AI resets the data center capex baseline, with U.S. power demand forecasts for 2030 revised up significantly

Overview

This J.P. Morgan hardware and networking industry report compiles the key views from recent third-party analysts on data center demand, project pipelines, and power supply bottlenecks. The main thesis is that AI infrastructure demand is pushing the scale and number of individual data center campuses to a new level, rewriting the traditional ~1 GW campus size with demand for AI campuses exceeding 10 GW; however, whether this demand can be realized increasingly depends on non-IT constraints such as power access, heavy electrical equipment, specialized construction labor, and regulatory approvals.

Core views

The core view is constructive but constraint-focused. On the demand side, expansion in AI servers, advanced chips, and cloud infrastructure has driven U.S. 2030 data center power demand forecasts up from 77 GW to 118 GW, with the aggressive scenario reaching 160 GW. On the supply side, constraints have expanded from a single chip issue to multiple components and stages: advanced packaging, HBM, DRAM/NAND, heavy electrical equipment, skilled trades, and grid interconnection all face long lead times. Regionally, off-grid gas solutions in West Texas and low-cost power markets such as Northern Mexico are becoming more attractive. On the risk side, local and state government opposition, inexperienced developers, and changing assumptions about infrastructure useful life could slow actual construction.

Analysis framework

The report uses a third-party industry update aggregation approach, synthesizing webinar content from institutions such as BloombergNEF, Gartner, and Datacenter Hawk to distill industry conclusions around demand forecasts, capacity pipelines, power interconnection, component lead times, and project cancellations/delays. The analysis focuses not on single-company financial models, but on identifying upward demand revisions and bottlenecks across the data center construction chain.

Methodology notes

• Industry supply-demand analysisDemand Forecast and Bottleneck Mapping

  Maps AI-driven data center power demand forecasts against constraints in the grid, equipment, labor, and approvals.

  This framework uses 2030 demand scenarios, annual capacity additions, grid load share, critical component lead times, and interconnection project status to assess the feasibility of data center expansion.

• Scenario analysisBase Case and AI-chip Scenario

  Estimates U.S. data center power demand using different assumptions for AI server and chip demand.

  Bloomberg’s base case expects U.S. data center power demand to reach 118 GW by 2030; the more aggressive AI-chip scenario, based on AI server forecasts from semiconductor vendors and OEMs, points to 160 GW.

Asset mapping & comparison

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

• IT hardware and telecom/networking equipment companies

  Benefit from server, networking, and infrastructure investment driven by AI data center capacity expansion.

  Strengths

  Demand forecasts have been revised up, individual AI campus scale has expanded significantly, and the demand base for hardware and networking equipment is higher.

  Weaknesses

  Some IT components still face lead times ranging from 3 to 48 months, which may limit the pace of revenue realization.

  Comparison

  Compared with traditional data centers, AI data centers require more high-speed networking, AI accelerators, and high-performance storage.

  Risks

  Power access, construction labor, and permitting delays may postpone equipment procurement and deployment.

• Semiconductor and memory supply chain

  AI server forecasts underpin the aggressive demand scenario, placing the related supply chain on an upward demand revision path.

  Strengths

  AI Accelerators, Advanced Packaging, HBM, and DRAM/NAND are all identified as critical components.

  Weaknesses

  Long lead times for Advanced Packaging, HBM, and DRAM/NAND indicate limited supply elasticity.

  Comparison

  Compared with the 3–12 month lead time for AI Accelerators, memory and packaging have longer lead times and more prominent bottlenecks.

  Risks

  If AI server shipments fall short of expectations or data center construction is hindered, chip demand scenarios may be revised down.

• Power utilities and grid equipment

  Growth in data center demand directly boosts demand for grid connection, power interconnection, and heavy electrical equipment.

  Strengths

  Under the base case, the U.S. adds about 10 GW of data center capacity per year, with the grid carrying about 80% of it.

  Weaknesses

  Grid interconnection and heavy electrical equipment have long lead times, and actual project construction lags far behind proposed applications.

  Comparison

  Lead times for non-IT constraints are generally longer than for most IT components, potentially making them the more critical bottleneck.

  Risks

  Local opposition, permitting delays, and insufficient grid carrying capacity may limit the deployment of new load.

