The AI boom has led to an increasing number of data centres being built. Dalvinder Kular, assistant editor at National Technology News, explores some of the challenges firms are facing around electricity.

When Amazon announced plans to build a new AI data centre campus in Texas, it was the 7.65GW natural gas power plant being built beside it that attracted much of the attention. The plant has been authorised to emit up to 33 million tonnes of carbon dioxide a year.

The project encapsulates one of the biggest challenges facing the technology industry. For years, the AI race has been defined by who could build the biggest models or buy the most powerful chips. Increasingly, however, access to electricity and the infrastructure needed to deliver it are becoming the real constraints.

In the UK, the problem is particularly pronounced. Ofgem has warned that around 140 proposed data centre projects could require 50GW of electricity, more than the country's current peak demand. The regulator has also raised concerns that the rapid growth in connection applications could delay renewable energy projects and other infrastructure needed to decarbonise the power system.

Elsewhere, the International Energy Agency (IEA) said that global electricity consumption from data centres increased by 17 per cent in 2025, with consumption from AI-focused facilities rising even faster. It expects overall data centre electricity consumption to double by 2030, while power use from AI-focused facilities could triple.

Infrastructure issues

Until recently, the biggest concern surrounding AI was access to advanced semiconductors. As GPU supply has improved, attention has shifted to the infrastructure surrounding those chips.

Matt Hawkins, chief executive of AI infrastructure provider CUDO Compute, argued that organisations are increasingly discovering that buying GPUs is just the start.

“There is a growing assumption that once you secure GPUs, you have solved the problem,” he said. “But the fact is GPUs alone do not create AI capacity.”
The infrastructure surrounding those chips, including substations, cooling systems, grid connections and power availability, is increasingly determining how quickly AI projects can be deployed.

Research commissioned by CUDO Compute found that 42 per cent of UK organisations believe high energy and infrastructure costs are contributing to concerns about an AI bubble, compared with 35 per cent in both the US and Europe.

Around a third of UK organisations cited shortages of affordable power or limited grid capacity as contributing to those concerns. Among AI-first organisations, the concern was even greater, with 51 per cent saying that high energy and infrastructure costs were contributing to fears about an AI bubble.

“The myth is that AI scaling is mainly a chip problem,” Hawkins said. “Increasingly, particularly in the UK, the constraints stretch far beyond silicon into substations, planning, cooling systems, utility access and power availability.”

The consequences could be significant for technology companies and their customers. Richard Price, board member for technology at the Energy Consultants Association (ECA), said the cost of powering AI will ultimately move through the technology supply chain and filter down to customers through additional fees.

“AI runs on energy – —and right now, that energy isn't cheap,” he said. “As data centre demand grows, those costs move through the supply chain.”
For SMEs, Price believes this could create a double financial hit, with businesses paying higher energy bills themselves while also facing increased costs for the SaaS platforms, cloud services and digital tools they rely on.

AI powering AI

The industry's response to rising energy costs is driving investment in technologies designed to reduce electricity consumption, secure alternative supplies and make facilities more flexible.

AI itself is now being used to manage AI infrastructure. Workload orchestration systems can determine when and where computing tasks should run, potentially shifting non-urgent workloads to periods when electricity is cheaper, less carbon-intensive or more readily available.

Sarah Beechey, practice leader for data centre and AI at advisory firm SHI, says this is changing the industry's definition of performance.

“For decades the focus was on achieving more performance, more capacity and more speed,” she said. “But AI is rewriting the equation. Now what might have taken a decade is happening in years, if not months, and the new metric isn't performance but performance per watt.”

She added that businesses are increasingly asking how much electricity their AI workloads consume, how efficiently infrastructure is being used and what the long-term cost of running those systems will be. Efficiency is becoming a commercial advantage, while businesses are also asking whether every workload needs the most powerful available AI model.

For companies deploying AI at scale, those choices can make a difference. Beechey said smaller, domain-specific models can sometimes perform a task with considerably less computing power than a general-purpose frontier model.

Finding and generating power

As demand continues to grow, data centre operators are looking for new sources of electricity. Renewable power purchase agreements are becoming increasingly important, allowing technology companies to secure long-term supplies of wind and solar power.

Battery storage is also proving popular. Large battery installations can provide backup power, help facilities avoid drawing electricity from the grid during periods of peak demand and potentially allow operators to participate in electricity markets.

Some technology companies are looking towards nuclear power because it can provide continuous, low-carbon electricity, matching the relatively constant demand profile of a data centre. Interest in small modular reactors (SMRs) has consequently grown, including in the UK, which has ambitions to develop a domestic SMR industry.

Gas turbines are attractive because they can provide large quantities of power relatively quickly. The Amazon project in Texas made headlines due to its scale. The planned 7.65GW gas facility would initially operate separately from the wider Texas grid and has attracted scrutiny because of its permitted emissions. Google has also floated the idea of a 933MW gas-fired facility in Texas to support an AI data centre campus.

For technology companies, sourcing their own power allows them to build AI infrastructure to a timeline that they, rather than energy grids, control. They are investing billions in renewable energy and promising to reduce their carbon footprints. Yet the urgency of supplying AI infrastructure is encouraging some operators to use natural gas because it can be deployed more quickly.

