The United States plans to regulate the global AI market more strictly. The current US administration under Joe Biden aims to limit the spread of US technology abroad, specifically concerning AI algorithms. They have proposed an “Export Control Framework for Artificial Intelligence Diffusion,” which is currently open for a 120-day comment period and could be enacted under Trump from April. This move is intended to strengthen US security interests, particularly against China.
Instead of maintaining a list of countries with special export rules, the US wants to reverse the approach. Chip designers like AMD, Nvidia, and Intel would only be allowed to sell their GPU accelerators for AI training in 18 specified countries without restrictions. These countries include Australia, Belgium, Denmark, Germany, Finland, France, Ireland, Italy, Japan, Canada, the Netherlands, New Zealand, Norway, South Korea, Spain, Sweden, Taiwan, and the United Kingdom. All other countries will be categorized into Tier-2 and “concerning” regions.
It is important to note that not only the headquarters of a company wishing to purchase accelerators is relevant, but also where the chips will be used. For example, if a German company wants to build a data center in Iceland, it must pass a security review by the US Bureau of Industry and Security (BIS) to obtain the “Universal Verified End User” (UVEU) status. This status allows a company to have up to 7% of its global computing capacity abroad.
Companies in Tier-2 countries can obtain the “National Verified End User” (NVEU) status under the same security conditions and are allowed to purchase up to 320,000 GPUs over the next two years. These limits apply across manufacturers, requiring coordination between AMD, Nvidia, and others.
For companies not operating in any of the 18 partner countries and without NVEU status, the limit is 50,000 GPUs per country. If a company operates data centers in two countries, the limit would be 100,000 GPUs. This limit applies over a two-year period.
Government agreements could potentially double the 50,000 GPU limit. These agreements focus on “export control, clean energy, and technology security.” Orders of up to 1,700 GPUs are exempt from these limits, which suffice only for smaller data centers. It remains uncertain how the US will address fragmented orders, potentially through shell companies.
Besides hardware export restrictions, the US also seeks to limit the spread of large AI models and their underlying training weights. A draft proposes a licensing process if a company wants to “export, reexport, or transfer” models requiring a certain computational power. This could apply to an AI model trained in a partner country and then licensed for a foreign AI assistant.
The key requirement is that the algorithm must have been developed using US technology. This could be argued for any AI model, considering the use of popular US electronic design automation (EDA) tools in chip development.
The proposed initial threshold is 100 quadrillion (1026) computational operations over the entire training period. Modern supercomputers with AI computing power of 200 exaflops (in INT8 data format) would require nearly two months of computing time to reach this level. This restriction targets the largest AI models.
This applies to closed models (“closed-weight”) but not to open-source ones. For closed models, the restriction is lifted if there is an open-source alternative.