DeepSeek Model Reportedly Trained on Nvidia Blackwell Chips Despite US Export Restrictions

Chinese AI startup DeepSeek has reportedly trained its latest model on Nvidia's advanced Blackwell chips despite US export restrictions barring such shipments to China, according to a senior Trump administration official cited by Reuters. The model could launch as soon as next week, raising questions about enforcement of US controls. DeepSeek, based in Hangzhou, used Nvidia's Blackwell processors, the company's most advanced AI chips, for training its new model. US policy explicitly prohibits Blackwell shipments to China due to national security concerns. The chips are believed to be located at DeepSeek's data center in Inner Mongolia, though the method of acquisition remains unclear. US intelligence indicates DeepSeek plans to remove technical markers that could reveal the use of American chips.

The revelation comes amid ongoing debates in Washington. China hawks argue that advanced chips risk bolstering Beijing's military AI capabilities, with former National Security Council official Chris McGuire stating, 'This shows why exporting any AI chips to China is so dangerous.' Others, including Nvidia CEO Jensen Huang, suggest limited sales could keep Chinese firms dependent on US technology. In December, the Trump administration approved sales of Nvidia's H200 chips to China but with restrictions, following an earlier reversal on scaled-down Blackwell variants. Major US AI firms including Google, OpenAI and Anthropic are preparing for DeepSeek's release, amid rumors of chip smuggling since late last year. The use of Blackwell chips by DeepSeek exposes vulnerabilities in US export control enforcement, particularly through covert supply chains that bypass official channels. This incident validates concerns from security experts who note China's reliance on smuggled hardware reveals gaps in domestic chip production but also accelerates AI progress via foreign tech.