Former OpenAI researcher Will DePue published forensic evidence indicating that leading Chinese AI labs have accessed and used Nvidia’s export-controlled Blackwell GPUs. The findings, verified independently by researchers at OpenAI and Anthropic, analyze published weights from six Chinese models released since April 2026, including models from DeepSeek, Moonshot AI, MiniMax, and Ant Group.

The forensic method examines initial random weight patterns set by PyTorch across GPU compute cores prior to training. While older Hopper architectures use 132 cores, Blackwell chips use 148 or 152 compute cores. DePue matched these exact core distributions to models such as DeepSeek-V4-Flash, DeepSeek-V4-Pro, Kimi K3, and Ant Group’s Ling-3.1-flash, later verifying the results on physical Blackwell hardware.

The report estimates these five training runs required approximately 12,000 Blackwell GPUs combined, representing over half a billion dollars in hardware value. This technical proof confirms long-held suspicions from US officials that export restrictions were being bypassed to power cutting-edge Chinese AI development.

Why it matters

  • Proves US export controls on advanced AI chips are being circumvented, which will likely trigger tighter supply chain regulations and compliance audits.

  • Establishes a novel forensic methodology using model weight artifacts to inspect and verify hardware backends of deployed LLMs.

  • Confirms Chinese labs like DeepSeek and Moonshot are training models on parity computing infrastructure despite international hardware sanctions.

Source: the-decoder.com