US Semiconductors: BofA raises its CY30 AI data-center systems opportunity to $2.2tn as agentic AI broadens compute demand.

Overview

BofA Global Research updates its US semiconductor outlook around a larger, more durable AI infrastructure cycle. Its central argument is that agentic AI increases token use and shifts demand beyond model training toward inference, orchestration and CPU-intensive workloads, supporting a broad expansion in data-center silicon and systems through 2030.

Core views

BofA raises its CY26-30E AI data-center systems TAM to $2.2tn from $1.8tn previously, lifting its projected annual growth rate to 40% from 33%. The report attributes the revision to a reinforcing cycle of consumer and enterprise agentic-tool adoption, competition among public and private frontier labs, and constrained chip supply. It argues that slower frontier-model progress or additional safety guardrails would not necessarily reduce infrastructure spending: they could raise compute requirements and shift the mix further from training toward faster-growing inference. The model spans hyperscaler capex, gigawatt deployment and silicon allocation. By CY30E, BofA expects about $1.5tn of AI accelerator TAM, roughly $400bn of AI networking TAM and approximately $1.3tn of data-center memory sales. AI connectivity is projected to grow at a 41% CAGR from $41bn in CY26E to $163bn in CY30E, including optical connectivity at $111bn and copper connectivity at $28bn. Networking and optics are expected to grow at 30-40%+ CAGRs as scale-up and scale-out networks expand and technologies such as CPO, NPO and AEC are adopted. In contrast, non-AI data-center systems are forecast to grow at only a 7% CAGR. The capacity analysis provides two checks on the demand outlook. From the chip-vendor bottom up, BofA expects roughly 22GW of incremental global data-center capacity in 2026E, 35GW in 2027E and about 212GW cumulatively over 2026-30E. Its operator-by-operator view projects capacity increasing from about 41GW in CY25 to about 219GW in CY30E, with annual additions reaching roughly 45GW. The top-down estimate is about 15% below the 253GW implied by the semiconductor model; BofA attributes the difference mainly to on-premise and sovereign deployments not yet assigned to named operators. Server CPUs are a material extension of the AI thesis. BofA forecasts total server CPU TAM of about $210bn in CY30E, up from about $61bn in CY26E, or 36% CAGR; AI CPUs are expected to account for about $180bn, growing at 44% CAGR. The report divides CY30E CPU TAM into roughly $30bn of traditional/IaaS workloads, $90bn of AI compute or head nodes, and $90bn of agentic AI CPU racks. It expects the initial concentration in GPU-attached compute nodes to give way to a roughly even split between compute/head-node and standalone agentic-rack demand by 2030, because agentic systems require sequential processing, orchestration, memory control and I/O alongside parallel GPU/XPU compute. In its CPU-share outlook, BofA sees ARM as the fastest value-share gainer as merchant and custom programs ramp. It forecasts CY30E value shares of roughly 22% for Intel, 31% for AMD, 38% for ARM merchant and 9% for custom ARM. Intel is expected to retain about 36% unit share despite a lower value share, while AMD is positioned by its high-core-count portfolio and x86 security and reliability features for AI head-node and agentic workloads. CPUs are expected to account for 7-8% of data-center systems value by CY30E, above less than 7% in 2024-25, as their importance in inference and agentic workloads rises. The report views cloud spending as increasingly supported by visible demand. Aggregate US hyperscaler, neocloud and China-cloud capex is estimated at about $1.0tn in 2026E, roughly double year on year, and $1.4tn in 2027E, up 37%; BofA expects $2-3tn by the end of the decade. It cites approximately $2.3tn of backlog at the four largest hyperscalers and about $370bn of 2026 capital raised by the top five hyperscalers, largely through long-dated debt and permanent equity. Although cloud capex is projected at roughly 105-110% of operating cash flow in 2026-28E and aggregate free-cash-flow margin at about -2% in 2026E and -3% in 2027-28E, BofA considers the spending sustainable because it is largely funded from operating cash flow and expects cash generation to improve from 2029 onward as infrastructure produces revenue and returns. BofA identifies frontier-lab revenue and token intensity as further support. It notes reported OpenAI revenue projections of $13bn in 2025, $30bn in 2026, $62bn in 2027 and $283bn by 2030, alongside roughly $665bn of cumulative compute spending through 2030. Agentic workflows are said to use 10x to 1,000x more tokens than a chat interaction, moving cloud economics from seat-based toward usage-based monetization. The report also argues that lower cost per token can increase usage rather than suppress capex, extending demand into inference, post-training, synthetic data, orchestration and enterprise automation. The semiconductor forecast is correspondingly stronger. BofA models 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 conditions, while PCs, smartphones and consumer units remain headwinds. It raises CY30 total-semiconductor sales to $3.4tn from $3.2tn previously, implying 19% CAGR from CY26E; memory is projected to reach about $2.0tn. The report expects CY26E memory sales to rise nearly 327% year on year, with DRAM up 328% and NAND up 341%, while logic demand benefits from AI accelerators and industrial markets recover after inventory digestion.

Analysis framework

BofA combines a top-down view of cloud operator capacity and capex with a bottom-up build from chip-vendor sales. It then allocates the resulting data-center spending across accelerators, CPUs, memory, networking, storage and optics, and tests the outlook against cloud backlog, financing, frontier-lab revenue and end-market semiconductor forecasts.

