No one anticipated that the CPU would become the next core player in AI computing power.

Meta's agent Muse quickly topped overseas application rankings, and the capital market has completed a narrative shift in a short period. The AI computing power pricing logic dominated by GPUs for two consecutive years is loosening. The task orchestration, tool scheduling, and cloud virtual machine load brought by AI agents have shifted server CPUs from a supporting role in computing power to a core bottleneck. Amid the market boom, industrial dividends are not universally rising. The supply chain fulfillment rhythm and commercial closed-loop capability will determine how much cake players can get in this round of the Agent wave.

When the market is still repeatedly gaming the GPU supply-demand relationship and the capital expenditure cycle of major manufacturers, the breakout of Meta Muse has brought a brand-new product paradigm and industrial logic to the AI industry.

Different from traditional passive-response conversational large models, Muse is an AI agent that can operate autonomously, capable of independently completing chained complex tasks such as web browsing, form filling, appointment booking, and cross-platform linkage, and can continue executing instructions in the background even after the user exits the program.

The rapid popularity of this phenomenal product is not only a breakthrough in C-end AI applications, but also directly ignites the strong expectation of the secondary market for AI computing power structure restructuring. The full takeoff of this round of AI market, with CPU, AI applications and computing power infrastructure all strengthening collectively, is essentially not a rotation of theme speculation, but a revaluation of industrial value brought by the iteration of AI load paradigm.

Restructuring of Computing Power Logic: From GPU's Solo Performance to CPU-GPU Collaborative Symbiosis

Over the past two years, the core computing power narrative of the AI industry has always revolved around GPUs. In the scenarios of large model training and basic conversational inference, the core demand is massive parallel matrix operations, and GPUs have become the absolute core with their super parallel computing capability.

Under this system, CPUs have long been reduced to a "supporting role", only responsible for basic work such as simple data distribution, equipment scheduling and basic control. The general computing power ratio in the industry has long been stable at 1 CPU corresponding to 4 to 8 GPUs, and the market has also habitually regarded GPU as the only core main line in the AI computing power track.

However, the large-scale implementation of AI agents has completely subverted this set of computing power allocation logic.

When an agent completes a full complex task, model generation only accounts for a very small part of the computing power consumption, and more computing power resources are used in serial general computing scenarios such as task planning, logical branch judgment, multi-round API tool invocation, web page rendering, permission verification, and task state persistence.

For this kind of refined, multi-step and highly logic-oriented tasks, the parallel computing advantage of GPU cannot be exerted at all, but it highly depends on the general computing capability of CPU. CPU has officially upgraded from an "auxiliary controller" to the core scheduling hub of the AI agent task link.

According to calculations by multiple industry institutions, with the continuous iteration and popularization of Agent products, the computing power ratio of CPUs and GPUs in AI data centers will continue to converge, gradually evolving from the traditional 1:4~1:8 to 1:1 or even a direction with a higher CPU proportion. In the exclusive scenario of agents with high concurrency and multi-tool linkage, CPU resource consumption has comprehensively exceeded that of GPU.

Relevant data from TrendForce shows that in traditional AI data centers, each gigawatt of power corresponds to about 30 million CPUs, while in the agent era, this figure is expected to soar to 120 million, and the increase in CPU computing power demand is nearly three times.

This is also the core industrial logic for the collective outbreak of the CPU sector in this round: the market no longer speculates on the GPU computing power gap in a single way, but trades in advance the long-term dividend of "computing power structure imbalance in the Agent era", and the value of supporting tracks such as CPU, memory, high-speed interconnection and complete servers is comprehensively revalued.

It needs to be clarified that the rising value of CPU does not mean that GPU is out of the stage. The optimal architecture of current AI agents is a heterogeneous collaboration mode where GPU is responsible for model generation and CPU is responsible for task orchestration. Insufficient CPU computing power supply will directly lead to idling and waste of high-end GPU resources. The balanced allocation capability of the computing power cluster has already become the core assessment indicator for leading cloud vendors and AI enterprises.

Three Levels of Industrial Chain Dividends: Differentiated Fulfillment Rhythm, Implemented Orders as the Only Touchstone

Following the transmission link of computing power, the dividends of this round of AI market can be clearly divided into three levels: core hardware, infrastructure support and application implementation. The three levels have completely different certainty of fulfillment, profit cycle, valuation elasticity and fluctuation risk, and the industrial dividends present obvious stepwise release characteristics.

Level 1: Server CPU

As the direct beneficiary target of this round of computing power restructuring, server CPU is the track with the most certain industrial dividend and the fastest implementation speed at present.

In the overseas supply chain, AMD, with its dual layout of EPYC server CPU + MI series GPU, is deeply bound to Meta's core computing power procurement system and becomes the core beneficiary target of this market; ARM and Qualcomm have developed dedicated server CPUs for AI agents in a targeted manner to seize the opportunity of the next-generation data center architecture upgrade; Intel continues to iterate its Xeon product line to fully adapt to the explosive general computing power demand in the agent scenario.

The domestic market focuses on the main line of domestic independent computing power. Leading manufacturers such as Haiguang Information, Loongson Technology and China Great Wall not only benefit from the policy dividend of independent and controllable domestic computing power infrastructure, but also continuously undertake the expansion orders of AI inference and agent computing power clusters from domestic cloud vendors, with clear performance fulfillment logic.

