This article was written as the crisis it describes was still unfolding. As of Saturday 30 May 2026, South East Water has restored supply to 15,000 customers across Kent but approximately 500 remain completely without water and a further 3,000 are experiencing intermittent pressure across Coxheath, Loose, Headcorn, Ulcombe, Kemsing and Benenden. Storage tank levels remain critically low. The company is still asking all customers to restrict use to drinking, cooking and hygiene. A Kent MP has said he has lost all faith in the company. Bottled water stations remain open.
Before anyone in Westminster or Brussels draws up another data centre masterplan, they should look at what is happening in Kent right now.
A pumping station near Charing failed under the pressure of a heatwave. Around 800 properties across Charing, Challock, Molash, Herne Bay, Whitstable, Coxheath and parts of Ulcombe lost supply or saw pressure collapse. South East Water asked all customers — not just garden hosers, not car washers — to restrict use to drinking, cooking and hygiene. Bottled water stations reopened at Challock Village Hall and Sainsbury’s Altira Park. Kent County Council, still absorbing the embarrassment, announced a new Kent Water Resilience Partnership on 28 May 2026 to impose strategic oversight on a system that had just demonstrably failed.
This was not an isolated incident. It was the latest expression of a structural condition that has been building for three years.
The Environment Agency’s own figures tell the story in Ml/d — megalitres per day, the unit that separates engineering from rhetoric. In 2022–23, a drought summer that recorded the driest spring in England since 1893 triggered a leakage spike across the industry; most companies saw increases compared to 2021–22, while South East Water alone recorded a 50 per cent increase in burst pipes as hot, dry ground shrank around its 9,000-mile network and pipe joints gave way. The 2023–24 reporting year brought a 3.7 per cent reduction, taking national total leakage to 2,690 Ml/d — described by the Environment Agency as the lowest level on record at that point. Progress, of a kind; but 19 per cent of all water put into supply was still being lost as leakage, and the Agency was explicit that a major step change was still required. The 2024–25 figures slowed further: just 2.7 per cent annual reduction across the sector, compared to 4.6 per cent the prior year; only four companies met their performance commitment level.
The three-year average for England and Wales from 2022 to 2025 sits at 2,967 Ml/d. That is 1,187 Olympic swimming pools of treated, pressurised, infrastructure-dependent water lost every single day. Thames Water, the UK’s largest supplier and a company currently restructuring under the weight of £18.7 billion in debt, leaked 570.4 Ml/d in 2023–24; more than 200 billion litres over the year, roughly a quarter of its entire supply. That was their best-ever performance. They still missed their regulatory target by a wide margin.
The longer view is worse. Since privatisation in 1992, England and Wales water companies have leaked a cumulative 41.4 trillion litres of treated water. Over the same period, those companies paid out £52.7 billion in dividends. Both figures come from Ofwat’s own published spreadsheets.
Into this system — leaking, stressed, ageing, already failing on warm bank holiday weekends — the government now proposes to introduce a new class of industrial consumer with a thermal physics problem that cannot be solved without very large quantities of cold water.
The dominant narrative around data centre expansion has been about electricity. Gigawatts. Grid capacity. Planning consent for new generation. It is not wrong; the EU’s data centre IT load sits around 10 GW today and is projected to reach 35 GW by 2030, consuming up to 9 per cent of Europe’s total electricity. Germany is already seeing grid congestion drive simultaneous negative and positive power prices in adjacent provinces because the transmission infrastructure cannot move electrons fast enough.
But the power debate has crowded out a more immediate constraint, one that cannot be solved by building more generation capacity or connecting to a new substation. It is a constraint written into the physics of the chips themselves.
Nvidia’s current Blackwell GPU generation produces up to 1,000 watts per chip, more than three times the thermal output of GPUs from seven years ago. AI rack densities have moved from a typical 15 kW per rack to 120–132 kW, with some deployments pushing beyond that. At those densities, air cooling is not merely inefficient — it is physically impossible. The volume of air required to remove that much heat at any practical velocity exceeds what engineering can deliver in a contained space. Liquid cooling is not an option or an upgrade; it is the only viable thermal management approach for AI-scale compute.
