Ask an AI chatbot a question, and something happens most people never think about. A few hundred miles away, inside a warehouse full of servers, machines run hotter — and cooling systems, some of which use water, kick in to compensate. That’s behind one of the fastest-growing questions in tech: how much water does AI use?
There’s no single number that answers this cleanly. Google, OpenAI, Microsoft, and independent researchers at universities including UC Riverside have all published figures, and they don’t agree — mostly because they’re measuring different things. Some count only the water evaporated inside a data center’s own cooling towers; others add the water used far away, at the power plant generating the electricity that data center runs on.
This article covers what’s actually known: how much water a single AI query uses, how ChatGPT and Gemini compare, why training a model is water-intensive, and whether AI’s growing water footprint is something to genuinely worry about.
How Much Water Does AI Use?
AI systems use water two ways: directly, to cool the servers that run them, and indirectly, through the water consumed by power plants generating their electricity. So how much water does AI use for a single query? Estimates range from a fraction of a milliliter to several hundred milliliters, depending on the model, the data center’s location and cooling system, the local climate, the electricity source, and even the time of day.
That’s a wide range, and it isn’t a sign that someone is exaggerating — it reflects real differences in what’s being measured. A 2025 first-party measurement from Google put its median Gemini text prompt at 0.26 milliliters of water, about five drops. A widely cited 2023 UC Riverside study, using different assumptions and older models, estimated that a 100-word response from GPT-4 could involve roughly 519 milliliters, close to a standard water bottle, once the electricity supply chain was included.
Both figures can be accurate at once: one measures water at the data center only, the other adds the water used upstream to generate power. That distinction explains almost every disagreement over AI water consumption. The main variables:
- The model — larger, more demanding models generally need more processing power
- Data center location — hot, dry climates need more cooling water than cooler regions
- Cooling technology — evaporative towers use far more water than closed-loop or air-cooling systems
- Electricity source — water-intensive thermoelectric power adds indirect use; wind and solar add much less
- Prompt and response length — longer outputs require more computation
- Direct vs. indirect accounting — on-site cooling water alone, or that plus electricity-related water
Why Does AI Need Water?
AI itself doesn’t touch water — the connection runs through the hardware underneath it. Your prompt is processed by GPUs or specialized AI chips packed into racks inside a data center. These chips work extremely hard, and like any electronics under sustained heavy load, they generate a lot of heat, which would damage the equipment within minutes if left unmanaged.
Data centers manage this heat with cooling systems, and many still rely on evaporative cooling: warm water absorbs heat from the server room, then travels to a cooling tower where much of it evaporates, carrying the heat away. It’s energy-efficient, but that evaporated water can’t be recovered locally — which is why researchers count it as “consumed” rather than merely “used.” Some facilities also use water for humidification, protecting electronics from static and corrosion.
Then there’s the less visible source: electricity generation. Producing the power that runs a data center, whether from gas, coal, or nuclear plants, often requires its own cooling water — so every watt-hour an AI model consumes can carry an indirect water cost before it reaches the server.
How Much Water Does One AI Query Use?
There’s no universal number here, and any source handing you one flat figure without context is oversimplifying.
The most transparent, first-party measurement comes from Google, which published a detailed technical report in August 2025. Using a comprehensive, full-stack methodology, Google found that a median Gemini Apps text prompt consumed 0.24 watt-hours of energy and 0.26 milliliters of water — a steep improvement over the prior year.
OpenAI has published a comparable figure: a 2025 blog post from CEO Sam Altman states that an average ChatGPT query uses about 0.000085 gallons of water, roughly 0.32 milliliters, alongside 0.34 watt-hours of electricity. OpenAI hasn’t released its full methodology, making the number harder to independently verify.
Independent academic researchers have generally landed on higher figures, particularly once indirect electricity-related water is included. The UC Riverside team behind the widely cited “Making AI Less Thirsty” research estimated water use per query could run into the tens of milliliters once the full electricity supply chain was factored in.
The spread exists because the number shifts with the model, prompt length, the specific data center, the day’s temperature, and whether the estimate stops at the data center’s fence line or follows the electricity back to the power plant.
How Much Water Does ChatGPT Use?
OpenAI has disclosed only limited detail about ChatGPT’s water use. The clearest available number is Sam Altman’s 2025 estimate of roughly 0.32 milliliters per average query, which appears to cover on-site operational water rather than electricity-related water, though OpenAI hasn’t confirmed the full scope.
