Back to atlas Power field guide2026

Understand the scale

From megawatts
to machine intelligence.

MW and GW describe the power a data center can draw at one moment. They are a capacity ceiling—not the amount of energy used over a day or year.

MW
1 megawatt

1,000 kilowatts
1,000,000 watts

× 1,000
GW
1 gigawatt

1,000 megawatts
1,000,000,000 watts

01

POWER VS ENERGY

How large is it?

Facility capacityEnergy at full loadApprox. IT power at PUE 1.2
100 MW2.4 GWh/day · 876 GWh/year≈ 83 MW for compute
1 GW24 GWh/day · 8.76 TWh/year≈ 833 MW for compute

PUE means Power Usage Effectiveness: total facility power divided by IT power. A PUE of 1.2 means every 1 MW used by computers needs about 0.2 MW more for cooling and electrical overhead.

02

ROUGH 2026 COST

What would it cost to build?

100 MW AI data center
$2.5–4.0B+

Rough fully equipped planning range

Shell, power & cooling
$1.0–1.5B
AI hardware & network
up to ~$2.5B
1 GW AI campus
$25–40B+

Usually delivered as many buildings and phases

Shell, power & cooling
$10–15B
AI hardware & network
up to ~$25B
These are order-of-magnitude estimates, not quotations. The model applies 2026 shell/core benchmarks and JLL's stated potential AI tech fit-out of up to $25M per MW. Land, a new power plant, grid upgrades, financing, taxes, and major fiber routes can add billions. Hardware generation and redundancy design can also move the total sharply.
PUBLIC-PROJECT CHECK

Google announced about $15B for a gigawatt-scale Vizag AI hub that also includes energy and subsea connectivity. OpenAI's broader Stargate commitment is $500B for 10 GW. These programs are not directly comparable, but they show why a single “cost per GW” can be misleading.

03

INSIDE THE CAMPUS

What does it actually contain?

Utility gridSubstation + UPSAI racksCooling + heat rejection

Grid & substations

High-voltage utility feeds, transformers, switchgear, and redundant distribution bring power safely to the campus.

UPS & backup

Battery UPS bridges short interruptions. Generators or on-site generation keep critical systems running during longer outages.

Cooling plant

Chillers, cooling towers, pumps, heat exchangers, and direct-to-chip liquid loops remove heat from dense AI racks.

AI compute racks

GPU or accelerator servers, CPUs, memory, rack power shelves, and high-speed interconnects perform model training and inference.

Network fabric

Ethernet or InfiniBand switches, optical transceivers, fiber, routers, and subsea or terrestrial links move data at very high speed.

Storage systems

Fast local flash, shared object storage, and backup systems continuously feed training data and preserve model checkpoints.

Safety & security

Physical access control, cameras, fire detection, clean-agent suppression, and cyber controls protect people and equipment.

Operations & DCIM

Sensors and Data Center Infrastructure Management software track temperature, power, water, failures, and capacity in real time.

SOURCE NOTES

Benchmarks behind the guide