US semiconductors and AI data-center infrastructure: BofA raises CY30 AI data-center systems TAM to $2.2tn as agentic AI broadens compute demand

Overview

This US semiconductor industry outlook raises BofA's long-term AI infrastructure forecast. Its central conclusion is that agentic AI expands token use and shifts demand toward inference, supporting a larger and more durable buildout across data-center compute, memory, connectivity and supporting semiconductor equipment.

Core views

BofA raises its CY26-30E AI data-center systems addressable-market forecast to about $2.2tn, from its prior $1.8tn, and lifts the implied CAGR to 40% from 33%. The firm attributes the increase to a reinforcing cycle: consumer and enterprise adoption of agentic tools, intensified competition among public and private frontier labs, and constrained chip supply. It argues that safety guardrails or slower frontier-model pacing would not necessarily reduce compute demand; they could raise the need for inference and other faster-growing workloads relative to training. Against this backdrop, BofA considers leading semiconductor valuations compelling relative to projected earnings growth, noting the SOX at 21x forward P/E versus a 24x median since the onset of ChatGPT and projected EPS growth potential above 40%. The data-center buildout is the quantitative core of the report. BofA projects total data-center systems TAM to rise to roughly $2.6tn by CY30E from $839bn in CY26E, compared with 11% CAGR for overall IT spending. Within this, AI systems reach about $2.2tn by CY30E. AI accelerators account for approximately $1.5tn, AI networking about $400bn, and data-center memory sales about $1.3tn within a roughly $2.0tn total memory market. AI connectivity is expected to grow from $41bn in CY26E to $163bn in CY30E, a 41% CAGR, including optical connectivity rising to $111bn and copper connectivity to $28bn. The report expects networking and optics to grow at 30-40%+ CAGR as scale-up and scale-out architectures expand and CPO, NPO and AEC adoption increases; non-AI data-center systems grow only about 7% annually. BofA cross-checks demand with bottom-up chip-vendor and top-down operator deployment models. Its bottom-up analysis implies approximately 212GW of incremental capacity during 2026-30E, after about 41GW installed through 2025. The operator model projects global AI/cloud capacity from around 41GW in CY25 to about 219GW by CY30E, with annual additions increasing from roughly 22GW in CY26E to 45GW in CY30E. The 219GW operator estimate is about 15% below the 253GW implied by the semiconductor model; BofA attributes the difference mainly to broader deployments not allocated to named operators, including on-premise and sovereign projects. It translates this into annual deployments of roughly 35-50GW through decade-end at approximately $50-60bn per GW. The report argues that CPUs become more important as AI moves from training toward inference and agentic workflows. Total server CPU TAM is forecast to reach about $210bn in CY30E from about $61bn in CY26E, a 36% CAGR; AI CPUs rise from $42bn to around $180bn, a 44% CAGR. Of the CY30E CPU market, BofA allocates roughly $30bn to traditional/IaaS workloads, $90bn to AI compute or head nodes, and $90bn to standalone agentic AI CPU racks. These standalone racks handle sequential processing, orchestration, memory control and I/O alongside GPU/XPU parallel processing. CPUs therefore expand to about 7-8% of overall data-center systems value, versus below 7% in 2024-25, while AI CPUs represent more than 80% of server CPUs over time. In BofA's CPU share outlook, ARM is the fastest gainer as merchant and custom programs ramp. By CY30E, the firm forecasts value shares of about 22% for Intel, 31% for AMD, 38% for merchant ARM and 9% for custom ARM; Intel nevertheless remains the largest by unit share at about 36%. Intel's value share declines from roughly 34% in CY26E, but BofA still forecasts 20% revenue CAGR through CY30E, aided by supply shortages, ASP expansion and a potential Coral Rapids ramp by CY28E. AMD retains an estimated 27-31% value share through CY30E, supported by high-end performance, core density and x86 security and reliability features. ARM gains through NVIDIA, Qualcomm and custom hyperscaler programs, although new ARM processors temper AMD's share gains beginning in CY27E. BofA sees cloud capital spending as high but sustainable. Its tracker forecasts aggregate top-US-hyperscaler, neocloud and China-cloud capex of about $1.0tn in 2026E, nearly double year on year, and $1.4tn in 2027E, up 37%; estimates for global cloud capex have risen by roughly 65-120% over the preceding year. Cloud capex excluding neoclouds is expected to consume roughly 105-110% of operating cash flow in CY26-28E, driving aggregate free-cash-flow margins to about -2% in 2026E and the -3% range in 2027-28E. BofA nevertheless views the investment as fundable through operating cash flow and expects improvement after 2029 as infrastructure begins generating revenue and returns. The top five hyperscalers had raised about $370bn in 2026, largely through long-dated debt and permanent equity, which BofA views as balance-sheet flexibility rather than evidence of immediate liquidity stress. Demand visibility is supported by frontier-lab monetization and customer commitments, in BofA's view. It cites OpenAI revenue expectations of $13bn in 2025, $30bn in 2026, $62bn in 2027 and about $283bn by 2030, alongside approximately $665bn of cumulative compute spend through 2030. It also cites Anthropic's annualized revenue run rate above $19bn, up $17bn year on year, as a proof point linking frontier-lab demand to cloud capacity and accelerator, HBM, networking, custom-silicon, power and thermal demand. Agentic workflows may generate 10x to 1,000x more tokens than chat interactions, creating usage-based cloud monetization and sustaining compute consumption even as efficiency improves. BofA also points to large backlogs and RPOs at Alphabet, Microsoft, Amazon, Meta and Oracle as support for out-year capex. The report updates the broader semiconductor outlook as well. BofA models total semiconductor sales of $3.4tn by 2030, versus $3.2tn previously, with 19% CY26-30E CAGR. It forecasts CY26E total semiconductor and core-semiconductor growth of 113% and 30% year on year, respectively, led by memory, data centers and improving industrial and automotive demand. Memory sales are forecast up nearly 327% in CY26E, with DRAM up 328% and NAND up 341%; compute and storage rises 50%, industrial 32%, automotive 12% and wired communications 29%, while wireless communications, consumer and smartphone demand remain weaker. BofA's named near-term top picks are NVIDIA, Intel, Micron, Marvell and Lam Research, with AMD, Applied Materials, Analog Devices and onsemi also favored.

