Politics
Report: Chinese AI Models Trained on Export-Banned Nvidia Blackwell Chips
A new forensic analysis of model weights suggests DeepSeek, Moonshot AI, MiniMax, and Ant Group trained their latest large language models on Nvidia Blackwell GPUs that U.S. export controls were meant to keep out of China, deepening scrutiny of Nvidia's compliance and its former CEO's ties to a Chinese intelligence operative.
SAN FRANCISCO â Four of China's most prominent artificial intelligence developers appear to have trained their latest large language models on Nvidia Blackwell graphics processing units, the advanced chips that the Trump administration restricted from export to China, according to a new forensic report by former OpenAI researchers.
The analysis, published by Will DePue's research group, found structural fingerprints in the published weights of models released by DeepSeek, Moonshot AI, MiniMax, and Ant Group that match the hardware configurations and initialization patterns associated with Nvidia's Blackwell GPUs. The report says the evidence comes entirely from the models' own public weights rather than from leaks or insider accounts.
"We recover traces of the models' initial values, and then search for the hardware configurations and initial configurations that reproduce them," the report states. "The fingerprint identifies how the weights were created and supports, without proving, the inference that Blackwell also performed the training."
The researchers also found that the Chinese companies began training on Blackwell chips only within the last six months, between April and October 2026. Earlier versions of the same models showed hallmarks of training on Nvidia's less powerful H200 or Hopper chips, which the company was still permitted to sell to China under existing rules.
The findings add a new dimension to the controversy surrounding former Nvidia CEO Jensen Huang, who resigned this week amid revelations about his relationship with Alice Huang, a Chinese intelligence operative. Alice Huang worked for Megaspeed, a company now suspected of facilitating illegal access to the restricted Blackwell chips.
Nvidia's potential corporate liability for export control violations likely factored into the board's decision to part ways with Huang, according to reporting by TechCrunch. The company has not commented publicly on the forensic report.
Separately, Bloomberg reported Friday that Aperia Cloud Services, a Singapore-based cloud partner of Nvidia with links to Alice Huang, purchased 22 servers from Dell containing Nvidia H800 chips, which were also banned from sale to China. The servers were reportedly destined for a data center in Shanghai. Dell cut business ties with Aperia in February 2025 over legal concerns about the deal.
Jensen Huang is scheduled to testify before the House Foreign Affairs Committee before Thanksgiving, according to Engadget. The hearing is expected to be closed-door, but lawmakers from both parties are certain to press him on the chip diversion, his relationship with Alice Huang, and what Nvidia knew about its partners' activities.
The episode raises fresh questions about the effectiveness of U.S. export controls aimed at preserving America's lead in AI computing, and about whether Nvidia and its intermediaries adequately policed the distribution of restricted hardware. For Bay Area readers, the stakes are both economic and strategic: Nvidia is one of the region's most valuable companies, and the integrity of its export compliance is now under congressional scrutiny.
The report's authors caution that their method supports but does not conclusively prove that Blackwell chips were used. Still, the pattern across four separate companies â all releasing models within months of one another â suggests a coordinated or at least widely shared access point to the restricted hardware.
Neither Nvidia nor the Chinese companies named in the report responded to requests for comment by publication time.