KEY POINTS
- Former OpenAI engineer Will DePue says six Chinese open-weights models were initialized on Nvidia Blackwell hardware.
- The method matches the initial weight patterns of published models against patterns generated by specific GPUs and seed numbers.
- DePue notes the result does not show where the chips were located, so US-based cloud access remains possible.
A former OpenAI engineer says several leading Chinese AI labs have trained models on Nvidia's Blackwell GPUs, which US export controls restrict from sale to China. The Register reported on the analysis, which was published by Will DePue, who now runs an AI research consultancy called Unemployment Labs.
DePue names DeepSeek, Moonshot AI, MiniMax and Ant Group as having gained access to the hardware. His technique relies on the way a model's weights are first set up. Those starting values come from a pseudorandom number generator, and according to DePue the resulting pattern depends on the exact GPU and the seed number used. He argues that traces of this pattern survive training and remain visible in the published model.
The report describes the process in more detail. Philox, a pseudorandom algorithm, produces the same initial values whenever it runs with the same seed. In PyTorch, initialization can run across many parallel GPU threads using Nvidia's cuRAND library. Because the output is deterministic, a researcher can pick a chip layout and a seed, generate the expected initial weights and compare them with a model's released weights.
DePue says he searched more than six trillion combinations of chip layout and seed number against open-weights models. He concluded that six Chinese models, including DeepSeek-V4-Pro, Kimi K3 and MiniMax-M3, were initialized on Blackwell systems. Four of them match Grace Blackwell parts, the GB200 and GB300 used in Nvidia's liquid-cooled rack-scale systems. He says this is the first public link between Blackwell and both MiniMax and Ant Group.
DePue also lists limits to the approach. Using Blackwell chips does not show where they were physically located, so some labs may have rented the hardware from US providers. The Cloud Security Act, one of several pending bills, is meant to close that gap. His report includes seed values tying DeepSeek-V3 to Nvidia's Hopper generation and DeepSeek-V4-Flash to Blackwell, so others can try to reproduce the findings. He claims multiple senior researchers, both independent and at frontier labs such as OpenAI and Anthropic, have confirmed the result.
The analysis follows another finding from a few weeks earlier. Israel-based security company Lava said it found 32 Nvidia B300 GPUs, part of the Blackwell Ultra generation, through publicly exposed monitoring endpoints with IP addresses geolocated to China. Lava noted that these GPUs are subject to US export controls.
The reports arrive during a renewed push in Washington to tighten chip rules. House Speaker Mike Johnson supports the idea, and at least four bills await consideration: the AI Overwatch Act, the Chip Security Act, the Match Act and the Cloud Security Act. The Register also noted that OpenAI and Anthropic have both complained about Chinese labs copying their models through distillation, a practice the US Cybersecurity and Infrastructure Security Agency has also noted.
