What is a data center?

What a data center does

A data center does three jobs at once: it computes, it stores and it connects. Servers run software, whether that is a bank's payment system, a video platform, a company's email or an AI model answering a question. Storage systems hold the data those programs need. Network equipment links the servers to each other and, through telecom carriers and internet exchanges, to the people using them.

What makes a building a data center rather than an office with a server cupboard is that it is engineered so those jobs never stop. Power arrives over more than one path and is backed up by batteries and generators. Heat from the equipment is removed continuously. Access is controlled, and the network has several independent routes out. The purpose of a data center is availability: keeping digital services running through power cuts, equipment failures and maintenance.

People search for what a data center is used for because the answer is now almost everything. Cloud computing, streaming, online shopping, mobile apps, government records, hospital systems and the training and serving of AI models all run in these buildings. The difference between them is mostly scale and ownership, which is what the types below describe.

What a data center looks like inside

From the outside most data centers are windowless, low and long, surrounded by fences, generator yards and cooling equipment on the roof or along the walls. Inside, the layout follows the flow of electricity and heat.

The data hall

The data hall, often called the white space, is where the computers live. Servers are mounted in standard metal cabinets called racks, set out in long rows. Rows face each other in pairs so that cool air is supplied to one aisle and hot exhaust is collected from the next, the hot-aisle and cold-aisle arrangement. In many halls one of the two aisles is enclosed with doors and a roof to stop hot and cold air mixing. Cables run overhead in trays or under a raised floor.

How much power one rack draws sets almost everything else. In the Uptime Institute's 2024 survey, average densities were rising gradually, with more facilities reporting 7–10 kW per rack, while operators building for AI were planning cabinets far above that, in some cases over 100 kW.

The power train

Large sites take power from the grid at high voltage through their own substation. Switchgear divides it into separate paths. Each path passes through an uninterruptible power supply (UPS), whose batteries bridge the seconds between a grid failure and the moment diesel or gas generators start and take the load. Power distribution units in the hall then feed each rack, normally from two independent sources so a single fault does not switch anything off. This electrical and mechanical plant is sometimes called the grey space.

Cooling

Nearly all the electricity a server uses becomes heat. Air-cooled halls use computer room air handlers that pass air over chilled-water coils; the water is cooled again by chillers, dry coolers or evaporative cooling towers outside. Denser AI equipment increasingly uses liquid cooling, in which coolant is piped to cold plates on the chips or the servers are immersed in fluid. In the atlas, 303 facilities are recorded with liquid cooling, and 151 of them are still under construction or announced.

Network and security

A meet-me room is where telecom carriers, internet exchanges and cloud on-ramps bring their fibre into the building and customers connect to them with cross-connects. Security is layered: perimeter fences, guarded entrances, mantraps, badge and biometric readers at each hall, cameras throughout and locked cabinets. Visitors rarely see the servers they came to inspect without an escort.

Types of data center

Data centers are usually grouped by who owns them and who uses them. The categories overlap, and one campus can serve several purposes.

Enterprise data centers

Built and run by one organisation for its own systems: a bank, a retailer, a university, a government department. They range from a few racks in a dedicated room to large private campuses. Many enterprises have moved part of this load into colocation and the cloud, but regulated industries still keep significant private capacity.

Colocation data centers

A colocation operator builds and runs the building, power, cooling and connectivity, and rents space and power to many customers who install their own equipment, by the rack, cage, suite or whole building. Colocation sites tend to be rich in network connections because many carriers and clouds meet there. Most of the operators in this atlas run colocation or wholesale facilities.

Hyperscale data centers

Hyperscale refers to the very large facilities built by or for the biggest cloud and internet companies such as Amazon Web Services, Microsoft, Google and Meta. They are designed to scale out with identical buildings on one campus, often tens to hundreds of megawatts each, and some are leased from wholesale developers rather than owned.

Edge data centers

Edge sites are small facilities placed close to users or devices to cut latency: telecom exchanges, modular containers at cell towers, or small halls in secondary cities. They are often under 1 MW. In the atlas, 125 facilities publish a capacity of 1 MW or less.

AI data centers

AI data centers are built around clusters of GPUs or other accelerators that train and run AI models. Because these chips draw so much power in a small space, AI sites need very large grid connections, liquid cooling and stronger floors. 558 facilities in the atlas are flagged as AI-ready by their operator or by the source document, and 287 of those are still under construction or announced, which is where most new AI capacity sits today. Prices and the census of AI compute campuses are tracked separately at computecensus.com.

