Models

Former OpenAI Researcher Claims Evidence of Sanctioned Nvidia Chips in Chinese AI Models

November 16, 2026 · 2 min read

Also published in Polski

blue waves digital wallpaper
Denis Sebastian Tamas / Unsplash

Will DePue, a former researcher at OpenAI, claims to have found technical evidence that leading Chinese AI laboratories have accessed the latest Blackwell hardware from Nvidia despite US export controls.

According to a forensic analysis by DePue, the companies DeepSeek, Moonshot AI, MiniMax, and Ant Group are using export-controlled Blackwell GPUs. The researcher states that six Chinese models released since April 2026 carry the distinct signature of the Blackwell generation. In contrast, previous models such as DeepSeek-V3 or Kimi K2.5 showed traces of the older Hopper chip generation.

The analysis is based on model weights. DePue explains that PyTorch software fills empty spaces with random numbers using a seed before training begins. Because different chip models have different numbers of cores, a characteristic pattern is created during this process. Hopper chips use 132 cores, while Blackwell chips use either 148 or 152 cores. DePue claims to have reverse-calculated this pattern and replicated the results for Kimi K3 and DeepSeek-V4-Flash on actual Blackwell chips.

Specifically, the analysis assigns B200 or B300 chips with 148 cores to MiniMax-M3 and DeepSeek-V4-Flash. The 152-core pattern of GB200 liquid-cooled flagship rack systems was found in Kimi K3, Kimi K3.1-Pro, DeepSeek-V4-Pro, and Ling-3.1-flash from Ant Group. DePue notes that the method was tested on 24 control models with known hardware, including Meta's Llama 3.1, and was correct in every case.

DePue estimates that five investigated training runs required approximately 12,000 Blackwell GPUs, which equals the computing power of about 30,000 H100 chips. He estimates that Kimi K3 and DeepSeek-V4-Pro required roughly 4,500 and 3,000 Blackwell processors respectively over about 70 days. With a unit price of around 70,000 US dollars, these individual clusters would cost between 200 and 300 million US dollars, with all five systems totaling over 500,000,000 US dollars.

While the report provides technical evidence of access to the hardware, DePue states it does not prove physical smuggling or a breach of US export laws, as Chinese firms can rent computing power in data centers outside China. He also noted that while an AMD MI300X chip could potentially be throttled to create a similar pattern, he believes he excluded AMD for DeepSeek-V4-Flash based on rounding differences.