Guest mode: you can read and try the quiz, but your answers will not be recorded.
Have you noticed how many more people around you seem to be using AI lately? Chatting, writing, drawing — even machines in factories are using it. When a tool moves from the laboratory to everyday life, does the way it uses data change too?
The China Centre for Information Industry Development (CCID) released a report, "Ten Major Trends in Innovation and Development of New Domains and New Quality Productive Forces", describing the country's AI data-use structure with statistics. They compared two kinds of data: training data (used to "teach" a model) and inference data (processed when a model actually answers questions or performs tasks). The report states that by June 2026, daily token calls exceeded 175 trillion, more than 1,000 times the level at the start of 2024.
According to the report, this is the first time inference data volume has exceeded training data volume in China. Put simply, most computing power used to go into "feeding the model" — repeated training and improvement; now more goes into "using the model" — answering questions and carrying out tasks. The report describes this turning point as a move from the model-training research stage into large-scale industrial application. At the same time, the update cycle of top models has shrunk from months to weeks or even days.
The report lists several signs of uptake: more than 500 excellence-level smart factories have been built nationwide, covering 31 provinces, regions and municipalities and more than 80% of major manufacturing categories; domestic computing power and large models have been applied in over 1,000 scenarios across 20 industries including steel and automobiles. By the first half of 2026, generative AI users in China exceeded 700 million, with penetration above 50%.
The report does not explain the specific statistical method or definition behind "inference data exceeding training data", nor does it provide the full report or a peer-reviewed source. These figures are therefore best read as one institute's description, not an independently verified conclusion.
Training and inference are two stages in a model's life cycle. Training uses large amounts of data to adjust a model's internal parameters so it learns patterns; inference is when a trained model receives new input and produces an answer. The same piece of text is "teaching material" during training and a "question" during inference. The two data volumes can therefore be counted separately, each reflecting where computing power goes at a different stage.
"Inference data exceeding training data" only shows a shift in how data is used; it does not mean model capability or accuracy has surpassed the training stage.
The report notes that top models' update cycles have shrunk from months to weeks or even days.
According to the CCID report, China's AI inference data volume has exceeded training data volume for the first time, described as a move from the model-training research stage into large-scale industrial application. The report also notes that more than 500 excellence-level smart factories have been built across 31 provinces, regions and municipalities, and that domestic computing power and large models have been applied in over 1,000 scenarios across 20 industries including steel and automobiles.
This article was researched and written from the following materials:
Primary research: 中國電子信息產業發展研究院(賽迪研究院,CCID)
News report: IT之家1 November 2026Read the original
AI editorial note: this article was rewritten by AI from the materials above and fact-checked. It is not a reproduction of the source article.
This item was auto-summarised and quizzed by AI from whitelisted science news sources. For learning only — use "Report a problem" below if anything looks wrong.