How Much Water Does AI Use?

The Hidden Physical Footprint of the Digital Cloud

When we interact with artificial intelligence—whether by generating an image, writing an email, or translating a document—the transaction feels entirely virtual. However, behind every algorithmic response lies a massive, highly physical infrastructure. This infrastructure consists of global networks of data centers housing hundreds of thousands of high-performance servers equipped with specialized silicon chips, such as Graphics Processing Units (GPUs) and Tensor Processing Units (TPUs).

These processors run continuously, drawing immense amounts of electricity and generating substantial heat. To prevent hardware from overheating, failing, or degrading, data centers must employ heavy-duty cooling systems. Yes, artificial intelligence uses a significant amount of water. In fact, water has emerged as one of the most critical, yet historically underreported, environmental costs of the modern AI boom.

To understand the scale of this resource consumption, we must look at the thermodynamic realities of computing, how water is utilized within data centers, and the indirect water costs embedded in the electrical grid that powers these systems.

Why and How AI Systems Consume Water

To understand why AI requires water, it is helpful to divide its water footprint into two distinct categories: direct consumption (onsite at the data center) and indirect consumption (offsite at the power plants that generate the data center's electricity).

                          ┌──────────────────────────────────┐
                          │     Total AI Water Footprint     │
                          └────────────────┬─────────────────┘
                                           │
                ┌──────────────────────────┴──────────────────────────┐
                ▼                                                     ▼
┌──────────────────────────────┐                      ┌──────────────────────────────┐
│      Direct Consumption      │                      │     Indirect Consumption     │
│     (Onsite Data Center)     │                      │    (Offsite Power Plants)    │
└──────────────┬───────────────┘                      └──────────────┬───────────────┘
               │                                                     │
 ┌─────────────┴─────────────┐                         ┌─────────────┴─────────────┐
 ▼                           ▼                         ▼                           ▼
Evaporative                 Chilled                   Hydroelectric               Steam Condensation
Cooling                     Water Loop                Evaporation                 in Thermoelectric
(Water lost to              (Closed-loop,             (Reservoir                  (Coal, Gas, Nuclear
atmosphere)                 minimal loss)             losses)                     cooling loops)

1. Direct Water Consumption (Onsite Cooling)

Data centers house thousands of densely packed server racks. As these servers execute complex mathematical calculations for AI training and inference, they convert electrical energy into thermal energy. If this heat is not rapidly dissipated, the processors will throttle their performance or suffer catastrophic thermal failure.

While there are various ways to cool a data center, the most energy-efficient methods historically rely on water. The primary onsite cooling mechanisms include:

• Evaporative Cooling Towers: This is the most common and water-intensive method. Warm water from the data center’s heat exchangers is pumped to a cooling tower, where it is exposed to outdoor air flow. A portion of the water evaporates, absorbing heat from the remaining water and carrying it away into the atmosphere. This cooled water is then routed back into the data center to absorb more heat. The evaporated water is "consumed" because it is lost to the local watershed and released as vapor.
• Chilled Water Systems: These systems use chillers (refrigeration units) to cool water, which is then circulated through heat exchangers inside the server rooms. While chillers can operate in a closed loop (reclaiming and reusing the same water), they require massive amounts of electricity to run the refrigeration compressors, which dramatically increases the data center's indirect water and carbon footprints.
• Adiabatic Cooling: This system combines dry cooling (using ambient air and fans) with evaporative cooling. When the outdoor temperature is low, the system uses air alone. When the outdoor temperature rises above a certain threshold, water is sprayed into the incoming air stream to cool it via evaporation before it hits the heat exchangers. This uses significantly less water than traditional cooling towers but still requires liquid volume during warm periods.

2. Indirect Water Consumption (Offsite Power Generation)

Data centers consume vast quantities of electricity. This electricity must be generated by utility companies, which themselves rely heavily on water.

Thermoelectric power plants (which burn coal, natural gas, or biomass, or utilize nuclear energy) boil water to create high-pressure steam that spins electricity-generating turbines. After passing through the turbines, this steam must be cooled back into water so it can be pump-recycled through the boiler. This cooling loop consumes billions of gallons of water daily, primarily through evaporation.

