Six of China’s newest AI models carry the numerical fingerprint of Nvidia Blackwell chips that US export rules bar from China, according to a forensic report published November 12 by former OpenAI researcher Will DePue. The six models come from DeepSeek, Moonshot AI, MiniMax, and Ant Group, and all six were released since April 2026.
DePue’s nonprofit, Unemployment Labs, says it searched more than 6 trillion combinations of chip layout and seed number to match traces left in the models’ published weights. For DeepSeek-V4-Flash, a Blackwell B300 chip reproduced 461 of 469 selected weight rows, 445 of them exactly, bit for bit. A rented Grace Blackwell GB200 system matched all 13 preserved rows of Moonshot’s Kimi K3.
No leak, no insider, no stolen document — just math on files the labs published themselves.
How the DeepSeek Blackwell chips fingerprint works
Every AI model starts as a pile of random numbers. Before training begins, PyTorch fills in those starting weights with a pseudorandom generator called Philox, and Nvidia’s cuRAND library spreads that work across the GPU’s processor cores, known as streaming multiprocessors. Each chip generation has a different core count, and that count shapes the pattern of the starting values.
Hopper chips, Nvidia’s previous generation, have 132 cores. Blackwell’s B200 and B300 have 148. The Grace Blackwell GB200 and GB300 have 152. Some of those starting values survive training untouched, so DePue hunted for them in the open weights the labs publish for anyone to download.
He also had a way to check his work. DeepSeek reused the same seed number, 953049583, for DeepSeek-V3 and DeepSeek-V4-Flash. Run that seed through a Blackwell B300 and you get V4-Flash’s starting weights. Run it through a Hopper chip and you get V3’s. Change either seed by one digit and every match breaks.
“Again, this evidence isn’t based on any insider information or leaking,” DePue wrote when he posted the findings on X, where the post passed 5.8 million views. “It’s based on a forensic analysis of the model weights released by the companies themselves.”
The method checked out against all 24 models whose makers have disclosed their hardware, including Nvidia’s own Nemotron-3 Ultra, which was partly trained on GB200s. Senior researchers at OpenAI and Anthropic independently verified the result, DePue says, though that claim is his own and hasn’t been confirmed by either company.
Which models matched, and which didn’t
Two models match the 148-core Blackwell layout: DeepSeek-V4-Flash and MiniMax-M3. Four match the 152-core Grace Blackwell layout: DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, and Ant Group’s Ling-3.1-flash. The Register reports this is the first public link between MiniMax or Ant Group and Blackwell use.
Older models from the same four labs still read as Hopper. DeepSeek-V3, released in December 2024, matches Hopper, and the switch to Blackwell starts with V4 in April 2026. Kimi K2 from July 2025 reads as Hopper; Kimi K3 from this July reads as Blackwell.
Not every Chinese lab made the jump. The newest open models from Alibaba’s Qwen, Zhipu, Tencent, Xiaomi, and StepFun still read as Hopper. And DeepSeek’s latest release, V4.1-Flash, showed no fingerprint at all after what the report calls an exhaustive search — a null result, not proof it trained on different hardware.
DePue also tried to put a price on the training runs. He estimates the five Blackwell runs used roughly 12,000 GPUs between them, with Kimi K3 needing about 4,500 and DeepSeek-V4-Pro about 3,000, each running for two to three months. At roughly $70,000 a chip, the largest cluster would cost $200 million to $300 million. He raised those figures on Thursday and says plainly that he isn’t confident in them, since they rest on run lengths and efficiency no lab has disclosed.
What the evidence can’t prove
A fingerprint shows which hardware created a model’s starting weights. It doesn’t prove that hardware did the full training run, and it says nothing about who owned the chips or where they sat. Blackwell systems can be rented from data centers outside China, so the same fingerprint could come from a cluster in Malaysia as easily as one in Beijing. TechCrunch notes that offshore rental has driven data center growth in Malaysia.
DePue tested the one obvious alternative. Specially configured AMD hardware, also restricted from export to China, could in theory leave a similar pattern. For DeepSeek-V4-Flash, he ruled it out: in 377 values where Nvidia and AMD results differ, the published weights matched Nvidia in 363 and AMD in none. For the other five models, he calls AMD very unlikely but hasn’t closed the door.
None of this proves a law was broken. The report doesn’t name Nvidia as a source of the chips, and it doesn’t claim any export rule was violated. What it does is replace years of unattributed claims with something anyone can reproduce — DePue published his code and the seed values, and an independent engineer has already released a CPU-only script that recovers the same patterns without renting Blackwell hardware.
Why Washington is paying attention
The report landed a day after Jensen Huang resigned as Nvidia’s CEO on November 12, ending a 33-year run amid a scandal involving alleged personal relationships and a woman US officials believe is a Chinese spy. Huang’s exit came as Congress pushes to tighten chip controls, with four bills in play: the AI Overwatch Act, the Chip Security Act, the Match Act, and the Cloud Security Act.
US officials and press reports had already alleged that DeepSeek and Moonshot AI were using Blackwell chips, without publishing technical evidence. White House technology adviser Michael Kratsios said earlier this year that Moonshot had used export-banned Nvidia chips, citing US intelligence sources. A Reuters tally of Chinese documents counted more than 23 billion yuan in AI data center projects tied to restricted Nvidia chips.
There’s also a second, smaller data point. Lava, an Israel-based security company, said it found 32 Nvidia B300 GPUs through publicly exposed monitoring endpoints with IP addresses geolocated to China. Those chips are part of Nvidia’s Blackwell Ultra generation and are subject to US export controls.
Why this matters
Export controls were supposed to slow China’s AI labs by starving them of top-tier chips. If DePue’s math holds up, four of the country’s leading labs got their hands on Blackwell hardware anyway — and the gap between Hopper and Blackwell in their newest models is visible in the weights themselves.
That doesn’t mean the controls failed. It could mean the chips were rented abroad, which is legal under current rules and exactly the loophole the Cloud Security Act aims to close. Or it could mean smuggling on a scale that would embarrass US enforcement. The fingerprint can’t tell those stories apart, and that’s the point: for the first time, the debate has a piece of evidence both sides have to answer, instead of competing leaks.
Nvidia reports third-quarter results on Tuesday, November 17, with interim CEO Colette Kress facing Wall Street for the first time. Expect questions about Blackwell, China, and what the company knows about where its chips end up. The answers may be thinner than investors want — a chipmaker can track a sale, but it can’t easily track a rental three hops down the chain.
Frequently asked questions
Did DeepSeek train its models on banned Nvidia Blackwell chips?
The report shows the starting weights of six Chinese models were created on Blackwell hardware. It supports, but does not prove, that Blackwell did the training, and it says nothing about where the chips were or whether any law was broken.
How does the Blackwell fingerprint in AI models work?
PyTorch spreads the random numbers that start a model’s weights across a GPU’s processor cores. Blackwell chips have 148 or 152 cores, while the older Hopper generation has 132, so traces of that layout survive training and can be read from the published weights.
Which Chinese AI models carry the Blackwell fingerprint?
DeepSeek-V4-Flash and MiniMax-M3 match the 148-core Blackwell layout, while DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, and Ant Group’s Ling-3.1-flash match the 152-core Grace Blackwell layout.
What happens next in the Blackwell chip export fight?
Congress is weighing four bills to tighten chip controls, including the Cloud Security Act, and Nvidia reports earnings on November 17. The report adds technical evidence to a debate that so far ran on press attributions and intelligence claims.
Sources: TechCrunch, The Decoder, The Register, Unemployment Labs
