AI inference drives CPU rerating, with AMD the biggest beneficiary but competition heating up
AI summary card
AI inference drives CPU rerating, with AMD the biggest beneficiary but competition heating up
Bank of America lifts its CY26-30E server CPU sales forecast to a 15% CAGR and points to a $71bn market size, keeps AMD at Buy and raises the target price to $310, while remaining cautious on INTC's valuation and ARM's commercial CPU opportunity.
- The server CPU industry sales forecast was raised from +13% CAGR and $59bn to +15% CAGR and $71bn, driven mainly by agentic AI workloads and stronger CPU pricing.
- AMD is seen as best positioned in the near term, benefiting from the Turin and Venice product lines, its existing position in cloud server CPUs, and 2.0-2.5GW compute buildouts starting in 2H26 from Meta, OpenAI, and others.
- INTC benefits from enterprise customer stickiness, improved server opportunities, and the scarcity value of manufacturing and packaging assets, but the report argues that the roughly $110bn increase in market cap over the past three months was excessive.
- ARM benefits from royalty upside driven by higher NVDA and cloud CPU IP adoption, but ARM AGI CPU lacks third-party benchmarks, and the commercial CPU market is crowded.
- The report argues that CPUs are becoming more important in AI inference, especially in the decode stage, for scheduling, routing, KV cache management, and low-latency output.
Report interpretation
Overview
This report centers on the 'rebirth of CPUs in the AI inference era' and argues that as AI computing gradually shifts from training to inference, the value of CPUs rises again in sequencing, low latency, scheduling, control, batching, routing, and memory management. Based on this, Bank of America raises its server CPU industry forecast and adjusts the valuation and target prices for AMD, INTC, and ARM respectively.
Core views
The core view is that the server CPU market will expand with agentic AI and cloud capex, but winners will be clearly differentiated. AMD is the most attractive thanks to its cloud CPU product pipeline and existing customer position; INTC has upside from enterprise customer stickiness and manufacturing/packaging assets, but the stock already reflects too much optimism; ARM benefits from ecosystem adoption in royalties, but its own/commercial CPU chip TAM may be squeezed by x86, NVDA Vera, QCOM, and cloud vendors' in-house chips.
Analysis framework
The report starts with the hardware workload differences between AI training and AI inference, breaks down the roles of CPUs, GPUs, memory, and networking across the load, prefill, and decode stages, and then combines server shipments, CPU configuration per server, ASPs, supply tightness, product roadmaps, and competition to derive the CY25-30E server CPU TAM and company revenue opportunities.
Methodology notes
CPUs account for about 5% of the overall AI data center TAM and rise with agentic AI inference demand.
The report expects CPUs to maintain roughly a 5% share of the about $1.4tn AI data center TAM, with the server CPU market growing from about $28bn in CY25 to more than $70bn in CY30E.
The load, prefill, and decode stages determine the division of labor between CPUs and accelerators.
The load stage is affected by disk I/O and CPU speed; prefill is matrix-intensive, with CPUs mainly handling tokenization, prompt formatting, scheduling, and memory structure setup; decode is highly sequential and latency-sensitive, with CPUs handling GPU kernel scheduling, sampling, safety checks, logits processing, KV cache routing, and token output.
The company's target prices are adjusted based on different earnings and sales bases and peer valuation ranges.
AMD's target price is based on 29x CY27E PE; INTC's target price is based on 3.7x CY28E EV/S; ARM's target price is based on 69x CY27E EPS and still sits within the 2x-3x PEG framework for IP peers.
Asset mapping & comparison
Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).
- AMDPrimary beneficiary
- Strengths
- Leading cloud server CPU position, strong Turin and Venice product lines, and benefit from compute buildouts by Meta and OpenAI; each GW of compute buildout could correspond to $15-20bn in net revenue opportunity.
- Weaknesses
- Still operating in a highly competitive server CPU and AI infrastructure market and must continue to execute on its product roadmap.
- Comparison
- Compared with INTC and ARM, the report sees AMD as best positioned near term, especially in cloud and public cloud instances.
- Risks
- ARM-based variants, cloud vendors' in-house CPUs, x86 competition, and supply chain or product execution risks.
- INTCBeneficiary, but rating remains cautious
- Strengths
- Enterprise and telecom customer stickiness is relatively strong, manufacturing and packaging assets are scarce and valuable, server CPU opportunities are improving, and ASPs are supported by tight supply and better product mix.
- Weaknesses
- The recent roughly $110bn increase in market cap is viewed as excessive; foundry opportunities may require incremental capex and delay breakeven; the product roadmap still needs improvement.
