Will DePue, an OpenAI researcher who contributed to Sora and GPT-5 post-training, published research today claiming forensic evidence that six AI models from China's top labs were initialized on NVIDIA's Blackwell GPUs, hardware that U.S. export controls prohibit from reaching the country.
The report, released through DePue's nonprofit Unemployment Labs, names DeepSeek-V4-Pro, DeepSeek-V4-Flash, Kimi K3, and MiniMax-M3 among the models that carry what he describes as a hardware fingerprint embedded in their published weights.
How the Method Works
The evidence comes entirely from the models' own published weights, rather than from leaks or insider accounts. The technique recovers traces of the models' initial values, then searches for hardware configurations and seeds that reproduce them. DePue's paper states that the fingerprint "identifies how the weights were created and supports, without proving, the inference that Blackwell also performed the training."
DePue claims he searched more than six trillion combinations of chip layout and seed number. Across that search space, the best accidental match should reach about 8 sigma; the real matches range from 49 to 1,385 sigma, with even the weakest single test hitting 20.
To validate the approach, the research tested 24 models where developers had publicly stated what hardware was used. In every case, the method agreed with the public announcement.
What the Results Claim
DeepSeek-V4-Flash and MiniMax-M3 match NVIDIA's Blackwell-class B200 or B300 chips, which have 148 streaming multiprocessors. DeepSeek's models read as Hopper-trained through V3, released December 2024, and as Blackwell from V4, released April 2026.
DePue estimates that the largest runs, Kimi K3 and DeepSeek-V4-Pro, used 2,000 to 3,000 Blackwell GPUs each. Across the four labs, he estimates the cost of the restricted GPUs exceeds $500 million.
DePue says the results were independently verified by senior researchers at OpenAI and Anthropic. Neither company has publicly commented.
The Export Control Context
The H100, A100, and the entire Blackwell family remain subject to a presumption of denial for direct export to China. NVIDIA's Blackwell B200 and upcoming Rubin architecture remain strictly prohibited for China export for at least 18-24 months after domestic launch, according to January 2026 Bureau of Industry and Security guidance.
The research does not prove where the chips physically are. Moonshot AI, which built Kimi K3, was accused by the White House this summer of accessing Nvidia Blackwell chips via a facility in Thailand. Sources told CNBC that Moonshot has sought access to clusters of 8,000 or more Blackwell GPUs hosted in Europe, describing the Chinese AI lab as "very aggressive" in its pursuit of more computing power.
Prior reporting from The Information, cited by outlets covering the Chinese AI industry, alleged that Blackwell chips were smuggled into China through countries that permitted their sale, then dismantled from data centers and shipped to China after clearing inspection.
Limitations
The fingerprint method identifies the hardware used for weight initialization, not necessarily for the full training run. DePue's own paper states the fingerprint "supports, without proving, the inference that Blackwell also performed the training." A lab could, in theory, initialize weights on one system and train on another, though there would be no obvious reason to do so.
The research has not been peer-reviewed. Unemployment Labs describes itself as a small, independent research group made up of ex-frontier-lab researchers and engineers.
DePue says more checkpoint forensics work is forthcoming. NVIDIA said in June it had already been operating in compliance with Department of Commerce guidance on export restrictions to Chinese-headquartered buyers. The company did not respond to a request for comment on DePue's claims.
