文｜硅碳变量，作者｜沈浪
根据TrendForce集邦咨询最新AI服务器产业研究，随着各AI服务厂商积极推出AI Agent 相关应用，服务器CPU需求自2025年下半年起稳步成长，预估2026年出货年增幅将达 31%。
2027年由于AI推理需求持续成长、Agentic AI服务渗透率提升，服务器CPU出货量有望再年增逾40%，不过以x86阵营主导的市场格局可能出现变化。
在GPU叙事霸屏的市场上，这条消息没有掀起太大浪花。
但站在采购与资本开支的视角重读这份预测，它指向另一层变化：AI基础设施的预算结构，正在从只买加速卡转向配齐一整套机柜算力，CPU在这本新账本里的权重正在明显抬升。
一笔账，从一行变成了三张清单
在大模型预训练为主的阶段，采购一台训练服务器，预算与评审的重心几乎都落在GPU上。CPU、内存、存储与网络更多被视作配套底座，按既定拓扑规格配套落地，很少会因为要不要多加几颗CPU单独论证。
行业长期通行1颗CPU承载8颗GPU的经典配比，在以浮点算力为标尺的账本里，CPU的计算能力几乎可以忽略。
但Agent负载把这道算术题改了。
集邦咨询Trendforce 4月24日发布博文，报道称在2026年第一季度财报电话会议上，英特尔首席执行官陈立武表示，伴随着AI工作负载重心从训练转向推理，推动CPU和GPU之间的比例关系平衡，从1:8向1:1转变，带动CPU需求激增。
行业数据显示，AI智能体时代每吉瓦所需的CPU核心数可能从3000万激增至1.2亿，增长4倍。
这并不意味着CPU的浮点算力追上了GPU，本质是负载结构发生了迁移：模型token生成这类密集矩阵计算，依然更适合交给GPU；而任务规划、请求路由、工具调用、多智能体之间的协调调度这类带有大量分支判断、外部交互的逻辑类工作，现阶段更多交由CPU承载。
当然，行业里也在尝试把部分轻量编排逻辑下沉到加速卡侧，只是尚未成为主流。
投行KeyBanc分析师表示，AMD EPYC服务器CPU 2026年全年产能近乎售罄，高端型号交期拉长至8–10周。
英伟达自家机架产品的规格，也侧面印证算力配比逻辑正在转向：GB200 NVL72以及面向 Agent场景的Vera Rubin NVL72，标准配置都是36颗CPU对应72颗GPU。
当然，这部分是和GPU深度绑定的Grace/Vera Arm核，和采购清单上独立的x86主机CPU 不能直接划等号，但它代表平台设计思路的转向：Agent、长上下文推理场景下，控制与编排资源不再是附带品。
单机柜内CPU配置规模大幅抬升，整机系统内存随之同步扩容；为支撑多节点并发调度与工具调用的网络交互，高速网络端口必须同步增加；整机功耗持续走高，供电容量、液冷散热方案都要重新核算。
落到云厂商采购环节，变化已经非常清晰：早年搭建训练集群，采购基本等同于确定采购多少张GPU，属于单一核心变量的单选题。
但面向Agent的推理集群，采购清单已经拉长为一整套系统方案：GPU算力、CPU核数、系统DRAM、KV缓存SSD、高速交换机、机柜配电与散热容量，所有环节必须整体匹配、统一规划，单一硬件指标已经无法决定集群交付能力。
过去核算AI基础设施TCO，预算重心几乎锚定在GPU，CPU更多是跟随整机方案配套、可适度压缩的附属项。而在Agent驱动的新负载下，CPU已经升级为需要独立评估规格、单独排产、单独商务议价的采购科目。
对芯片与服务器供应商来说，这是一块明确的增量市场；对云厂商和大模型企业这类采购方而言，它意味着多了一层需要权衡的成本变量，也新增了一个议价抓手。
一条被市场低估的支线
CPU增量需求的底层，已经分化出三类诉求完全不同的买家。
一类是云厂商深度绑定自家软件栈和自研适配Agent负载的Arm算力；一类是面向AI原生创业与模型客户、CPU-GPU配比灵活调整的通用算力；还有一类面向传统企业客户，优先考虑生态兼容性的x86算力。
不同路线对应的负载诉求、迁移成本以及议价空间差异很大，也会影响增量红利最终流向谁。
此前AI赛道的估值定价，重心基本锚定在GPU产业链：HBM、先进封装、光模块、液冷、机柜集成，每个环节都诞生了高估值标的。CPU长期游离在AI的主线叙事之外，Intel和AMD的股价弹性，整体不及GPU上下游的标的。
这套定价逻辑正在随Agent负载带来的算力配比重构，慢慢发生偏移。
倘若TrendForce的预判能够落地，服务器CPU产业链有望走出一条久违的增长曲线。而这一轮增量红利最终流向哪些玩家，会是资本市场持续研判的核心命题。
一种可能的分配路径是，Arm阵营拿下本轮增量里的大头。AWS Graviton、阿里云倚天、Google Axion这类云厂商自研CPU，叠加NVIDIA机柜平台中的Grace、Vera，共同推高Arm架构服务器CPU的出货规模，连带拉动配套代工、先进封装与内存需求。
对台积电和封测厂商而言，这是独立于GPU订单之外的一条新需求曲线，但它所消耗的工艺与封装资源和高端加速卡并不完全重合。
另一种分配路径则是x86阵营依托稳固的存量基本盘，在通用推理和企业混合云场景守住份额更长时间。Intel Xeon 6、AMD第六代EPYC，都已经面向Agent负载做针对性优化。
成熟的软件生态、遍布行业的渠道网络以及长期沉淀的企业客户服务链路，是Arm方案短期内很难补齐的短板。在这个情景里，即便Agent相关负载规模化落地，x86也不一定会快速丢失阵地。
还有一层容易被忽略的供应链传导效应：CPU需求抬升或将重塑服务器整机赛道的话语权格局。
在AI训练集群主导的阶段能不能拿到GPU配额，往往是整机厂商拿到订单的核心前提。而随着CPU、内存、电源、散热在采购清单里的优先级同步抬升，整机厂商与系统集成商的角色，有机会从被外界简单贴上标签的组装厂向方案设计者回归。
