Summary
China’s AI sector is shifting from a focus on large-model training and raw computing capacity toward deploying and commercializing AI agents. The China Telecom Research Institute expects inference to account for 80% of China’s computing market by 2029, with agents potentially driving nearly tenfold annual growth in demand over the next two to three years.
From building models to running agents
China’s artificial intelligence industry is moving into a new phase. Rather than competing primarily on the size of foundation models or the amount of raw computing capacity available for training, companies are increasingly focused on deploying AI agents and turning them into commercial products and services.
That is the central conclusion of a report from the China Telecom Research Institute, the research arm of the state-owned telecommunications carrier. The report was carried by China Central Television. Its projections point to a change in how demand for AI infrastructure is generated: instead of being concentrated in large model-training campaigns, demand could increasingly come from the continuous operation of systems used by businesses and consumers.
AI agents typically do more than generate a single response. They may interpret an instruction, break a task into steps, call software tools, retrieve information, perform an action and then evaluate the result. Each additional step can require another model invocation. As agents become integrated into workflows, the resulting demand is likely to be more frequent and operational in nature than the demand created by a one-time training run.
The report expects agents to drive close to tenfold annual growth in China’s computing demand over the next two to three years. The summary available does not specify the industries, deployment volumes or methodology behind that forecast. It should therefore be treated as an institutional projection rather than a measure of realized market growth. Still, it highlights where Chinese industry research expects the next infrastructure bottleneck to emerge.
Inference becomes the central cost question
The report’s most striking forecast is that inference will account for 80% of China’s computing market by 2029, overtaking training-related demand. Inference is the process of running a trained model to produce an output. It is required each time a user or software system calls the model, making it a recurring operating expense rather than a predominantly upfront development cost.
Training remains resource-intensive and will continue to support the development of new models, specialized systems and multimodal capabilities. But inference has a different economic profile. Training resources may be concentrated in a limited number of major projects, while inference costs accumulate as usage expands across customer-service systems, enterprise software, office automation and other applications.
If the forecast is broadly correct, the way companies evaluate AI infrastructure will have to change. Peak performance and training efficiency will remain relevant, but buyers and operators will also pay closer attention to latency, cost per request, energy consumption, reliability and the ability to coordinate models with external tools. Capacity planning will become less about a single large training cycle and more about supporting variable, recurring workloads.
The shift also raises operational questions. An agent with access to internal data or business software may need to perform several actions before completing a task. Organizations will need controls over what the system can access, which actions require approval and how each operation is recorded. In financial, payments or asset-management environments, permission separation, key management, audit trails and human review can be as important as model quality. The report does not discuss a specific custody or wallet provider, and there is no basis to attribute any particular product strategy to the report.
Spending moves closer to the application layer
The China Telecom Research Institute report says Chinese technology companies are expected to spend close to 600 billion yuan, or about $89 billion, on AI this year. It puts that amount at more than one-tenth of total investment in the country.
The figure indicates that AI has become a broad investment category rather than a narrow research program limited to a handful of technology companies. Yet spending levels alone do not establish that commercial deployments are working at scale. Moving an agent from a demonstration to a production environment can require extensive work on data quality, access permissions, exception handling, cybersecurity and accountability.
The economics of deployment may also differ significantly from the economics of model development. A training expense is often concentrated in the period when a model is created or updated. Inference expenses continue as users make requests and as agents perform more complex tasks. Enterprises assessing an agent project may therefore need to account not only for model access, but also for monitoring, human review, maintenance, security controls and integration with existing systems.
Reliability will be another important constraint. A conventional software feature may follow a predefined path, while an agent can make intermediate decisions based on changing information. That flexibility can make it useful, but it can also create unpredictable costs or errors. Wider adoption will depend on whether organizations can limit those risks without eliminating the benefits of automation.
Europe follows a different infrastructure timetable
The same reporting compares China’s industry shift with Europe’s effort to expand AI infrastructure through large centralized facilities. The European Commission opened bidding in July for up to seven AI gigafactories. The program has a planned budget of 30 billion euros, with around 10 billion euros expected from public funds and 20 billion euros sought from private investors.
Only roughly 1 billion euros has been committed so far, according to the report. Applications are due by November, awards are expected in early 2027, construction is planned to begin that year and the machines are expected to be operational by the middle of 2028. Those milestones remain dependent on the bidding process, financing, construction, equipment delivery and access to sufficient power.
Europe’s approach emphasizes the creation of large-scale shared infrastructure and the mobilization of private capital. It is not necessarily a direct substitute for the Chinese model described in the report. China’s projected spending reflects activity by technology companies and a rapid shift toward applications, while Europe’s initiative is a public-private infrastructure program still moving through the selection and funding process.
The gap between the program’s intended budget and its committed financing is therefore significant. The plan may eventually expand Europe’s access to computing resources, but its eventual impact will depend on whether the facilities are delivered on schedule, whether they can attract sustained demand and whether their capacity matches the needs of both model training and inference workloads.
The competition enters an operations phase
Taken together, the developments point to a broader change in the AI market. Model capability remains important, but it is no longer the only measure of competitiveness. The next phase will also be defined by how efficiently systems can run, how securely they can interact with business software and how consistently they can deliver useful results at scale.
For infrastructure operators, that means optimizing not only chips and data centers but also scheduling, energy use, model routing and inference costs. For enterprises, it means evaluating agents as ongoing operational systems rather than as isolated software purchases. For regulators and institutional users, questions around data access, automated decisions, accountability and cross-system actions will become increasingly material.
The China Telecom Research Institute’s 2029 forecast is not a settled outcome. It is, however, a clear statement about the direction of travel: AI infrastructure demand may increasingly be driven by everyday use rather than by the initial construction of models. As agents move from demonstrations into business processes, the industry’s success metrics are likely to expand from model size and benchmark scores to include inference economics, operational reliability and governance.
Source: link