Volume I, No. XLII
Commerce moves AI chip controls from the shipping dock to the data center's customer list
A Commerce interim final rule makes overseas data centers answer for who rents their compute, adds a chip license requirement for Indonesia, Malaysia, Singapore, Thailand and Brazil, and formally repeals the 2025 AI Diffusion Rule. It takes effect on publication, set for November 19.
American chip controls have long rested on an awkward premise: that a government can decide who uses a computer by deciding where the computer is shipped. The premise holds while a buyer must own the chips. It weakens once a company can rent their output from a data center in another country. The Bureau of Industry and Security’s interim final rule, filed for public inspection on November 16 and scheduled for publication on November 19, is Washington’s answer1. It does not try to license cloud services as such. Instead it attaches a license requirement to the hardware whenever the exporter knows the data center will sell a large enough share of its compute to a company headquartered in China, Macau or another Country Group D:5 destination1. The control still turns on the shipment, but the question now asked of the data center is who its customers are, not just where it stands.
The instrument: an end-use control built on “knowledge”
The rule is issued under the Export Control Reform Act of 2018 (ECRA), 50 U.S.C. 4801-4852, which BIS cites for its authority to regulate exports, reexports and in-country transfers of items subject to U.S. jurisdiction (section 1753, 50 U.S.C. 4812) and to require licenses for them (section 1754, 50 U.S.C. 4813)1. Section 1762 of the Act, codified at 50 U.S.C. 4821, takes functions exercised under ECRA outside most of the Administrative Procedure Act, which is why the rule is effective on publication without prior notice and comment1,4. Comments are still invited, on docket BIS-2026-0628, for 60 days after publication1. By this journal’s count that is January 18, 2027.
The central provision is a new section, 15 CFR 744.24. It forbids exporting, reexporting or transferring covered chips without a license when the exporter has “knowledge” that they are destined to a data center, anywhere in the world, “that is providing or will provide access to more than 12,000,000 ‘aggregate TPP’” to or for companies headquartered, or with an ultimate parent headquartered, in Macau or a D:5 country1. TPP is total processing performance as defined in the technical notes to ECCN 3A090; aggregate TPP is the sum across a group of chips1. The threshold can be met at one site or through “linked data centers,” a term the rule extends to “digitally-linked infrastructure, not just physically-linked infrastructure”1.
Four features give the control its reach. First, it binds whoever holds the chips, whether or not that person was party to the original export1. Second, the usual de minimis rules for U.S. content in foreign-made goods cannot be used where anyone in the chain knows of the prohibited end use1. Third, a change of customers counts: if a data center that received chips lawfully later begins serving a D:5-headquartered company above the threshold, that change “is considered a ‘transfer (in-country)’” requiring authorization, and exporters shipping without a license must tell recipients so in writing1. Fourth, “knowledge” is given a practical edge. Exporters must be able to produce the foreign data center’s know-your-customer diligence, including checks for direct or indirect ownership or control by D:5 or Macau persons, and a data center “declining or being unable to provide information sufficient to verify compliance” is a red flag that “presumptively provides you with ‘knowledge’” unless resolved1.
The limits are as instructive. The control reaches the remote customer only through the hardware supplier and the data center operator; the D:5 company renting compute is not itself the regulated party. And it exempts computing used solely for a “commercial recommender system,” defined as a tool that ranks, filters or suggests products or content, provided the operator reports the customer’s name, the aggregate TPP supplied and its diligence to BIS1. Any system that generates text, images, code or other content falls outside the exemption, even if it also recommends1.
The mechanism: five new license destinations and a reporting exit
Separately from the end-use rule, the regulation adds a destination-based license requirement under the Regional Stability controls in 15 CFR 742.6(a)(6)(iii)(A)(1) for exports to or within Indonesia, Malaysia, Singapore, Thailand and Brazil1. The preamble calls them countries with “an unacceptable risk of diversion”1. Applications will be judged case by case, including on “the effectiveness of a destination’s export control regimes and coordination with BIS”1. The rule’s summary adds that the Secretary of Commerce may designate further countries and that Commerce will consider consequences for countries that tolerate evasion, among them “import restrictions” and Entity List designations1.
