Divergent Compute.AI Economic Think Tank

Markets / Industries / Cooling & Electrical

Industry view · Cooling & Electrical

AI made silicon cheap to buy and impossible to cool or power

The bottleneck in artificial intelligence has migrated from chips to the unglamorous physical layers beneath them: thermal management and electrical distribution. Data-center electricity use jumped 17% in 2025 and high-power transformer lead times now stretch to as long as five years, turning cooling and electrical suppliers into the gatekeepers of the AI build-out.

17%
2025 jump in data-center electricity use (vs 3% global)
IEA
~945 TWh
projected data-center electricity demand by 2030
IEA
$15B
Vertiv AI-infrastructure order backlog, end-2025
Vertiv / DCF
120kW
per-rack power of NVIDIA GB200 NVL72, air-cooling not viable
NVIDIA / The Register

01 · The thesis

The constraint moved downstream, and the money followed

For two years the AI trade was about GPUs. In 2025-2026 it is about whether you can power and cool them. NVIDIA's GB200 NVL72 rack draws roughly 120kW and ships with no air-cooled variant — direct-to-chip liquid cooling is mandatory, not optional. That single architectural fact has converted thermal management from a facilities line-item into a strategic chokepoint, and pulled an entire electrical supply chain — transformers, switchgear, UPS, busways — into multi-year backlog.

The scarce input is no longer compute; it is the grid interconnect and the equipment that conditions power on its way to the chip. The IEA reports wait times for transformers and cables have doubled in three years, and analysts warn a meaningful share of planned 2026 data-center projects face delay or cancellation for lack of electrical hardware. In that world, the companies that make CDUs, cold plates, medium-voltage switchgear and grid transformers hold pricing power that the chip-buyers do not.

1Grid & Transformers

The five-year wait

High-power transformer lead times have blown out from 24-30 months pre-2020 to as long as five years, with costs up 70-100% by 2025 — the hardest physical constraint on AI build-out.

GE Vernova, Hitachi Energy, Siemens Energy
2Power Distribution

Grid-to-chip integration

Switchgear, UPS and busway makers are co-designing reference architectures with NVIDIA, collapsing the boundary between the substation and the rack.

Eaton, Schneider Electric, ABB, Vertiv
3Thermal — Liquid

Air cooling hits its ceiling

At 120kW+ per rack, direct-to-chip and immersion cooling shift from niche to mandatory; the segment is the fastest-growing layer of the AI stack.

Vertiv, Schneider/Motivair, ZutaCore, JetCool
4AI-Optimized Ops

Software wringing out the watts

Reinforcement-learning controllers tune chillers, pumps and setpoints in real time — DeepMind cut Google's cooling energy ~40% — turning cooling itself into an AI workload.

Hyperscaler in-house teams, DCIM vendors
5Heat Reuse & Siting

Where the heat goes next

Waterless and warm-water designs plus heat-reuse are emerging as siting and sustainability differentiators as water and power permits tighten.

ZutaCore, Iceotope, colocation operators
Pace of AI disruption by stage — Divergent Compute assessment

02 · Public players & exposure

Who routes through, who gets routed around

We plot the listed players on two editorial axes — how exposed each is to AI disruption, against how ready its data, brand and position are to be the answer. The figures in the table are sourced; the placement is our read.

