Data center capex continues to be revised upward, with strength maintained across the AI semiconductor value chain

Overview

This report focuses on the AI data center investment cycle within semiconductors, semiconductor capital equipment, and IT hardware. JPMorgan notes that announcements from cloud vendors and hyperscalers over the past quarter indicate AI infrastructure buildout plans continue to exceed expectations, driving upward revisions to data center capex forecasts. The report argues that AI inference demand, rising compute intensity from reasoning/agentic workloads, and continued capacity constraints at hyperscalers and foundation model labs are the core fundamentals supporting multi-year AI infrastructure growth.

Core views

The core view is that AI-related data center capex remains in an upward revision cycle. CY2026 data center spending is expected to grow 63%, above the prior 52% forecast; CY2027 is initially expected to grow 40%. NVDA and AVGO have already disclosed order or demand visibility extending into CY2027, with customers locking in capacity in advance. The report also emphasizes that the CY2027 forecast assumes limited increases in inflation and commodity costs, and assumes a higher mix of custom ASICs relative to GPUs; these factors could lower the per-GW cost of incremental capacity and create room for further upward revisions.

Analysis framework

The report uses cloud data center capex forecasts, North American data center power capacity plans, AI accelerator order visibility, and semiconductor company customer activity as its main analytical threads, and maps capex growth through to segments of the AI semiconductor value chain including GPUs, custom ASIC/XPU, HBM/eSSD, semiconductor equipment, and EDA.

Methodology notes

• Industry cycle analysisData center capex forecasting

  Use hyperscaler and cloud vendor capex plans to assess the strength of AI infrastructure demand.

  Upward revisions in capex growth and absolute spending typically imply stronger support for demand in servers, accelerators, storage, networking, semiconductor equipment, and design software.

• Supply-demand visibility analysisOrders, backlog, and capacity reservation

  Use disclosed backlog, orders, and advance customer capacity reservations to judge future revenue visibility.

  The report cites long-term demand or agreements from companies such as NVDA, AVGO, and MU as evidence supporting continued demand into CY2027.

• Technology substitution and structural changeMix shift between GPUs and custom ASIC/XPU

  Compare share shifts between general-purpose GPUs and hyperscaler in-house or custom AI XPUs.

  The report believes custom AI ASIC/XPU will continue to gain share within AI accelerator TAM in CY2026/CY2027, while GPU demand remains supported by multi-year AI compute expansion.

Asset mapping & comparison

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

• NVDA

  Core beneficiary in AI data center GPUs

  Strengths

  Blackwell and Vera Rubin demand visibility extends into CY2027, benefiting from inference demand, rising compute intensity, and customers locking in capacity early.

  Weaknesses

  The report believes custom ASIC/XPU will gain share relative to GPUs, which may affect GPUs' relative share within AI accelerator TAM.

  Comparison

  Versus AMD, NVDA remains the dominant merchant GPU supplier; versus AVGO/MRVL, NVDA is more concentrated in GPU acceleration.

  Risks

  Capex revisions fall short of expectations, insufficient supply chain capacity, customers shift to in-house ASICs, slowing AI demand growth.

• AMD

  Second-largest merchant GPU supplier and AI acceleration beneficiary

  Strengths

  MI400/MI500 data center GPUs are positioned to capture AI demand over the coming years.

  Weaknesses

  Relative to NVDA, ecosystem and market share are still in catch-up mode.

  Comparison

  The report positions AMD as the #2 merchant GPU supplier and includes it among AI/cloud top picks.

  Risks

  Product timing, pace of customer adoption, competition with NVDA, changes in AI GPU demand mix.

• AVGO

  Key beneficiary in custom AI ASIC/XPU

  Strengths

  Benefits from customer projects such as Google/Anthropic TPU, Meta MTIA, OpenAI ASIC XPU, and Softbank/ARM ASIC XPU, and has disclosed FY2027 AI revenue above $100 billion.

  Weaknesses

  Sensitive to the timing of hyperscaler custom projects and customer concentration.

  Comparison

  Relative to GPU suppliers, AVGO benefits more from rising share of custom ASIC/XPU.

  Risks

  Delays in custom projects, changes in customers' internal solutions, capex or AI ASIC penetration below expectations.

• MRVL

  Beneficiary in cloud custom ASIC/XPU and connectivity-related areas

  Strengths

  Benefits from the ramp of ASIC XPUs such as Trainium and Maia, as well as hyperscaler custom chip expansion.

  Weaknesses

  Business realization depends on customer ramp timing and project scale.

  Comparison

  Similar to AVGO in benefiting from custom ASIC/XPU, but with a different project mix and scale.