There is an exit for American firms. The requirement does not apply where the ultimate consignee or end user is headquartered, or has its ultimate parent headquartered, in the United States, provided the exporter files a report under new 15 CFR 743.101. The report lists every party and ultimate parent, the ECCN at paragraph level, product name and SKU, TPP, quantity and date, and is due by the first business day of the month after the transaction1.
Licensing for D:1, D:4 and D:5 destinations and the United Arab Emirates continues1. China and Macau keep a case-by-case policy for chips below 21,000 TPP and 6,500 GB/s of DRAM bandwidth, with a presumption of denial for everything else bound for D:5 or for D:5-headquartered entities1. Front-end fabricators and chip packagers also lose a route around the presumption that an advanced logic chip is controlled: designers located in, or owned from, D:1, D:4, D:5 or the five new countries can no longer use the diligence steps in Note 1 to 3A090.a to rebut it1. Shipments already en route on publication may proceed under the old terms if completed within 30 days1.
The effect: compliance moves to the operator
The burden falls first on chip exporters and on data center operators in Southeast Asia and Brazil, who must now document their customers’ ownership to keep receiving hardware. BIS itself expects the paperwork to be modest. It estimates about 100 additional license applications a year and a net increase of roughly 970 report submissions, and it classifies the rule as not significant under Executive Order 128661. Existing operators serving D:5-headquartered customers above the threshold have 90 days from the effective date to obtain a license “or face future export restrictions”1.
The case for the rule rests on recent experience. Asia Times reported in June that Megaspeed, a Singapore company spun off from a Chinese gaming firm in 2023, had a Malaysian subsidiary that bought nearly $2 billion of advanced Nvidia chips, shipped to data centers in Malaysia and Indonesia that appeared to serve Chinese clients remotely5. That reporting followed BIS’s May 31 guidance, which said a license had been required since November 17, 2023 for advanced chips going to D:5- or Macau-headquartered entities “even if the entities themselves are located outside” those places2. Greenberg Traurig advised clients afterward that a recipient’s physical location is not enough to settle the licensing question, and that screening should identify parent ownership and headquarters before shipment7.
The opposing view comes chiefly from the chipmakers. Nvidia lobbied against a remote-access rule, Gizmodo reported, arguing that it would harm the United States by pushing China to accelerate domestic chipmaking6. The same report described a draft covering seven countries, including Norway and South Africa6. The published text names five1.
The longer view: from allocations to diligence
The rule ends a two-year detour. The Framework for Artificial Intelligence Diffusion, effective January 13, 2025 with compliance due May 15, 2025, would have sorted the world into tiers and capped most countries at a cumulative 790,000,000 TPP from 2025 to 20273. It also controlled AI model weights under a new ECCN 4E0913. In May 2025 BIS said it would not enforce the rule’s new compliance requirements2. The new regulation removes the allocations, License Exceptions AIA and LPP, the Data Center validated end-user authorization and ECCN 4E091, while keeping License Exception ACM and the worldwide scope of the advanced computing foreign direct product rule1.
What replaces the diffusion framework is narrower in geography and deeper in diligence: no quota for most of the world, but a duty on operators to know who sits behind their tenants. The approach descends from the October 2023 controls, which, BIS now says, were always meant to stop D:5 entities “setting up cloud servers or data centers in other countries” to train models1.
One question the text leaves open is practical. The regulatory text exempts commercial recommender systems from the new prohibition on condition of reporting1. The preamble, discussing the conforming change to 744.23, says operators providing remote access for such systems “may continue providing these services for 90 days, but must obtain a license following expiration of this period”1. Operators will need BIS to say which sentence governs before the 90 days run out.