On November 16, 2026, the US Department of Commerce issued an interim rule requiring a license for the export, re-export, and transfer of AI chips and servers to data centers worldwide if there is a risk that the computing power will be used by companies from China, Russia, Iran, North Korea, and 18 other countries in the D:5 group. For the first time, control extends not only to physical delivery but also to the remote use of computing power through intermediaries—i.e., renting GPUs in third countries.

What happened
The threshold is defined quantitatively: license control applies to data centers with a total performance exceeding 12 million TPP, which roughly corresponds to a cluster of 750 Nvidia H100/H200 accelerators. The control covers the export, re-export, and transfer of AI chips and servers, and indirect access through intermediaries is explicitly named within the rule's scope: renting such capacity to companies with D:5 risk now requires a license. Existing operators of affected data centers must obtain licenses within 90 days, while exporters and operators are simultaneously required to verify the ownership structure of their clients. The document is drafted as an interim rule.
Context
This measure changes the very logic of export control: the question is no longer "where was the chip physically delivered," but "who can remotely reach the computing power." The trigger was the practice where Chinese labs, deprived of direct access to advanced Nvidia chips, rented computing power on American hardware in third-country data centers—the new rule closes exactly this channel. The document also repeals the AI Diffusion Rule of the Biden administration, presented in January 2025. This is the first time the US has tied export control not to a specific chip deal, but to the total performance of a data center and the beneficial user.
Why this matters for the industry
For the industry, the economics of GPU rental change: it ceases to be a neutral service, as deals with a risk of use by D:5 companies face a licensing barrier. In the coming weeks, pauses and renegotiations of terms in third-country rental chains are likely, as operators and exporters determine whether their installations fall under the 12 million TPP threshold. Some providers will likely refuse clients from D:5, while others will integrate the licensing process into onboarding: questionnaires on ownership structure and separate pricing for "verified" access. A niche opens up for compliance automation and computing power metrology—the access verification layer itself becomes a product.
Why this matters for users
For readers following AI infrastructure, this means that access to advanced Nvidia chips for model training will become more difficult and expensive for teams under sanctions restrictions. Developers in the US and allied countries are not directly prohibited from anything—basic access remains—however, cloud providers will pass new obligations on to clients in the form of questionnaires when renting and restrictions on booking large clusters. For users and companies from Russia and other D:5 countries, legal access to large rental clusters abroad is narrowing.
What is currently unknown / limitations
Confidence in the conclusions is limited: the rule is interim, and the evidence base consists of two publications (TechCrunch and Reuters via Yahoo News); the primary text of the Department of Commerce document is not quoted in the materials. The equivalence of "12 million TPP — approximately 750 Nvidia H100/H200 chips" is an approximate conversion from media sources, not a verified technical specification, so the threshold boundaries for specific data center configurations are currently unknown. It is unclear how the licensing practice will be implemented in reality, whether the interim rule will become permanent, and whether bypass schemes through sub-threshold clusters will emerge.
Sources
- TechCrunch — US tightens AI chip export controls to deter Chinese companies from accessing compute remotely
- Reuters (via Yahoo News) — US issues rule to protect advanced AI chips from being accessed remotely
Author
Look at AI, editorial team