  Risks

  Customer project delays, intensifying competition, supply chain execution risk.

• MU

  Beneficiary in AI storage and memory

  Strengths

  Benefits from HBM memory and eSSD demand, and has signed the industry's first five-year CSA.

  Weaknesses

  The memory cycle may still be affected by pricing, supply-demand dynamics, and capacity expansion.

  Comparison

  Relative to compute chip companies, MU is more exposed to AI server memory and storage attach demand.

  Risks

  Memory price volatility, HBM competition, changes in AI server configurations.

• AMAT, LRCX, KLAC

  Beneficiaries in the semiconductor capital equipment value chain

  Strengths

  Expansion in AI chips and advanced manufacturing lifts demand for wafer fabrication, process control, and equipment.

  Weaknesses

  Equipment orders may lag end-market capex and are affected by fab expansion timing.

  Comparison

  Relative to chip design companies, equipment companies are more leveraged to AI supply-side capacity expansion and process investment.

  Risks

  Fab capex volatility, export controls, delays in customer capacity expansion.

• CDNS, SNPS

  Beneficiaries in the EDA value chain

  Strengths

  An increase in custom ASIC/XPU raises chip design complexity and demand for EDA tools.

  Weaknesses

  Revenue growth depends on the number of design projects, licensing models, and customer budgets.

  Comparison

  Relative to hardware companies, EDA benefits more from increased custom chip design activity.

  Risks

  Changes in customer design budgets, license renewals, competition, and regulatory risk.

Key data

• CY2026 data center capex growth63%Revised upward from the previous 52% forecast.
• Initial CY2027 data center capex growth40%The report believes it could still be revised upward later as visibility and confidence improve.
• Incremental data center capex in 2026More than $200 billionThe largest annual increase as of the report date, exceeding the 2025 record.
• Planned North American data center power capacityMore than 140GW by the end of 2025Supports future installed capacity expansion.
• Installed North American data center capacity48GW in 2025; 60GW+ in 2026E; 75GW+ in 2027EThe chart shows total capacity increasing from about 10GW in 2019 to 78GW in 2027E.
• Top 4 U.S. CSP annual data center capex YoY76% in 2024; 79% in 2025; 63% in 2026E; 40% in 2027EBased on annual capex and YoY growth shown in the chart.
• Tier 2 CSP and Neoclouds capex YoY108% in 2025; 66% in 2026E; 39% in 2027EThe chart shows growth for second-tier cloud vendors and neoclouds moderating from high levels but remaining strong.
• NVDA demand visibilityBlackwell and Vera Rubin order/demand visibility of over $1 trillion through CY2027Used by the report to support the view of sustained AI compute demand growth.
• AVGO AI revenue outlookFY2027 AI revenue above $100 billionThe report mentions Broadcom's discussion of AI revenue.
• Price date used in the report2026-04-24 closing priceUnless otherwise noted, company prices discussed in the report are as of the April 24, 2026 close.

Impact & implications

If data center capex continues to be revised upward, the AI semiconductor demand chain could see sustained catalysts. Beneficiaries include NVDA's data center GPUs, AMD's MI400/MI500 data center GPUs, AVGO and MRVL's cloud/hyperscaler custom AI ASIC XPUs, MU's HBM and enterprise SSDs, as well as semiconductor equipment and EDA companies. The report also implies that bottlenecks in AI infrastructure buildout are expanding from standalone chip demand to power capacity, supply chain capacity, advance customer capacity reservation, and overall system cost.

Risks

• AI infrastructure capex growth may come in below the current upwardly revised expectations.
• Bottlenecks in power, data center construction, supply chain capacity, or deliveries may limit installed capacity expansion.
• If inflation, commodities, and construction costs rise above assumptions, the per-GW cost of incremental capacity may increase.
• Rising share of custom ASIC/XPU may change the relative growth elasticity of GPU suppliers.
• Changes in demand, budgets, or technology roadmaps at hyperscalers or foundation model labs may affect order visibility.
• The report contains forecasts and estimates, and future actual results may differ materially from current views.

What to watch

• Whether cloud vendors and hyperscalers continue to revise capex guidance upward.
• Order, backlog, customer activity, and capacity commentary from NVDA, AVGO, MRVL, AMD, and MU.
• Supply improvement for GPUs, XPUs, HBM, eSSD, and advanced packaging in the AI server supply chain.
• Changes in North American data center power capacity, construction progress, and conversion rates.
• Share shifts between GPUs and custom ASIC/XPU in CY2026/CY2027.
• Whether market expectations for 2026/2027 capex growth continue to move higher during 2026.