Thirsty Machines: The AI Data Centre Boom
- Richard
- 2 hours ago
- 5 min read
The world is racing to build new data centres to power artificial intelligence, and with them comes a huge, often hidden thirst for energy and water. While these facilities enable breakthroughs in everything from medical research to language models, they are also turning into major industrial water users, especially in regions already facing drought and heat stress.
The AI-fuelled data centre boom
AI models demand vast computing power, particularly during training and large-scale inference, and that power has to live somewhere – in sprawling clusters of servers inside data centres. Global data centre electricity consumption reached roughly 460–470 terawatt-hours in 2022 and is projected in some scenarios to more than double to around 1,000–1,100 terawatt-hours by the mid‑2020s, putting the sector on par with the world’s largest power‑consuming countries.
Much of this growth comes from hyperscale AI facilities that pack tens or hundreds of thousands of specialised chips into dense racks. These chips run hot, so operators must move enormous amounts of heat away from equipment, every second of every day, making cooling the main driver of both electricity and water use.
How much water are we talking about?
The numbers are eye‑catching once you convert them into everyday terms.
- A typical 100‑megawatt data centre used for AI workloads in the US consumes around 2 million litres of water every day, roughly equivalent to the daily water use of about 6,500 households.
- On an annual basis, the World Economic Forum has estimated that each megawatt of data centre capacity can require around 25.5 million litres of water for cooling, so a 100‑megawatt site can approach 2.5 billion litres per year.
- Globally, data centres already consume on the order of 560 billion litres of water annually, and that figure could rise to about 1,200 billion litres by 2030 as AI‑heavy sites roll out.
- In one widely cited example, a single large data centre in Iowa used about 1 billion gallons (around 3.8 billion litres) of water in 2024, enough to cover all residential water needs in the state for several days.
These volumes include both “direct” water used on site and “indirect” water embedded in electricity generation. Where power comes from water‑intensive thermal plants, the water footprint of the electricity feeding the data centre can rival or exceed what the cooling towers consume.
Why data centres need so much fresh water
Most large data centres maintain tight temperature and humidity ranges to keep servers operating reliably and to avoid corrosion, condensation, or component failure. To achieve this at scale, operators rely on a mix of cooling technologies, many of which use freshwater directly or indirectly.
Key reasons so much fresh water is used include:
- Evaporative cooling towers
Many facilities use evaporative cooling, where water is sprayed or circulated so that a portion evaporates, carrying heat away and cooling the remaining water. This approach is energy‑efficient compared with purely mechanical chillers but requires continuous “make‑up” water to replace what evaporates and what is discharged to prevent mineral build‑up.
- Strict water quality requirements
Cooling systems need relatively clean water to avoid scale, corrosion and biological growth in pipes, heat exchangers and towers. As a result, operators often draw treated drinking‑quality water from municipal systems rather than using untreated surface water, which increases pressure on local potable supplies.
- Redundant, always‑on operation
Data centres are designed for near‑continuous uptime, with cooling systems running 24/7, often with redundancy. Even at partial IT load, cooling systems circulate and treat water constantly, so consumption does not fall in proportion to compute usage.
- Indirect water in electricity
Conventional power plants, particularly coal, gas and nuclear, withdraw large amounts of water for steam generation and cooling. When a data centre draws hundreds of megawatts from such grids, the off‑site water needed to produce that electricity becomes a big part of its overall water footprint.
In hotter, drier climates, the trade‑off between water and energy becomes especially stark. To save water, operators can lean more on energy‑intensive mechanical chillers; to save electricity, they can favour evaporative and adiabatic systems that consume more water. AI‑heavy sites push both dimensions hard, because high‑density racks can demand far more cooling per square metre than traditional enterprise data halls.
The geography of a thirsty cloud
Location choices for new AI data centres increasingly intersect with water stress. Analyses of new US builds since 2022 show that roughly two‑thirds of these sites are being sited in regions already experiencing high levels of water stress, drawn by cheap land, favourable tax regimes and existing power infrastructure.
In places like the American West, Australia and parts of Europe, this can pit industrial cooling needs against agriculture and residential supply. For example, Australia now hosts more than 250 data centres, and experts there have warned against relying on drinking water for cooling as more AI‑oriented facilities are proposed in already dry regions.
Because companies rarely disclose precise water usage by site, communities often struggle to understand or contest the cumulative impact of multiple new facilities. As heatwaves and droughts become more frequent under climate change, the tension between digital growth and water security is likely to intensify.
Promising paths to lower water and energy use
Despite these challenges, there is rapid innovation aimed at reducing both the water and energy footprints of AI data centres.
Some of the more promising directions include:
- Advanced liquid and direct‑to‑chip cooling
Instead of cooling entire rooms with chilled air, direct‑to‑chip and other liquid‑cooling approaches move heat away right at the processor, using coolants that are recirculated in closed loops. These systems can cut water consumption by 20–90 percent in suitable climates and layouts, while also reducing facility‑level power needs by up to around 18 percent compared with conventional air cooling.
- Zero‑water and dry cooling designs
Emerging “zero‑water” facilities pair closed‑loop liquid cooling inside the data hall with dry coolers or air‑cooled heat exchangers outside, virtually eliminating evaporative water losses. Some operators report potential savings of over 125 million litres of water per site each year and water usage effectiveness (WUE) approaching zero in favourable conditions.
- Circular and non‑potable water use
New projects experiment with using reclaimed wastewater, industrial effluent, or seawater where available, treating it for cooling rather than tapping drinking‑water networks. Circular water strategies also look at cascading heat to district heating networks and re‑using process water multiple times before discharge, reducing both withdrawals and thermal pollution.
- Cleaner power and smarter siting
Shifting data centres onto grids with more solar and wind cuts the indirect water footprint of electricity generation, since these technologies use far less water than fossil fuel plants. Pairing new builds with robust water‑availability assessments and community consultation can steer AI infrastructure toward regions that can better support its resource demands.
Conclusion
AI itself may play a role, as operators deploy machine‑learning systems to optimise cooling, predict failures and balance water‑versus‑energy trade‑offs hour by hour. If these technical and planning advances spread quickly enough, the next wave of AI data centres could become considerably less thirsty than the first – but only if water stops being an afterthought and is treated as a core design and policy constraint from the outset.
🤖
Title image credit: Unsplash: https://unsplash.com/@geoffreymoffett
Comments