Additionally, hydroelectric power plants experience significant water loss through reservoir evaporation. Therefore, every kilowatt-hour (kWh) of electricity a data center draws from the local grid carries an "embedded" water cost, which varies widely depending on the local energy mix.

Quantifying the Water Footprint: From Queries to Training

Because tech companies historically guarded their resource-use metrics closely, researchers have had to reconstruct AI's water footprint using public utility reports, corporate environmental disclosures, and thermodynamic modeling. Landmark studies—such as those led by researchers at the University of California, Riverside—have shed light on both the micro-level (individual prompts) and macro-level (training runs) water costs of large language models (LLMs).

The Micro Level: What Does a Single Query Cost?

Every time a user inputs a prompt into an AI tool like ChatGPT, Claude, or Gemini, the query is routed to a physical GPU clustered in a data center. The computational work required to generate a multi-paragraph response, write a block of code, or synthesize an image translates directly to milliliters of water.

• The Baseline Estimate: Research indicates that a standard session of 10 to 50 conversational exchanges with a state-of-the-art LLM (like GPT-4) consumes roughly 500 milliliters (approximately 16.9 fluid ounces) of water. This is equivalent to a standard personal water bottle.
• Variable Factors: This is not a static figure. The actual water footprint of a single query depends heavily on:
  • Model Size and Complexity: Larger models with hundreds of billions of parameters require more computation per token generated, escalating the heat output.
  • The Location of the Server: A query routed to a data center in a hot, arid climate (such as Arizona) will consume far more water via onsite evaporative cooling than the same query routed to a facility in a cooler climate (such as Ireland or Sweden).
  • The Time of Day: Queries processed during the heat of the afternoon require more evaporative cooling than those processed during the cool of the night.

The Macro Level: Training Large-Scale Models

Training an AI model is an incredibly intensive process that runs thousands of high-power processors continuously for weeks or months. This is where massive chunks of water are consumed in concentrated bursts.

• GPT-3 Training: Researchers estimate that training OpenAI's GPT-3 model in Microsoft's state-of-the-art US data centers directly consumed roughly 700,000 liters (185,000 gallons) of clean freshwater for cooling. If indirect water consumption from electricity generation is factored in, the total water footprint easily doubles or triples.
• Next-Generation Models (GPT-4, LLaMA-3, Gemini): As models have scaled up by orders of magnitude, their training footprints have grown proportionally. While precise figures for these newer models are proprietary, environmental researchers estimate that training models of this scale requires millions of liters of water, equivalent to the volume needed to fill several Olympic-sized swimming pools or manufacture thousands of electric vehicle batteries.

Direct vs. Indirect Water Use: A Comparative Look

To manage their environmental impact, developers must balance the trade-off between using water directly onsite or indirectly through grid power. The table below outlines how these two paths compare:

The Water-Energy Nexus and Geographic Variations

The relationship between water and energy in data centers is known as the Water-Energy Nexus. In data center design, engineers evaluate two primary efficiency metrics:

1. Power Usage Effectiveness (PUE): The ratio of the total energy entering a data center to the energy delivered to the computing equipment. A PUE of 1.0 is perfect, meaning all energy goes to computing, not cooling or auxiliary systems.
2. Water Usage Effectiveness (WUE): The ratio of water consumed at the facility to the electricity delivered to the computing equipment, measured in liters per kilowatt-hour ($L/kWh$).

There is often a direct, opposing trade-off between these two metrics:

$$\text{To lower PUE (use less electricity)} \longrightarrow \text{Use evaporative cooling} \longrightarrow \text{WUE spikes (uses more water)}$$ $$\text{To lower WUE (use less water)} \longrightarrow \text{Use dry, mechanical chilling} \longrightarrow \text{PUE spikes (uses more electricity)}$$

The Role of Geography and Climate

Because of these trade-offs, the geographic location of an AI data center radically alters its real-world water footprint.

