AI, energy and data centres: why infrastructure design now matters

A recent story in The Guardian highlighted a growing challenge: how to meet the rising electricity demands of AI data centres while maintaining resilient energy systems and long‑term climate goals.

Data centres already account for a significant and growing share of national electricity use (around 2% in 2024*), but what is less widely discussed is why demand is rising so sharply, and what solutions are being sought to ameliorate the increase.

According to the UK’s National Energy System Operator, grid connected data centres consumed 7.6 terawatt hours (TWh) of electricity in 2024, around 2% of total UK demand.

By 2035, that figure is projected to rise to between 20 and 41 TWh*, driven largely by AI computing workloads.

Part of the issue is that these changes aren’t just a result of scaling of AI use, they are driven by fundamental energy infrastructure issues, both inside and outside the data centre.

Traditional enterprise data centres were designed around rack power densities of 5–15 kilowatts (kW). The first generation of AI computing systems raised rack power demand to 20–40 kW. Today’s AI data centres commonly operate at 50–120 kW per rack.

That represents:

• a 230–900% increase over traditional enterprise racks, and

• a 25–500% increase over early GPU based systems

Industry roadmaps suggest the next generation of AI infrastructure could reach up to 600 kW* per rack — a further 400–1,100% increase on today’s AI deployments.

At that point, single AI data centres could be using the same energy as a steelworks, or aluminium smelter.**

Much of today’s data centre infrastructure is based on legacy AC power architectures, originally designed for far lower power densities.

Electricity typically passes through multiple stages of AC to DC and DC to DC conversion before reaching processors. Each step introduces small efficiency losses, and when working at scale these losses accumulate, and are ultimately released as heat.

The impacts are measurable:

• Cooling systems can consume 7–30% of total facility power, depending on efficiency.

• Hyperscale AI data centres can draw 10–50 times more electricity per square metre than typical commercial buildings.

• Grid connections of hundreds of megawatts per site are becoming more common, intensifying competition for capacity.

In this context, rising energy demand is not just a function of “more AI” with facilities working at peak 24/7. It is also a result of how power is delivered inside data centres.

Compound Semiconductor Applications Catapult’s latest report, The AI Data Centre Power Crisis: Why 1500 V DC Architectures Matter, highlights how alternative power delivery approaches could reduce the cumulative impact of AI infrastructure.

One option gaining attention is high voltage DC (HVDC) power distribution, particularly 1500 volt DC architectures using Solid State Transformers (SSTs). By reducing the number of conversion steps and increasing distribution voltage, these systems can:

• cut power conversion losses,

• reduce heat generation,

• lower cooling demand, and

• support higher density compute more efficiently

• with smaller footprint

• and easy integration into renewable power sources.

There is a layer of nuance to these technologies, whilst we have direct data around current usage, to suggest a single fixed energy saving figure from high voltage DC architectures would be a fool’s errand, there are too many contributing factors and interdependencies, some of which remain unknown.

Instead, the clear evidence shows that reducing power conversion stages and operating at higher distribution voltage cuts electrical losses and heat generation. We know that in AI data centres operating at tens or hundreds of megawatts, even single percentage efficiency improvements can translate into substantial reductions in electricity use, cooling demand and long-term operating cost. This, in turn, helps retain the return on investment when moving to the next generation of gigawatt datacentres. Every percentage matters.

At the scale of modern AI facilities, even single percentage efficiency improvements can translate into meaningful reductions in total electricity use, operating cost and pressure on grid infrastructure.

CSA Catapult’s latest report does not present this as a quick fix or a substitute for wider energy planning. Rather, it points to the importance of early engineering choices that will shape the energy footprint of AI systems for decades.

Much of the public debate currently focuses on a single question: how much power will AI use?

The evidence suggests a broader question is now equally important: how intelligently, efficiently and robustly can that power be delivered?

As AI continues to scale, system level design which spans power electronics, thermal management and infrastructure integration will play a decisive role in determining whether future data centres become an escalating burden on energy systems, or a more manageable and efficient part of the landscape.

*Read the CSA Catapult report: The AI Data Centre Power Crisis: Why 1500 V DC Architectures Matter