There is no single fixed amount of water that ChatGPT uses per question. The best-known academic estimate suggests roughly 500 millilitres for a conversation of about 20 to 50 queries—equivalent to around 10–25 mL per query if divided evenly—but that is a scenario-based estimate, not a universal meter reading. The real amount can be far lower or higher depending on the model, data centre, cooling system, local weather, electricity source and length of the response.
Updated 23 August 2026. This guide separates what researchers have estimated from what cannot yet be known precisely.
| Question | Best evidence-based answer |
|---|---|
| How much water does one short ChatGPT query use? | No reliable universal figure is available. |
| What does the widely quoted 500 mL figure mean? | About 20–50 queries under the assumptions in a 2023 University of California, Riverside estimate—not one prompt. |
| How much water did GPT-3 training use? | The same research estimated about 700,000 litres of direct freshwater for one particular training scenario. |
| Why do estimates differ? | A 2025 Lawrence Berkeley National Laboratory study found workload-level water use can vary by more than 10,000-fold. |
Why there is no universal per-query water number
A ChatGPT prompt does not open a tap beside your device. It sends work to computing infrastructure, and the environmental footprint depends on where and how that work is performed. Even two identical prompts may be routed to different hardware or locations. A five-word answer and a long reasoning-heavy response also demand different amounts of computation.
Researchers therefore estimate water use from variables such as server energy, power usage effectiveness, water usage effectiveness, local temperature, cooling technology and the water intensity of the electricity grid. Each boundary produces a different answer. A figure that counts only water evaporated at the data centre cannot be compared directly with one that also includes water used to generate electricity.
This uncertainty is not a reason to dismiss the footprint. It is a reason to label the assumptions and avoid converting one study into a timeless claim about every prompt.
Where ChatGPT’s water footprint comes from
AI services can have two principal operational water pathways:
- Direct water use: water used at a data centre, often to remove heat through evaporative cooling.
- Indirect water use: water associated with producing the electricity consumed by servers and supporting equipment.
It also matters whether a report measures water withdrawal or water consumption. Withdrawal is water taken from a source; some may be treated and returned. Consumption is the portion not returned promptly to the same watershed, often because it evaporates. Neither measure alone describes local ecological stress: using a litre in a water-scarce basin can have different consequences from using one where water is plentiful.
The 2025 Lawrence Berkeley National Laboratory analysis showed why simple averages mislead. Across workload scenarios, water use differed by more than four orders of magnitude. Server efficiency, grid water intensity, utilisation, cooling, infrastructure and climate all mattered.
What the “500 mL for ChatGPT” estimate actually says
The often-repeated bottle-of-water analogy came from research publicised by the University of California, Riverside in April 2023. Its authors estimated that a conversation involving roughly 20–50 questions and answers could consume about 500 mL of water, depending on when and where the service ran.
Dividing 500 mL by that range gives a rough 10–25 mL per exchange. That arithmetic can help readers understand the scale of the study, but it should not be presented as a measured current rate for ChatGPT. The estimate used specific assumptions about models, facilities, grids and conditions available at the time. AI hardware and data-centre practices continue to change, while providers do not publish enough workload-level data to calculate a precise live figure.
The same study estimated that training GPT-3 in one Microsoft data-centre scenario could directly consume around 700,000 litres of freshwater. That training estimate is separate from the water associated with answering a user’s prompt. Training builds the model; inference is the repeated computation used when people interact with it.
Six factors that can change the answer
- Model and response size. A compact model producing a short answer generally requires less computation than a larger model generating a long one.
- Hardware efficiency. Newer accelerators can complete the same work with less electricity, although total demand may still grow as usage increases.
- Server utilisation. Equipment running efficiently at high utilisation spreads supporting overhead across more useful work.
- Cooling system. Evaporative cooling can save electricity but consume water; air cooling may use less on-site water but more energy.
- Weather and location. Temperature, humidity and local water scarcity affect cooling needs and the consequences of water use.
- Electricity mix. Power sources have very different water footprints, so the same server workload can have a different indirect impact by region and hour.