• Hot and Arid Regions (e.g., Arizona, Texas, Utah): In dry climates, evaporative cooling is highly efficient at lowering PUE, but it consumes vast quantities of scarce local freshwater. Because these regions already suffer from water stress, data centers can place a severe burden on local municipal drinking water supplies and agricultural irrigation.
• Cool and Temperate Regions (e.g., Northern Europe, Pacific Northwest): Data centers in these areas can utilize free air cooling (outside air) for the majority of the year. This minimizes both water consumption and mechanical electricity use, making the overall environmental footprint much lower.
• Grid Composition Matters: If a data center is located in a region with a highly green power grid (dominated by wind, solar photovoltaic, and run-of-river hydro), its indirect water consumption drops significantly. Conversely, a data center running on a grid powered by coal or traditional nuclear power plants will have a much higher indirect water footprint.

Environmental and Societal Implications

The rapid, unchecked expansion of AI infrastructure has sparked growing concern among environmental scientists, policymakers, and local communities. Several critical issues have emerged:

Local Watershed Strain

Data centers require millions of gallons of water daily, often drawing directly from municipal potable water lines—the same water used by local residents for drinking, hygiene, and farming. During droughts or seasonal dry spells, this shared reliance can lead to localized water crises, political friction, and legal battles between tech conglomerates and local communities.

Water Quality and Treatment Chemicals

Cooling systems do not just use water; they process it. To prevent algae growth, scale buildup, and pipe corrosion, water in cooling towers is treated with biocides, algaecides, and mineral-inhibiting chemicals.

Eventually, as water evaporates, the concentration of these chemicals and minerals in the remaining water rises. This highly concentrated wastewater—known as blowdown water—must be discharged into local sewer systems or water bodies. If not properly managed, blowdown water can introduce thermal and chemical pollution into local ecosystems.

Transparency and Regulatory Gaps

For years, major technology companies treated water usage data as proprietary business information. While most hyperscale cloud providers (such as Microsoft, Google, Meta, and AWS) now publish annual sustainability reports detailing their total enterprise-wide water consumption, pinpointing the exact water footprint of specific AI products remains difficult. Furthermore, local governments often grant generous water-use permits to attract technology investments, sometimes overriding local conservation guidelines.

Strategies for Mitigating AI’s Water Footprint

As public awareness and regulatory pressures grow, the tech industry is actively exploring and implementing technologies to curtail AI's liquid dependency. These efforts span hardware, software, and infrastructure architecture.

                  ┌─────────────────────────────────────────┐
                  │      AI Water Mitigation Strategies     │
                  └────────────────────┬────────────────────┘
                                       │
         ┌─────────────────────────────┼─────────────────────────────┐
         ▼                             ▼                             ▼
┌───────────────────┐         ┌───────────────────┐         ┌───────────────────┐
│     Hardware      │         │     Software      │         │  Infrastructure   │
│    Innovations    │         │    Optimization   │         │    Sourcing       │
└────────┬──────────┘         └────────┬──────────┘         └────────┬──────────┘
         │                             │                             │
  • Immersion cooling          • Smaller models              • Recycled water
  • Direct-to-chip             • Load-shifting to            • Rainwater capture
    liquid cooling               cooler regions/nights       • Air-cooled designs

1. Advanced Cooling Technologies

• Direct-to-Chip Liquid Cooling: Instead of cooling entire rooms with air or chilled water loops, this technology routes a closed-loop dielectric liquid or water line directly across a cold plate attached to the CPU or GPU. This liquid absorbs heat directly from the chip and is cycled to an external radiator. Because it is a completely closed loop, water loss to evaporation is virtually zero.
• Immersion Cooling: In this cutting-edge approach, entire server blades are submerged in a bath of non-conductive, dielectric fluid. The fluid absorbs heat directly from all components. In single-phase immersion cooling, the heated liquid is pumped to an external heat exchanger and returned. In two-phase immersion, the fluid boils at a low temperature, vaporizes, condenses on a cold plate, and drips back down. This method is incredibly energy-efficient and consumes almost no water onsite.

2. Sourcing Alternative Water

To avoid depleting municipal drinking water supplies, progressive data centers are shifting to non-potable water sources:

• Recycled/Reclaimed Water: Using treated municipal wastewater that is unfit for drinking but perfectly suitable for cooling towers.
• Rainwater Harvesting: Capturing and storing rainfall onsite to supplement cooling systems.
• Industrial Water: Sourcing brackish water, agricultural runoff, or industrial process water.

