Researcher finds Nvidia chip fingerprints in six Chinese AI models

Will DePue says Nvidia chips initialized weights in models from DeepSeek, Moonshot AI, MiniMax and Ant Group. The analysis does not establish where the chips were or prove an export-control violation.
In this article
Will DePue of Unemployment Labs says he found Nvidia Blackwell hardware fingerprints in the published weights of six Chinese superintelligence (SI) models. The models come from DeepSeek, Moonshot AI, MiniMax and Ant Group.
Blackwell chips are subject to US restrictions on exports to China. DePue’s method points to the hardware used to initialize the models' weights—their starting numerical parameters—not necessarily the hardware used throughout training. It does not establish the chips' location, ownership or access arrangements, or prove smuggling or a violation of export controls.
Analysis
How the fingerprints work
Model weights are numerical parameters adjusted during training. The method compares published weights with starting patterns generated using a seed—a starting value that reproduces the same numbers—and a particular chip layout. It uses the PyTorch software framework and Nvidia’s cuRAND library.
DePue says unused parts of a model can retain their initial values, while even heavily trained parts can retain a statistical trace. He says he reproduced the starting patterns for DeepSeek-V4-Flash and Kimi K3 exactly on Blackwell chips. In the researchers' comparisons, changing the numbers of compute units and processing threads broke the correlation.
DePue says the method also correctly identified all 24 control models whose hardware was known. An artificially configured AMD MI300X chip could potentially imitate a layout pattern, but he says chip-specific rounding differences ruled out AMD for DeepSeek-V4-Flash.
What the comparison can establish
The method requires models with publicly available weights and depends on the hardware and software used to initialize them. As Tom’s Hardware explains, matching a layout cannot distinguish exact chip variants by characteristics such as memory capacity or bandwidth.
Chinese companies can rent computing capacity outside China, though the analysis does not establish whether that happened for these models. DePue has published verification code and starting values so others can reproduce comparisons. He says senior researchers at OpenAI and Anthropic independently checked and confirmed his findings.
Follow AveNews
Earlier in Technology
CompaniesItaly’s privacy regulator fines Meta €805 million
The authority ordered Meta to turn off Facebook’s gallery suggestions by default within 60 days and seek separate consent, including from users who had already granted gallery access.
3 min
SoftwareUS cybersecurity agency lists exploited UnrealIRCd flaw
The agency added CVE-2026-114713 to its catalog on 13 November and requires federal civilian agencies to apply fixes by 16 November.
2 min
SoftwareWorkday outage blocks employees from pay and benefits tasks
Workday put some systems into maintenance after testing revealed inconsistent information that could cause account and task errors, the company said. The November 13 outage also blocked timekeeping and leave requests.
2 min
Sources
The RegisterChinese AI labs are training models on Nvidia's Blackwell despite export controls, says ex-OpenAI engineer (opens in a new tab)
The DecoderFormer OpenAI researcher says he found Nvidia Blackwell fingerprints hidden inside six Chinese AI models (opens in a new tab)
Tom's HardwareResearchers allege DeepSeek, Kimi, and other Chinese AI models were trained using banned Nvidia Blackwell GPUs (opens in a new tab)