Positioning — editorial assessment, not a sourced metric. Bubble = approximate relative scale.
CompanyStanceThe sourced fact
Vertiv HoldingsVRTCategory leaderPosted $10.2B revenue in 2025 (+28% YoY) and ended the year with a $15B order backlog, Q4 orders up 252% YoY; liquid-cooling revenue more than doubled in Q1 2025.
EatonETNGrid-to-chip playRecord 2025 sales of $27.4B; data-center revenue grew ~40% in Q4 with orders up ~200%, and Electrical Americas backlog hit a record $13.2B.
Schneider ElectricSU.PADesign lock-inAcquired control of Motivair in early 2025 and launched an end-to-end liquid-cooling portfolio including a 2.5MW CDU; Motivair tech runs in six of the world's top-ten supercomputers.
GE VernovaGEVTransformer gatekeeperCalled Q4 2025 its largest-ever hyperscaler quarter in Electrification; Q1 2026 data-center orders of $2.4B already exceeded the full 2025 total, with total backlog at $163B.
ABBABBN.SWMV switchgearLaunched the AI-ready MegaFlex UL 415V UPS for large-scale data centers in June 2025, competing with Eaton in medium-voltage switchgear for AI loads.
Hitachi EnergyHTHIYCapacity-constrainedUndertaking 2025-2026 transformer capacity expansions to ease grid-component shortages, including large multi-billion-dollar grid-infrastructure commitments with hyperscalers.
ZutaCorePRIV-ZUTATwo-phase challengerRaised a $100M+ Series C in 2026 from Mitsubishi Electric, Carrier Ventures and Samsung Ventures (~$600M valuation) for waterless direct-to-chip two-phase cooling.
IceotopePRIV-ICEPrecision liquidClosed a $26M Series B in 2025, taking total funding to ~$81.4M — the highest-funded pure-play in the data-center cooling-systems category per Tracxn.
JetCoolPRIV-JETMicroconvective betRaised $17M for microconvective direct-to-chip cold plates targeting higher heat-transfer efficiency at the chip surface.
Legacy air-cooling OEMsAIRStranded architectureAt 120kW+ racks the GB200 reference design has no air-cooling variant, structurally eroding demand for air-only CRAC/CRAH product lines.
The map is Divergent Compute’s editorial positioning, offered as a lens, not a measurement. Every figure in the right-hand column is drawn from a named source — see Sources.

03 · The two clocks

The spend, and the payoff

Three timers are running against the AI build-out — and cooling and electrical suppliers sit on all three.

Disclosed order backlogs / contracted figures, end-2025 to Q1-2026. Sources: Vertiv, Eaton, GE Vernova investor disclosures.

The interconnect clock. Average grid-connection wait times in primary data-center markets now exceed four years, and Microsoft has acknowledged GPUs sitting idle in inventory because it cannot find electricity to power them. Compute is no longer the binding constraint; the wire is.

The transformer clock. Lead times for high-power transformers have stretched from 24-30 months before 2020 to as long as five years, with costs up 70-100% by 2025. Analysts warn that more than half of some 2026 U.S. data-center plans risk delay or cancellation for lack of electrical equipment.

The thermal clock. New greenfield AI builds in 2025-2026 are specifying 250-400kW per cabinet row as baseline. Direct-to-chip and immersion cooling are projected to grow from roughly $5.3B in 2025 to over $32B by 2032 — the fastest-compounding physical layer of the stack.

04 · Private flagships

The AI-native challengers

The companies attacking this industry AI-first, with disclosed funding where available:

Vertiv

Power & thermal full-stack leader

The clearest pure-play on AI's physical layer, spanning UPS, busway, CDUs and liquid cooling; liquid-cooling revenue is guided to ~40% CAGR through 2028.

Public (NYSE: VRT); $10.2B 2025 revenue, $15B backlog

Eaton

Grid-to-chip power architecture

Partnered with NVIDIA on the Beam Rubin DSX grid-to-chip platform and raised incremental capacity investment to ~$1.5B for transformers, switchgear and distribution.

Public (NYSE: ETN); record $27.4B 2025 sales

Schneider Electric

Electrical + liquid cooling via Motivair

Used the early-2025 Motivair deal to bolt a credible liquid-cooling portfolio onto its dominant electrical-distribution franchise, betting on design lock-in.

Public (EPA: SU); Motivair control acquired 2025

GE Vernova

Transformers and grid equipment

Owns one of the scarcest inputs in the chain; its electrification backlog has more than quadrupled in four years and is set to double again by 2028.

Public (NYSE: GEV); $163B total backlog

ZutaCore

Waterless two-phase cooling

Strategic-investor-backed challenger pushing dielectric two-phase direct-to-chip cooling that roughly halves cooling energy and removes water from the loop.

~$200M total; ~$600M valuation (2026 Series C)

Iceotope

Precision immersion / liquid

British precision-liquid specialist positioned for the density classes where direct-to-chip alone struggles; highest-funded pure-play cooling startup.