混合Agent负载对整机拓扑、资源配比以及功耗平衡的定制化要求，会放大系统层面的设计价值。
当然还有一层风险值得留意：倘若资本市场提前按照CPU超级周期给相关标的定价，但 Agent商业化落地节奏不及预期，估值回撤的速度可能和此前上涨一样迅速。
这条增长支线现阶段仍然停留在预期层面，尚未充分兑现到业绩里，相关定价更需要锚定可验证的出货与落地数据，而非单纯的叙事。
X86的租金，Arm的增量
服务器CPU市场十几年来的格局，很大程度上建立在一笔生态租金之上。
大量数据库、中间件、行业业务系统长期围绕这套架构开发，客户切换架构的迁移成本高昂，构成长期稳定的壁垒。当然，租金的前提是厂商持续投入芯片架构与平台能力迭代，并非单纯依靠历史积累坐享红利。
Intel与AMD长期拿下近九成市场份额，背后并不来自芯片性能的绝对碾压。真正的壁垒是围绕x86生长起来的完整软件生态：全球绝大多数企业软件、数据库、中间件与运维工具链，长期以x86作为默认适配基准，跨架构迁移需要承担不菲的改造与验证成本。
一旦客户切换芯片架构，就要直面整套软件栈的兼容改造压力。高昂的迁移成本构成了x86的护城河，也是这份生态租金的底层支撑。
Arm并不是没有发起过挑战，从早年的Calxeda再到后来的Ampere，Arm服务器芯片已经试了十几年，始终难以在传统企业市场形成规模化落地，其核心阻碍就是x86垒起的软件生态高墙。
当然，早期尝试还叠加硬件、资金、客户落地节奏等多重约束，生态迁移成本依然是最难翻越的一环。
但这一轮突破的路径已经不一样：不再正面冲击存量厚重的传统企业市场，转而从云厂商或AI原生客户这类可以自主把控整套软件栈的场景切入。这类主体有条件在业务起步阶段就基于Arm做适配，不必去撬动已经固化的存量应用生态。
比如，云厂商在底层做操作系统、调度组件与运行时环境的深度定制，把架构适配、兼容性验证的成本在体系内部消化，以此绕开Arm长期面对的外部生态短板。
而NVIDIA走出了另一条路径：它并非完全不对外提供独立Arm CPU，但是在GB200、Vera Rubin这类NVL72整机柜方案里，Grace、Vera和GPU以Superchip的形式深度耦合打包交付。
客户采购的是一套面向Agent负载优化的整机AI机架，这部分Arm CPU的出货规模直接和机柜内GPU的交付量绑定。GB200到Vera Rubin两代机架迭代、订单上量的过程，也同步带动配套Arm CPU的出货同步增长。
可以看出，市场正在走向分化：x86守住存量生态的基本盘，Arm借Agent带来的增量市场打开新的增长空间。
31%、40%的增幅看起来很有吸引力，不过拆解这份预测不难发现，它依托于三组还没有落地验证的假设。
第一个假设是Agent能够实现规模化商业化落地。当下云厂商与大模型企业都在密集推出Agent相关产品，但用户付费意愿、任务端到端完成率以及企业客户长期调用规模都还没有经历完整商业周期的检验。
倘若Agent长期停留在演示效果亮眼和付费转化偏弱的状态，服务器CPU的增量空间就会相应收缩，Arm依托这一赛道扩张的叙事也会随之承压。
第二个假设是CPU、GPU的资源配比会稳定在高CPU消耗的区间。现阶段Agent负载确实带来显著的主机侧开销，但推理框架与底层架构的迭代随时可能重塑负载分布。
第三个假设是Arm生态的碎片化问题不会拖累扩张节奏。Graviton、Axion、Ampere、Grace/Vera等不同Arm平台，彼此并不能做到完全兼容，软件往往需要针对不同平台单独适配调优。
在云厂商封闭的自有生态内，这类适配成本可以内部消化；但当Arm方案走出闭环，面向外部ISV与传统企业客户推广时，碎片化就会转化为明确的迁移与维护成本。
对产业投资者而言，更有价值的做法是盯住三组前置高频指标：云厂商硬件采购清单里 CPU与GPU的配比变动、Agent产品的真实调用规模与付费转化效率、Arm架构服务器 CPU在季度出货中的占比变化。
一线订单与业务运营数据，远比任何机构的远期预测更贴近产业真实景气度。
把TrendForce这份预测放回商业语境里，它本质指向一个朴素的变化：AI算力的投入逻辑正在从只采购加速卡，转向搭配一整套均衡的算力机柜。
GPU依旧是算力采购里最大的支出项，但CPU、内存、网络、供电这些此前被视作配套的环节，开始在预算分配中获得独立权重。
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
It's a vinyl silicon carbon variable.
According to the latest AI server industry study of TrendForce, with the active introduction of AI Agent-related applications by AI service providers, the demand for CPUs has grown steadily since the second half of 2025, with an estimated 31 per cent annual increase in departures in 2026.
In 2027, as a result of the growing demand for AI reasoning and the increased penetration of Agentic AI services, it is expected that the delivery of CPU from the server will increase by more than 40 per cent a year, although the market pattern led by the x86 camp may change.