Liquid cooling means cold plates mounted directly on processors, with coolant circulating through a distribution unit before the heat is rejected somewhere. That somewhere, in most practical large-scale deployments, means evaporative cooling towers; systems that reject heat to atmosphere by evaporating water. Large hyperscale facilities are currently consuming between 11,356 and 18,927 cubic metres of water per day. The AI chip in a data centre is not just a power consumer. It is a water consumer, structurally and continuously, at a rate that scales directly with compute density.
The industry’s preferred rebuttal at this point is the closed loop. Advanced direct-to-chip cooling systems circulate coolant in a sealed circuit, rejecting heat through dry radiators rather than evaporative towers. No open water consumption; problem solved. The rebuttal is half right and half thermodynamics. A closed loop is only closed until it meets the ambient air temperature. In rejecting heat into a 35°C summer atmosphere; the precise conditions under which a Kent pumping station failed and hands-off flow restrictions spike across English catchments; dry-air radiators lose efficiency rapidly. The compensating mechanism is adiabatic assist: spraying water onto external cooling fins to reduce the incoming air temperature through evaporation. The closed loop opens, on the hottest days, in the most water-stressed conditions, at exactly the moment of peak AI compute load. The wet-bulb temperature is not a design parameter that lobbying can vary.
Cooling infrastructure already accounts for roughly 38 per cent of an average data centre’s energy footprint, before a single byte of user data has been processed. For AI deployments at current rack densities, that proportion rises further. Microsoft, Google, Amazon and Meta are collectively spending more than $380 billion on AI infrastructure in 2025. Every dollar of that spending creates derivative demand for thermal management. Every thermal management system creates derivative demand for water.
Google, large enough to run parallel infrastructure strategies simultaneously, has publicly adopted a bifurcated approach: liquid cooling for the highest-density AI workloads, and what it describes as “advanced air cooling” only in Texas and Nevada facilities — explicitly chosen to navigate water scarcity. A hyperscaler is selecting site geography partly based on water availability. That is not a future planning consideration. It is a current operational constraint, acknowledged by the largest builders in the industry.
The second rebuttal the industry reaches for is geographic displacement: move the water-intensive training workloads to Iceland or Finland, where cold ambient air and abundant hydro reduce the thermal problem considerably. The argument has merit for training, the long, batch-oriented jobs that can tolerate latency. It has no merit for inference, which is the workload that actually powers the products people use. Every AI search query, every real-time financial decision, every automated content moderation call is an inference job. Inference is latency-sensitive; it must execute close to the end user to remain usable. The compute serving London cannot live in Reykjavik. The facilities that must sit in or near the major European urban centres; London, Frankfurt, Amsterdam, Paris; are precisely the facilities in the most water-stressed catchments, competing with the densest domestic populations for the most constrained supply. The industry is hydrologically trapped by its own need for speed.
A single large AI data centre consuming 15,000 cubic metres of water per day is a significant industrial draw. Now consider the UK government’s ambitions, replicated across Europe.
The European Commission’s AI Continent Action Plan, launched in the wake of the February 2025 Paris AI Action Summit, aims to triple Europe’s data centre capacity by 2030. The UK, post-Brexit and anxious not to fall behind, is pushing its own expansion agenda. Every new facility proposed; and there are dozens in various stages of planning; needs a site, a grid connection, and a water supply. The site and the grid connection get planning scrutiny. The water supply, until very recently, has been treated as an overhead to be resolved in the detail phase.
It cannot be resolved in the detail phase if the water is not there.
In England, the abstraction licensing system administered by the Environment Agency is the mechanism through which industrial water access is controlled. Any abstraction of more than 20 cubic metres per day from a surface or groundwater source requires a licence. A data centre consuming 15,000 cubic metres per day is not applying for a minor licence variation; it is entering a competition for a resource that regulators have already determined is over-allocated in many catchments across southern and eastern England; the precise geography where data centre demand is concentrated.
The Environment Agency cannot grant new or amended abstraction licences that adversely affect the rights of existing abstractors. In catchments that are already over-licensed, new applicants face strict conditions or outright refusal. There are no new licences available for most year-round surface water abstractions across much of England and Wales; the only route to a new supply is to trade abstraction rights from an existing licence holder, a market that is thin and not designed for the volumes an AI data centre requires.
Then there are hands-off flow conditions. Many existing abstraction licences carry hard limits that require the holder to stop abstracting entirely when river flows fall below a defined threshold. These are not discretionary; they are licence conditions, legally enforceable. During the drought conditions of July 2025, approximately 780 hands-off flow restrictions were simultaneously in force across England, rising by 130 in a single week. A data centre with an abstraction licence in a stressed catchment does not have a guaranteed water supply. It has a licence to abstract water when there is enough of it; which is precisely not when cooling demand peaks.
