AI GPU Infrastructure

What AI GPU infrastructure is

AI GPU infrastructure is the physical system that makes artificial intelligence possible: the GPUs that perform the calculations, the high-speed networks that connect them into clusters, the storage that feeds them data, and the data centre, power and cooling systems that keep the whole thing running.

It is easy to think of AI as software. In practice, an AI model is a very large mathematical object that has to be computed somewhere, by hardware that exists in a building, drawing electricity from a grid. Many of the major advances in AI over the past decade have been enabled by increases in available compute, alongside advances in models, data and software — and making compute available is a construction and energy problem as much as a software one.

As AI deployment scales, one of the increasingly important constraints is serviceable compute capacity: powered, cooled and connected infrastructure where GPU systems can operate effectively. Chips, models, data and skilled people all remain scarce in their own right. But capacity is the layer that is hardest to add quickly, because it depends on physical assets and connections that take years rather than months to deliver. It is one of the AI infrastructure focus areas Nuway works across.

Why AI runs on GPUs

A CPU is designed to do a small number of complicated things quickly, one after another. A GPU is designed to do an enormous number of simple things at the same time.

AI workloads are overwhelmingly the second kind of problem. Training and running a neural network is, at bottom, vast quantities of matrix multiplication — the same simple operation repeated across billions of numbers. That work divides neatly across thousands of parallel processing cores, which is exactly what a GPU provides. It is why hardware originally built for rendering graphics turned out to suit machine learning, and why GPU supply became a strategic concern for governments rather than a consumer-electronics story.

Two further properties matter as much as raw core count: memory bandwidth, because a processing core waiting for data is doing nothing, and interconnect speed, because large AI workloads often need to be distributed across multiple GPUs, making interconnect performance increasingly important.

Training and inference are different problems

Training is building the model. It is a single large job that can run for weeks across many GPUs simultaneously, all of which must stay tightly synchronised. It is bursty, power-intensive, and highly sensitive to interconnect performance — one slow link slows the whole cluster.

Inference is using the trained model to answer a query. Inference workloads can range from relatively small requests to demanding multimodal and reasoning workloads, but they typically need to serve requests responsively and at scale. Inference is more sensitive to latency and runs continuously rather than in campaigns.

This distinction shapes real infrastructure decisions. Training is generally less location-dependent than inference, so it can more readily be sited where power and land are available — though data location, network capacity, regulation, resilience and operational support all still influence where training capacity can practically be built. Inference tends to benefit from proximity to the users or the data it serves. As AI moves from being built to being used, demand shifts from the first pattern toward the second, which changes where facilities are needed, not only how many.

How a GPU cluster is built

From chip to rack

A modern AI system is assembled in layers. Individual GPUs are mounted on boards, boards go into compute trays, trays are installed in a rack, and racks are grouped into clusters.

NVIDIA's GB200 NVL72 illustrates how far this has moved from conventional IT. A single rack holds 72 Blackwell GPUs and 36 Grace CPUs across 18 compute trays, connected by nine NVLink switch trays through a passive copper backplane, and it is liquid-cooled by design rather than by retrofit. NVIDIA's rack documentation sets out that configuration, and NVIDIA describes the result as behaving like one very large GPU rather than 72 separate ones, with 130 terabytes per second of communication between them inside the rack.

The point is not the specification. It is that the rack has become the unit of AI compute. Historically you bought servers and installed them in whatever rack you had. Now the rack, its power delivery, its cooling and its internal network are designed together as one system — and a building either accommodates that system or needs to be adapted to do so.

The network is part of the computer

Where a cluster spans multiple racks, the network between them stops being plumbing and becomes part of the machine. During training, GPUs must repeatedly exchange results and stay in step; if that exchange is slow, expensive processors sit idle waiting. This is why AI clusters use dedicated high-bandwidth fabrics — InfiniBand or AI-optimised Ethernet — rather than standard enterprise networking, and why interconnect design is one of the harder parts of building a cluster well.

Storage and data movement

Training data has to reach the GPUs fast enough to keep them busy, which requires high-throughput parallel storage and careful data pipeline design. Storage is a comparatively small share of a data centre's electricity — the IEA puts it at around 5%, against roughly 60% for servers — but it is a common place for an otherwise well-specified cluster to underperform.

What makes a data centre AI-ready

Density, cooling and power delivery are what separate a conventional facility from an AI-capable one. These are also the AI-ready data centres that AI workloads increasingly depend on.

Power density

Density is the clearest dividing line between a conventional data centre and an AI-capable one. AI racks draw substantially more power than the racks most existing halls were designed around — a difference of scale rather than degree.

That matters because density is not a property of the equipment alone. A hall designed around conventional rack loads has power distribution, busbars, floor loading and cooling all sized for that assumption. Raising density means revisiting each of them.

Cooling

Air cooling has a practical ceiling. Past a certain heat load per rack, moving enough air becomes impractical and the heat has to be carried away by liquid instead. Direct-to-chip liquid cooling circulates coolant through cold plates sitting directly on the processors; NVIDIA's rack-scale systems are built around it as standard, with liquid cooling manifolds integral to the rack.

