A detailed analysis report from a foreign source claims that China's top four AI labs—DeepSeek, Moonshot, MiniMax, and Ant Group—used NVIDIA's most advanced Blackwell chips to build their latest models. The core proof comes from the "hardware fingerprint" in the publicly released weights of the model; specifically, it's the numerical pattern left by the thread layout determined by the number of streaming multiprocessors (SMs) on the GPU when PyTorch initializes random weights. Methodology (the core technical basis of the report): PyTorch initializes the weights using a Philox pseudo-random number generator before training begins. The thread allocation in this process depends on the number of SMs (Streaming Service Providers) on the GPU: Hopper series (H100 SXM / H200 / H800): 132 SMs (approximately 270,336 threads) Blackwell B200/B300: 148 SMs (approximately 303,104 threads) Grace Blackwell GB200/GB300: 152 SMs (approximately 311,296 threads) The report analyzed several open-source models released in 2026 and found that: DeepSeek-V4-Flash, MiniMax-M3: match 148 SMs (B200/B300) DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, Ling-3.1-flash (Ant Group): match 152 SMs (GB200/GB300) These earlier lab models (such as...) DeepSeek-V3, Kimi K2.5, MiniMax-M2.7, and Ling-3.0-flash) match Hopper's 132 SM, switching to the Blackwell layout for new models starting in mid-2026. Detailed analysis report: https://(unemploymentlabs.com)/checkpoint-forensics