A researcher says he can trace six leading Chinese models back to Blackwell GPUs that were never supposed to reach China. It lands in the same week Nvidia's CEO walked out the door.
AdvertisementI have covered the chip war long enough to know that most "banned chips in China" stories are one anonymous source and a lot of adjectives. This one is different, because it does not rely on a tip. It relies on the models themselves.
The analysis comes from Will DePue, who leads a nonprofit research group called Unemployment Labs. As TechCrunch reported on Friday, he looked at the published weights of open models from DeepSeek, Moonshot AI, MiniMax and Ant Group. His conclusion is that the most closely watched ones, including the Kimi K3 releases and DeepSeek's V4 family, show signs of having been trained on Blackwell hardware.
What the research actually claims
According to a report summarising the analysis, six models were flagged: DeepSeek-V4-Flash, DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, MiniMax-M3 and Ant Group's Ling-3.1-flash. Four of them reportedly match Nvidia's GB200 or GB300 systems, and two match B200 or B300 chips.
| Model | Lab | What the analysis reportedly flags |
|---|---|---|
| DeepSeek-V4-Pro | DeepSeek | Blackwell fingerprint; one of the two largest runs, estimated at roughly 2,000 to 3,000 GPUs |
| DeepSeek-V4-Flash | DeepSeek | Starting weights reproduced exactly on a rented B300 |
| Kimi K3 | Moonshot AI | Blackwell fingerprint; 13 preserved rows reproduced on a rented GB200 |
| Kimi K3.1-Pro | Moonshot AI | Blackwell fingerprint |
| MiniMax-M3 | MiniMax | Blackwell fingerprint |
| Ling-3.1-flash | Ant Group | Blackwell fingerprint |
How do you "fingerprint" a chip from a model?
My understanding, in plain terms: the weights of a trained model still carry traces of how the hardware did its maths at the start of training. DePue says he recovered the random seed used for one model, then ran it on real Blackwell hardware. A B300 reproduced DeepSeek-V4-Flash's untouched starting weights bit for bit. A rented GB200 matched all 13 preserved rows he had for Kimi K3. That is a much harder thing to wave away than a leaked rumour.
Why this is not an open-and-shut case
Here is where I want to slow down, because the headlines will not. Having Blackwell fingerprints does not automatically mean a law was broken. TechCrunch points out that these chips could have been in an offshore data center, which is a popular way for Chinese firms to get compute. Malaysia has become a hub for exactly that. Epoch AI estimates 275,000 Nvidia accelerators were imported into Malaysia across 2024 and 2025.
Epoch also estimates that about two million advanced chips reached China in those two years through schemes designed to hide their final destination. Those are estimates, not audited counts, so treat the exact number with care.
Then there is the US side of the story. In February, a senior Trump administration official told Reuters that DeepSeek's latest model was trained on Blackwell chips at a facility in Inner Mongolia. DeepSeek has reportedly denied using Blackwell, and has pointed to older H800 chips and Huawei's Ascend processors. White House technology adviser Michael Kratsios has also said Moonshot used export-banned chips, citing US intelligence. So the new analysis lines up with what Washington has been saying for months. It does not settle the argument about how the chips got there.
Banned chips keep showing up in China's own paperwork
Separate reporting from Reuters this week adds weight. One review found more than 23 billion yuan, around $3.2 billion, of Chinese AI computing projects and equipment deals that explicitly name restricted H100, H200 and H800 processors. Another Reuters exclusive found that a Chinese university with deep ties to the military bought H200 chips last month despite being on a US blacklist.
The H200 is allowed to be exported under the current Trump administration policy, but only through a route that lets the US government take a 25% cut. Blackwell is not allowed at all. Meanwhile, Reuters has reported that roughly 600,000 H200 chips are piling up in Hong Kong, waiting for China to approve imports. Nobody I have read can fully explain why Beijing is holding them back.
Nvidia is caught in the middle, and Jensen Huang is gone
The timing could hardly be worse for the company. Nvidia announced on November 12 that Jensen Huang had resigned as president, CEO and board member, with Colette Kress named interim CEO. This followed weeks of reporting about his ties to Alice Huang, an executive at Megaspeed, a Singapore-based company that is one of Nvidia's biggest customers in Southeast Asia. US officials have reportedly described her as a Chinese intelligence asset. Megaspeed denies wrongdoing, and Nvidia has said it found no evidence of chip smuggling.
On Friday, Bloomberg reported that a company with links to Alice Huang was involved in smuggling banned Nvidia chips to a data center in Shanghai. These are allegations drawn from news reports and officials, and none has been proven in court. But the pressure is no longer only on Nvidia. Commerce Secretary Howard Lutnick is also reported to be under scrutiny over whether his department missed suspicious buyers.
My take
When I read this story, the thing that stuck with me was not the smuggling. It was the gap in the rules. Washington restricts who owns chips and where they sit, yet a Chinese lab can apparently rent the same silicon one country over and train a frontier model. If that is legal, then the ban has been protecting less than people assumed. If it is not, then enforcement has been far too slow.
I would also caution readers against the opposite mistake. This is one researcher's analysis. It is detailed and testable, which I like, but it needs independent replication before anyone treats it as final. Expect the labs, Nvidia and regulators to push back hard.
Harsh Choudhury, Founder & CEO, Newszest.in
What to watch next
Three things will tell us where this goes. First, whether other researchers can reproduce DePue's results. Second, whether the Justice Department and Congress widen their probes into Megaspeed and the Shanghai data center. Third, whether Washington moves to regulate who can use chips and not just who can buy them. Nvidia has also talked about adding location-tracking features to its chips to make smuggling harder, though that is not yet a finished fix.
Bottom line
If DePue is right, the most capable open models coming out of China were built, at least in part, on the very chips Washington tried to keep away from them. That does not prove smuggling, and it does not prove the models depend on Blackwell to stay competitive. It does show that the export wall has more doors in it than policymakers admitted. This story is still moving, and I will keep updating it as new details land. As always, do your own research.
Frequently asked questions
Were China's top AI models really trained on banned Nvidia chips?
A forensic analysis by Will DePue, published on November 12, 2026, says six open-weight Chinese models carry fingerprints of Nvidia Blackwell chips. It is strong technical evidence, but it does not prove a law was broken, because the chips could have been in legal offshore data centers.
Which Chinese AI models are named in the analysis?
DeepSeek-V4-Flash, DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, MiniMax-M3 and Ant Group's Ling-3.1-flash.
Why are Nvidia Blackwell chips banned in China?
The US restricts exports of its most advanced AI chips to China because they can speed up military and intelligence uses of AI. Blackwell is barred, while the older H200 can be exported under conditions that include a 25% cut for the US government.
Is it illegal for Chinese companies to rent banned chips abroad?
Not necessarily. Current rules focus on who the chips are shipped to and where they are installed, not always on who rents the computing power. That gap is exactly what US China hawks want closed.
What does this mean for Nvidia and Jensen Huang?
Jensen Huang resigned as CEO and board member on November 12, 2026, and Colette Kress is interim CEO. Nvidia faces scrutiny over its ties to Megaspeed, a Singapore-based customer that denies wrongdoing.
