AI Agents Ignite the “Multiplier Effect” in Computing Power: An Industrial Paradigm Revolution from “Selling Resources” to “Producing Intelligence”

In the past week, the heat in the computing power sector has not faded, but the spotlight has quietly shifted. If previously the market’s core concern was whether chips were “sold out” and whether computing centers were “full,” then the most watched topic in the past week comes from a deeper variable—AI agents are reshaping the underlying logic of the entire computing power industry with a “multiplier effect.” From the 2026 China Artificial Intelligence Computing Power Development Assessment Report jointly released by IDC and Inspur Information, to the assertion at the 2026 Apsara Conference that “the total amount of machine thinking will reach 1,000 times that of humans,” to the intensive releases of new products and strategies for the agent era by leading vendors such as Lenovo, Inspur, and AMD, a clear signal is emerging: the yardstick for the computing power industry is shifting from “how much computing power do you have” to “how much intelligence can you produce.”

This is not an ordinary technological iteration, but a systemic reconstruction involving demand structure, hardware value distribution, business models, and even national strategic positioning. To understand this transformation, one must begin with why AI agents possess a “multiplier effect.”

I. AI Agents Rewrite the Computing Power Equation: Token Consumption Growth Is 37 Times the Growth of Agents

The key to understanding this round of exploding computing power demand lies in a variable that had not been fully discussed before: the “multiplier effect” of AI agents. Over the past two years, the core topics of the large-model industry have been parameter scale, training data volume, and model leaderboard rankings. The subtext was that AI is a race to “build engines”—whoever builds a larger and stronger engine holds the initiative. What AI agents rewrite is precisely this equation.

An AI agent refers to an AI system capable of autonomous planning, tool invocation, multi-step task execution, and self-correction based on feedback. Its fundamental difference from traditional conversational AI is this: conversational AI is “one question, one answer,” and the interaction ends after a single exchange; an AI agent, by contrast, is like a digital employee—after receiving a task, it autonomously breaks it down into steps, invokes search engines or databases, writes and executes code, checks results, and retries after finding errors until the task is complete. This process may involve anywhere from a few to dozens of reasoning rounds, and each round consumes tokens.

In the 2026 China Artificial Intelligence Computing Power Development Assessment Report, IDC provided two highly striking figures: the number of active AI agents worldwide will grow from 79.4 million in 2026 to 2.216 billion in 2030, a compound annual growth rate of 129.8%; over the same period, global token consumption will grow at a compound annual growth rate of 4,822.6%, roughly 37 times the growth rate of the number of agents. This means that a single agent does not simply complete one invocation and exit; through multi-round reasoning, tool calls, and cross-system collaboration, every task expansion triggers exponential token consumption.

The industrial implications behind this figure deserve deeper unpacking. Suppose an agent receives the task “analyze a company’s financial data over the past five years and generate investment advice.” It may need to sequentially: call a database to obtain financial report data, write a Python script to calculate financial ratios, retrieve peer company data for benchmarking, generate charts, write an analysis report, self-check data consistency, and correct calculation errors. The entire process may involve dozens of model inference calls, consuming dozens or even hundreds of times more tokens than a user directly asking, “Is this company worth investing in?” When hundreds of millions of agents operate simultaneously in finance, manufacturing, healthcare, government affairs, and other scenarios, the scale of token consumption will expand geometrically.

As Huang Shan, Strategic Technology Director of Lenovo China’s Infrastructure Business Group, pointed out: “In the past, users asked questions and models answered, and token consumption followed a linear trajectory; now, to make AI execute multi-step tasks and self-verify like a digital employee, token consumption is no longer a simple multiple relationship, but shows several-fold or even exponential explosive growth.” He further used a vivid metaphor: “In the past, we measured computing power demand like calculating how many customers a restaurant serves per day; now, in the agent era, every ‘customer’ orders a full set meal and keeps asking for more dishes—the restaurant’s table turnover rate is on a completely different order of magnitude.”

