An AI data centre is a facility designed to run artificial intelligence training and inference at scale. It packs far more compute into each rack than a standard data centre, uses liquid cooling to remove the heat that GPUs generate, and draws power loads measured in tens or hundreds of megawatts. The building, the cooling and the grid connection are all sized for one job: keeping racks of accelerators running at full load.
That definition matters because the label gets used loosely. A site advertised as "AI-ready" may still be air-cooled, low-density and stuck in a grid connection queue. The difference between a real AI data centre and a repurposed legacy hall comes down to four measurable things: rack power density, cooling method, available power, and the date that power can actually be delivered. This guide explains each one, then covers cost, water, location and how the market is shifting.
What makes a data centre AI-ready?
AI-readiness is a spec set, not a marketing claim. Four properties decide whether a facility can host AI workloads at scale.
Rack power density. A traditional enterprise rack draws 5 to 10 kW. An AI training rack can draw 50 to 150 kW, because each rack holds dozens of GPUs running at sustained high utilisation. A facility built for 8 kW racks cannot host AI compute without a full rebuild of its power and cooling.
Cooling method. Air cooling stops working efficiently above roughly 30 to 40 kW per rack. AI data centres use liquid cooling instead, in one of three common forms: direct-to-chip cold plates, immersion cooling, or rear-door heat exchangers.
Available power. AI campuses are increasingly sized in hundreds of megawatts. The constraint is rarely the building. It is whether the grid can deliver that load.
Connection date. A site with a signed grid connection two years out is more valuable than a larger site that energises in seven. For AI buyers racing to deploy, the energisation date often decides the deal.
If a listing cannot answer those four questions with verified numbers, it is not yet AI-ready.
How is an AI data centre different from a standard data centre?
A standard data centre supports a broad mix of workloads: web hosting, enterprise applications, storage, and general cloud compute. Power and cooling are sized for moderate, varied demand. An AI data centre supports a narrow, brutal workload profile instead. Training a large model means thousands of GPUs running near full power for weeks, generating heat that air alone cannot move.
Three differences stand out.
The first is density. AI racks concentrate ten to twenty times the power of a conventional rack into the same floor space, so the whole electrical and cooling design changes.
The second is cooling. Liquid replaces or supplements air, which means new plumbing, coolant distribution units, and in some cases a different building shell.
The third is power procurement. A 100 MW AI campus needs its own substation-scale connection and often a multi-year grid agreement. New AI builds are now commonly planned at 100 to 500 MW per campus, with the largest projects targeting a gigawatt or more. Standard colocation sites rarely face that hurdle.
How much power does an AI data centre use?
AI data centres are defined by their appetite for electricity. A single high-density AI rack can use more power than a small office building. Campus-scale AI sites are now planned at 100 MW and beyond, and a wave of announced projects target the gigawatt scale, the equivalent of a large power station serving a single campus.
Power demand is the binding constraint in most markets. In the United Kingdom, the West London and Slough corridor is grid-locked, and connection dates can slip by years. In Australia, grid connection queues and water access shape where new capacity can land at all. A buyer underwriting an AI site is really underwriting a power schedule. The compute hardware can be ordered. The megawatts cannot.
This is why energisation dates and grid connection status sit at the centre of any serious AI capacity search.
Why does AI need liquid cooling?
GPUs run hot. A rack of AI accelerators can produce more heat than air can carry away at reasonable airflow and temperature. Once density passes roughly 30 to 40 kW per rack, air cooling becomes inefficient and eventually impossible without enormous fan energy.
Liquid carries heat far more effectively than air. Three approaches dominate. Direct-to-chip uses cold plates that sit directly on the hottest components, carrying heat away in a sealed loop. Immersion sees servers sit in a bath of non-conductive fluid, which absorbs heat across the whole board. Rear-door heat exchangers use a liquid-cooled door on the back of the rack to cool the air as it exits.
Each method changes the building, the maintenance model, and the water or coolant supply. For a buyer, the cooling method is a hard spec, not a preference.
How much does an AI data centre cost?
AI capacity is more expensive to build per megawatt than standard capacity, because the power, cooling and density requirements are higher. Standard hyperscale capacity runs roughly 10 to 12 million US dollars per megawatt to build, while AI-ready, high-density capacity is commonly put at 20 million or more, before the cost of the compute hardware itself. Public figures vary widely and depend on region, power price, land, and whether you are building, buying, or leasing.
The honest answer is that real numbers are hard to find. Operators guard pricing, and public sources give vague ranges. Cost per megawatt, cap rates and acquisition multiples are the figures buyers actually need to underwrite a deal, and those rarely appear in open listings. Getting verified access to gated pricing is usually the only way to model an AI site accurately.
Does AI use a lot of water?
Many large data centres use water for cooling, either directly in evaporative systems or indirectly through the power they consume. A large evaporatively cooled site can draw several million litres a day, and AI sites, with their higher heat output, can intensify that demand. Water access has become a live planning issue in several markets, including parts of Australia, where the sector could reach a significant share of some cities' supply within a decade.
Cooling method affects water use. Closed-loop and air-assisted designs reduce or remove direct water draw, while open evaporative systems consume more. When assessing an AI site, the water dependency of its cooling design belongs on the checklist alongside power and density.
How is the data centre market changing for AI?
AI has shifted the centre of gravity in data centre development. Demand has moved from steady, moderate-density colocation towards large, power-hungry AI campuses. Operators are racing to secure powered land and grid connections, and the best sites surface in public planning and grid filings long before any sale process begins.
For buyers, this means the search has changed. The question is no longer "where is there a data centre" but "where is there deliverable power, at the right density, on a timeline I can build to." For sellers and developers, a site that can demonstrate verified power, density and a credible connection date is worth far more to AI buyers than a generic hall.
Frequently asked questions
What is an AI data centre in simple terms?
It is a building designed to run artificial intelligence at scale, with racks that draw far more power than normal, liquid cooling to manage the heat, and a grid connection sized for very large electrical loads.
What makes a data centre AI-ready?
Four verified specs: rack power density (often 50 to 150 kW per rack), liquid cooling, large available power, and a confirmed energisation date.
How much power does an AI data centre use?
Campus-scale AI sites are commonly planned at 100 to 500 MW, with the largest projects targeting a gigawatt or more. A single AI rack can use more electricity than a small office building.
Why do AI data centres need liquid cooling?
Because GPUs generate more heat than air can remove efficiently once rack density passes roughly 30 to 40 kW. Liquid carries heat away far more effectively.
Is an AI data centre different from a cloud data centre?
Yes. A general cloud or colocation site supports varied, moderate workloads. An AI data centre is built for sustained, high-density GPU compute, which changes its power, cooling and grid design.
How much does AI data centre capacity cost?
It costs more per megawatt than standard capacity. Standard hyperscale build cost is commonly around 10 to 12 million US dollars per megawatt and AI-ready capacity 20 million or more, though public figures are unreliable. Cost per MW, cap rate and acquisition multiples are the numbers buyers need, and they usually sit behind verification.