What the framework says
The new Framework creates a three-tier licensing hierarchy. Favored countries such as the 10 EU nations will be able to purchase AI chips, including the most powerful, without restriction. Most countries fall into the middle tier, subject to export licensing restrictions on how much computing power they can get hold of. And then there are coutnries that already can’t buy AI chips from the US, including obvious candidates such as China and Russia.
For countries in the middle tier, if an individual order doesn’t exceed a “collective computation power up to roughly 1,700 advanced GPUs,” — the sort of GPU power used by a university or medical institute — no export license will be required. These sales won’t count against national chip quotas.
As for LLMs, sales of the most powerful proprietary models will also be restricted outside of the favored countries. The US Department of Commerce’s Bureau of Industry and Security (BIS), which drafted the framework, defines the restricted models as those built using closed (as opposed to open-source) weights using more than 10^26 computational operations.
