Research Says Chinese AI Models Show Signs of Training on Export-Banned Nvidia Chips
A new analysis of openly released AI models from Chinese companies found evidence that some were trained using Nvidia’s Blackwell processors, which U.S. export controls prohibit from being shipped to China.
Will DePue, a former OpenAI researcher who leads the nonprofit Unemployment Labs, examined the model weights of large language models released openly by DeepSeek, Moonshot AI, MiniMax, and Ant Group. He identified hardware-related patterns in releases including Moonshot’s Kimi K3 releases and DeepSeek’s V4 models.
The findings do not establish where the training took place, which chips were used for every stage of development, or whether any export-control rules were broken. The models could have been trained in part on Nvidia hardware located outside mainland China, including at offshore data centers. Chinese companies have used such facilities to obtain computing capacity, with Malaysia emerging as one destination.


