US Export Rules Now Follow AI Chips Into Overseas Data Centers

The U.S. Commerce Department will require licenses to export, reexport or transfer advanced AI chips and servers to data centers anywhere in the world if those facilities may provide compute to companies in China or other restricted countries, according to TechCrunch. The requirement appears in an interim final rule from the Bureau of Industry and Security, which states that its purpose is to stop existing restrictions on sensitive American technology from being bypassed through remote computing arrangements in third countries.
The rule addresses remote access. A customer in a restricted jurisdiction could use compute housed in a facility that never receives hardware from that customer. The Bureau's stated aim is to bring protections for remote access to advanced computing in line with the controls that already govern physical shipments of sensitive U.S. technology. The remote-access provisions target the D:5 group of embargoed countries, which includes China, Russia, Iran and North Korea, along with 18 others.
The threshold is measured in total processing performance (TPP), a single score for how much raw computing power a system can deliver. The rule applies to data centers, or any network, capable of supplying more than 12 million TPP. According to the same report, that ceiling corresponds to roughly the maximum computing power of about 750 Nvidia H100 or H200 chips. The threshold therefore captures large clusters rather than individual servers, although the figure is expressed as a performance ceiling, not a unit count.
Indirect access is covered too. The rule reaches arrangements that run through intermediaries or brokers, and it requires exporters and data center operators to monitor customers' ownership structures and activities. The purpose of that monitoring is to confirm that a customer is not run by anyone from the D:5 group. For operators, know-your-customer diligence, the process of verifying who a client really is, becomes a compliance obligation tied to beneficial ownership, meaning the people who ultimately own or control a company.
Five additional jurisdictions receive specific treatment. The rule adds license requirements for exports, reexports or transfers of advanced semiconductors to Malaysia, Singapore, Indonesia, Thailand and Brazil, which the rule describes as presenting higher-than-normal risk of diversion, meaning chips redirected to unauthorized users. The license requirements will not apply to chips destined for companies based in the United States.
Existing data center operators that provide covered services must obtain licenses within 90 days of the rule taking effect. The rule does not apply to commercial recommendation systems, such as those used in e-commerce or on social media. Those systems make up a large share of production AI workloads, and the carve-out keeps the rule aimed at general-purpose compute.
The rule also rescinds the Biden administration's AI Diffusion Rule, issued in January 2025. That earlier framework set country-level limits on exports of advanced AI chips and restricted transfers of certain powerful AI models. The new interim final rule replaces that country-based design with a control keyed to the compute itself and to who ultimately operates it.
The broader context here is a change in where the control point sits. Export controls on hardware were built around a physical object crossing a border, which is easy to define and to inspect. A rule that follows compute wherever it is hosted requires exporters and operators to verify ownership and end use across a chain of brokers, cloud tenants and foreign facilities. Whether that verification can be done reliably at scale is a question the 90-day licensing window will test in practice.
In my view, the most useful reading for infrastructure teams is the mechanics. The TPP threshold, the broker language and the ownership-monitoring duty all require operators to inventory their capacity and their customers, not just their shipments. Operators with multi-tenant GPU fleets in the named countries, or with customers that have indirect ties to the D:5 group, will have the most work to do before the deadline. Those that already maintain detailed customer due diligence will find the transition more manageable than those that do not.


