China Telecom Research Institute: China's AI Focus Shifts From Large Models to Agents

China Telecom Research Institute released AI Infrastructure Development in the Agent Era (2026) on Saturday, September 12, according to CCTV News and Securities Times. The report's central claim is that China's AI industry is shifting its centre of gravity away from large models and compute competition toward agents deployed at scale and applications that can actually be monetized.

The report argues that AI is entering a new phase in which agents are the dominant form. Agents hide the underlying complexity and make the technology broadly accessible: consumers and enterprises do not need specialist skills to have an agent complete a full, multi-step task.

On the numbers, Rao Shaoyang, director of the institute's Industry and Enterprise Strategy Research Institute, expects China's annual token consumption to reach about 1x10^17 in 2026 and to exceed 3.5x10^19 by 2030, a compound annual growth rate of nearly 12 times, a pace that pushes AI usage intensity straight into infrastructure planning.

Compute demand shifts with it. Rao said the agent boom will keep driving rapid growth in Chinese compute demand, and that inference will far outstrip training: by 2029, inference is expected to account for 80 percent of China's compute market.

That changes how infrastructure has to be planned. Training is a concentrated, one-off capital expense; inference is recurring spending that scales with users and with the number of agents running, and it puts very different demands on power, network latency and scheduling efficiency.

Pairing deployment at scale with application monetization in the same finding also signals that the industry's attention is moving from model leaderboards to billable usage. The token and inference-share forecasts effectively anchor AI's commercial value to call volume rather than rankings.

For Chinese cloud providers and telecom operators the implication is direct: whoever can scale inference cost, agent orchestration and industry workflows is best positioned to capture the incremental demand agents create. At the same time, consumption in the 10^17 range means pressure on energy and compute supply will grow alongside it.

The next things to watch are how quickly agents penetrate government, finance and manufacturing workflows, and whether the capital spending of carriers and cloud providers on inference compute can keep pace with the growth curve the report describes.