Chinese AI Models Use Blackwell GPUs, Analysis Reveals

Will DePue indicates that he has identified, within six Chinese models released since April 2026, a characteristic initialization fingerprint of Nvidia's Blackwell GPUs. He specifies that his method has been validated by researchers from OpenAI and Anthropic and has made his code available. Several systems exhibit specific configurations and costs exceeding $500 million, without any evidence of illegal chip trafficking or the location of the machines being demonstrated.

The method does not establish an infraction and does not locate the GPUs

The technique described only provides information about the type of hardware that initialized the weights of the models, without confirming a complete training on those same systems. Another technical hypothesis mentioned is the use of a throttled AMD MI300X to mimic the pattern, a possibility that Will DePue claims to have ruled out for DeepSeek‑V4‑Flash due to rounding differences. The published data does not reveal the location of the GPUs or their owner. Chinese companies can also rent computing power outside of China. Access to Nvidia's top hardware, subject to export restrictions, is mentioned, but no physical trafficking or violation of U.S. legislation is proven.

Five sessions estimated at 12,000 Blackwell GPUs and over $500 million

Will DePue estimates that approximately 12,000 Blackwell GPUs are required for the five training sessions analyzed, a level presented as equivalent to 30,000 H100s. In this total, Kimi K3 would have utilized nearly 4,500 processors over about 70 days, and DeepSeek‑V4‑Pro around 3,000. Based on a projected unit price of around $70,000, each cluster would range between $200 million and $300 million, bringing the total for the five systems to over $500 million.

Models associated with 148 or 152 cores according to the fingerprints

The proposed associations link DeepSeek‑V4‑Flash and MiniMax‑M3 to B200 or B300 chips, whose initialization would reflect 148 cores. DeepSeek‑V4‑Pro, Kimi K3, Kimi K3.1‑Pro, and Ant Group's Ling‑3.1‑flash would exhibit a pattern of 152 cores characteristic of GB200 or GB300 systems. These signatures are deduced from the fact that the H100/H800 have 132 cores, the B200/B300 have 148, and the GB200/GB300 have 152. Will DePue claims to identify the Blackwell generation in six models released since April 2026, while earlier versions like DeepSeek‑V3 and Kimi K2.5 would refer to the Hopper generation.

Fingerprinting procedure derived from weights and announced verifications

The clues come from the models themselves: billions of weights initialized by PyTorch via a seed, distributed by the GPU among its computing cores. The difference in cores per architecture induces an initialization pattern, with intact values remaining for unused symbols and a low statistical echo in trained areas, according to Will DePue. He claims to have reproduced the pattern of DeepSeek‑V4‑Flash and Kimi K3 on actual Blackwell hardware and has uploaded a verification code. He adds that senior researchers from OpenAI and Anthropic have independently confirmed his results, that randomness has been statistically ruled out, and that 24 control models with known hardware, including Llama 3.1 on H100, have been correctly identified. He presents his work as the first technical proof of such access.

Overview of Chinese players and external convergences

Will DePue attributes the use of Blackwell GPUs to DeepSeek, Moonshot AI, MiniMax, and Ant Group, following a period where American officials and media already suspected DeepSeek, and for some, Moonshot AI, without published evidence. His conclusions overlap with Dylan Patel's analysis, which suggests that the architecture of DeepSeek‑V3 remained suited to the Hopper, while that of DeepSeek‑V4 would target both Nvidia's Blackwell and Huawei's Ascend. At the same time, Alibaba's Qwen3.8 and the models from Zhipu, Tencent, Xiaomi, and StepFun would only show Hopper signatures, while Meituan's LongCat‑2.0 would point to Huawei hardware.