OpenAI employee claims GPU fingerprint evidence shows four Chinese AI labs trained models on banned Blackwell chips

On November 13, developer Will Depue (@willdepue), who identifies as an OpenAI employee, claimed on his personal account to have discovered a "major finding": the latest models from four leading Chinese AI labs—DeepSeek, Moonshot, MiniMax and Ant Group—may have been trained on NVIDIA's most advanced Blackwell chips, which are banned from export to China, and said he had "conclusive forensic evidence." The claim was subsequently amplified by multiple developers.

Confirmed

• The source is Will Depue, who says he is an OpenAI employee, but this statement was a personal act without OpenAI's endorsement.
• The four named organizations are DeepSeek, Moonshot, MiniMax and Ant Group.
• His method is called "checkpoint forensics": the core premise is that nearly all AI accelerators leave identifiable "GPU fingerprints" in model weights after training, so analyzing these fingerprints in public weights can reveal the training hardware. unemploymentlabs published an analysis with an illustrated table mapping models like DeepSeek to restricted Blackwell GPU fingerprints.
• NVIDIA Blackwell chips fall under product categories the US bans from export to China.

Unconfirmed

• The "forensic evidence" itself has not been independently verified; multiple reposts explicitly flagged it as an "unverified claim."
• The four named companies and NVIDIA have not publicly responded, so whether any actual illicit procurement or use occurred cannot be determined.
• The reliability of inferring training hardware from weight fingerprints is itself debated in the community.

Why it matters

• If the allegations hold, it would mean China's top labs obtained banned top-tier training chips through some channel, directly challenging the enforcement of US export controls against China and potentially triggering regulatory and geopolitical ripple effects.
• Even if ultimately unverified, the "checkpoint forensics" approach—reverse-engineering training hardware from public weights—is worth watching: it offers the community a new audit tool, while its methodological reliability will become a focal point of debate.

2026-11-13 ~ 2026-11-13 · 10 related posts

• [source] OpenAI researcher claims weight forensics show Chinese labs trained on banned Blackwell GPUs — willdepue · 2026-11-13
• [source] Checkpoint forensics claims top Chinese AI models were trained on export-banned Blackwell GPUs — ResultBackground2450 · 2026-11-13
• Ex-OpenAI Researcher: DeepSeek, Kimi, MiniMax, Ling All Trained on Nvidia Chips — FateOfMuffins · 2026-11-13

7 near-duplicate retellings: DimitrisPapail · cis_female · aidenybai · wightmanr · marvinvonhagen · felipesuch · felipesuch