From 2028, the position tightens further. The Environment Agency gains enhanced powers under the Environment Act 2021 to vary or revoke abstraction licences without compensation when abstraction is putting environmental conditions in jeopardy. The Cunliffe Report of 2025 and the subsequent White Paper commit to tighter abstraction licensing rules across the board, with a full review of all licences through the late 2020s and early 2030s. A data centre that secures an abstraction licence in 2026 cannot treat that licence as permanent infrastructure.
The failure mode is specific and fast. Cooling interruption at a facility running Blackwell-generation GPUs at 130 kW per rack does not result in graceful degradation. Chip junction temperatures have tight thermal envelopes; sustained overtemperature causes permanent electromigration damage to interconnects. The data centre does not reboot. It replaces hardware. The same bank holiday heatwave that collapsed South East Water’s pumping station near Charing; the same meteorological event that dries ground, shrinks soils, triggers hands-off flow conditions and spikes domestic water demand; is precisely the event that drives AI compute to its maximum thermal output. The two failure modes are not independent. They are correlated, driven by the same weather system, competing for the same stressed resource.
The question of whether the water was ever formally assessed for the new builds has a specific and documented answer: for the most prominent fast-tracked approvals, it was not.
All three large-scale data centre projects approved by the Labour government through 2024 and 2025; at Iver in Buckinghamshire, Abbots Langley in Hertfordshire, and a third site; were consented by then-Secretary of State Angela Rayner without a full Environmental Impact Assessment. Local councils had concluded, on the basis of developer assurances, that no EIA was required. The government upheld that screening decision on appeal, overriding local authority refusals on green belt grounds.
The legal challenge brought by Foxglove and Global Action Plan argued the decisions had failed to consider “obviously material factors such as the development’s water needs, electricity demand and associated carbon emissions.” Leigh Day, acting for the claimants, stated the Inspector had failed to consider whether the data centre’s use of natural resources, “including the massive quantities of drinking water required to keep it operational”, would be likely to have a significant effect on the environment.
In January 2026, the Secretary of State conceded the point. The Buckinghamshire consent was accepted for quashing on the grounds of a “serious logical error”: the EIA screening had relied on mitigation measures that were not, in fact, secured within the consent itself. The Hertfordshire approval, consented on identical terms by the same Secretary of State, now sits under precisely the same legal exposure.
The government’s own 2025 data centre policy paper, published in the midst of this litigation, acknowledges that “vast amounts” of water may be used depending on cooling method, notes “limited transparency and data on water consumption,” and states that industry is “being encouraged to contact their proposed water and wastewater supplier early in the planning process.” Encouraged. For facilities that, at AI hyperscale density, will consume more water per day than a small town.
The power station analogy is instructive here. A coal or gas generation facility operating cooling towers requires an abstraction licence with hands-off flow conditions, a thermal discharge consent for warm blowdown water returned to a river or estuary, and a full environmental permit under the Industrial Emissions Directive. The regulatory framework exists because regulators learned in the 1960s what concentrated industrial heat rejection does to aquatic ecology. A large AI data centre rejects comparable quantities of waste heat through comparable mechanisms. The planning frameworks under which the consented facilities were approved were written for office buildings. The Environment Agency is being asked to regulate what is functionally a power station that applied for permission as a warehouse.
The United Kingdom’s abstraction licensing framework finds its European equivalent in the Environmental Flows regime embedded in the EU Water Framework Directive. The mechanism is the same; the institutional machinery is larger and the legal consequences of non-compliance more severe.
Under the WFD, member states are legally obligated to bring water bodies to “good ecological status.” Environmental flows are the binding minimum water volumes that must remain in rivers to maintain that status. They are not advisory targets. A data centre that abstracts below E-flow thresholds is not just breaching its own permit; it is potentially triggering member state liability for WFD compliance failures, with European Commission legal proceedings to follow.