3. Software-Level Optimizations

Software developers can also play a major role in reducing physical resource usage:

• Spatiotemporal Load Shifting: Moving heavy training workloads dynamically across a global network of data centers to match favorable local environmental conditions. For instance, a tech company can route intensive model-training computations to data centers located in regions experiencing nighttime hours, colder seasonal climates, or periods of excess carbon-free, low-water energy generation.
• Model Efficiency Improvements: Designing smaller, more efficient neural networks (such as distilled models or mixture-of-experts architectures) reduces the sheer number of floating-point operations required per query, leading to lower heat generation and energy consumption at the silicon level.

The short answer

Yes, artificial intelligence (AI) can use water, but there is no single amount that applies to every AI system or every request. Water is used mainly to cool the data centers that run AI models, and indirectly to generate the electricity those data centers consume. The amount depends on the model, the hardware, the length and complexity of the task, the data center’s cooling design, the local climate, the electricity mix, and whether water is counted as water withdrawn or water actually consumed.

A single AI interaction may be associated with a small amount of water use, while training and operating large models at data-center scale can involve substantial water demand. Public estimates should therefore be treated as order-of-magnitude approximations, not as a universal “water price” for every prompt. A claim that an AI query uses a particular number of milliliters is incomplete unless it states what system, workload, location, accounting method, and time period were used.

The most accurate general answer is:

• AI uses water primarily for cooling computing equipment.
• AI can also use water indirectly through electricity generation and hardware manufacturing.
• The water associated with one request varies widely and may be difficult to measure.
• Large-scale training, model serving, and data-center expansion can create meaningful local water pressures, especially in water-stressed regions.
• AI does not inherently require drinking-quality water or a fixed amount of water; engineering choices can greatly change its water footprint.

Why does AI use water?

AI software runs on physical computers. Modern models, particularly large language models and systems that process images, audio, or video, often run on specialized processors such as graphics processing units or other accelerators. These chips perform enormous numbers of calculations and release heat while operating.

If that heat is not removed, the equipment can slow down, fail, or suffer a shorter operating life. Data centers use several kinds of cooling systems to transfer heat away from servers. Many of those systems use water either directly or through an industrial cooling loop.

Direct cooling in the data center

A common approach is evaporative cooling. Warm air or warm water is exposed to conditions that allow some water to evaporate. Evaporation removes heat efficiently, much as perspiration cools the human body. The evaporated water is no longer immediately available for reuse at that location, so it is generally counted as water consumption in water-footprint accounting.

Other facilities use chilled-water systems, cooling towers, or combinations of air and liquid cooling. Water may circulate repeatedly through a closed loop, but a facility may still need to add water to replace evaporation, leaks, and water discharged to prevent the buildup of minerals and contaminants.

Some newer facilities use direct-to-chip liquid cooling, in which a coolant carries heat directly from processors. This can reduce dependence on traditional air cooling and may improve energy efficiency, but it does not automatically eliminate water use. The coolant loop may still reject heat through a cooling tower or another water-dependent system. In some designs, heat is rejected primarily through air, reducing operational water consumption at the site.

Indirect water use from electricity

Data centers consume electricity, and power generation can also require water. Thermal power plants often use water for steam production and cooling. The amount and type of water use depend on the fuel, generating technology, cooling system, local climate, and whether the plant withdraws and returns water or consumes it through evaporation.

Hydroelectric facilities interact with water differently from thermoelectric plants. Reservoir evaporation can be included in some water-footprint calculations, although assigning all reservoir evaporation to electricity production is methodologically difficult. Wind and solar photovoltaic generation generally have low operational water requirements compared with many thermal power plants, but their manufacturing and construction still have environmental footprints.

Consequently, a data center’s water footprint cannot always be inferred from its on-site cooling system alone. A facility that uses little water on site may rely on an electricity supply whose generation has a higher indirect water footprint, while a facility with water-intensive cooling may use electricity with comparatively low water requirements.

Water used in hardware manufacturing

AI also depends on a supply chain that includes semiconductors, servers, memory, networking equipment, buildings, and backup power systems. Semiconductor manufacturing uses highly purified water for cleaning and processing wafers. Manufacturing equipment and construction materials also have embedded water and energy requirements.

This is sometimes called embodied or upstream water use. It is distinct from the operational water used while a model is being trained or served. Most public discussions focus on data-center operation because it is easier to connect to a particular workload, but a complete life-cycle assessment would consider both operational and supply-chain impacts.

