Will Depue says forensic analysis of weights found Chinese labs using NVIDIA Blackwell chips
Former OpenAI researcher Will Depue posted on X on November 12 and published a report saying that, through forensic analysis of publicly available model weights, he found that the latest models from DeepSeek, Moonshot AI, MiniMax and Ant Group were initialized on NVIDIA Blackwell chips. Exports of these chips to China are subject to U.S. licensing restrictions.
Depue said that the initial weight layout of the models carries fingerprints related to specific GPUs and random seeds, and that these remain after training. He matched more than six trillion combinations, saying that six models (including DeepSeek-V4-Pro, Kimi K3 and MiniMax-M3) matched Blackwell systems, with four matching GB200/GB300. He said he reproduced some of the initial weights on real hardware (B300 bit-for-bit matched DeepSeek-V4-Flash, and GB200 had a correlation of 0.999995 with Kimi K3), and estimated that large models used 2,000 to 3,000 Blackwell GPUs. The report said the results were independently validated by senior researchers at OpenAI and Anthropic.
Original source:
Title: Top Chinese Frontier AI Models Carry the Fingerprint of Export-Controlled NVIDIA Blackwell GPUs
Publisher: Unemployment Labs (Will DePue)
Publication date: 2026-11-12
URL: unemploymentlabs.com/checkpoint-forensics
Source nature: author report and X post
Verification status: The report summary and the original X post text were actually read in this instance; verification date 2026-11-12
Supporting content: Supports the forensic claim that DeepSeek, Moonshot, MiniMax and Ant Group models match Blackwell chips and the hardware reproduction description
ABAB AI Insight
This method relies on fingerprints in public weights rather than internal information or logistics records. It can indicate the type of hardware used for initialization, but it cannot by itself prove the chips’ physical location, procurement path, or whether export controls were violated.
The report compares recent models with earlier versions said to have used Hopper, and provides estimates of statistical significance and hardware reproduction results. The specific scope and method of the independent validation were not detailed in the public post.
If such forensic analysis holds up, it could provide publicly available material for export-control enforcement, but it would need to be cross-checked against supply-chain, customs and company disclosures. At present, it is only the researchers’ claim, and no formal response from the relevant regulators or companies has been seen.
ABAB News · Cognitive Law
Public model weights can retain initialization fingerprints related to training hardware, enabling retrospective forensic matching.
Forensic results point to the type of hardware and do not automatically amount to complete proof of the chips’ physical location or of export-control violations.