AI 컴퓨팅 투자가 가속기 카드 중심에서 균형 잡힌 컴퓨팅 랙 중심으로 옮겨가고 있으며, AI 서버에서 CPU가 더 이상 부차적인 구성 요소가 아니라는 분석이 제기됐다.

According to TrendForce's latest research on the AI server industry, demand for server CPUs has grown steadily since the second half of 2025 as AI service providers roll out AI agent applications. Shipments are forecast to rise 31% year over year in 2026. In 2027, as AI inference demand keeps growing and agentic AI services penetrate further, server CPU shipments could grow by more than 40% year over year again, although the market structure, currently dominated by the x86 camp, may shift.

In a market dominated by the GPU narrative, this forecast has made little splash. Read from the perspective of procurement and capital expenditure, however, it points to a further change: the budget structure of AI infrastructure is moving from buying accelerator cards alone to equipping a complete rack of compute. The weight of CPUs in this new ledger is rising noticeably.

**One Line Item Becomes Three Checklists**

In the pretraining-dominated phase of large models, budget and review attention for a training server were centered almost entirely on GPUs. CPUs, memory, storage, and networking were treated as supporting infrastructure, deployed to established topology specifications, and rarely justified separately on the question of whether a few more CPUs were needed.

The industry has long used the classic ratio of one CPU serving eight GPUs. In a ledger measured by floating-point performance, the CPU's computing power is almost negligible.

Agent workloads change this arithmetic.

On April 24, TrendForce published a blog post reporting that on Intel's first-quarter 2026 earnings call, CEO Lip-Bu Tan said that as AI workloads shift their center of gravity from training to inference, the CPU-to-GPU ratio is rebalancing from 1:8 to 1:1, driving a surge in CPU demand. Industry data suggests that in the AI agent era, the number of CPU cores required per gigawatt could jump from 30 million to 120 million, a fourfold increase.

This does not mean CPUs have caught up with GPUs in floating-point performance. Rather, the nature of the workload has migrated. Dense matrix computation, such as token generation in models, remains better suited to GPUs. Logic-heavy work involving extensive branching, external interaction, task planning, request routing, tool invocation, and coordination and scheduling among multiple agents is currently handled mostly by CPUs. Some in the industry are trying to move lightweight orchestration logic onto the accelerator side, but this has not yet become mainstream.

Investment bank KeyBanc's analysts said that AMD EPYC server CPU capacity for 2026 is nearly sold out, with lead times for high-end models stretching to 8–10 weeks.

Nvidia's own rack-level products also indirectly confirm the shift in compute ratios. The standard configuration of the GB200 NVL72, and of the Vera Rubin NVL72 designed for agent scenarios, pairs 36 CPUs with 72 GPUs. These are Grace and Vera Arm cores tightly coupled with the GPUs, so they cannot be equated with the standalone x86 host CPUs on a procurement list. Still, they represent a shift in platform design philosophy: in agent and long-context inference scenarios, control and orchestration resources are no longer an afterthought.

CPU configurations per rack rise sharply, and system memory must expand in step. Multi-node concurrent scheduling and tool-call networking require more high-speed network ports. Overall power consumption keeps climbing, so power supply capacity and liquid cooling solutions must be recalculated.

For cloud providers, the change in procurement is now clear. In the early days of building training clusters, procurement essentially meant deciding how many GPUs to buy, a single-choice question driven by one core variable. For agent-oriented inference clusters, however, the procurement list has expanded into a complete system solution: GPU compute, CPU core count, system DRAM, KV cache SSDs, high-speed switches, and rack power distribution and cooling capacity. Every element must be matched and planned holistically, and no single hardware metric can determine a cluster's delivery capability.

In the past, when accounting for AI infrastructure TCO, budget focus was anchored almost entirely on GPUs, and CPUs were a secondary item that followed the overall system plan and could be trimmed. Under the new agent-driven workloads, CPUs have been upgraded to a procurement line item that requires independent specification evaluation, separate production scheduling, and separate commercial negotiation.

For chip and server suppliers, this is a clear incremental market. For buyers such as cloud providers and large-model companies, it adds a cost variable to weigh and a new lever in price negotiations.

**An Undervalued Side Track**

The underlying demand for incremental CPUs has split into three buyer groups with very different needs.

The first is Arm compute that cloud providers deeply tie to their own software stacks and custom-adapt for agent workloads. The second is general-purpose compute for AI-native startups and model customers, where the CPU-to-GPU ratio is adjusted flexibly. The third is x86 compute for traditional enterprise customers, who prioritize ecosystem compatibility.

Workload requirements, migration costs, and bargaining power differ greatly across these routes, and they will shape who ultimately captures the incremental windfall.

Until now, valuations in the AI sector have been anchored mainly to the GPU supply chain: HBM, advanced packaging, optical modules, liquid cooling, and rack integration each produced high-valuation names. CPUs have remained largely outside the mainstream AI narrative, and the stock price elasticity of Intel and AMD has generally lagged GPU-related upstream and downstream companies.

As agent workloads restructure compute ratios, this pricing logic is gradually beginning to shift.

If TrendForce's forecast materializes, the server CPU industry could see a long-absent growth curve. Which players capture this round of incremental returns will be a central question for capital markets to keep assessing.

One possible allocation path is that the Arm camp takes the bulk of this round's increment. In-house CPUs from cloud providers, such as AWS Graviton, Alibaba Cloud's Yitian, and Google Axion, combined with the Grace and Vera chips in Nvidia's rack platforms, could together push up shipments of Arm-architecture server CPUs and drive demand for associated foundry, advanced packaging, and memory. For TSMC and packaging and testing providers, this would be a new demand curve separate from GPU orders, although the process and packaging resources it consumes do not fully overlap with those used for high-end accelerator cards.

