The Commerce Department is drafting a rule that would block Chinese artificial intelligence companies from renting computing time on advanced Nvidia chips housed in Southeast Asian data centers, according to people familiar with the effort cited by Tom's Hardware, with a version of the rule possibly shared with trade groups as soon as this month.
The measure, developed by the department's Bureau of Industry and Security, would for the first time treat remote, cloud-based access to export-controlled AI chips the same way Washington already treats a physical shipment of the hardware. Existing rules bar Nvidia from selling its most advanced processors directly into mainland China, but they say little about a Chinese firm logging into a GPU cluster in Bangkok or Singapore and renting compute by the hour — a gap officials and lawmakers now describe as the industry's most consequential loophole.
A gap widened by policy reversals
The loophole traces back to last year, when the Trump administration shelved the Biden-era AI Diffusion Rule, a tiered licensing framework that would have required cloud providers to vet and report end users renting export-controlled compute. Without it, Chinese firms spent roughly a year routing purchases and rentals through subsidiaries and resellers in Singapore and Malaysia. The Bureau of Industry and Security moved to close part of that gap on May 31, when it issued guidance clarifying that chips ultimately controlled by Chinese or Macau-based owners require a license regardless of which country the buying entity is registered in — a clarification that followed reporting that Singapore-based Megaspeed had routed nearly $2 billion in Nvidia chips to a Malaysian data center serving Chinese clients.
The new draft rule goes further, aiming at remote access rather than ownership: a Chinese company would not need to buy or import a single chip to fall under the rule if it is simply renting time on one sitting offshore. Analysts note the underlying demand is unlikely to fade on its own — Chinese-made accelerators from Huawei are closing the raw-performance gap with Nvidia's prior-generation chips, but still lag badly on the software tooling that keeps engineers locked into Nvidia's ecosystem.
The case that accelerated the rule
The push gained urgency after Michael Kratsios, director of the White House Office of Science and Technology Policy, alleged in July that Chinese startup Moonshot AI had accessed Nvidia GB300 chips housed in Thailand to help train its Kimi K3 model, in a case first reported by CNBC. Kratsios said Moonshot had built what he called a platform to conduct large-scale distillation against U.S. models while switching between methods of access to avoid detection.
A sophisticated internal platform to conduct large scale distillation against U.S. models, allowing them to quickly switch between multiple methods of access to avoid detection.
Michael Kratsios, director, White House Office of Science and Technology Policy
Moonshot has not publicly detailed how it obtained the compute it used. Congress had already moved to give BIS clearer authority to act on cases like it: the House passed the Remote Access Security Act, which would redefine cloud-based access to controlled technology as a regulated export, by a 369-22 vote on H.R. 2683 in January. A Senate companion bill remains pending in the Banking Committee. Export-control lawyers say that gap matters: without RASA's passage, it remains legally uncertain whether BIS has clear statutory authority to police purely remote, intangible access to compute rather than the transfer of a physical good — an ambiguity the draft rule will have to navigate.
Industry pushback and what comes next
Major American cloud providers, including AWS, Microsoft Azure and Google Cloud, have privately raised concerns that the rule would impose heavy compliance burdens on their overseas facilities and could push customers toward non-U.S. cloud platforms. Nvidia and Southeast Asian data-center operators face similar uncertainty: neither Singapore nor Malaysia currently restricts data centers based on the identity of the compute customer, meaning the rule would rely on the extraterritorial reach of U.S. export licensing to have any effect.
Commerce has not confirmed a publication date, and any formal rule would likely go through a comment period before taking effect. But officials working on the draft have signaled they hope to share language with industry before the end of September, giving cloud operators and chipmakers a narrow window to weigh in before a policy that would reshape how compute is bought, sold and rented across the region.
