On this page · 5 sections ▾
Will DePue, a former OpenAI researcher who now runs the nonprofit research group Unemployment Labs, published an analysis on November 13 claiming that several leading Chinese open-weight AI models were initialized on Nvidia Blackwell GPUs. Those chips are barred from export to China. The findings, covered by TechCrunch, The Register and Tom's Hardware, do not establish that any law was broken.
Key facts
- DePue analyzed published weights from DeepSeek, Moonshot AI, MiniMax and Ant Group, according to TechCrunch.
- He says he searched more than six trillion combinations of chip layout and seed number, per The Register, and found six Chinese models initialized on Blackwell systems, including DeepSeek-V4-Pro, Kimi K3 and MiniMax-M3.
- Four of the six match Grace Blackwell parts, the GB200 and GB300, The Register reports.
- Tom's Hardware lists DeepSeek V4, Kimi K3.1, MiniMax-M3 and Ling-3.1-Flash among the models matched.
- The Register says this is the first public link between MiniMax or Ant Group and Blackwell use.
How the method works
When a model is set up, its weights receive pseudo-random starting values. Using PyTorch and Nvidia's cuRAND library, that work is split across a GPU's shader modules and threads, so a given seed produces a pattern tied to the chip layout. Tom's Hardware notes that Hopper chips have 132 shader modules, Blackwell B series 148 and Blackwell GB parts 152.
DePue says traces of that pattern survive training. Researchers reproduced the initialization on matching hardware and found the weight distribution matched the Chinese models. Changing the shader and thread counts broke the correlation, according to Tom's Hardware. For DeepSeek V4, the researchers also say AMD hardware did not reproduce the weight pattern.
The report includes seed values linking DeepSeek-V3 to Hopper and DeepSeek-V4-Flash to Blackwell so others can attempt to reproduce it. DePue claims senior researchers, including some inside OpenAI and Anthropic, have confirmed the result. That claim is his own and is not independently verified in the coverage.
What the evidence does not show
The limits are significant. The method only works on open-weight models and depends on the hardware and software stack used. It identifies chip families and approximate card counts, not individual card variants, Tom's Hardware reports. It also cannot show where the GPUs were located.
TechCrunch points out that the chips could have been in an offshore data center, a popular route for Chinese firms that it says has driven growth in Malaysia. Epoch AI estimates 275,000 Nvidia accelerators were imported to Malaysia in 2024 and 2025. Tom's Hardware says that, as far as it can tell, renting Blackwell hardware abroad is not currently illegal for Chinese companies. The Register adds that DePue acknowledged the possibility of access through US providers.
Why it matters
The findings land in a charged policy setting. TechCrunch reports that Nvidia CEO Jensen Huang left the company this week after admitting an affair with Alice Huang, an executive at Megaspeed, a company implicated in smuggling Nvidia chips into China. The Register says Megaspeed is reportedly under investigation by US authorities and that the Trump administration said on Thursday it was unsure whether Alice Huang was a spy. Beijing denies the allegation, and the FBI is investigating.
The Register also lists four bills aimed at limiting China's access to leading AI technology: the AI Overwatch Act, the Chip Security Act, the Match Act and the Cloud Security Act. Sam Bresnick of Georgetown's Center for Security and Emerging Technology told TechCrunch earlier this year that it makes no sense to block chips in China but allow them in Malaysia.
