AI Heat Needs a Neighbor

The detail that makes the BBC's report on Exmouth Leisure Centre stick is its size. A washing-machine-sized data centre, computers surrounded by oil, and a heat exchanger are being used to warm a public pool to about 30C for roughly 60% of the time. Deep Green provides the unit to the council-run centre for free, sells the compute to clients, and told the BBC it would refund the electricity cost of running the unit.

My read is that this is a better sustainability story than most AI-infrastructure press releases because the physics and the buyer are both visible. The pool needs heat. The servers make heat. The hard part is not the slogan. It is whether enough real compute demand can be placed where someone nearby can use the heat.

Answer Snapshot

Question	My read
What happened?	BBC News reported that Deep Green's small data centre at Exmouth Leisure Centre uses mineral-oil cooling and a heat exchanger to help warm the pool.
Why it matters	It turns data-centre waste heat into a local heating input instead of treating heat only as a cooling problem.
Who benefits if it works?	Leisure centres and other sites with steady heat demand, plus compute customers willing to use a distributed high-performance computing setup.
The catch	The heat is free to the pool only if the compute side has paying workloads, and the pool still needs backup heat when the servers are not enough.
My thesis	AI heat becomes useful infrastructure only when compute demand and heat demand are designed as neighbors.

The Clever Part Is Plumbing

The basic engineering is refreshingly plain. BBC describes computers in a white box surrounded by oil; the hot oil is pumped through a heat exchanger to warm pool water. CIBSE Journal's case study adds useful detail: the Exmouth installation used immersion cooling, the whole computer is cooled in oil instead of by blowing cold air over components, and pipes are lagged so more heat is reused.

That matters because it keeps the idea from sounding magical. The data centre is still consuming electricity. Nearly all of that energy becomes heat. The difference is that the heat is moved into a water-heating job the leisure centre already had, instead of being rejected into the air by a conventional cooling setup.

CIBSE reported expected savings of £22,000 in heating bills over the next year, while Ars Technica's writeup put the expected saving at about £20,000 and said Deep Green claimed a 62% reduction in the pool's gas heat usage. I would not treat those numbers as universal. I would treat them as site-specific evidence that, under the right conditions, the heat stream is large enough to matter.

Swimming Pools Are A Good First Customer

A pool is a more believable heat buyer than a vague community promise. It is already wet, already plumbed, already paying for heat, and already suffering when energy costs rise. The BBC article says Exmouth Leisure Centre expected its energy bills to rise by £100,000 that year, and it connects the story to a wider pressure point: BBC News had reported that 65 swimming pools had closed since 2019, with rising energy costs cited as a significant reason.

That does not make the server box a substitute for a heating system. Ars notes that the pool still has a gas boiler to boost temperature when required. That caveat makes the story more credible, not less. The honest pitch is not "servers replace boilers everywhere." It is "some sites can offset a meaningful share of heat demand when compute is colocated with the load."

This is where I think the public narrative often gets too smooth. Heat reuse is not the same thing as zero-impact compute. It is better thought of as double use: if the computation is going to happen anyway, design the facility so the waste output displaces a local heating input.

The Business Model Is The Weak Point To Watch

Deep Green's own model depends on selling the compute. The BBC says the company charges clients for computing power used for artificial intelligence and machine learning. TNW later reported that Deep Green wanted to scale to 100-150 swimming pools after an Octopus Energy investment, but also quoted the company saying it needed more corporates to stop using traditional data centres and use its servers so it could give heat to communities.

That is the constraint I find most important. The heat host may be delighted, but the system does not work because a pool has spare space. It works only if Deep Green can keep the servers occupied with workloads that tolerate this deployment model. A Hacker News discussion around the Ars article surfaced the same concern: commenters asked who would buy small, local colocation capacity and whether Deep Green could aggregate enough distributed compute demand.

I do not think that objection kills the idea. It makes the execution bar visible. A distributed data-centre network has to sell reliability, support, security, connectivity, hardware maintenance, and workload fit. The heat story may win attention, but the compute customer is still buying infrastructure.

The AI Context Makes This Less Niche

When this BBC story ran in March 2023, it could read like a clever local energy hack. In 2026, it reads more like an early version of a bigger infrastructure question. Deep Green's current site now frames the company around AI and high-performance-computing colocation, with live capacity in Manchester. It says its DG01 Manchester site donates heat to Move Urmston leisure centre, saving around £80,000 a year on community heating costs and reducing CO2 emissions by 100-150 tonnes.

The reason this matters is that data-centre demand is no longer a background technical issue. The International Energy Agency's Energy and AI report says data centres accounted for around 1.5% of global electricity consumption in 2024, or 415 TWh, and projects that data-centre electricity consumption will more than double to around 945 TWh by 2030. The same report stresses that local impacts can be far more pronounced than the global share suggests.

That is why heat reuse should not be treated as the whole answer. It is one design lever inside a larger energy problem. Locating compute near useful heat loads can reduce waste, but it does not erase electricity demand, grid constraints, capital requirements, or the emissions profile of the power supply.

Distance And Temperature Decide A Lot

The strongest skeptical point is not that waste heat is fake. It is that waste heat is often awkward. The Ars writeup summarized the broader challenge well: distributing heat can involve cost, local district-heating access, technology constraints, and recipients who may need heat pumps because the available heat is not always hot enough on its own. Physical closeness matters; so does the temperature of the heat stream.

That is why the pool example is unusually tidy. A leisure centre has constant water circulation, a large thermal mass, and a heating load that can accept steady contribution. A distant housing project, an office district, or a seasonal heat user may be much harder to serve. The same server output can look valuable in one building and nearly useless across the wrong boundary.

My Bottom Line

I like the Exmouth story because it narrows a giant AI-energy debate into a concrete loop. Instead of saying data centres should be sustainable in the abstract, it asks whether one box of servers can sit beside one real heat load and reduce one real bill.

But I would not promote this as a universal fix. The practical lesson is more disciplined: put compute where its waste heat has a buyer, prove the workloads are real, keep backup systems honest, and measure the net result at the site. If the AI boom is going to keep filling buildings with hot machines, the least convincing answer is to pretend the heat disappears. The more useful answer is to design neighborhoods for it.