On September 21, International Data Corporation (IDC) and Inspur Information jointly released the 2026 China Artificial Intelligence Computing Power Development Assessment Report. As the industry’s first research report focused on agent infrastructure, it notes that intelligent agents are reshaping computing infrastructure, driving an exponential surge in Token consumption through a multiplier effect, and sharply revising upward the growth trajectory of China’s intelligent computing power. The report projects that China’s intelligent computing power capacity will reach 2,576.5 EFLOPS in 2026, up 87.9% year over year, and further rise to 10,524.3 EFLOPS by 2030.
Why can agents expand computing demand to this extent? The report outlines a transmission chain: the number of active agents worldwide is expected to grow from 79.4 million in 2026 to 2.216 billion in 2030, a CAGR of 129.8%. As tasks move from simple information retrieval and content generation to complex tasks requiring multi-turn reasoning, tool invocation, and cross-system collaboration, Token consumption is amplified exponentially. Over the same period, global Token consumption is projected to post a CAGR of 4,822.6%, roughly 37 times the growth rate of active agents.
Rising demand is also reflected in market size. China’s AI computing power market is expected to reach $51.5 billion in 2026, up 37% year over year, and $150.1 billion by 2030, with a CAGR of 31.8% from 2025 to 2030. Supply, however, will not be easy: the report expects the global AI compute demand fulfillment rate to fall from 79% in 2024 to 71% in 2027 and recover only to 77% by 2030, meaning supply-demand tensions will persist over the long term.
Simply expanding computing capacity is no longer sufficient to meet sustained growth in intelligence demand. The report notes that in the agent era, AI infrastructure is shifting from scale expansion to competition over system efficiency. Through full-system coordination across computing, storage, networking, data centers, terminals, and models, scheduling, and Token operations, the core of competition is moving from “how much computing power you have” to “how efficiently you can continuously produce intelligence.” The computing industry is also evolving along two lines: capability-oriented intelligent computing, which pushes the ceiling of intelligence through architectures such as supernodes, and capacity-oriented intelligent computing, which improves Token output efficiency through diverse compute collaboration, inference decoupling, and scheduling optimization.
This aligns with the direction of the StarWar GPU Computing Platform. As different workloads—inference, planning, tool invocation, and others—interweave on the same infrastructure, unified compute resource management and dynamic scheduling will directly determine whether resources can be kept fully utilized, turning “efficient intelligence production” from a goal into a practical engineering problem.
The report also observes uneven industry adoption. Software development and office tasks have already entered large-scale application because of high digitalization, clear task boundaries, and verifiable results. By contrast, in manufacturing, healthcare, and energy and power, the share of companies that have truly entered core business and formed an “execution-feedback-optimization” loop remains 0%; while 70.0% of telecom operators have integrated some business processes, only 3.3% have entered a core-business closed loop.
At the city level, evaluation logic is also changing. For the first time, the report incorporates agent application deployment, embodied intelligence industry layout, large language model development, and inference computing infrastructure into its indicators for city-level AI computing power development. Beijing, Hangzhou, and Shenzhen rank in the top three. As computing power is elevated to the same strategic infrastructure status as water conservancy, power grids, and railways, whether it can be converted stably and efficiently into intelligence supply will become the real dividing line for the next stage.