Statistical Analysis Links Chinese AI Models to Banned Nvidia Blackwell Silicon
Independent researchers have found strong statistical evidence that major Chinese AI models were trained using export-restricted Nvidia Blackwell hardware.
Independent research organization UnemploymentLabs has released a statistical analysis indicating that several prominent open-weight Chinese AI models were trained using Nvidia's export-restricted Blackwell GPUs. According to Tom's Hardware, the models implicated in the findings include DeepSeek V4, Kimi K3.1, MiniMax-M3, and Ling-3.1-Flash. TechCrunch reports that the research also encompasses models developed by Ant Group and Moonshot AI, underscoring the widespread accessibility of advanced silicon despite U.S. trade restrictions designed to protect American AI leadership.
The identification method relies on examining the initial weights of large language models, which undergo a randomization step during initialization using the cuRAND software stack. Because this computational workload is distributed across hardware shader modules and processing threads, the resulting weight distributions retain a mathematical fingerprint of the underlying architecture. Nvidia Hopper chips feature 132 shader modules, while Blackwell B-series and GB-series chips feature 148 and 152 modules respectively, alongside a thread count per shader module of 2,048.
By replicating the initialization process using identical hardware configurations and software stacks, the researchers observed that final statistical distributions precisely matched those of the Chinese models. According to Tom's Hardware, altering the shader module and thread counts caused the statistical correlation to break down entirely. The researchers also tested and ruled out alternative hardware, such as AMD accelerators, by analyzing actual weight values.
Despite the precision of the mathematical analysis, the researchers emphasize that their methodology cannot pinpoint the physical location of the hardware or determine how it was accessed. TechCrunch notes that utilizing offshore data centers, such as facilities in Malaysia, has become a popular method for Chinese firms to access advanced compute power without violating domestic territorial restrictions. Epoch AI estimates that hundreds of thousands of advanced Nvidia accelerators entered Malaysia throughout 2024 and 2025.
The findings arrive amid heightened scrutiny over enforcement gaps in U.S. export controls and Nvidia's corporate compliance. According to TechCrunch, White House technology adviser Michael Kratsios previously stated that Moonshot utilized export-banned chips, while independent estimates suggest millions of advanced processors have entered China through alternative routing schemes. The intersection of offshore compute access, smuggling allegations, and software-stack fingerprinting continues to complicate efforts to restrict foreign access to cutting-edge semiconductor hardware.