Four Chinese AI labs used restricted Nvidia Blackwell chips in 2026

Will DePue said Nov. 13 that forensic weight signatures indicate four leading Chinese artificial intelligence developers trained six frontier models on Nvidia's export-controlled Blackwell processors.

The finding covers six open-weight releases from DeepSeek, Moonshot AI, MiniMax and Ant Group. Those four firms account for roughly 70% of China's open-source frontier model downloads this year.

The technical analysis rests on more than 6 trillion weight initialization patterns. It reveals that the Chinese labs abandoned Nvidia's older Hopper architecture in favor of Blackwell systems during model initialization throughout 2026.

DePue, a former OpenAI researcher, established the non-profit research consultancy Unemployment Labs earlier this year. His team evaluated the mathematical values assigned to neural networks before active training cycles commence.

Standard machine learning frameworks such as PyTorch distribute initial pseudorandom values across parallel hardware cores using the Philox algorithm. The algorithm relies on a fixed seed to populate empty weight matrices.

A GPU splits that workload across physical streaming multiprocessors — the distinct hardware blocks that run parallel calculations. A full initialization sequence requires 2,048 compute threads per multiprocessor block.

Different GPU architectures feature fixed multiprocessor layouts. That structural difference alters where identical pseudorandom numbers land across a tensor grid, the multidimensional array that holds a model's operational weights.

Nvidia's older Hopper H100 and H800 accelerators use 132 multiprocessors with 270,336 active threads. The newer B200 and B300 processors feature 148 multiprocessors running 303,104 threads, the report said.

The Grace Blackwell GB200 and GB300 systems contain 152 multiprocessor blocks supporting 311,296 threads. Those exact thread counts leave an indelible digital impression inside the resulting model files, DePue said.

The study examined six recent model checkpoints released since April 2026. Four systems matched the 152-block Grace Blackwell design: DeepSeek-V4-Pro, Moonshot AI's Kimi K3, Kimi K3.1-Pro and Ant Group's Ling-3.1-flash.

The remaining two systems, DeepSeek-V4-Flash and MiniMax-M3, aligned with the 148-block B200 and B300 hardware layout. Both models showed deterministic thread placements unique to the standalone Blackwell architecture.

Unemployment Labs confirmed the match on physical Blackwell hardware for DeepSeek-V4-Flash. The model matched 445 out of 469 test rows against a physical B300 chip, achieving a 94.9% layout alignment.

Physical testing against Nvidia's Hopper H100 and AMD's Instinct MI300X produced zero row matches. That divergence confirmed the baseline mathematical distribution originated specifically on Blackwell silicon, the researchers said.

Successive model releases trace when Chinese labs transitioned their compute environments. DeepSeek's V3 model showed an older Hopper signature when the Hangzhou-based laboratory published it in December 2024.

DeepSeek's subsequent V4 models switched entirely to Blackwell layouts by April 2026. Moonshot AI's Kimi K2.5 registered Hopper traces in January 2026 before its K3 release adopted Grace Blackwell in July 2026.

MiniMax shifted from Hopper on its M2.7 release in April 2026 to Blackwell on M3 in June 2026. Ant Group followed an identical trajectory, migrating from Hopper in August 2026 to Grace Blackwell in October 2026.

The technical trail does not appear across all Chinese developers. Models published by Alibaba Group Holding, Tencent Holdings, Xiaomi and Zhipu AI continue to reflect Hopper hardware configurations.

Nvidia's commercial position faces mounting scrutiny following the revelation. The US Department of Commerce banned shipments of advanced Blackwell GPUs to China and Chinese-owned corporate entities in 2024 to curb military computing parity.

The technical disclosure follows the abrupt resignation of Nvidia Chief Executive Officer Jensen Huang on Nov. 11. Huang stepped down after federal inquiries targeted third-party distribution channels in Southeast Asia.

Those inquiries focus on Singapore-based Megaspeed International, a server vendor that reportedly leased high-performance compute clusters to Chinese clients. Chinese entities operated approximately $3.4 billion in computing projects using restricted hardware this year.

Chinese artificial intelligence laboratories have disputed allegations of illicit hardware acquisition. Moonshot AI stated publicly that its model training relies entirely on compliant infrastructure procured through authorized international channels.

The forensic report does not settle where the Blackwell hardware physically operated. The digital fingerprints verify the compute architecture used at initialization, but do not prove physical smuggling into mainland China.

The computational footprints remain fully consistent with Chinese developers renting Blackwell capacity remotely. Offshore cloud facilities operating across Singapore and Malaysia offer compute clusters without physical hardware transfer, the report said.

The Bureau of Industry and Security has not confirmed whether offshore remote leasing constitutes an actionable export violation under current rules. Existing export regulations focus primarily on physical commodity transfers across customs borders.

The US Senate plans to review offshore compute enforcement on Dec. 3 during hearings on the proposed AI Overwatch Act. The measure aims to prohibit foreign access to American advanced computing assets worldwide.