AI’s infrastructure boom runs into a water constraint
Server cooling fans in a datacenter. Photo by Winston Chen on Unsplash.
The global expansion of artificial intelligence has encountered a physical constraint that has little to do with chips, models or investment: many of the regions attracting new AI data centres are already struggling with water scarcity.
The tension extends from drought-sensitive parts of Spain and India to water-stressed regions of the United States. Britain has become another prominent example. In July, Water UK, the trade body representing the water industry, warned that government plans for AI growth had failed to account properly for the water required by data centres.
More than three-quarters of Britain’s facilities are concentrated in the water-stressed south and east of England. Some proposed sites in Affinity Water’s region have requested as much as three million litres a day.
The warning reflects a much wider collision. Governments see domestic computing capacity as essential to economic competitiveness, technological sovereignty and national security. Yet dense racks of AI processors generate immense quantities of heat, and some data centres remove that heat using cooling systems that evaporate water.
Water UK estimates that data centres in England currently use 6.6 million litres of drinking water a day. If capacity triples by 2030, that figure could rise to 19.8 million litres unless cooling practices change. This is a scenario based on projected expansion, rather than a firm forecast, but it shows why water demand must be considered before new AI infrastructure is approved.
The numbers will differ by country, climate and facility type. The underlying conflict, however, is increasingly global: AI infrastructure is being expanded in places where households, agriculture and industry already compete for limited water supplies.
Can the sector continue growing without draining those supplies? The short answer is that data centres can sharply reduce or even eliminate water consumption from cooling during normal operation. That does not make AI waterless. Electricity generation, semiconductor fabrication, construction and other facility operations can all carry a water footprint.
The engineering response spans the computing stack. Chip designers are enabling systems to work with warmer coolant; startups are moving liquid closer to the silicon or immersing entire servers; hyperscalers are building closed loops and substituting recycled water for drinking water. Together, these approaches suggest that water consumption is becoming a design constraint rather than an issue to address after a data centre has been built.
NVIDIA: letting AI hardware run hotter
NVIDIA’s forthcoming Vera Rubin platform starts by changing the temperature at which its rack-scale systems can be cooled. The associated reference architecture is designed for single-phase liquid cooling supplied at up to 45°C, capturing heat directly from chips and networking equipment.
In suitable conditions, that heat can be released through outdoor dry coolers using ambient air. Conventional cooling towers evaporate water, while chillers use electricity to drive compressors. A sufficiently warm cooling loop can reduce the need for both.
NVIDIA says the architecture could cut facility cooling-water consumption from approximately 2.6 million US gallons per megawatt annually in a conventional cooling-tower system to near zero. The company is also publishing its DSX AI factory reference design to help data-centre operators build around the new platform.
“Near zero” refers specifically to water consumed by the facility’s cooling system. It excludes water used to manufacture chips, construct the centre or generate electricity. Actual performance will depend on local climate, workload and facility design.
Nevertheless, Rubin makes thermal management part of the computing architecture rather than a system added afterwards. If the design performs as intended, operators could deploy high-density AI hardware in a wider range of water-constrained locations without relying as heavily on evaporation.
Corintis and Microsoft: moving cooling into the chip
Swiss startup Corintis is going closer still. Founded in 2022 as a spin-off from the Swiss Federal Institute of Technology Lausanne, or EPFL, the company develops microfluidic cooling technology that uses microscopic liquid channels to target hotspots in powerful processors.
Most direct-to-chip cooling systems attach a cold plate to a processor and circulate liquid through it. Microfluidic cooling instead brings the coolant closer to the silicon. Corintis uses simulation software to optimise channel layouts, including branching structures inspired by patterns such as leaf veins.
Microsoft collaborated with Corintis to optimise the bio-inspired channel design for an experimental in-chip cooling system. In Microsoft’s tests, the prototype removed heat up to three times more effectively than conventional cold-plate technology and reduced the maximum temperature rise inside the silicon by 65%.
The result belongs to Microsoft’s broader research programme rather than representing a commercially deployed Corintis product. It nevertheless demonstrates what closer integration between chip and cooling design could achieve.
Extracting heat nearer its source could support higher chip densities and reduce the energy required to move coolant. It may also help processor designers manage increasingly concentrated hotspots as AI chips become more powerful.