Methodology notes

• Industry AnalysisSupply-demand framework

  Bottom-up and top-down data-center gigawatt deployment analysis

  The report estimates capacity additions from chip-vendor demand and separately from named data-center operators, using the comparison to assess whether projected infrastructure demand is plausible.

• Industry AnalysisUpstream-Midstream-Downstream Transmission

  AI workload demand translated into cloud capex and semiconductor content

  The analysis links agentic AI usage and cloud commitments to data-center capacity, then to demand for accelerators, CPUs, memory, networking, optics and related components.

• Valuation methodsP/E and PEG Valuation

  Forward P/E and PEG-based valuation references for semiconductor stocks

  The report compares sector and company valuation multiples with historical ranges and projected EPS growth when discussing the attractiveness of selected chip stocks.

Asset mapping & comparison

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

• NVIDIA (NVDA)

  Top pick and leading AI accelerator supplier

  Strengths

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

  Comparison

  Expected to remain the dominant accelerator vendor despite share gains by AMD and Broadcom.

  Risks

  Lumpy AI projects, power constraints, competition, China shipment restrictions and government scrutiny are cited risks.

• Intel (INTC)

  Top pick with agentic CPU and foundry catalysts

  Strengths

  Expected to retain enterprise CPU strength; Coral Rapids could narrow the performance gap from CY28E.

  Weaknesses

  Value share is projected to decline to about 22% by CY30E.

  Comparison

  BofA expects Intel to retain the largest CPU unit share at roughly 36% by CY30E.

  Risks

  Foundry yield and ramp risk, lack of external foundry customers, PC weakness and CPU share loss.

• AMD (AMD)

  Favored AI CPU/GPU and server-processor beneficiary

  Strengths

  High-end performance, leading core counts and x86 security and reliability features support AI-node positioning.

  Weaknesses

  PC, embedded and console markets are described as slower-growth cyclical exposures.

  Comparison

  BofA sees AMD holding roughly 27-31% CPU value share through CY30E.

  Risks

  Rack-scale product execution, project timing, uneven customer spending and manufacturing concentration.

• Micron Technology (MU)

  Top pick and memory-cycle beneficiary

  Strengths

  The report expects exceptionally strong memory growth driven by AI and HBM demand.

  Risks

  Memory ASP declines, China competition, share loss and softening demand across data center, smartphone or PC markets.

• Marvell Technology (MRVL)

  Top pick tied to XPU, networking and custom ASIC growth

  Strengths

  BofA highlights custom ASIC visibility and AEC/CPO/scale-up share-gain opportunities.

  Weaknesses

  Exposure to cyclical legacy storage, enterprise networking and carrier markets.

  Comparison

  Faces competition from merchant AI vendors and incumbent ASIC supplier Broadcom.

  Risks

  Loss of custom-ASIC visibility and intensified competition in AI compute.

• Lam Research (LRCX)

  Top pick with memory and logic wafer-fabrication-equipment exposure

  Strengths

  Etch/deposition leadership, share-gain potential and memory/logic WFE growth support the case.

  Weaknesses

  Near-term cost inflation and tariff concerns are noted.

  Comparison

  BofA expects potential share gains across memory and logic.

  Risks

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

Key data

• AI data-center systems TAM, CY30E~$2.2tnRaised from $1.8tn previously; +40% CAGR from CY26E versus +33% previously.
• Cloud and neocloud capex~$1.0tn in 2026E; ~$1.4tn in 2027EApproximately 2x year-on-year growth in 2026E and +37% in 2027E.
• Global AI/cloud data-center capacity~41GW in CY25 to ~219GW in CY30EOperator-based estimate; annual additions rise to about 45GW by CY30E.
• Server CPU TAM, CY30E~$210bnUp from ~$61bn in CY26E, a 36% CAGR; AI CPUs contribute about $180bn.
• AI accelerator TAM, CY30E~$1.5tnBofA expects NVIDIA to maintain roughly 70%+ AI accelerator share over time.
• Total semiconductor sales, CY30E$3.4tnRaised from $3.2tn previously; implies 19% CAGR from CY26E.

Impact & implications

BofA’s thesis favors a broad AI infrastructure cycle rather than a narrow accelerator-only buildout. It sees rising agentic inference and orchestration needs expanding demand for CPUs, memory, networking, optics and power-related content alongside AI accelerators, while backlog and contracted commitments improve visibility for cloud spending.

Risks

• Execution constraints in power availability, data-center construction, financing and component supply could limit deployment.
• Frontier-lab monetization faces lower-margin and high-cash-burn challenges.
• Cloud capex is projected to materially depress near-term free-cash-flow margins.
• PC, smartphone and consumer semiconductor units remain end-market headwinds.

What to watch

• Cloud and neocloud capex plans, customer commitments and backlog conversion.
• The pace of data-center gigawatt additions across hyperscalers, neoclouds and sovereign projects.
• Frontier-lab revenue growth, compute spending and agentic AI token consumption.
• The shift from AI head-node CPUs toward standalone agentic CPU racks.
• Memory pricing, supply conditions and the pace of HBM, networking and optical adoption.