美国半导体：CPU重生：智能驱动增长，但需警惕竞争

CPU Reborn: agentic drives growth, butwatch for competition Price Objective Change 17 April 2026 Raising server CPU forecasts, ests/POs for AMD, INTC, ARMWe raise server CY26-30E server CPU industry sales forecasts to +15% CAGR towards $71bn from +13%/$59bn driven by: 1) accelerating spend from hyperscalers to addressemerging agentic AI workloads, and 2) stronger CPU pricing on supply tightness andimproved mix. We further raise our estimates and POs on Buy-rated AMD (to $310 from$280), Neutral-rated ARM (to $180 from $155) and Underperform-rated INTC (to $48from $40). Near-term, we see AMD as best-positioned with its leading pipeline (current-gen Turin, upcoming Venice). INTC benefits also, but we see the recent ~$110bn marketcap addition over 3 months as excessive. Longer-term, we believe CPU market getsmore competitive with several ARM-based variants including NVDA Vera, ARM AGI, andmany internal cloud options (Amazon Graviton, Google Axion, Meta/ARM, MSFT Cobalt). EquityUnited StatesSemiconductors Vivek AryaResearch AnalystBofASvivek.arya@bofa.com Duksan JangResearch AnalystBofASduksan.jang@bofa.com Michael ManiResearch AnalystBofASmichael.mani@bofa.com AMD: reiterate Buy, raise estimates/PO to $310 Wecontinue to like AMD’s strong CPU product pipeline and incumbency in cloud serverCPUs, as well as the upcoming first 2.0-2.5 GW of compute capacity (GPUs, CPUs, NICs)buildout beginning 2H26 at Meta, OpenAI, and others (potentially Anthropic). Each GWinstallment represents $15-20bn of net revenue for AMD, resulting in >60% YoY DataCenter growth in both CY26/27E. Our new $310 PO is based on 29x CY27E PE (vs. 27xprior) on larger addressable market. Liam PharrResearch AnalystBofASliam.pharr@bofa.com INTC: raise estimates/PO to $48, remain Underperform See inside for Glossary We like INTC’s improving opportunity in server CPU, EPS accretion from recent fab buy-outand scarcity value of its manufacturing & packaging assets for external customers.However, the ~$110bn market cap increase over the last 3 months seems excessive to us.On a conceptual sum-of-parts basis: 1) Any external manufacturing business, wafers orpackaging should not be valued more than the ~10x EV/S of leader TSMC. We model$3bn/$6bn in CY27/28E (MediaTek TPU packaging, Apple M processors, other ASICs), andanything more will require incremental capex, shifting out foundry breakeven & wipingcash accretion, 2) Core Products can generate $2-$2.50 but should also be valued no morethan the ~16x CY28E PE of profitable fabless peers AMD/AVGO/NVDA. Last, we areskeptical of durable benefits from Tesla Terafab beyond some IP sharing & packaging,since Tesla is creating a rival fab, and since it already has foundry relationships withSamsung & TSMC. Our new $48 PO is now based on 3.7x CY28E EV/S (vs. 3.5x CY27E) aswe look out into further foundry opportunity, toward high-end of 1.7x-4x historical range. ARM:reiterate Neutral, CPU chip TAM limitedWe like the upside in ARM royalties from rising NVDA/cloud CPU IP adoption that could offset headwinds in mid/low-end Android sales. However, on the chip side, we are yet tosee any third-party benchmarking of ARM AGI CPU, which enters a very crowded field ofincumbents and larger rivals. ARM’s recent CPU event featured Softbank/Stargate andMeta, but did not feature other hyperscale customers. Successes of NVDA Vera and x86is also competition for any new merchant incumbent in our view, limiting ARM’s actualopportunity. Medium/longer-term we expect ARM to leverage Softbank Graphcoreassets to enter AI accelerators, another growing but crowded field. Our new $180 PO isbased on 69x CY27E EPS (vs. 60x prior) on larger addressable market, but still within 2x-3x PEG framework for IP peers. BofASecurities does and seeks to do business with issuers covered in its researchreports. As a result, investors should be aware that the firm may have a conflict ofinterest that could affect the objectivity of this report. Investors should consider thisreport as only a single factor in making their investment decision.Refer to important disclosures on page 17 to 21. Analyst Certification on page 16. PriceObjective Basis/Risk on page 15.12960058 Timestamp: 17 April 2026 12:53AM EDT Contents The Rise of CPUs in AI Inference3CPUs in Training4CPUs in Inference4CPUs in Prefill Inference5CPUs in Decode Inference6 Server CPU Update8 Server CPU growth towards $70bn+ TAM8Server Unit Growth vs. Server CPU Unit Growth9Server CPU Competition9Server CPU Roadmap10 End Market Poster11 The Rise of CPUs in AI Inference AI training and AI inference perform fundamentally different workloads, and thus requirefundamentally different hardware/semis with varying degrees of roles for CPUs, GPUs,memory, and other compute/networking components. Particularly, GPUs or accelerators typically excel in parallel compute-intensive workloads,such as in AI training. Meanwhile, CPUs excel in sequential and latency-sensitive workloads, as well as indi