American regulators are seriously concerned about a new practice that allows Chinese artificial intelligence developers to circumvent export sanctions on advanced Nvidia GPUs. Instead of directly shipping processors to China, companies rent computing power in foreign data centers, keeping the physical equipment outside U.S. jurisdiction.
The essence of the scheme is simple and elegant: servers with Nvidia chips are placed in Malaysia, Thailand, or other Southeast Asian countries, while Chinese engineers gain remote access to them to train their models. This allows them to bypass the formal restrictions that Washington has been progressively tightening since 2022.
How the loophole works and who has used it
The key feature of current rules is that they regulate the physical export of equipment. Remote access to chips located outside China has long remained outside the scope of control. Industry giants are actively exploiting this. Based on my data, Alibaba and ByteDance are already renting servers with Nvidia in Southeast Asia, obtaining the necessary computing resources without violating the letter of the law.
The U.S. Bureau of Industry and Security is now trying to systematize this problem. The agency is compiling lists of countries through which banned chips could physically reach China, as well as states where Chinese companies gain remote access to them. The case of Alibaba is telling: the corporation works with Singapore-based Megaspeed, which provides access to GPUs in Malaysia. At the same time, the chain is deliberately complicated—through a Cayman Islands structure, whose ultimate beneficiary is Alibaba itself.
Kimi K3 and escalating tensions
The discussion sharply intensified after the release of the open-source model Kimi K3 from Moonshot AI. In July, White House Office of Science and Technology Policy Director Michael Kratsios publicly stated that the startup had rented servers with Nvidia GB300 in Thailand to train its algorithms. Although he provided no technical evidence, the very fact of such a statement at the highest level speaks volumes. The situation is further fueled by accusations of distilling Anthropic models, which the Chinese side categorically denies.
Notably, Kimi K3 with 2.8 trillion parameters is catching up with leading American systems in a number of tests. To optimize GPU cores, developers used the Nvidia H200—a chip that formally should not have been in their possession. A similar story previously emerged around DeepSeek, which was suspected of using shell companies to gain access to data centers in the region. At the time, Nvidia insisted that all deliveries were legal.
Legislative response and Nvidia's position
The U.S. Congress is trying to close this loophole through the "Remote Access Security Act." The House of Representatives approved the document back in January, and it extends export controls to cloud computing. If the bill passes the Senate, restrictions will apply not only to physical deliveries but also to the ability to use computing power from China.
Nvidia, naturally, is resisting further tightening. The company reasonably points out that current rules themselves allow the creation of cloud infrastructure abroad, and partners undergo vetting. The manufacturer has already calculated losses from previous bans: licensing the H20 cost $5.5 billion, and the Chinese market has been effectively lost. Although in January 2026 Washington softened its stance by allowing H200 deliveries to select clients, the overall trend remains tough.
My analysis: We are witnessing a classic arms race in the AI sector. Restrictions stimulate the development of illegal and semi-legal schemes rather than stopping Chinese developers. In fact, U.S. attempts to isolate China from advanced chips are merely accelerating the creation of parallel infrastructure in Southeast Asia, which is becoming a new testing ground for technological confrontation. In the long term, this could lead to fragmentation of the global AI market and a reduction in the effectiveness of American sanctions.