DeepSeek and five other Chinese AI labs allegedly subscribed in bulk to their American rivals and trained on their outputs, extracting capabilities worth billions since 2024 according to the FBI and the NSA, which may explain their rapid catching up with U.S. models. Facing U.S. export restrictions on high-end Nvidia chips, Chinese labs developed algorithms that drastically reduce the complexity of the attention mechanism, enabling them to achieve better results with fewer resources. According to data from Larridin, Chinese models such as GLM 5.2 and Kimi 2.6 handle roughly 75% of everyday engineering tasks at one-fifth the cost of U.S. models, with Chinese open-source models accounting for 41% of total downloads on Hugging Face in 2024. Ramp reported that the share of businesses using Chinese models in their AI spending rose from 4.5% in January to 6.1% in July, though U.S. models still hold a several-month lead on the most complex tasks.
Source: Read the original article