How data-centre cooling changes water use
Computers turn most of their electricity into heat. Data centres must move that heat away reliably, but cooling systems make different trade-offs. Evaporative towers reject heat efficiently by evaporating water. Air-cooled or dry systems reduce direct water consumption, yet may require more electricity, especially in hot weather. Closed-loop liquid cooling can transfer heat efficiently inside a facility, but the final heat-rejection stage still determines much of the site’s water demand.
Facility operators can also use reclaimed water, operate cooling differently during hot periods or site infrastructure in cooler regions. The best choice depends on the local grid and watershed rather than a single global rule. For a closer look at the hardware side of heat removal, see Dillo’s guide to aluminium extrusions in computer cooling systems.
Training, inference and scale are different questions
Large training runs attract striking headline numbers because thousands of accelerators may operate for weeks. Yet a model may then answer millions or billions of requests. To understand its lifetime impact, analysts need both the one-off training footprint and the cumulative inference footprint.
There is also a difference between a prompt, an exchange and a conversation. A request can include a long document; an exchange includes the generated answer; a conversation may preserve extensive context over many turns. Comparing figures without matching these units creates false precision.
At system level, efficiency gains can be offset by rapidly increasing demand. This is sometimes called a rebound effect: a cheaper, faster service is used more often. That is why both per-task efficiency and total annual water consumption matter.
Is ChatGPT’s water use environmentally harmful?
The answer depends heavily on location, timing and scale. Water consumption can be especially consequential in drought-prone or water-stressed areas. The United Nations Environment Programme’s 2026 guidance highlights design, cooling technology and site selection as important sustainability levers, particularly where facilities rely on water-based cooling in hot or water-stressed regions.
Water is only one part of responsible AI. Energy, carbon, land, materials and electronic waste also deserve attention, along with the service’s social value. Dillo’s overview of ethical AI and responsible automation explains why transparent reporting and accountable decisions matter alongside technical efficiency.
How to evaluate an AI water-use claim
Before sharing a dramatic number, check whether the source answers these questions:
- Does it identify the model, hardware, place and date?
- Is the unit a prompt, exchange, conversation or training run?
- Does it count direct cooling water, indirect electricity water or both?
- Does it distinguish water consumption from withdrawal?
- Is it a measured value, a provider disclosure or a modelled estimate?
- Does it give a range and explain uncertainty?
A claim that omits most of those details may still start a useful conversation, but it should not be treated as a precise universal fact.
What users and providers can do
Individuals can reduce unnecessary computation modestly by writing clear prompts, combining related questions, requesting an appropriate answer length and avoiding repeated regenerations that add no value. Those choices should remain proportionate: personal prompt habits are not a substitute for infrastructure-level action.
The biggest levers belong to providers and policymakers. They include efficient models and hardware, transparent workload-level reporting, water-aware siting and scheduling, reclaimed-water systems, lower-water electricity, responsible cooling design and safeguards for stressed watersheds. Reporting should disclose both annual totals and intensity metrics so that falling per-query use does not conceal rising overall demand.
Frequently asked questions
Does one ChatGPT prompt use 500 mL of water?
No. The widely cited estimate was about 500 mL for approximately 20–50 questions and answers under particular 2023 assumptions.
How much water does ChatGPT use per question?
No universal measured value exists. Dividing the UCR estimate gives roughly 10–25 mL per exchange, but that range is an illustration of one modelled scenario, not a current fixed rate.
Why does an online AI service need water?
Its servers produce heat, and some data centres use water in cooling. Producing the electricity that powers the servers can also consume water.
Is it always drinking water?
No. The source varies by facility and can include municipal freshwater or reclaimed water. The source and local water scarcity are essential context.
Can AI water consumption be reduced?
Yes. More efficient models and hardware, better utilisation, low-water cooling, reclaimed water, careful siting, cleaner power and transparent measurement can all help.
The bottom line
So, how much water does ChatGPT use? A defensible shorthand is that one influential study estimated about half a litre for 20–50 exchanges, while the true amount for any particular query is unknown and highly variable. Treat that figure as a useful order-of-magnitude illustration—not a universal conversion rate. The clearest path forward is better provider disclosure paired with efficient computing, water-aware infrastructure and honest reporting of uncertainty.