Editorial illustration for Former OpenAI researcher maps six Chinese AI models to Nvidia Blackwell
Ex-OpenAI Researcher Maps Chinese AI to Nvidia Blackwell
Will DePue spent his days at OpenAI building Sora, Sora 2, GPT-4o, o3, and GPT-5. Now running the nonprofit Unemployment Labs, he says he's turned that experience toward a different question: what hardware is actually powering China's AI labs. His answer, laid out in a forensic analysis posted to X, names four companies, DeepSeek, Moonshot AI, MiniMax, and Ant Group, as users of Nvidia's Blackwell GPUs, chips covered by US export controls meant to keep them out of Chinese hands.
This isn't the first time the suspicion has surfaced. US officials and journalists have floated the idea that DeepSeek and Moonshot AI had found a way to Blackwell hardware, but nobody had put forward technical proof. DePue says his method changes that, and that researchers at OpenAI and Anthropic have already checked his work independently and come back with the same conclusion.
The case rests on something buried inside the models themselves, specifically the random numbers that fill a model's billions of weights before training even begins. Those starting values, generated by the seed system built into PyTorch, apparently carry a signature tied to the hardware that produced them.
Why this matters
DePue's method, if it holds up under scrutiny from researchers at OpenAI and Anthropic, gives the field something it's lacked: a way to check export-control compliance claims against the actual silicon signatures baked into model behavior, rather than taking lab statements at face value. For developers and founders building on top of DeepSeek, Kimi, or Ling models, the core finding is practical, not just geopolitical. Core counts of 148 or 152 map to specific Nvidia chips (B200, B300, GB200) with specific memory bandwidth and interconnect profiles, which shapes how these models were trained and what kind of scale their makers can sustain.
Training runs north of half a billion dollars aren't something you run on smuggled chips in small batches. We'd want independent replication before treating this as settled fact, since one researcher's forensic read, however credible his resume, isn't a government audit. But if six labs show the same hardware fingerprint, that's a pattern worth tracking closely, not a one-off anomaly to shrug off.
Related Reading
Browse all OpenAI news →
Research & Benchmarks Microsoft AI Chief: "If It Isn't Safe We Shouldn't Build It
Research & Benchmarks Google's FACTS benchmark shows 70% factuality ceiling across four tests
Research & Benchmarks Nvidia's DLSS 4.5 beta adds 6x Multi Frame Generation for RTX 50 GPUs
Related Topic OpenAI researcher quits, citing distrust over ad‑driven engagement metrics
LLMs & Generative AI
Related Topic OpenAI launches GPT-Image 1.5 with precise editing for enterprise visuals
LLMs & Generative AI
From Archives Google DeepMind Staff Tap UK Government to Escalate Unionization Fight
From Archives Astra's coding test results dropped 40% undetected by safety checks
From Archives ChatGPT adds new Wallet to let AI agents shop and check out for you
LLMs & Generative AI
From Archives Amplitude uses OpenAI's Codex to ship product improvements in 1–2 days
LLMs & Generative AI
Common Questions Answered
What did Will DePue discover about Chinese AI labs and Nvidia Blackwell GPUs?
Will DePue's forensic analysis found technical evidence that leading Chinese AI labs including DeepSeek, Moonshot AI, MiniMax, and Ant Group have access to Nvidia's Blackwell GPUs despite US export controls designed to prevent this. He identified specific hardware fingerprints embedded in the behavior of six Chinese AI models that correspond to Blackwell chip specifications.
How does DePue's method for detecting Nvidia Blackwell usage work?
DePue's forensic analysis identifies silicon signatures baked into model behavior by mapping specific core counts to particular Nvidia chips, such as core counts of 148 or 152 that correspond to specific B200 variants. This method allows researchers to check export-control compliance claims against actual hardware evidence rather than relying on lab statements at face value.
What is Will DePue's background and current role?
Will DePue previously worked at OpenAI where he contributed to building major AI systems including Sora, Sora 2, GPT-4o, o3, and GPT-5. He now runs the nonprofit organization Unemployment Labs, where he has applied his AI expertise to investigate hardware usage in Chinese AI laboratories.
Why is DePue's detection method significant for export control compliance?
DePue's method provides the first technical way to verify whether companies are actually complying with US export controls on advanced chips, rather than simply accepting their claims. Senior researchers at OpenAI and Anthropic have already reviewed and validated his methodology, lending credibility to this approach for monitoring hardware compliance.
What practical implications does this discovery have for developers using Chinese AI models?
For developers building on top of DeepSeek, Kimi, or Ling models, understanding the actual hardware powering these systems helps them make informed decisions about model capabilities and performance characteristics. The discovery also raises important considerations about the reliability and geopolitical implications of relying on these platforms for AI development.
Further Reading
- Papers with Code Benchmarks - Papers with Code
- Chatbot Arena Leaderboard - LMSYS
