Where the world's
compute draws power.
A sourced index of hyperscale, colocation and gigawatt-class AI campuses across — countries. Every capacity figure is labelled as either IT load or grid power, and converted with an adjustable PUE model to estimate how many Blackwell-class GPUs it could host.
Power by region, and what it could run
About DC Atlas
DC Atlas tracks facilities: hyperscalers' own campuses, colocation and wholesale operators, telco and sovereign AI builds, and the new generation of dedicated AI campuses. Where an operator has published a capacity, the record carries it along with what that figure measures. Where nothing has been published, the record says undisclosed rather than guessing.
Totals only count facilities with disclosed capacity, so real figures are higher. Planned and under-construction capacity is included by default; use the status filter to see only what is operating today.
Records with a Verified badge were checked against the linked sources on . Compiled records come from public reporting and each has a one-click source search.
Facilities by country
| Country | Region | Facilities | Disclosed grid MW |
|---|
Update log
The dataset is refreshed automatically every week: new announcements, status changes and capacity revisions are researched, checked against sources and logged here.
Methodology & FAQ
What is the difference between grid power and IT load?
Grid power (sometimes called utility or contracted capacity) is everything the site draws from the grid or its own generation. IT load (critical load) is what reaches the servers, storage and network gear. The difference goes to cooling, UPS and transformer losses, lighting and so on. Operators announce one or the other, and AI campuses are usually quoted in grid gigawatts. Each record here says which one it carries.
What is PUE, and why can I change it?
Power Usage Effectiveness is grid power ÷ IT load. A PUE of 1.20 means 20% overhead. The default "by facility type" scenario uses 1.20 for hyperscalers, 1.25 for dedicated AI campuses (liquid-cooled, newer), 1.40 for colocation and 1.50 for telco and enterprise sites. The uniform 1.20 / 1.40 / 1.60 scenarios show how sensitive the totals are to that assumption.
How is "Blackwell-equivalent GPUs" calculated?
IT load ÷ 120 kW per GB200 NVL72 rack × 72 GPUs. That's about 600 GPUs per MW of IT load. The figure is a yardstick for scale, not an inventory: real sites mix CPUs, storage, older accelerators, TPUs and Trainium, and they keep headroom.
Why do some gigawatt campuses look smaller than their headlines?
Headlines often quote an eventual multi-phase ambition. Where a firmer phase is published (for example, a first phase or an independent estimate of what is energised), the record uses the firmer figure and mentions the larger ambition in its note.
Where does the data come from?
From operator announcements, regulatory and utility filings, satellite-based estimates (Epoch AI's data center work in particular), and reputable trade press. Map coordinates are city or campus level. Please verify anything before relying on it for investment or engineering decisions.