~$81.4M total funding (2025 Series B)

05 · Signals

What moved, and what to watch

2025

Liquid cooling crosses from optional to mandatory

NVIDIA's GB200 NVL72 ships with a pre-plumbed liquid manifold and no air-cooled variant at ~120kW/rack, forcing the chassis and facilities ecosystem to pivot.

2025

Transformer scarcity becomes the build-out's hard ceiling

IEA flags transformer and cable lead times doubling in three years; high-power units now quoted at up to five years against sub-18-month AI deployment cycles.

2025-26

Electrical orders go vertical

Eaton data-center orders up ~200% in Q4 2025; GE Vernova's Q1-2026 data-center orders alone exceed its entire 2025 total; Vertiv Q4 orders up 252%.

2026

Strategic capital floods cooling startups

ZutaCore raises $100M+ from Mitsubishi, Carrier and Samsung; incumbents (Schneider/Motivair) buy rather than build to close the thermal gap.

2026

Power, not silicon, gates deployment

Hyperscaler capex approaches ~$650B with grid-connection waits over four years; Microsoft admits idle GPUs waiting on electricity.

06 · The exposure read

Who’s defensible, who’s at risk

AI rewards clean, structured advantage and punishes friction. The line runs through who owns the data, the brand and the customer — and who is merely a step the technology can route around.

Defensible

  • Full-stack power-and-thermal incumbents — Vertiv, Eaton and Schneider convert AI density into multi-year backlogs and pricing power, because hyperscalers now buy integrated grid-to-chip systems, not components.
  • Transformer and grid-equipment makers — GE Vernova and Hitachi Energy own the scarcest physical input; quadrupled backlogs and 70-100% price increases accrue to whoever can actually ship steel.
  • Liquid-cooling specialists with strategic backers — ZutaCore, Iceotope and JetCool ride the shift from air to direct-to-chip, with incumbents now paying up to acquire or fund them.
  • AI-ops software layers — the same reinforcement learning that drives the load can cut cooling energy ~40%, rewarding DCIM and control vendors that turn thermal management into a model.

At risk

  • Air-only cooling OEMs — product lines built around CRAC/CRAH and 40kW-rack assumptions face structural demand erosion as reference designs above 120kW ship liquid-only.
  • Data-center developers without secured power — projects relying on uncommitted grid interconnects or undelivered transformers risk multi-year delay or outright cancellation.
  • Sub-scale cooling startups without strategic capital — in a market where incumbents are acquiring and $100M rounds are setting the pace, thinly funded challengers risk being out-built and out-distributed.
  • Hyperscalers exposed to the electrical bottleneck — idle GPUs waiting on electricity convert capex into stranded assets, pressuring returns until power and cooling catch up.
The AI narrative is written in silicon, but the binding constraints of 2025-2026 are copper, dielectric fluid and grain-oriented steel. Until transformer lead times compress and liquid cooling scales, the cooling-and-electrical layer — not the GPU — sets the clock speed of the entire build-out. That makes it the least glamorous and most defensible position in the AI economy.

Sources

Where this comes from

IEA — Data-centre electricity use surged in 2025  ·  IEA — Energy and AI, Energy demand from AI  ·  Vertiv AI infrastructure surge / $15B backlog — Data Center Frontier  ·  Vertiv $15B backlog & liquid-cooling detail — Seeking Alpha  ·  Eaton Q4/FY2025 record results press release  ·  Eaton 200% data-center order growth, Beam Rubin DSX — TIKR  ·  Schneider Electric + Motivair liquid-cooling portfolio  ·  GE Vernova backlog surges on AI/grid demand — Yahoo Finance  ·  Transformer shortage delaying data centers — Energy News Beat  ·  NVIDIA 120kW DGX GB200 NVL72 rack — The Register  ·  NVIDIA GB200 NVL72 cooling requirements — ToneCooling  ·  ZutaCore $100M Series C — SiliconANGLE  ·  Iceotope $26M Series B — The Next Web  ·  DeepMind AI cuts Google cooling energy 40% — Google DeepMind