Datacenter Power Market Analysis — May 2026

Market Size & Growth

There is no single clean TAM figure in the signal set — and that absence is itself a finding (see Reverse-Hype). A transparent bottom-up proxy for the *AI/datacenter-attributable* revenue pool across the six tracked L1 vendors in 2025: GE Vernova FY2025 revenue ≈ $37B with multi-GW data-center turbine reservations [7eb603b0-de70-429a-b62c-1839c244d6c6]; Eaton Q3 2025 ≈ $7.27B/quarter with datacenter its fastest-growing end market [151e8e03-1d01-4bf4-9e9e-76ac82009a90]; Schneider 9M 2025 ≈ €30B with double-digit datacenter growth [62e94c87-61f2-4972-aac7-7e3328bd305c]; Vertiv FY2025 guided ≈ $10B [c763ccd4-bbaa-40a2-b372-67dabbb8b7bb]; Bloom FY2025 ≈ $1.9B [41044b02-1474-49ff-b957-8f1673479c8a]; Generac FY2025 ≈ $4.5B [f11ad94c-4afe-46c8-aade-76861e6d2413]. Datacenter-attributable slices of these put a defensible **lower bound near $20–25B/yr in directly disclosed AI-power equipment + on-site generation revenue**, growing fast.

• **Current market size (USD):** ~$20–25B/yr disclosed AI/datacenter-attributable equipment & power revenue across tracked L1 vendors (synthesized lower bound; constituent sources cited above).
• **3-year CAGR:** ~28% blended, anchored by Vertiv +29–34% YoY [300316a7-61c1-4e59-a464-708bed15cf8c; c763ccd4-bbaa-40a2-b372-67dabbb8b7bb], Eaton datacenter orders growth well above 30% rolling-12-month [8c2e3280-1cf9-4b89-b8d5-90ce4096898c], Bloom +38.6% Q1 2025 [a571fda2-9491-4116-9b99-d0cc59e292ee].
• **Key drivers:** GPU rack power density (142 kW/rack on GB300 NVL72 [0a8f1612-9dab-4213-87d2-324b0ff197e2], roadmap to 1 MW+ racks); grid interconnect scarcity pushing on-site generation; multi-year backlog conversion (Vertiv backlog ≈ $8.5B [c763ccd4-bbaa-40a2-b372-67dabbb8b7bb]).

Supply Chain

• **Upstream:** HA-class gas turbines (GE Vernova 7HA), solid-oxide fuel-cell stacks (Bloom), MV transformers & switchgear, copper/grain-oriented electrical steel, 800 VDC power semiconductors, liquid-cooling CDUs/manifolds (Motivair, Vertiv 360AI).
• **Downstream:** hyperscalers (Microsoft, Meta, AWS, Oracle), neoclouds, colo operators (Equinix, Digital Realty), greenfield off-grid AI campuses (Fermi HyperGrid, Amarillo [4d3aa15e-4b45-4bc0-903a-4926c0906900]).

`[turbine / SOFC / grid + transformer + switchgear OEM] → [MW-scale power + thermal integration & prefab skid] → [hyperscaler / colo / AI campus]`

Trend Lines (3 trends, ranked by importance)

**1. The 800 VDC rack-power re-architecture is a forced, ecosystem-gated redesign — not an incremental upgrade.**

• Evidence: "Eaton was named among NVIDIA's power ecosystem partners developing 800 VDC data center power architecture to support 1 MW+ AI racks" [4fee1ff0-170e-49fb-85b0-b1926ebb4491].
• Why now: Between Computex May 2025 [7c05d0c4-7b56-4399-94ad-91014439f151] and OCP October 2025, NVIDIA published a fixed 800 VDC spec tied to the 2027 Rubin/Kyber roadmap; at GB300 NVL72's 142 kW/rack [0a8f1612-9dab-4213-87d2-324b0ff197e2] legacy 415V AC distribution physically cannot carry the conductor current. This is a 2025 spec-freeze event, not generic AI progress.
• Implication: Vendors with NVIDIA-validated 800 VDC reference designs (Vertiv 360AI, Eaton, Schneider) get socketed into 2027 buildouts now; non-aligned power vendors are designed out before the RFP in the next 6–12 months.

