Quantcast
Business & Tech

Report: Leading Chinese AI Models Were Trained on Nvidia Chips China Wasn't Supposed to Have

Report: Leading Chinese AI Models Were Trained on Nvidia Chips China Wasn't Supposed to Have
PhotoJensen Huang gives a keynote address at the NVIDIA Berlin GPU Technology Conference (GTC) on October 21, 2026 in Berlin, Germany. (Photo by Sean Gallup/Getty Images)

The Jensen Huang/Nvidia scandal may go much deeper than just an affair with a Chinese spy and an affair with a subordinate or two, as a new research report finds that China's biggest AI models appear to have been trained on Nvidia chips that were barred from export.

The plot thickens with the Jensen Huang honeypot scandal fallout, and now it's not just mere speculation that the CEO's affair with Chinese intelligence operative Alice Huang could have led to the stealing of critical AI chip components.

As TechCrunch reports, via a new report by former OpenAI researchers Will DePue, the hallmarks of Nvidia's Blackwell Graphics Processing Units (GPUs) can be found in the unique structures of large language models that have been released openly by Chinese firms DeepSeek, Moonshot AI, MiniMax, and Ant Group. These GPUs were among the pieces of hardware that the Trump administration had placed strict export controls on, in order to keep them out of China's hands.

"The evidence comes entirely from the models’ own published weights, rather than from leaks or insider accounts," the report says, with "weights" referring to these internal AI structures. "We recover traces of the models’ initial values, and then search for the hardware configurations and initial configurations that reproduce them... The fingerprint identifies how the weights were created and supports, without proving, the inference that Blackwell also performed the training."

The report further finds that the Chinese companies only began training their models on Nvidia's GPUs in the last six months — between April and October 2026 — and earlier versions of the models had hallmarks of training from Nvidia's less powerful H200 or Hopper chips, which the company was still allowed to sell to China.

While not proving anything entirely, all this points to the possibility that Alice Huang and the company she worked for, Megaspeed, succeeded in gaining illegal access to those Blackwell chips.

As TechCrunch notes, "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 [the board's] decision to part ways with Huang."

Bloomberg reported separately on Friday that another company with links to Alice Huang was involved with smuggling Nvidia chips to a data center in Shanghai. The company, a Singapore-based cloud partner of Nvidia called Aperia Cloud Services, reportedly purchased 22 servers from Dell that contained Nvidia H800 chips, which were banned from sale to China — and Dell reportedly cut business ties with Aperia in February 2025 over legal concerns around this deal.

Jensen Huang is now set to testify before the House Foreign Affairs Committee sometime before Thanksgiving, as Engadget reported Thursday, and while it's likely to be a closed-door hearing, Huang is guaranteed to face questions about all of the above from both Republicans and Democrats.

Previously: Jensen Huang Resigns, Nvidia Shares Tumble

Join the conversation

Comments

Comments load automatically as this section approaches.

—
Report: Leading Chinese AI Models Were Trained on Nvidia Chips China Wasn't Supposed to Have
Share this

Start your morning with SFist.