top of page
Sage Logo Full_edited.png
Search

Washington Closes the AI Chip “Remote Access” Loophole — What It Means Beyond Nvidia

Writer: Sage
Sage
4 days ago
3 min read

On November 16, 2026, the U.S. Commerce Department's Bureau of Industry and Security (BIS) issued an interim final rule that closes one of the most consequential gaps in American export-control policy: the ability of companies in China and other restricted countries to access advanced AI computing power remotely, through cloud data centers, without ever owning the underlying chips. The rule lands just days after Nvidia's board accepted the resignation of CEO Jensen Huang amid a widening controversy over how advanced AI chips have continued to reach China despite years of export restrictions. For any organization that builds on, resells, or depends on AI infrastructure, this is a regulatory moment worth watching closely — well beyond the chip industry itself.


What the new rule actually closes

Export controls on advanced AI chips have historically focused on physical hardware: who can buy an Nvidia H100 or H200, and where it can be shipped. But as AI workloads increasingly run in the cloud, that focus left an obvious workaround — a restricted buyer could simply rent compute time in a data center located somewhere with looser rules, rather than import chips directly. BIS's new interim final rule is built to close exactly that loophole.

Under the rule, exporting, re-exporting, or transferring advanced AI chips and servers to any data center or network worldwide now requires a license if that facility could end up providing compute to companies based in China, Russia, Iran, North Korea, or the roughly twenty other countries and territories in the U.S. government's “D:5” group of embargoed destinations. The threshold is defined technically: any cluster above 12 million in total processing performance, which works out to roughly 750 H100- or H200-class chips. Operators and exporters are also now expected to monitor the ownership and activity of their customers, closing off the practice of routing access through intermediaries or brokers. Five additional jurisdictions — Malaysia, Singapore, Indonesia, Thailand, and Brazil — were added to the license-requirement list, which Commerce said reflects “a higher-than-normal risk of diversion.” Existing data-center operators providing covered services have 90 days to secure the required licenses. The rule also formally rescinds the Biden administration's AI Diffusion Rule from January 2025, replacing a tiered, country-by-country framework with this more targeted, compute-and-ownership-based approach.


Why now: the Nvidia leadership shake-up

The timing is not a coincidence. Nvidia's board accepted Jensen Huang's resignation on November 11 and announced it the following day, naming CFO Colette Kress as interim CEO and independent director Harvey C. Jones as the company's first independent board chair. The departure followed weeks of reporting on Huang's personal conduct, alongside separate and more consequential scrutiny: congressional investigators and industry analysts have examined allegations that large volumes of restricted AI chips — some estimates put the figure as high as 1.5 million GPUs — reached China through diversion schemes despite existing export controls. Lawmakers from both parties have opened inquiries and called for testimony, and Nvidia itself has signaled it is pulling back from the Chinese market, where chip sales had already fallen to roughly 8% of revenue as controls and Beijing's push for domestic alternatives took hold. Commerce's new rule is the policy response to that same underlying problem: hardware controls alone cannot prevent diversion once compute can be accessed remotely.


The takeaway for risk and compliance teams

The practical lesson here extends well past semiconductor companies. Any business that rents GPU capacity from a cloud provider, resells AI infrastructure, integrates AI systems into a supply chain, or is evaluating an AI vendor in diligence now has a new category of exposure to map: where does the underlying compute actually sit, who owns the data center it runs in, and could that facility plausibly serve a customer in a restricted jurisdiction? Vendor risk questionnaires built around physical hardware provenance will need to expand to cover compute sourcing and remote-access arrangements, and teams managing cloud or AI vendor relationships should flag the 90-day licensing window now, before it becomes a surprise contractual gap. More broadly, this is a clear signal that export-control enforcement is catching up to how AI is actually delivered — as a service, not just as a shipment — and governance frameworks that still treat AI risk as a hardware problem are already behind.


Sources: TechCrunch, NPR, and CBS News reporting on the Commerce Department's November 16, 2026 interim final rule and Nvidia's leadership transition.

 
 
 

Recent Posts

See All

Comments


bottom of page