Monday 16 November
Global

Your region, kept on this device.

Ex-OpenAI engineer claims Chinese AI models were trained on export-controlled Nvidia Blackwell GPUs

Severity
Medium
Sources
1

Will DePue reports that the weights of six Chinese open-weights models, including DeepSeek-V4-Pro, Kimi K3 and MiniMax-M3, match PyTorch/cuRAND Philox initialisation patterns for Blackwell, with four matching GB200 or GB300 configurations, after searching over six trillion chip-and-seed combinations. The claim is unverified, and the report gives seed values for DeepSeek-V3 (Hopper) and DeepSeek-V4-Flash (Blackwell) so others can reproduce it. Defenders should note that Blackwell use does not establish physical location, so cloud-access controls are a relevant concern.

For security analysts · Written by AI from 1 source · Wrong?

What to do

  1. No action needed.

Written by AI from the sources, checked against them

Full story timeline · 1 update

  1. Fri 13 NovemberChinese AI labs are training models on Nvidia's Blackwell despite export controls, says ex-OpenAI engineerThe Register (all)

Who reported · 1 outlet

The Register (all)

Who and what is in it

Companies

Report a problem