The other path is that the x86 camp, relying on its entrenched installed base, holds its share for longer in general-purpose inference and enterprise hybrid cloud scenarios. Intel Xeon 6 and AMD's 6th-generation EPYC have both been specifically optimized for agent workloads. The mature software ecosystem, extensive channel networks, and long-established enterprise customer service chains are shortcomings that Arm solutions will struggle to close in the short term. In this scenario, even if agent-related workloads are deployed at scale, x86 may not quickly lose ground.

There is also a supply chain ripple effect that is easy to overlook: rising CPU demand may reshape influence within the server systems segment. During the phase when AI training clusters dominated, securing GPU allocation was often the core prerequisite for server makers to win orders. As CPUs, memory, power supplies, and cooling rise in priority on procurement lists, server makers and system integrators have a chance to shift from being labeled simple assemblers to becoming solution designers.

The customization demands that mixed agent workloads place on system topology, resource ratios, and power balance will amplify the value of system-level design.

There is also a risk worth monitoring. If capital markets price related stocks ahead of a CPU super-cycle, but the commercialization of agents proceeds more slowly than expected, valuations could retreat as quickly as they rose.

This growth side track remains, for now, at the level of expectation and has not been fully reflected in earnings. Valuations should be anchored to verifiable shipment and deployment data rather than to narrative alone.

**The x86 Rent, the Arm Increment**

For more than a decade, the server CPU market's structure has rested largely on an ecosystem rent.

A large number of databases, middleware, and industry business systems have long been developed around this architecture. The high cost for customers to switch architectures creates a durable, stable barrier. That said, the rent depends on vendors continuing to invest in chip architecture and platform capabilities, rather than simply living off historical accumulation.

Intel and AMD have long held close to 90% of the market, and this is not due to an absolute performance advantage in chips. The real barrier is the complete software ecosystem built around x86: most enterprise software, databases, middleware, and operations toolchains worldwide have long used x86 as the default baseline, and migrating across architectures requires substantial rework and validation costs.

Once a customer switches chip architectures, it must face the compatibility overhaul of the entire software stack. High migration costs form x86's moat and underpin this ecosystem rent.

Arm has not been absent from challenges. From early attempts like Calxeda to later efforts like Ampere, Arm server chips have been tried for more than a decade, yet they have never achieved large-scale adoption in the traditional enterprise market. The core obstacle has been the software ecosystem wall built by x86.

Early attempts were also constrained by hardware, capital, and customer adoption pacing, but ecosystem migration cost has remained the hardest hurdle to clear.

The path for this round of breakthroughs is different. Rather than frontally attacking the entrenched traditional enterprise market, Arm is entering through cloud providers and AI-native customers, who control their entire software stack and can adapt to Arm from the outset of their businesses, without needing to dislodge an established legacy application ecosystem.

For example, cloud providers can deeply customize the operating system, scheduling components, and runtime environments at the infrastructure level, absorbing the costs of architecture adaptation and compatibility validation within their own systems. This sidesteps the external ecosystem shortcomings that Arm has long faced.

Nvidia has taken a different route. It does not entirely refuse to offer standalone Arm CPUs, but in rack-scale NVL72 solutions such as GB200 and Vera Rubin, Grace, Vera, and the GPUs are deeply integrated and delivered as Superchips.

Customers buy an AI rack optimized for agent workloads, so the shipment volume of these Arm CPUs is directly tied to the GPU deliveries within the rack. As the GB200 and Vera Rubin rack generations iterate and orders ramp up, shipments of the associated Arm CPUs grow in tandem.

The market is clearly diverging: x86 holds the base of its established ecosystem, while Arm opens new growth space through the incremental market created by agents.

Growth rates of 31% and 40% look attractive, but a closer look at the forecast shows that it rests on three assumptions that have not yet been validated.

The first assumption is that agents can achieve large-scale commercial deployment. Cloud providers and large-model companies are currently rolling out agent-related products at a rapid pace, but user willingness to pay, end-to-end task completion rates, and the long-term calling volume of enterprise customers have not yet been tested through a full commercial cycle.

If agents remain impressive in demonstrations but weak in paid conversion, the incremental room for server CPUs will shrink accordingly, and the narrative of Arm expanding on the back of this track will also come under pressure.

The second assumption is that the CPU-to-GPU resource ratio will stabilize at a high-CPU-consumption range. Agent workloads currently do impose significant host-side overhead, but iterations in inference frameworks and underlying architectures could reshape workload distribution at any time.

The third assumption is that fragmentation within the Arm ecosystem will not hold back expansion. Different Arm platforms, such as Graviton, Axion, Ampere, and Grace/Vera, are not fully compatible with one another, and software often needs separate adaptation and tuning for each platform.

Within the closed ecosystems of cloud providers, these adaptation costs can be absorbed internally. But when Arm solutions move beyond those closed loops and are promoted to external ISVs and traditional enterprise customers, fragmentation turns into explicit migration and maintenance costs.

For industry investors, the more useful approach is to track three leading, high-frequency indicators: changes in the CPU-to-GPU ratio in cloud providers' hardware procurement lists, the real calling volume of agent products and the efficiency of their paid conversion, and the share of Arm-architecture server CPUs in quarterly shipments.

Frontline orders and business operating data are far closer to real industry conditions than any institution's long-range forecast.

Placing TrendForce's forecast in a commercial context, its essential message is a simple one: AI compute investment is shifting from buying accelerator cards alone to pairing them with a balanced rack of compute.

GPUs remain the largest expense in compute procurement. But CPUs, memory, networking, and power, once treated as supporting components, are beginning to earn independent weight in budget allocation.