Ex-OpenAI researcher finds Nvidia Blackwell fingerprints in six Chinese AI models
The Decoder · rss · 2026-11-13
Former OpenAI researcher Will DePue says he has produced the first technical evidence that DeepSeek, Moonshot AI, MiniMax, and Ant Group trained on Nvidia's latest Blackwell GPUs despite US export controls. His forensic method reads characteristic chip signatures directly from published model weights, covering six Chinese models and five training runs he estimates cost over half a billion dollars in hardware. If validated, weight fingerprints could become a new compliance-audit tool for chip export controls.
More from Infra
- Nvidia China GPU funnel under scrutiny: ~20% of compute revenue may be China-linked, guidance cut looms — SumitGup · 2026-11-14
- NVIDIA touts SGLang partnership: 5 GPU generations, 60+ day-zero model launches — NVIDIAAI · 2026-11-14
- Qwen3.5-9B runs offline on a stock PS4 at 39 tok/s using its own MTP head — leonardo_alegria · 2026-11-14
- llama.cpp 1.0 commits credit a locally-run DeepSeek on Mac Studio as coding assistant — krishnan · 2026-11-13
- llama.cpp hits v1.0.0 after 3.5 years, promising semver for its C API — krishnan · 2026-11-13
- ACM SIGOPS publishes Seoul Report mapping systems research amid AI and hardware shifts — anaklimovic · 2026-11-13