Skip to content

LAYER 07 DATA CENTER MW · PUE

Last revised · MAY 13, 2026

What it takes to power a dense AI data center.

As racks get denser, data-center design shifts from floor space to substations, conversion losses, and cooling.

200 MW

campus scale

1.15

PUE target

7 yr

grid queue (PJM)

765 kV

substation feed

Native unit

MW · PUE

What constrains it

New AI capacity depends on power delivery, interconnect timing, and cooling infrastructure.

FIG. L07 · SIGNATURE DATA CENTER POWER MIX

Fit overview · pinch to zoom

FIG. L07
Power sources for new AI capacity, 2026–30
PROJECTED
Combined-cycle gas
40%
Aeroderivatives
18%
Reciprocating engines
12%
Solar + battery
10%
Bloom fuel cells
8%
Nuclear / grid
12%
Half of new capacity by 2030 will be behind-the-meter — expensive, but unblocked by interconnection queues and able to come online in months instead of years.
Projected supply mix for new AI capacity, 2026–30. The useful question is not just how much power arrives, but what can be delivered on time.

What this layer does

The data-center layer turns racks into a site plan. Power enters, changes form several times, and leaves as heat. Understanding that path explains why substations, cooling, and connection queues matter so much.

Follow the megawatt

Follow the megawatt

A data center is the machine around the machine.

The figure above is about supply. This guide shows what that supply has to survive before it becomes usable AI capacity inside the hall.

01 · Bring power in

The first constraint is permission to consume power.

New AI capacity starts with firm site power: interconnection approvals, transformers, substations, and often a waiting period before electrons ever reach the building.

Grid access

02 · Turn it into rack power

The building spends power before the GPUs see it.

Power is stepped down, converted, distributed through shelves and bus bars, and measured against PUE. Every conversion layer eats into the site’s useful IT capacity.

PUE matters

03 · Remove the heat

Dense racks become a thermal-design problem.

Rack-scale AI systems use liquid-cooling manifolds and cold plates because the waste heat is too concentrated to treat as ordinary room air.

Cooling is load

Why the figure matters

Supply mix changes because grid timing is slow.

When new connections lag, developers reach for behind-the-meter generation, storage, and other ways to make the site deployable sooner.

Useful ratio

PUE tells you how much non-compute overhead remains.

A PUE of 1.15 means every 1.00 MW of IT load needs about 1.15 MW at the facility boundary. The closer that ratio is to 1.0, the less power disappears into overhead.

Reader takeaway

The building is part of the accelerator.

If power arrival, distribution, or cooling slips, installed GPUs do not become useful capacity on schedule.

The suppliers are sold ahead of the sites

The firms that make this layer work are already working through order books that extend past the buildings being permitted now. GE Vernova ended Q1 2026 with $163 B of backlog and booked $2.4 B of data-center electrification orders that quarter — more than its Electrification segment booked in all of FY2025. Siemens Energy reported a Q2 FY2026 order backlog of €154 B, and Vertiv raised its FY2026 revenue guide to $13.5–14.0 B, roughly 30% above FY2025.

Outside the fence the picture is the same: Quanta Services reported a Q1 2026 backlog of $48.47 B and MasTec an 18-month backlog of $20.3 B. The seven-year PJM queue is one timer on a site. The order books in front of the turbine, switchgear, and substation suppliers are another, and the two clocks run on the same projects.

The hyperscalers rent half of what they run

Digital Realty sits on roughly 8 GW of total IT capacity — about 3 GW in place and more than 5 GW of future development (DLR March 2026 investor presentation). Q1 2026 backlog reached $1.8 B of annualized GAAP base rent at 100% share (DLR Q1 2026 release). The landlord is selling forward, too.

Iron Mountain leased 32 MW through April and has 400 MW energizing over the next 24 months (IRM Q1 2026 release). Equinix raised FY2026 capex guidance to roughly $4.1 B (EQIX Q1 2026 release). Three REITs, three order books, all feeding the site map the chapter just walked.

The site-walk above describes the building. It does not name the owner. The REIT layer is the missing landlord in the tour.

Heat removal is the third clock

Modine’s Q3 FY2026 data-center sales rose 78% year over year, and management guided FY2026 net-sales growth of 20–25% (Modine Q3 FY2026 release). The cooling order book is already running ahead of the buildings.

Munters booked an approximately BSEK 2.0 modular AI cooling order in late April with deliveries from early 2027 through Q1 2028 (Munters press release, 2026-04-24). nVent ended Q1 2026 with a $2.6 B backlog and raised its FY2026 organic guide to 21–23% (nVent Q1 2026 release). Same projects, separate suppliers.

The chapter already counted two clocks on a site. Heat removal is the third. Same project, separate order book.

A fast chip is only useful if the building around it can feed and cool it.