Former OpenAI researcher says banned Blackwell chips trained top Chinese AI models
A former OpenAI researcher says freely released models from China's top labs carry a hidden fingerprint of Nvidia Blackwell chips that were banned from China.

Watch the short
Under a minute, read by an AI anchor. Made . The story has been updated since, and the text below is the current version.
YouTube loads only when you press play.
A former OpenAI researcher, Will DePue, says the open-weight models released by four of China's top AI labs were initialized on Nvidia Blackwell systems. US rules bar those chips from export to China. DePue, who worked on Sora and on training OpenAI's models before leaving, laid out the claim in an X post and a longer report from Unemployment Labs, a nonprofit he runs. He names six models: DeepSeek-V4-Flash and V4-Pro, Moonshot's Kimi K3 and K3.1-Pro, MiniMax-M3, and Ant Group's Ling-3.1-flash. He says the evidence comes only from weights the labs published themselves, not from leaks or insiders.
The method depends on how a model's weights are set before training begins. Labs fill every weight with a value that looks random but comes from a formula that starts from a single seed number, so the same seed always yields the same values. A GPU splits that work across thousands of parallel workers, and how it divides the job depends on the chip's design. DePue's report counts 132 processing cores on Nvidia's older Hopper chips, 148 on the B200 and B300, and 152 on the GB200 and GB300. Two chips given the same seed produce the same numbers laid out in different patterns. That pattern is the fingerprint.
According to the report, the fingerprint survives training because training changes most weights but not all of them. Some weights are never touched, such as those tied to words that rarely appear in training data. Others keep a fainter trace, since values that started high tend to stay relatively high.
To recover the fingerprint, DePue searched more than six trillion combinations of seed and chip layout against the published weights. He then checked each match on a separate batch of weights he had not used to find it. Next he ran ordinary PyTorch on real hardware. A Blackwell B300 regenerated DeepSeek-V4-Flash's untouched starting weights bit for bit. An H100 reproduced none of them, and a rented GB200 reproduced Kimi K3's to a correlation of 0.999995. Changing the seed by one broke the match. Four of the six models, he says, match the GB200 and GB300 parts used in Nvidia's top-end rack systems.
He argues that chance cannot explain the matches. Across six trillion tries, the best accidental match should reach only about 8 sigma, meaning eight standard deviations above what random noise would produce. The real matches range from 49 to 1,385 sigma, and the weakest single test hits 20. He also says the method named the right hardware in all 24 cases where a developer has publicly said what it used, including Meta's Llama 3.1, which ran on H100s. His timeline shows DeepSeek's models reading as Hopper through V3, released in December 2024, and as Blackwell from V4 in April.
The X post and the report differ in tone. The post calls the evidence conclusive. The report says the fingerprint shows which chip set a model's starting numbers, not which chip did the training. It also concedes that a specially configured AMD chip could mimic the layouts. DePue says a numerical test points to Nvidia for DeepSeek-V4-Flash and calls AMD very unlikely for the other five. The method works only on models with public weights, so the leading American models behind paid apps and APIs cannot be checked this way.
The weights cannot show who owned the chips or where they sat. The report says the finding alone does not prove smuggling or a broken law. It notes that Blackwell systems can be rented from data centers outside China, and that US rules generally allow Chinese firms to rent chips abroad even though selling them to those firms is banned.
The claim adds physical-chip evidence to allegations that US officials have made for months without technical proof. In April, a senior Trump administration official said DeepSeek V4 was trained on a Blackwell cluster in Inner Mongolia. DeepSeek denied it, saying it used Nvidia H800s and Huawei Ascend 910C chips. In July, White House science adviser Michael Kratsios said Moonshot had acquired GB300 servers and used GB300s in Thailand. A person familiar with Moonshot's purchasing told Bloomberg the company has a route to Blackwell chips through Southeast Asia, but would not say whether that meant renting or buying. CNBC later reported that Moonshot discussed renting clusters of more than 8,000 Blackwell GPUs from European cloud providers.
DePue estimates that the two largest runs, Kimi K3 and DeepSeek-V4-Pro, used 2,000 to 3,000 Blackwell GPUs each, based on an analysis of release schedules and model details. He prices each cluster at $150 million to $200 million. He puts the total across these labs above half a billion dollars, before counting experimental compute, and suspects the real figure runs into the billions.
The claim arrives as Nvidia faces scrutiny over China. Jensen Huang resigned on Thursday after 33 years, with CFO Colette Kress running the company on an interim basis. Huang has agreed to testify behind closed doors before three House committees before Thanksgiving. Not everyone accepts DePue's conclusions. Dylan Patel, posting as @dylan522p, called the claim nonsense and said the code Moonshot and DeepSeek open-sourced is only for Blackwell. A watcher of Chinese AI posting as @teortaxestex called a few thousand Blackwells 'utterly insufficient' for keeping up at the frontier.
The openness that makes these models popular is also what made the reading possible, since anyone can download the weights and check them. DePue's post warns that now that the method is public, any lab can erase many of these signatures. His code is available to reviewers on request rather than published. The senior researchers at OpenAI and Anthropic who he says verified the result have not been named. He says more findings are coming.
What changed
Six open-weight models from Chinese labs are claimed to have been initialized on Nvidia Blackwell systems, which US rules bar from export to China.
Type: Research · Applies to: DeepSeek-V4-Flash, DeepSeek-V4-Pro, Kimi K3, MiniMax-M3, Ling-3.1-flash
Why it matters
If the claim holds, it would suggest that restricted hardware reached Chinese labs and shaped open models, which could draw tighter export enforcement and closer scrutiny of where AI models come from.
How the story developed
Adds that the researcher says six Chinese models show traces of Blackwell hardware.
Adds the specific claims: six named Chinese models initialized on Blackwell systems, a 24-for-24 match on models with disclosed hardware, a 2 to 3K GPU estimate, and critics' views on scale.
Original source
Further reporting
- thestateofai.comAs Nvidia Faces Scrutiny Over China, a Researcher Says Banned Blackwell Chips Trained China's Top AI Models | State of AI
- letsdatascience.comDeepSeek V4 Trained on Smuggled Nvidia Blackwell Chips, U.S.…
- link.mail.beehiiv.comrebuttal
- link.mail.beehiiv.comrequest for verifiers
- link.mail.beehiiv.comteortaxesTex
- link.mail.beehiiv.comscale note
- link.mail.beehiiv.comzephyr
- link.mail.beehiiv.comeliebakouch
- the-decoder.comChinese models show signs of restricted chips (TD+)
Where we saw it
Latent Space, State of AI (Melissa Rosenthal), THE DECODER
AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works