technology 6 min read

Nvidia fallout: Commerce closes third-country GPU loophole for China

The Commerce Department has closed a critical export-control loophole letting Chinese firms lease Nvidia GPUs from data centers in Singapore, Malaysia, Indonesia and other transit countries. The move reshapes the gray-market cloud AI economy—and comes amid a wider scandal that already cost Jensen Huang his job.

  • NVIDIA
  • US-China Tech Rivalry
  • BIS
  • China Semiconductor Export Controls
  • Cloud GPU Computing

Closing the door on cloud-bridged chips

The U.S. Commerce Department has closed one of the most consequential loopholes in American export policy — the ability of Chinese companies to bypass chip bans by renting Nvidia computing power from data centers in third countries.

The interim final rule, published Monday by the Bureau of Industry and Security, requires a license to export or reexport advanced Semiconductors to Indonesia, Malaysia, Singapore, Thailand or Brazil, plus any country the Secretary designates later. The language is blunt: it targets entities “headquartered in countries of concern” that use U.S.-built Super Intelligence chips to train cutting-edge AI models abroad, then consume those models remotely from China.

This is not a minor administrative tweak. It is the first substantive policy response from Washington since the scandal surrounding Nvidia CEO Jensen Huang erupted last week — a scandal that ended with his resignation over allegations of affairs with Alice Huang, a Megaspeed executive U.S. intelligence believed to be a Chinese spy.

Megaspeed, a Singapore-based company at the center of the controversy, had been leasing Nvidia GPUs to Chinese clients precisely because Beijing could not buy them directly. That business model now faces an immediate compliance reckoning.

Who wins and who loses

Chinese cloud-AI firms are the clearest losers in this round. Alibaba and ByteDance, named in a letter from Sen. Maria Cantwell (D-Wash.), ranking Democrat on the Senate Commerce Committee, have relied on overseas data centers loaded with advanced Nvidia silicon to train frontier models. The new rule means every transaction through the listed countries now runs through a licensing gate — one that carries an enforceable presumption of denial for end-users tied to China’s military-civil fusion strategy, as the department put it.

Singapore’s hyperscale cloud footprint shrinks almost overnight as a strategic asset for Beijing. The island state hosted major data-center build-outs specifically because its proximity to Southeast Asian demand and favorable regulatory environment made it a natural hub for cross-border compute trading. Thai and Brazilian operators will feel similar pressure; both jurisdictions have courted Western chip manufacturers looking for alternative markets.

Nvidia itself sits in an awkward middle ground. The company lost its chief executive during a scandal that grew out of its commercial relationship with Megaspeed, yet the new rules also tighten restrictions on where its products can physically reside when used by restricted parties. The company’s revenue from data-center sales will face downward pressure in countries previously treated as safe conduits. On the other hand, the rules may reduce competitive pressure from Chinese firms that were using subsidized compute to narrow the AI gap faster than Nvidia had anticipated.

Megaspeed is worst hit of all. The company’s entire value proposition rested on bridging the geographic gap between embargoed chips and Chinese demand. Its Singapore base, once an advantage, is now exposed. Treasury Secretary Scott Bessent publicly blamed Commerce for not placing Megaspeed on the entity list sooner — a rebuke that signals internal friction and suggests further sanctions are likely.

How the loophole operated

The architecture was deceptively simple. A Chinese firm would sign up for cloud compute in Singapore, Malaysia or a similar jurisdiction. The provider would pre-install Nvidia H100-class chips, which U.S. law barred from direct sale to China. The Chinese customer would then run training jobs remotely through VPNs or other network arrangements, never taking physical possession of the hardware but consuming its computational output.

Under prior rules, this gray-market arrangement existed in a legal blind spot. Export controls governed the movement of chips across borders, not the delivery of compute services derived from those chips. As long as the hardware stayed physically outside China, and the end-user agreement pointed to a non-Chinese subsidiary, the deal cleared existing regulatory gates.

Senator Cantwell flagged exactly this gap in her November 15 letter to Commerce Secretary Howard Lutnick. She wrote that Chinese firms were exploiting “a loophole in U.S. export controls by using advanced American chips in foreign data centers even when those chips are restricted from the People’s Republic of China,” noting that BIS had issued warnings about certain arrangements without triggering enforcement action.

What happens next

The interim final rule leaves several questions open. The text does not enumerate every type of cloud service it covers — it refers broadly to “advanced semiconductors” used for “training advanced SI models.” Whether inference-only workloads qualify, whether existing contracts are grandfathered, and how aggressively the rule applies to smaller Southeast Asian providers remain unclear. The Commerce Department said additional countries may be added after consultation with State and the National Security Council.

That consultation matters. Secretary of State Marco Rubio warned President Donald Trump last spring that Jensen Huang maintained a close relationship with Alice Huang, whom intelligence agencies considered a Chinese operative. The New York Times reported the warning on November 8. Within days, Huang resigned. Now the NSC is directly shaping the rule’s geographic scope — signaling that this crackdown is part of a broader strategic effort, not a standalone regulatory adjustment.

China will push back. Expect Beijing to accelerate domestic GPU development and explore alternative routing through non-listed countries, including potentially African and South Asian jurisdictions that have not yet been designated. Some Chinese firms may shift to lower-tier chips that fall below the rule’s thresholds, though that limits training throughput for frontier models.

For Nvidia’s competitors — AMD, Intel and China’s own Huawei — the effect is mixed. The crackdown reduces near-term demand from the gray-market channel but also removes a competitive advantage some Chinese firms held over Western startups that lacked access to subsidized overseas compute.

Why this matters beyond chip politics

This rule change reflects a fundamental shift in how Washington treats computational power as an export-controlled good in its own right, rather than only controlling the physical chips themselves. The distinction matters because it transforms every cloud provider operating in the listed countries into a de facto extension of U.S. export enforcement. Companies that built their businesses on serving Chinese clients through regional data centers now face either compliance overhauls or exit.

It also raises a longer-term question about the feasibility of AI containment in an era where silicon can travel thousands of miles via fiber optic cables. If the United States can criminalize the consumption of advanced compute produced on its technology regardless of geography, the traditional boundary between domestic and foreign AI development blurs significantly. Other nations — particularly allies hosting hyperscale data centers — will need to decide whether they want to become enforcing agents in a contest that has increasingly become an economic and ideological fault line.