Researcher says Nvidia Blackwell chips trained leading Chinese open-weight AI models
For a model-risk officer at a regional bank
Will DePue, a former OpenAI researcher who leads the nonprofit research group Unemployment Labs, found evidence that leading open-weight AI models from Chinese companies were trained on Nvidia Blackwell chips banned from export to China, TechCrunch reported.1
DePue, a former OpenAI researcher, analyzed the weights of open models released by DeepSeek, Moonshot AI, MiniMax and Ant Group, according to TechCrunch; the weights still reflect the configurations of the chips used to train them.1 TechCrunch said DePue found compelling evidence that Moonshot's Kimi K3 releases and DeepSeek's V4 models bear fingerprints of training on Blackwell chips; the report cites no published methodology, and neither company's response is given.1
TechCrunch said this does not mean the models were trained entirely on Nvidia chips or that export laws were necessarily violated, because the chips could have been in an offshore data center, a common way for Chinese firms to access compute.1
White House technology adviser Michael Kratsios said earlier this year that Moonshot had used export-banned Nvidia chips, citing U.S. intelligence sources, according to TechCrunch.1
Epoch AI, a nonprofit research organization, estimates that about two million advanced chips entered China in 2024 and 2025 through schemes designed to hide their final destination, and that 275,000 Nvidia AI accelerators were imported to Malaysia in those years.1
Blackwell GPUs are banned from export to China or Chinese-owned firms, while the Trump administration allows exports of the less powerful H200 under a route that lets the U.S. claim a 25% tariff, TechCrunch said.1
Banks and fintechs that adopt free Chinese open-weight models may inherit questions about how those models were trained, which could matter for vendor due diligence, sanctions and export-control exposure, and regulator scrutiny. Know which systems run these files and what you can document about their origin.
Our take. If DeepSeek V4 or Kimi K3 runs anywhere in your bank, the model file itself now carries evidence of training on chips the U.S. barred from China. The download was free; the provenance question is not.
Action. Model risk: Flag DeepSeek V4, Kimi K3 and other Chinese open-weight models in the model inventory, including vendor products built on them.
What our records say
NVIDIA sells GPUs, DGX systems and the licensed NVIDIA AI Enterprise stack so banks can train, customise and serve models, including in-house models and agents, in their own data centres or in the cloud.1 NVIDIA designs GPUs and AI systems and sells them with a licensed software stack for training, customising and serving models, and banks run the stack in their own data centres or rent it in the cloud.1
From our records
- NVIDIAvendor · banks using it (6) BNY, Amex, Capital One · alternatives SambaNova
- ChatGPT Enterprise (OpenAI)vendor · banks using it (17) Morgan Stanley, BNY, Customers Bank · alternatives Microsoft Copilot, Claude, Gemini Enterprise