Retrofitting high-density AI infrastructure into an existing facility can require significant changes to power distribution, cooling and other building systems, making feasibility highly site-specific.

What existing data centres need to support AI infrastructure

This is the crux, and it explains why AI capacity can be scarce even where data centres are plentiful. The UK government's assessment is direct: the UK has a significant and mature data centre market, but it is not yet optimised for AI, because most existing facilities are geared toward general-purpose enterprise computing and lack the density, energy integration and technical design needed for high-intensity AI workloads.

Many facilities built for conventional workloads can support AI infrastructure, but doing so may require substantial upgrades — to power distribution and capacity, to cooling, to networking, and sometimes to the structure of the building itself. What is scarce is therefore not floor space, but floor space that is already AI-ready or can be economically brought to that standard.

AI factories

"AI factory" is the term that has emerged for a facility purpose-built to produce AI output — a site designed from the outset around dense GPU clusters, liquid cooling and large, reliable power, rather than a general-purpose data centre with AI equipment installed in it.

The framing is deliberate. A conventional data centre stores and serves data. An AI factory takes in data and electricity and produces trained models and inference at scale. Like any production facility, its performance and efficiency depend on compute utilisation, power availability, cooling capacity, and the ability to keep the infrastructure operating reliably at high utilisation.

The main constraints: power, grid and land

Electricity demand

Data centres accounted for around 1.5% of global electricity consumption in 2024, or 415 terawatt-hours, and the IEA projects that roughly doubling to around 945 TWh by 2030 — slightly more than Japan's entire electricity consumption today. Consumption has been growing around 12% a year since 2017, more than four times faster than electricity demand overall.

Scale at the individual site level is equally striking: the IEA notes that a typical AI-focused data centre uses about as much electricity as 100,000 households, and that the largest under construction will use twenty times that.

Grid connection

Power existing somewhere is not the same as power being deliverable to a specific site on a specific date, and this is where AI infrastructure projects most often encounter delay.

The IEA estimates that around 20% of planned data centre projects globally are at risk of delay because of grid strain, and that building new transmission lines takes four to eight years in advanced economies, with lead times for critical components such as transformers and cables having doubled in three years.

Great Britain illustrates the problem clearly. The connections queue had grown to over 700 GW — around four times what the country is projected to need by 2030 — with shovel-ready projects waiting up to ten years. Following reform, NESO confirmed a re-ordered pipeline in December 2025 prioritising deliverable projects, covering 283 GW of generation and storage and 99 GW of transmission-connected demand, with data centres explicitly among the future needs the new process is designed to serve.

For AI infrastructure projects, grid connection status can materially affect project deliverability and timing.

Siting and planning

Site selection for AI infrastructure resolves to a short list of practical questions: how much power can be delivered here, and when; is there a viable route for heat rejection; is there fibre; will it obtain planning consent; and can the grid connection, the building and the equipment all be ready in the same window. Sites satisfying all of these simultaneously are uncommon, which is why site identification and development has become a discipline in its own right.

AI infrastructure in the UK and Europe

The UK has set out a specific and quantified position. The UK Compute Roadmap forecasts that the UK will need at least 6 GW of AI-capable data centre capacity by 2030 — a threefold increase on capacity available today — and aims to establish nationally significant sites each capable of serving at least 500 MW, with at least one AI Growth Zone scaling beyond 1 GW by 2030.

On the public research side, the same roadmap commits over £1 billion to expand the AI Research Resource twenty-fold, from 21 AI exaFLOPS in 2025 to 420 AI exaFLOPS by 2030, within a compute package of up to £2 billion. AI Growth Zones are the delivery mechanism on the private side: designated locations with streamlined planning and prioritised grid access, with bidding opened to local and regional authorities in February 2025 and particular focus on deindustrialised areas with existing power access, set out further in Delivering AI Growth Zones.

The European picture points in a similar direction. The European Commission's AI Factories initiative is establishing AI-optimised supercomputing facilities across member states, and the UK Compute Roadmap notes that the US, UAE and several European nations are each building dedicated AI campuses that will exceed 1 GW.

There is now clear public-policy support for expanding AI compute capacity alongside private-sector demand, creating a more defined strategic framework for AI infrastructure development in the UK and Europe.

How Nuway Capital works in AI compute infrastructure

Nuway Capital originates and develops commercial opportunities across AI compute infrastructure, GPU compute and data centres, working alongside infrastructure partners and specialists, including the NuSphere Alliance.

In practice that means working on the layers described above — the compute, the facilities, and the power and connection questions that determine whether a project is deliverable — with partners who bring the specialist capability in each. Nuway does not manufacture GPUs, operate cloud GPU services or operate energy infrastructure.

More on how Nuway Capital works and the team behind it.

Common questions about AI GPU infrastructure

If you would like to discuss AI compute infrastructure, get in touch.