Calls to halt new datacentres often focus on tales of their gargantuan appetite for water and electricity. Such arguments are magnified in current conditions. Drought was declared throughout large parts of England in July. Wildfires across the parched countrysides of France and Spain forced 320,000 people to flee. The Danube fell too low to cool Hungary’s key nuclear reactor, while downstream Romania blasted rocks to improve the river’s water flow.
In the US, politicians have blamed Virginia’s 11GW of datacentre capacity for an estimated daily drain of three million gallons (11.3 million litres) from each of its groundwater budget regions, arguing a single hyperscale datacentre can consume 300,000 gallons (1.1 million litres) a day.
Meanwhile, there has been a lot of noise around how much artificial intelligence (AI) could increase datacentre water consumption. The International Energy Agency has estimated energy use of an 100MW hyperscale AI facility as equivalent to 100,000 households. Water UK has estimated UK datacentres use about 6.6 million litres of water daily and that AI might drive water demand up 129% by 2050, versus 2025.
But what is the scale of water use by datacentres that are coming online now, and in the UK in particular?
This article looks at the water consumption of modern datacentres and tries to separate alarming headlines from real-world engineering practices. It examines how different cooling technologies, server efficiencies and geographic locations shape the environmental footprint of digital infrastructure.
Datacentre water use varies
While some US datacentres are among the most voracious water users, datacentre water consumption varies a lot. The Universities of California and Illinois have estimated datacentre water use varying by 10,000x, citing 1,000x differences in water consumption per kilowatt-hour (kWh) of server electricity consumed and 10x differences in server workload efficiency.
Equations change according to server utilisation rates, types of workloads, cooling to the processor, heat rejection externally to the datacentre, and so on. Datacentres may consume little water for cooling – far less than for irrigation, for example – while higher density servers need more cooling, and older servers are less efficient.
Aaron Binkley, Digital Realty’s sustainability director, says 75% of its 300-plus datacentres don’t use water for cooling at all. “They’re using air cooled or closed-loop systems that don’t evaporate water,” Binkley says. “The biggest water use in our air-cooled datacentres is landscape irrigation – no different than an office, hotel or any other building with landscape.”
A big problem is, however, that there’s no standard way to know how much water datacentres use. The Uptime Institute finds only 50% of operators so far calculate or report their water use, even though water usage effectiveness (WUE) – annual water use in litres divided by IT energy consumption in kWh – isn’t difficult to measure or monitor.
Cooling system type matters
To make sense of datacentre thermal management, it helps to divide cooling into two distinct stages: Primary cooling (how heat is pulled away from the chips and servers) and heat rejection (how that heat is ultimately expelled outside the facility).
Primary cooling includes:
- Air cooling that relies on internal server fans to blow ambient room air directly over heatsinks attached to the CPU/GPU, and is used in traditional, lower-density workloads.
- Air-conditioned air via computer room air conditioner (CRAC) or air handler (CRAH) units to mechanically chill and pressurise the air before it enters the server aisles (typically via raised floors or hot/cold aisle containment). This is often found in standard enterprise datacentres.
- Liquid cooling, which uses a dielectric fluid or water-glycol mix that has a much higher thermal capacity than air. Sub-categories include direct-to-chip (pipes liquid to a cold plate on the processor) and immersion cooling (where servers are literally submerged in a bath of dielectric fluid). This is absolutely required by modern high-density AI clusters (GPUs) where air is physically no longer capable of handling the thermal load.
Facility-level heat rejection comes into play once heat is captured from servers (by air or liquid) and must be dumped outside.
Heat rejection methods include:
- Dry cooling, in which heat is transferred to a closed loop of liquid, which passes through outdoor radiators (dry coolers) where industrial fans blow ambient air across the coils to reject heat. There is zero water consumption (closed loop) as long as outdoor air temperatures are not extreme.
- Hybrid chiller-and-dry-cooler is dry cooling with the capability to add mechanical chilling, which uses compressors and refrigerant cycles. When ambient temperatures spike too high for the dry coolers to maintain safe operating thresholds on their own, valves automatically route the loop through the chillers to artificially sub-cool the liquid.
