The US just closed the cloud backdoor to China's AI ambitions the same week Nvidia's CEO stepped down over an alleged espionage scandal.

The Summary

The Signal

For two years, the US has been playing export control whack-a-mole with China's AI industry. Ban the chips, watch them show up via Singapore cloud providers. The Commerce Department just swung the hammer at the moles themselves.

The interim rule targets what officials call "rapid advancement in cloud computing and SI infrastructure" that lets Chinese companies rent GPU clusters in countries not subject to US restrictions. Instead of smuggling physical H100s across borders, Chinese AI labs were spinning up instances in Singapore data centers. Same compute, different jurisdiction.

"Rapid advancement in cloud computing infrastructure made it increasingly possible to access advanced American computing capabilities remotely, including through data centers located outside countries subject to US export restrictions."

The five countries now requiring reexport licenses read like a map of Southeast Asia's data center boom:

  • Singapore: Regional cloud hub, home to major AWS, Google, and Microsoft facilities
  • Malaysia and Thailand: Emerging markets racing to build AI infrastructure
  • Brazil and Indonesia: Large economies with minimal chip export oversight until now

The timing is not subtle. Hours before the rule dropped, Nvidia's stock fell 3.9% in premarket trading as Jensen Huang's resignation became official. The alleged espionage scandal that forced him out involved chip exports, the kind of thing that makes regulators move fast and ask questions later.

What's less obvious: this rule doesn't just hurt Chinese AI labs. It hurts every startup in Singapore, São Paulo, or Bangkok trying to train models on frontier hardware. Want to fine-tune Llama on 128 H100s? Better hope your cloud provider's export compliance paperwork is in order. The collateral damage zone just expanded from China to half the developing world.

"The US will have to proceed very carefully on its stance to limit chip exports to China, as the race for AI dominance continues."

Mary Lovely at the Peterson Institute called it: the US has to proceed carefully here. Too aggressive, and American chip companies lose the entire ASEAN market. Too loose, and Chinese labs train the next generation of models via rental GPUs. The rule tries to split that difference by requiring licenses, not outright bans, but licenses mean delays, compliance costs, and uncertainty.

The backdrop matters. An upcoming Trump-Xi-Putin summit looms, with Lovely setting a "low bar" for outcomes. When three leaders with conflicting AI strategies sit down, chip export policy becomes a bargaining chip itself. This rule gives Trump leverage, but it also signals how far the US thinks it needs to go to maintain its compute advantage.

The Implication

If you're building AI infrastructure outside the US, Europe, or Japan, this rule just made your compliance budget a line item. Cloud providers in the five named countries will need export licenses for every advanced chip deployment, which means longer lead times and higher costs. Expect those costs to flow downstream to customers.

For Chinese AI companies, the cloud loophole is closing but not closed. Licenses create friction, not walls. The real question is whether China's domestic chip industry can close the gap before the next round of restrictions drops. Watch Huawei's next-gen AI accelerators and SMIC's yield rates. That's where the long game plays out.

Sources

Bloomberg Tech | Business Insider Tech