In May 1995, the Bureau of Export Administration published the text of its supercomputer safeguards plan, which a foreign buyer of a US supercomputer had to sign before the export was licensed. Machines were graded by a Composite Theoretical Performance figure in millions of theoretical operations per second. The plan reached past the loading dock: the consignee had to limit physical and computational access to authorized persons, issue no passwords to restricted nationals, and perform no work on the machine "on behalf of restricted nationals." Singapore and South Korea were among the destinations it covered. The interim final rule that the Commerce Department released on 16 November 2026 returns to that question of who works on the machine, at a time when the user of a computer is usually a cloud customer on another continent.
The rule is RIN 0694-AK50, "Addressing Remote Access to Advanced Semiconductors," a 132-page agency text from the Bureau of Industry and Security (BIS). The Federal Register filed it for public inspection on 16 November as document 2026-23529, with publication set for 19 November. The rule takes effect on that publication date, and comments are due 60 days later. This post reads the agency text itself and leaves aside press summaries of it.
Practical takeaway. From 19 November 2026, a data center anywhere in the world needs a BIS license for US-controlled AI chips if it knowingly gives more than 12,000,000 aggregate TPP to companies headquartered in Country Group D:5 or Macau, or whose ultimate parent is, unless the compute serves only a commercial recommender system. An operator already providing that service has 90 days from the effective date to obtain the license.
What section 744.24 prohibits, and who must hold the license
New section 744.24(a)(1) prohibits the knowing export, reexport or in-country transfer of advanced chips "for use in a data center to or within any destination worldwide" that will provide access to more than 12,000,000 aggregate TPP to companies headquartered in, or whose ultimate parent is headquartered in, Macau or a Country Group D:5 destination. TPP is the total processing performance metric of ECCN 3A090. By Quentir's arithmetic, an Nvidia H100 SXM, with about 990 teraflops of dense FP16 Tensor Core performance (NVIDIA lists 1,979 with sparsity), rates at roughly 15,800 TPP (990 times a 16-bit operand length), so the threshold corresponds to about 760 such chips. The count covers one site or several linked data centers, a term the rule defines to include "digitally-linked infrastructure, not just physically-linked infrastructure." An operator cannot stay under the threshold by spreading one customer across several halls.
The rule also reaches intermediaries. BIS states that "for an entity" covers a data center that knowingly serves a non-D:5 customer which in turn trains models for, or supplies models to, a D:5 company. Note 1 to the section applies the prohibition to whoever holds the chips, whether or not that party took part in the original shipment, and paragraph (e) treats a change from a permitted to a prohibited end use as an in-country transfer. An exporter shipping without a license on or after 19 November must tell the recipient that any such change requires BIS authorization. Note 2 bars reliance on the de minimis rules once US-origin chips are built into foreign-made servers, and Note 3 expects foreign data centers to examine direct and indirect ownership and control by D:5 or Macau persons, which BIS says is meant to defeat restructurings and spinoffs.
Which uses are exempt: recommender systems under section 744.24(b)(2)
The prohibition does not apply when the compute serves only a "commercial recommender system," defined as a tool "exclusively designed and exclusively used to rank, filter, or suggest" products, services or content. The definition excludes any system that generates text, images, audio, video or software, and any system that incorporates such a model even if it also recommends. A provider relying on the exemption must report the D:5 company, the summed TPP and the due diligence performed through a BIS portal, and BIS shares the filing with the End-User Review Committee. The exemption therefore protects online retail and social-media ranking while keeping generative assistants of the same platforms inside the prohibition. Every other application is reviewed case by case under section 744.24(c), and paragraph (b)(3) gives data centers that already serve D:5 customers 90 days from the effective date to obtain a license or "face future export restrictions."
Why Indonesia, Malaysia, Singapore, Thailand and Brazil now need a license
The second change amends section 742.6(a)(6)(iii)(A)(1) to require a license for advanced chips shipped to or within Indonesia, Malaysia, Singapore, Thailand and Brazil, which BIS describes as presenting a higher-than-normal risk of diversion. The Secretary may add destinations. Shipments to an ultimate consignee or end user headquartered in the United States, or whose ultimate parent company is headquartered there, are excluded if the exporter files a report under new section 743.10, listing every party and parent company, the ECCN, the product and SKU, its TPP and the quantity. Those reports are monthly and fall due on the first business day of the month after each transaction. License applications for the five countries are judged case by case, with BIS evaluating each destination on the effectiveness of its export control regime and its cooperation on enforcement.