Corintis raised $24 million in a Series A round in September 2025. Its challenge is manufacturability: laboratory performance must translate into reliable mass production and integration with semiconductor packaging. Cooling systems must operate continuously under demanding conditions and be serviceable across global data-centre fleets.
The company illustrates a broader change in AI infrastructure. As processors become more powerful and heat densities rise, cooling is moving from the building towards the server and, ultimately, into the design and packaging of the chip itself.
Microsoft: closing the data-centre cooling loop
Microsoft is also applying direct-to-chip cooling at facility level. Its latest AI data-centre design circulates liquid between servers and cooling equipment in a closed loop. Once the system has been filled during construction, the water is continuously reused rather than lost through evaporation.
The company estimates that eliminating evaporative cooling can avoid more than 125 million litres of water consumption annually at each data centre using the design. Microsoft began introducing the architecture in new projects in August 2024 and says it will be used at forthcoming facilities in Spain, including new AI data centres in Zaragoza — a region where prolonged drought has made industrial water use particularly sensitive.
The cooling design forms part of a broader water strategy that reached a significant milestone in 2025. Microsoft says it became “water positive” five years ahead of its 2030 target, replenishing more water than it withdrew across its global operations during the year.
The company also reduced the water-use intensity of its data centres by 25% against a 2022 baseline, passing the halfway point towards its target of a 40% reduction by 2030.
Those figures measure two different forms of progress. Water-use intensity shows how much water Microsoft withdraws relative to the computing capacity it operates. “Water positive” compares its global withdrawals with the estimated benefits of projects such as wetland restoration, watershed protection and more efficient irrigation.
Neither means that every Microsoft data centre returns more water to the same watershed from which it withdraws it. Global replenishment accounting can balance consumption in one location against water restored or saved elsewhere. That distinction is particularly important when a facility operates in a region facing seasonal shortages.
There is still water inside Microsoft’s closed-loop system, and its “zero water” description applies only to cooling during normal operation. Water may still be used elsewhere at the site, including in kitchens and toilets.
Closed-loop cooling also introduces a trade-off. Replacing efficient evaporative cooling with dry coolers or air-cooled chillers can increase electricity demand, especially on hot days. Microsoft estimates that its new design produces a nominal increase in annual energy use compared with evaporative cooling.
Whether that trade-off is justified depends on local water stress, climate, electricity supply and the carbon intensity of the grid. A design that makes sense in a water-scarce region may be less attractive where water is abundant but electricity is carbon-intensive or constrained.
Submer: putting entire servers in liquid
Barcelona-based Submer removes air from the equation more completely. Its immersion-cooling systems place servers in a non-conductive fluid that absorbs heat directly from their components. Pumps transfer the heat to an external exchanger, eliminating many of the fans and air-conditioning systems used in conventional facilities.
Submer is developing its first owned and operated 56MW data centre in Barcelona. The company says the facility will showcase liquid-cooling technology and support computing densities above 150kW per rack or immersion tank.
It has not publicly provided enough technical detail to verify the project’s eventual operational water consumption, so it is premature to describe the Barcelona facility itself as consuming no water for cooling.
In July 2025, Submer also signed a memorandum of understanding with the government of Madhya Pradesh to explore the joint development of up to one gigawatt of liquid-cooled, AI-ready data-centre capacity in India. The agreement covers design standards, supply-chain development and workforce skills.
Submer says the proposed facilities would use no water for cooling, but the announcement describes an ambition under an agreement rather than completed or fully financed infrastructure.
The projects show why water-efficient cooling has international relevance. Spain has experienced prolonged drought, while India is simultaneously expanding its digital infrastructure and managing severe pressure on water supplies in many regions. If AI capacity is to grow in such markets, conventional assumptions about cooling may no longer be workable.
Immersion cooling supports dense computing in a smaller footprint but is less straightforward as a retrofit. Operators must consider hardware compatibility, maintenance procedures, the coolant’s environmental properties and how captured heat is ultimately discharged. Removing a cooling tower does not make heat rejection disappear.
Submer’s technology therefore addresses one part of the problem. Its wider value will depend on whether immersion cooling can be deployed reliably and economically across different climates, electricity systems and data-centre designs.
