Report: 6 Chinese AI Models Started on Nvidia Blackwell

§ 01What the report claims#

• The post. DePue posted the result on X at 21:29 UTC on November 12. The post had about 533,000 views when we read it at 22:49 UTC.
• The claim. The report says six models released since April 2026 “contain numerical patterns matching Blackwell hardware”, while earlier models from the same labs match Nvidia’s older Hopper generation.
• The evidence. It comes from the weights the labs published themselves. The report says it uses no leaks and no insider accounts.
• Other labs. The report says the newest models it could read from Alibaba, Zhipu, Tencent, Xiaomi and StepFun match Hopper.

Sigma is the report’s measure of how far a match stands above chance. It says the best chance match across its whole search would reach about 8 sigma.

Two notes on the table. The report says Kimi K3.1-Pro starts from the same seed as Kimi K3 and may be K3 with further training. And it dates each model by when its weights were posted: our own page on Kimi K3.1 records that Moonshot released that model on October 23 and its weights on November 4.

§ 02How the method works#

• Starting values. Before training, a model’s weights are filled with numbers from a random number generator. The generator starts from one number, the seed. The same seed gives the same numbers every time.
• The layout. PyTorch splits that job across many parallel threads on the GPU, and the number of threads depends on how many processor units the chip has. DePue’s thread says: “Hopper has 132, Blackwell has 148 and Grace Blackwell has 152.”
• Why the chip shows. The same seed therefore puts the same numbers in different positions on each chip class.
• What survives. Training changes most weights. DePue’s thread says some are never touched and the rest still correlate with where they started.
• The search. Pick a seed and a chip layout, generate the starting values, and compare them with the published weights. The report says it searched roughly 6 trillion combinations.

§ 03What was checked on real hardware#

• DeepSeek-V4-Flash. The report says a Blackwell B300 running unmodified PyTorch with the seed 953049583 reproduced 461 of 469 selected rows of the model’s weights, 445 of them bit for bit.
• The control. A Hopper H100 with the same seed reproduced DeepSeek-V3’s preserved values and none of V4-Flash’s rows. Changing the seed by one reproduced nothing on either machine.
• Kimi K3. A rented GB200 reproduced all 13 of the model’s preserved rows with the seed 20256586.
• Two of six. The report’s summary says the match was confirmed on physical Blackwell chips for DeepSeek-V4-Flash and Kimi K3. It names no other model there.
• Known answers. For 24 models whose developers have named their hardware, the report says its reading agrees each time. It calls this “a consistency check rather than a measured accuracy rate”.

§ 04What the report says it cannot show#

• Where the chips are. The report says: “The weights also cannot reveal who owned the chips or where they were located.”
• Rented abroad. It adds: “Blackwell systems can be rented from data centers outside China.”
• No finding of a violation. Of its own evidence, the report says “it does not establish that chips were smuggled into China or that an export-control violation occurred”.
• The start, not the whole run. In its words: “The method identifies the hardware that generated the starting weights, not every chip that later helped train the model.”
• An AMD alternative. The report says an AMD MI300X limited to 152 or 148 of its processing units reproduces the same layouts. A second test excludes AMD for DeepSeek-V4-Flash. For the rest, its words are “consider it very unlikely for the other five models”, which is not the same as excluded.
• Models it cannot read. Closed models have no published weights. Some frameworks leave no pattern. The report found no fingerprint in DeepSeek-V4.1-Flash.

§ 05Where the thread goes further than the report#

• Conclusive, in the thread. The thread opens: “I’ve found conclusive forensic evidence that the top 4 Chinese AI labs”.
• Conclusive, in the report. Its conclusion calls the evidence strong for the six models and says “for DeepSeek-V4-Flash the evidence is conclusive”.
• Started on, or trained on. The thread says the labs built their newest models on Blackwell. The report says the fingerprint “supports, without proving, the inference that Blackwell also performed the training”.
• Export rules. A reply in the thread says “it’s clear that the export bans are deeply failing”. The report makes no finding that a rule was broken.
• Who verified it. The thread says “it has been independently verified by senior researchers at OpenAI and Anthropic”. The report names no outside verifier. We did not look for a statement from either company.
• Code. Another reply says: “I’ll be sharing a GitHub repo and full paper soon with implementation details”. Until then, nobody outside can re-run the search from his code.

§ 06The scale estimate#

The report estimates how many Blackwell GPUs each training run needed. It assumes each model was trained on the hardware that started it, which the report itself says is not proven.

The ranges are wide. The report prices a Blackwell GPU at roughly $70,000, puts the largest clusters at $150 million to $200 million, and the five together at more than half a billion dollars. These are its estimates, not measurements.

§ 07What we did not check#

• We ran none of the tests and downloaded no model weights.
• We read the report’s summary, findings, limits, verification and conclusion sections. We did not read all of its method section or its second appendix, look at its figures, or open the sources it cites, including the US export rule.
• One search of X at 22:41 UTC found no post since the thread from accounts of Nvidia, DeepSeek, Moonshot, MiniMax or Ant Group. We guessed the account names, and we did not check the companies’ own sites.
• We do not know whether any of the four labs holds a license, or where any of these chips are.

§ 08Where CellCog sits#

CellCog is our product, an AI employee platform. Its agents run on Anthropic’s Claude models, and it runs none of the six models in the report. Two things a reader should know. Anthropic, whose researchers DePue says verified the result, is our model provider. And our Privacy Policy names MiniMax as the provider of our speech and voice cloning models, through fal.ai. The MiniMax model in the report’s Table 1 is MiniMax-M3, a language model. OpenAI, where DePue’s profile says he worked, makes the image model behind this page’s pictures.

§ 09Sources#

• Will DePue: Top Chinese Frontier AI Models Carry the Fingerprint of Export-Controlled NVIDIA Blackwell GPUs, Unemployment Labs, November 12, 2026, read at 22:42 UTC that day.
• Will DePue: thread on X, November 12, 2026, 21:29 UTC.
• CellCog: Privacy Policy, for the MiniMax line.