OTAKAR.AI

Data Centers by the Numbers

The physical economics of AI compute — power, water, land, and money, in one place

Infrastructure economics: September 2026 · Figures are industry-reported ranges; exact numbers vary by hyperscaler, market, and cooling architecture

1The physical stack, in one picture

1. LandSite + zoning

2. GridInterconnect queue

3. PowerSubstation, switchgear

4. RacksGPUs, servers

5. HeatByproduct of compute

6. CoolingAir, liquid, water

The punchline: a data center is not primarily an IT problem. It is an industrial power-and-heat problem that happens to have servers in it. Every dollar and every year of delay traces back to one of two constraints: how many megawatts you can get, and how fast you can remove the heat they generate.

2The rack: where density exploded

THE MACHINERY

Rack power draw is set by GPU count and generation, not by the building. A GB200 NVL72 packs 72 GPUs and 36 CPUs into a single liquid-cooled rack; the next generation multiplies that further.

THE INTUITION

Think of it as swapping a household appliance for a small industrial furnace, and putting dozens of them in one room. The building didn't get bigger — the thing inside it got radically denser.

facility IT load (MW) = rack count × power per rack (kW) ÷ 1,000

Rack type	Power per rack	Cooling required
Traditional enterprise rack	7–10 kW	Air
NVIDIA H100 (8-GPU servers)	40–50 kW	Air / hybrid
NVIDIA H200 (8-GPU servers)	50–70 kW	Liquid-assisted
NVIDIA GB200 NVL72	120–140 kW	Direct-to-chip liquid
Vera Rubin NVL144 (targeted, 2026–27)	up to 600 kW	Liquid, required
OCP 2025 concept racks	up to 1 MW	Liquid, required

A 14x jump in one generation: a standard enterprise rack draws single-digit-to-low-teens kW. A GB200 NVL72 draws 120–140 kW. That single number is why the entire industry is redesigning around AI workloads rather than adapting existing buildings.

3The number that reorganizes everything: PUE

THE MACHINERY

Power Usage Effectiveness is total facility energy divided by the energy that actually reaches IT equipment. A PUE of 1.0 is theoretical; a PUE of 1.5 means the building spends an extra 50% of IT load on cooling, distribution loss, and overhead.

THE INTUITION

PUE is the tax the building itself charges on top of the compute you actually wanted. Every tenth of a point above 1.0 is pure overhead — power spent moving and removing heat rather than running GPUs.

PUE = total facility energy (kWh) ÷ IT equipment energy (kWh)

Facility class	Typical PUE
U.S. average, 12-month trailing	~1.4
Typical enterprise facility	1.5–1.6
New-build target (industry benchmark)	≤1.2
Leading hyperscale (Google fleet-wide)	1.09
Best individual sites (Dublin, Eemshaven)	1.07–1.08

4The other utility bill: water

THE MACHINERY

Water Usage Effectiveness measures liters of water consumed per kWh of IT energy. Evaporative cooling towers trade electricity for water; closed-loop liquid and immersion cooling trade water for capital and complexity.

THE INTUITION

PUE and WUE pull in opposite directions. A facility can post an excellent power number by evaporating enormous amounts of water — the efficiency didn't disappear, it just moved from the electric meter to the water main.

WUE = water consumed (liters) ÷ IT equipment energy (kWh)

Cooling method	WUE (L/kWh)	PUE
Air cooling	~0	1.4–1.8
Evaporative cooling	1.5–2.5	1.1–1.3
Liquid / direct-to-chip	near 0	1.05–1.2

The trade nobody advertises: globally, average WUE runs near 1.8–1.9 L/kWh. U.S. data centers consumed roughly 17 billion gallons of water directly for cooling in 2023 — on top of the far larger water footprint embedded in the electricity generation itself.

5What a megawatt costs to build

THE MACHINERY

Construction cost is quoted per MW of IT capacity, not per square foot. Shell-and-core cost has risen from $7.7M/MW in 2020 to roughly $11.3M/MW in 2026 — before AI-specific fit-out.

THE INTUITION

The building is almost the cheap part now. What you're really pricing is transformers, switchgear, and liquid-cooling plumbing — the industrial guts, not the walls and roof.

total build cost ≈ $/MW (market rate) × MW of IT capacity

Market	Shell-and-core, $/MW
Low-cost U.S. (Texas, Ohio, Nevada)	$8M
Global average, 2026	$11.3M
Northern Virginia, Silicon Valley	$14–18M
Tokyo, Singapore (priciest markets)	~$15M
All-in AI build (incl. GPU fit-out)	$30–40M

AI tenant fit-out — liquid cooling, higher-density power distribution, GPUs — can add up to $25M/MW on top of shell-and-core.

6Where the capex actually goes

THE MACHINERY

Power infrastructure — switchgear, transformers, medium-voltage cable, UPS, and interconnection — typically consumes 40–50% of total construction cost, more than the building envelope itself.

THE INTUITION

You are not really building a data center. You are building a substation with a very expensive computer room attached to it.

