On September 12 state broadcaster CCTV carried a report from the China Telecom Research Institute arguing that China’s AI industry has stopped competing on large models and raw compute, and started deploying and selling agents. Bloomberg picked up its main findings.
The institute is the research arm of the state-owned carrier China Telecom. That parentage matters. This is not a brokerage forecast written to move a stock. It is a call from a company that has to rack, power and procure against its own number.
The number is 80%. By 2029, the report says, inference will account for 80% of China’s computing-power market, overtaking demand from training. It also expects agents to drive close to tenfold annual growth in the country’s computing demand over the next two to three years.
Training is a purchase. Inference is a meter.
The distinction is not technical. It is an accounting line.
Training builds a model once, and lands in capital expenditure. Inference happens every time somebody uses one, and lands in operating expense. When a national compute market flips from training-led to inference-led, the money has stopped going into building the thing and started going into keeping it running.
What runs continuously is agents. What agents run against is workflows: tickets, reconciliations, reviews, first-pass screening, on-call rotations, first drafts.
The report also sizes the spend. Chinese technology companies are on course to put close to 600 billion yuan, about $89B, into AI this year, which it puts at more than a tenth of all investment in the country. That ratio is the sentence underneath the forecast. AI is no longer a sector capex programme in China. It is a tenth of national investment.
Why a compute-mix forecast is a white-collar forecast
“Inference reaches 80%” sounds like a data-centre story. It is a payroll story wearing a data-centre costume.
A company buying training capacity is buying research capability, which has close to no relationship with its headcount plan. A company buying inference capacity is buying the unit cost of having machines do specific work. That bill arrives monthly, and its denominator is tasks, not people.
This year the same shift already showed up in how software is priced. We covered Salesforce moving pricing off the seat on September 9: the old contract sold one login per person, the new one sells work completed. The China Telecom report is the supply side of that same trade. When 80% of a compute market serves per-task inference, buyers have already stopped purchasing per head.
Two other pieces sit alongside it. Zscaler packaged tier-one security triage as a product, which is a whole job function shipped as a service. Cognition’s run rate nearly doubled inside a year on a story about engineers moving into the overseer seat. Both of those burn inference, and both burn it continuously.
Europe is building for the same demand on a slower clock
The report has a natural control group.
The European Commission opened bidding in July for up to seven AI gigafactories, a €30B programme with roughly €10B of public money and €20B hoped for from private investors. About €1B of that is actually committed. On the report’s own numbers, Chinese technology companies spend that much on AI roughly every five days.
The schedule: applications close on 12 November, awards are expected in early 2027, construction starts that year, and the machines are due to run by mid-2028. On paper that lands a year ahead of China’s crossover point, provided nothing else slips.
Things have already slipped. Bidding moved from May to July. The evaluation criteria were delayed more than once. Interest fell from about 70 interested companies to roughly ten expected bidders. And the money is the harder problem: most of the public half depends on a 2028-2035 budget that member states have not agreed.
For cohort context: €1B committed against $89B spent in a year is not a gap in ambition. It is a gap in arithmetic. Europe is choosing where to put buildings while China is forecasting what it costs to run them.
What to do with this if you are not building data centres
The report does not put anyone out of work. It does put a dated line on the table, and that line is useful for career decisions.
First, if inference reaches 80% by 2029, then every year between now and then raises demand for the work of getting agents running inside a real business without incident. That is not research. It is deployment, integration, evaluation, monitoring, exception handling and compliance audit. The industry calls it engineering. In practice it is closer to operations plus quality management, and supply is nowhere near demand.
Second, the workflows burning that inference are the most standardised stretch of white-collar work: junior analysis, first-pass screening, first-level review, reconciliation, templated writing, tier-one support. There is a simple way to locate yourself. Cut a normal week into tasks, then count how many have a fixed input format and a right answer. That share is roughly your exposure to somebody else’s inference bill.
Third, geography buys less protection than it looks like it does. The report is about China’s market, but the tilt toward inference tracks unit cost, not borders. Europe’s buildings arriving two years later does not mean Europe’s job categories move two years later. Buying inference does not require the machine to be nearby.
