An organization called Unemployment Labs issued an analysis on November 12, saying that they had traced the public weights of several Chinese open source models and found digital fingerprints that were consistent with Nvidia’s Blackwell architecture.
Named are DeepSeek, Dark Side of the Moon (Kimi), MiniMax and six models recently released by Ant Group. The study concluded that the models were initialized with high-end hardware subject to U.S. export controls.
The principle can be simply understood as: when different GPUs perform floating-point operations, they will leave subtle and reproducible characteristics in the values. The residual characteristics in the weights can be used to infer which generation of hardware is used. The researchers said they ruled out the possibility of other hardware and verified this method using models with known training hardware.
The same report also acknowledged limitations: fingerprints can point to the architecture, but they cannot trace where the chips are or how they came about, or prove whether export rules were violated. Legal overseas computing power rental and overseas data centers may leave the same traces.
So this study is more like a puzzle piece than a chain of evidence. The timing of its release is also sensitive: the U.S. Department of Commerce is finalizing new regulations to block "remote core use," and Nvidia has just experienced high-level changes. This type of technology traceability will most likely be used more in policy discussions in the future.
For ordinary developers, the impact is not big at the moment, and the weights can still be downloaded. What really needs attention is that if subsequent controls are tightened to the level of cloud computing power, will domestic teams be blocked from renting cards for training overseas?