Demand driven by running models instead of training them shifts spending toward inference chips, data centers and token-priced services, where Chinese brokers already report rising compute rental prices.
Running AI models in China now consumes more data than training them, according to a report from the China Center for Information Industry Development (CCID), a research institute under the Ministry of Industry and Information Technology (MIIT). The report, Ten Trends in New-Domain, New-Quality Innovation and Development (《新域新质创新发展十大趋势》), was released on October 31, according to an IT Home report citing CCTV News and republished by Sina Tech.
China’s average daily token calls exceeded 175 trillion as of June 2026, more than 1,000 times the level at the start of 2024, according to CCID. Tokens are the small units of text that AI models read and write, and Chinese officials now use them as the standard measure of how much AI is being used.
The report frames the crossover as a change of phase: Chinese AI has moved from building models in the lab to deploying them across industry. It also says the release cycle for top models has shrunk from months to weeks or even days.
Daily token calls rose by a quarter in three months
The new figure extends a run of official numbers. In March, Liu Liehong, director of the National Data Administration, said average daily token usage had passed 140 trillion, up 40% from 100 trillion at the end of 2025, Yicai reported at the time. CCID’s 175 trillion for June implies growth of roughly a quarter in the three months that followed.
The CCID report placed the token numbers alongside deployment figures. China has built more than 500 "excellence-level" smart factories covering all 31 provincial-level regions and more than 80% of manufacturing industry categories, and domestic computing power and models have been applied in more than 1,000 scenarios across 20 industries, including steel and autos, according to the report.
The usage figures sit inside a wider push on data supply. By the end of 2025, China had built more than 100,000 high-quality datasets totaling more than 890 petabytes, Liu said in March, according to Yicai, and the National Data Administration has designated 2026 the "Year of Data Element Value Release."
Separate data from the China Internet Network Information Center put the number of generative AI users in China at more than 700 million by the end of the first half of 2026, a penetration rate above 50%, in a report released at the China Internet Basic Resources Conference in September. EastFrontier covered the earlier phase of this growth in July, when daily token calls reached hundreds of trillions.
Inference demand is tightening compute supply
The shift matters for where money flows. Training a model is a large but occasional expense; inference, the work of answering each user request, scales with every call and runs continuously. When inference dominates, demand moves toward chips and data centers built to serve requests cheaply and quickly.
Brokers are already linking the report to compute scarcity. Zhongyuan Securities said agents that carry out tasks on their own sharply increase how often models are called, leaving computing power in short supply and pushing up rental prices for AI accelerator cards, according to Jiemian. It added that domestic compute is tighter than overseas because of chip shortages.
Markets reacted on the first trading day after the report. The STAR 50 index, which tracks the largest companies on Shanghai’s technology board, was up 4.01% as of 2:06 pm on November 2, Jiemian reported, with Roborock up 18.28% and image-sensor maker GalaxyCore up 13.84%.
The method behind the crossover is not public
The reports do not explain how CCID measured "data volume" for inference and training, or whether the token count covers private deployments inside companies as well as public services. Token counts measure activity; none of the reports ties them to revenue for model developers or cloud providers.
The figures to watch are the National Data Administration’s next update on daily token calls and quarterly results from listed cloud and compute providers. If token growth keeps compounding while card rental prices rise, the pressure will fall on suppliers of domestic inference chips to close the gap.