On the morning of November 4, in Wuzhen, two blue books sat on a lit table, one about the world and one about China. The publisher was the Chinese Academy of Cyberspace Studies, an institution that rarely trends on social media.
This was the tenth year in a row it had done this, one book a year since 2017. A decade on, the reports read like an almanac of the internet, one page per year, and stacked together they show which way the curve bends.
Two much-quoted numbers this year sit close together, in the same paragraph, separated by a single comma. Yet they describe two entirely different layers of business, forty-odd times apart.
Two Books in the Tenth Year
The World Internet Development Report looks global; the China Internet Development Report looks domestic. The first traces how countries write artificial intelligence into national strategy; the second does the accounting inside the industry. One leans narrative, the other statistical.
The division of labour is clear: one handles trends, one handles numbers. Almost everything pulled out and quoted comes from the second, because it offers scale that can be compared across countries and makes for an easy headline.
The timing is deliberate too. It was the third day of the Wuzhen summit, the panels were mostly done, and the press was still on site. Release numbers then, and the next day's pages are more or less set, with a distribution path already in place.
The Layer of Twenty-odd Billion
The first number: in 2025, the market for enterprise-grade AI agents in China jumped to 21.2 billion yuan. All three qualifiers in that sentence matter. Drop one and the meaning changes, and readers can no longer compare it with anything.
Enterprise-grade means sold to companies, not the thing on your phone. 2025 is a year that has already happened, not a forecast. An agent is the kind that breaks a task into pieces, calls tools and finishes the job.
What is 21.2 billion yuan? Spread across the country, it is less than 20 yuan per person for a year. Within the whole AI industry it is a thin layer, thin enough to be covered by the larger number beside it.
Thin, but not hollow. It reflects companies paying real money: customer service, reporting, code, approval workflows. The payer is clear and the demand is clear. It simply has not reached the stage of wide deployment.
The Layer of Nine Hundred Billion
The second number: in 2025, the market for embodied intelligence in China was about 915 billion yuan, up 20.4 percent year on year. The key phrase in that sentence is year on year, not quarter on quarter, and not cumulative.
Embodied intelligence sounds awkward, but unpacked it means intelligence with a body: robots that walk, arms that grasp, cars that drive themselves. The emphasis is on doing, not only on talking.
A market of more than 900 billion yuan contains far more than algorithms. It contains motors, reducers, sensors, batteries, complete machines, production lines, plus installation and after-sales. Hardware is inherently heavier than software, with longer chains, slower payment and real inventory risk.
The 20.4 percent growth rate is one of the few figures in the report that can be compared directly. It shows the market is still growing, but not wildly. For an asset-heavy industry, that pace is already considerable.
More Than a Multiple Apart
Put the two together and they differ by more than forty times. Many readers take the first glance to mean embodied intelligence is forty times hotter than agents. That reading is wrong, and it is the kind of error that gets forwarded and starts to sound true.
They do not measure the same thing. One measures what companies pay for software; the other measures output across an entire hardware supply chain. The first is a procurement budget, the second is industrial output, with several layers in between.
An analogy: a restaurant pays 2,000 yuan a year for software subscriptions and 90,000 yuan for its fridges and stoves. You cannot conclude that the restaurant cares more about fridges, only that most of its money sits in equipment.
What can genuinely be read from the pair is this: agents still sit high in the stack, light, thin and close to money, but limited in scale. Embodied intelligence sits below, heavy, thick and long-chained, and any movement there is large.
That also explains why model companies keep looking toward hardware. The ceiling on the software layer is hard to push through by selling interfaces alone. You have to go down, to where things can be installed, powered and touched.
Looked at another way, the two numbers mark two ways of making a living. One relies on people writing rules, wiring interfaces and running processes; the other on factories, supply chains and on-site installation. The barriers, rhythms and margins all differ.
The Hand Reaching Out in a Pharmacy
At the same expo, Ant's LingBot model demonstrated a scene: in a mock retail pharmacy, a robotic arm took a box of medicine from a shelf, glanced at the packaging and put it in a bag, not quickly.
Boxes vary in size, packaging reflects light, and the arrangement is not tidy. Trivial for a person, hard for a machine: it must first recognise what the object is, then decide the angle, the force, and which slot to place it in.
According to Xinhua, the system is already in use in several pharmacies in Shanghai's Pudong district. The report gives no specific number of stores. What can be confirmed is that the scene has moved from a booth to a counter, from demonstration to shift work.
The scene connects the two layers neatly: above, a model recognising and deciding; below, the arm, the sensors and the whole supply chain. Remove either end and nothing lands, nothing earns, and there is no repeat business.
The Rulers Behind the Numbers
The report contains another set of figures: by June 2026, 924 generative AI services had completed filing and 573 applications or features had completed registration. The two counts are not the same thing.
Another set covers usage: in March 2026, average daily token calls in China rose to 140 trillion, more than a thousandfold increase in two years. The figure is written into the blue book, with both definition and date spelled out.
Filings, call volume and market size are three different rulers. Filings measure compliant supply, calls measure intensity of use, market size measures money. Mix them and you get strange conclusions.
Only by reading all three together can you avoid mistaking an industry's temperature for the mercury in one particular thermometer. That is the first rule for reports like this: check the definition, then the size, and only then the ranking.
Back to those two numbers. They are not there to be compared but to be layered. See clearly which layer is thin and which is thick, and you know where the business actually grows, where the money moves, and who is really paying.
Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.