This structural change has directly pushed up the growth slope of China’s intelligent computing power. The report estimates that China’s intelligent computing power scale will reach 2,576.5 EFLOPS in 2026, a year-on-year increase of 87.9%, and will further reach 10,524.3 EFLOPS by 2030; the compound annual growth rate from 2025 to 2030 has been raised from 46.2% in the previous forecast to 50.3%. A 4.1 percentage point upward revision may not seem large, but its essence is not a more optimistic assumption, but a correction to the shape of the demand curve—what agents bring is not linear extrapolation, but an entirely new demand paradigm. Linear extrapolation means “whatever growth occurred in the past will roughly continue in the future”; a paradigm shift means “the growth curve itself has changed shape,” from a gentle exponential curve to a steep one. Over five- or ten-year horizons, this difference will produce scale differences of several times or even dozens of times.

From a global perspective, this trend is equally significant. Gartner predicts that by 2028, at least 15% of daily work decisions will be made autonomously by agents, and by then global annual AI token consumption will reach hundreds of times the current level. Data from the China Academy of Information and Communications Technology shows that in the first half of 2026, China’s average daily AI token call volume exceeded 500 trillion, more than six times the level in the same period of 2025, far exceeding previous general industry expectations.

II. The Inference Side Becomes the Main Battlefield: The Token Consumption Ratio Between Training and Inference Has Exceeded 6:4

Another profound change driven by agents is the shift in computing power demand structure from training-dominated to inference-dominated. The significance of this shift is no less than the internet’s transition from “dial-up” to “always-on broadband”—it means computing power is no longer a one-time capital expenditure, but a continuous operational expense.

During the Apsara Conference, Huang Shan revealed that the token consumption ratio between inference and training has now exceeded 6:4, and computing power demand on the inference side is growing rapidly. Data provided by Ji Zhaohui, Vice President of Marketing for AMD Greater China, further confirms this trend: in 2025, the workload split between AI inference and training was 5:5; by 2026, inference had risen to 60%; and by 2030, it is expected to exceed 80%. In other words, the investment focus of computing power infrastructure in the future will shift from “building large clusters to train large models” to “broad coverage for massive inference.”

Gartner data reveals the order-of-magnitude difference in consumption between agents and traditional conversational bots: token consumption for agent applications is 5 to 30 times higher than for conversational bots, while IDC predicts that by 2027, AI token consumption among the Global 2000 will increase 1,000 times compared with today. What does 1,000 times mean? If a large enterprise currently consumes an average of 1 billion tokens per month for AI, then by 2027 that will swell to the level of 1 trillion, and the corresponding computing power cost will leap from hundreds of thousands of yuan to hundreds of millions.

This change places entirely different requirements on computing power infrastructure. Traditional training clusters pursue extreme single-card performance and low-latency interconnection; their design philosophy is to “connect ten thousand cards into one card,” emphasizing high bandwidth, low latency, and strong consistency—typical examples being NVIDIA’s NVLink and InfiniBand networks. Agent inference scenarios, however, must handle load characteristics such as “fragmented requests, large traffic fluctuations, and greater emphasis on fine-grained resource partitioning.” Requests may come from different regions, applications, and priority levels; traffic may fluctuate dozens of times within a day; and each request has a different sensitivity to latency. This requires computing power infrastructure to have stronger elastic scheduling capabilities and finer resource partitioning granularity—for example, splitting a single GPU into multiple virtual instances and allocating them on demand to different agent tasks.

Liu Jun, Chief AI Strategy Officer of Inspur Information, summarized this divergence at the 2026 Artificial Intelligence Computing Conference into two clear development dimensions: “capability-oriented intelligent computing” for the intelligent frontier, supporting breakthroughs in the upper limits of intelligence and answering “how strong can intelligence ultimately be”; and “capacity-oriented intelligent computing” for the popularization of intelligence, achieving adequate supply of intelligence and answering “how much intelligence can we obtain at how low a cost.” The two are not separated from each other, but form a positive cycle of “breakthrough at the frontier—popularization and deployment—new demand generation—further breakthroughs.”

