What Is an AI Data Centre? The Definitive Explainer

An AI data centre is a facility purpose-built to host the GPU and accelerator hardware used to train and run artificial-intelligence models. Where a conventional data centre is optimised to serve web traffic, store files and run business applications, an AI data centre is engineered around one constraint: cramming as much parallel compute as possible into each rack, and removing the enormous heat that compute produces. The defining characteristic is power density. A traditional rack draws roughly 5–10 kW; an AI rack packed with GPUs draws 40–120 kW — and the next generation of accelerator clusters pushes higher still. Everything else that makes an AI data centre distinct — liquid cooling, reinforced electrical infrastructure, grid-constrained site selection — follows from that single number.

This page is a neutral, technical explainer of what AI data centres are, how they differ from the facilities that came before them, where the UK's AI compute actually sits, and why their energy appetite has made on-site generation a board-level question rather than a sustainability footnote.

The short definition

In one sentence: an AI data centre is a high-density compute facility built to run GPU and accelerator workloads — model training and inference — that demand far more power and cooling per square metre than the general-purpose IT a traditional data centre was designed for. The hardware inside is dominated by graphics processing units (GPUs) and specialised AI accelerators rather than conventional CPUs, because the mathematics of neural networks (large matrix multiplications, run massively in parallel) maps onto GPU architecture far more efficiently than onto a standard server processor.

How an AI data centre differs from a traditional data centre

The two facility types share a postcode and a perimeter fence, but very little else once you look inside. The differences cascade from power density outward into cooling, electrical design, building structure and even where the facility can be built at all.

Characteristic	Traditional data centre	AI data centre
Dominant hardware	CPU servers, storage arrays	GPU / accelerator clusters
Rack power density	5–10 kW	40–120 kW (and rising)
Primary workload	Web, storage, enterprise apps, virtualisation	Model training and inference
Cooling method	Air cooling (CRAC/CRAH, hot/cold aisle)	Direct-to-chip or immersion liquid cooling
Electrical infrastructure	Standard PDU and busbar provision	Reinforced distribution, high-amperage busways
Site selection driver	Connectivity, latency, market proximity	Grid capacity, available megawatts, water

Power and grid demand

The headline difference is appetite. A single AI training hall can demand tens of megawatts continuously — comparable to a small town. Because AI accelerators run flat-out for long stretches, the load is not only high but sustained, with little of the diurnal variation a mixed-workload facility shows. This is why the binding constraint on AI data centre development in the UK is no longer land or fibre but available grid connection capacity. Operators routinely face multi-year queues for new high-voltage connections, and grid availability now dictates where the next generation of facilities can physically be built.

Cooling

You cannot air-cool a 100 kW rack effectively. Beyond roughly 30–40 kW per rack, conventional hot-aisle/cold-aisle airflow runs out of headroom, and AI facilities move to liquid cooling — either direct-to-chip cold plates that carry coolant straight to the GPU die, or full immersion where servers sit in a dielectric fluid. Liquid carries heat away far more efficiently than air, but it changes the building: coolant distribution units, manifolds, leak detection and, in some designs, significant water consumption for heat rejection. Cooling is no longer a facilities afterthought; it is a core design discipline.

Building and electrical structure

High density concentrates weight and heat. Floors must carry heavier racks, electrical rooms must house far more transformer and switchgear capacity per square metre, and the ratio of "white space" (the IT floor) to plant has shifted. An AI data centre is, in effect, a power-and-cooling plant with some computers in it — the reverse of the perception most people hold.

Training versus inference — the two AI workloads

Not all AI compute is the same, and the distinction matters for power planning. Training is the one-off (but enormous) process of building a model: feeding it vast datasets and adjusting billions of parameters over weeks of continuous computation across thousands of GPUs. Training is power-hungry, bursty at the cluster level, and increasingly clustered in a small number of very large facilities because it benefits from being co-located.

Inference is the day-to-day running of a finished model — answering a query, generating an image, classifying a document. Each inference event is smaller than a training run, but inference happens billions of times a day and is far more sensitive to latency, which pushes it geographically closer to users. As AI adoption matures, inference is expected to dominate total AI energy consumption — the model is trained once but queried endlessly. This split is one reason edge and distributed AI infrastructure is growing alongside the headline mega-campuses; you can read how facility purpose shapes design across our HPC and AI data centre vertical.

