US BIS Caps Overseas AI Chip Access at 12 Million TPP in New Export Rule
The US BIS issued a rule capping remote access to American AI chips at 12 million TPP, adding licenses for five countries and a 90-day transition window.
AI SummaryAI
- BIS issued an interim final rule on November 15 capping remote AI compute at 12 million aggregate TPP.
- New license requirements cover exports to Indonesia, Malaysia, Singapore, Thailand and Brazil.
- Existing data centers get 90 days from Federal Register publication to obtain authorization.
- The rule PDF still shows a placeholder instead of an effective date and an unfinished ID number.
BIS Places Remote Compute Under Export Control
The US Commerce Department’s Bureau of Industry and Security (BIS) has issued an interim final rule that extends advanced-chip export controls to large-scale remote computing capacity supplied by overseas data centers, and adds license requirements for five countries including Singapore. The measure, announced by the agency, targets companies headquartered in restricted regions, chiefly China and other D:5 group countries, or in Macau, and is designed to stop existing technology limits from being bypassed through remote-computing arrangements in third-country facilities.
The control threshold is set at more than 12 million in aggregate Total Processing Performance (TPP), a measure of chip computing capability rather than a count of chips. A license is required for any export, reexport or in-country transfer when the supplier knows the destination data center will make compute above that threshold available to companies in restricted regions. The rule also addresses distributed setups: multiple data centers linked physically or digitally that pool resources must have their performance measured on a combined basis, so splitting capacity across sites does not dodge the test.
A second track draws lines by destination. Designated advanced computing items bound for Indonesia, Malaysia, Singapore, Thailand and Brazil now face new license requirements, with reexports and in-country transfers included and applications reviewed case by case. An exclusion survives where the end customer, or its ultimate parent, is headquartered in the United States. Exporters must report the parties, product specifications, quantities and final destination of covered transactions every month.
The remote-compute provisions carve out commercial recommendation systems built solely to rank, filter and recommend products, services or content, provided the operator files the users served, aggregate performance and due-diligence details. Generative AI does not qualify: any system producing text, images, audio, video or code from user inputs falls outside the exception, even when it also carries recommendation features. Data centers already serving large-scale compute before the rule take effect have 90 days from publication in the Federal Register to obtain authorization. Supply-chain firms must in turn obtain customer-screening information from overseas data centers, and a facility that cannot or will not provide adequate verification is treated as a red flag to be addressed.
A Rule With No Date Filled In Yet
BIS dated its announcement November 15 and framed the purpose as keeping US restrictions on sensitive technology intact against remote access routed through data centers outside restricted countries. The agency’s logic is that equipment installed abroad can be reached over a network without the chips ever crossing a restricted border, and that cloud computing and emerging superintelligence infrastructure have made remote access to advanced American compute practical at scale. BIS paired the restrictions with a stated aim to encourage lawful commercial adoption of US technology worldwide.
The published text, however, is not final on its own timeline. The rule document leaves the effective date as “INSERT DATE OF PUBLICATION IN THE FEDERAL REGISTER”, keys the comment deadline to the same publication, and carries the regulatory identification number “BIS-2026-[XXXX]”, with “CUI, Deliberative Process” markings repeated across pages. The announcement itself lists no country-by-country application table and no operator-by-operator duties, so it does not by itself establish a blanket ban on using AI chips through foreign cloud providers.
A parallel legislative track remains separate. House records show the Remote Access Security bill, H.R. 2683, which would bring remote access into the US export-control framework, passed the House 369 to 22 on January 12 and was referred the next day to the Senate Banking, Housing and Urban Affairs Committee; its status is listed as passed House. NVIDIA (NVDA) has acknowledged exposure to this policy area in its own filings: its quarterly report dated as of July 26 states that restrictions on providing products and services supporting third-party applications and models could materially affect its business, operating results and financial condition. That is a general risk disclosure, not a comment on this specific rule.
The shift is structural: control moves from where a chip physically sits to whose workloads it ultimately serves, which is exactly the gap the third-country rental route had opened. Compliance duties, from monthly reporting to customer vetting that now reaches into data-center operations, take practical effect even though the effective date is still an unfilled placeholder, leaving suppliers and overseas operators preparing against a clock that has not started. The threshold figures in export rules are compliance tests, not market support and resistance levels, and we read them as such; the event that matters next is the Federal Register publication that fixes day one of the 90-day transition and starts the comment clock.
Primary sources
- announced by the agency · bis.gov
- rule document · bis.gov
AI-generated, AI-reviewed, under COINOTAG editorial oversight.

