ANALYSIS

Forensic Analysis Finds Nvidia Blackwell Traces in Six Chinese AI Models

M Marcus Rivera Nov 14, 2026 2 min read
Engine Score 8/10 — Important

Alleged chip-ban evasion in Chinese AI models, high novelty

Editorial illustration for: Forensic Analysis Finds Nvidia Blackwell Traces in Six Chinese AI Models
  • Former OpenAI researcher Will DePue says forensic analysis found Nvidia Blackwell GPU signatures hidden inside six Chinese AI models.
  • DeepSeek, Moonshot AI, MiniMax, and Ant Group models released since April 2026 carry the pattern, despite US export controls on Blackwell chips.
  • The method reads weight-initialization patterns tied to each GPU’s compute-core count: 132 for Hopper chips versus 148 or 152 for Blackwell.
  • DePue says senior researchers at OpenAI and Anthropic have independently confirmed his findings, and he has published verification code on GitHub.

What Happened

Former OpenAI researcher Will DePue says a forensic analysis of model weights found the first technical evidence that Chinese AI labs have access to Nvidia’s export-controlled Blackwell GPUs, The Decoder reported on November 13, 2026. DePue, who worked on OpenAI’s Sora and GPT-5 before founding Unemployment Labs, says the pattern appears in six models released since April 2026.

Why It Matters

The finding adds technical substance to long-standing suspicions that DeepSeek and Moonshot AI were training on restricted hardware, following Nvidia telling Congress it could track chips to stop smuggling and documents tying banned Nvidia chips to a Shanghai data center. US officials had suspected chip access but had not published proof.

Technical Details

DePue’s method reads weight-initialization patterns left by PyTorch, which fills untrained weight slots with random values distributed according to each GPU’s compute-core count: 132 cores for Hopper chips like the H100 versus 148 or 152 cores for Blackwell’s B200/B300 and GB200/GB300. He mapped DeepSeek-V4-Flash and MiniMax-M3 to 148-core chips, and DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, and Ant Group’s Ling-3.1-flash to the 152-core GB200/GB300 pattern, estimating the five runs combined used roughly 12,000 Blackwell GPUs.

Who’s Affected

The findings implicate DeepSeek, Moonshot AI, MiniMax, and Ant Group in apparent violations of US export controls, intensifying scrutiny already facing Nvidia over how its restricted chips reach Chinese AI labs.

What’s Next

DePue has published his verification code on GitHub for independent review, and he says senior researchers at OpenAI and Anthropic have already confirmed the method, setting up further scrutiny of the named labs’ hardware sourcing.

Share

Enjoyed this story?

Get articles like this delivered daily. The Engine Room — free AI intelligence newsletter.

Free · No spam · Unsubscribe anytime