How Much Water Does AI Use? Per-Prompt Estimates Explained

Updated September 19, 2026 with current AI-inference estimates, EU reporting requirements and clearer water-accounting boundaries.

There is no single defensible water figure for an AI prompt. Google estimates that the median Gemini Apps text prompt in May 2025 consumed 0.26 mL of water for data-center cooling. A separate academic model estimated roughly 10-50 mL for a medium-length GPT-3 response when both on-site cooling water and off-site water consumed in electricity generation were included. Those figures describe different models, workloads, locations and accounting boundaries, so they should not be presented as competing measurements of the same thing.

Scope: those per-prompt estimates concern model inference: generating answers from already trained models. They do not describe model training, fine-tuning or a continuously utilized AI/HPC facility. Facility planning must use the actual mix of training and inference, hardware, utilization and load profile rather than multiplying a prompt estimate by an assumed request count.

For a data-center buyer, the more useful question is not whether every prompt consumes a fixed number of milliliters. It is where water enters the system, how much the facility needs during normal and peak conditions, which water source it uses, and whether the cooling design fits the local climate and watershed.

How much water does an AI prompt use?

Published estimates vary by more than an order of magnitude because they answer different questions.

An operator-reported fleet estimate: 0.26 mL for a median Gemini prompt

Google reported that a median Gemini Apps text prompt in May 2025 used 0.24 Wh of energy and 0.26 mL of water. The energy calculation uses internal telemetry across Google's production serving fleet and includes accelerator power, host CPU and memory, idle capacity and data-center energy overhead. The water figure is not a meter reading for an individual prompt: Google derives it from per-prompt energy and the prior year's average freshwater WUE for the data centers serving the models. It covers water consumed to cool machines and associated data-center infrastructure. This is a point-in-time fleet estimate, not a figure for every prompt or future model. See Google's technical paper on AI inference.

A modeled operational estimate: about 10-50 mL for a medium-length GPT-3 response

Li, Yang, Islam and Ren modeled the operational water footprint of GPT-3 using public energy, Power Usage Effectiveness (PUE), WUE and electricity-generation data. PUE compares total data-center energy with the energy used by IT equipment. For a request with up to roughly 800 input words and 150-300 output words, their location-specific estimates ranged from about 7 mL to 48 mL when on-site cooling and off-site electricity-generation water were combined. That is the source of the widely repeated claim that a 500 mL bottle corresponds to roughly 10-50 responses.

This was a model of GPT-3, not direct telemetry from a current ChatGPT service. The paper's assumptions, including energy per request and regional grid-water intensity, are part of the result. Read the current version of Making AI Less Thirsty.

What a defensible estimate must disclose

A per-prompt figure is only meaningful when it identifies:

• the model and type of task;
• input and output length, batching and server utilization;
• the location and time period;
• whether it counts on-site cooling water, electricity-generation water, hardware manufacturing, or a combination;
• whether the figure is measured, operator-estimated or modeled.

Without those boundaries, a precise-looking number is usually not comparable with another precise-looking number.

Why does AI use water?

Nearly all electricity used by IT equipment becomes heat. That heat must move through a complete chain: from chips to a server or rack cooling system, from that system to a facility loop, and from the facility to the outdoor environment.

Water can appear in three different parts of the footprint:

• On-site operational water. Cooling towers and adiabatic systems may evaporate water while rejecting heat outdoors. Blowdown and water treatment can add to water withdrawal.
• Off-site operational water. Some electricity-generation technologies withdraw or consume water. This impact depends on the power source and regional grid mix.
• Embodied water. Semiconductor fabrication and equipment manufacturing use water. This belongs to a lifecycle assessment, not the facility's annual cooling-water figure.

Water withdrawal and water consumption are also different. Withdrawal is water taken from a source; consumption is the portion not returned to the immediate water environment, often because it evaporates. A project may have a high withdrawal rate but return much of that water, or a lower withdrawal rate with a larger consumed share. Both matter, but for different reasons.

Direct and indirect water must be reported separately

Lawrence Berkeley National Laboratory estimated that U.S. data centers consumed about 66 billion liters of water on site in 2023. The same report estimated an indirect water footprint of nearly 800 billion liters associated with electricity use, based on regional grid mixes. Its national average was 4.52 L/kWh for indirect water consumption, compared with an average site WUE just above 0.36 L/kWh in 2023.

These are U.S. modeled averages, not universal conversion factors. Electricity-water intensity varies sharply by generation technology and location, and the study did not incorporate facility-specific power purchase agreements or behind-the-meter generation. Reducing on-site water therefore does not automatically eliminate the wider water footprint. See the 2024 U.S. Data Center Energy Usage Report.

What WUE measures - and what it misses

Water Usage Effectiveness relates annual site water use to annual IT energy use. A common expression is:

WUE = annual site water use / annual IT equipment energy

It is usually expressed in liters per kWh, or equivalently cubic metres per MWh. Lower values mean less site water per unit of IT energy within the stated measurement boundary.