For Southeast Asian data-center markets, the exclusion separates two kinds of builder. A US hyperscaler, or its local subsidiary, building in Johor or Batam keeps its supply under a monthly reporting duty, and an operator with neither a US headquarters nor a US ultimate parent needs a license. A separate amendment, section 742.6(a)(6)(iii)(A)(2), requires a license for Country Groups D:1, D:4 and D:5, excluding destinations also in Country Groups A:5 or A:6, and for the United Arab Emirates. For the UAE, no license is needed when the ultimate consignee and all end users are approved UAE entities listed in supplement no. 8 to part 740 and approved for the items. That approved-entity route builds on the July 2026 rule that moved the UAE into Country Group A:5, which Quentir examined in How One Export-Control Table Opened a New AI Compute Route.
What the rule removes: the AI Diffusion Rule and the foundry presumptions
The rule formally rescinds the regulatory text of the Framework for Artificial Intelligence Diffusion, published on 15 January 2025 at 90 FR 4544, which BIS says it had never enforced and whose rescission it announced in May 2025. The worldwide Regional Stability license requirement and the model-weights control that rule created are removed, with conforming changes to ECCNs from 3A001 to 5D992. The rule also reduces the ways in which front-end fabricators and chip packagers can overcome the existing presumptions on advanced-node logic chips when the chip designer is located in a destination that needs a license for SI chips. Following Executive Order 14434, the text uses "Super Intelligence" (SI) where earlier rules said "artificial intelligence." A savings clause lets shipments already en route on 19 November proceed under their previous eligibility if they are completed within 30 days.
How Quentir Reads It
The rule extends US export jurisdiction from the chip to the customer relationship of a foreign data center. Its operative facts are ownership, aggregate compute and the purpose of a workload, all of which sit in cloud contracts, know-your-customer files and capacity-allocation systems. A Malaysian operator, a Brazilian colocation provider and a European cloud with a Singapore region now have to read their customer lists through the definition of an ultimate parent headquartered in D:5, and they carry that duty for as long as they hold the chips. BIS enforces the rule through what operators know about their customers: the prohibition applies when a party has "knowledge" of a prohibited end use, and under section 744.24(a)(2) the Secretary or the Under Secretary can notify a specific operator, or amend the regulations, so that a license becomes required for particular customers or destinations. Physical verification of where a chip operates is the approach Congress has considered in the Chip Security Act, which Quentir discussed in Could a 5G Radio Inside the GPU Package Meet the Chip Security Act?. The two approaches can coexist, and the IFR does not depend on location hardware.
Three consequences reach beyond export compliance. Governments in the five countries now have a commercial reason to strengthen their own export control enforcement, because BIS ties case-by-case licensing to that enforcement and threatens import restrictions or a downgraded status for countries that tolerate evasion. Recommender-only infrastructure gains a regulatory advantage over general-purpose clusters, which may shape how platforms partition their hardware. And the 60-day comment period is the one formal opening for operators to contest the 12,000,000 TPP threshold or the definition of linked data centers before the rule hardens in practice.
The rule, the UAE reclassification and the Chip Security Act debate belong to one line of compute-control analysis that Quentir has followed since July, and the All-access membership opens that whole archive under one subscription.
Sources: U.S. Department of Commerce, Bureau of Industry and Security, "Addressing Remote Access to Advanced Semiconductors," interim final rule, RIN 0694-AK50, agency text released 16 November 2026 (summary; preamble sections I to VI; savings clause; new sections 743.10 and 744.24; amended section 742.6); Office of the Federal Register, public inspection document 2026-23529, filed 16 November 2026, scheduled for publication on 19 November 2026; BIS, Framework for Artificial Intelligence Diffusion, 90 FR 4544, 15 January 2025; Bureau of Export Administration, Simplification of Export Administration Regulations, proposed rule, 60 FR 25268, 11 May 1995, Supplement No. 4 to part 742 (supercomputer safeguards plan); NVIDIA, H100 Tensor Core GPU specifications (H100 SXM). The H100 comparison is Quentir's own calculation from the ECCN 3A090 TPP formula and is not stated in the rule. Federal Register citations for the new rule will be available after publication. Links checked 17 November 2026.
Published intelligence, built to inform your own decisions. Published: November 17, 2026.