Research Finds Export-Banned Nvidia Chips Used to Train Leading Chinese AI Models
Analysis of open-weight models shows that several advanced Chinese AI systems were trained on Nvidia Blackwell GPUs that are prohibited for export to China.
A new study indicates that leading open-weight AI models developed by Chinese firms were trained on Nvidia Blackwell GPUs, which are subject to U.S. export restrictions. The research, led by former OpenAI researcher Will DePue of Unemployment Labs, examined the weight structures of models released by DeepSeek, Moonshot AI, MiniMax and Ant Group and identified patterns consistent with training on the banned chips.
White House technology adviser Michael Kratsios previously said that Moonshot AI used export-banned Nvidia chips, citing U.S. intelligence sources. Nonprofit Epoch AI estimates that roughly two million advanced Nvidia chips entered China in 2024-2025 through schemes designed to conceal their final destination, while about 275,000 Nvidia AI accelerators were imported to Malaysia during the same period.
The findings do not prove that the models were trained exclusively on Nvidia hardware or that export-control laws were violated, as some chips may have been used in offshore data centers. Nonetheless, the export bans aim to prevent Chinese labs from accessing the most powerful AI processors.
"If we care about the U.S. tech edge, it doesn’t make sense to block chips in mainland China while allowing the same access via Malaysia," said Sam Bresnick of the Georgetown Center for Security and Emerging Technology.
Investigations have linked Megaspeed, a company led by Alice Huang, to the smuggling of banned Nvidia chips to a data center in Shanghai. Bloomberg reported that a firm associated with Alice Huang was involved in the illicit shipments. Short-selling firm Culper previously warned that Nvidia faces a "China problem" due to significant revenue from Chinese demand, whether through remote data centers or illicit channels.
While some analysts suggest that withholding chips could spur domestic Chinese chip production, others, such as Stanford researcher Graham Webster, argue that the scale of demand makes that outcome unlikely. The reasons for the current hold-up on chip shipments to China remain unclear, with speculation ranging from preparation for a future trade deal to concerns about surveillance capabilities.
U.S. companies currently have limited obligations to monitor end-use of controlled AI chips. However, Nvidia’s leadership faces heightened scrutiny after Jensen Huang’s recent departure from the company amid allegations of personal and professional ties to Alice Huang and potential involvement in chip smuggling.