DATA CENTER · POWER · REFERENCE · Published May 30

AI data center power crisis, mid-2026: US construction fell for the first time since 2020. More than half of planned 2026 builds are expected to slip.

Bloomberg, February 25, 2026: US data center capacity under construction fell to 5.99 GW at end-2025, down from 6.35 GW at end-2024 — the first decline since 2020. The drivers: permit, zoning, and power-procurement delays, compounded by a domestic shortage of transformers, switchgear, and batteries that has forced reliance on imported equipment. More than half of US data centers planned for 2026 are expected to be delayed. Hyperscalers (Alphabet, Amazon, Meta, Microsoft) are still on track for over $650 billion in AI infrastructure capex this year — capital is not the constraint, the ability to physically deploy it is. By 2030, BCG projects AI data centers alone will consume electricity equivalent to two-thirds of all US residential demand. The shape of the AI build-out is now an energy story.

5.99 GW under construction · end-2025
$650B+ hyperscaler 2026 capex
2/3 of US homes by 2030
TL;DR 30-second version · free
  1. 01 The numbers (Bloomberg + BCG, Feb-May 2026): US data center capacity under construction was 5.99 GW at end-2025, down from 6.35 GW at end-2024 — the first annual decline since 2020. More than half of US data centers planned for 2026 are expected to be delayed, per Bloomberg. Hyperscaler aggregate capex is on track at $650B+ (Alphabet + Amazon + Meta + Microsoft). The gap between announced AI infrastructure plans and physically-deliverable capacity has widened sharply through 2025-2026 — and capital is not the constraint.
  2. 02 The bottleneck moved from silicon to power. Through 2023-2024 the constraint was H100s and the related supply chain. By mid-2026 the binding constraints are utility power availability, transformers (24-36 month lead times for large-scale), switchgear, batteries, and grid interconnection. Rack power density tells the same story — racks once at 30-40 kW are now hitting hundreds of kW with megawatt-range designs in late development. The grid was designed for none of this.
  3. 03 What changes operationally: long-lead capacity contracts get re-bid at premium pricing; behind-the-meter generation (gas, solar+storage, fuel cells, SMRs eventually) enters the conversation for serious AI tenants; colocation capacity gets locked up by hyperscalers; the energy-stack vendors — utility-scale batteries, transformer manufacturers, transmission build, power-purchase-agreement structuring — get a multi-year tailwind. The picks-and-shovels theme shifted from chip equipment to grid equipment.
DEEP ANALYSIS · free while in beta
READING AS
FOR YOU

Efficiency is now a capacity win, not just a cost win. Model routing (use smaller models when sufficient), prompt caching (30-60% reduction at scale), structured outputs (15-25%), and request batching all reduce watts per request. Add per-watt instrumentation to your AI feature dashboards. When evaluating architectural choices, ask the energy question alongside the cost question. The teams who run this discipline are positioned for capacity-constrained scaling.

FOR YOU

The picks-and-shovels theme migrated from chip equipment to grid equipment. Utility-scale batteries (Fluence, Tesla Energy, BYD), transformer manufacturers (Hitachi Energy, Siemens Energy, GE Vernova), switchgear (Hubbell, Eaton, ABB, Schneider Electric), advanced cooling (Vertiv, Schneider, Stulz), SMR developers (NuScale, Oklo, X-energy, BWXT), gas turbine (GE Vernova, Siemens Energy, Mitsubishi Heavy), uranium fuel cycle (Cameco, Centrus). Public markets are partially pricing this; private allocations have room.

FOR YOU

Power is now a strategic procurement question. If your business runs significant AI workloads, get the AI capacity conversation onto the same agenda as the energy procurement conversation. 2026-2027 capacity that you do not lock now may not be available later. Behind-the-meter generation (gas, solar+storage) enters the conversation for serious AI tenants. Site selection for new facilities should weight utility responsiveness as heavily as land cost.

FOR YOU

AI infrastructure capex has a different shape now. Long-lead-time equipment (transformers 24-36mo, switchgear similar) needs to be capital-committed earlier in the project lifecycle. PPAs are decade-scale commitments. The IRR profile of new AI data center projects depends on power-price assumptions over the lifecycle of the contract. Build energy-price scenarios into project models; the 5-year assumption is no longer a flat extrapolation.

FOR YOU

AI infrastructure is now an energy-policy story. State-level utility interconnection reform, federal SMR licensing pace, transmission build-out (FERC Order 2023, planned transmission corridors), domestic transformer + grid-equipment manufacturing all matter to the AI buildout. If your organization has policy reach, the highest-leverage interventions in 2026-2027 are on the grid + power side, not directly on AI.

