Washington moves to cut off rented computing power for Chinese firms, and puts five more countries under scrutiny
A new US rule requires data centres worldwide to obtain a licence if they can supply AI computing power to companies from China or other embargoed countries.
The US Commerce Department has issued an interim rule meant to stop Chinese companies from getting around export restrictions by remotely renting computing power in data centres outside the US. According to TechCrunch, a licence will now be needed to export or transfer AI chips and servers to data centres anywhere in the world if they can supply computing power to companies from the so-called D:5 group, which includes China, Russia, Iran and North Korea. Access through intermediaries is also covered, and operators must keep an eye on their customers' ownership structures.
The rule applies to data centres or networks with more than 12 million TPP, a measure of computing power that TechCrunch says corresponds to roughly 750 Nvidia H100 or H200 chips. Existing operators get 90 days to obtain a licence; recommendation systems for e-commerce or social media are exempt. In addition, a licence will be required to sell advanced chips to Malaysia, Singapore, Indonesia, Thailand and Brazil, countries the department says carry a higher risk of diversion, unless the chips are destined for US companies. The rule also formally rescinds the Biden administration's AI Diffusion Rule of January 2025.
The Register places the measure in the aftermath of Nvidia chief Jensen Huang's resignation and the affair surrounding Megaspeed, an Nvidia customer that built data centres in Malaysia. Counterpoint analyst Neil Shah told the site that the burden now shifts from chipmakers to cloud operators, who must show that their computing power is not going to embargoed countries. In the short term he expects a bottleneck for Chinese tech firms, and in the long term faster growth of home-grown alternatives such as Huawei's Ascend chips.