Top Chinese AI labs built their newest models on Nvidia's Blackwell chips (twitter.com)

A recent forensic analysis has revealed that four major Chinese AI labs—DeepSeek, Moonshot, MiniMax, and Ant Group—have utilized NVIDIA’s advanced Blackwell chips, which are currently prohibited from export to China, for training their latest AI models. This discovery, based on unique hardware signatures found in model weights, indicates a significant breach of export controls, as the analysis involved matching over six trillion configurations of chip layouts and seed numbers to publicly available AI models. The forensic examination confirmed that several models, including DeepSeek-V4-Pro and Kimi K3, were initialized with configurations exclusive to Blackwell GPUs, supporting correlation values that suggest intentional mapping rather than coincidence. The implications of this finding are substantial for the AI/ML community and international trade laws. With individual model training runs reportedly consuming thousands of Blackwell GPUs, the estimated expenditure exceeds $500 million, indicating a substantial investment in restricted technology. This incident not only raises concerns about compliance with export regulations but also highlights the lengths to which entities may go to circumvent these controls. It casts light on the competitive landscape in the AI sector, where access to cutting-edge technology can dramatically influence capabilities and advancements. This situation is poised to spark discussions about regulatory enforcement and the geopolitical dimensions of AI development.