Six models from 4 Chinese labs initialized on Blackwell, he claims
Will Depue posted on X at 21:29:42 UTC on Nov. 12 (6:29:42 a.m. Nov. 13 Japan time) his own analysis claiming that the newest models of China's four leading AI labs, DeepSeek, Moonshot AI, MiniMax and Ant Group, were built on NVIDIA's Blackwell GPUs (chips used for AI computing). He described Blackwell as chips "barred from export to China." He said the result came from analyzing the model weights (the numerical values a model learns in training) that the companies themselves released, and stressed that it was not based on insider information or leaks.
He named six models as the subjects of the analysis, claiming that six models, including "DeepSeek-V4-Pro," "Kimi K3" and "MiniMax-M3," were initialized (the setup of weights before training begins) on Blackwell systems. Four of the six models matched the GB200 and GB300 of "Grace Blackwell," used in NVIDIA's liquid-cooled rack-scale systems, he said. The post names these three models and "DeepSeek-V4-Flash," which appears in his description of the verification; the names of the remaining models are not in the post.
This is a third-party analysis, and it has not been acknowledged by the four companies, NVIDIA or the U.S. government. The opening of the post says the four companies "built" their newest models on Blackwell, but what he presented as verification is a match of which GPU was used at the initialization stage. The post does not address how the chips were obtained, where they are located, or whether all of the training was done on Blackwell.
I have a major result to share. I’ve found conclusive forensic evidence that the top 4 Chinese AI labs—DeepSeek, Moonshot, MiniMax and Ant Group—have built their newest models on NVIDIA’s most advanced Blackwell chips, which are barred from export to China. This sounds insane, but I discovered that almost every AI model carries a hidden signature of the GPU that trained it. When a model is first set up, the starting values of its weights are laid out in a pattern that depends on the exact GPU. Each combination of GPU and seed number produces a unique pattern—a sort of fingerprint of the hardware used to create the model. Traces of that pattern survive training into the published model. Again, this evidence isn’t based on any insider information or leaking. It’s based on a forensic analysis of the model weights released by the companies themselves. I searched more than six trillion combinations of chip layout and seed number and matched them to open-weights models. In doing so, I found conclusive evidence that six Chinese AI models—including DeepSeek-V4-Pro, Kimi K3, and MiniMax-M3—were initialized on NVIDIA Blackwell systems. Four of the models match Grace Blackwell, the GB200 and GB300 parts in NVIDIA’s flagship liquid-cooled rack-scale systems. I then verified these results on real hardware. Running unmodified PyTorch with the seeds I’d recovered, a Blackwell B300 reproduced DeepSeek-V4-Flash’s untouched starting weights bit for bit, and a rented GB200 reproduced Kimi K3’s to a correlation of 0.999995. A Hopper H100 reproduced none of DeepSeek-V4-Flash’s, and changing the seed by one broke the match. (DeepSeek’s seed is 953049583 by the way.) The method also was verified to match 24 known open-weights models to the hardware that developers said they used. It is exceedingly unlikely that these results are a coincidence. Across six trillion combinations, the best accidental match should reach only about 8 sigma (i.e., eight standard deviations above what random chance would produce). The real matches range from 49 to 1,385 sigma, and even the weakest single test hits 20. These were data-center-scale training runs. With some analysis of release schedules and model details, I estimate that the largest, Kimi K3 and DeepSeek-V4-Pro, used 2,000 to 3,000 Blackwell GPUs each. Clusters of this size cost on the order of $150 to $200 million each to buy, and across these labs alone would add up to more than half a billion dollars of export-controlled GPUs, not counting experimental compute or otherwise. It’s quite likely that we see multi-billion dollar expenditure on export-controlled GPUs. I am highly confident in this result, and it has been independently verified by senior researchers at OpenAI and Anthropic. You can read the full report at https://t.co/YOtzND0Jqk. There is much more I’ve identified about these models through other forms of checkpoint forensics like this, and will have more to share soon.
GPU "fingerprint" left in initial weights is the clue, he says
According to Depue, when an AI model is first set up, the starting values of its weights are laid out in a pattern determined by the GPU used. Each combination of GPU and seed number (the starting point for random numbers) produces a unique pattern that acts like a "fingerprint" of the hardware used to create the model, he said. He claims these traces survive training and remain in the published model.
He said he searched more than six trillion combinations of chip layout and seed number and matched them against open-weights models. He wrote that the same method also matched 24 known open-weights models to the hardware their developers said they used. The names of the 24 models are not in the post.
Starting weights reproduced on B300 and GB200, he says
He said he also verified the results on real hardware. Using the recovered seeds and running unmodified PyTorch, the machine learning framework, a Blackwell-generation "B300" reproduced DeepSeek-V4-Flash's starting weights bit for bit, he said. A rented GB200 reproduced Kimi K3's starting weights to a correlation of 0.999995, he said. By contrast, a Hopper-generation "H100" reproduced none of DeepSeek-V4-Flash's starting weights, and changing the seed by one broke the match, he said.
On the chance of coincidence, he wrote that across six trillion combinations, the best accidental match should reach only about 8 sigma (eight standard deviations above what random chance would produce). The real matches ranged from 49 to 1,385 sigma, and even the weakest single test reached 20 sigma, he said. These are his calculations, and ORIENS has not checked how they were done.
2,000 to 3,000 GPUs each for the two largest models, over $500 million in total, he estimates
Based on release schedules and model details, he estimated that the largest models, Kimi K3 and DeepSeek-V4-Pro, were each trained using 2,000 to 3,000 Blackwell GPUs. Clusters of this size cost on the order of $150 million to $200 million each to buy, and across the four labs alone export-controlled GPUs would exceed $500 million, he said. That excludes experimental compute and other uses, and he also wrote that it is "quite likely" that spending on export-controlled GPUs reaches the multi-billion-dollar level. The post does not explain how the estimates were calculated.
Views of the labs, NVIDIA and U.S. government unconfirmed
ORIENS found no official view on the claims from the four companies, DeepSeek, Moonshot AI, MiniMax and Ant Group, or from NVIDIA or the U.S. Commerce Department, as of Nov. 13.
Depue wrote in the post that the result was independently verified by senior researchers at OpenAI and Anthropic. This is his account, and ORIENS could not confirm that either company has announced such verification. He also linked a detailed report, but when ORIENS tried to retrieve it on Nov. 13, the report page returned an error and its contents could not be checked. His X profile reads "prev @ openai" and says he worked on the video generation model "Sora" and other projects.
Separately, CNBC reported on Nov. 5 that Moonshot had held talks to rent Blackwell chips at data centers in Europe, which ORIENS also reported. That is a separate matter from this analysis, and any link between the two is unknown.
He said he has identified much more about these models through similar analysis and will share more soon.
