China's AI industry is moving from competing on large models and raw computing power to deploying and selling AI agents, according to a report from the China Telecom Research Institute, the research arm of the state-owned carrier, carried by state broadcaster CCTV on Saturday. The report expects agents to drive close to tenfold annual growth in the country's computing demand over the next two to three years, and puts inference at 80% of China's computing-power market by 2029, overtaking training demand. For business, the shift means the money in AI is moving from building models once to paying for them every time they are used.
What the state report says
The sharpest figure in the report is the split between training and inference. Training builds a model once; inference is what it costs every time somebody uses one, which makes it an operating expense rather than a capital one. By 2029, inference is expected to account for 80% of China's computing-power market, ahead of training. Chinese technology companies are expected to spend close to 600 billion yuan, about $89B, on AI this year, which the report puts at more than a tenth of all investment in the country. That spending level is the backdrop against which the agent forecast is made.
The mechanics behind the forecast are straightforward. An agent does not just answer a query once; it runs a sequence of steps, calls tools, checks results and repeats the cycle, so each task consumes far more inference capacity than a single chat response. That is why the report ties agent adoption directly to computing demand rather than to model quality. The same logic changes who pays: inference is billed per use, so the cost lands on the operating budget of the company running the agent, not on the capital budget of the lab that trained the model. The report does not name specific agent products or vendors, and it does not give a breakdown by industry.
China's forecast arrives while Europe is building for the same demand on a different clock. The European Commission opened bidding in July for up to seven gigafactories, a EUR 30B programme with about EUR 10B of public money and EUR 20B hoped for from private investors. Roughly EUR 1B of it is actually committed. Chinese technology companies are on course to spend that much on AI about every five days this year. 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 a year before China's inference crossover, assuming nothing else slips.
What this means for business
For companies adopting AI, the practical consequence is that the cost centre moves. A firm that bought a model licence or ran a pilot now faces a per-use bill that scales with how often agents run, so budgeting shifts from a one-off project to a recurring operating line. The difference is sharpest by size: a small company can start with a single agent on one process and watch the inference bill directly, while a large enterprise running agents across support, sales and back office needs usage controls before rollout, because tenfold growth in demand is also tenfold growth in spend if nothing is capped. Procurement questions change accordingly — not only which model, but how many calls a task takes and who pays for them.
What the report does not settle is timing and availability. It is a forecast about running models, not a product announcement, and it names no vendors, prices or deployment dates. Europe's programme carries its own uncertainties: bidding moved from May to July, the evaluation criteria were delayed more than once, and interest fell from about 70 companies to roughly ten expected bidders. Most of the public half of the EUR 30B depends on a budget for 2028 to 2035 that member states have not agreed. Before committing, buyers should ask vendors where inference runs, what a task costs at current usage, and what happens to the price if demand grows as the report predicts.
The marker to watch is the 2029 inference share. If China's computing market does cross to 80% inference, the operating-expense model becomes the default way AI is bought, and vendors that optimise models for running rather than building will hold the margin. If the share stalls below that level, the agent-driven demand forecast was early. In Europe, the nearer test is 12 November, when bids for the gigafactories close and the number of actual bidders shows whether the programme can reach the scale it targets.
