Will DePue, a former research engineer who has participated in the development of OpenAI Sora, GPT-4o and other models, published a survey on the social platform The companies involved include DeepSeek, Moonshot AI, MiniMax and Ant Group. A total of six models were determined to use the Blackwell architecture for weight initialization. Since these high-end chips are subject to U.S. export controls to China, the research results have also drawn attention to how Chinese AI companies obtain computing resources.
Unemployment Labs, a research organization founded by DePue, pointed out in a report that before an AI model can start training, it must first set initial values for billions or even hundreds of billions of parameters. These seemingly random numbers are actually affected by the GPU computing architecture and random seeds (Seed). Different chips have different numbers of computing units, and even if the same seed is used, the parameter arrangement may be different. Some values have not been modified in subsequent training, so they still retain the original arrangement characteristics, forming what researchers call "GPU hardware fingerprints", which can be used to infer the type of hardware that originally established the model weights.
After analyzing about 6 trillion combinations of chip configurations and random seeds, the research team found that the weight characteristics of DeepSeek-V4-Flash and MiniMax-M3 are consistent with Blackwell B200 and B300; DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro and Ant Group Ling-3.1-flash are consistent with the Grace Blackwell GB200 and GB300 series. In comparison, earlier versions such as DeepSeek-V3 and Kimi K2.5, as well as models such as Alibaba Qwen and Zhipu GLM covered by the study, still exhibit the characteristics of the previous generation Hopper architecture. This method mainly identifies the number of computing units of the GPU, such as Hopper's 132 SMs, Blackwell's 148 SMs, and Grace Blackwell's 152 SMs. It cannot further distinguish all chip models in the same group.
In order to confirm the results, DePue used actual NVIDIA B300 and GB200 chips to regenerate the initial weights of the model and successfully reproduced the values retained in DeepSeek-V4-Flash and Kimi K3; after switching to H100 or changing the random seed, the same results could not be obtained. He also announced the verification code and said it had been independently confirmed by senior researchers from OpenAI and Anthropic. Other researchers have publicly stated that they have successfully reproduced some comparison results, but the two companies have not yet seen a formal verification statement. The full scope of application of this method still needs to be tested by more research. However, hardware fingerprints can identify the initialization characteristics of model weights, and cannot directly prove that the same chip is used throughout the model training process, let alone determine whether the device is located in China or an overseas data center. DePue also acknowledged in the report that the evidence was insufficient to determine how the chips were obtained and whether anyone violated U.S. export controls. Although the research results provide new technical clues that Chinese AI companies may have access to regulated high-end GPUs, the actual hardware sources, training locations and procurement methods of the relevant companies have not yet been confirmed.