In the GPU-backed market, the news did not raise much.
However, rereading the forecast from the perspective of procurement and capital expenditure points to a different level of change: the budget structure of AI infrastructure is shifting from purchasing only accelerators to completing a full set of machine counters, and the weight of CPU in this new book is clearly rising.
A check, from one line to three lists.
During the pre-training phase of LLM, the procurement of a training server, the budget and evaluation focus fell almost exclusively on GPU. CPUs, memory, storage and networks are more often seen as complementary sub-seases, which, in accordance with the established scaling-up specifications, rarely come to the ground because of the need to add additional CPUs to stand alone.
The industry has long adopted a classic CPU matching of eight GPUs, and the computing capacity of CPUs is almost negligible in the book with floating-point calculators.
But Agent's load changed the math.
In the first quarter of 2026, Chief Executive Intel Chen Zhiwo stated that, along with the shift in focus of AI from training to reasoning, promoting a balance in the ratio between CPU and GPU, from 1:8 to 1:1, driving a surge in CPU demand.
Industry data show that the core number of CPUs per giva in the AZAI agent era is likely to increase fourfold from 30 million to 120 million.
This does not mean that the CPU floats after GPU, essentially because the load structure has shifted: Model token generates such intensive matrix calculations, which are still more suitable for GPU, while coordinated task planning, request route, tool call, and multi-AI agent, such as logical type work with extensive branch judgement and external interaction, are more carried by CPU at this stage.
Of course, the industry is also trying to sink some of the light organization logic to the acceleration card side, although it is not yet in the mainstream.
According to the Bank's KeyBanc analyst, the AMD EPYC server CPU was almost out of production in 2026 and the high-end model was extended to 8-10 weeks.
The specifications of YVD's own racks, and the geometrical logic of side-provers, are moving towards: GB200 NVL72 and Vera Rubin NVL72 for Agent. The standard configuration is 36 CPUs for 72 GPUs.
