China's most powerful open AI models were built on hardware the United States bans from export to China. That is the claim of new forensic research reported by TechCrunch on 13 November 2026 — and the evidence is written into the models themselves, in their publicly released weights. Six leading models from DeepSeek, Moonshot AI, MiniMax and Ant Group carry what researchers call the fingerprint of Nvidia's Blackwell GPUs. Here is what was actually found, how the fingerprint works, and where the claim stops short.
What the research found
The analysis comes from Unemployment Labs, a nonprofit research group led by Will DePue, a former OpenAI researcher. DePue examined the released weights of six open-weight models and found patterns consistent with Nvidia's Blackwell chips — the current generation of AI accelerators that the US bans from export to China or to Chinese-owned firms. The six models are:
- DeepSeek-V4-Flash and DeepSeek-V4-Pro (DeepSeek)
- Kimi K3 and Kimi K3.1-Pro (Moonshot AI)
- MiniMax-M3 (MiniMax)
- Ling-3.1-flash (Ant Group)
All six were released since April 2026. Earlier models from the same companies match Nvidia's older Hopper generation instead — the shift to Blackwell fingerprints began only in the last six months, between April and October 2026.
How a chip leaves a fingerprint in a model
When a model is first created, its starting weights are generated on a specific piece of hardware, and the layout of that hardware leaves traces in the numbers. Blackwell chips have distinctive thread counts — 148 streaming multiprocessors (303,104 threads) on the B200/B300, and 152 (311,296 threads) on the GB200/GB300. By searching for the hardware layout that best reproduces those leftover traces, the researchers matched four of the models to Grace Blackwell systems and two to Blackwell chips. The strongest single result: for DeepSeek-V4-Flash, one Blackwell chip regenerated the model's starting weights exactly — in 377 values where Nvidia and AMD results differ, the published weights matched Nvidia in 363 and AMD in none. The team also confirmed the match on physical Blackwell chips for DeepSeek-V4-Flash and Kimi K3.
What it does not prove
This is where careful reading matters. The report itself calls the Blackwell patterns "strongly suggestive" of initialisation on Blackwell — it identifies the hardware used to generate the models' initial weights, and supports, but does not prove, that Blackwell chips also did the training. It says nothing about where the chips were: Blackwell systems can be rented from data centres outside China, and the export rules do not cover offshore use. So the finding is not, on its own, proof that any export-control law was broken. TechCrunch's write-up is firmer than the report on the training claim — the report stops at initialisation.
The scale of the surrounding trade is still striking. Epoch AI, another nonprofit research organisation, estimates that about two million advanced chips entered China in 2024 and 2025 through schemes designed to hide their final destination, including 275,000 Nvidia AI accelerators imported to Malaysia in those two years. Separately, the Israeli security firm Lava found 32 export-controlled Blackwell Ultra GPUs through publicly exposed monitoring endpoints geolocated to China.
Why it lands now
The timing is impossible to ignore. Nvidia CEO Jensen Huang left the company this week after admitting to a relationship with Alice Huang, an executive of Megaspeed — a company implicated in smuggling Nvidia chips into China. TechCrunch reports that Nvidia's potential corporate liability for export-control violations likely factored into the board's decision to part ways with him. White House technology adviser Michael Kratsios said earlier this year that Moonshot had used export-banned Nvidia chips, citing US intelligence sources. Nvidia, DeepSeek, Moonshot, MiniMax and Ant Group had not responded to the findings at the time of publication.
Meanwhile the rules themselves have a loophole built in: the Trump administration allows Nvidia to export the less powerful H200 chip to China, provided shipments take a circuitous route that lets the US claim a 25% tariff. Blackwell, by contrast, is banned outright.
Bottom line
Six of China's flagship AI models carry the technical signature of the very chips Washington tried to keep out of China — the first forensic, weight-based evidence of its kind, built entirely from the models' own public weights rather than leaks. It does not prove a law was broken, and it does not say where the chips sat. But it shows the export-control wall has a large hole in it, and it lands in the same week Nvidia lost the CEO who spent years arguing that wall should come down. Expect this story to get louder: Huang is due to testify before Congress by Thanksgiving.
Sources: TechCrunch, the Unemployment Labs report as described by TechCrunch and secondary coverage. Hero photo: server racks at NERSC, by Derrick Coetzee, CC0 public domain, via Wikimedia Commons.
