What is a datacentre?

A datacentre is a building whose job is to keep computers running without interruption. It takes in electricity, turns almost all of it into computation and heat, carries the heat back out, and connects the machines inside to the networks outside. Everything else in the building, from the generators in the yard to the locks on the doors, exists to protect those three flows: power in, heat out, and data in and out.

That description fits a server room in an office basement and a campus the size of a small town. What separates them is scale, and the way the scale is paid for. This piece describes the parts every datacentre shares, how its capacity is measured and sold, who owns and runs it, and what changes when the tenant is an AI cluster rather than a general-purpose cloud.

The parts every datacentre shares

Every datacentre, whatever its size, is built from the same layers.

• The site and the grid connection. Land, planning consent and, above all, a connection to the electricity grid, with a substation that steps transmission voltage down to something the campus can use. On a new site this is usually the slowest part to deliver.
• The electrical plant. Transformers, switchgear, uninterruptible power supplies that bridge the seconds before the backup generators start, the generators themselves, and the busways that carry power along the rows of racks.
• The cooling plant. The equipment that moves heat from the chips to the outside air or to water: air handlers in older halls, liquid loops and coolant distribution units in newer ones, and dry coolers, cooling towers or chillers outside.
• The white space. The data halls where the racks stand, with raised floors or overhead trays for cables and pipework.
• The network. Fibre routes into the campus from more than one direction, meet-me rooms where carriers and customers interconnect, and the switching inside the halls.
• Security and operations. Fences, guards, access control and fire suppression, and the staff and software that watch all of it every hour of the year.

How capacity is measured

Datacentres are not measured in square metres but in megawatts of critical IT power: the power available to the servers, storage and network, before the overhead for cooling and electrical losses. A campus is described by the critical IT load it can deliver, and space is leased by the kilowatt per month. When a developer says a building has a given number of megawatts, that is the number to anchor on; floor area follows from how densely the racks are packed.

The same unit lets a lease be set against the economics of the computers inside it. Our piece on converting $/kW-month to $/GPU-hour works through the arithmetic, and it explains why a lease for a powered shell sits so far below a contract for the GPUs in the same megawatt.

Who owns them, and who runs them

Ownership comes in several shapes. Enterprises still run some of their own facilities. Colocation providers own buildings and rent out space, power and cooling to many tenants, either a cage or a few racks at a time, known as retail, or whole halls on long leases, known as wholesale. Hyperscalers build most of their own capacity and lease the rest. Power landlords, many of them former crypto miners, own energised shells and lease them to the operators that install the GPUs. Neoclouds do some of each: they lease halls, build their own, and fill them with GPU clusters that they rent out by the hour or under contract.

Those choices decide who carries which risk. The owner of the building carries the risk of the site, the grid connection and construction. The owner of the computers carries the risk that the chips lose their value before they have paid for themselves. The two are often different companies, bound together by a lease.

What changes when the tenant is an AI cluster

A cloud datacentre is built for flexibility: many tenants, many kinds of workload, modest power per rack and air cooling. An AI datacentre is built for one product at the highest possible output, and nearly every layer changes as a result.

• Density. The newest GPU racks draw many times the power of a cloud rack. Air cannot carry that much heat away, so AI halls are built for liquid cooling from the start. Our piece on power and cooling explains how the power gets in and the heat gets out.
• The network. Thousands of GPUs have to behave like one machine during training, which needs a dedicated fabric of switches and optical links far denser than a cloud hall's. See networking and optics.
• Tenancy. A hall is often leased to a single anchor tenant, sometimes on a contract that runs for years, rather than to many small ones.
• The critical path. The GPUs are rarely the slowest part to arrive. Grid connections, substations and utility works take years in many markets, which is why sites that already have power command a premium, and why capacity moves to wherever electricity can be delivered soonest.

The Primer calls the result an AI factory: a datacentre run as a production line, with electricity in and tokens out.

Why investors watch the building

Three features make the building, and not just the chips, an investment question.

First, time to power is revenue. A hall that is energised months earlier starts earning months earlier, and on the contract prices of the past year that difference outweighs almost any saving in operating cost. Developers with sites that already have power have been able to lease them faster and on better terms.

Second, the scale of spending is large and visible. Microsoft, Alphabet, Amazon and Meta paid

$201.9B

SEC EDGAR 10-Q/10-K (XBRL): purchases of property and equipment, cash-flow statement (calc.)As of 30 Oct 2026Source

(as of 30 Oct 2026) for property and equipment in the latest quarter, from their filings, and much of it goes into datacentres and the equipment inside them. Their suppliers, from transformer makers to cooling specialists, report the same demand in their order books.

Third, the lives do not match. A building and its electrical plant are depreciated over decades; the computers inside are replaced every few years, with each new generation of chips. Whoever owns the building is betting on the site and its tenants, and whoever owns the GPUs is betting on how long they keep earning, the subject of the economics track.

Where to go next

The rest of this track follows the stack upward: what a neocloud is, what an AI factory is, power and cooling, and networking and optics. Our Global AI Datacentre Monitor tracks the AI datacentre sites themselves.

Sources

• AI Infrastructure Primer 2026, Compute Terminal: the chapters "The value chain", "Deployment: how GPUs actually get installed" and "Site selection and energy economics".
• Microsoft, Alphabet, Amazon and Meta: quarterly and annual reports (10-Q and 10-K, SEC EDGAR), purchases of property and equipment; how we compute it.
• Compute Terminal Global AI Datacentre Monitor, our register of AI datacentre sites.