Where is the UK's AI compute?

The question "where are the AI servers located" has a clear answer in the UK, and it follows two forces: where the grid has capacity, and where research and commercial demand concentrate. Roughly 75–80% of the UK's data centre estate sits in London and the Thames Valley — the "L" in Europe's FLAP-D market (Frankfurt, London, Amsterdam, Paris, Dublin) — and AI-specific compute is layering onto that existing geography while spilling into new grid-rich locations.

• Cambridge and the East of England: the UK's research-AI heartland. The University of Cambridge runs significant HPC capability, Arm Holdings is headquartered here, and Kao Data's Harlow campus (around 20 miles south) is one of the country's leading purpose-built AI-compute sites, hosting high-density GPU deployments.
• Slough and the Thames Valley: Europe's densest data centre cluster — "Data Centre Alley" along the M4 — home to Equinix, VIRTUS, Digital Realty, Global Switch and Ark. The existing colocation density and SSEN grid make it a natural host for AI capacity where connections allow.
• London Docklands and West London: the carrier-neutral interconnection core (Telehouse, Equinix, LINX) in E14/E16, plus heavy hyperscale build-out around Hayes and Park Royal.
• Regional growth hubs: Manchester, Cardiff/Newport (Next Generation Data's vast Imperial Park facility), Leeds, Birmingham, Bristol and Edinburgh all carry growing compute footprints, often where grid headroom is more readily available than in the saturated South East.

Layered on top of private investment is UK Government policy. The Government has designated AI Growth Zones — areas earmarked for accelerated planning and grid connection to attract AI infrastructure — and funds the AI Research Resource, a national programme of publicly-backed supercomputing capacity for science and public research. Both are general policy frameworks rather than fixed addresses, and the specific sites continue to evolve, but together they signal that AI compute siting is now treated as national infrastructure. Our UK data centre locations map sets out each major cluster and its host DNO in detail, and our Cambridge data centre solar page covers the research-compute corridor specifically.

The energy problem AI data centres create

AI's compute density translates directly into an electricity problem. UK data centres already consume on the order of 12 TWh per year, growing 8–12% annually, and AI training and inference could add a further 4–6 TWh by 2030. That growth collides with three pressures at once: a constrained grid that cannot connect new load quickly; volatile wholesale electricity prices that make industrial-and-commercial tariffs of 18–32p/kWh routine; and tightening corporate climate commitments.

The largest operators have made those commitments public. Microsoft targets 100% carbon-free energy by 2030, Google has committed to 24/7 carbon-free energy by 2030, and Amazon, having met its 100% renewable-matching goal, is working toward round-the-clock matching. The bar has moved from annual renewable matching to hourly carbon-free energy (24/7 CFE) — proving that every hour of consumption is met by carbon-free generation, not merely offset on an annual average. For a facility running a flat 24/7 IT baseload, that is a genuinely hard standard to meet.

Why on-site solar and PPAs matter for AI compute

On-site rooftop solar PV will never power an AI data centre on its own — roof area constrains generation to roughly 5–15% of annual load on a typical site. But that fraction is unusually valuable, because of how a data centre consumes power. With a flat, round-the-clock IT baseload, virtually every kilowatt-hour the panels produce is consumed on site in the same instant it is generated. That means close to 100% self-consumption, no export losses, and no battery-cycling penalty — which makes data centre solar the lowest-LCOE rooftop megawatt-hours in the country, at roughly 3–5p/kWh against 18–32p/kWh grid retail.

For the operator that arithmetic does three things at once: it shaves the highest-priced peak-daytime grid units, it generates REGOs (Ofgem's annual Renewable Energy Guarantees of Origin) that underpin a market-based Scope 2 zero, and — when paired with EnergyTag granular certificates — it contributes hourly-matched carbon-free energy toward a 24/7 CFE claim. Where rooftop alone cannot close the gap, off-site corporate Power Purchase Agreements (PPAs) bring additional matched renewable supply. On-site solar and PPAs are complementary tools in the same decarbonisation stack; our PUE and sustainability guide sets out exactly how each interacts with Scope 2 accounting and CFE reporting.

