China touts AI milestone: models generate more tokens than used in training

According to a government-backed report released by the China Centre for Information Industry Development on Saturday, the amount of AI inference data generated in China has now surpassed the data used to train those very systems. The report revealed that daily token calls in China averaged 175 trillion, a staggering increase of over 1,000 times from early 2024.

Inference, the process of running trained AI models to respond to user requests, is heavily reliant on tokens, with inference usage providing a good indicator of the growing demand for AI services. While the report did not disclose the exact amount of data used for AI training in China, the China Academy of Information and Communications Technology estimated the average daily token consumption, including private cloud deployments, to be around 500 trillion tokens as of June.

The National Data Administration in China predicts that the demand for computing power to provide AI inference could reach three times that required for training models in the future. The Chinese government has been placing significant emphasis on the token economy, with local governments like Beijing announcing policies to boost token consumption and development through token factories and distribution platforms.

This milestone was included in a report by Beijing-based think tank CCID, which listed 10 key trends in "new domains and new-quality capabilities", highlighting the growing importance of sectors like AI, quantum computing, and aerospace technology for China's great power competition. Chinese technology companies have been investing heavily in AI infrastructure to keep up with the demand.

For instance, the capital expenditure on AI infrastructure among 13 major Chinese firms is expected to more than double to 932 billion yuan in 2026 and exceed 1.2 trillion yuan by 2027. Companies like Alibaba have been working on making inference more cost-efficient through innovations like DeepInfer, an open-source framework designed to run AI models more efficiently and reduce computing costs.

However, the surge in AI demand has also brought financial pressures as companies spend heavily on the necessary infrastructure.