- Comparison
- Compared with AMD, INTC's growth is expected to lag the overall market; compared with TSMC, its external manufacturing and packaging business should not deserve a higher valuation.
- Risks
- Concerns around Diamond Rapids, foundry cash flow drag, limited Tesla Terafab synergies, and competition from Samsung and TSMC.
- ARMRoyalty beneficiary, but chip TAM is limited
- Strengths
- Higher NVDA and cloud CPU IP adoption can lift royalties, and in the long term it may leverage SoftBank's Graphcore assets to enter AI accelerators.
- Weaknesses
- ARM AGI CPU still lacks third-party benchmarks, exposure at hyperscale customers is limited, and the commercial CPU market is crowded.
- Comparison
- Compared with x86, NVDA Vera, and cloud vendors' in-house Arm chips, ARM faces a smaller serviceable market as a new commercial CPU supplier.
- Risks
- Weakness in mid- and low-end Android, slower-than-expected adoption of commercial CPU customers, and intense competition in AI accelerators.
- METADownstream demand and potential customer
- Strengths
- Meta is mentioned as a participant in compute buildouts starting in 2H26 and is also a potential participant in the cloud/in-house CPU ecosystem.
- Weaknesses
- The report does not assign META a direct rating or target price.
- Comparison
- As a hyperscaler, META may both buy external CPUs/GPUs and strengthen its bargaining power through in-house or partner chips.
- Risks
- Changes in in-house chip roadmaps and external procurement cadence would affect supplier revenue realization.
Key data
- Server CPU industry sales forecast+15% CAGR to $71bnRaised from the prior +13% CAGR and $59bn estimate; covers CY26-30E.
- AI accelerator market growth+137% CAGRFrom 2022 to 2025, driven by AI training demand; server CPUs grew only +5% CAGR over the same period.
- 2025 compute spending mixAI accelerators 87%, CPUs 13%In 2021, AI accelerators were 26% and CPUs were 74%, showing a shift in value toward accelerators during the training era.
- Server CPU TAM outlookCY25 about $28bn to CY30E $70bn+CPUs are expected to account for about 5% of the overall AI data center TAM.
- Server shipment outlookCY25 about 17mn units to CY30E about 25mn units, +8% CAGRMainly driven by agentic AI demand and capacity shifting toward server chips.
- AMD server CPU sales outlookCY30E about $24bn, about 34% shareCY25-30E CAGR is about +20%; the report believes AMD will continue to lead in cloud, especially public cloud instances.
- ARM server CPU sales outlookCY30E about $18bnShare is expected to rise from about 13% in CY25 to about 25% in CY30E, with CY25-30E CAGR of about +37%.
- INTC server CPU sales outlookCY28E about $22bnRaised from about $13bn in CY25, supported by enterprise customer stickiness, ASPs, and tight supply.
- AMD target price$310Buy maintained, previous target price was $280.
- INTC target price$48Underperform maintained, previous target price was $40.
- ARM target price$180Neutral maintained, previous target price was $155.
Impact & implications
In investment terms, the AI infrastructure chain is no longer just a single-line beneficiary of GPUs and accelerators; CPUs, memory pooling, networking, and scheduling capability are gaining importance in inference. For equities, AMD's cloud CPU and data center growth leverage are clearer; INTC has operational and asset improvement, but valuation already reflects much of it; ARM's IP royalty leverage is better than its commercial CPU chip opportunity.
Risks
- The server CPU market is becoming crowded, with competition from x86, ARM commercial vendors, NVDA Vera, QCOM, and cloud vendors' in-house solutions such as Amazon Graviton, Google Axion, Meta/ARM, and MSFT Cobalt.
- If agentic AI inference demand or cloud capex falls short of expectations, the server CPU TAM expansion thesis could weaken.
- Tight CPU supply and ASP improvement may come at the expense of PC capacity, and supply-chain adjustments carry execution risk.
- If INTC's foundry and packaging businesses require additional capex, they could delay foundry breakeven and offset cash generation.
- ARM AGI CPU lacks third-party benchmarks, and actual customer adoption and merchant TAM remain uncertain.
What to watch
- AMD Venice launch in 2H26 and the pace of cloud customer adoption.
- INTC 18A/14A nodes, Diamond Rapids production ramp, and monetization of cooperation with the NVDA Rubin platform.
- ARM AGI CPU third-party benchmarks and hyperscaler customer disclosures around 2027.
- Whether agentic AI workloads continue to lift CPU scheduling, memory management, and low-latency decode demand.
- Progress in the real deployment of CXL memory pooling, KV cache management, and CPU-attached memory in inference systems.
- Whether server CPU ASPs, unit shipments, and CPUs per server meet the CY25-30E forecasts.