The U.S. issued a rule today to regulate Chinese companies renting Nvidia computing capacity overseas. The threshold is low—about 760 H100s. Above that, they’ll need a license from the U.S. 🧐
I went through the original rule posted by the Department of Commerce. It applies no matter which country the data center is in: companies must obtain a license if they provide more than 12 million TPP of computing capacity to companies headquartered in China, Macau, or other restricted countries, or to companies whose parent companies are based in those places.
TPP is the unit the U.S. uses to measure chip computing power for export controls.
Based on the specs for Nvidia’s H100 SXM on its official website, 12 million TPP works out to about 758 cards.
Capacity pooled across several data centers counts, and so does capacity connected over a network. Companies already providing computing capacity to Chinese firms will face export restrictions if they don’t get a license within 90 days.
Also, importing top-tier U.S. AI chips into these five countries—Malaysia, Singapore, Thailand, Indonesia, and Brazil—will require a license. U.S. companies can be exempt if they report the imports.
The Wall Street Journal’s article about Megaspeed yesterday was about this loophole: the chips are located in Malaysia, and Chinese companies rent and use them remotely. The Journal said this was legal under the rules in force at the time.
The Biden administration’s rules were withdrawn last May, leaving a gap for 18 months. Today’s rule fills it. It hasn’t yet been published in the Federal Register, and the effective date field is still blank..
I went through the original rule posted by the Department of Commerce. It applies no matter which country the data center is in: companies must obtain a license if they provide more than 12 million TPP of computing capacity to companies headquartered in China, Macau, or other restricted countries, or to companies whose parent companies are based in those places.
TPP is the unit the U.S. uses to measure chip computing power for export controls.
Based on the specs for Nvidia’s H100 SXM on its official website, 12 million TPP works out to about 758 cards.
Capacity pooled across several data centers counts, and so does capacity connected over a network. Companies already providing computing capacity to Chinese firms will face export restrictions if they don’t get a license within 90 days.
Also, importing top-tier U.S. AI chips into these five countries—Malaysia, Singapore, Thailand, Indonesia, and Brazil—will require a license. U.S. companies can be exempt if they report the imports.
The Wall Street Journal’s article about Megaspeed yesterday was about this loophole: the chips are located in Malaysia, and Chinese companies rent and use them remotely. The Journal said this was legal under the rules in force at the time.
The Biden administration’s rules were withdrawn last May, leaving a gap for 18 months. Today’s rule fills it. It hasn’t yet been published in the Federal Register, and the effective date field is still blank..

