US officials are probing whether Chinese AI start-up DeepSeek bought advanced Nvidia semiconductors through third parties in Singapore, circumventing US restrictions on sales of chips used for artificial intelligence tasks, people familiar with the matter said.
DeepSeek recently released a chatbot, called R1, that in some respects performs as well as comparable tools from the US, suggesting that China is further ahead in the AI race than previously believed. Some prominent engineers have marveled at R1’s capabilities, and DeepSeek has touted the tool’s low cost and efficiency, prompting rivals to speculate whether it was built on the back of Western technology.
Officials in the White House and Federal Bureau of Investigation are also trying to determine whether DeepSeek used intermediaries in the Southeast Asian nation to purchase Nvidia chips that the US has banned from sale to China, said the people, who requested anonymity to relay private conversations.
DeepSeek didn’t immediately respond to a request for comment. An Nvidia spokesperson said in a statement that “we insist that our partners comply with all applicable laws, and if we receive any information to the contrary, act accordingly.” The company issued a statement on DeepSeek earlier this week that indicated it believes the Chinese company didn’t violate US restrictions.
An FBI spokesperson declined to comment, while a representative for the White House didn’t immediately respond to a request for comment. The Singapore embassy in Washington couldn’t be reached outside of office hours.
Howard Lutnick, US President Donald Trump’s pick to lead the Commerce Department, suggested on Wednesday that DeepSeek evaded US export controls.
See also: Apple forecast cheers investors after mixed holiday results
“Nvidia’s chips, which they bought tons of, and they found their ways around it, drive their DeepSeek model,” Lutnick, who would enforce semiconductor trade restrictions as the Commerce head, told senators in his confirmation hearing Wednesday. “It’s got to end. If they are going to compete with us, let them compete, but stop using our tools to compete with us. So I’m going to be very strong on that.”
R1, released earlier this month by Hangzhou-based DeepSeek, can mimic the way humans reason, posing a potential threat to the market dominated by OpenAI and US rivals and leading to a US$1 trillion market wipeout.
The breakthrough triggered a debate about the effectiveness of US attempts to cut China off from advanced technology. The restrictions, covering a variety of chips and the tools used to make them, are aimed at slowing China’s efforts to develop its semiconductor industry and preventing Beijing from accessing AI that could give the country a military edge.
See also: Alibaba touts new AI model superior to DeepSeek’s and Meta’s
While DeepSeek hasn’t fully identified which AI semiconductors it used to develop its models, its researchers said in a paper that the V3 model, released last month, was trained on 2,048 of Nvidia’s H800 chips. Nvidia created the H800 for the Chinese market after the Biden administration cut off the country’s access to a range of more powerful versions.
US officials then banned H800 and other Nvidia chips for China in October 2023, prompting Nvidia to design yet another, less-capable chip called the H20 for that market. Trump administration officials are now in early conversations about whether to restrict the H20 chips as well, Bloomberg has reported, continuing a policy debate that began at the end of President Joe Biden’s term.
At the same time as the US has expanded the types of chips that cannot be sold to China, it’s also increased the geographic scope of the trade rules to cover more countries.
In 2023, the Biden administration imposed restrictions on more than 40 countries that officials worried could serve as intermediaries to get chips to China, including most of the Middle East and some of Southeast Asia — but not Singapore.
Earlier this year, Biden officials expanded the rules even further to capture most of the world, with the exception of a handful of US allies. Large shipments to Singapore, as is the case for most countries, now require a license. Shipments of less than around 1,700 of the processors only require notification.
Singapore accounts for about 20% of Nvidia’s revenue, according to regulatory filings. But according to those documents, “most shipments associated with Singapore revenue were to locations other than Singapore, and shipments to Singapore were insignificant.”
The Nvidia spokesperson said that the revenue associated with Singapore doesn’t indicate diversion to China. “Our public filings report ‘bill to’ not ‘ship to’ locations of our customers,” he said. “Many of our customers have business entities in Singapore and use those entities for products destined for the US and the West.”
The top Democrat and Republican on a China-focused panel in the US House of Representatives cited Nvidia’s Singapore revenue in a letter Wednesday to National Security Advisor Mike Waltz.
“Countries like Singapore should be subject to strict licensing requirements absent a willingness to crack down” on shipments to China, Representatives John Moolenaar and Raja Krishnamoorthi wrote.