The most cited outside estimate traces to the UC Riverside research, developed with input from The Washington Post, which modeled the resources needed for GPT-4 to write a 100-word email at a typical U.S. data center — about 519 milliliters once both on-site cooling and electricity-related water were included.
That gap between 0.3 milliliters and 519 milliliters isn’t a contradiction; it comes down to accounting boundaries, and without OpenAI’s full methodology, the two figures can’t be cleanly reconciled. Until AI companies converge on a shared, audited reporting standard — something Google’s 2025 disclosure was partly meant to encourage — comparisons between providers will keep running into this problem.
How Much Water Does AI Training Use?
Training a large AI model carries a different kind of water cost than everyday use — a concentrated event rather than an ongoing drip. It runs enormous volumes of calculations across thousands of GPUs, often continuously for weeks or months, generating far more heat than typical day-to-day inference (the model actually answering prompts).
The most-cited estimate, again from UC Riverside researchers, found that training GPT-3 in Microsoft’s U.S. data centers directly consumed around 700,000 liters of water for on-site cooling, rising to an estimated 5.4 million liters once indirect, electricity-related water was added.
GPT-3 is now a comparatively small, older model, and newer frontier models train with far more computing power — so footprints have likely grown since, though no company has released an updated, verified figure. Training happens once or occasionally during retraining; inference happens with every prompt, which is why researchers increasingly see inference as the dominant, fastest-growing share of AI’s water use.
AI Water Consumption vs Other Everyday Activities
Numbers like “519 milliliters per response” are hard to picture without context.
- A single Gemini text prompt (0.26 mL) is about five drops of water.
- The higher UC Riverside estimate for a 100-word response (519 mL) is roughly one bottle of water.
- A cotton T-shirt is commonly estimated at around 2,700 liters across its full growing and production chain, per the Water Footprint Network and World Wildlife Fund — thousands of times more than even the higher per-query AI estimates.
- A typical load of laundry uses roughly 150 liters.
These comparisons help put a single query in perspective, but they mislead if stretched too far. A T-shirt is a one-time purchase; an AI query repeats billions of times daily, which is why researchers focus more on aggregate, regional totals than any per-query figure.
Why AI Data Centers Consume So Much Water
AI GPUs Generate Significant Heat
A single high-end AI accelerator chip can draw several hundred watts or more under full load, and a large data center may house tens of thousands running continuously — throwing off heat that must be removed to keep hardware from failing.
Cooling Systems Require Resources
Evaporative cooling towers remain common: water absorbs heat and evaporates into the atmosphere, carrying it away efficiently. Every gallon that evaporates, though, is drawn from the local supply and not returned. Newer facilities increasingly use direct-to-chip liquid cooling or closed-loop systems that use far less water.
Data Center Location Matters
A facility in a hot, dry climate typically needs more cooling water than one in a cooler region. Some companies favor cooler climates partly for this reason, though power infrastructure, land, and connectivity often outweigh water in site selection.
Electricity Generation Can Have a Water Footprint
Thermoelectric power plants — coal, gas, and nuclear — typically use water for their own cooling, much like a data center does. A facility drawing from a grid dominated by these sources inherits a share of that water use, which is often larger than the direct, on-site figure — why estimates counting only on-site water look far smaller than those tracing the full supply chain.
Is AI Water Consumption a Serious Environmental Problem?
A fair answer holds two things at once: the per-query number is small, and the aggregate, localized impact is real.
Even the higher per-query estimates are modest next to other everyday water uses. The real concern isn’t any single prompt; it’s billions of prompts arriving daily, processed in facilities concentrated in specific places, many already water-stressed. A 2025 study from the Houston Advanced Research Center, a Texas-based nonprofit research institute, projected that data centers across the state could consume roughly 49 billion gallons of water in 2025 alone, potentially climbing sharply by 2030. Communities near major hubs — including parts of Arizona, Georgia, and Northern Virginia — have raised concerns about competing with agricultural and residential water needs during droughts.
A June 2026 report from the United Nations University’s Institute for Water, Environment and Health projected that by 2030, the water footprint of global AI data centers could rival the basic annual domestic water needs of the 1.3 billion people living in Sub-Saharan Africa. A separate December 2025 estimate from researcher Alex de Vries-Gao, founder of the research firm Digiconomist, put AI’s global water footprint for 2025 at several hundred billion liters. Both are forecasts built on growth and efficiency assumptions, not direct measurements.
