According to the latest AI server industry research from TrendForce, as AI service providers ramp up their Agent-related applications, demand for server CPUs has climbed steadily since the second half of 2025, with shipment growth projected to reach 31% year-on-year in 2026.

In 2027, continued growth in AI inference demand and rising penetration of Agentic AI services are expected to push server CPU shipments up by more than 40% year-on-year once again. The market landscape long dominated by the x86 camp, however, may be about to change.

In a market dominated by the GPU narrative, this news barely caused a ripple.

But read this forecast through the lens of procurement and capital expenditure, and another layer of change comes into view: the budget structure of AI infrastructure is shifting from buying accelerator cards alone to equipping a complete rack of compute, and the CPU's weight in this new ledger is clearly rising.

01 One Account Has Grown from a Single Line into Three Lists

In the era dominated by large-model pretraining, when a training server was procured, the budgeting and review almost invariably centered on GPUs. CPU, memory, storage, and networking were largely treated as supporting foundations, deployed according to established topology specifications, and rarely did anyone need a separate justification for adding a few more CPUs.

The industry long followed the classic ratio of one CPU to eight GPUs. In a ledger measured by floating-point compute, the CPU's computing power was almost negligible.

But Agent workloads have upended this arithmetic.

In a blog post published on April 24, TrendForce reported that during the Q1 2026 earnings call, Intel CEO Lip-Bu Tan said that as the focus of AI workloads shifts from training to inference, the CPU-to-GPU ratio is being rebalanced from 1:8 toward 1:1, driving a surge in CPU demand.

Industry data shows that in the AI agent era, the number of CPU cores required per gigawatt may surge from 30 million to 120 million—a fourfold increase.

This does not mean CPU floating-point compute has caught up with GPUs. The essence is a shift in workload structure: intensive matrix computations such as model token generation remain better suited to GPUs, while logic-heavy work involving extensive branching, external interaction, and coordination—such as task planning, request routing, tool invocation, and scheduling among multiple agents—is at this stage more often handled by CPUs.

The industry is, of course, also trying to push some lightweight orchestration logic down to the accelerator side, but this has yet to become mainstream.

KeyBanc analysts said that AMD EPYC server CPU capacity for the full year 2026 is nearly sold out, with lead times for high-end models extended to 8–10 weeks.

The specifications of Nvidia's own rack products also indirectly confirm that the logic of compute allocation is shifting: the GB200 NVL72 and the Vera Rubin NVL72 aimed at Agent scenarios both come standard with 36 CPUs paired with 72 GPUs.

These, of course, are Grace/Vera Arm cores deeply bound to GPUs and cannot be directly equated with standalone x86 host CPUs on procurement lists. But they represent a shift in platform design thinking: in Agent and long-context inference scenarios, control and orchestration resources are no longer accessories.

The scale of CPU configuration within a single rack has risen sharply, and system memory has expanded accordingly. To support network interaction for multi-node concurrent scheduling and tool invocation, high-speed network ports must increase in tandem. As overall rack power consumption continues to rise, power supply capacity and liquid cooling solutions must all be recalculated.

When it comes to cloud vendors' procurement, the change is already very clear: in the early days of building training clusters, procurement was basically equivalent to deciding how many GPUs to buy—a single-variable multiple-choice question.

But for Agent-oriented inference clusters, the procurement list has lengthened into a complete system solution: GPU compute, CPU core count, system DRAM, KV cache SSD, high-speed switches, rack power distribution, and cooling capacity. Every link must be matched and planned as a whole, and no single hardware metric can any longer determine a cluster's delivery capability.

In the past, when calculating the TCO of AI infrastructure, the budget focus was almost anchored on GPUs, while CPUs were mostly accessory items that followed the overall system solution and could be appropriately compressed. Under the new Agent-driven workloads, CPUs have been upgraded to a procurement category 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 vendors and large-model companies, it means an additional cost variable to weigh and a new bargaining lever.

02 A Branch Line the Market Has Undervalued

Underlying the incremental CPU demand, three types of buyers with completely different demands have already differentiated.

One type is cloud vendors deeply binding their own software stacks and self-developed Arm compute adapted to Agent workloads; one type is general-purpose compute with flexibly adjusted CPU-GPU ratios for AI-native startups and model customers; and another type is x86 compute for traditional enterprise customers that prioritizes ecosystem compatibility.

Different routes correspond to very different workload demands, migration costs, and bargaining space, and will also affect who ultimately captures the incremental dividends.

Previously, valuation pricing in the AI track was basically anchored in the GPU industry chain: HBM, advanced packaging, optical modules, liquid cooling, and rack integration—each segment produced high-valuation targets. CPUs long remained outside the main AI narrative, and the stock price elasticity of Intel and AMD overall lagged behind targets in the GPU upstream and downstream.

This pricing logic is gradually shifting as Agent workloads restructure compute allocation.

If TrendForce's forecast materializes, the server CPU industry chain is expected to embark on a long-awaited growth curve. And where this round of incremental dividends ultimately flows will be a core question for continuous analysis in the capital market.

One possible distribution path is that the Arm camp captures the bulk of this round of incremental growth. Cloud vendors' self-developed CPUs such as AWS Graviton, Alibaba Cloud Yitian, and Google Axion, together with Grace and Vera in Nvidia's rack platforms, jointly drive up shipment scale of Arm-based server CPUs, in turn pulling demand for supporting foundry, advanced packaging, and memory.

For TSMC and packaging and testing companies, this is a new demand curve independent of GPU orders, but the process and packaging resources it consumes do not completely overlap with those for high-end accelerator cards.

