Export-banned Nvidia chips are behind China’s most powerful AI models—uh oh, Jensen!

The leading open-weight models developed by Chinese companies were trained on Nvidia chips that the US government restricted from shipping to its largest geopolitical rival, a new research effort claims.

The apparently widespread access to Blackwell GPUs among Chinese frontier labs shows the magnitude of the problem facing the leading chip designer. Nvidia CEO Jensen Huang left the company this week after admitting to his affair with Alice Huang, executive of Megaspeed, a company implicated in smuggling Nvidia chips into China. Alice Huang has also been alleged to be an agent of China’s intelligence service.

While Jensen Huang’s relationships with a customer and two employees have drawn the most public interest, Nvidia’s potential corporate liability for export control violations likely factored into its decision to part ways with Huang.

US export controls are intended to protect the country’s leadership in AI capabilities against China, but Huang has long sought to open up that market. Blackwell GPUs are banned from export to China or Chinese-owned firms, while the Trump administration allows Nvidia to export the H200, a less powerful chip, provided shipments take a circuitous shipping route that allows the US to claim a 25% tariff.

Will DePue, a former OpenAI researcher who leads the nonprofit research group Unemployment Labs, analyzed the weights—the unique structure—of large language models released openly by DeepSeek, Moonshot AI, MiniMax and Ant Group. The weights still reflect the configurations of the chips used to train them, and DePue was able to find compelling evidence that the most closely watched models, including Moonshot’s Kimi’s K3 releases and DeepSeek’s V4 models, bear fingerprints from training on Blackwell chips.

When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.