Six anchored facts about the May 2026 capacity picture. Each is load-bearing for the planning conversations downstream.

Feb 2026

Bloomberg · 5.99 GW under construction · >50% planned delayed

Bloomberg, Feb 25, 2026: US data center capacity under construction fell to 5.99 GW at end-2025 from 6.35 GW at end-2024 — first decline since 2020. More than half of US data centers planned for 2026 are expected to be delayed. Binding constraint: physical infrastructure availability, not capital.

capacity
Mar 2026

Hyperscaler 2026 capex: $650B+

Alphabet, Amazon, Meta, Microsoft combined AI infrastructure capex on track for $650B+ in 2026, per earnings disclosures + analyst consensus. Capital is not the constraint; ability to physically deploy it is.

capex
2030 proj.

2/3 of US homes' electricity by 2030

BCG May 2026 analysis: AI data centers alone will consume electricity equivalent to two-thirds of total US residential demand by 2030. The grid was not designed for this load curve.

demand
Trend

Rack power density: 30-40 kW → MW range

Racks once at 30-40 kW are now hitting hundreds of kW; megawatt-range rack designs in late development. Cooling architectures shift from air to liquid to immersion to direct-to-chip.

engineering
Constraint

Transformer lead time: 24-36 months

Large-scale grid transformer lead times have stretched from ~12 months pre-pandemic to 24-36 months in 2026. The component is the long pole in many planned data center buildouts.

supply
Constraint

1+ GW data centers becoming the unit of measure

Multiple announced AI data center projects now exceed 1 GW capacity per campus. The unit of compute infrastructure is now the gigawatt — comparable to a small nuclear reactor — not the megawatt.

scale

The AI infrastructure constraint stack has reorganized through 2025-2026. The shape below contrasts where the bottlenecks were and where they are now, and why the energy-stack vendors became the new picks-and-shovels of the AI buildout.

BEFORE
How the constraint stack looked in 2023-2024
  • Compute silicon (H100s, B200s) was the binding constraint — supply-allocated, premium-priced
  • Networking + interconnect (NVLink, InfiniBand) was secondary but visible
  • Data center capacity was assumed available — colocation operators had spare power
  • Utility power was treated as commodity input, not strategic resource
  • Transformers, switchgear, batteries were assumed available on commercial lead times
  • Rack densities were 20-40 kW — well within existing cooling architectures
→
AFTER
How the constraint stack looks mid-2026
  • Compute silicon is plentiful (NVIDIA + AMD shipping; competitive supply)
  • Networking is solved-enough for current scales
  • Data center capacity is the binding constraint — power-allocated, not commodity
  • Utility power is the strategic resource — hyperscaler-PPA-locked, location-determining
  • Transformers + switchgear at 24-36 month lead times — long-pole for new buildouts
  • Rack densities 100s of kW to MW — driving liquid + immersion cooling, redesigning entire facilities

The AI infrastructure buildout in 2026-2030 is no longer a chip-supply story; it is a grid-and-power story. Capital is plentiful; the ability to physically deploy at announced timelines is constrained. The vendors that benefit are the ones that solve power supply (utilities, generation, batteries, transformers) and the ones that improve compute-per-watt (efficient chips, model routing, prompt caching). Compute-per-watt becomes the right metric, not compute-per-dollar.

DEEP READ 4 sections · cited primary sources · technical review pending

01 The construction-pace decline — what Bloomberg actually measured

Bloomberg, February 25, 2026: capacity under construction at US data centers fell to 5.99 gigawatts at the end of 2025, down from 6.35 GW at the end of 2024. It was the first annual decline in active construction since 2020 — and it landed at the moment AI compute demand was at its sharpest growth rate. The drivers Bloomberg cites: permit and zoning delays, power-procurement constraints, and a domestic shortage of transformers, switchgear, and batteries that has forced reliance on imported electrical equipment.

Bloomberg further reports that more than half of US data centers planned for 2026 are expected to be delayed. "At risk" means different things by project: some sites slip from 2026 to 2027-2028 because the utility interconnection date moved out; some reduce scope (a planned 1 GW campus becoming 400 MW initially with later phases); some cancel entirely because the right power purchase agreement could not be secured. Industry-analyst estimates of the specific aggregate at risk (e.g., Tech Insider has pointed at 7 GW) sit on top of the Bloomberg pace numbers as additional context.