Of course, this part of the Grace/Vera Arm, which is deeply bound to the GPU, and the separate X86 host CPU on the procurement list, cannot be assigned the same number directly, but it represents a shift in the design thinking of the platform: Agent, long context reasoning, and control and organization of resources is no longer an accessory.
The size of the CPU configuration in the stand-alone cabinet has been substantially increased, with the simultaneous build-up of the entire system; the high-speed network port has to be increased in order to support multinodes and to interact with the tool call network; and the overall capacity of the power supply and the liquid cool-dispersion programme has to be recalculated.
The change has become very clear when it comes to the cloud manufacturer ' s procurement chain: in the early years, the training cluster was built, and the procurement was essentially the same as determining how many GPUs were procured, which was the single subject of a single core variable.
However, the reasoning cluster for Agent has been extended to a full set of system options: GPU calculus, CPU numbers, system DRAM, KVcacheSSD, high-speed switches, machine cabinets, and heat-dispersion capacity, all of which must be fully matched and planned, and single hardware indicators are no longer able to determine cluster delivery capacity.
In the past, IAI infrastructure TCO was accounted for, with a budget focus almost anchored in GPU, and CPU was more of a moderately condensable subsidiary accompanying the whole machine programme. Under the new load driven by Agent, the CPU has been upgraded to the procurement subject of the need for independent assessment specifications, stand-alone release, and stand-alone commercial bargaining.
For chip and server suppliers, this is a clear incremental market; for buyers such as cloud manufacturers and LLM enterprises, it means an additional set of cost variables to be weighed and a new bargaining power added.
A feeder that the market underestimated.
At the bottom of the CPU incremental demand, three different types of buyers have been divided.
One category is the Arum calculus for cloud manufacturers to bind their own silos to depths and self-study to fit the Agent load; another is the generic calculus for Ai's original entrepreneurship and model customers, CPU-GPU matching with flexibility; and another one is for traditional business customers, with priority given to eco-compatibility x86.
The wide variations in load claims, relocation costs and bargaining space associated with different routes can also affect the ultimate flow of incremental dividends.
Prior to the valuation pricing of the AI track, the centre of gravity was anchored in the GPU industrial chain: HBM, advanced seals, photomodules, liquid cooling, and cabinet integration, each of which produced a high valuation mark. CPU has long been away from AI ' s main line narrative, and the stock price elasticity of Intel and AMD is not as large as the up and down of the GPS.
This pricing logic is slowly shifting with the weighting structure of the calculus of the Agent load.
If TrendForce's prognosis could land, the CPU industry chain on the server was expected to emerge from a long-standing growth curve. This round of incremental dividends to which players will ultimately flow will be at the heart of ongoing capital market research.
One possible distribution path is that the Arm camp takes the big part of the increase. AWS Graviton, Alibaba Cloud, Google Axion, and others, such as cloud manufacturers, studied CPU, superseding Grace, Vera, from the NVIDIA cabinet platform, and jointly pushed up the volume of CPU from the Arm architecture server, along with the pull-up support, advanced sealing and memory requirements.
This is a new demand curve independent of the GPU order for the build-up and seal manufacturers, but the process it consumes does not fully overlap with the containment resources and high-end accelerators.
Another distribution path is for x86 camps to hold their shares longer in common reasoning and in a mixed cloud landscape. Intel Xeon 6 and the sixth generation of AMD EPYC have been targeted and optimized for Agent ' s load.
A mature software ecology, a network of industry-wide channels and a long-term settled business customer service link are short boards that are difficult to fill in the short term. In this scenario, even if Agent's associated load size were to land, x86 would not be likely to lose its position quickly.
There is also an easily neglected supply chain conductor effect: CPU demand to lift or remodel the voice of the entire track of the server.
The availability of GPU quotas at the AI training cluster-led stage is often a central prerequisite for the purchase of orders by the entire manufacturer. As the priority levels of CPU, memory, power supply, and heat spread in the procurement list are synchronized, the role of the whole plant and the system integrator provides an opportunity to return from the assembly plants that are simply labelled by the outside world to programme designers.
Combining the customization requirements of the Agent load to the whole machine, the resource ratio and the power balance would magnify the design value at the system level.
There is, of course, a further risk to be noted: if capital markets pre-pricing the target in accordance with the CPU supercycle, but the commercialization of Agent is at a slower pace than expected, valuation retreats may be as rapid as before.
This branch of growth is still at the expected level at this stage and has not yet fully materialized into performance, and the associated pricing is more in need of anchoring verifiable exit and landing data than mere narratives.
