On November 12, ex-openAI researcher Will DePue (who’s since started his own research lab, Unemployment Labs) posted forensic evidence on X indicating that China’s biggest AI labs — DeepSeek, Moonshot AI, MiniMax, and Ant Group — have trained open-source models on Nvidia’s Blackwell chips. Advanced chips, like Blackwells, are what allow large AI models to run, and to be trained. Since 2022, the U.S. has placed export controls on Blackwells, among other advanced chips. DePue’s results showed that all of the Chinese labs in question switched to Blackwells sometime in mid 2026.
On October 29, the New York Times reported that Nvidia CEO Jensen Huang had a close relationship with Chinese businesswomen and alleged spy Alice Huang. Last Thursday, after admitting to the affair, Jensen Huang stepped down as CEO. This ordeal renewed scrutiny over Jensen’s stance on China’s access to Nvidia’s chips, given that he’d publicly expressed hope that U.S. export controls would be loosened, or lifted, and was adamant that selling Blackwells to China was not against American interests. (He later qualified this position, telling journalists earlier this year that China should not be able to access America’s “latest and greatest” chips.) This week, Nvidia announced the creation of a committee to oversee compliance with US export controls.
I spoke with DePue to better understand his results, how he found them, and what this means for the US-China AI race. Our conversation has been edited and condensed for clarity and brevity.
Get 14 day free trial