Can Britain power an AI boom without running dry?

England has spent the summer in drought just as the government races to build data centres for AI. The concern is real, many of the headlines are wrong, and the fixes already exist. Banning data centres would solve neither the water problem nor the energy one.

This August, almost three-quarters of England was officially in drought after the driest July in 190 years, and the Prime Minister chaired an emergency COBR meeting on the response. As reservoirs fell, the water industry raised the alarm about the strain new data centres could add to supplies.

At the same time, the government wants Britain to be an AI power, and AI runs on data centres: big, hot buildings that use a great deal of electricity and, sometimes, a great deal of water. So the question is fair. Can we have both?

Our answer, as engineers, is yes. The concern is legitimate, and the problem is solvable, because a data centre is a building: a very hot, very hungry one, but a building all the same. Our industry has been here before.

The numbers, without the panic…

Globally, data centres use a small slice of the world’s electricity. The International Energy Agency puts demand at around 485 TWh last year, roughly doubling to 950 TWh by 2030: about 3% of global demand. Locally, it’s a different story. In Ireland, data centres used 23% of all metered electricity last year.

Water is similar. The famous “bottle of water per chat” estimate counted water used at power stations and came from an older model, at roughly half a litre per 10 to 50 responses. Google’s own on-site figure is about five drops per prompt. A single question is trivial. Billions of them, served from a few giant campuses in hot, dry places, are not.

In England, data centres use about 0.2% of non-household mains water. The catch is timing. Environment Agency data from around 200 sites shows a clear summer peak, as hybrid systems switch to water cooling in hot weather: exactly when rivers and homes need it most.

Why everyone reaches for water…

Evaporative cooling won because it is the cheapest way to move heat. Evaporating a single litre of water soaks up roughly 0.7 kWh of heat, the same reason sweating cools you down, while a chiller has to do that job with a compressor running on electricity. With cooling historically accounting for up to 40% of a data centre’s electricity, and grid connections in short supply, every kilowatt saved on cooling can go to servers instead.

The industry’s own scorecard rewarded it too. The headline efficiency metric, PUE (power usage effectiveness), compares a site’s total energy with the energy that reaches its servers. Water doesn’t appear in it at all.

Going waterless isn’t free: fans and chillers use more electricity, and that electricity carries its own water footprint at the power station. But wind and solar use almost no water, so as they replace thermal power stations, the balance tips further towards dry cooling every year.

The fixes are already here…

The biggest shift is happening inside the server. AI chips now run so hot that air can’t cool them, and the liquid cooling that replaces it barely needs to evaporate anything.

Warm liquid cooling is the breakthrough. It sounds backwards, but it’s simple physics: heat only flows from something hotter to something cooler. Older data centres needed air or water colder than a summer’s day, so in warm weather they had to evaporate water or run power-hungry chillers to get there. NVIDIA’s newest servers can be cooled with a water-glycol mix at up to 45°C, hotter than a hot tub. Outdoor air is almost always cooler than that, so simple heat rejection plant (in effect, large radiators outside the building) can carry the heat away for most of the year, while the same coolant circulates round a sealed loop. In the right climate, the water used for cooling falls to almost nothing. Better still, heat at that temperature is warm enough to be recovered and put to use, warming nearby buildings instead of the sky.

The big operators are switching. Microsoft says its new Fairwater AI campus in Wisconsin uses closed-loop cooling and consumes about as much water a year as a neighbourhood restaurant. That’s the company’s claim, not an independent audit, but the direction is clear.

Nobody should be cooling servers with drinking water. Research for the UK water industry this year called for fewer barriers to non-potable alternatives, such as treated wastewater.

One warning. Some of the most efficient immersion cooling has relied on PFAS, the so-called forever chemicals, and the dominant supplier has phased out production. Swapping a water problem for a pollution problem isn’t progress.

For the engineers: coolant entering the chip at 45°C leaves at around 55°C. UK air almost never approaches 45°C, so dry coolers can run without chillers in nearly every hour of the year, and the 55°C return is hot enough, through a heat exchanger, to feed a low-temperature heat network or be lifted efficiently by a heat pump. The coolant isn’t harmless: glycol and its inhibitors need proper leak detection and licensed disposal, but a sealed loop avoids the constant chemical dosing, blowdown and Legionella management that cooling towers demand.

From waste heat to warm homes…

Almost every kilowatt-hour a data centre uses ends up as heat. Think of a 100 MW campus as a 100 MW heater that never switches off.

Europe has shown what to do with it. In Odense, waste heat from Meta’s data centre warms about 11,000 households. Near Helsinki, two new Microsoft data centres could eventually supply about 40% of the area’s heat demand.

The UK is catching up fast. In west London, construction of the Old Oak and Park Royal heat network, built around local data centres, is due to start this month. In September the government announced £90 million for heat networks, including a scheme to carry data centre heat to around 650,000 London homes. And a study for the Mayor found that over 300,000 London homes could be heated from existing data centres alone.