Liu Jun further explained that the core metrics of capability-oriented intelligent computing are “maximum single-cluster computing scale” and “training efficiency,” with users mainly being leading large-model companies and national research institutions; the core metrics of capacity-oriented intelligent computing are “unit token cost” and “inference throughput,” with users covering small and medium-sized enterprises and developers across all industries. The two types of intelligent computing centers also differ significantly in hardware selection, network architecture, cooling solutions, and even site selection—capability-oriented centers tend to be deployed in western hubs with abundant energy and cool climates, while capacity-oriented centers are closer to densely populated eastern user areas to reduce network latency.

This divergence also explains why the computing power leasing market has recently been extremely active. For most enterprises, the cost-effectiveness of building their own capacity-oriented intelligent computing center is not high—hardware iterates too quickly, operations and maintenance are too complex, and demand fluctuates too much. Renting computing power and paying by token has become a more rational choice. This is precisely the industrial background for the rise of the “selling tokens” model.

III. CPU Returns to the Core of Computing Power: CPU Core Demand per Gigawatt of Data Center Has Quadrupled

The paradigm shift brought by agents is reshaping the value distribution pattern of the entire AI hardware industry chain. Over the past two years, market attention was almost entirely occupied by GPUs, and NVIDIA’s market value myth made “GPU equals AI” a common perception. But as AI shifts from “training-dominated” to “inference- and agent-driven,” the CPU is leaping from an auxiliary supporting role to the core hub of AI systems. This change is rewriting the power map of the semiconductor industry.

AMD CEO Lisa Su explicitly pointed out during the company’s Q1 2026 earnings call that in AI infrastructure, the CPU-to-GPU ratio is gradually converging from the traditional 1:4 to 1:8 toward 1:1, and in agent-intensive deployment scenarios, the number of CPUs may even surpass GPUs. The logic is that, in addition to GPUs and accelerators, complex AI agent workflows require large numbers of CPUs to handle task scheduling, tool execution, and orchestration; orchestration alone can account for 50% to 90% of total system latency. In other words, GPUs handle “computing,” while CPUs handle “managing”—when the workload of “managing” rises exponentially with agent complexity, the strategic value of CPUs naturally rises.

Arm’s calculations show that the number of CPU cores required per gigawatt of data center has surged from about 30 million to 120 million, a fourfold increase. Morgan Stanley estimates that by 2030, Agentic AI could bring an incremental CPU market of $32.5 billion to $60 billion, corresponding to an overall server CPU market size that could exceed $100 billion. If this prediction comes true, it means the server CPU market will double within five years, with almost all incremental demand coming from AI agent-related needs.

The direct consequence of supply-demand imbalance is comprehensive tightening of prices and lead times. Server CPU delivery lead times for Intel and AMD have extended from the usual 1 to 2 weeks to 8 to 12 weeks, with some Intel products taking up to 6 months. Intel’s CFO expects server CPUs to be in certain short supply in 2026, with supply pressure potentially lasting until 2028. The server industry is shifting from “demand determines growth” to “supply determines how much can be delivered,” and CPUs have become a new bottleneck in AI computing power expansion after GPUs and HBM.

This bottleneck has also spawned new industrial opportunities. On the one hand, ARM-architecture server CPUs are accelerating their penetration into the data center market, and their high energy efficiency is particularly prominent in inference scenarios—shipments of self-developed ARM chips such as Amazon Graviton, Google Axion, and Alibaba Yitian 710 continue to climb. On the other hand, the RISC-V architecture is also beginning to emerge in AI accelerators and edge inference scenarios, providing domestic chips with a new path around x86 and ARM licensing restrictions. Diversified competition in CPU architectures is becoming one of the most dynamic variables in the AI computing power industry chain.

IV. From “Selling Resources” to “Producing Intelligence”: A Fundamental Reconstruction of the Computing Power Service Business Model

The most profound change brought by agents may be that the business model of computing power services is undergoing fundamental reconstruction. In the past, the core logic of computing power services was “selling resources”—billing by GPU card-hour, storage capacity, and network bandwidth. Customers bought “bare computing power” and had to solve model deployment, performance tuning, troubleshooting, and a series of other problems themselves. In the agent era, customers want “directly usable intelligence,” paying by token and settling by results—“Intelligence as a Service.”

