American regulators have zeroed in on a scheme that allows Chinese artificial intelligence developers to use advanced Nvidia chips without formally violating export bans. This is not about smuggling hardware, but about renting computing power in data centers outside the PRC—in Malaysia, Thailand, and other countries in the region. Physically, the servers remain on foreign territory, but Chinese engineers gain remote access to them to train their models.
The Mechanics of the Loophole
Since 2022, Washington has been consistently tightening rules on the supply of the most powerful AI accelerators to China. However, current restrictions primarily target the export of physical equipment. Remote use of chips via cloud servers has long remained outside the scope of control. Chinese companies have exploited this.
Based on my data, even giants like Alibaba and ByteDance have been drawn into the orbit of such schemes. They rent Nvidia GPU capacity in Southeast Asia, where AI infrastructure is growing at explosive rates. The U.S. Bureau of Industry and Security is now systematically cataloging this problem, compiling lists of transit countries and states where Chinese entities gain remote access to banned chips. Notably, the second scheme in itself is not always a violation of existing export rules.
A telling example is Alibaba, which operates through the Singapore-based entity Megaspeed. The ultimate beneficiary of this chain is a Cayman Islands company owned by Alibaba. Such a multi-layered jurisdictional structure is designed to obscure the trail and complicate the legal classification of the transaction.
The Kimi K3 Case and Nvidia's Position
The issue gained public prominence after the release of the open-source model Kimi K3 by Moonshot AI with 2.8 trillion parameters. In July, Michael Kratios, director of the White House Office of Science and Technology Policy, stated that the startup rented servers with Nvidia GB300 in Thailand to train this model. However, he did not provide convincing technical evidence. The Chinese side rejected the accusations, calling them slander.
Notably, a similar story previously emerged around DeepSeek, which was suspected of using shell companies to access data centers. Nvidia then stated that DeepSeek had legally purchased H800 chips.
Legislative Response and the Economics of the Issue
The U.S. Congress is trying to close this loophole through the "Remote Access Security Act." The document, approved by the House of Representatives in January, extends export controls to the use of technology via the internet and cloud services. Restrictions could now apply not only to the supply of a chip but also to the very possibility of leveraging its power from China. The bill still needs to pass the Senate.
Nvidia opposes such an expansion. The company reasonably points out that current rules deliberately allow the creation of cloud infrastructure outside controlled jurisdictions. The manufacturer has also calculated the losses: bans have already cost it approximately $5.5 billion due to H20 licensing, and the subsequent permission to supply H200 to select clients does not compensate for the loss of the second-largest market.
My analysis: This situation is a classic example of an "arms race" in regulation. Each new restriction spawns more sophisticated circumvention schemes, and in the end, American companies themselves suffer, losing market share to competitors from other jurisdictions. While the U.S. fights remote access, China will continue to build its own computing capacity, which in the long term could render export restrictions obsolete.