Nvidia Blackwell chips found in Chinese AI models, study says
A new study says DeepSeek's V4 and Moonshot's Kimi K3 carry numerical fingerprints of banned Nvidia Blackwell chips, hidden in the models' own published weights.

Chinese AI labs can't buy Nvidia's Blackwell chips, yet several of their newest models appear to have been trained on them. That is the claim of a new study that read the evidence straight out of the models themselves.
The report, published on Thursday, November 12, comes from Will DePue, a former OpenAI researcher who leads the nonprofit Unemployment Labs. He analysed the weights of open models from DeepSeek, Moonshot AI, MiniMax and Ant Group. Weights are the billions of adjustable numbers that decide how a model answers questions, and they can carry traces of the hardware used in training. Six models released since April 2026 show patterns that match Blackwell, while earlier models from the same companies match the older Hopper generation, which China is allowed to buy.
How did researchers find Blackwell chips in the models?
The team recovered traces of the random starting values a model has before training begins. Those values come from an algorithm called Philox, and they depend on the chip's configuration. By searching for setups that reproduce them, the researchers matched DeepSeek-V4-Flash and MiniMax-M3 to B200 or B300 chips, and DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro and Ling-3.1-flash to GB200 or GB300 chips. For some models they confirmed the match on real Blackwell hardware. They also tested 24 models whose makers had disclosed their hardware, and the method agreed every time.
The report estimates Kimi K3 and DeepSeek-V4-Pro each needed thousands of Blackwell GPUs for two to three months. It puts the final training runs at roughly $200 million to $300 million for some clusters, using about $70,000 per GPU.
Does this prove US export rules were broken?
No. The findings don't show the models were trained only on Nvidia chips, or that any law was broken. Some Chinese firms rent computing remotely from data centres in countries such as Malaysia, where Blackwell exports are permitted. Models from Alibaba, Zhipu, Tencent, Xiaomi and StepFun read as Hopper-trained, and Meituan's LongCat-2.0 used AI ASIC superpods.
The timing is awkward for Nvidia. Jensen Huang stepped down last week after reports of a relationship with a former Megaspeed executive whom US authorities suspect of spying for Beijing, and lawmakers are again pushing for tighter limits on chip sales to China.
Apps worth a look
Tap the right tiles, add them up, hit the target. Easy to learn, hard to master. Free, for all ages.
Talk to an AI conversation partner; every sentence comes back corrected and explained in your language. 9 languages, A1 to C2.
Open EDIFACT and X12 messages, see every segment explained, convert to JSON/XML, export PDF and Excel. Runs locally, no upload.
Pick who it is for, the occasion and their personality. The search engine reads product descriptions from local shops and ranks what fits. Free, no sign-up.
More from this day
See all →
Comet 220P/McNaught suddenly got 1,000 times brighter. Here's why

Coral Hart used AI to write 370 books in 20 months

WhatsApp Zaps: self-deleting photos and videos for any chat




