American regulators are seriously concerned about a new scheme that allows Chinese companies in the field of artificial intelligence to bypass export restrictions on advanced Nvidia chips. Instead of physically importing processors into the PRC, developers rent computing power in data centers abroad while maintaining full control over training their models.
The equipment is physically located, for example, in Malaysia or Thailand, and Chinese engineers connect to it remotely. This allows them to use the most powerful GPUs for AI training without violating the letter of the law, although undermining its spirit.
How the "gray" scheme works
Since 2022, the United States has consistently tightened export rules for the most powerful AI chips to China. However, the restrictions primarily concern the physical supply of equipment, leaving aside the issue of remote access to it. Chinese giants are actively exploiting this loophole. According to my data, Alibaba and ByteDance have long been renting servers with Nvidia in Southeast Asia, obtaining the necessary computing resources.
The U.S. Bureau of Industry and Security is now trying to systematize this problem. The agency maintains lists of countries through which prohibited chips can physically enter the PRC, as well as states where Chinese companies gain remote access to them. The example of Alibaba's cooperation with Singapore's Megaspeed is telling. The scheme is deliberately convoluted: the Chinese firm gains access to chips in Malaysia through a chain of legal entities, including structures in the Cayman Islands, whose ultimate beneficiary is Alibaba.
Kimi K3 as a catalyst for discussion
This topic gained particular resonance after the release of the open model Kimi K3 from Moonshot AI. In July, Michael Kratios, director of the White House Office of Science and Technology Policy, stated that the startup allegedly rented servers with Nvidia GB300 in Thailand to train its models. True, he never provided public technical evidence. Meanwhile, Moonshot AI has already accused the U.S. administration of defamation, and the model itself, with 2.8 trillion parameters, showed results comparable to the best American counterparts.
A similar story previously arose around DeepSeek, which Washington suspected of using shell companies to access chips. Nvidia then stated that the Chinese startup used legally purchased H800.
Attempt to close the loophole
The U.S. Congress has already proposed a solution — the "Remote Access Security Act." The document, approved by the House of Representatives in January, expands export controls to cover the use of technology via the internet and cloud services. If the law is passed by the Senate, restrictions will apply not only to the physical supply of chips but also to the very possibility of utilizing their computing power from China.
Nvidia, naturally, opposes this. The company emphasizes that current rules specifically allow the creation of cloud infrastructure outside controlled jurisdictions, and its partners undergo strict vetting. The manufacturer has also warned that export bans have already deprived it of its second-largest commercial market, and the costs associated with restriction-related measures (e.g., H20 licensing) are estimated at $5.5 billion.
My view: This situation clearly demonstrates the limits of unilateral export restrictions in a globalized economy. As long as demand exists, schemes to satisfy it will also exist. Washington is trying to fight not the technology itself but its spread, which in the long term looks like a losing strategy that stimulates the development of China's own chips and accelerates the fragmentation of the technological landscape.