Top Chinese AI labs trained on banned Nvidia Blackwell chips, researcher finds

Top Chinese AI Labs Trained on Banned Nvidia Blackwell Chips

Researchers have found evidence that leading Chinese AI laboratories trained their models on Nvidia Blackwell GPUs — chips subject to U.S. export restrictions designed to prevent advanced AI compute from reaching China.

The discovery emerged through analysis of open-source code and training methodologies. Dylan Patel of SemiAnalysis confirmed the finding is straightforward to verify: DeepSeek and other Chinese labs openly source their model details, and examining the code for roughly ten seconds reveals Blackwell fingerprints in the training pipeline.

The significance lies not in sheer compute volume but in efficiency. Chinese open-source models require far fewer Blackwell chips than the massive clusters Western labs deploy. The hosts estimate the total investment at probably in the "billions of dollars" range — substantial in absolute terms but modest relative to Nvidia's total revenue.

The math works because Chinese labs avoid the computational overhead that constrains American counterparts. U.S. companies must reserve GPU capacity for inference to serve live models, cutting into available training compute. Chinese open-source models, by contrast, get served from separate cloud infrastructure, freeing all purchased compute for training. Additionally, distillation techniques and experimental training runs consume far less compute than brute-force scaling.

This efficiency gap, combined with evidence of policy circumvention, sharpens the ongoing tension around Nvidia's open-source AI ecosystem. The Blackwell leak undercuts the geopolitical intent of U.S. export controls while also fueling the broader debate over whether open-source AI democratizes capability or simply exports American competitive advantage to potential rivals.