Latest
Nvidia has warned its customers in recent months that proposed US legislation on Chinese companies’ remote access to its AI chips could expose both the chipmaker and its buyers to US “legal action,” CNBC reported on Tuesday, citing two people with direct knowledge of the matter. The warning concerns the Remote Access Security Act, which would give the US government broader authority to control access to cloud services in the same way as physical hardware; the bill passed the House of Representatives in January but has not passed the Senate. Nvidia also told buyers that, for now, it would not sell its chips to customers who rent more than 50 percent of their GPUs to Chinese companies, the sources said. The Financial Times reported earlier this month that Nvidia had verbally told data-centre clients to direct at least 51 percent of their services to non-Chinese customers under future contracts. CNBC also reported on November 5 that the Chinese lab Moonshot AI sought access to Nvidia chips through data centres in Europe; the Trump administration had accused the company in July of accessing the chips through centres in East Asia.
The Bureau of Industry and Security released the interim final rule on Monday, saying it was written by and at the direction of the Treasury Department and the National Security Council to ensure existing restrictions “cannot be circumvented through remote computing arrangements involving data centers in third countries.” Under the rule, a licence is required when a person knowingly supplies covered US chips or servers to a data centre anywhere in the world that provides more than 12 million aggregate total processing performance (TPP) to companies headquartered in Macau or a Country Group D:5 destination, including China, or whose ultimate parent is headquartered there. Licences are also now required to export, reexport or transfer in-country advanced semiconductors to Indonesia, Malaysia, Singapore, Thailand or Brazil, which BIS deems present a higher-than-normal risk of diversion; the requirement does not apply where the chips are destined for companies headquartered in the United States, or with a US-headquartered parent, provided reporting requirements are met. Data centres, overseas foundries and chip exporters must perform due diligence on customers, including checking they are not owned, directly or indirectly, by persons on the US restricted export list. The rule formally rescinds the Biden-era AI Diffusion Rule, which was never enforced. It takes effect on publication in the Federal Register, with comments open for 60 days and existing data-centre arrangements given 90 days to obtain licences, according to summaries of the rule text.
The Commerce Department moved on Monday to close a loophole that has allowed Chinese companies to lease Nvidia’s computing power from locations in other countries, CNBC reported. The action is the first US policy response to the scandal at Nvidia that led chief executive Jensen Huang to resign last week, and it follows criticism of Commerce Secretary Howard Lutnick, who has been accused of not doing enough to prevent possible diversion of Nvidia’s most advanced chips to China in violation of US export controls. CNBC’s report, published as breaking news, gave no instrument, country list or effective date, and the department had published no rule text that this run could find.
The shape of the measure has been reported over the past week. The Bureau of Industry and Security has been finalising rules that would require companies in seven countries, reported as Brazil, Indonesia, Malaysia, Norway, Singapore, South Africa and Thailand, to seek approval before importing Nvidia’s Blackwell-class processors, with importers required to state where the chips are going and how they will be used, according to The Information. The rule as issued names five countries, not the seven reported last week. US companies operating data centres in those countries could be exempt from licensing but would still have to report which countries their customers are in and how much computing capacity each receives, and no remote access could go to customers in China or Russia.
Existing controls ban direct exports of advanced AI chips to China, and guidance issued in May already required a licence for advanced chips where the parent company is based in China or Macau, regardless of where the purchasing entity is incorporated. Renting computing power from an overseas data centre sat outside that framework, which is the gap the new rules are aimed at. Chinese AI developers have used overseas capacity to train models on hardware they cannot buy, and the reports last week that Washington was considering action on offshore capacity were among the pressures behind today’s slide in Chinese chip names, covered separately in this edition.