Yu Fenghui, a senior researcher at the Pangu Think Tank, pointed out that the most significant change in the current computing power service industry is that the delivery model is shifting from “selling resources” to “selling tokens,” the competitive dimension is shifting from scale to efficiency, and the industry competition metric is changing from “how many cards do you have” to “token production efficiency per watt.” Integrated software-hardware optimization, inference engine tuning, and intelligent scheduling for peak shaving and valley filling have become key capabilities. He further analyzed that this shift places entirely new requirements on the capability structure of computing power service providers: in the past, what was compared was capital strength and procurement capability—whoever could obtain more GPU cards and secure electricity at a lower price had an advantage; now, what is compared is technical capability and operational capability—whoever can produce more effective tokens with the same hardware, lower inference costs further, and ensure service stability will win customers.

Huang Shan’s summary of this shift is even more direct: “The standard for measuring AI computing power is gradually shifting from ‘how much does a token cost’ to ‘how much value has a token created.’” He further broke down the cost structure of tokens—infrastructure accounts for about one quarter, of which electricity accounts for about 10%, while software-hardware co-optimization affects more than 50% of token cost. This means that token production efficiency is influenced not only by chip hardware, but also by computing, communication, memory access, storage, inference optimization, and the scheduling capability of the entire system. Competition in AI computing power is moving from single-point hardware capability to system capability. Taking inference engine optimization as an example, he explained: “With the same model and the same hardware, a deeply optimized inference engine can increase throughput by 3 to 5 times, equivalent to reducing token cost by 60% to 80%. This cannot be achieved by buying more expensive cards; it depends on software and algorithm prowess.”

Continuous demand-side release is already directly reflected in orders. Recently, listed companies have intensively signed large computing power service contracts: the GPU computing power service contract signed by a wholly owned subsidiary of Parallel Technology with a customer has a total tax-inclusive amount of no more than 1.913 billion yuan; the computing power service contract signed by a wholly owned subsidiary of Kanghui Pharmaceutical has a total amount of approximately 1.720 billion yuan, with a five-year contract term. The common features of these contracts are long terms, large amounts, and pricing units based on tokens or computing power service volume, marking that computing power services are moving from “project-based” to “subscription-based,” and from “one-time transactions” to “long-term operations.”

At the same time, on-device AI is becoming an important path to reducing token costs. Ji Zhaohui offered a vivid metaphor: deploying agents on local new computing devices such as AI PCs and agent hosts, and handing only heavy tasks to cloud-based large models, can reduce enterprise token consumption by more than 80%—“This is like most of the time we buy ingredients and cook at home, and only go out to eat on important occasions.” Related research shows that the latest locally deployable Qwen3.8 model (27B parameters) has reached 80% to 90% of the capability of the most advanced cloud-based large models, and the feasibility of on-device AI is rising significantly. This means that the future computing power architecture will present a layered pattern of “cloud-edge-device collaboration”: the device side handles high-frequency, lightweight, privacy-sensitive inference tasks; the cloud handles low-frequency, heavy, massively parallel training and complex inference tasks; and edge nodes undertake scheduling and caching functions in between.

This trend also has profound implications for the chip industry. The rise of on-device AI will drive demand growth for NPUs, low-power CPUs, and high-bandwidth memory, while the expansion of cloud inference will benefit high-efficiency GPUs, customized ASICs, and high-speed interconnect chips. The computing power chip market will move from “GPU dominance” to a pattern of “diverse heterogeneous coexistence.”

V. Computing Power Leaps to National Strategic Infrastructure, and the Logic of City Competition Undergoes Fundamental Change

The sustained, rigid computing power demand brought by large-scale agent operation is pushing computing power from an industrial concept to national strategic infrastructure on the same level as water conservancy, power grids, and railways. During the “15th Five-Year Plan” period, direct investment in computing network construction will increase by 4 trillion yuan, a scale comparable to traditional major infrastructure such as water conservancy at 6 trillion yuan, power grids at 5 trillion yuan, and railways at 4 trillion yuan. The establishment of this strategic positioning means that computing power is no longer merely an “internal affair” of the technology industry, but a overarching issue concerning national competitiveness, industrial security, and economic lifelines.

