The focus of China’s artificial intelligence (AI) industry is rapidly shifting from foundational model training to the large-scale deployment and commercial monetization of AI agents. According to a recent industry study, this structural transformation is projected to trigger a massive, exponential surge in token consumption over the coming years.

A comprehensive research report titled “Research Report on AI Infrastructure Development in the Era of Intelligent Agents,” released by the China Telecom Research Institute under the state-owned telecommunications giant China Telecom, highlights that China’s AI landscape has officially entered a new developmental phase where intelligent agents serve as the primary product form. Unlike traditional generative AI tools that merely respond to static queries, AI agents function as autonomous software systems capable of independent reasoning and executing complex multi-step workflows. Whether automatically planning multi-destination travel itineraries or managing sophisticated enterprise customer relations, these applications empower everyday users to integrate advanced AI capabilities into their daily workflows and personal lives without requiring specialized technical expertise.

This momentum was vividly illustrated earlier this year when “OpenClaw,” an office automation AI agent developed by Austrian engineer Peter Steinberger, sparked a massive nationwide craze across China. Major digital ecosystems and portals, including Tencent and Baidu, rushed to support OpenClaw installations, spawning a thriving market for third-party setup services charging up to 199 yuan (approx. $28) per installation. Although regulatory authorities introduced heightened security scrutiny around March, the fundamental concept of the AI agent successfully penetrated mainstream consumer and enterprise consciousness.

Driven by the proliferation of these agent-based applications, token consumption is experiencing unprecedented growth. Tokens serve as the fundamental data units processed by large language models during training and inference phases, breaking text down into manageable segments while acting as the core metric for commercial service pricing. Rao Xiaoyang, Director of the Industrial and Enterprise Strategy Research Institute under the China Telecom Research Institute, projects that China’s annual token consumption will reach 10 trillion tokens by the end of 2026, before skyrocketing past 350 trillion tokens by 2030. 

Concurrently, the underlying demand structure for computing power (computing infrastructure) is undergoing a fundamental shift. The report forecasts that inference workloads will account for approximately 80% of total computing demand, decisively outpacing training requirements. While model training represents a one-time capital expenditure to build a foundational model, inference occurs continuously every time an end-user interacts with an AI application. For corporate stakeholders, inference expenses transition from capital expenditures (CapEx) to ongoing operational expenditures (OpEx). Reflecting this aggressive infrastructural buildout, total AI investments by major Chinese technology enterprises are anticipated to approach 600 billion yuan (approx. $83 billion) this year alone, as telecommunications operators and cloud giants restructure their commercial models around universal token ecosystems.

[Copyright (c) Global Economic Times. All Rights Reserved.]