AI data centers

What do AI data centers do?

Tracked cohort on the record

They hold the computers that train and run AI, and give those computers the power, cooling, and wiring they need.

An AI data center is a building packed with powerful computers whose job is to train and run AI.

The building is there to give those computers three things at a huge scale: steady electricity, cooling to carry away the heat they make, and fast wiring to move data between them. That is really all it does.

We track the real sites that do this, and we keep what has been announced separate from what is actually up and running.

✓ Last verified not yet on record1 source

Find out instantly when it changes.

An AI data center does one job, at a huge scale. It runs the computers that train and use AI, and it gives those computers everything they need to keep going.

Inside are rows of servers packed with GPUs. These are the chips that do the heavy math AI needs.

Training a big AI model means running thousands of these chips together for days or weeks. Answering people's questions, called inference, means running them all the time. Both turn electricity into work and heat.

So the building around the chips is there to supply three things.

Power. A big site can use as much electricity as a small city, so a strong power hookup and sometimes its own power plant are a big deal.

Cooling. The chips get very hot, and that heat has to be carried away, using air, water, or a sealed loop of liquid.

Wiring. Training splits the work across many machines, so they are tied together with very fast links.

The words also cover a few types. Hyperscale sites are the huge single-company sites where most AI runs. Colocation sites rent space to many companies. Enterprise sites serve one company. Edge sites sit close to users to make things load faster.

We track mostly the big, AI-scale sites, along with their size and who runs them.

What we actually track

The definition is general. The record is specific. This is the set of AI-scale sites we track, with what has been announced kept separate from what is up and running today.

Campuses: 103
Facilities: 3,431
Operators: 861
Announced vs energized GW: 128.0 / 5.6

The full US count → The national map →

The main types of data center

The word covers a few kinds of site. These are the ones people usually want to tell apart. Full definitions are in the glossary.

Hyperscale

A huge site run by one company, like a cloud or AI company, for its own work. This is where most AI runs.

Colocation

A building where a company rents out space, power, and cooling to other companies that bring their own computers.

Enterprise, or on-site

A data center one company runs for its own use, usually smaller than a hyperscale site.

Edge

A small site placed close to users to make things load faster, instead of putting all the computing in one big place.

How we know this

The definitions here are plain descriptions of how the industry works, linked to our own glossary. The counts of sites, size, and operators come straight from the records on DEPLOY.

Lawrence Berkeley National Laboratory (US DOE)2024-12-01

We track the real AI-scale sites one by one, and we keep what has been announced separate from what is actually running. The counts here are the set we track, not a full census of the whole industry.

Common questions

What is an AI data center?

A building full of powerful computers, mostly GPU servers, built to train and run AI. It is different from a regular data center mainly in how much it packs in. AI puts far more power and heat into each rack, which is why it needs so much electricity and cooling. We track the real AI-scale sites one by one.

What is the difference between a data center and an AI data center?

A data center is any building that holds computers and storage. An AI data center is built for the packed GPU clusters that train and run AI, so it uses far more power and makes far more heat per rack. That is why AI has raised new worry about power and water. The basic job, running and cooling computers, is the same.

What are the types of data centers?

The main types are hyperscale (huge single-company sites, where most AI runs), colocation (rented space for many companies), enterprise or on-site (one company's own site), and edge (small sites near users to make things load faster). What we track leans to the big, AI-scale end, and we cover those sites one by one.

What is a hyperscale data center?

A very large data center run by one company, usually a cloud or AI company, for its own work, built to grow to a huge amount of computing and power. Hyperscale sites are where most big AI training and use happens, and they are behind most of the recent jump in data-center electricity use.

How many AI data centers are there?

We track a specific, reviewed set of AI-scale sites: big flagship sites with a full capacity record, plus the wider set of colocation and hyperscale buildings, run by hundreds of companies. The exact counts are on the how-many pages, where we keep what has been announced separate from what is actually running.

Keep reading

How many AI data centers are actually running?Answer Data centers How many AI data centers are in the US?Answer Data centers What are the biggest AI data centers?Answer Data centers Which companies operate the most AI data centers?Answer Data centers All verified answersHub Every standing question DEPLOY tracks.

Ask the registry

Track this answer

Find out instantly when it changes.

Canonical URL: /what-do-ai-data-centers-do · machine-readable mirror: /what-do-ai-data-centers-do.md