- Evaporative cooling (open-loop /cooling towers), in which warm water is sprayed downward against an upward flow of air in a cooling tower. The resulting evaporation strips heat away. This creates high levels of water consumption (~1.85 litres per kWh) but is very effective even in hot weather.
- Adiabatic cooling is a hybrid. On normal days, it acts like dry cooling (using just air and fans). On hot peak days, a fine mist of water is sprayed into the incoming air stream to pre-cool it via evaporation before it hits the dry coils. This results in intermittent or low water consumption as it is only used when ambient air temperatures spike.
How prevalent?
While legacy datacentres (70% to 80%) still rely on traditional air-conditioning and water-guzzling evaporative cooling towers for lower-density racks, modern hyperscalers have shifted toward dry cooling to cut water use. Meanwhile, extreme-density AI workloads (40kW to 100kW+ per rack) are making a pivot to direct-to-chip liquid cooling.
Newer datacentres typically favour closed-loop cooling, where cooling fluid re-circulates and is not consumed, confirms Uptime Institute analyst Jacqueline Davis. Water UK says closed-loop systems consume an average 0.1 litres per kWh.
Meanwhile, Davis says adiabatic cooling in mixed or cooler climate regions, such as the Nordics, achieves WUEs of 0.3-0.4.
An Uptime Institute survey found that 34% of datacentres use air-cooled chillers, followed by direct expansion (DX), evaporative towers and dry coolers. DX uses a refrigerant loop to absorb heat from the air and release it outside.
Elsewhere, in drier, hotter climates, you need water for humidification; relative humidity must be at least 20% to prevent electrostatic shock and associated equipment damage, warns Davis. That’s because if air becomes too dry (dropping below that 20% mark), it loses its ability to dissipate static electricity.
Air cooling when possible
Binkley confirms Digital Realty’s standard global design has used air cooling since “at least 2013”. Meanwhile, 25% of its portfolio uses evaporative or adiabatic cooling “of some sort”. Its 2025 reporting calculated a 0.59 WUE (15.7% better than 2024) and 45% recycled water use.
Digital Realty added 34% more sites by 2025, but total water use rose only 3% since 2023. Three Singapore facilities use 100% non-potable water; 15 Virginia sites use 80% recycled water, itself potentially then reused or retreated. Half of Digital Realty’s water-cooled sites use mainly recycled water – typically 90% to 98% of the total water consumed by that site. Its Madrid, Los Angeles, Phoenix and South Africa datacentres scored worst in terms of water efficiency and environmental stress.
“But even our most water-intensive site barely hits the low end of the range of predominant estimates. Even when it’s a fully water-cooled legacy system,” Binkley says. “If you’re in Norway, you probably have better efficiency, reducing water consumption. The more efficient you are, the less heat you have to reject. We work to minimise water use across the portfolio.”
Uptime Institute’s Davis says that by upgrading and packing workloads onto fewer, modern, high-density servers – slashing the total amount of physical hardware by half – an enterprise can drop its direct energy and cooling requirements by well over half.
She says that if done properly, strategies such as DX, air-cooled chillers with integrated economisers [to save water], pumped refrigerant systems, and northern-climate or adiabatic-assisted cooling can consume almost zero water.
“Closed loop using no evaporative cooling is a close second, using little water except for top-up for chemistry maintenance, equipment service and so on,” she says. “Many operators are installing dry cooling or adiabatic external closed-loop chiller systems with economisers. Typically, these systems only need water during the 200 to 300 hottest hours of the year.”
Local circumstances can block change
Digital Realty’s Binkley notes that while modern cooling towers and chillers are more efficient and operators are trying to reduce water use overall, often the infrastructure isn’t there for recycled water.
“We can’t build a whole recycled water system ourselves,” Binkley says. “And cooling replacement is expensive and has carbon footprint. So, a site that’s evaporative cooled today might stay that way, barring major regulatory change or similar.”
