In short
- A hyperscale campus is built and run by one operator for its own cloud or AI, not rented out rack by rack.
- Cloud campuses run 100–750 MW; AI gigacampuses plan for 1–5 GW.
- Buildings repeat: 30–200 MW each, energised in phases as the substation grows.
- The grid, not the land, usually limits how fast a campus can grow.
What makes a data center hyperscale
Two things: scale and ownership. A hyperscale campus is designed, built and operated by a cloud or AI company such as Amazon, Microsoft, Google, Meta or Oracle for its own services. It uses standardised building designs repeated across the site and across the world, with power and cooling sized for tens of thousands of racks.
Colocation providers also build very large campuses, sometimes for a single hyperscale tenant. The line is blurred in practice; the term usually describes the size and the single-operator model rather than who owns the land.
Four sizes of data center
| Type | Power | Typical racks | Model |
|---|---|---|---|
| Enterprise · edge | 0.1–5 MW | 5–15 kW, air cooled | One building or room, owner-operated |
| Colocation | 5–50 MW | 10–30 kW | Multi-tenant, sold by rack, cage or suite |
| Cloud hyperscale | 100–750 MW | 15–40 kW | Several 30–150 MW buildings, one operator |
| AI gigacampus | 1–5 GW | 100 kW and more, liquid cooled | Buildings of 150 MW or larger, on-site power |
These are typical ranges, not definitions. The scale ladder on the home page shows the same steps.
Eight parts of every campus
- HV substation: transmission-voltage connection, transformers in N+1, room for more bays.
- Data halls: repeated buildings of 30–200 MW IT, single or multi-storey.
- Cooling plant: chillers, dry coolers or towers, and CDUs for liquid-cooled racks.
- Backup and bridge power: generators and batteries; on-site generation to energise before the grid is ready.
- Fibre and meet-me rooms: diverse routes to two or more carriers.
- Water and treatment: make-up water for evaporative systems or a closed loop.
- Roads and laydown: heavy access for transformers and space to build later phases.
- Security and operations: perimeter, gatehouse, stores and the operations centre.
Only about a quarter of a greenfield site sits under data halls; the land budget shows where the rest goes.
AI changes the scale
AI training clusters want tens of thousands of GPUs on one network, close together. That pushes rack densities past 100 kW, moves cooling to liquid and makes single buildings of 150 MW or more normal. Because the grid cannot deliver gigawatts quickly, AI campuses increasingly carry their own generation for the first phases. Epoch AI tracks 93 AI sites with 13.7 GW of IT power as of October 2026.
The campuses that set the scale
| Campus | Operator | Location | Size |
|---|---|---|---|
| Hyperion | Meta | Richland Parish, Louisiana | 2 GW first, up to 5 GW |
| Stargate Abilene | OpenAI · Oracle · Crusoe | Abilene, Texas | About 1.2 GW |
| Colossus 2 | xAI | Memphis, Tennessee | 946 MW IT |
| New Carlisle | Amazon | New Carlisle, Indiana | 910 MW IT |
| Fairwater Atlanta | Microsoft | Fayetteville, Georgia | 636 MW IT |
| Prometheus | Meta | New Albany, Ohio | 471 MW IT |
Figures as published; sources are listed in the landmarks section.
Questions about hyperscale
How big is a hyperscale data center?
Typically 100 to 750 MW on one campus for cloud, and 1 GW or more for the largest AI campuses, built in phases of several buildings.
Who are the hyperscalers?
The large cloud and AI operators: Amazon Web Services, Microsoft, Google, Meta, Oracle and, in Asia, Alibaba, Tencent and ByteDance, plus new AI builders such as xAI and OpenAI's partners.
What is the difference between hyperscale and colocation?
Hyperscale campuses are run by one operator for its own services. Colocation sites rent space and power to many customers, although some large colocation campuses are built for one hyperscale tenant.
How long does a hyperscale campus take to build?
A first building can be live in about two to three years if power is available. Full build-out over several phases often takes five years or more, paced by the grid.
Sources
- DCD (Hyperion), DMC (Abilene, Prometheus), Epoch AI database as of October 2026 (IT power).
- Size ranges are typical industry values, not formal definitions.