What is an AI data center, and how is it different from an ordinary one?

An AI data center is a building full of accelerators, meaning GPUs or specialised AI chips, drawing electricity at a density conventional data centers were never designed to handle. That single difference, power per rack, drives everything else about how they are built, cooled and sited. The International Energy Agency puts all data center electricity use at about 415 TWh in 2024, roughly 1.5 per cent of global demand.

What makes it different from an ordinary data center?

Density, and the consequences of density.

A conventional server rack draws a certain amount of power and gives off a corresponding amount of heat, and the building's electrical supply and air conditioning are designed around it. An accelerator rack draws several times that. Multiply across a hall and the electrical capacity required per square metre rises to a level most existing facilities cannot supply.

Heat follows power exactly. Every watt of electricity that goes in comes out as heat that has to be removed, and above a certain density moving air past the components stops being sufficient. That is the reason liquid cooling has gone from exotic to standard in new builds, and the reason retrofitting an older facility is often impractical rather than merely expensive.

Networking is the third difference and the least visible. A training job runs across many machines at once, so the speed at which those machines can talk to each other becomes a limit on the whole job. Specialised high-bandwidth interconnect is part of what distinguishes these facilities from a building full of ordinary servers.

What is inside one?

Four systems, and the hard engineering is in the last two.

Compute, which is racks of accelerators plus the conventional servers that feed them data and coordinate the work. Storage, because training needs vast datasets delivered fast enough to keep expensive chips busy, which makes storage throughput a limiting factor rather than an afterthought.

Then power. A facility needs a connection to the grid at the right capacity, transformers and distribution to get it to the racks, and uninterruptible power supplies and generators to ride through interruptions. At accelerator densities all of that equipment is larger and more expensive per square metre than in a conventional build.

And cooling, which is the system most likely to determine whether a building can host this workload at all. Air handling, chillers, heat exchangers, and in new builds a liquid loop that runs to the rack or to the chip itself. The cooling plant is frequently a substantial share of both the capital cost and the floor area.

How much electricity does this actually use?

The best public numbers come from the IEA's Energy and AI report, and they are worth quoting precisely because they are so often inflated.

Data center electricity consumption was about 415 TWh in 2024, which the agency puts at roughly 1.5 per cent of global electricity demand. Consumption has grown about 12 per cent a year over the previous five years, considerably faster than overall electricity demand.

The base case projects about 945 TWh by 2030, just under 3 per cent of global electricity consumption. That is roughly a doubling of the sector over six years.

One caveat belongs with every one of those figures. They cover data centers in total, including search, streaming, cloud storage, enterprise software and everything else that runs in a rack. Quoting 415 TWh as AI's electricity use is wrong, and it is one of the most common errors in coverage of this subject.

Where the growth is concentrated

Two countries, overwhelmingly. The United States and China together represent close to 80 per cent of the projected global growth in the IEA's analysis.

US consumption is expected to increase by as much as 240 TWh, about 130 per cent above 2024 levels. China's is expected to grow by 175 TWh, about 170 per cent above its own baseline.

That concentration explains why a subject that sounds global is in practice decided by grid connection queues in a handful of American states and by electricity planning in China. It also explains why the political argument about data centers is loudest in exactly those places.

How uncertain are these projections?

Considerably, and the report says so by publishing a range rather than a number.

Its Lift-Off case, which assumes faster AI adoption, puts 2035 demand around 45 per cent above the base case and exceeding 1,700 TWh. Its High Efficiency case, which assumes stronger efficiency gains, reaches around 970 TWh by 2035.

Those two scenarios differ by roughly a factor of 1.75 for the same year. Anyone quoting a single 2035 figure for data center electricity is choosing a scenario, and the honest form of the claim names which one.

What the buildings are actually optimised for

Two different workloads, which pull the design in different directions.

Training is one long job spread across many accelerators, running at sustained high power for weeks. It cares about interconnect bandwidth and about reliability, because a failure partway through is expensive, and it does not care where in the world it happens.

Inference is a constant stream of short requests, measured in tokens in and out. It cares about latency, which means being near the user, and about scaling up and down with demand, which produces a far more variable electrical load.

A facility tuned for one is not tuned for the other, which is why operators increasingly separate them and why announcements specify which kind of capacity is being built.

Why they get built where they get built

Power availability, first and by a distance. The binding constraint on a new facility is usually how quickly the site can be connected to enough electricity, which is why announcements cluster near generation and existing transmission rather than near population.

After power come land at the right scale, water or another cooling resource, network connectivity, and tax treatment. Latency to end users matters for inference and barely at all for training, which is why training capacity can be sited in places no consumer product would choose.

The local politics follow from the economics. A large facility is a very large capital investment that employs a substantial temporary construction workforce and a modest permanent technical staff, because automation is the point of the design. That asymmetry between money in and jobs created is why incentives offered to attract these buildings attract scrutiny.

What to watch rather than the headline number

Three things that tell you more than a TWh figure.

Grid interconnection timelines, because a site with chips and no connection is a car park. Queues measured in years are the real limit on how fast capacity arrives, regardless of how much capital wants to be deployed.

Efficiency per unit of work, rather than total consumption. The IEA's own High Efficiency scenario shows how much of the 2035 range is decided by this, and the useful question about a new facility is what it delivers per watt rather than how many watts it draws.

And whether new capacity brings new generation with it. A facility supplied by newly built generation has a different effect on a grid and on prices than one drawing from existing supply, and that distinction is where most of the substantive local argument sits.

The short version

An AI data center is defined by power density rather than by size or by software. Accelerators draw far more per rack than general-purpose servers, which forces more electrical capacity, liquid cooling and specialised networking, and makes most older facilities unsuitable.

On electricity, use the IEA figures and attribute them accurately: about 415 TWh for all data centers in 2024, roughly 1.5 per cent of global demand, projected to about 945 TWh by 2030 in the base case, with the United States and China accounting for close to 80 per cent of the growth.

This article reports published projections and is not a forecast. Figures were verified on 20 October 2026.