China Says Its AI Now Answers More Than It Learns
China's artificial intelligence systems now generate more inference data than was used to train them, according to a government backed report that treats the crossover as a coming of age for the country's AI industry.
The report came from the China Centre for Information Industry Development, which sits under the Ministry of Industry and Information Technology, and it cites official figures showing average daily token calls above 175 trillion. That is more than 1,000 times the volume recorded in early 2024.
But the token is the unit that matters here. Every prompt, document and image a model processes is broken into tokens, so a token call is one measure of how much work an AI system is actually doing for paying users rather than how much it might do.
But the report does not say how much data China used to train its models, which means the headline comparison cannot be checked from the document itself. The inference figure is a published number and the training figure is not.
The second half of the picture comes from a different body. The China Academy of Information and Communications Technology, another think tank under the same ministry, put average daily token consumption at about 500 trillion by mid year once private cloud deployments are counted, which is nearly three times the public cloud figure in the newer report.
And the crossover has become policy language. Officials have started using the phrase token economy, and Beijing's municipal government has announced plans to boost token consumption through token factories and token distribution platforms, which is an unusual thing for a government to industrialise.
The mechanism behind the surge is ordinary adoption rather than a single breakthrough. Chat assistants, coding tools, customer service bots and document analysis all bill by consumption, so the more companies embed them the more tokens move, and inference is the cost line that grows with use rather than the one off expense of training a model.
So the milestone carries a commercial reading as much as a technical one. It suggests Chinese enterprises are now paying to run AI at scale, which is the stage where revenue appears and where the bill for compute arrives at the same time.
So the bill is already visible. Rhodium Group projected in a report two months ago that capital spending on AI infrastructure by 13 major Chinese companies would more than double to 932 billion yuan, about USD 139 billion, this year, and exceed 1.2 trillion yuan in 2027. Alibaba, Tencent and Baidu together swung to a combined negative free cash flow of 16 billion yuan in the first half of the year, against a positive 170 billion yuan a year earlier.
Alibaba's own numbers show the demand side of that trade. Revenue from its AI cloud and computing services rose 45 per cent year on year to USD 7.1 billion in the quarter that ended at mid year, and the company set up a unit called Token Foundry in the middle of the year to push consumption internally and commercially.
What is missing is the cost per token. Neither report gives a price, and without it the token count says how much AI is being used but not whether anyone is making money from it. That gap is where the next set of numbers will be argued over.
China's daily token calls now exceed 175 trillion, more than 1,000 times the level in early 2024
A government backed report says inference data now exceeds the data used for training for the first time
The report does not disclose how much training data China uses, so the comparison cannot be verified from it
A separate ministry think tank puts total daily consumption near 500 trillion tokens including private cloud
Rhodium expects AI infrastructure spending by 13 major Chinese firms to double to about USD 139 billion this year
Source: South China Morning Post