• Off-grid gas and low-cost power regions

  When the grid cannot support aggressive AI demand, demand for off-grid generation and alternative regions rises.

  Strengths

  Off-grid gas campuses in West Texas are expanding rapidly, while Northern Mexico is more attractive due to lower-cost energy.

  Weaknesses

  Off-grid solutions may face constraints related to fuel, environmental issues, permitting, and supporting infrastructure.

  Comparison

  Compared with traditional data center siting that depends on the grid, power availability is increasingly determining location preference.

  Risks

  Policy, environmental review, and community opposition may affect progress in off-grid or cross-regional projects.

Key data

• Base forecast for U.S. data center power demand in 2030118 GWAbove the previous 77 GW forecast from 2H25.
• AI-chip scenario for U.S. data center power demand in 2030160 GWBased on AI server forecasts from semiconductor vendors and OEMs.
• Annual capacity additions under the base caseabout 10 GW/yearAround 80% is expected to be carried by the grid, with the remainder off-grid.
• Annual grid absorption requirement under the AI-chip scenario19 GW/yearThe report views this target as difficult, increasing the role of on-site gas generation.
• Proposed power interconnection capacity in Texas434 GWAbout 5.1x the state’s peak demand, covering all interconnection applications rather than only data centers.
• Contracted power interconnection capacity in Texas134 GW，31%ERCOT-related pipeline data cited by Gartner.
• Approved-for-construction power interconnection capacity in Texas39 GW，9%Far below the proposed scale.
• Actual power interconnection capacity under construction in Texas11 GW，2%Shows a clear gap between proposed projects and actual construction.
• AI Accelerators lead time3–12 monthsRelatively shorter among IT-related components.
• Optics lead time6–12 monthsSupply cycle for networking and optics.
• Advanced Packaging lead time24–36 monthsAdvanced packaging is a major bottleneck for AI hardware expansion.
• HBM lead time30–36 monthsSupply constraints in high-bandwidth memory.
• DRAM/NAND lead time36–48 monthsLong lead times for memory components.
• Heavy electrical equipment lead time36–60 monthsIncludes transformers, turbines, and switchgear, among others.
• Skilled trades lead time48–60 monthsIncludes electricians, refrigeration engineers, plumbers, and pipefitters, among others.
• Grid interconnection lead time48–84 monthsThe longest lead-time segment cited in the report.
• Projects blocked or delayed in 1Q26about $130 bnReflects rising resistance at the local and state government levels.

Impact & implications

In terms of investment implications, the report reinforces the view that the AI data center capex cycle is still moving upward, benefiting industry chains tied to data center construction, networking equipment, AI accelerators, optical components, advanced packaging, HBM, memory, heavy electrical equipment, and power infrastructure. At the same time, the scarcest constraints may shift from compute hardware to power and project execution capability, meaning investors need to distinguish between orders or proposed pipelines and projects that are actually contracted, approved, under construction, and connected to the grid.

Risks

• Insufficient power availability could become the primary constraint on data center expansion.
• Long lead times for heavy electrical equipment, skilled trades, and grid interconnection could delay projects.
• In regions such as Texas, the large gap between proposed interconnection capacity and actual capacity under construction requires careful differentiation of pipeline quality.
• Opposition from local communities and state governments is increasing, and project cancellations or delays have already risen significantly in 2026.
• Some projects are being advanced by inexperienced developers, creating higher execution risk.
• If hyperscalers shorten their assumptions for infrastructure useful life, demand for new data center buildings could decline.

What to watch

• Whether U.S. data center power demand for 2030 continues to move from the 118 GW case toward the 160 GW scenario.
• Changes in the share of annual new data center capacity carried by the grid versus off-grid solutions.
• Conversion rates in Texas interconnection applications from contracted, approved, and under-construction capacity to actual grid-connected capacity.
• Whether lead times improve for AI Accelerators, HBM, DRAM/NAND, advanced packaging, and optical components.
• Whether supply of heavy electrical equipment such as transformers, turbines, and switchgear remains tight.
• The speed of project deployment in alternative markets such as off-grid gas campuses in West Texas and Northern Mexico.
• Approval, restriction, and cancellation trends from local and state governments toward new data centers.