This change in strategic positioning has also profoundly affected the evaluation system of city competition. In the 2026 China Artificial Intelligence Computing Power City Ranking, Beijing, Hangzhou, and Shenzhen ranked in the top three. Notably, the report for the first time included agent application deployment, embodied intelligence industrial layout, large-model development, and inference computing power infrastructure as key evaluation indicators for urban AI computing power development—in the past, judging a city’s AI strength meant looking at “how many computing centers it built”; now it means looking at “what it has done with computing power.” This shift in the evaluation system is essentially a question about “computing power effectiveness”: building a computing center that no one uses is vastly different in value from building one that drives industrial upgrading.

The pattern differences among the three leading cities correspond precisely to three different development paths: Beijing is “full-chain,” with complete industrial chains on both the supply and demand sides—it has chip companies such as Cambricon and Moore Threads, large-model and application giants such as Baidu and ByteDance, and top research institutions such as the Chinese Academy of Sciences and Tsinghua University, forming a complete loop from basic research to industrial application. Hangzhou is “scenario-driven,” with an AI innovation enterprise cluster and internet industrial foundation叠加 embodied intelligence and other frontier investment; Alibaba Cloud, Zhejiang Lab, and Unitree constitute a diverse ecosystem from cloud computing to robot manufacturing, and the richness of application scenarios is its core competitiveness. Shenzhen is “manufacturing-base,” and its identity as an AI server and computing chip manufacturing base has been converted into a direct advantage during the window of the agent explosion; Huawei Ascend, Tencent, BYD, and other enterprises form synergies in chips, cloud services, and intelligent terminals, and its hardware manufacturing capability and supply chain efficiency are its unique endowments. There is no superiority or inferiority among the three paths, but together they reveal a trend: single-point advantages are no longer enough; real competition occurs in the closed-loop efficiency of “supply—scenario—industry.”

From a broader perspective, the strategicization of computing power infrastructure also brings new governance challenges. How to avoid duplicated construction, how to promote cross-regional scheduling, how to ensure energy supply, and how to balance data security and circulation efficiency are all urgent issues. The launch of the national integrated computing network monitoring and scheduling pilot verification platform is precisely an institutional response to this series of challenges. In the future, computing power may well operate like electricity, adopting an operating model of “unified scheduling, zonal pricing, and cross-regional trading,” and the institutional design of this model will profoundly affect the spatial pattern of China’s digital economy.

Conclusion

From “providing computing power” to “producing intelligence,” the transformation condensed in these seven words captures the core paradigm revolution of the computing power industry in the agent era. As Liu Jun said, competition in AI computing in the Agent era has shifted from single-point hardware performance contests to a contest of full-stack system capabilities and industrial ecosystems—“In the past, when customers bought computing power, they had to solve model deployment and performance tuning themselves; now customers want directly usable intelligence and a continuous, stable token supply.”

The deeper significance of this transformation is that it redefines the value chain of the computing power industry: chips are no longer the endpoint, but the starting point; computing centers are no longer the purpose, but the means; the real value lies in converting computing power into intelligence, and then converting intelligence into productivity. When agents move from conversational tools to production practice, and when computing power becomes a factor of production that can be produced at scale like water and electricity, the significance of this transformation will far exceed the computing power industry itself, profoundly reshaping how thousands of industries operate and compete. For enterprises, the question is no longer “whether to use AI,” but “how to obtain the highest-quality intelligence at the lowest cost”; for cities, the question is no longer “whether to build computing centers,” but “how to make computing power truly serve industrial upgrading”; for countries, the question is no longer “whether to develop computing power,” but “how to build a sustainable autonomous advantage in global computing power competition.” The answers to these questions will determine the map of the digital economy over the next decade.