From definition to deployment

Understanding what an AI data centre is — a high-density, power-hungry, liquid-cooled GPU facility under intense scrutiny over its energy footprint — leads naturally to the practical question every operator now faces: how to generate clean power on the asset itself. We are a UK specialist in exactly that. Whether you operate a hyperscale campus, a colocation hall or a research-compute site, our AI data centre solar page covers feasibility, zero-export G99 configuration, capital allowances and integration with high-density facilities. The definition is the easy part; engineering the generation to match a 24/7 GPU load is where specialism earns its keep.

How much power does an AI data centre use?

There is no single answer, because AI data centres range from a single high-density hall inside an existing colocation building to a purpose-built campus drawing more electricity than a small city. The useful way to think about consumption is at three scales: the rack, the facility, and the national grid.

At the rack level, the defining figure is density. A conventional IT rack draws 5–10 kW; an AI rack packed with GPUs draws 40–120 kW, and next-generation accelerator clusters push beyond that. Because AI accelerators run flat-out for long training and inference cycles, that load is sustained around the clock rather than peaking during business hours — a flat 24/7 baseload with little diurnal variation.

At the facility level, a single large AI training hall can demand tens of megawatts continuously. A hyperscale AI campus can be specified at 100 MW or more of IT load — which, allowing for cooling and electrical overhead, translates into a grid connection comparable to a substantial town. This is why the binding constraint on new UK AI development is no longer land or fibre but available grid connection capacity, with operators facing multi-year queues for new high-voltage connections.

At the national level, UK data centres already consume on the order of 12 TWh per year, growing 8–12% annually. AI training and inference are forecast to add a further 4–6 TWh per year by 2030. The table below frames typical continuous power draw by facility scale — figures are indicative ranges, not fixed specifications.

Facility scale	Typical IT load	Indicative grid demand	Rough equivalent
Single AI hall (in existing DC)	1–10 MW	1.5–14 MW incl. cooling	A few thousand homes
Dedicated AI / HPC site	10–40 MW	14–56 MW incl. cooling	A small town
Hyperscale AI campus	100 MW+	140 MW+ incl. cooling	A small city

One nuance shapes the long-term picture: training versus inference. Training a large model is an enormous one-off burst — thousands of GPUs running for weeks — but a finished model is then queried billions of times a day. As AI adoption matures, inference is expected to dominate total AI energy consumption, which spreads demand across more, smaller deployments closer to users. For operators, the practical consequence of all three scales is the same: a large, continuous, hard-to-decarbonise electricity bill that makes on-site generation a board-level question. Our data centre solar cost guide and battery storage page set out how on-site generation offsets that load.

Why power consumption makes on-site generation a board issue

The scale of AI power draw collides with three pressures at once: a constrained grid that cannot connect new load quickly; volatile wholesale prices that keep commercial tariffs at 18–32p/kWh; and tightening corporate climate commitments that have shifted the bar from annual renewable matching to hourly 24/7 carbon-free energy. A facility running a flat round-the-clock GPU baseload is one of the hardest loads in the economy to decarbonise on an hourly basis.

On-site rooftop solar will not power an AI data centre alone — roof area limits generation to roughly 5–15% of annual load. But that fraction is unusually valuable, because a flat 24/7 baseload means virtually every kilowatt-hour generated is consumed on site the instant it is produced: close to 100% self-consumption, no export losses, no battery-cycling penalty. That makes data centre rooftop the lowest-LCOE solar in the country at around 3–5p/kWh against 18–32p/kWh grid retail. It also generates REGOs for market-based Scope 2 accounting and, paired with EnergyTag granular certificates, contributes toward a 24/7 CFE claim. Where rooftop cannot close the gap, off-site corporate PPAs add matched supply. As a supplier-neutral, data-centre-dedicated installer, we model exactly how much of your measured load on-site PV can offset — explore our AI data centre solar service or request a free desk feasibility study.