AI inference drives a server CPU supercycle

Overview

This report presents J.P. Morgan's proprietary server CPU shipment model, splitting demand into three categories: AI server headnode CPUs, agentic AI server CPUs, and general-purpose server CPUs. The report argues that as AI applications shift from training to inference, rising needs around GPU and AI accelerator orchestration, storage, query scheduling, and security will push server CPUs into a new high-growth cycle.

Core views

The core views are: first, high growth in AI accelerator shipments combined with rising CPU attachment rates will drive exponential growth in demand for AI headnode CPUs; second, agentic AI servers will become a breakthrough incremental source of demand for general-purpose server CPUs; third, Arm-architecture CPUs will rapidly gain share in AI headnode and agentic AI servers, but x86 server CPUs are still expected to achieve a 25% CAGR in 2025-2028.

Analysis framework

The report uses a combination of top-down and bottom-up methods. The top-down section anchors on Gartner's 2024 server CPU data and combines J.P. Morgan's semiconductor team's forecasts for AI accelerators and CoWoS, accelerator-to-CPU ratios, and general-purpose server CPU growth assumptions to estimate total demand. The bottom-up section breaks demand down by Intel, AMD, Arm, and ASIC-related CPU suppliers, and cross-validates using the three demand buckets of AI headnode, general-purpose server, and agentic AI.

Methodology notes

• Demand forecastingThree-bucket server CPU demand model

  Split server CPU demand into AI headnode CPUs, agentic AI CPUs, and general-purpose server CPUs.

  This framework is used to distinguish traditional server upgrades, CPUs bundled with AI accelerators, and newly added orchestration server demand in the inference era, thereby explaining why total CPU demand growth is higher than in past cycles.

• Cross-validationTop-down and bottom-up triangulation

  Use total-demand forecasts and vendor/architecture breakdown forecasts to validate each other.

  The report first derives total CPU demand from total server volume, AI accelerators, and attachment rates, then validates it using shipment breakdowns across x86, Arm, and ASIC-related CPU suppliers.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• AMD

  Related beneficiary from rising x86 server CPU share

  Strengths

  The report notes that AMD continues to gain market share in x86 server CPUs and is poised to benefit from accelerating demand growth for x86 server CPUs.

  Weaknesses

  The rapid increase in Arm architecture share in AI headnode and agentic AI servers may limit x86's long-term share in certain AI workloads.

  Comparison

  Compared with Arm-related CPUs, AMD benefits from the improving x86 cycle; compared with the traditional general-purpose server cycle, this round of growth is driven more by AI inference and agentic AI.

  Risks

  Faster-than-expected Arm substitution, changes in AI server architecture, and easing CPU supply-chain tightness leading to downward revisions in ASP assumptions.

• TSMC

  Beneficiary of AMD and AI CPU supply chains

  Strengths

  AMD's rising server CPU share benefits its key foundry, while expanding demand for AI server CPUs supports advanced-node demand.

  Weaknesses

  Exposed to client product timing, capacity allocation, and fluctuations in end demand for AI servers.

  Comparison

  Compared with downstream ODMs, TSMC is closer to the high-value chip manufacturing link.

  Risks

  AI accelerator or CPU demand falls short of expectations, or pricing or capacity assumptions change.

• ASPEED, Lotes, Unimicron

  Highly correlated beneficiaries of the server CPU cycle

  Strengths

  The report believes server CPU, BMC, substrate, and socket demand are highly correlated, and all three benefit from unit demand, content upgrades, and potential price increases.

  Weaknesses

  Demand is highly dependent on server shipments and customer platform upgrade cycles.

  Comparison

  Compared with memory vendors, these companies are more directly exposed to upgrades in server motherboards, management chips, and interconnect structures.

  Risks

  Server CPU shipments miss expectations, product price increases fail to materialize, or platform specifications change.

• SK hynix, Samsung Electronics

  Beneficiaries of DRAM and high-capacity memory demand

  Strengths

  Rising demand for DIMM and SOCAMM may benefit memory vendors.

  Weaknesses

  The memory cycle remains affected by supply-demand conditions, pricing, and capital expenditure.

  Comparison

  Compared with CPU-related components, memory vendors' earnings are also affected by the DRAM industry cycle.

  Risks

  Memory price volatility, customer inventory adjustments, and compliance concerns related to investment-banking conflict disclosures.

Key data

• Total server CPU shipments26 million units in 2025, 68 million units in 2028Equivalent to a 38% CAGR in 2025-2028.
• Server CPU TAMAbout US$100bn in 2028The report assumes CPU prices rise by about 10% per year, driving TAM to a 53% CAGR in 2025-2028.
• AI accelerator shipments7.6 million units in 2024, 32.5 million units in 2028Based on J.P. Morgan semiconductor team's CoWoS-related forecasts.
• AI server headnode CPU demandAbout 74% CAGR in 2025-2028Driven jointly by growth in AI accelerator demand and rising CPU attachment rates.
• agentic AI CPU demand155% CAGR in 2025-2028The report believes this is a key growth driver of total server CPU demand.
• Arm server CPU shareAbout 22% in 2025, about 43% in 2028Arm share in AI headnode CPUs is expected to reach about 90% by 2028.
• x86 and Arm server CPU growthx86 at 25% CAGR, Arm at 72% CAGRForecast period is 2025-2028.

Impact & implications

The report believes the server CPU supercycle will benefit the semiconductor manufacturing, packaging and testing, BMC, substrate, connector, PCB, server ODM, and DRAM supply chains. Key beneficiaries include TSMC, ASE, ASPEED, Lotes, Unimicron, Tripod, Wiwynn, Inventec, as well as memory vendors such as SK hynix and Samsung Electronics; AMD's rising share in x86 server CPUs also benefits its related foundry and packaging/testing supply chain.

Risks

• AI accelerator shipments come in below forecasts, weakening demand for AI headnode CPUs and agentic AI CPUs.
• The accelerator-to-CPU ratio does not decline as assumed in the report, causing CPU attachment-rate gains to fall short of expectations.
• There may be forecasting errors in Arm architecture penetration, x86 supply shortages, or customers' in-house CPU roadmaps.
• If the assumption of about 10% annual server CPU price increases cannot be realized, the 2028 TAM forecast may be revised down.
• If agentic AI applications are adopted more slowly than expected, incremental demand for general-purpose servers may be overestimated.

What to watch

• Changes in AI accelerator shipments and CoWoS supply.
• The evolution of accelerator-to-CPU ratios on platforms such as Nvidia NVL72, Google TPU, and AWS Trainium.
• The actual pull from agentic AI inference workloads on external storage, query orchestration, and CPU servers.
• The pace of Arm server CPU share gains in AI headnode and agentic AI servers.
• Changes in AMD's market share in x86 server CPUs, as well as order performance at supply-chain names such as TSMC and ASE.