Will DePue, a former OpenAI engineer who now runs Unemployment Labs, published a report on 12 November 2026 arguing that six open-weight Chinese AI models from DeepSeek, Moonshot AI, MiniMax and Ant Group were initialized on Nvidia Blackwell GPUs, a chip class that U.S. export controls bar from China. The finding is the author’s own analysis and has not been independently verified, and the report itself says it does not show that any export rule was broken. TechCrunch and The Register covered it on 13 November 2026.
What the report claims
- The report, “Top Chinese Frontier AI Models Carry the Fingerprint of Export-Controlled NVIDIA Blackwell GPUs”, was published on 12 November 2026 and revised the same day. It analyses published model weights, not leaks or insider accounts.
- It reports Blackwell-class matches for six models released since April 2026: DeepSeek-V4-Flash and MiniMax-M3 (matching the 148-multiprocessor B200/B300 layout), and DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro and Ant Group’s Ling-3.1-flash (matching the 152-multiprocessor GB200/GB300 layout).
- Earlier models from the same labs, including DeepSeek-V3 and the Kimi K2 series, match Nvidia’s older Hopper generation, which the report says cannot separate H800, H100 and H200 chips.
- The author says the evidence is conclusive for DeepSeek-V4-Flash and strong for the other five.
How the weight-fingerprinting method works
Training starts from random initial weights. According to the report, these are generated by the Philox algorithm through PyTorch and Nvidia’s cuRAND library, and the output depends on a seed and on the number of streaming multiprocessors on the GPU, which sets the thread layout. The author searched roughly six trillion seed-and-layout combinations for ones that reproduce the initial values still visible in the finished weights, checked matches on weights held out of the search, and says he reproduced the patterns on physical H100, B300, GB200 and MI300X hardware. The report publishes seed values so other researchers can attempt to reproduce the results.
What the report does not show
- It does not establish where the chips were located, how the labs obtained access, that Blackwell did all of the training, or that any export control was violated. The report says its export-status flags describe hardware classes and are not a finding that any law was broken.
- It identifies the hardware that generated the initial weights, which is not necessarily all the hardware used in training. Only models with published weights and PyTorch-style initialization can be tested.
- Its GPU-count estimates (about 12,000 Blackwell GPUs across the five final training runs, with a stated range of 7,000 to 22,000) rest on assumptions about efficiency and run length, and the author says he is not confident in them.
- Some models show no chip signature at all, including Meituan’s LongCat-2.0 and DeepSeek-V4.1-Flash.
Coverage and reactions
According to TechCrunch, the findings do not mean the models were trained entirely on Nvidia hardware or that export law was necessarily broken, because the chips could have been in an offshore data centre. TechCrunch also reports that Blackwell GPUs are banned from export to China or Chinese-owned firms, while the administration allows exports of the less powerful H200 under conditions. The Register reports that DePue says senior researchers confirmed the result, which the outlet did not independently verify. Neither outlet reported a response from Nvidia or the named labs at the time of writing.
The finding arrives amid several separate threads on chip diversion and remote access that CASRAI has covered, including reports of Moonshot AI seeking Blackwell access in Europe and the Commerce rule on third-country data centres. This post concerns only the weight-analysis report.
What it means for researchers and research offices
- Treat it as an allegation with a reproducible method. The report is self-published and not peer reviewed. Its seeds are published, so independent replication is possible, and conclusions should wait for it.
- Remote compute is the live compliance question. Both outlets note that location and access are not settled by the weights. Offices that broker or buy cloud GPU time, or run collaborations with Chinese labs, should check that their export-control and research-security reviews cover access to controlled hardware, not only its physical shipment.
- Provenance of open weights. Institutions adopting open-weight models from the named labs should follow the official response before drawing compliance conclusions from this report alone.
- Method transfer. If the technique holds up, initialization fingerprints could become a tool for auditing model provenance and reproducibility claims, which is a methods question for the AI research community as well as for regulators.
