US requires licence for data centres selling large-scale AI computing to Chinese companies
The United States requires data centres anywhere in the world to obtain an export licence if they can sell large amounts of AI computing to companies from China, Russia, Iran or other restricted countries. Chip exports to Malaysia, Singapore, Indonesia, Thailand and Brazil will also need a licence.
The Bureau of Industry and Security (BIS), the US Commerce Department’s export control agency, published an interim rule on Monday 16 November that extends export restrictions on AI chips to remote access. Until now, Chinese companies have been able to rent Nvidia’s most advanced chips legally from data centres in South-East Asia and other countries, even though they are not allowed to buy the chips directly, The New York Times reported.
How the rule works
A licence is required to export chips or servers to any data centre that can sell computing to companies in country group D:5, TechCrunch reported. The group includes China, Russia, Iran, North Korea and 18 other countries under a US arms embargo.
The licence threshold is 12 million TPP. TPP (total processing performance) is the metric US export controls use for chips’ computing power. TechCrunch wrote that the threshold roughly equals the maximum computing power of 750 Nvidia H100 or H200 chips. AI Times reported that the capacity of interconnected data centres also counts towards the total.
The rule also covers computing obtained through intermediaries. Exporters and data centres must establish their customers’ ownership and activities. The rule does not apply to computing used only for recommendation systems in online shops and social media.
The department justifies the new licence requirement for five countries on the grounds that chips passing through them are more likely than usual to end up at unauthorised destinations. TechCrunch and AP say the requirement covers exports to these countries; the NYT and Ars Technica say it covers companies headquartered there. The requirement does not apply to chips bound for US companies, provided certain notification conditions are met, AI Times reported. The rule also restores a licence requirement when a Chinese company wants advanced chips made by, for example, TSMC or Samsung, the NYT reported.
The sources differ on when the rule takes effect. AP said it took effect immediately, the NYT said in 90 days, and Ars Technica said 90 days after its publication in the Federal Register, the US government’s official journal. TechCrunch and AI Times reported that data centres already in operation have 90 days after the rule takes effect to obtain a licence.
Enforcement stays with the Commerce Department
Remote access came into focus because of Megaspeed, a Singapore-based data centre company. US authorities are investigating the company over suspected smuggling of Nvidia chips to China. The NYT reported that the company was founded by Alice Huang, whom authorities regard as an agent of Chinese intelligence. Jensen Huang, who had been in a relationship with Alice Huang, resigned as Nvidia’s chief executive last week. China denies spying, and Megaspeed denies smuggling.
The Treasury Department and the National Security Council wrote the rule and also directed the work, the department’s press release said. Commerce Secretary Howard Lutnick had pushed for a lighter version, the NYT reported. Analysts quoted by the NYT said enforcement nevertheless remains with the Commerce Department, whose means of enforcing the rule are still unclear.
Chris McGuire, a researcher at the Council on Foreign Relations think tank and a former Biden administration official, told the NYT that, if enforced, the rule would upend the business of South-East Asian data centre companies, known as neoclouds. ‘This effectively kills the business of neoclouds like Megaspeed that exist to provide remote access to large amounts of A.I. compute to China,’ McGuire said.