Spain has moved furthest and fastest in converting this pressure into domestic regulation. The Spanish Ministry for Ecological Transition (MITECO) has advanced a Draft Royal Decree transposing the EU’s Energy Efficiency Directive that goes materially beyond the Directive’s own requirements. Under the draft: any data centre over 1 MW must publish an absolute water reduction strategy annually, detailing potable versus recycled water use; any data centre over 100 MW must demonstrate that its facility sits in the top 15 per cent of global performance for Water Usage Effectiveness, alongside Power Usage Effectiveness, energy reuse factor and renewable energy ratio. Grid connection permits are conditional on demonstrated compliance. The Council of State report is the only remaining procedural step before Council of Ministers approval; the direction is not in doubt.
The practical effect of the top-15-per-cent WUE requirement for large facilities is that the bar is set against the global frontier; not the EU average, not the national average, but the best performers worldwide. A facility that clears that bar in 2026 may not clear it in 2028 as the frontier moves. And even a facility that clears it faces the Ebro and the Guadalquivir, river basins operating under structural drought protocols where local authorities retain the power to shut off industrial water intake entirely when E-flow thresholds are breached. No WUE score prevents a dried-up riverbed from refusing to supply coolant.
The Netherlands presents a different species of the same genus. The country’s nitrogen crisis; which froze construction projects across the nation through a sequence of court rulings when it was determined that cumulative nitrogen emissions were breaching environmental law; established a legal precedent that European water lawyers are now applying to water. Dutch water boards and permit issuers including Rijkswaterstaat are actively auditing existing industrial abstraction permits for WFD compliance. A data centre in Amsterdam or Groningen cannot treat a historical water permit as a stable asset. If the local water table or canal system drops below the critical ecological baseline, those permits can be unilaterally altered or revoked without compensation. The nitrogen crisis showed that when the legal wall arrives, it arrives without warning and without a grace period.
The industry’s response to this regulatory environment has been instructive. Microsoft and DigitalEurope; the lobby group whose members include Amazon, Google and Meta; successfully lobbied EU lawmakers in 2024 for a confidentiality carve-out in the implementing regulation under the Energy Efficiency Directive. Individual data centre environmental data; energy use, water consumption, the metrics that would allow communities, researchers and local authorities to understand what was being proposed; was classified as commercially sensitive. The Commission incorporated language drafted by Microsoft and DigitalEurope almost verbatim. A senior Commission figure subsequently instructed national authorities that they were “obliged to keep confidential all information and key performance indicators for individual data centres.” Ten leading legal scholars told the investigative journalism cooperative Investigate Europe that the clause likely violates the Aarhus Convention, the international treaty guaranteeing public access to environmental information.
The same companies that cannot tell you how much water their existing facilities consume are the ones applying for planning permission and abstraction licences to build facilities that will consume tens of thousands of cubic metres per day from catchments already operating under hands-off flow restrictions.
Europe is approximately three years behind the United States in the confrontation between data centre expansion and community water rights, but the trajectory is visible.
In the American West, the collision is already acute. Water use is the single most cited reason for local opposition to proposed data centre projects, mentioned in more than 40 per cent of contested cases. In 2025, local opposition led to the delay or cancellation of projects totalling $156 billion. Twenty-five data centre projects were cancelled due to local opposition in 2025 alone — 21 of them in the second half of the year, as awareness spread between communities that had successfully blocked or delayed proposals.
The agricultural dimension is direct. In The Dalles, Oregon — long a concentration of hyperscaler data centres along the Columbia River — farms and orchards have reported severe droughts and declining aquifers. In rural Georgia, communities near Meta facilities report depleted water supplies. In Arizona, data centres compete directly with agriculture for groundwater in a state already managing structural deficit. In August 2025, Tucson City Council unanimously rejected an Amazon-linked data centre on water grounds alone.
The Spanish conflict is sharper still. Amazon’s data centres in Aragon are expected to consume enough water to irrigate 233 hectares of corn — one of the region’s primary crops. In a basin already strained by water-intensive export agriculture, local opposition has coalesced under the name Tu Nube Seca Mi Río — your cloud is drying my river. The name is not metaphor. It is a description of the hydrological mechanism.
One in five US data centres in 2021 was already sited in water-stressed regions — the developers following cheap land, low power costs, favourable tax treatment and a regulatory environment that had not yet caught up. The communities that noticed first are now the template for the communities that will notice next.
The industry is not without a technical response to the water constraint, and fairness requires it to be heard before it is weighed.