For buyers (enterprise AI tenants, GPU-as-a-service customers, model training shops), the consequence is real: GPU availability through 2026-2027 will reflect deployed capacity, not planned capacity. The premium for guaranteed capacity contracts has widened. The negotiation leverage for reserved instances and long-term commitments has shifted toward the cloud providers that can actually deliver megawatts on schedule.

  • Under construction · end-2024 6.35 GW (Bloomberg)
  • Under construction · end-2025 5.99 GW (Bloomberg, first decline since 2020)
  • Planned 2026 capacity at risk "more than half" expected to be delayed (Bloomberg)
  • Hyperscaler 2026 capex $650B+ aggregate (Alphabet + Amazon + Meta + Microsoft per Bloomberg)
PRIMARY SOURCE Bloomberg · US Data Center Construction Drops as Permit, Power Delays Slow Projects (Feb 25, 2026)

02 Why transformers + switchgear became the long pole

Large-scale grid transformers (10-100 MVA, the size class that connects a 100+ MW data center to the utility grid) had pre-pandemic lead times of roughly 12 months. By 2026 those lead times are 24-36 months. The reasons stack: pandemic-era manufacturing disruption, demand from the broader US grid build-out (renewables interconnection, EV charging infrastructure, residential demand growth), labor and material costs, limited number of global manufacturers (Hitachi Energy, Siemens Energy, GE Vernova, Hyundai Electric, Hubbell — collectively the global supply), and the long capital cycle to add manufacturing capacity.

Switchgear (the electrical control + protection equipment that distributes utility power inside a facility) shares the constraint, with similar lead times and similar concentrated supply. Batteries (for backup + grid services) are less constrained but still subject to supply discipline. Together, these three components — transformers, switchgear, batteries — are the long pole in physically deploying a new AI data center campus. A site with utility interconnection approval, land, permits, and capital but no transformer is a site with no data center.

What this means for the energy-stack vendors: a multi-year demand tailwind. Hitachi Energy and Siemens Energy have both signaled record backlogs and capacity expansion through 2027-2028. Hubbell and Eaton (switchgear) report similar dynamics. Fluence (utility-scale batteries) is benefiting from data center backup + grid services demand. The vendors that can ship transformers in 18 months instead of 36 will price accordingly.

  • Transformers (10-100 MVA) Pre-pandemic 12mo lead → 2026 24-36mo lead. Hitachi, Siemens, GE Vernova, Hyundai, Hubbell.
  • Switchgear Similar lead-time stretch. Hubbell, Eaton, ABB, Schneider Electric.
  • Utility-scale batteries Less constrained but supply-disciplined. Fluence, Tesla Energy, Wartsila, BYD.
  • Multi-year tailwind Record backlogs through 2027-2028; price discipline favors vendors who can deliver.

03 Rack densities, cooling, and the redesign of the facility

Data center engineering through 2010-2020 optimized for racks at 5-20 kW with air cooling. NVIDIA H100 deployments pushed densities to 30-50 kW; B200 and Blackwell-generation systems are arriving at 80-130 kW per rack; the next generation pushes to 200-400 kW and beyond, with megawatt-range rack designs in late development. Air cooling becomes structurally impossible at these densities — the heat flux exceeds what forced air can remove.

The cooling architecture shift is multi-stage. First, liquid-cooled rear-door heat exchangers (a relatively small change to the rack). Then direct-to-chip liquid cooling (significant retrofit, but well-established). Then immersion cooling (full submersion of compute in dielectric fluid; cleanest thermal profile, biggest facility change). The hyperscalers are deploying all three; the open question is which mix wins for which workload. Single-phase immersion is the bet among multiple operators for next-generation builds.

The facility-level redesign is equally substantial. Megawatt-rack power distribution requires busbar designs more akin to industrial substations than traditional rack PDUs. Floor loading constraints get tested by liquid-filled immersion tanks. Heat reuse becomes economically viable (district heating, adjacent process heat) at gigawatt-scale facilities. The data center of 2027 will look meaningfully different from the data center of 2020 — and the engineering know-how to design at this scale is itself a scarce resource.