Policy is following. London’s draft Plan would require new data centres to show heat recovery informed by a waste heat utilisation plan, and under heat network zoning, data centres can be required to connect to a heat network.

The catch is time and place. A data centre makes heat all year, homes want it mostly in winter, and heat can’t travel far without losses. The data centre, the network and the homes have to be planned together. That’s a building services problem, and one our industry knows how to solve.

What the government is planning…

So far, the government’s data centre plans are about speed: power, planning and price.

Since January, the largest data centres can seek consent as nationally significant infrastructure, the route used for power stations. AI Growth Zones in Oxfordshire, the North East, North and South Wales and Scotland promise faster planning and grid connections, and from April 2027 data centres in those zones are due to get cheaper electricity.

The grid is the real bottleneck. In March, ministers proposed priority in the connection queue for strategically important data centres, prompting housebuilders to warn that new homes could lose out. Ofgem says 315 data centres are waiting for a grid connection, seeking around 73 GW between them, and has proposed a refundable fee to weed out speculative bids.

Since taking office in July, Prime Minister Andy Burnham has put AI at the centre of government, with a cabinet-level AI minister and a new AI Taskforce. One of his government’s first concrete steps touching data centres has been on heat: the £90 million for heat networks described above.

What’s missing is measurement. Data centres still have no duty to report their energy or water use, and the national policy statement for data centres, which would set the ground rules for the biggest schemes, had yet to go out to consultation by the summer. MPs on the Environmental Audit Committee have opened an inquiry, and the Energy Secretary has told them future demand is “inherently uncertain”.

Speed matters: grid queues are holding up homes as well as servers. But speed without measurement risks repeating a mistake our industry has spent years correcting in buildings: approve on the model, then discover the performance gap.

Elsewhere, the rules are catching up…

Other countries are further ahead on measurement, though not all in one direction.

European Union. In September the Commission adopted a rating scheme giving large data centres A to G grades for energy and water efficiency, built from data they have actually reported.

Ireland. New data centres can connect to the grid only on conditions: larger sites must bring their own generation or storage, and all but the smallest must match 80% of their demand with new Irish renewables.

Germany. Rules can loosen as well as tighten. In June, Germany’s cabinet approved an amendment that eases several efficiency limits and replaces the duty to use waste heat with a cost-benefit analysis.

Further afield. Singapore allocates new capacity to the greenest bids. China requires new large sites to reach a PUE of 1.25 or lower. The US has no national rules, so states and cities are passing their own pauses.

The direction of travel is clear: away from design-stage promises, towards measured performance. Measured, not modelled is the same shift the draft London Plan and the UK Net Zero Carbon Buildings Standard are bringing to the rest of the built environment.

The case for a ban, and why it falls short…

Calls to halt data centres are getting louder. In the US, a federal moratorium bill has been introduced, and local opposition groups doubled to 833 in early 2026. In Scotland, campaigners have renewed calls for a moratorium, and this month the Scottish Parliament opened an inquiry into hyperscale data centres.

The critics are right about more than the industry likes to admit. Water companies can’t yet say how much extra demand data centres will add to their plans, the UK has no reporting duty, and the IEA warns that data centres could push up electricity prices in some places if growth isn’t managed.

But read what they actually ask for. The US bill would pause new data centres until Congress passes AI legislation; the Scottish campaign wants a pause until the government defines a green data centre. The demand is for rules, and on that, engineers agree. Where we part company is the route.

A ban doesn’t make demand disappear; it moves it somewhere with weaker rules and a dirtier grid. It freezes the oldest, thirstiest buildings in place. And it gives away the leverage that let Ireland and Singapore win clean power and efficiency commitments from developers queuing to build.

Our own industry proves the point. Buildings are still the biggest gap in the UK’s path to its 2030 climate target. Nobody’s answer is to stop building; it’s Part L, Passivhaus, LETI and verified, in-use performance. The answer to data centres isn’t “stop”. It’s “prove it”.

Five questions to ask of any data centre…

Whether you sit on a planning committee, live near a proposed site or are designing the building next door:

1. What will its measured energy and water use be, and will they be published every year?
2. Where does the cooling water come from? It shouldn’t be drinking water.
3. What happens on the hottest day of the year? That’s when water demand peaks.
4. Where does the heat go? Ask for the waste heat plan.
5. Where does the electricity come from? Look for new clean generation, not just certificates.

Why this matters to us…

Data centres are buildings, and buildings are what we do. The engineering that makes a data centre a good neighbour is the engineering we practise every day: fabric-first design, low-carbon heat and heat networks, and performance proven at the meter rather than promised in a model. The principles that make a school or a home work for the people inside it will make digital infrastructure work for the community around it.

Britain is going to build a lot of data centres. The job now is to make sure they’re good neighbours.

Planning a data centre, or a scheme that could put a neighbour’s waste heat to work? Bringing this thinking to data centres is a natural next step for us, and the earlier we’re involved, the cheaper the big decisions are to change. We’d love to hear from you.