The United States plans to restrict Chinese AI developers' access to foreign data centers
- The USA is preparing a law to block remote access to AI accelerators.
- The restrictions will affect data centers in Thailand, Singapore, and Malaysia.
- Nvidia continued shipments of H200 accelerators to China, earning hundreds of millions.
- The Chinese model Kimi K3 was trained in a Thai data center, raising suspicions in the US.
A bill is being prepared in American corridors that will allow limiting access for Chinese artificial intelligence developers to computing resources located outside of China. Initiated by the US Department of Commerce, it concerns data centers equipped with modern accelerators made in the USA but located in Thailand, Singapore, and Malaysia.
The current set of export control rules, effective from the beginning of 2025, formally prohibits the supply of such accelerators to 'unfriendly' countries. However, during Donald Trump's previous presidential term, the agency did not enforce compliance, and major manufacturers, particularly Nvidia, actively lobbied for the easing of restrictions. The company was able to continue shipments of H200 accelerators to China and earned several hundred million US dollars in 2023.
Remote access as a new vulnerability
New restrictions are aimed not only at the physical delivery of equipment but also at the possibility of remote use of computing power. Currently, Chinese AI companies can rent servers in Asian data centers where American accelerators are installed, without directly violating export rules. According to U.S. officials, such practices contribute to the rapid development of domestic models, among which Kimi K3 from Moonshot AI, trained in Thailand, stands out.
The U.S. security service suspects that Kimi K3 uses technologies obtained through the distillation of American AI models, which raises concerns about the leakage of know-how. Therefore, the Department of Commerce plans to discuss with stakeholders next month possible tools to restrict remote access, although legally the agency currently lacks the authority to regulate such connections.
The problem is that current legislation only covers the physical movement of accelerators, not their remote use over the network. Nevertheless, American regulators are considering additional measures, including sanctions against companies providing such services, and tightening control over the supply of cloud services.
If the initiative is implemented, Chinese developers will have to seek alternative ways to obtain computing resources, which may slow the development of their AI models. For the U.S., this is a step towards protecting technological advantages and limiting access to advanced computing solutions, which have become an important element of geopolitical competition in recent years.
At the same time, the restrictions may impact the global cloud services market, as many international providers use American accelerators in their data centers. It is expected that the discussion of the initiative will affect the interests of not only the U.S. and China but also partner countries in Southeast Asia, where key infrastructure facilities are located.
Source: 3DNews