WUE is useful, but it is not a complete water-risk score. Before comparing two figures, check:

• whether the numerator is water input, withdrawal or consumption;
• whether potable and non-potable sources are separated;
• which equipment and facility functions sit inside the boundary;
• the annual IT load, climate and operating conditions;
• whether the number is measured, modeled or a design target;
• the site's local water stress and peak make-up-water demand.

For one current reference point, the LBNL study estimated a U.S. average site WUE just above 0.36 L/kWh in 2023 and modeled 0.45-0.48 L/kWh after 2023. That does not make either range a universal benchmark or a target for a particular facility. Compare WUE together with absolute annual volume, peak water-delivery rate, source type and local seasonal scarcity. A lower WUE at a water-stressed site is not automatically preferable to a higher WUE supplied by reclaimed water in a water-abundant location.

Does liquid cooling reduce water consumption?

Liquid cooling and water consumption are related, but they are not the same decision.

Direct-to-chip cooling can move a large share of rack heat into a liquid loop efficiently. A coolant distribution unit then transfers that heat to a facility system. The IT loop can be sealed and recirculating, yet the outdoor plant can still reject heat through an evaporative cooling tower. In that case, the servers are liquid-cooled but the facility still consumes water.

Conversely, a facility can use a closed loop with dry coolers and avoid routine evaporation for heat rejection. The tradeoff is that dry systems may require more fan or compressor energy during hot conditions, particularly when the required fluid temperature is low. The local climate, equipment temperature limits and power-water priorities determine the result.

The U.S. Department of Energy's cooling-water guidance illustrates the full chain: rack cooling, facility loops and the final outdoor heat-rejection stage must be assessed together.

Cooling options for water-constrained AI sites

Dry heat rejection

Dry coolers reject heat to outdoor air without routine evaporative water use. They are strongest when the facility-water temperature remains sufficiently above the outdoor-air temperature for efficient heat transfer. The hottest outdoor conditions used as the design basis can increase fan power, require larger heat exchangers, or trigger mechanical assistance.

Adiabatic assistance

An adiabatic system uses water only when ambient conditions require extra heat-rejection capacity. This can reduce annual water demand compared with continuous evaporative cooling, but the project still needs a peak water rate, storage and refill strategy, water-quality requirements and control sequence.

Evaporative cooling towers

Cooling towers can reject heat efficiently, especially in hot conditions, but require make-up water for evaporation and blowdown. Water quality and the number of times dissolved minerals are concentrated before water is discharged affect the withdrawal rate. The lowest-energy option is not automatically the lowest-water option.

Free cooling and economizers

When outdoor conditions permit, economizers can reduce compressor operation. The number of useful hours depends on the site's weather, equipment limits, supply temperatures, load and controls. A label such as "northern Europe" or "Middle East" is not a substitute for an hourly climate simulation.

How to choose between dry, hybrid and evaporative cooling

The right architecture depends on which constraint the project cannot violate. Use this as a screening guide, not as a substitute for site modelling:

• If minimum routine water consumption is the priority: start with dry heat rejection. Check hot-day capacity, electrical demand, equipment size and any temperature-related derating.
• If the site needs low annual water use without designing the whole system around the hottest hours: assess dry cooling with adiabatic assistance. Define exactly when water assistance starts and how the site operates if stored water is unavailable.
• If hot-day energy efficiency or a constrained equipment footprint dominates: evaporative heat rejection may remain an option, but only after checking water entitlement, drought restrictions, make-up rate, treatment and discharge.
• If the workload supports warm-water liquid cooling: use the higher return temperature to improve dry-cooling potential, then assess the residual air-cooled load and final heat-rejection plant separately.

A site decision should also test legal water entitlement, connection capacity, seasonal source reliability, potable and reclaimed-water availability, wastewater and blowdown permits, community demand, water-price exposure, local watershed stress and the operating plan for a water-supply interruption. Annual liters alone can hide an unacceptable peak demand or drought risk.

What the EU requires data centers to report

Commission Delegated Regulation (EU) 2024/1364 applies to data centers with installed information technology power demand of at least 500 kW. The regulation defines that demand as the sum of the nominal demand of the networks, servers and storage equipment installed in the computer-room floor area. Reporting operators must communicate annual information and key performance indicators to the European database. The regulation includes total water input, potable water input and WUE, alongside energy, cooling, renewable-energy and heat-reuse data.

The first deadline was September 15, 2024; annual reporting moved to May 15 from 2025 onward. Individual data-center information is treated as confidential in the database, while aggregated national and Union-level data are public. The European Commission is still developing the rating scheme and work toward minimum performance standards, so operators should follow the current Commission guidance rather than an outdated expected launch date. See the official regulation and the Commission's data-center energy-performance page.

