A US administration official confirmed to Reuters that DeepSeek’s upcoming model, which could be launched in the coming weeks, was trained on Nvidia Blackwell chips – the company’s flagship AI processor family currently restricted from export to China. According to the source, the chips were allegedly located in a computing cluster in China’s Inner Mongolia region, despite the existing controls.
The administration emphasized that “we’re not shipping Blackwells to China” – making clear that US policy prohibits delivery of these highest-performance chips to Chinese entities. However, officials did not disclose how DeepSeek may have obtained them or how US authorities became aware of their location and use. Nvidia, the US Department of Commerce, and DeepSeek did not respond to requests for comment.
The export restrictions are part of a broader US strategy to limit Chinese technology firms’ access to the most advanced training and computing tools, seen as critical in the global AI race. Washington previously revoked licenses even for reduced-performance versions of such chips, arguing that top-tier processors should remain out of reach of China’s fast-developing AI ecosystem.
The Chinese embassy in Washington firmly rejected allegations of violations, criticizing the US export policy as “drawing ideological lines” and “overstretching the concept of national security.” A spokesperson for China’s foreign ministry added that Beijing has clearly communicated its position on US export restrictions but does not have detailed information regarding the specific circumstances reported by Reuters.
Officials note that the DeepSeek case presents challenges for current export-control mechanisms, especially if chips reach China through indirect channels or gray-market pathways. Some experts and policymakers warn that such incidents could undermine US export controls designed to slow the rapid progress of Chinese AI firms and their potential civilian and military applications.
DeepSeek, founded in 2023 in Hangzhou, quickly gained global attention for high-performance large language models comparable to leading Silicon Valley systems and reportedly more cost-efficient to train. Its models, including R1 and subsequent V3 iterations, have been widely used in consumer applications and digital services, with adoption extending beyond China.
Reports about the alleged use of Blackwell chips by DeepSeek also emerge amid broader tensions over technological competition – including claims that some Chinese AI firms have used “distillation” techniques on models such as Claude to accelerate development of their own systems without full access to training resources.