A forensic analysis of publicly released AI models has found what US officials and press reports had only alleged until now: six recent models from four of China's leading AI labs carry numerical fingerprints matching Nvidia's export-controlled Blackwell GPUs. The report, published on November 12 by researcher Will DePue through Unemployment Labs, says the evidence comes entirely from the models' own published weights, not from leaks or insider accounts.
The six models are DeepSeek-V4-Flash and DeepSeek-V4-Pro, both released in April 2026; MiniMax-M3 from June; Moonshot AI's Kimi K3 from July and Kimi K3.1-Pro from November; and Ant Group's Ling-3.1-flash from October. Two match Blackwell B200 or B300 chips, while the other four match the larger GB200 or GB300 systems. Earlier models from the same companies, including DeepSeek-V3 and Kimi K2.5, read as built on Nvidia's older Hopper generation, suggesting the switch to Blackwell began only in the last few months.
How a weight pattern reveals the chip
The method turns on how a model begins life. Before training, a model's billions of weights get initial values from a pseudorandom number generator called Philox. The software splits that work across parallel threads on the GPU, and the thread count depends on how many processing units, or streaming multiprocessors, the chip has: 132 on Hopper chips, 148 on Blackwell B200 and B300 chips, and 152 on GB200 and GB300 chips. That difference changes where each starting number lands, leaving a layout pattern in the weights.
DePue's team recovered those preserved starting values and searched for the seed and hardware layout that reproduce them. The matches came back at 49 to 1,385 sigma, against a chance benchmark of about 8 sigma. For two models the match was confirmed on real hardware: a single Blackwell B300 reproduced DeepSeek-V4-Flash's initial weights, and a rented GB200 reproduced Kimi K3's to a correlation of 0.999995. It agreed with all 24 models whose hardware developers had publicly disclosed.
Data-centre scale, and the limits of the evidence
The report estimates the training runs were data-centre scale, needing roughly 12,000 Blackwell GPUs across five runs, with the largest, Kimi K3 and DeepSeek-V4-Pro, each using a few thousand chips for two to three months. The biggest clusters would cost roughly $200 million to $300 million each, though DePue says he is not confident in the estimates.
The report is equally careful about what the fingerprint does not prove. It identifies the hardware that created a model's starting weights and supports, without proving, the inference that Blackwell chips also did the training. It says nothing about where the chips were located, how access was obtained, or whether any export control was violated. Blackwell systems can be rented outside China, and the results are, in its own words, not a finding that any law was broken. The only alternative is a specially configured AMD MI300X chip, also export-restricted; ruled out for DeepSeek-V4-Flash and called very unlikely for the other five.
Not every Chinese lab has switched
The pattern is not universal. The newest readable models from Alibaba, Zhipu, Tencent, Xiaomi and StepFun still read as Hopper-trained, and Meituan's LongCat-2.0 shows no Nvidia fingerprint. Blackwell B200, B300, GB200 and GB300 chips require a US export licence for shipment to China, and the report says no such licences have been reported. Neither Nvidia nor any of the four labs had commented at the time of publication. The findings land as the Senate summons Nvidia's former chief executive Jensen Huang to testify over the company's China ties.



