X Discussion: Ex-OpenAI Researcher Will DePue Publishes a Checkpoint Forensics Report Claiming Six Open-Weight Models from DeepSeek, Moonshot, MiniMax and Ant Group Carry NVIDIA Blackwell Initialization Fingerprints; the Report Says It Does Not Prove Export-Control Violations

Core Background and Industry Pain Points

Blackwell is NVIDIA's most advanced data-center GPU, and exporting it to China requires a U.S. license. U.S. officials and press reports had alleged that some Chinese labs were using Blackwell, but without public technical evidence. On November 12, former OpenAI researcher Will DePue published "Top Chinese Frontier AI Models Carry the Fingerprint of Export-Controlled NVIDIA Blackwell GPUs." His X post drew about 2.05 million views, 20,000 likes and nearly 3,000 reposts, making it one of the most discussed AI posts of the past 12 hours.

Architecture Highlights and Underlying Mechanism

The idea: PyTorch generates initial weights with cuRAND's Philox algorithm and splits the work across threads, and the total thread count depends on the GPU's SM count. Hopper (H100/H200/H800) has 132 SMs, B200/B300 has 148 and GB200/GB300 has 152. The same seed puts its random numbers in different positions on different chips. Some vocabulary rows never touched in training keep nearly their initial values, so the seed and chip type can be recovered. The author says a B300 reproduced DeepSeek-V4-Flash's initial weights bit for bit from the recovered seed.

Authoritative Benchmarks and Measured Comparisons

There are no model benchmarks here, only the author's statistics: across about 6 trillion combinations, chance matches reach about 8 sigma, while the six models' strongest evidence ranges from 49 to 1,385 sigma; the method agreed with all 24 models whose developers disclosed training hardware, including NVIDIA's Nemotron-3 Ultra. These are the author's figures, and the companies named have not publicly confirmed them.

Developer Guide and Getting Started

Keep three points in mind. First, the fingerprint shows which class of chip initialized the weights; the author writes that it supports, without proving, that training also ran on Blackwell. Second, the report cannot establish where the chips were, how they were obtained, or whether any rules were broken. Third, the author says he is not confident in his GPU-count and cost estimates. For developers, these open weights remain available as before; the bigger significance is that tracing training hardware from public weights has been done systematically for the first time and may be used for model provenance work.