RealityHackerOpen in RealityHacker ⇢
Science · AI in Drug Discovery · published 2026-11-19T00:00:00+00:00 · via Scientific American

Chinese AI Models Show Signs of Training on Banned Nvidia Blackwell Chips

Image via Scientific American
Image via Scientific American

Independent researcher Will DePue analyzed the initial weights of several top Chinese AI models and found evidence suggesting they were trained on Nvidia's Blackwell chips, which are restricted for export to China. The study indicates that specific hardware signatures from these chips can be detected in the models' data, even after training. The findings point to potential circumvention of U.S. export controls by companies including DeepSeek, Moonshot AI, MiniMax, and Ant Group.

Expanded Detail

In November 2026, independent researcher Will DePue published findings suggesting that several leading Chinese AI models—including those from DeepSeek, Moonshot AI (Kimi), MiniMax, and Ant Group—were likely trained on Nvidia’s Blackwell chips, which are banned for export to China under U.S. export controls. DePue’s analysis focused on the initial weights of these models, which are set during training using a random seed processed across GPU processors. Because Blackwell chips have 148 or 152 streaming multiprocessors—more than the 132 in Hopper chips—the distribution of random numbers during initialization differs between chip generations. By identifying patterns in weights that changed minimally during training, DePue matched them to the unique hardware signatures of Blackwell chips, even though the models were trained after the fact. This forensic method implies that the models’ foundational data was shaped by restricted hardware, despite the U.S. ban. The findings follow earlier reports, including a New York Times article from late October 2026, which revealed U.S. intelligence suspicions about a Singapore-based executive with ties to Nvidia’s former CEO Jensen Huang, who resigned amid the scandal. While DePue does not claim definitive proof of illegal access, his work suggests that Chinese AI labs may have circumvented export restrictions through third-party channels or offshore infrastructure. The exact mechanisms of chip access—whether through intermediaries, cloud rentals, or illicit supply chains—remain unknown. It is also unclear whether these models were trained entirely on Blackwell hardware or if hybrid setups were used. The implications for U.S. national security policy and global semiconductor governance are significant, but no formal investigation or legal action has been announced.

Context

The discovery could intensify scrutiny on global AI supply chains and the effectiveness of export controls in the age of decentralized computing. If confirmed, DePue’s findings suggest that even tightly regulated hardware can leave detectable traces in AI models, potentially enabling new forms of digital forensics to track the provenance of AI systems. This may prompt governments to develop more sophisticated monitoring tools or revise compliance frameworks for AI training infrastructure. The case may also influence how tech firms manage their global operations, especially those with cross-border data flows or cloud services. A comparable precedent is the 2018 U.S. sanctions on ZTE, where export restrictions were circumvented through third-party suppliers, leading to broader regulatory reforms. In this case, the stakes are higher: AI models trained on restricted hardware could be used in sensitive domains like defense, surveillance, or biotechnology, raising concerns about technological parity and strategic risk. While the exact impact on China’s AI ambitions remains uncertain, the findings may accelerate efforts by the U.S. and allies to tighten oversight of AI infrastructure. What’s next could include more public research into model forensics, increased pressure on cloud providers to audit hardware usage, or new international agreements on AI development standards. Whether this leads to policy shifts or remains a technical curiosity depends on how governments and institutions respond to the growing challenge of tracing AI’s digital origins.

Expanded detail and Context are AI-generated analysis; the linked article remains the authoritative source.
Read the full article at Scientific American →
Related stories
Older adults today show superior memory compared to previous generations · Medical Diagnostics
This summary is Al-enhanced to contain extended analysis and broader social context. The original is {NAME); the linked article is the authoritative source. Original headline: “Chinese AI models reveal clues pointing to restricted Blackwell chips.” Browse more stories.