This risks prolonging reliance on fossil fuels, a concern particularly relevant in the UK, where the government is simultaneously attempting to decarbonise electricity generation.

Cooling concerns

One of the biggest changes is taking place inside the data centre itself. The hardware used for AI such as GPUs processors generate substantially more heat than traditional computing equipment, making cooling an increasingly important part of the energy equation.

Liquid cooling, once a specialist technology, is becoming a much more important part of AI infrastructure. Direct-to-chip systems can remove heat more efficiently than conventional air cooling, allowing higher-density computing while reducing the electricity required by cooling systems.

Amazon is one example of a company attempting to address the issue. The company said its global data centres used 0.12 litres of water per kilowatt-hour in 2025, which it claims is more than seven times more efficient than the industry average. Around 90 per cent of the time, its cooling systems use outside air rather than water, with water-based evaporative cooling reserved primarily for hotter periods.

Using water for cooling during the hottest months can sometimes consume less energy than replacing it with electricity-intensive chillers. Amazon said free-air cooling is used for most of the year, while evaporative cooling during the hottest hours can avoid the additional electricity demand associated with mechanical cooling. The figures are company-reported, however, and comparisons depend on how water use and industry averages are measured.

Data centres as energy assets

There are also attempts to turn data centres from passive consumers of electricity into active participants in the energy system.

Rolf Bienert, managing and technical director of the non-profit OpenADR Alliance, believes data centres could eventually operate as virtual power plants. A facility equipped with renewable generation and batteries could produce and store some of its own electricity, while adjusting its computing workloads according to conditions on the grid.

During periods of high electricity demand, it could draw on stored power or reduce non-essential workloads. At other times, Bienert said it could use surplus renewable electricity to charge its batteries or perform computationally intensive tasks.

“Done well, this doesn't just reduce a single facility's draw on the network; it can actively support the grid in the area around it,” he said. “This comes at a cost for the data centre though – —its own production and storage are not cheap. But the cost hits the actual business and not the unsuspecting consumer.”

Bienert said data centre owners need to communicate with grid owners to determine the impact of a facility on the network. They also need to exchange signals about capacity, price and demand in real time, using open, interoperable protocols rather than closed, proprietary ones.

Heat highways

Data centres produce a lot of heat: almost all the electricity consumed by computing equipment eventually becomes heat that has to be removed. Researchers and infrastructure companies are increasingly looking at whether it can be captured and reused through “heat highways”.

Analysis by EnergiRaven and Viegand Maagøe estimates that projected UK data centre growth could generate enough recoverable waste heat to provide heating for between 3.5 million and 6.3 million homes by 2035, depending on factors including facility design and efficiency. That is an estimate of technical potential, rather than a forecast of homes that will actually be connected to heat networks.

The opportunity is greatest where data centres are located close to housing developments, as in Manchester. Some 5,000 homes are planned in the Victoria North development, and 14,000–20,000 in Adlington. The city is home to more than a dozen existing data centres, with four new facilities planned. Subject to the necessary heat-network infrastructure and commercial arrangements, these facilities could supply heat to some of the planned homes and areas affected by fuel poverty.

The Nordic countries have already experimented with integrating waste heat from data centres, power stations and other industrial facilities into district heating systems. For example, the waste heat generated by Finland’s LUMI supercomputer is used to heat hundreds of households in the neighbouring town Kajaani. The UK is trying to develop similar infrastructure but has struggled to connect sources and consumers.

Peter Maagøe Petersen, director and partner at Viegand Maagøe, said heat should be treated as a national infrastructure opportunity.

“In addition to heating homes, heat highways can also reduce strain on the electricity grid and act as a large thermal battery, allowing renewables to keep operating even when usage is low, and reducing reliance on imported fossil fuels,” he said.

“With denser cities than its Nordic neighbours, and a wealth of waste heat on the horizon, the UK is a fantastic place for heat networks.”

If heat highways are implemented, the location of data centres becomes increasingly important. Moodie believes the North of England could benefit from data centre investment because of its access to renewable generation.

“The future may involve data centres being built where power is available rather than where the existing technology clusters are concentrated,” he said.

Large quantities of wind power are generated in Scotland and northern England, but insufficient transmission capacity can prevent that electricity from reaching demand centres further south. Putting data centres closer to where energy is produced could reduce this pressure.

For years, the AI race has been measured by who could build the biggest models or secure the most powerful chips. Increasingly, however, access to electricity is becoming just as important.

While power consumption per AI task is falling rapidly, the number of AI tasks being performed is rising even faster. As a result, total electricity consumption continues to increase. That is forcing technology companies into difficult trade-offs. The same organisations investing billions in renewable energy, more efficient cooling systems and smarter infrastructure are, in some cases, also turning to natural gas and dedicated power generation to keep pace with demand.

If data centres simply compete for limited grid capacity, they risk putting greater pressure on electricity networks and slowing wider decarbonisation efforts. But if they are built closer to renewable generation, make better use of waste heat and become more flexible in how they consume electricity, they could become part of the solution rather than simply another source of demand.

The next stage of AI will depend on more than advances in computing. As power becomes an increasingly scarce resource, the organisations best placed to benefit may be those that can secure reliable, affordable and lower-carbon electricity as effectively as they can develop the technology itself.

---