Another distribution path is that the x86 camp relies on its solid existing base to defend share for a longer time in general-purpose inference and enterprise hybrid cloud scenarios. Intel Xeon 6 and AMD's sixth-generation EPYC have both been optimized specifically for Agent workloads.

Mature software ecosystems, channel networks spanning industries, and long-established enterprise customer service links are shortcomings that Arm solutions will find difficult to make up for in the short term. In this scenario, even if Agent-related workloads are deployed at scale, x86 may not necessarily lose ground quickly.

There is also a supply chain transmission effect that is easily overlooked: rising CPU demand may reshape the balance of power in the server system sector.

In the stage dominated by AI training clusters, whether one could secure GPU allocations was often the core prerequisite for system vendors to win orders. As the priority of CPUs, memory, power supplies, and cooling in procurement lists rises in tandem, system vendors and integrators have an opportunity to return from being simply labeled assemblers by the outside world to solution designers.

The customized requirements of hybrid Agent workloads for system topology, resource allocation, and power balance will amplify the value of system-level design.

Of course, there is another risk worth noting: if the capital market prices relevant targets in advance according to a CPU super-cycle, but the pace of Agent commercialization falls short of expectations, valuations could retract as quickly as they previously rose.

This growth branch line currently remains at the expectation level and has not yet been fully reflected in performance. Relevant pricing should be anchored more to verifiable shipment and deployment data rather than pure narrative.

03 X86's Rent, Arm's Increment

The landscape of the server CPU market over the past decade or more has been built to a large extent on a kind of ecosystem rent.

A large number of databases, middleware, and industry business systems have long been developed around this architecture, and the migration cost for customers switching architectures is high, constituting a long-term stable barrier. Of course, the premise of rent is that vendors continue to invest in chip architecture and platform capability iteration, rather than simply relying on historical accumulation to enjoy the dividends.

Intel and AMD have long held nearly 90% of the market share, and this does not come from an absolute crushing of chip performance. The real barrier is the complete software ecosystem that has grown around x86: the vast majority of enterprise software, databases, middleware, and operations toolchains worldwide have long used x86 as the default adaptation baseline, and cross-architecture migration entails considerable transformation and validation costs.

Once customers switch chip architectures, they must directly face the pressure of compatibility transformation across the entire software stack. High migration costs form x86's moat and are the underlying support for this ecosystem rent.

Arm has not failed to mount challenges. From Calxeda in the early years to Ampere later, Arm server chips have been tried for more than a decade, yet it has always been difficult to achieve large-scale deployment in the traditional enterprise market. The core obstacle is the high wall of the software ecosystem built by x86.

Of course, early attempts were also constrained by multiple factors such as hardware, funding, and customer deployment pace, and ecosystem migration cost remains the hardest link to overcome.

But the path to breakthrough in this round is different: instead of directly attacking the traditional enterprise market with its heavy existing base, it starts from scenarios such as cloud vendors or AI-native customers that can control the entire software stack themselves. Such entities are able to adapt based on Arm from the outset of their business and do not need to pry open an already solidified installed application ecosystem.

For example, cloud vendors deeply customize operating systems, scheduling components, and runtime environments at the underlying layer, absorbing architecture adaptation and compatibility validation costs internally, thereby bypassing the external ecosystem shortcomings that Arm has long faced.

Nvidia has taken another path: it does not entirely refrain from offering standalone Arm CPUs externally, but in whole-rack solutions such as GB200 and Vera Rubin NVL72, Grace and Vera are deeply coupled and delivered together with GPUs in the form of Superchips.

What customers purchase is a complete AI rack optimized for Agent workloads, and the shipment scale of these Arm CPUs is directly tied to the delivery volume of GPUs in the rack. The process of iterating from GB200 to Vera Rubin across two rack generations and ramping orders also drives synchronous growth in shipments of supporting Arm CPUs.

It can be seen that the market is moving toward differentiation: x86 holds the existing ecosystem base, while Arm opens new growth space through the incremental market brought by Agents.

The 31% and 40% growth rates look attractive, but breaking down this forecast reveals that it rests on three sets of assumptions that have not yet been verified in practice.

The first assumption is that Agents can achieve large-scale commercialization. At present, cloud vendors and large-model companies are densely launching Agent-related products, but users' willingness to pay, end-to-end task completion rates, and long-term invocation scale by enterprise customers have not yet been tested through a complete business cycle.

If Agents remain for a long time in a state where demo effects are impressive but paid conversion is weak, the incremental space for server CPUs will shrink accordingly, and the narrative of Arm expanding through this track will also come under pressure.

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

The third assumption is that the fragmentation problem of the Arm ecosystem will not drag down the pace of 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 different platforms.

Within cloud vendors' closed proprietary ecosystems, such adaptation costs can be absorbed internally; but when Arm solutions move outside the closed loop and are promoted to external ISVs and traditional enterprise customers, fragmentation will translate into clear migration and maintenance costs.

For industry investors, a more valuable approach is to watch three sets of leading high-frequency indicators: changes in the CPU-to-GPU ratio in cloud vendors' hardware procurement lists, the real invocation scale and paid conversion efficiency of Agent products, and changes in the share of Arm-based server CPUs in quarterly shipments.

Frontline orders and business operations data reflect the true state of industry prosperity far more accurately than any institution's long-term forecast.

Placing TrendForce's forecast in a business context, it essentially points to a straightforward shift: the logic of AI compute investment is moving away from purchasing only accelerator cards toward equipping a complete, balanced compute rack.

GPUs remain the largest expenditure item in compute procurement, but CPUs, memory, networking, and power supply—components previously regarded as supporting elements—are beginning to gain independent weight in budget allocation.