Forensic analysis of the weights of AI models has led to the discovery of multiple Chinese AI labs likely training large models with NVIDIA Blackwell GPUs. This is significant as it implies that advanced American hardware has managed to find its way into restricted locations despite existing export controls. This is done by looking for certain hardware fingerprints in the initial weights of the AI models as opposed to finding supply chain related proof.
Hidden hardware fingerprints in DeepSeek and Moonshot AI weights point to restricted NVIDIA Blackwell access
AI models from DeepSeek and Moonshot AI were investigated. (DeepSeek-V4-Flash DeepSeek-V4-Pro MiniMax-M3 Kimi K3 Ling-3.1-flash). They identified certain initialization patterns within the models consistent with the structure of NVIDIA Blackwell GPUs. These patterns can be matched to help researchers infer the computing hardware that was used to train the model without physically accessing the hardware servers.
According to estimates, each training run of these models has used clusters in the range of 3000-4500 Blackwell GPUs. The cumulative total footprint (across analyzed models) is estimated to be in the range of 7000-22000 GPUs. Conversely, there are no GPU fingerprints in the cases of Alibaba Qwen3, or Meituan LongCat-2.0 (suggesting different hardware or ASICs). Also, the DeepSeek-V4.1-Flash is based on a new architecture, which obscures the GPU initialization footprints.
Unemployment Labs has said that this analysis is providing technical indication of access to advanced computing hardware, but that does not mean we know where the chips were. They've also emphasized that this does not mean we know how access was gained, or that export controls were broken. We've been following Blackwell AI GPUs pretty closely and can refer back to our previous coverage on Moore Threads MTT S5000 GPU Runs for background on different hardware options.
It's confirmed that there are AI models that have hidden hardware signatures associated with them suggesting that there's a lot of NVIDIA Blackwell being used. The discovery is based on forensic analysis of the weights of the models (as opposed to actual observation of hardware usage or supply chain logs). This is important as it means we don't know the exact means of acquisition and where the chips were physically located.
Comments (0)