Renting AI compute in a third country is now covered by U.S. export control in the same way shipping the chips is. On Nov. 16 the Commerce Department’s Bureau of Industry and Security (BIS) released an interim final rule that requires a license before advanced AI chips go to any data center, anywhere, that gives large amounts of computing power to companies headquartered in Macau or in Country Group D:5, the group that includes China. The same rule adds a license requirement for chips bound for Indonesia, Malaysia, Singapore, Thailand and Brazil.
What the rule does
The BIS announcement says the rule was “written by and at the direction of the Department of Treasury and the National Security Council” and is meant to stop existing chip restrictions being bypassed through remote computing in third countries. The rule text, filed as RIN 0694-AK50, makes four changes.
A threshold for remote access
A new section of the Export Administration Regulations, 744.24, bars exporting, reexporting or transferring advanced chips and servers without a license when the sender knows they are headed to a data center that provides, or will provide, more than 12,000,000 “aggregate TPP” to or for a company headquartered in Macau or a D:5 country, or one whose ultimate parent is. TPP is total processing performance, the measure BIS already uses to classify chips. The total counts capacity at one site or across linked data centers, and the rule says linking includes digital connections as well as physical ones. The rule gives the threshold only in TPP and does not convert it into a number of chips.
The control follows the hardware after delivery. If a data center that received chips lawfully later starts serving such a customer above the threshold, BIS treats the change as a transfer and the operator needs authorization. Data centers already providing that capacity must obtain a license within 90 days of the effective date “or face future export restrictions”. The phrase “to or for” also reaches a customer from outside those countries that trains models for, or supplies models to, a D:5 company.
Five more destinations
Chips going to Indonesia, Malaysia, Singapore, Thailand or Brazil now need a license, reviewed case by case. BIS says it added the five because of concern that controlled chips are being diverted from them without authorization. The requirement does not apply when the ultimate consignee or end user is headquartered in the United States, or has a U.S. ultimate parent, as long as the exporter files a monthly report naming every party and its parents, the destination, each product’s name, SKU, classification and TPP, the quantity and the date. Reports are due on the first business day of the month after the transaction.
Foundries and chip packagers
The rule narrows the ways a foundry or packaging company can show that an advanced logic chip it makes for a designer in a restricted destination falls outside the controls.
The AI Diffusion Rule is formally gone
The rule removes the regulations behind the January 2025 Framework for Artificial Intelligence Diffusion: the country tiers, the Artificial Intelligence Authorization license exception, the data center validated end-user program and the export control on model weights (classification 4E091). BIS announced the rescission in May 2025, and the new text says of that framework: “The AI Diffusion Rule has never been enforced.”
Due diligence becomes the cloud customer’s problem too
The rule turns on what the sender knows, and BIS spells out what it expects. An exporter is expected to obtain, and be able to hand over, information from the foreign data center about that data center’s “know your customer” checks, including whether any customer is directly or indirectly owned or controlled by persons in Macau or a D:5 country. A data center that declines or is unable to provide that information is treated as a red flag that presumptively gives the exporter knowledge.
That puts the paperwork on the customer’s desk. A data center operator that wants its next shipment of GPUs has to show its supplier who its tenants are and who owns them. In our reading, every company renting large blocks of GPU capacity outside the United States should expect ownership questionnaires from its providers, whatever its own nationality.
One exception exists. The prohibition does not apply when the computing power is used only for a “commercial recommender system”, meaning a tool that ranks, filters or suggests products, services or content. Generative systems are excluded from that definition even when they also make recommendations, and the operator has to report the customer’s name, the total TPP supplied and the diligence it performed. The document is not consistent on this point: one passage in the preamble says operators serving recommender systems may continue for 90 days and must then obtain a license, while section 744.24 says the prohibition does not apply to them once the report is filed.
What it means for the buying committee
U.S.-headquartered enterprises keep access to capacity in the five countries, at the price of a monthly report. If the company owns the servers in a colocation site in Singapore or Malaysia, its supplier will ask for the corporate chain up to the ultimate parent. If it rents from a cloud provider there, the provider receives the chips, so in our reading the provider’s headquarters decides whether the exclusion applies.
Enterprises headquartered elsewhere, including in the UK and the EU, get no such exclusion in those five countries. Chips shipped to them there need a license, and the rule states no timeline for the case-by-case review. Plans for a Southeast Asian or Brazilian AI region should now carry a licensing step. Sovereign AI projects such as the NVIDIA, Rebellions and SK Telecom racks planned for 2028 show how long such builds already take without one.
Companies with a parent in a D:5 country or Macau, and vendors that train or host models for one, face the hard limit. Above the threshold their provider needs a license for the underlying chips, and BIS says no license exception overrides the prohibition except where it notes otherwise.
Model builders lose one constraint: with classification 4E091 removed, the January 2025 export control on model weights no longer exists. BIS says it prefers destination and end-use license requirements for weights. Buyers comparing open-weight options can read our piece on how Mistral licenses the Large 4 weights.
BIS expects modest volumes. Its paperwork estimate is about 100 additional license applications and about 970 additional report submissions a year. For buyers the cost sits in contract terms and lead times for capacity abroad. Building at home continues alongside: this month Google announced a $17 billion data center in Wyoming.
What to do before the rule is published
The rule takes effect on the day it appears in the Federal Register, and comments are due 60 days later. The copy on display leaves both dates blank. Shipments already en route on the publication date may complete under the old terms within 30 days.
Three checks are worth making this week. List every non-U.S. site where the company runs or rents controlled GPUs, and mark the ones in the five named countries. Ask each provider how it screens customers for D:5 ownership and whether it serves any above the threshold, because a provider that cannot answer may find its own chip supply held up. Add a contract clause that requires notice if the provider applies for, or is refused, a license covering the capacity you use. Companies that want the recommender-system wording clarified have the comment period to ask.
Source: Bureau of Industry and Security