Does AI Waste Water If It Returns Through the Water Cycle?

AI data centres can consume water locally even though evaporation does not destroy it. The US Department of Energy explains that cooling towers lose water through evaporation and discharge some water as blowdown, while make-up water replaces those losses. That is why “does AI waste water?” is really a question about local availability, not whether water vanishes from the planet.

Claims that an AI prompt “uses” a particular amount of water sound simple, but the underlying accounting is not. Results vary according to the model, hardware, location, weather, electricity supply, cooling system and which stages of the supply chain are counted.

The useful question is therefore not whether AI permanently removes water from the planet. It is where, when and how water is withdrawn or consumed to provide AI services.

Water withdrawal and water consumption are different

Water withdrawal means taking water from a river, reservoir, aquifer or public supply. Some or most of it may later be returned.

Water consumption generally refers to water that is not returned promptly to the same local source, often because it has evaporated. This distinction helps explain why two data centres can withdraw similar volumes but have very different effects on local availability.

How does AI use and consume water?

1. AI computation produces heat. Training and inference run on servers whose processors and supporting equipment use electricity and release heat.
2. A cooling system moves that heat. Air, water or another coolant carries heat away from the chips and server racks.
3. The facility rejects the heat. Some sites use predominantly dry cooling; others evaporate water because evaporation can remove heat efficiently.
4. Consumed water must be replaced. Evaporative systems need make-up water, while electricity generation and semiconductor manufacturing can create additional off-site water demand.

Water pathway	What happens	Is the water consumed locally?
Closed coolant loop	Coolant circulates between servers and a heat exchanger.	Usually little or no routine evaporation inside that loop, although the wider facility can still use water.
Evaporative cooling	A cooling tower releases some water as vapour while removing heat.	Yes. Make-up water replaces the evaporated volume.
Blowdown or discharge	A controlled stream leaves a cooling system to manage dissolved minerals.	It may be treated, reused or discharged under local rules; this is different from evaporation.
Electricity and chip production	Water can be used away from the data centre to generate power and manufacture semiconductors.	It depends on the energy source, factory, location and accounting boundary.

Does water used by AI go back into the water cycle?

In cooling towers, evaporated water leaves the immediate cooling system as water vapour. Neither the DOE cooling-tower guidance nor the LBNL report says that evaporated water returns to the same catchment when it is needed.

The practical issue is therefore local and time-sensitive: water can be unavailable to the supplying system even though it has not disappeared from the planet.

How much water does an AI prompt use?

There is no universal figure.

A useful estimate must explain what is included, where the computation runs and whether it counts only on-site cooling or a wider lifecycle footprint.

What official data shows about data-centre water use

The numbers are material at data-centre scale, but they are not per-prompt figures. A 2024 US Department of Energy report, prepared by Lawrence Berkeley National Laboratory, estimated 66 billion litres of direct water consumption across US data centres in 2023. It projected that hyperscale facilities could directly consume 60-124 billion litres in 2028, depending on growth and cooling choices.

This does not mean AI alone consumed 66 billion litres. The estimate covers whole data centres, including non-AI workloads. Dividing it by an estimated number of prompts would create false precision because the model, facility, cooling system, weather and electricity mix are not known.

The same LBNL report estimated an indirect water footprint of nearly 800 billion litres in 2023 from the electricity used by US data centres.

Why WUE cannot tell the whole story

Water usage effectiveness, or WUE, describes facility water use relative to computing energy. It can help operators compare performance, but a single number cannot show whether water was taken from a stressed catchment, whether recycled water was available or what happened in the supply chain.

A facility can improve its operational WUE while still expanding total water demand because it is serving more workloads.

Microsoft reports that its FY25 global WUE for owned, controlled and fully operational data centres was 0.27 litres per kilowatt-hour of IT energy.

Questions businesses should ask AI suppliers

Most organisations will not receive a reliable per-prompt water figure for every AI product. Procurement teams can still ask questions that reveal whether a supplier understands and manages the issue:

• Does the reported figure cover withdrawal, consumption or both?
• Does it include only on-site cooling, or also electricity and hardware manufacturing?
• Which locations, models, dates and workload types does the estimate represent?
• What cooling system is used, and does it rely on evaporation?
• Is potable, non-potable or recycled water used?
• How does the provider assess seasonal and catchment-level water stress?
• Are replenishment claims separated clearly from operational consumption?
• Can the supplier report both efficiency and total water consumption as demand grows?

Water should also be considered alongside cost, carbon, security, privacy and service quality.

AI water use FAQ

Does evaporated water from AI come back as rain?

Evaporation sends cooling water into the atmosphere, rather than destroying it. The cited DOE and LBNL sources do not give a timetable or location for its return as rain.

Is there one water-use figure for every AI prompt?

There is no universal figure.

Does closed-loop cooling eliminate AI water use?

Not necessarily. Other facility operations, electricity generation and semiconductor manufacturing may still have water impacts.

What is the most useful way to judge an AI service?

Look for transparent boundaries, location-specific data and total consumption alongside efficiency metrics. There is no honest universal number for AI water use, but better disclosure can show whether a provider is reducing pressure where it actually matters.

Related reading

• Should we tax AI data centres?
• How to keep your critical thinking sharp when using AI