The most honest summary: AI’s water footprint isn’t a crisis at the level of any single chatbot conversation, but it’s a real, growing infrastructure issue in the specific regions where data centers cluster — a concentration that national averages tend to obscure.
How AI Companies Are Reducing Water Usage
Microsoft’s 2025 Environmental Sustainability Report states the company reached “water positive” status for fiscal year 2025 — replenishing more water globally than its operations consumed — five years ahead of its 2030 target, crediting expanded direct-to-chip cooling across roughly 90% of its newer data centers. It also reports cutting average water-use intensity from 2.3 to about 0.27 liters per kilowatt-hour.
Google’s picture is more complicated: the company replenished 64% of the freshwater it consumed in 2024, up from 18% the year before, even as total consumption rose. It has also published detailed methodology for measuring water and energy per AI prompt — a transparency step other companies haven’t fully matched.
Industry-wide, companies are pursuing direct-to-chip and immersion cooling, closed-loop systems that recirculate water instead of continuously evaporating fresh supply, reclaimed and non-potable water, cooler-climate siting for new facilities, and workload scheduling shifted to cooler hours. These approaches genuinely cut water intensity — the open question is whether efficiency gains can outpace how fast AI infrastructure itself is expanding.
Can AI Become More Water Efficient?
The trajectory so far suggests yes, at least per query. Google’s reported figures show a 33-fold drop in energy use per median Gemini prompt over a single year, driven by software efficiency work, better model architectures, and custom hardware — and water use tends to track energy use closely.
Further gains could come from smaller, more targeted models; better inference optimization; more efficient AI-specific chips; low- and zero-water cooling designs already being deployed at scale; and smarter, climate-informed data center design.
The catch is a rebound effect: even if each query grows more water-efficient, total AI water use can still rise if the number of queries — and data centers built to serve them — grows faster than efficiency improves. Whether AI’s water footprint truly decouples from its growth is likely to be the central story here.
Frequently Asked Questions About How Much Water AI Uses
How much water does AI use?
It depends on the model, the data center, and whether direct cooling water alone or electricity-related water is included. Google’s 2025 measurement put its median Gemini prompt at 0.26 milliliters; a UC Riverside estimate for a longer GPT-4 response came to around 519 milliliters.
How much water does one ChatGPT query use?
OpenAI’s Sam Altman stated in 2025 that an average query uses about 0.000085 gallons, or roughly 0.32 milliliters. Independent researchers estimate higher figures when electricity-related water is included.
Why does AI need water?
AI servers generate heat while processing prompts, and many data centers cool that heat with water-based evaporative towers. AI also carries an indirect water cost through electricity, since power plants often use water for their own cooling.
Does training AI use more water than using AI?
Training GPT-3 is estimated to have used around 700,000 liters on-site, or 5.4 million liters including indirect electricity use — a large, one-time cost. Because inference happens with every prompt at massive scale, it’s now the larger, faster-growing share of AI’s water footprint.
Is AI water consumption bad for the environment?
Per-query use is generally small. The bigger concern is concentrated demand in specific, often water-stressed regions where data center construction is outpacing local water infrastructure planning.
Do AI data centers use drinking water?
Some do. Many have historically drawn on municipal potable water, though Microsoft and Google are increasingly shifting toward reclaimed and recycled sources.
Can AI companies reduce water consumption?
Yes — direct-to-chip cooling, closed-loop systems, reclaimed water, and cooler-climate siting can meaningfully cut water intensity. Microsoft reports reaching water-positive status for its global operations as of fiscal year 2025.
Does every AI model use the same amount of water?
No. Water use varies with model size, prompt length, the data center’s cooling system, local climate, and the electricity source powering it. Larger, more demanding models generally require more processing and more cooling.
The Bottom Line
How much water does AI use? The honest answer is that it depends — on the model answering you, the data center handling the request, the weather that day, and whether the figure counts only on-site cooling or the full electricity supply chain too. What’s clear is that AI’s water footprint is real, growing as adoption scales, and concentrated in ways that matter most to the communities near data centers rather than to any individual user’s daily habits.
The responsible path forward looks less like panic and more like transparency: more companies publishing verified, comparable numbers the way Google and Microsoft have started to, continued investment in low-water cooling technology, and siting decisions that account for local water stress before ground is broken.
