The report shows that the large-scale operation of intelligent agents is bringing sustained and rigid computing power demand. By 2026, the scale of China's intelligent computing power will reach 2576.5 EFLOPS, a year-on-year increase of 87.9%; By 2030, it will reach 10524.3EFLOPS, with a compound annual growth rate of 50.3% from 2025 to 2030. During the same period, the compound annual growth rate of China's general computing power reached 27.7%, which was significantly higher than the previous forecast. The market size of artificial intelligence computing power in China is expected to reach 51.5 billion US dollars by 2026, a year-on-year increase of 37%, and will increase to 150.1 billion US dollars by 2030. The report predicts that by 2031, artificial intelligence will create a cumulative global economic value of $22.5 trillion, far exceeding the scale of any previous generation of general-purpose technology.

Intelligent agents are becoming an important incremental investment in artificial intelligence. IDC predicts that the number of active intelligent agents worldwide will increase from 79.4 million in 2026 to 2.216 billion in 2030, with a compound annual growth rate of 129.8%; The compound annual growth rate of global token consumption during the same period reached 4822.6%, which is about 37 times the growth rate of the number of active agents. The number of active intelligent agents in China is expected to increase from 4.95 million in 2026 to 197 million in 2030, with a compound growth rate of 151.2%. The penetration rate of intelligent agent deployment is 81.0%, which is higher than the global leading level.

The application of intelligent agents has entered the stage of 'using intelligent agents' from' discussing intelligent agents', and it is expected that the overall explosive growth will be concentrated in the next two years, "said Zhou Zhengang, Vice President of IDC, in an interview with reporters.

The report also reminds that the supply growth rate in the physical world is difficult to match the demand growth rate in the digital world. Due to factors such as longer production expansion cycles for core components such as high-end chips, HBM (high bandwidth memory), and server CPUs, as well as overlapping energy constraints, the global demand for artificial intelligence computing power is expected to continue to decline from 79% in 2024.

Next year will be the trough of computing power satisfaction rate, which will gradually ease thereafter, "said Zhou Zhengang. It is expected to be 71% in 2027 and rise to 77% in 2030; The supply-demand gap will widen from $38.9 billion in 2024 to $380.9 billion in 2030. The frequency of token and cloud service price adjustments will significantly accelerate in 2026, with an expected annual increase of 30% to 50%.

In his view, AI computing has entered a new stage of system capability competition, and industry competition is no longer limited to peak performance competition of single chips, but is evolving towards a dual line parallel direction of "capability based intelligent computing" and "capacity based intelligent computing". Among them, "capability based intelligent computing" expands the high-speed interconnection domain and unified memory addressing through the Scale up architecture represented by supernodes, enhances the scale and complex task capability of single machine carrying models, and breaks through the upper limit of intelligence; Capacity based intelligent computing improves token output efficiency, reduces unit token costs, and achieves intelligent scale supply through system innovation such as multi computing power collaboration, inference decoupling, caching, and scheduling optimization.

The report points out that intelligent agents are driving the shift of artificial intelligence infrastructure from single point performance optimization to task level system collaborative optimization. Data shows that 94.3% of surveyed companies believe that supernodes are an important capability of artificial intelligence infrastructure. At the same time, storage systems are moving from data storage to intelligent data supply; The network is evolving towards a low latency, high reliability connection intelligent agent task execution chain; The data center is restructuring towards high-density, continuously operating intelligent production systems, with single cabinet power expected to exceed 300 kilowatts by the end of 2026. Native liquid cooling is gradually becoming the core system design concept for complete cabinets and even the entire data center.

From the perspective of industry implementation, code development agents and office agents have taken the lead in entering the stage of large-scale application due to their high degree of digitization, clear task boundaries, and verifiable results; Traditional industries such as manufacturing, healthcare, and energy and power have begun to embed artificial intelligence into specific business processes, but overall they are still in the exploratory pilot stage. The report shows that 47.2% of manufacturing companies are piloting intelligent agent applications, and embodied intelligence is starting to drive artificial intelligence from the digital space to the physical world.

Intelligent agents will eventually move towards a closed loop of multi-agent collaboration, self verification under human supervision, and self circulation. Industries such as healthcare and manufacturing still need to gradually solve problems such as process adaptation and result verification, "said Zhou Zhengang.

The report shows that by 2026, the urban pattern of artificial intelligence computing power in China will be further reshaped, with Beijing, Hangzhou, and Shenzhen forming the first tier. Among them, Shenzhen, with its strong computing power supply industry foundation and rapid growth in intelligent agent applications, has surpassed Shanghai and entered the top three.

The report suggests that enterprises should follow the gradual evolution ladder of intelligent agents and build a hybrid computing architecture of "self built Token Factory+public MaaS+edge computing power"; Industrial planners should promote the transformation of intelligent computing infrastructure from a "construction oriented" to a "landing oriented" approach, using standards as the first priority and demonstration as the driving force to promote the industrialization of super nodes and native liquid cooling, and support the coordinated maturity of the intelligent agent industry and computing infrastructure.