What determines how much water an AI request uses?

The water associated with AI is not a fixed property of the word “AI.” It is the result of a chain of technical and environmental conditions.

Model and task complexity

A short text classification task generally requires less computation than generating a long answer, creating a high-resolution image, transcribing a long recording, or producing video. Larger models may require more computation per request, although efficient hardware, caching, batching, quantization, and other techniques can reduce the cost of serving them.

The relevant workload is often measured in tokens for language systems, but tokens alone do not determine water use. Hardware utilization, memory access, input length, output length, and the way requests are grouped together also matter. Two requests with the same token count can have different energy and water footprints if they run on different systems or under different operating conditions.

Training versus ordinary use

Training is the process of adjusting a model’s parameters using large datasets. It can require many processors running for extended periods, so it may create a large concentrated demand for electricity and cooling. Training is only one part of a model’s environmental footprint, however.

Inference, or serving the trained model to users, occurs every time the system generates a result. A single inference may be much less computationally intensive than full training, but millions or billions of inferences can make total operational impacts significant. An analysis that counts training but ignores routine use is incomplete; an analysis that estimates a single request without considering the scale of deployment is also incomplete.

Data-center design and climate

Cooling requirements vary with outdoor temperature, humidity, altitude, building design, and equipment density. A data center in a cool climate may be able to use outside air for part of the year. A facility in a hot or humid location may need mechanical refrigeration or more intensive cooling.

Water availability also affects design choices. Operators may use reclaimed wastewater, rainwater, seawater in specialized systems, or municipal supplies, depending on local infrastructure and treatment requirements. The source matters environmentally and socially: using recycled industrial water is not equivalent to drawing scarce potable water from a stressed basin, even if both are reported as “water use.”

Operational efficiency and utilization

The efficiency of processors, servers, and cooling equipment affects how much electricity and water are associated with a task. Better utilization can reduce wasted capacity, while idle servers still consume energy. Software optimization can reduce the number of calculations required, but efficiency improvements may also make AI cheaper to operate and increase total demand. This is sometimes described as a rebound effect: lower resource use per task does not necessarily lower total resource use if the number of tasks grows substantially.

Water accounting method

Water reporting uses terms that are easy to confuse:

A large withdrawal is not automatically equivalent to a large consumption figure. For example, water may pass through a cooling system and be discharged back to a source, although the discharge can affect temperature or quality. Conversely, a smaller withdrawal can produce substantial consumption if much of it evaporates.

Does AI use a lot of water?

The answer depends on the scale and the comparison. An individual request is usually not the most important unit for assessing water stress. The more significant questions are how many requests a service handles, where the computing occurs, how much water the data center consumes, and whether the facility is located in a basin where water is already scarce.

At the level of a large data center, water demand can be substantial because the facility operates continuously and may contain large numbers of high-power servers. AI workloads can increase power density: a rack designed for accelerated computing may produce more heat than a conventional enterprise-computing rack. If the cooling system uses evaporation, that heat can translate into greater water consumption.

However, not every AI data center has the same water profile. Some use air cooling for suitable conditions, closed-loop systems, reclaimed water, or designs intended to minimize potable-water demand. Others may have a higher water footprint because of climate, cooling technology, power source, or local operating practices.

It is also important not to compare an AI service with a vague category such as “the internet” without defining the boundary. AI is one workload among many in the data-center sector. Traditional cloud storage, video streaming, search, online games, scientific computing, and business applications also require electricity and cooling. AI can be unusually computationally intensive for some tasks, but the environmental comparison depends on what service is being compared and what alternatives are being considered.

Why published estimates differ

People often encounter sharply different estimates for the water used by an AI prompt. Such disagreement does not necessarily mean that one number is fabricated. Estimates may differ because they count different things.

A study might divide a data center’s estimated water consumption by the number of model outputs during a period. Another might include the water used to generate electricity. A third might estimate only direct cooling. They may also use different assumptions about model size, response length, hardware, location, season, and whether training is allocated across later users.

There are practical measurement problems as well. AI providers may not disclose the precise location of workloads, the model used for a request, cooling-water volumes, power-purchase arrangements, or the number of requests processed. Workloads can move between data centers as demand, electricity prices, and capacity change. A request may also be handled by multiple services rather than one isolated model.

