Does AI Waste Water? Cooling, the Water Cycle and Real Usage

Updated 25 July 2026: AI does not destroy water, but data centres can consume it locally when cooling systems evaporate water. That matters because the water may leave a catchment when local supplies are already under pressure. The complete footprint can also include water used to generate electricity and manufacture semiconductors.

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 It is where, when and how water is withdrawn or consumed to provide AI services.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 AI data centre cooling uses water

AI servers generate heat during inference, when a trained model answers a request, and during training. Operators must remove that heat reliably to protect equipment and maintain performance.

Cooling designs include air cooling, evaporative systems and liquid cooling that moves heat away from chips. The right choice depends on hardware density, climate, energy efficiency, local water conditions and facility design.

Closed-loop direct-to-chip cooling

In a closed-loop system, coolant circulates repeatedly instead of being continuously discarded. Direct-to-chip cooling carries that coolant close to heat-producing components, which can handle dense AI equipment more effectively than relying on room air alone.

Microsoft says its cited direct-to-chip design removes heat with zero water evaporation in that cooling loop. It also reports a 25% reduction in water-use intensity by 2025 against a 2022 baseline, measured as water withdrawals per megawatt.

That does not mean every closed-loop facility has zero water use. Whole-facility cooling, electricity generation, maintenance and semiconductor manufacturing remain separate questions. The technology reduces one specific source of consumption rather than making the entire AI service water-free.

Why the water cycle does not erase local impacts

Evaporated water remains part of the global water cycle and may eventually return as precipitation. However, it may return later, in a different place or to a different catchment.

This is why “the water comes back” is not a complete answer. A litre consumed during a dry period can still intensify competition between homes, agriculture, industry and ecosystems, even if that water later falls as rain elsewhere.

Corporate replenishment projects can improve watersheds and water stewardship. For example, Google reports that its 2025 projects replenished about 7.7 billion gallons, equivalent to around 78% of freshwater consumption. That figure does not prove that consumed water returned immediately to the same catchments from which it was taken.

How much water does one AI prompt use?

There is no universal figure. A Lawrence Berkeley National Laboratory analysis found that workload-level water use can vary by more than 10,000-fold. Drivers include server efficiency, utilisation, cooling design, infrastructure efficiency, climate and the water intensity of the electricity grid.

Published estimate	Reported water consumption	Boundary and limitation
Google Gemini Apps text prompt	0.26 mL for the median prompt in May 2025	A provider-reported estimate, not independently verified, that derives water consumption using Google’s 2024 fleetwide average WUE. It also reported 0.24 Wh and included accelerators, CPUs, idle capacity and data-centre overhead.
Mistral Le Chat response	45 mL for a 400-token response	A lifecycle estimate that excludes the user’s device. Mistral described it as a first approximation.

These numbers are not like-for-like. They concern different providers, models, dates, response boundaries and methodologies. Mistral’s figure comes from a lifecycle analysis, while Google’s Gemini estimate measures a particular production workload.

Quoting either number as the water cost of all AI prompts would create false precision. 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.

UK water supply and data centre planning

The issue is increasingly relevant to UK infrastructure policy. The Environment Agency’s National Framework summary says England may require up to five billion additional litres per day for public water supply by 2055.

It identifies computer data-centre cooling as an emerging pressure and says siting, cooling technology and recycled-water options should be considered. In other words, national totals are not enough. A project’s effect depends heavily on the condition of the specific catchment in which it is built.

This is one reason proposed facilities face questions about utility capacity and local benefit. My guide to AI data centre water, power and planning backlash examines that wider infrastructure tension.

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. WUE may also exclude water associated with electricity generation and chip manufacturing, both of which the European Environment Agency identifies as parts of AI’s water footprint.

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. The infrastructure behind apparently cheap AI is part of the broader reason AI compute is not yet simply cheaper than people.

AI water use FAQ

Does AI permanently destroy water?

No. Water remains within the global water cycle. However, evaporation can make it unavailable to the original catchment at the place and time it is needed, which is why local consumption still matters.

Is every AI prompt responsible for 45 mL of water?

No. The 45 mL figure is Mistral’s first lifecycle approximation for a specific 400-token Le Chat response. Google reported 0.26 mL for its median Gemini Apps text prompt under a different methodology. They should not be treated as equivalent measurements.

Does closed-loop cooling eliminate AI water use?

Not necessarily. Microsoft’s cited design has zero water evaporation within its direct-to-chip cooling loop. 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.