The core threshold of this track is not concept speculation, but the access qualification of the leading supply chain. Whether it can pass the strict stability, virtualization and energy efficiency ratio tests of major manufacturers and obtain verifiable batch implementation orders is the core standard to distinguish real growth targets from pure theme targets. Targets without order support will most likely present a pulse market and lack sustainability.

Level 2: Computing Power Infrastructure

The large-scale expansion of CPU computing power clusters will drive the demand explosion of the entire computing power hardware industrial chain. Supporting tracks such as server ODM complete machines, high-speed optical modules, memory chips, storage media, power supply and heat dissipation will usher in a mid-term performance release cycle.

The core characteristics of this type of "seller of shovels" track are strong performance certainty, stable cash flow and moderate valuation elasticity. Compared with chip design enterprises, the performance of hardware supporting vendors follows the implementation of capital expenditure of major manufacturers, with a more logical logic, but there is a certain lag. It belongs to the core direction of the second relay of the market and is the focus of institutional capital layout for the mid-term AI market.

Level 3: AI Agent Application

The AI application layer is the track with the greatest imagination but the highest uncertainty in this round of market, and it is also the core foothold of the long-term industrial value in the future. The overall pattern presents "giants occupying positions, vertical sectors breaking through, and most falling into involution".

One category is native C-end agents of overseas technology giants such as Meta and Google, which rely on massive traffic entrances to achieve rapid popularization and have extremely strong capability to reach users, but they have the pain points of large R&D investment, high computing power cost and difficulty in making profits in the short term, and are currently in the stage of burning money to expand the market.

The other category is AI agents in vertical industries, focusing on segmented scenarios such as enterprise office, e-commerce automation, industrial data retrieval, fiscal and tax process optimization, and government services. Their core value lies in reducing costs and increasing efficiency for enterprises, and their commercial implementation logic is more solid.

However, the industry generally has the core pain point of easy to make Demo but difficult to monetize. At present, the vast majority of AI agent products are still in the function demonstration stage, and user willingness to pay, SaaS charging mode, service authority boundary and data compliance system are not mature. The short-term application market is dominated by emotional premium. Only vertical application vendors that have successfully run through the payment closed-loop and solved the real rigid demand of the industry can lock in long-term dividends, and most general-purpose applications will fall into homogeneous involution.

From the perspective of capital pricing rhythm, the market logic is clear and fixed: first trade the expectation of CPU chip orders, then implement the performance of hardware infrastructure, and finally screen AI applications with mature commercialization.

Risk Boundary: Expectation Goes Ahead, Industrial Implementation Speed Lags Behind Market Boom

The boom of this round of AI market is rising comprehensively, but the valuation iteration speed of the secondary market has far exceeded the actual industrial implementation rhythm, and multiple potential risks cannot be ignored.

The first is the risk of overdrawn valuation expectation. This round of market is catalyzed by the user explosion of Meta Muse, which is a typical advance trading of long-term demand. The current market has fully overdrawn the expectation of agent computing power expansion. If the subsequent growth of agent users slows down and Meta and leading cloud vendors lower their capital expenditure guidelines, high-level computing power targets may usher in a rapid valuation correction.

The second is the risk of industry supply surplus and price war. The four overseas giants AMD, Intel, ARM and Qualcomm are competing on the same stage, and domestic domestic CPU manufacturers continue to iterate technology and release production capacity. In the next 2-3 years, the supply of the server CPU track will explode intensively, and price competition will most likely appear in the industry, squeezing the gross profit margin of enterprises and accelerating the industry reshuffle.

The most core industrial hidden worry is that AI agents are still in the early budding stage. At present, agents generally have problems such as low reliability of multi-step task execution, high error rate of tool invocation, high risk of data privacy compliance and insufficient user trust. From Demo-level products to large-scale commercial use and universal popularization, long-term technical polishing and scenario verification are still required, and the industrial implementation speed will most likely fail to keep up with the market speculation rhythm.

Conclusion: In the Era of Balanced Computing Power, Dividends Are Only Reserved for Players with Delivery Capability

The explosive popularity of Meta Muse has officially kicked off the industrial curtain of the AI Agent era. The two-year-long unilateral computing power narrative dominated by GPU has completely ended, and the AI industry has officially entered a brand-new stage of CPU-GPU heterogeneous collaboration and balanced development of computing power, with server CPU successfully returning to the core stage of the computing power architecture.

But the dividends of the AI industry are never evenly distributed. The short-term market competition focuses on supply chain access and order fulfillment capability, the mid-term market relies on the continuous implementation of computing power infrastructure capital expenditure, and the long-term value depends on the commercial closed-loop capability of AI applications.

The AI industry has bid farewell to the extensive development stage of "competing for parameters and concepts", and entered the era of systematic engineering competition where computing power architecture, software orchestration, product experience and commercial realization compete in all dimensions. Under the noisy market, only players who abandon theme speculation, identify the reality of industrial implementation, deeply cultivate technology and have real delivery capabilities can truly gain the long-term dividends of this round of AI agent wave.

【Risk Warning】 This article is only an in-depth analysis of industrial logic and does not constitute any investment advice. The AI sector market fluctuates greatly, and there is uncertainty in industrial implementation. Please treat the market boom and industry development rhythm rationally.

This article is from "ChaoXun Wang", author: Felix, published with authorization from 36Kr.