Report: China's AI Is Now Answering More Than It Learns

China's daily AI inference activity has overtaken training, according to an industry estimate reported by Pandaily, pointing to growing demand for computing power to serve users. ^[1]

CCID, a Chinese industry body, estimated that the crossover happened in June, Pandaily said. It put daily inference activity above 175 trillion token calls, more than 1,000 times the level in early 2024. ^[1]

Inference is the work a trained model does when it responds to a prompt. Training is the process of adjusting the model's internal settings using data so that it learns patterns. The reported crossover suggests that using models is becoming a larger activity than building them. ^[1]

That would make the economics of answering questions increasingly important. A model needs resources to be built, but it also needs computing power every time someone uses it. At sufficient scale, that repeated work can become an enormous demand of its own.

Counting pieces of text

The reported 175 trillion figure counts "token calls". ^[1] Tokens are the small pieces of text a model processes, and one question and its answer can involve many of them.

The size of the estimate is striking. The reported increase of more than 1,000 times means activity has risen by more than three orders of magnitude since early 2024. ^[1] But a count of text activity cannot, by itself, tell us how many chips were occupied or how much electricity they consumed.

CCID's counting method, coverage and basis for comparing inference with training have not been established. The estimate has not been independently verified, and the date and year of its release are also unconfirmed. No direct CCID statement or company response is available to establish whether spending or chip deployments have changed. ^[1]

The recurring computing bill

A successful model brings a recurring computing bill alongside the cost of developing it.

If the reported growth reflects a sustained rise in use, providers would have a stronger reason to focus on the cost and speed of serving answers. Computing capacity matters, as do the chips doing the calculations and the memory supplying the data they need. The practical challenge is to keep answering requests without making each answer too slow or too expensive.

That is why the estimate matters beyond the headline number. Growth in inference could change which computing investments pay off. Building a more capable model and running a popular service are related businesses, but they place different demands on the systems behind them.

The next test is whether CCID publishes a reproducible account of the activity it counted, and whether Chinese cloud and AI companies disclose investment decisions that show how rising use is changing their computing needs.