Forensic Analysis: Top Chinese AI Models Fingerprinted to Export-Banned NVIDIA Blackwell GPUs
elie · x · 2026-11-16
Unemployment Labs publishes a forensic report claiming that six models released since April 2026 by DeepSeek, Moonshot AI, MiniMax and Ant Group carry numerical patterns in their published weights matching NVIDIA's export-controlled Blackwell GPUs, while those companies' earlier models match the older Hopper generation — suggesting leading Chinese AI labs have trained newest models on restricted hardware.
- Method: Evidence comes entirely from the models' own released weights — the team recovers traces of initial values and searches for hardware configurations and seeds that reproduce them; matches for DeepSeek-V4-Flash and Kimi K3 were also confirmed on physical Blackwell chips. The fingerprint shows the weights were created on Blackwell, supporting (without proving) that training used it too.
- Models flagged: DeepSeek-V4-Flash/V4-Pro (B200/B300, GB200/GB300), Kimi K3 and K3.1-Pro (GB200/GB300), MiniMax-M3 (B200/B300), Ant Ling-3.1-flash (GB200/GB300).
- Alternative: specially configured AMD hardware (also export-restricted) is the only alternative identified — excluded for DeepSeek-V4-Flash and deemed very unlikely for the rest.
- Limits: the findings establish technical evidence of access to advanced compute but not chip location, how access was obtained, or export-control violations. Prior allegations about DeepSeek and Moonshot lacked technical evidence.
Elie, who shared it as a weekend read, says "the science behind the analysis looks legit."
More from Infra
- Inside SGLang Summit 2026: 23 sessions in one track and a striking open-source turnout — ying11231 · 2026-11-16
- EPYC 32-core + 512GB DDR4 + 44GB VRAM: is a DIY local LLM rig still worth it? — sitytitan · 2026-11-16
- Samsung open-sources LittleBit: 13B LLM under 1GB with sub-bit weights and XOR compute — dr_cintas · 2026-11-16
- Samsung to invest $1.9B in Austin to make chips for SpaceX Starlink satellites — elonmusk · 2026-11-16
- Jensen Huang rumored plotting return to Nvidia CEO as Musk, Sweeney back public letter — The Verge AI · 2026-11-16
- Open-source ComfyUI node offloads text encoders to a spare GPU on another machine — randomlyfixed · 2026-11-16