Two-phase immersion cooling — where servers are submerged in dielectric fluids that boil at low temperatures, carrying heat away as vapour before condensing and recirculating — is genuinely water-minimal at the chip level. No evaporative loss; the cooling medium stays in a closed cycle. A lifecycle assessment published in Nature in May 2025 by Microsoft researchers found that switching from air cooling to advanced liquid cooling reduces water consumption by 31 to 52 per cent across the full facility lifecycle. Specialist firms including Iceotope and LiquidStack are deploying immersion systems commercially in high-density AI environments, and the industry consensus is that evaporative cooling will eventually be phased out in new builds.
The timeline is the problem. The EU’s AI Continent Action Plan aims to triple data centre capacity by 2030. Immersion cooling at hyperscale requires purpose-built facilities; it cannot be retrofitted to the majority of existing or consented buildings currently in the planning pipeline. The mix through the late 2020s will remain predominantly evaporative and adiabatic-assist systems — precisely the technologies that compete most directly with domestic water supply under heatwave conditions.
Desalination has been proposed as a longer-term structural solution. The concept has a certain engineering elegance: draw seawater from an estuary or coastal intake, use waste heat from the data centre to assist the desalination process, produce potable or process water as a byproduct. Patents for combined data centre and desalination co-location exist. England’s Environment Agency now supports desalination as a component of national water resilience planning, and Anglian Water has contracted the technical development of proposed plants in the driest parts of the country.
The constraint is everything downstream of the concept. Reverse osmosis desalination requires roughly 3 to 4 kWh per cubic metre of product water — so a data centre solving its water problem through desalination is simultaneously enlarging its power consumption, in a sector already projected to consume 9 per cent of Europe’s electricity by 2030. Coastal and estuarine siting brings marine licensing, brine discharge consents, and habitat regulation assessments; England’s major estuaries — the Thames, the Medway, the Humber, the Severn — are designated Special Protection Areas and Sites of Special Scientific Interest in their lower reaches. The planning and consent timeline for a new coastal desalination plant serving a major data centre cluster runs to a decade or more.
Liquid nitrogen cooling, occasionally surfaced as a water-free alternative, has not scaled commercially beyond laboratory and specialist high-performance computing environments. The logistics of continuous cryogenic supply to facilities consuming the volumes an AI hyperscale requires are prohibitive; the concept remains an engineering curiosity rather than a deployable solution at the scale and timeline the expansion agenda demands.
The technology roadmap points in the right direction. It does not arrive in time for the expansion agenda currently being approved.
Europe frames data centre expansion as a sovereignty project. The AI Continent Action Plan, the Cloud and AI Development Act, the tripling of processing capacity — all of it is presented as the continent reducing its dependency on infrastructure it does not control, building the physical substrate of a digital future on European ground.
The argument is not wrong in principle. Sovereign infrastructure requires sovereign physical plant. But sovereignty requires more than a planning permission and a grid connection. It requires stable power, skilled operators, and — the ingredient currently missing from every ministerial briefing and every hyperscaler feasibility study — reliable, legally secure, climatically resilient access to cold water.
The UK has a water distribution system leaking nearly 3,000 megalitres per day through Victorian-era pipes, abstraction licences that can be suspended by drought conditions in the catchments most attractive to data centre developers, and an incoming regulatory tightening that will make existing licences less secure, not more. The EU has the Water Framework Directive as a hard legal ceiling, member states that are actively auditing industrial water permits for compliance, and a Spanish regulatory environment that has set the WUE bar at the global frontier for any facility above 100 MW.
The data centre industry’s response — lobbying environmental data into commercial confidentiality, seeking to apply “the simplification omnibus principle” to water policy, framing minimum performance standards as a burden — does not suggest an industry that has absorbed the scale of the constraint it is operating under.
A pumping station failed near Charing on a warm bank holiday weekend. Hundreds of properties lost pressure. Bottled water stations opened. A county council announced an emergency oversight body. Meanwhile, the government is approving expansion of the most thermally intensive compute infrastructure ever built into the same water catchment areas, under the same meteorological conditions, competing for the same resource.
The cloud runs on water. The infrastructure to supply that water is already failing under current demand. The AI ambitions of both the UK and the EU assume a physical resource base that the engineering record of the last three years suggests cannot be assumed.
The raindrops haven’t formed yet.
The Sovereign Auditor covers supply chain security, digital sovereignty, and infrastructure policy—with particular focus on Isle of Man jurisdiction and Crown Dependency issues.
Payments via PayPal. Credentials delivered by email. No Substack. No Stripe. No middlemen.
