China’s AI Race Is Shifting From Bigger Models to Scalable Agents

China’s AI sector is increasingly being framed as a deployment race: turning foundation models into agents that can perform useful work across software, telecoms, manufacturing, and consumer services. The central question is no longer only who can train the largest model, but who can operate AI systems at high volume with acceptable cost, speed, reliability, and safeguards.

The headline shift: from training capacity to inference demand

A China Telecom Research Institute report forecasts that inference will represent 80% of China’s computing-power market by 2029, overtaking training-related demand. The report also projects that agents could drive nearly tenfold annual growth in computing demand over the next two to three years. 59

The same research stream estimates China’s annual token consumption could reach 100 quadrillion in 2026 and exceed 35 quintillion by 2030. Those figures are projections rather than observed outcomes, but they illustrate why serving models repeatedly may become as strategically important as training them. 6

Inference is the work performed when a trained model responds to a prompt or carries out a step in a task. Agents can make that workload much more intensive than a single chatbot response: they may retrieve information, call software tools, execute multi-step workflows, verify outputs, and repeat the process. At scale, each of those actions consumes compute and adds operational complexity.

Why agents change the economics of AI

Training a frontier model is typically a concentrated infrastructure expense. Deploying an agent is an ongoing service operation. Its costs can recur through model inference, data retrieval, networking, storage, monitoring, and human review or escalation.

That changes the commercial test. A capable agent is not necessarily a viable product if it is too expensive to run, too slow for the workflow, or too unreliable to be trusted with business systems. The most valuable measure becomes the cost and quality of a completed task—not simply a model’s headline benchmark performance.

This creates strong incentives to improve serving efficiency and operational discipline. In practice, providers and enterprises will need to focus on:

• selecting the smallest model that can safely handle a task;
• routing work to the appropriate model or tool;
• reducing repeated computation through caching and efficient serving;
• connecting agents to trusted, task-specific data;
• testing tool use and end-to-end workflows, not only model outputs;
• monitoring failures, access controls, and audit trails; and
• ensuring that people can intervene when an agent reaches a high-risk decision or action.

The strategic implication is not that frontier-model training stops mattering. Training remains important for capability, independence, and access to advanced models. But when capable models are broadly available, sustained execution—cheap, dependable inference connected to real systems—can become the harder differentiator.

Beijing’s policy focus is application diffusion

China’s policy signals reinforce the move from standalone models to repeatable deployments. The Ministry of Industry and Information Technology’s AI-plus-software plan calls for AI-powered programming tools to reach more than 20,000 major software enterprises by 2028, alongside 100 benchmark agent-based applications across key industries and at least five high-quality open-source projects. 3

The intended 2030 direction is a more intelligent, secure software ecosystem, including intelligent upgrades to critical software systems. 37 A separate MIIT plan for AI and information communications calls for more than 30 high-value scenarios by 2028, as well as representative applications and specialized agents; it also links deployment to stronger network and compute infrastructure. 17

Another MIIT and National Data Administration initiative explicitly calls for industry-specific high-value AI scenarios, specialized data sets, and specialized models or agents where autonomous planning and execution are needed. It also calls for tailored evaluation data sets and assessment systems for industry models and agents. 18

Taken together, these initiatives suggest a policy model aimed at diffusion: establish reusable scenarios, data, software tools, and reference deployments that can be adopted across sectors. The opportunity is faster adoption. The risk is that success gets measured by the number of pilots rather than whether systems deliver durable, safe, and economical results.

Spending and industrial scale raise the stakes

Reporting based on the China Telecom Research Institute analysis puts Chinese technology companies’ AI spending this year at close to 600 billion yuan. Because public reports can use different definitions of AI spending, that aggregate should be treated as an estimate rather than a fully standardized measure. 9

At the national level, China expects AI-related industries to exceed 10 trillion yuan by the end of the 2026–2030 Five-Year Plan period. 45 That framing matters: AI agents are being positioned not merely as a cloud-service category, but as an enabling layer for industrial modernization, software development, communications, and consumer services.

Europe is building compute capacity, but deployment matters too

Europe is pursuing its own infrastructure response. The European Commission has opened a call for up to seven AI Gigafactories, supported by up to €10 billion in EU and national funding and expected to unlock at least €20 billion in private investment. 50

The initiative is significant for European compute capacity and technological sovereignty. Yet compute construction alone does not resolve the deployment challenge. The competitive value of that capacity will depend on whether it supports accessible, affordable inference; useful domain data; software integration; and trustworthy applications for European organizations.

What the next phase of the AI race will reward

China’s agent push points to a broader global transition. AI advantage will increasingly depend on converting compute into useful, auditable work at sustained scale. That puts a premium on chips and electricity, but also on serving efficiency, data access, workflow integration, security controls, evaluation, and human accountability.

The 80% inference figure should be read as a directional forecast, not a settled fact. Still, its message is clear: as AI moves out of demos and into continuous operations, the winners may be less defined by the largest training run than by the ability to run agents reliably enough—and cheaply enough—to make them part of everyday work. 56