Analysis Detects Blackwell GPU Traces in Six Chinese AI Models
The analysis links the fingerprints to use during model initialization, but cannot locate the chips or establish whether U.S. export rules were violated.

Unemployment Labs’ Nov. 12 analysis identified digital patterns in public weights from six recent Chinese AI models that match NVIDIA’s Blackwell architecture, pointing to use of hardware subject to U.S. export controls during model initialization.
The models came from DeepSeek, Moonshot AI, MiniMax and Ant Group. The analysis examined publicly available weights, using forensic tracing to detect hardware signatures. It says researchers ruled out other hardware as a match and validated the method against models whose training hardware was known.
The analysis presents the matches as evidence that Blackwell architecture had begun to be used in model training. The signatures do not reveal where the chips were located or how they were obtained, and the findings do not establish that any U.S. export-control rules were violated.