Former OpenAI researcher Will DePue publishes forensic analysis claiming weight-initialization patterns show six Chinese AI models were trained on Nvidia Blackwell GPUs

Will DePue, a former OpenAI researcher now at Unemployment Labs, published a method that infers which GPU generation trained a model from the statistical residue of PyTorch's random weight initialisation. He says Hopper chips leave a 132-core pattern and Blackwell chips a 148 or 152-core pattern. On that basis he links six Chinese models to Blackwell: DeepSeek-V4-Flash, DeepSeek-V4-Pro, MiniMax-M3, Kimi K3, Kimi K3.1-Pro and Ant's Ling-3.1-flash. He reports that it correctly classified all 24 control models and that he reproduced results for two of the six on real Blackwell hardware. Verification code is on GitHub, but the coverage gives no direct links to it or to the report, so I have not been able to check either.

If the method holds up, it gives regulators and Nvidia's competitors a cheap, repeatable test that needs no access to a lab's servers. It would be the first public technical sign that leading Chinese labs have trained on Blackwell-class chips, which bears directly on whether US export controls are working. It does not show where the chips are, who bought them, or whether the purchases broke any law. They could have been used through overseas cloud rentals or bought through legal channels, so it does not prove smuggling. Any policy or enforcement decision would need separate evidence on supply chains.

The "first public evidence" framing and the figures of about 12,000 GPUs and more than $500M of hardware are estimates built on top of an inference, not measurements. The claim that OpenAI and Anthropic researchers independently confirmed the method rests on DePue's own post on X, with no named confirmers. Headlines implying smuggling go beyond what a fingerprint of the chip type can show. The reproduction on real hardware covers only two of the six models.

1. 01Independent replication of the method by named researchers outside DePue's circle, especially on the four models not reproduced on real Blackwell hardware, and any public confirmation or rebuttal from OpenAI or Anthropic staff.
2. 02Responses from DeepSeek, Moonshot, MiniMax and Ant on how they accessed Blackwell, such as through overseas cloud rental, plus any statement from Nvidia on its own compliance.
3. 03Whether US regulators or Congress cite the analysis in export-control hearings, investigations or new rules, and whether SemiAnalysis or other chip-supply analysts publish matching evidence on where the GPUs are located.