AWS: reducing water consumption in existing data centres
Not every data centre can be rebuilt around 45°C coolant, microfluidic channels or immersion tanks. Much of the infrastructure expected to support rising AI demand already exists or is being developed using more conventional architecture.
AWS is therefore combining liquid cooling for new AI racks with more incremental water-management measures across its estate.
Its in-row heat exchanger captures heat close to high-density hardware before it spreads through the room. AWS expects the system to reduce water use by about 9% compared with an evaporatively cooled, air-based data centre once it is fully deployed. The company also uses real-time operational data and sensors to identify leaks and adjust cooling systems.
Where water is still required, AWS is increasingly replacing drinking water with treated wastewater. It reported that 24 data centres were using recycled water in 2023. A forthcoming Hong Kong facility is being designed to use recycled water for cooling in cooperation with local authorities; as of the company’s latest update, the system remained in the detailed design phase.
That approach is less dramatic but potentially significant. Optimising existing facilities and changing their water source may deliver earlier savings than waiting for an entirely new infrastructure generation.
It may also be more practical in regions where operators cannot redesign a data centre from the ground up but can reduce its demand for drinking water.
Recycled water does not eliminate consumption, and it may still have competing local uses. But it can reduce pressure on potable supplies and demonstrates that the water question is not limited to cooling technology. It also concerns where water comes from, how efficiently it is managed and whether facilities are designed around the conditions of their local watershed.
What does ‘zero water’ actually mean for an AI data centre?
Emerging cooling technologies make it possible to envisage AI data centres that consume little or no water for cooling during normal operation. They do not make AI waterless.
Electricity generation can carry a substantial water footprint, while semiconductor fabrication requires large quantities of ultra-pure water. Cooling loops need an initial fill, construction consumes water and facilities still require water for other operational purposes.
Corporate “water positive” targets introduce another layer of complexity. Microsoft’s achievement means that its estimated global replenishment exceeded the water withdrawn by its operations in 2025. It does not mean that the company stopped consuming water or that every local watershed received back the water taken from it.
Replenishment projects may produce genuine benefits, but their location, timing and measurement determine whether they relieve pressure around a particular data centre.
The terminology surrounding data-centre water consumption can obscure other important differences. “Zero water for cooling” is not the same as zero water across a facility, supply chain or energy system. A percentage reduction against an evaporatively cooled reference design cannot be compared directly with a closed-loop claim, while replacing drinking water with treated wastewater changes the source rather than eliminating consumption.
Companies do not always distinguish consistently between water withdrawal and water consumption either. Withdrawal measures the amount taken from a source, even if some is later returned. Consumption refers to water that is evaporated, incorporated into a product or otherwise unavailable for immediate reuse.
Without a common definition and clearly stated baseline, apparently comparable claims may describe different outcomes.
Which data-centre cooling technology uses the least water?
There is no universally preferable cooling system.
Evaporative cooling can reduce electricity use but consumes water. Dry cooling conserves water but may require more electricity, particularly in hot weather. Direct-to-chip systems handle very high heat densities, while immersion cooling removes more air-cooling equipment. Recycled water protects drinking supplies without ending water consumption.
Local conditions determine which trade-off is defensible. Climate, watershed stress, electricity availability, grid emissions, chip density and the age of the facility all influence the calculation.
Those conditions must increasingly be considered before new capacity is approved. Data-centre developers may need to disclose expected water withdrawals and consumption, the source and quality of that water, seasonal demand, cooling technology and the consequences of drought restrictions.
Governments pursuing AI sovereignty will also have to decide where additional computing capacity can be built without intensifying competition for an already constrained resource. National capacity targets mean little if the required infrastructure cannot be supported by local power and water systems.
Water can no longer be treated as an unlimited utility arriving at the site. It is becoming a global AI infrastructure constraint that reaches back through land-use planning, server architecture and semiconductor design.
For governments and companies building AI capacity — particularly in water-stressed regions — the question is no longer simply how many accelerators they can install, but how much computing they can produce for every litre they can responsibly use.
Liked this article? You can support our independent journalism via our page on Buy Me a Coffee. It helps keep MoveTheNeedle.news focused on depth, not clicks.