Commerce Closes the Remote-Compute Loophole for Advanced AI Chips
BIS released an interim final rule on 16 November requiring licences when overseas data centres give Chinese or D:5 firms over 12 million TPP of AI compute.

On this page
Key takeaways
- The Bureau of Industry and Security released an interim final rule on 16 November 2026 covering remote access to advanced AI chips and servers.
- A licence is required when a data centre anywhere in the world provides more than 12 million aggregate TPP of compute to a company headquartered in Macau or a Country Group D:5 destination.
- Twelve million TPP is roughly the maximum computing power of 750 Nvidia H100 or H200 chips.
- Existing data-centre operators have 90 days from the rule's effective date to obtain a licence, and exports to five Southeast Asian and Latin American countries face new licence requirements.
- The rule rescinds the Biden-era AI Diffusion Rule and leaves its Federal Register publication and effective dates to be set.
The U.S. Bureau of Industry and Security released an interim final rule on Monday 16 November 2026 that extends export controls from the physical shipment of advanced AI chips to the remote use of those chips in data centres abroad. The rule is aimed at a gap that has worried Washington for more than a year: a Chinese company can rent computing capacity in a third country without ever taking delivery of a restricted chip.
Related reading: FCC Tells California to Stop Blocking AT&T's Copper Retirement · California Subpoenas Nvidia Over Whether Its Chips Reached China.
The rule requires a licence when a person knowingly exports, reexports or transfers covered AI chips or servers to a data centre anywhere in the world that provides, or will provide, more than 12 million aggregate total processing performance (TPP) of computing access to a company headquartered in Macau or a Country Group D:5 destination, or whose ultimate parent is headquartered there.
D:5 is the U.S. list of countries subject to a UN arms embargo. It includes China, Russia, Iran and North Korea, along with 18 others. The 12 million TPP threshold is roughly the maximum computing power of 750 Nvidia H100 or H200 chips.
Why remote access became the loophole
Existing export controls already require a licence to ship advanced AI chips to companies based in, or owned by, a D:5 country. The new rule treats remote access as a transfer of the same technology. BIS says the control is a logical extension of the shipment rules and that it is meant to stop companies from using third-country data centres to train advanced models.
The measure had been reported as a draft in the days before publication, with a shifting list of third countries and no final text. The published rule goes further than those reports in one respect: the data-centre control covers the whole D:5 group, not only the handful of countries named for the separate destination licences.
The rule lands days after Nvidia co-founder Jensen Huang resigned as chief executive amid a scandal involving an alleged Chinese spy linked to one of the chipmaker's customers. The U.S. government has reportedly wanted to add that customer, Megaspeed, to a special entity list over allegations that it bought thousands of Nvidia chips to supply compute to Chinese businesses. Reuters reported last week that Chinese AI companies were rushing to secure computing capacity overseas ahead of tighter restrictions.
BIS says the rule was written by, and at the direction of, the Treasury Department and the National Security Council. That phrasing gives the control a broader political mandate than a routine BIS update and signals that the Treasury's national-security tools are now part of the chip enforcement stack.
Linked data centres and brokers are in scope
The rule is not limited to a single building. The 12 million TPP threshold can be reached at one data centre or across linked data centres, including facilities that are digitally or physically connected, share resources, balance workloads or replicate data as a unified whole. That closes the obvious workaround of splitting a large cluster across several sites.
It also covers indirect access. A data-centre customer outside D:5 can bring a transaction into scope if it trains models for, participates in training models for, or supplies models to a D:5 company. Exporters and foreign data-centre operators are expected to run know-your-customer checks on direct and indirect ownership or control by Macau or D:5 persons.
A refusal or inability to provide enough verification information is treated as a red flag that presumptively supplies the knowledge that triggers the licence requirement, unless the exporter resolves the question before proceeding. Compliance also continues after shipment: moving covered equipment from a permitted use to a prohibited one counts as an in-country transfer that needs authorisation.
Existing data-centre operators that were already providing significant compute to covered companies before publication must obtain a licence within 90 days of the effective date to keep operating, or face future export restrictions.
Five countries get separate licence requirements
The rule adds licence requirements for exports, reexports and transfers of covered advanced semiconductors to or within Indonesia, Malaysia, Singapore, Thailand and Brazil. Commerce says those countries present a higher-than-normal risk of diversion.
Transactions for end users headquartered in the United States, or whose ultimate parent is U.S.-headquartered, are excluded from those five-country requirements, but only if the exporter files monthly reports. The reports must identify the parties and parents, destinations, end users, product classifications, names and stock-keeping units, TPP, quantities and transaction dates, and are due by the first business day of the following month.