X86 rent, Arm's increase
The pattern of the CPU market for more than a decade is largely based on an ecological rent.
A large number of databases, intermediates and industry business systems have been developed around this architecture for a long time, and the costs of moving the client-to-peer structure are high and constitute barriers to long-term stability. Of course, the rent is premised on the continued investment of the manufacturer in the chip architecture and platform capabilities, rather than simply building dividends from history.
Intel and AMD have been taking a market share of nearly 90 per cent over a long period of time, and are not behind the absolute crushing of chip performance. The real barrier is the complete software ecology that grows around x86: the vast majority of the world ' s business software, databases, intermediates and transport tool chains, with x86 as the default adaptation benchmark for the long term, and cross-structured migrations entail considerable adaptation and certification costs.
Once the customer has switched the chip structure, the compatible retrofit pressure of the entire software warehouse is presented directly. The high relocation costs constitute the moat of x86 and support this ecological rent.
Arm didn't launch a challenge from Calxeda in the early years to Ampere, Arm Server core. It has been trying for more than a decade, and it has been difficult to form a scaleed landing in the traditional business market, with the core impediment being the software eco-wall from base x86.
Of course, the cost of ecological transport remains one of the most difficult to overwhelm, as early attempts have multiple constraints such as hardware, finance, and the down-the-ground rhythm of clients.
But the path to this round of breakthroughs has changed: instead of hitting the traditional market of heavy-stocked enterprises, it has moved from a cloud manufacturer or an AI-based customer to a scenario where you can take control of the entire suite. Such subjects are in a position to do so at the start-up stage of their operations on the basis of Arm ' s suitability and do not have to pry the already solid stock application ecology.
For example, cloud manufacturers customize their operating systems at the ground floor, their movement control components and the operating environment in depth, so as to absorb the costs of fit and compatible validation within the system, thus circumventing the external ecological slab that Arm has long faced.
While NVIDIA has come out of another path: it does not completely provide an independent Arm CPU, in NVL72 whole-box programmes such as GB200, Vera Rubin, Grace, Vera and GPU are delivered in depth in the form of Superchip.
The client procured a complete AI rack for Agent ' s load optimization, which is a direct delivery scale for Arm CPU and a binding delivery of GPU within the cabinet. The process of GB200 to Vera Rubin, overlaying the order, also synchronized the increase in the delivery of the Arm CPU.
As can be seen, markets are becoming fragmented: x86 holds the fundamentals of stock ecology, and Arm opens up new growth space by taking advantage of the incremental market brought by Agent.
The 31 per cent and 40 per cent increases appear to be attractive, although it is not difficult to disassembly this projection to find that it is based on three sets of unverified assumptions.
The first assumption is that Agent can achieve the commercialization of scale. The current cloud company and the LLM enterprise are both intensely rolling out Agent-related products, but none of the user paid's will, the end-to-end completion rate of the mission, and the long-term size of the enterprise's clients have yet to undergo a full business cycle test.
If Agent stays in a state where the performance is bright and paid is weak, the incremental space of the CPU on the server will shrink accordingly, and the narratives of Arm relying on this expansion of the track will be put under pressure.
The second assumption is that the resource ratio of CPUs and GPUs will stabilize in high CPU consumption areas. At this stage, it is true that Agent loads lead to significant side costs for the mainframe, but it is possible that the iterative structure of the reasoning framework and the bottom structure may at any time reshape the load distribution.
The third assumption is that the fragmentation of the Arm ecology will not slow down the pace of expansion. Different Arm platforms, such as Graviton, Axion, Ampere, Grace/Vera, are not fully compatible with each other, and software often needs to be tailored to different platforms.
Within the closed cloud-planted home-grown ecology, such adaptation costs can be internalized; however, when the Arm programme moves out of the closed loop and reaches out to external ISV and traditional business clients, fragmentation translates into clear transport and maintenance costs.
For industrial investors, it would be more valuable to target three sets of forward HF indicators: Changes in the ratio of CPUs to GPUs in the cloud manufacturer hardware purchase list, the true call size of Agent products versus paid conversion efficiency, and changes in the share of the Arm architecture server CPUs in quarterly deliveries.
The data on first-line orders and operations are far closer to the real industry than any agency's long-term projections.
To put TrendForce back in business, it is essentially a simple change: The input logic of AI computing is moving from purchasing only accelerators to combining a balanced set of machine cabinets.
The GPU remains the largest expenditure item in power procurement, but CPU, memory, network and power supply were previously seen as supporting elements and began to obtain independent weights in budget allocations.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.