**2. Time-to-power overtakes $/kW as the deciding procurement axis, and behind-the-meter generation becomes a financed asset class.**

• Evidence: "Fuel-cell power capable of being deployed within ~90 days of a site decision" [17af21a6-27e4-4a9c-937c-52f5eea6d0ec].
• Why now: The capability shift is financial, not technical — Brookfield's October 2025 commitment of up to $5B to fund and deploy Bloom platforms [2b1a430b-a3a8-441b-ae07-f02168962a54] turned behind-the-meter power into a project-financed vehicle, alongside GE Vernova turbine reservations for Fermi's ~11 GW Texas campus [4d3aa15e-4b45-4bc0-903a-4926c0906900]. Off-grid AI siting is now underwritable.
• Implication: Power moves from a hyperscaler capex line to an off-balance-sheet financed input; expect more 90-day SOFC and reserved-turbine campuses announced through 2026 while PJM-class queues stay >5 years.

**3. The bottleneck migrated from silicon to the electrical long-lead chain — driving vertical integration of rack/power/skid manufacturing.**

• Evidence: "expanding North American electrical manufacturing capacity for transformers, switchgear and modular power to meet data center and grid demand" [8c2e3280-1cf9-4b89-b8d5-90ce4096898c].
• Why now: 2024–2026 MV transformer lead times blew out to 130+ weeks [5b0a3681-e14c-4c8f-98b8-b3ad8d9da306]; the response was M&A — Eaton's $1.4B Fibrebond prefab acquisition [b9e9b09c-2d23-439c-947a-bc97cb72ad36] and Vertiv's $200M Great Lakes rack/cabinet buy [1d8de43f-43ee-40b2-8b48-e7a62390db76], both in 2025, to compress on-site build time.
• Implication: Awards through 2026 are decided by lead-time and integrated-skid throughput, not component price; further consolidation of rack/skid manufacturers is likely.

Key Inflection Points (Watch List)

• First production 800 VDC racks shipping ahead of the 2027 Rubin/Kyber roadmap [7c05d0c4-7b56-4399-94ad-91014439f151] — slippage resets the whole power-vendor selection.
• Vertiv book-to-bill staying >1 against ≈ $8.5B backlog [c763ccd4-bbaa-40a2-b372-67dabbb8b7bb]; a drop below 1 is the first crack in the supercycle.
• Brookfield's $5B Bloom facility drawdown pace [2b1a430b-a3a8-441b-ae07-f02168962a54] vs. Bloom realized FY2025 ≈ $1.9B [41044b02-1474-49ff-b957-8f1673479c8a].
• First hyperscaler campus actually *energized* off-grid (Fermi HyperGrid, ~11 GW target [4d3aa15e-4b45-4bc0-903a-4926c0906900]) — proof, not reservation.
• FERC Order 2023 interconnection-reform implementation: if queues clear, the behind-the-meter premium compresses [51251fff-baae-4fde-a526-c130d39b028b].
• MV transformer lead times breaking back below ~50 weeks [1a081a8c-b29c-42c8-aa8c-7e0c6c387a3b] — the single best signal that the constraint is relaxing.

Reverse-Hype Warnings

**Overhyped:** The pure-play "AI power supercycle" equity narrative. GE Vernova has ~4x'd since its April 2024 spin (~$140 → ~$520) [09fdfc29-2b9b-43da-bb10-fedd901222dd], but gas-turbine slots "sold out through 2028" are *reservations*, not non-cancelable revenue [bb99bbf8-9b45-4312-92be-59a28f33f9b9], and the same equity story ignores that Generac's consolidated growth is diluted to a ~8–12% ceiling by its residential half regardless of datacenter demand [2f51ddb4-e55e-4147-a4bd-d4a5959bf62a]. The fuel-cell thesis is doubly fragile: SOFC capex remains $7–9K/kW installed vs. $1–1.5K/kW for CCGT [51251fff-baae-4fde-a526-c130d39b028b] — Bloom wins on time-to-power, not economics, so any acceleration of interconnect-queue clearing (FERC Order 2023) erodes the entire pitch. Backlog-as-destiny is the consensus error: every L1 name is capacity-rationing today, which flatters order books that have not yet survived a single demand-air-pocket.

**Underrated:** The orchestration/software layer is the weak leg across *every* vendor — Vertiv mid-pack on DCIM [3fdf14ac-3a20-4b5f-bb20-a231a8578de9], Schneider's EcoStruxure IT only ~45-tier [97027062-3743-4500-950a-81000a791d64] — and Vertiv's quiet December 2025 Waylay acquisition [823ac77c-4115-4faa-ab14-abad495ac273] signals that whoever first delivers ML-driven power orchestration at MW scale captures the recurring, cycle-resistant margin the hardware build-out does not. Equally underrated: the transformer/switchgear lead-time wall is structurally tighter than the gas-turbine bottleneck because it throttles every datacenter regardless of generation source [5b0a3681-e14c-4c8f-98b8-b3ad8d9da306] — the market over-watches turbines and under-watches the substation.