For two years, the story of Nvidia in China was a story about Washington: what the Bureau of Industry and Security would allow, and what it would not. That story just flipped. The binding constraint on China's access to advanced AI chips is now increasingly coming from Beijing itself, and the workaround Chinese labs are using to route around both governments runs through data centers in Malaysia, Thailand and Singapore, not the mainland.

The Quota Nobody Can Use

In January 2026, BIS shifted its review of Nvidia H200 and AMD MI325X export licenses to China from a presumption of denial to case-by-case approval, provided the chips sit under a 21,000 TPP and 6,500 GB/s memory-bandwidth threshold 1. Nvidia began shipping H200s into China in July under that framework, and by mid-August ByteDance and Tencent had each taken delivery of roughly 10,000 units 3. That sounds substantial until set against the licenses themselves: ByteDance and Tencent are each cleared to buy up to 75,000-100,000 H200s, meaning the chips that have actually arrived represent only about 13 percent of what Washington has approved 4.

Beijing's Own Bottleneck

The gap is not a Washington story anymore. China's National Development and Reform Commission has directed ByteDance and Tencent to keep the bulk of their US-licensed allotments in Hong Kong rather than bring them onto the mainland, exploiting the fact that Hong Kong sits outside the mainland customs border while still allowing engineering teams to reach the hardware over cross-border networks 3. The catch is that Hong Kong's entire installed data-center base runs to roughly 581 megawatts, and fully exercising the licensed allowance would require power equivalent to more than a fifth of that entire estate 7. Beijing has, in effect, approved more chips than its own preferred storage location can plug in. Hong Kong's data-center operators and power utilities are now the practical ceiling on how much of Nvidia's China business actually materializes, not the State Department.

The Cloud Loophole Doing the Real Work

While licensed hardware sits underused in Hong Kong, a separate and faster channel has emerged: renting, rather than buying. US export rules control who owns a restricted chip, not who remotely accesses its compute over a network, so a Chinese company barred from importing Nvidia's Blackwell-generation GB300 can still lease time on a GB300 cluster sitting in a third country 2. ByteDance has worked with Singapore-headquartered cloud provider Aolani to reach compute hosted in Malaysia, and the White House's own science and technology office has said Moonshot AI accessed GB300 servers in Thailand to help train its 2.8-trillion-parameter Kimi K3 model 5. Bloomberg has reported Kimi runs in part on a 20,000-GPU cluster provisioned through Alibaba, which holds a 36 percent stake in Moonshot 6. None of this requires a single chip to cross into Chinese customs territory, and none of it currently requires an export license at all.

Washington's Fix Is Stalled, and the Clock Is Public

Congress already has an answer drafted: the Remote Access Security Act would extend BIS authority to cover remote access to controlled compute, closing the gap that lets a foreign company rent rather than own its way around the rules. The House passed it 369-22 back in January, but it has sat without a Senate vote since 2. The bill's timing has become newly political. A US-China AI dialogue is scheduled for September 24, the first formal AI talks between the two governments since President Trump's second term began, and a fresh investigative report on a blacklisted Chinese server maker's American subsidiary exporting Nvidia's most advanced chips through an Entity List loophole has added friction to that agenda days beforehand 2. If the Senate moves RASA before or shortly after the summit, cloud providers in Malaysia, Thailand and Singapore would suddenly carry the same compliance burden — customer verification, end-use checks — that chip exporters already carry.

The Investment Read

Three things are worth tracking separately, because they cut against each other. First, Nvidia's realized China revenue will likely keep understating the chips it has technically been cleared to sell, since Hong Kong's power ceiling — not demand — is now the limiting factor; investors modeling a China re-acceleration off the January licensing shift should treat the 13-percent delivery rate as the near-term run-rate, not a transitional blip. Second, Southeast Asian data-center capacity tied to Chinese AI workloads — the Malaysia and Thailand facilities named in US findings — carries the most direct exposure to RASA: a Senate vote would compress a genuine growth channel for regional colocation operators almost overnight. Third, Hong Kong's power and real-estate buildout, previously a story about generic AI-infrastructure demand, now has a specific and dated catalyst in the mainland's own licensing overhang, which should support a case for accelerated grid and data-center investment in the territory regardless of how the Washington-Beijing dialogue on September 24 resolves. Whichever government moves first to close its own gap will decide who ends up holding the capacity.