  • Rack density evolution 5-20 kW (2010-2020) → 30-50 kW (H100) → 80-130 kW (Blackwell) → 200-400 kW + MW (next gen)
  • Cooling shift Air → rear-door liquid → direct-to-chip liquid → immersion (single-phase, then potentially two-phase)
  • Power distribution Traditional rack PDUs replaced by busbars approaching industrial-substation design
  • Heat reuse Becomes economically viable at gigawatt scale: district heating, adjacent process heat, agriculture

04 Behind-the-meter generation and the SMR question

Faced with utility-power constraints, hyperscalers and large AI tenants are increasingly looking at behind-the-meter generation — power generated on-site, not delivered through the utility grid. Options: natural gas turbines (fast to deploy, carbon footprint), solar + storage (location-dependent, scaling well), fuel cells (Bloom Energy and competitors, niche but real), and small modular reactors (SMRs — the longer-horizon high-payoff bet).

Microsoft's September 2024 announcement of a 20-year PPA to restart Three Mile Island Unit 1 (Constellation Energy) marked a strategic shift: hyperscalers willing to underwrite the long-tailed capital structure of nuclear power to secure decade-scale capacity. Amazon, Google, Meta have followed with various nuclear-power deals (PPAs, equity stakes in SMR developers, site-adjacent capacity contracts). The bet: SMRs at 50-300 MW scale, deployed adjacent to data centers, deliver carbon-free baseload that the utility grid cannot.

The SMR timeline is the open question. NuScale Power, Oklo, X-energy, GE Hitachi BWRX-300, TerraPower, Westinghouse AP300 — multiple SMR designs are in NRC licensing or under construction, but first commercial operation dates range from 2027 (Oklo aggressive timeline) to 2030-2035 (more conservative estimates). Through 2027-2028 the realistic behind-the-meter mix is natural gas + solar/storage; SMRs ship in measurable quantity later in the decade.

  • Natural gas turbines Fastest to deploy, carbon-heavy. Gap-filler through 2027-2028.
  • Solar + storage Location-dependent. Sun Belt, Texas, Southwest plays scaling well.
  • Fuel cells (Bloom etc) Niche but real for specific deployment patterns. ~5-20 MW per site practical.
  • SMRs The 2028-2035 high-payoff bet. NuScale, Oklo, X-energy, BWRX-300, TerraPower, AP300.
  • Nuclear restart Three Mile Island Unit 1 (Microsoft/Constellation), others in negotiation. Decade-scale capacity bet.

Six places where the AI infrastructure narrative is misleading or where operational pain is being understated.

  1. announced-vs-deployed HIGH

    Treating announced data center capacity as deployable capacity

    Hyperscaler and AI-first company data center announcements through 2024-2026 routinely cite "X gigawatts by 2026/2027" headlines. The May 2026 reality: roughly one-third of those gigawatts is actively under construction. Treating announced capacity as deliverable capacity in your 2026-2027 product planning understates the actual GPU availability you can rely on. The 30-50% slip rate is a structural feature for the next 18-24 months, not a temporary disruption.

    DO In capacity planning, discount announced 2026-2027 buildout by 30-50%. Use deployed-capacity numbers (under construction + commissioned) as the realistic input, not press-release totals.
  2. compute-per-dollar HIGH

    Continuing to optimize compute-per-dollar instead of compute-per-watt

    Through 2023-2024, dollars were the binding constraint and compute-per-dollar was the right metric. As power becomes the binding constraint and dollars become plentiful, compute-per-watt (or compute-per-megawatt) is the right metric. The teams that retain a compute-per-dollar mindset will be outcompeted by teams that pick more efficient models, route to smaller models when sufficient, cache aggressively, and design feature P&Ls around energy consumption rather than just inference cost.

    DO Add per-watt metrics to your AI infrastructure dashboards. When evaluating model choices and architectural patterns, optimize for compute-per-watt at constant quality.
  3. utility-as-commodity HIGH

    Treating utility power as a commodity input

    Through the cloud era, utility power was a commodity — you bought compute from AWS/Azure/GCP and they handled energy procurement. That mental model breaks when you scale to megawatts. Power is now a strategic resource: PPA-locked years in advance, location-determining for new facilities, scarce in specific geographies. Enterprise AI tenants negotiating major capacity contracts in 2026 should treat power supply as a first-class procurement question, not a cloud-vendor implementation detail.

    DO In any 2026+ capacity contract over 10 MW equivalent, ask the provider for the underlying power supply structure: where is it, what is the PPA term, what happens to your capacity if their PPA renegotiates.
  4. smr-timeline-optimism MEDIUM

    Counting on SMRs to deliver behind-the-meter power before 2030

    SMR press coverage in 2025-2026 has been optimistic; the operational reality is that meaningful SMR generation will come online 2028-2032 in the best-case scenarios, later in the conservative ones. Hyperscalers signing SMR deals are securing 2030s capacity, not 2026-2027 capacity. The realistic behind-the-meter mix through 2028 is natural gas + solar/storage + fuel cells, with SMRs as a longer-horizon bet.

