Science & Tech · AI · 3 hrs ago
Chinese AI models show evidence of training on restricted Nvidia chips
Former OpenAI engineer Will DePue says Chinese AI models show signs of training on Nvidia Blackwell chips.
His analysis examined published model weights from DeepSeek, Moonshot AI, MiniMax and Ant Group.
The United States bars exports of these powerful GPUs to China as part of controls intended to limit China's access to advanced AI technology.
DePue says the weights retain a hardware fingerprint from the chips used in training, which he matched against chip and random-seed combinations.
The results suggest Chinese labs may have accessed restricted computing power, raising questions about how well export controls work.
But the findings do not establish where the chips were located or prove that export rules were broken; the models may not have been trained entirely on those GPUs.
The debate over tighter controls is continuing, with several US bills awaiting consideration.
A forensic analysis by former OpenAI engineer Will DePue found signs that models from four Chinese AI companies were trained using Nvidia Blackwell GPUs.
The companies named are DeepSeek, Moonshot AI, MiniMax and Ant Group.
DePue said he compared published model weights with GPU-generated patterns after searching more than six trillion combinations of chip layout and seed number.
The findings do not establish that the models were trained entirely on Blackwell chips or that export-control laws were violated.
The chips could have been accessed through offshore data centers, which Chinese companies have used to obtain computing power.
- Who
- Will DePue, a former OpenAI engineer, conducted the analysis. The models are from DeepSeek, Moonshot AI, MiniMax and Ant Group.
- What
- A forensic analysis found signs that the companies’ models carry fingerprints of Nvidia Blackwell GPUs.
- When
- The reports were published on November 13, 2026.
- Where
- The models were developed by Chinese AI companies. The location where the chips were used is not established.
- Why
- DePue compared published model weights with patterns associated with GPU hardware and seed values.
This story does not have two clearly opposing sides.
Epoch AI estimates that about two million advanced chips entered China through schemes designed to hide their final destination.
White House technology adviser Michael Kratsios said Moonshot had used export-banned Nvidia chips, citing US intelligence sources.
Reports published DePue’s analysis finding Blackwell GPU fingerprints in models from four Chinese AI companies.
- Researcher
- Will DePue, former OpenAI engineer
- Companies named
- DeepSeek, Moonshot AI, MiniMax and Ant Group
- Chip generation
- Nvidia Blackwell
- Search scale
- More than six trillion chip-layout and seed-number combinations
- Malaysia imports
- 275,000 Nvidia AI accelerators in 2024 and 2025
2 sources
What people are saying
This sounds insane, but I discovered that almost every AI model carries a hidden signature of the GPU that trained it
I searched more than six trillion combinations of chip layout and seed number and matched them to open-weights models
If we care about the U.S. tech edge, it doesn’t make any sense to me that you would not allow Chinese companies to access chips in mainland China or Hong Kong and Macau, but then say, actually, if you just go over to Malaysia, you can do whatever you want
Posts on X
I have a major result to share. I’ve found conclusive forensic evidence that the top 4 Chinese AI labs—DeepSeek, Moonshot, MiniMax and Ant Group—have built their newest models on NVIDIA’s most advanced Blackwell chips, which are barred from export to China. This sounds insane,