Questions to put into an AI data-center cooling brief

A credible proposal should answer these questions before quoting annual water performance:

1. What hardware is being cooled? Define rack loads, liquid and residual-air shares, OEM temperature limits and expected utilization.
2. Where does the heat go? Trace the path from chips and rack air to the final outdoor heat-rejection equipment.
3. What is the water boundary? Separate initial system fill, cooling consumption, make-up water, blowdown, storage and off-site electricity-related water.
4. What happens in the hottest design conditions? State cooling capacity, peak water demand, storage duration, refill logistics and operating mode during a water interruption.
5. Which water can be used? Identify potable, reclaimed or treated sources, required chemistry, treatment responsibility and the discharge route for blowdown.
6. What are the energy consequences? Model annual PUE and absolute energy with the same climate file, load profile and redundancy state used for water analysis.
7. How will performance be measured? Define meter locations, reporting boundary and whether the quoted result is a target, simulation or measured outcome.

What a cooling proposal should commit to

Before comparing bids, require each supplier to state the same operating cases and boundaries. At minimum, the proposal should identify:

• predicted annual water consumption, peak hourly and daily make-up demand, and WUE at defined IT-load profiles and climate conditions;
• PUE and WUE for the same load, climate, redundancy state and measurement boundary;
• available cooling capacity and redundancy in dry mode and any water-assisted mode;
• the operating sequence, capacity and energy consequence when cooling water is unavailable;
• water-quality limits, treatment chemicals, blowdown quality and responsibility for tanks, refill, treatment and permits;
• metering points, commissioning tests, seasonal performance checks and data ownership;
• which figures are simulations, design targets or contractually guaranteed values, including tolerances and remedies.

A numerical comparison is only useful when it uses the project's rack schedule, site coordinates, climate file, availability target, water constraints and operating profile. A generic worked example can illustrate a method, but it should not be presented as the expected performance of a real site.

How ModulEdge approaches water-aware cooling

ModulEdge selects the thermal architecture against the actual hardware, rack density, climate and site constraints. A closed loop is only one part of that design; the outdoor heat-rejection method determines whether routine evaporation is required. The configurations below are manufacturer reference architectures described in current product brochures, not universal performance guarantees.

The MDC-500 reference architecture keeps the IT equipment air-cooled. Active rear-door heat exchangers transfer rack exhaust heat into a closed chilled-water loop, and three modular air-cooled chillers provide external heat rejection with integrated free cooling. The reference design does not use an evaporative cooling tower, but the product brochure does not claim a universal PUE or WUE.

The MDC-2000 reference uses warm-water direct liquid cooling for the dominant AI heat load, with a separate air-side path for heat that remains in room air. Its external plant is dry-first, with tank-fed adiabatic assistance when ambient conditions require it. The reference strategy can use stored water for intermittent adiabatic operation rather than continuous mains make-up. Whether a project can omit a permanent cooling-water connection depends on modeled annual and peak consumption, storage autonomy, refill availability, water quality and the required operating mode during a water-supply interruption. Fire protection, sanitation and maintenance-water requirements must be assessed separately.

Frequently asked questions

How much water does one AI prompt use?

There is no universal figure. Google estimated 0.26 mL of cooling water for a median Gemini Apps text prompt in May 2025. A separate academic model estimated roughly 10-50 mL for a medium-length GPT-3 response when on-site cooling and electricity-generation water were combined. The model, workload, location and accounting boundary must accompany the number.

Why do AI water-use estimates differ so much?

They may cover different models, response lengths, hardware, utilization, climates and electricity mixes. Some count only on-site cooling; others add water consumed in electricity generation or equipment manufacturing. Measured, operator-estimated and modeled figures should not be treated as equivalent.

Does ChatGPT use 500 mL of water every few prompts?

The widely repeated bottle comparison comes from a modeled GPT-3 scenario, not a direct measurement of current ChatGPT. The study estimated that 500 mL could correspond to roughly 10-50 medium-length responses depending on deployment location and time, including both on-site and off-site operational water.

Does direct liquid cooling eliminate data-center water use?

No. Direct liquid cooling moves heat efficiently from chips into a liquid loop, but the facility must still reject that heat outdoors. A closed IT loop can feed an evaporative tower, a dry cooler, a chiller or a hybrid system. The final heat-rejection design determines routine cooling-water consumption.

What is WUE?

Water Usage Effectiveness is annual site water use divided by annual IT equipment energy, usually expressed as L/kWh or m³/MWh. The reported boundary and water category matter. WUE does not include the full electricity-generation or manufacturing water footprint unless a separate metric explicitly adds them.

Can an AI data center operate without evaporating water for cooling?

Yes. A design can use closed loops and dry heat rejection so routine cooling does not depend on evaporation. That does not mean the entire facility has zero water use, and it can increase energy use or equipment size during hot conditions. Microsoft, for example, describes a new zero-water-evaporation cooling design while noting that water remains necessary for administrative uses.

Which EU data centers must report water performance?

Under Commission Delegated Regulation (EU) 2024/1364, operators of data centers with installed IT power demand of at least 500 kW must report annual indicators including total water input, potable water input and WUE.