How big are data centers?

Data center size is quoted in megawatts of power, not square metres, because power is the scarce input and the thing customers pay for. The figure usually means IT load, the power available to the servers themselves, though some developers quote total campus or grid capacity.

Of the 6,017 facilities in the atlas, 1,701 have a published power figure. Their median is 15 MW; 59% are 20 MW or smaller, and 207 exceed 100 MW. The direction of travel is clear in the medians by status: 9 MW for operational facilities, 50 MW for those under construction and 90 MW for announced projects.

The largest published campuses

The biggest numbers belong to campuses that are not finished. These are the twelve largest published capacities in the atlas. 12 of the twelve are flagged AI-ready in their source, and none of them is operating yet at its full published size; campuses of this scale are built in phases over several years.

Among facilities already operating, the largest published figures are Switch The Citadel Campus (TAHOE RENO 1) (850 MW), IndoData Center Bogor (500 MW), EdgeCore Phoenix (Mesa) Data Centers (496 MW), Switch The Core Campus (Las Vegas) (495 MW) and STACK NVA02 (420 MW). The full list of projects is on the projects tracker.

How much power do data centers use?

The International Energy Agency estimates that data centres consumed around 415 terawatt-hours of electricity in 2024, about 1.5% of global electricity use, and projects consumption roughly doubling to around 945 TWh by 2030 in its base case, with the United States and China accounting for nearly 80% of that growth.

In the United States the share is already higher. A Lawrence Berkeley National Laboratory report for the Department of Energy found that data centers used about 4.4% of US electricity in 2023 and could use between 6.7% and 12% by 2028, or 325 to 580 TWh.

Inside a facility, servers take the largest share. The IEA puts servers at around 60% of demand in modern data centres, while cooling ranges from about 7% in efficient hyperscale sites to more than 30% in less efficient enterprise ones. The industry's efficiency measure is power usage effectiveness (PUE): total facility energy divided by the energy used by IT equipment. A PUE of 1.0 would mean every watt reaches the servers. The Uptime Institute's 2024 survey put the average reported PUE at 1.56, flat for five years.

A useful way to translate megawatts into energy: a facility drawing 100 MW without pause for a full year would use 876,000 MWh (100 MW × 8,760 hours). Real sites run below their rated capacity, so actual consumption is lower, but the arithmetic explains why a single campus can matter to a regional grid, and why grid operators are writing new rules for large loads. Australia's market rule-maker, for example, has drafted connection standards for data centres of 30 MW and above; that and other rules are tracked on the Australia page.

How much water do data centers use?

Water is used mainly for cooling. Evaporative cooling towers and adiabatic coolers save electricity by letting water evaporate, which consumes it; closed-loop and air-cooled designs use far less water on site but more power. There is also indirect water use at the power stations that generate the electricity.

The same Berkeley Lab report estimated that US data centers directly consumed about 66 billion litres (roughly 17 billion gallons) of water in 2023, with far more used indirectly through electricity generation. Litres per day, the WUE metric, cooling designs and the rules now written around water are covered in the guide how much water does a data center use. Water has become a planning issue in places where data centers cluster. In Malaysia's Johor, the state government said it would stop approving Tier 1 and Tier 2 data centres because of their water use and now requires reclaimed water for new projects. See the Malaysia page for the full list of restrictions there.

Reliability tiers

The most widely used reliability scale is the Uptime Institute's Tier Classification System. Tier I provides basic capacity with a UPS, cooling and a generator but no redundancy. Tier II adds redundant capacity components. Tier III is concurrently maintainable: any component can be taken out for maintenance without shutting down IT. Tier IV is fault tolerant, with physically separate systems so that a single failure does not interrupt IT. 1,049 facilities in the atlas list an Uptime Tier certification among their credentials; most are Tier III.

Where are data centers?

Data centers cluster where power, fibre, land and customers meet: around major cities, internet exchanges and, increasingly, wherever large grid connections can be found. The atlas covers 183 countries. The ten with the most facilities are below.

Browse every country on the countries index, see them in space on the interactive map, or follow what is being built on the projects tracker, which currently lists 581 projects.

Questions

The IEA, Berkeley Lab, US Department of Energy, Uptime Institute and Malaysian and Australian documents quoted above are listed on the Sources page.