Analysis framework

BofA combines a top-down estimate of cloud-operator capex and gigawatt capacity with a bottom-up model of GPU/XPU vendor sales and silicon content across compute, memory, networking, storage and optics. It then tests capex sustainability using operating cash flow, free-cash-flow margins, financing, customer backlog and frontier-lab demand indicators, and translates the demand outlook into CPU share and broader semiconductor end-market forecasts.

Methodology notes

• Industry AnalysisSupply-demand framework

  AI data-center supply-demand and capacity analysis

  The report compares chip-vendor-implied installations with operator capacity plans, then links demand from agents, cloud commitments and frontier labs to constraints in power, construction and chip supply.

• Industry AnalysisUpstream-Midstream-Downstream Transmission

  AI infrastructure value-chain analysis

  BofA traces cloud and frontier-lab spending through accelerators, CPUs, memory, networking, optics, power and semiconductor equipment.

• Valuation methodsP/E and PEG Valuation

  Forward P/E and PEG-based price-objective frameworks for mentioned stocks

  The report's stock appendices use earnings multiples, historical valuation ranges and, for Broadcom, a PEG framework to explain individual price objectives.

• Valuation methodsSOTP (Sum-of-the-Parts) Valuation

  Micron sum-of-the-parts valuation

  BofA values Micron's cyclical memory business using P/B and its AI HBM business using P/E, then combines the components in its price objective.

Asset mapping & comparison

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

• NVIDIA (NVDA)

  Top pick and projected AI accelerator share leader

  Strengths

  BofA expects roughly 70%+ AI accelerator share over time and cites leadership in AI compute and networking.

  Weaknesses

  Exposure to cyclical gaming and lumpy AI-project demand.

  Comparison

  Projected to retain the largest accelerator share versus AMD, Broadcom, Marvell and others.

  Risks

  Power availability, China shipment restrictions, competition, capital-return deceleration and government scrutiny.