    DO In capacity planning through 2028, count SMRs as 0 GW available. Plan for natural gas + solar/storage as the realistic behind-the-meter sources for that window.
  5. cooling-design-lockin MEDIUM

    Locking in cooling architecture before next-gen rack densities are clear

    Rack densities are still evolving — 80-130 kW today, projected 200-400 kW and beyond in 2026-2027. A facility designed for 80-130 kW air or rear-door liquid cooling may be undersized for 2027-2028 generation hardware. The right architectural choice depends on the multi-year horizon: 2-3 year facility lifecycle (rear-door liquid is enough), 5+ year (direct-to-chip or immersion to be safe), 10+ year (immersion + heat reuse + serious facility-level redesign).

    DO In new facility design, plan for at least one rack-density generation beyond today's hardware. The cost of over-designing is small; the cost of stranded assets is large.
  6. grid-modernization-skepticism MEDIUM

    Underestimating utility-side grid modernization timelines

    Adding gigawatts of AI-data-center load to existing utility transmission and distribution infrastructure requires utility-side investment that takes years. Some utilities are moving faster (Texas ERCOT, certain Southeast utilities); others are slower (California, parts of the Northeast). Site selection that ignores utility responsiveness is selecting for delay. The geographic concentration of AI infrastructure in the Texas-Virginia-Oregon-Arizona belt is partly a function of which utilities can deliver megawatts on credible timelines.

    DO In site selection, evaluate the local utility's recent track record on large interconnection requests. Past performance is the best predictor.

Three actions worth taking this quarter regardless of where you sit on the AI infrastructure plan.

  1. 1

    Re-plan 2026-2027 AI compute capacity against deployed-capacity reality

    Take your 2026-2027 AI infrastructure plan and re-baseline against deployed (not announced) hyperscaler capacity. Apply a 30-50% discount to announced 2026-2027 buildouts. If your product depends on GPU availability ramps tied to those plans, you have a planning gap. Either lock long-term reserved capacity contracts now (premium pricing but guaranteed delivery), or reduce your compute-dependence assumptions.

  2. 2

    Add compute-per-watt to your AI infrastructure metrics dashboard

    Make energy efficiency a first-class metric alongside cost. Per-feature: watts consumed per request, per million tokens, per agent loop. Per-deployment: PUE, rack power utilization, model-routing efficiency (how often a smaller model is sufficient). The teams that build these dashboards in 2026 will outcompete teams optimizing only for dollar cost as the power constraint binds.

  3. 3

    Talk to your colocation provider about 2027-2028 capacity now

    If you operate any on-prem or colocated AI infrastructure, get your provider in a conversation about 2027-2028 capacity, pricing, and any contractual commitments needed to secure it. Hyperscalers are buying available capacity at premium. Your existing colocation contracts may not renew at the same scale or price. Lock contracts now where you have leverage; build a Plan B for sites that may not be renewable.

Six developments that will reshape the AI data center picture over the next 12-24 months.

Transformer manufacturer capacity expansion

Hitachi Energy, Siemens Energy, GE Vernova, Hyundai Electric capex announcements through 2026-2027 will determine whether the 24-36 month transformer lead time eases or worsens. Watch for new fab announcements + capacity adds.

SMR first commercial operations

NuScale (CFPP exit / pivot), Oklo (Aurora plant), X-energy (Dow Chemical partnership), GE Hitachi BWRX-300, TerraPower Natrium. First commercial operation dates ranging 2027-2032 will reshape the behind-the-meter conversation.

Hyperscaler PPA + nuclear-restart announcements

Microsoft/Three Mile Island, Amazon/Talen, Google/Kairos — the cadence and scale of these decade-horizon power deals tells you how seriously hyperscalers treat the power constraint.

NVIDIA / AMD rack density roadmaps

When next-gen NVIDIA and AMD rack designs publish power specs, facility planners get the data they need to size cooling and power. Watch GTC + AMD launch announcements through 2026-2027.

Utility interconnection process reform

FERC and state PUC actions to streamline large-load interconnection (the queue process that gates new data center grid connections) will measurably affect deployment pace. Multiple proceedings active in 2026.

Immersion + advanced cooling adoption curve

Single-phase immersion is becoming the default for next-generation hyperscale builds. Watch which hyperscalers commit publicly — sets the supply ecosystem for cooling equipment vendors.