The US is closing the route by which Chinese companies have used American AI chips without buying them: renting the computing power from data centres abroad. A rule from the Commerce Department makes any data centre, anywhere, that supplies a large volume of computing to a company headquartered in China, Macau or another country in the US’s Country Group D:5 come under a US licence, and it puts five countries where chips are thought to slip through, Indonesia, Malaysia, Singapore, Thailand and Brazil, under a licence requirement of their own. It lands on the day Nvidia reports its first results since Jensen Huang’s resignation.
Markets in Asia read it as bad news for the chip trade. Taiwan’s main index fell 2.31% on Tuesday, the Taiwan Stock Exchange’s figures show, and its index of electronic parts makers dropped 6.9%.
The text, filed with the Federal Register on Monday and in force from its publication on Thursday, goes further than closing a single loophole. Until now the controls were mostly about who buys the chips. A foreign data centre was caught when a Chinese firm was the purchaser, or its exclusive or main remote user, taking more than half its computing, the rule notes. Now the seller of the chips needs a licence if it knows the data centre will offer that much computing to Chinese firms, and a data centre that already does so has 90 days to obtain one or face “future export restrictions”, the rule says.
The line is drawn by size. The licence is needed for access to more than 12m units of “aggregate TPP”, the government’s measure of processing power, at one site or across sites linked together. The rule says its aim is to stop companies “headquartered in countries of concern” from using American chips “to train advanced SI models”, in the new official term for AI. Recommendation systems of the kind online shops and social media use are left out.
The five named countries get a broader rule. Every export of advanced chips to them now needs a licence, unless the buyer or its parent company is headquartered in the US and the exporter files the reports the rule requires. Commerce also warns governments that it will be “actively monitoring for countries that tolerate evasion”, with import restrictions and a downgrade in their export-control status among the possible penalties.
What makes the rule bite is a gap that Nvidia describes in its own filings. The company reports its sales by where each buyer has its headquarters, and in the quarter to 26 July, customers based outside the US accounted for 38% of its revenue, up from 30% a year before, its latest quarterly report shows. But it adds that “the end customer and shipping location may be different from our direct customers’ headquarters location”, and that “as we do not have physical control of our products after sale, we must also rely on the compliance programs of our customers and partners.” The new rule moves part of that burden onto the data centres themselves, which must now check who owns their customers.
The rule also tidies up. It formally scraps the AI Diffusion Rule of January 2025, the Biden administration’s worldwide licensing scheme, which Commerce says “has never been enforced”. In its place is a narrower, harder line aimed at China rather than at the whole world.
Our view: the rule matters more for where AI data centres get built than for Nvidia’s next quarter. The exemption for American-owned buyers means the largest cloud companies can keep building in the five countries with paperwork rather than permission. The losers are local and Chinese-owned operators in the five countries, whose business case rested on serving clients Washington no longer wants served. Nvidia reports after the US close tonight, its first results without Huang, as we wrote on Sunday; how much of its guidance already assumed this rule is likely to be among the first questions its executives face. Comments on the rule close 60 days after Thursday’s publication, and the industry will use them.