• Intel (INTC)

  Top pick with agentic CPU and foundry catalysts

  Strengths

  Expected relative strength in traditional enterprise workloads, potential ASP expansion, and a Coral Rapids ramp by CY28E.

  Weaknesses

  Server CPU value share is projected to decline to about 22% by CY30E.

  Comparison

  BofA expects Intel to remain largest by CPU unit share but trail ARM merchant and AMD in CY30E value share.

  Risks

  Foundry yield and ramp risk, limited external customers, mature-PC weakness and CPU share loss.

• Micron Technology (MU)

  Top pick benefiting from AI-memory demand

  Strengths

  The report forecasts strong memory demand, including HBM, DRAM and NAND growth.

  Weaknesses

  Memory earnings remain exposed to ASP cyclicality.

  Comparison

  BofA separates cyclical memory from AI HBM in its sum-of-the-parts valuation.

  Risks

  Memory ASP declines, Chinese competition, share loss and softer data-center, smartphone or PC demand.

• Marvell Technology (MRVL)

  Top pick linked to custom ASIC, connectivity and optical growth

  Strengths

  Potential visibility improvement in major customer ASIC programs and share gains in AEC, CPO and scale-up switching.

  Weaknesses

  Legacy storage, enterprise networking and carrier markets retain cyclical exposure.

  Comparison

  BofA sees networking strength and ASIC upside offsetting cyclical risks.

  Risks

  Reduced visibility on key ASIC projects and competition from merchant vendors and Broadcom.

• Lam Research (LRCX)

  Top pick positioned for memory and logic wafer-fabrication spending

  Strengths

  Etch and deposition leadership, memory-cycle exposure, share gains and stronger foundry/logic mix.

  Weaknesses

  Near-term cost inflation and tariff concerns.

  Comparison

  BofA expects potential share gains across memory and logic.

  Risks

  Slower capex, delayed memory capacity additions, share loss, customer consolidation and China exposure.

Key data

• AI data-center systems TAM~$2.2tn by CY30ERaised from $1.8tn previously; 40% CAGR from CY26E versus 33% prior.
• AI accelerator TAM~$1.5tn by CY30EThe largest component of projected AI data-center systems spending.
• Cloud and neocloud capex~$1.0tn in 2026E; ~$1.4tn in 2027EApproximately 2x year-on-year in 2026E and +37% in 2027E.
• Global AI/cloud data-center capacity~41GW in CY25 to ~219GW by CY30EOperator-based estimate; annual additions reach ~45GW in CY30E.
• Server CPU TAM~$210bn by CY30EUp from ~$61bn in CY26E, implying 36% CAGR; AI CPUs reach about $180bn.
• Total semiconductor sales$3.4tn by 2030BofA's updated forecast, versus $3.2tn previously; 19% CY26-30E CAGR.

Impact & implications

BofA believes agentic AI broadens the infrastructure cycle beyond training clusters by increasing inference, token processing and CPU demand. Its framework favors continued spending across accelerators, memory, networking, optics and semiconductor equipment, while treating cloud backlogs, frontier-lab monetization and capacity deployment as evidence that elevated capex can persist into the out-years.

Risks

• BofA identifies execution risks in power availability, data-center construction, financing and component supply.
• Elevated cloud capex is expected to push free-cash-flow margins negative through 2028E before anticipated out-year improvement.
• Frontier-lab growth supports demand but lower margins and higher cash burn remain profitability challenges.
• The report notes continued weakness in PCs, smartphones and consumer end markets, which offsets stronger AI, memory and data-center demand.

What to watch

• Cloud and neocloud capex trajectories, especially whether annual deployment reaches the report's 35-50GW range.
• Frontier-lab revenue growth, compute commitments and token-intensive agentic workload adoption.
• Hyperscaler RPO and backlog conversion, including Alphabet, Microsoft, Amazon and Oracle disclosures.
• AI accelerator share trends, CPU program ramps and the shift from head-node CPUs to standalone agentic CPU racks.
• Power, construction, financing and component-supply constraints that could affect data-center deployment timing.