered in a restricted country, and customers outside those countries that train models for, or provide models to, a restricted-country company.
Why are five more countries on the list?
Indonesia, Malaysia, Singapore, Thailand and Brazil now need a licence for covered advanced chips. STR Trade reported that BIS says these countries "present a higher-than-normal risk" of diversion, and that other countries may be added later. Katten says applications will be reviewed case by case, weighing each country's export-control regime and cooperation with BIS. These are not bans, IBTimes stressed.
Companies headquartered in the United States, or whose ultimate parent is, are exempt from the five-country requirement if they meet reporting conditions. Katten says the exporter must still file monthly reports with BIS by the first business day of the following month. The carve-out matters commercially: U.S. cloud providers can keep building capacity in Southeast Asia and Brazil under reporting, while foreign-owned operators face licensing for the same kind of shipment.
What changes for data-centre operators?
The compliance burden falls on suppliers and operators. Katten's summary says suppliers must obtain, and be able to produce, information on a data centre's customer diligence, including direct and indirect ownership or control by restricted-country persons. If a data centre refuses or cannot provide it, the supplier is presumed to know of a prohibited end use unless it resolves the red flag first. Building U.S. chips into foreign-made servers does not avoid the rule, and a change in use after export, such as a data centre that begins serving a restricted-country customer at covered scale, is treated as a transfer needing authorisation.
Existing arrangements get a grace period. Katten and IBTimes both say data centres already serving restricted-country companies at covered scale have 90 days after the effective date to obtain a licence. A narrow exclusion covers compute used exclusively for commercial recommender systems of the kind run by online retailers and social media companies, but systems that generate text, images, audio, video or software do not qualify, according to IBTimes.
When does it take effect?
Sources differ. STR Trade gave an effective date of 19 November 2026, a comment deadline of 19 January 2027, and a savings clause letting shipments already en route on 19 November proceed if completed by 21 December. Katten and IBTimes instead tied the effective date to publication in the Federal Register and did not give a date, with comments due 60 days after publication. The two descriptions are compatible if publication falls on 19 November, but none of the four accounts read confirms it independently.
What is the context?
The rule arrives while Congress weighs related legislation. Manufacturing Dive noted that Congress is considering the Remote Access Security Act, which would make clear that export controls apply to remote and cloud-based access. Sen. Dave McCormick (R-Pa.), quoted on that bill, said that "bad actors can train AI models by accessing advanced chips under the jurisdiction of the US." The Commerce rule reaches the same loophole by regulation.
BIS also signalled how it will enforce. According to STR Trade, the agency says it will monitor countries that tolerate evasion and may consider import restrictions, new licence requirements, Entity List designations and temporary export prohibitions on foreign persons involved.
What we don't know yet
None of the articles read for this report carried reactions from Nvidia, Chinese authorities, or the governments of the five newly listed countries; IBTimes listed those reactions among its open questions. It is also unclear which chip models fall under the rule beyond its general description, how TPP is measured across linked campuses in practice, and how BIS will decide which countries "tolerate" evasion. The 12 million TPP threshold, the effective date and the comment period should be confirmed against the Federal Register notice.
For chip designers and cloud operators, the direction is clear even where the details are not. Controls that once followed the chip now follow the customer and the workload, and the burden of proving who is renting a cluster, and for what, moves to the data-centre operator and its suppliers.