For these reasons, a responsible estimate should state:

1. The system boundary: direct cooling only, data-center operation, electricity generation, or full life cycle.
2. The workload: training, inference, batch processing, image generation, and so on.
3. The time period: a single request, a day, a training run, or annual operation.
4. The location: climate, water source, electricity grid, and local water conditions.
5. The accounting metric: withdrawal, consumption, or another defined measure.
6. The uncertainty: whether the value is measured, modeled, or inferred from assumptions.

A number presented without these details can create false precision. It is more informative to describe a range and explain its drivers than to imply that every request has an identical water cost.

What are the environmental concerns?

The central concern is usually not that AI consumes water in the abstract, but that water use can be concentrated in particular places and times. A data center may draw from the same watershed, utility system, or groundwater reserve that supports households, agriculture, ecosystems, and other industries. Even a facility with a modest annual footprint can be controversial if demand peaks during a drought or if local infrastructure is limited.

The quality and temperature of returned water also matter. Water withdrawals can affect aquatic habitats, and discharges require appropriate treatment and regulatory oversight. Construction of new data centers may add demands for municipal water, electrical infrastructure, land, and backup systems.

At the same time, water use should be assessed alongside other impacts rather than treated as the sole environmental metric. AI systems also involve greenhouse-gas emissions, mineral extraction, electronic waste, land use, noise, and energy demand. Reducing water use through more energy-intensive cooling might shift impacts rather than eliminate them. A sound assessment considers trade-offs across water, energy, emissions, reliability, and local ecological conditions.

How can AI water use be reduced?

Reducing water use generally requires action at the infrastructure, software, and governance levels.

Data-center and cooling measures

Operators can reduce water demand by using cooling systems that rely more heavily on outside air or dry heat rejection where climate and equipment allow. Direct-to-chip cooling can improve thermal management, and closed-loop designs can reduce ongoing makeup-water requirements. Reclaimed or non-potable water can replace potable supplies when treatment, reliability, and local environmental conditions make that appropriate.

These options are not universally interchangeable. Dry cooling may use more electricity in hot weather, while evaporative cooling can reduce electricity use but consume more water. The best design depends on local water scarcity, grid conditions, climate, equipment, and reliability requirements.

More efficient models and hardware

Quantization, pruning, distillation, smaller specialized models, improved scheduling, and efficient accelerators can reduce the computation required for a given result. Caching repeated responses and avoiding unnecessary regeneration can also reduce workload. These measures lower resource use per task, although total savings depend on whether demand increases as services become cheaper or faster.

Better siting and transparency

Locating facilities where water is more abundant, using low-water electricity, and coordinating with local utilities can reduce pressure on vulnerable watersheds. Public reporting should distinguish withdrawals from consumption and disclose enough information for communities and researchers to evaluate impacts. A low annual average can conceal seasonal stress, so timing and local basin conditions should be reported where possible.

User choices

Individuals cannot control the data-center design behind a service, but they can avoid unnecessary repeated requests, select a smaller model when it is adequate, and use text rather than more computationally intensive media generation when that meets the need. These choices may reduce marginal demand, but they should not be presented as a substitute for infrastructure-level responsibility. The largest effects generally come from how providers design, power, site, and operate their systems.

How to interpret claims about AI and water

When evaluating a statement such as “one AI prompt uses a glass of water” or “AI uses almost no water,” ask what was actually measured. The claim may refer to direct evaporative cooling, include electricity generation, average a facility across many workloads, or allocate a share of training to each user. Each approach can be useful for a particular purpose, but none is a universal answer.

The most defensible interpretation is that AI has a variable water footprint. It is real because computing produces heat and because the supporting electricity and hardware supply chains can require water. It is variable because technical and geographic conditions differ greatly. The meaningful policy question is therefore not merely whether AI uses water, but whether providers are reducing consumption, avoiding stressed watersheds, using appropriate non-potable sources, measuring impacts transparently, and accounting for the growth of demand.

For high-stakes decisions about a specific facility, project, or local water supply, general estimates are not enough. The relevant data should come from the operator, utility, environmental regulator, or an independent water and infrastructure assessment that defines its methods clearly.