That reporting duty turns the exemption into a paperwork regime. It also gives BIS a transaction-level view of where advanced chips are going in the five countries, which could feed future enforcement or narrower controls.
The rule covers data centres that provide more than 12 million aggregate TPP. It does not set a limit on the number of chips a single facility can hold, and it does not apply to consumer cloud services that stay below the threshold.
The recommendation-system exception is narrow
The prohibition does not apply when computing access is used exclusively for a commercial recommender system, and the required information is submitted to BIS for sharing with the End User Review Committee. The exception covers systems designed and used only to rank, filter or suggest commercial products, services or content.
It excludes systems that generate novel text, images, audio, video, software or other content, including recommendation products that incorporate such models. The submission must identify the covered company, report the total TPP provided and describe the due diligence that establishes exclusive recommender use.
In practice, the exception is for advertising and e-commerce ranking, not for general-purpose AI. A company that runs a recommendation model and a chatbot in the same data centre cannot claim the exception for the whole facility.
The AI Diffusion Rule is formally gone
The rule rescinds the Biden administration's AI Diffusion Rule from January 2025. That framework tried to impose country-level limits on advanced AI chip exports and to restrict transfers of certain powerful AI models. The Trump administration said in May 2025 that it would throw the framework out, arguing that it burdened U.S. companies and allies, but the formal rescission had not been completed.
The new rule removes the model-weights control and its foreign-direct-product rule, the AI Diffusion country tiers, the Artificial Intelligence Authorization and Low Processing Performance licence exceptions, and the Data Center Validated End User Authorization. It replaces them with destination and end-use controls tied to computing capacity and the ultimate user.
The result is a regime that is narrower in some ways and broader in others. It drops the country-tier system that allies disliked, but it reaches into data centres in countries that are not themselves subject to the shipment controls.
What is still unresolved
The on-display text leaves the Federal Register publication date, effective date and comment deadline as placeholders. That means the rule is announced but not yet on a compliance clock. BIS has not said whether the interim final rule will take effect on publication or after a short delay, and the comment period could produce changes before enforcement begins.
The legal basis is also likely to be tested. The rule stretches the Export Administration Regulations to cover computing access rather than hardware shipment, and it relies on a knowledge standard that puts a due-diligence burden on exporters and data-centre operators. Courts have not reviewed this specific theory.
Enforcement will be the harder test. BIS can police chip shipments at the border, but it cannot easily inspect every data centre in a third country. The rule's know-your-customer requirements and 90-day licensing window give it a paper trail, not a guarantee.
What to watch
The first signal will be the Federal Register notice, which will set the effective date and the comment deadline. After that, watch whether BIS publishes guidance on what counts as "significant compute" for the 90-day transition and how it will treat shared or sublet data-centre capacity.
The second signal is the reaction from cloud providers and allied governments. Singapore, Malaysia and Brazil have all courted AI data-centre investment, and the new licence requirements could slow projects that were counting on U.S. chips. The U.S. reporting exemption for American-headquartered end users may soften the impact, but it adds a monthly filing burden that smaller operators may struggle to meet.
For chipmakers, the rule is another layer of compliance on top of the shipment controls that already shape their China business. For Chinese AI companies, it narrows the overseas route that many had been rushing to secure. The loophole is closing, but the rule's reach will depend on how aggressively BIS enforces it.
The rule's date and comment period were still placeholders in the on-display text on 16 November. The 12 million TPP comparison to 750 H100 or H200 chips is an approximation based on the export-control metric, not a BIS statement about any specific configuration.
Tags
Frequently asked questions
- What does the new BIS rule cover?
- It covers exports, reexports and transfers of advanced AI chips and servers to data centres anywhere in the world that provide, or will provide, more than 12 million aggregate TPP of computing access to companies headquartered in Macau or Country Group D:5, including China and Russia.
- When does the rule take effect?
- The on-display text leaves the Federal Register publication date, effective date and comment deadline as placeholders. BIS released the rule on 16 November 2026, but the compliance clock starts when the dates are set in the Federal Register.
- Which countries are affected?
- The data-centre control covers Macau and Country Group D:5 destinations, which include China, Russia, Iran and North Korea. Separate licence requirements apply to Indonesia, Malaysia, Singapore, Thailand and Brazil.
- Does the rule ban Chinese companies from using overseas data centres?
- It does not ban the activity outright. It requires a licence for covered transactions above the computing threshold and sets a case-by-case review, with a narrow exception for commercial recommendation systems.