Cost component	Range, $/MW
Shell and core	$3–5M
Mechanical / cooling systems	$2–4M
Electrical / power distribution	$3–5M
Fire suppression and controls	$0.5–1M
Land, interconnection, permitting (separate)	$1–10M+

7Anatomy of a megawatt

Illustrative at PUE 1.5, a common facility average; a leading-edge liquid-cooled site (PUE 1.1) shrinks the cooling and loss bars substantially.

8The real bottleneck isn't chips — it's the grid

THE MACHINERY

FERC targets 8–11 months for a generation project to clear the interconnection queue. Actual timelines in high-demand zones run 36–84 months — longer than the 12–24 months it takes to build the data center itself.

THE INTUITION

You can pour concrete and stack racks faster than you can get permission to plug them in. The constraint has flipped from "can we build it" to "can we get power to it."

Grid / region	Queue or wait
FERC target (generation interconnection)	8–11 months
PJM Interconnection, average	~40 months
Active data-center load zones, U.S.	36–48 months
National U.S. queue (all fuels)	2.2–2.6 TW waiting
ERCOT (Texas) large-load queue	410 GW; 87% data centers
New 50 MW site, London / Amsterdam	~8–10 years

Scale check: the U.S. interconnection queue now holds nearly twice the capacity of all currently installed U.S. generation. The bottleneck isn't a shortage of proposed power — it's the speed at which any of it can be approved and connected.

9The workaround: build power like it's the product

THE MACHINERY

Faced with multi-year grid delays, developers increasingly build generation "behind the meter" — on-site, bypassing the interconnection queue entirely. Over 130 GW of behind-the-meter capacity has been proposed for U.S. data center projects.

THE INTUITION

If the utility can't connect you in time, you become your own utility. It costs more per megawatt-hour, but "more expensive and on time" beats "cheaper and four years late."

Technology	Deployment time
Fuel cells	~90 days
Gas turbines (heavy-duty)	3–7 years

behind-the-meter power: $100–165/MWh vs. grid: $90–95/MWh

Gas accounts for over 80% of the proposed behind-the-meter pipeline — the market is selecting for speed, not fuel type.

10Floor space inverts

THE MACHINERY

Because AI racks draw an order of magnitude more power each, a fixed number of megawatts now fits into far fewer racks — and a much smaller building footprint per megawatt.

THE INTUITION

Data centers aren't getting bigger buildings — they're getting smaller, denser ones that draw vastly more power per square foot than anything built a decade ago.

Facility type	Racks for 100 MW	Sq ft per rack
Legacy enterprise (8 kW/rack)	~12,500	25–30
GB200 NVL72 (120–140 kW/rack)	~700–800	30–50

11The scale of the buildout

THE MACHINERY

The five largest hyperscalers are projected to spend $745–775B in capex in 2026 alone. Roughly 100 GW of new data center capacity is projected between 2026 and 2030 — doubling today's installed global base.

THE INTUITION

This is being described as the largest five-year surge to the U.S. grid since the 1980s. The AI buildout is now, functionally, an electricity infrastructure project that happens to run models.

Metric	Value
Top-5 hyperscaler 2026 capex	$745–775B
New global capacity, 2026–2030	~100 GW
U.S. data centers, share of electricity (2024)	~4%
U.S. data centers, projected share (2028)	7–12%

12The one-megawatt mental model

Think like an infrastructure investor, not an IT buyer:

A megawatt is not merely a unit of compute capacity.
It is a unit of land, grid queue position, heat to be removed, water rights, and ultimately years of lead time.

Cost per useful compute-hour ≠ chip price alone

It depends on rack density, PUE, WUE, market-specific $/MW, interconnection timeline, and whether power is grid-sourced or built behind the meter.

Five questions to ask about any data center deal

Where is the power coming from? Grid-queued, already interconnected, or behind-the-meter — and how long did that take?

What's the PUE and WUE together? A good power number achieved through heavy water use isn't free efficiency.

What's the rack density, and does the site support it? Air-cooled buildings cannot simply be retrofitted for 120+ kW racks.

Is the $/MW figure shell-and-core or all-in? The gap between $11M and $35M per MW is the AI fit-out.

What happens to the queue position if the project slips? Interconnection slots, transformers, and switchgear all have multi-year lead times of their own.

Rules of thumb

The grid is now slower than construction. A 12–24 month build can sit behind a 36–84 month interconnection queue.

Power infrastructure is the site. It's 40–50% of construction cost — more than the building.

Density cuts both ways. Fewer racks and less floor space, but each rack becomes a liquid-cooling and power-delivery problem.

PUE and WUE trade against each other. Chasing one in isolation just moves the cost to a different meter.

"Announced megawatts" and "energized megawatts" are different numbers. 30–50% of the pipeline typically slips a year or more.

Source note: figures are drawn from industry and analyst reporting (JLL, BloombergNEF, Carbon Direct, NREL, hyperscaler sustainability disclosures, and infrastructure trade press) current as of 2026. Individual facilities vary substantially by market, cooling architecture, and disclosure practice.