iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
AI uses water, but there is no universal amount per prompt. Its footprint can include water consumed to cool data centers, water used to generate their electricity, and water used to manufacture chips. The figure depends on the model, workload, location, cooling system and accounting boundary.
Where AI’s water footprint comes from
When people ask how much water an AI prompt uses, they often mean the water needed to cool the computers that process it. That is only one part of the footprint. The OECD identifies three main pathways:
- Data-center cooling: Some facilities use water directly to remove heat from servers. The amount varies with cooling technology, climate and operating conditions.
- Electricity generation: Power plants can use water to generate electricity. This is an indirect footprint associated with the computing workload, not water delivered to the data center for cooling.
- Semiconductor manufacturing: Making chips also uses water. This is part of the hardware’s broader lifecycle footprint, not the same measure as water consumed while running a model.
The OECD’s 2022 review explains these pathways and notes that water consumption is less well understood than operational energy use and greenhouse-gas emissions: OECD, “The AI footprint”.
Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Scan for outdated or missing drivers - takes under a minuteDriver Scan →These components cannot simply be added unless their boundaries are clear. For example, a figure for direct cooling water and another that already includes water used to produce electricity could overlap if combined carelessly.
#1 Best Overall
How much water does an AI prompt use?
One published estimate gives a useful example, not a universal rate. In a 2025 technical report, Google researchers measured median Gemini Apps text serving in production and reported 0.26 milliliters of water and 0.24 watt-hours of energy per prompt, using the study’s defined functional unit and full-stack measurement approach. The finding applies to the measured service and text prompts; it should not be treated as a figure for every model, prompt, or data center. The authors also describe how location, season and cooling technology affect water demand.
Read the 2025 preprint or the technical report for the study’s scope and method. Its number is not directly comparable to estimates with different workloads or accounting boundaries.
In particular, a short text exchange and a different task or workload are not interchangeable units. A per-prompt number also does not tell you how much water a provider uses overall: total use depends on the number and type of tasks, where the computing runs, and the infrastructure serving them.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Fix the driver behind crashes, sound loss and screen glitches3Clear out junk files and repair common Windows errorsRank #2
Why training and serving estimates are different
Training is the computation used to build a model. Serving, or inference, is the computation used to respond to requests after the model is deployed. They are different workloads, so an estimate for training cannot be divided or repurposed into a reliable per-prompt figure without additional evidence and assumptions.
A widely cited 2023 preprint by Li, Yang, Islam and Ren estimated that training GPT-3 in specified Microsoft U.S. data centers consumed 700,000 liters of direct freshwater. That is the authors’ model-based estimate, not an independently measured audit of the training run. The authors estimated that the modeled consumption would be three times higher if the training took place in Microsoft’s Asian data centers. Those estimates concern a particular training workload and location scenario, not the water footprint of each later prompt.
See “Making AI Less ‘Thirsty’” for the authors’ assumptions and analysis. Comparing its training estimate with Google’s later serving measurement would not be a controlled comparison of the two companies’ models.
Why water use varies by data center
Cooling design and local conditions matter. Google’s 2025 technical report says air-cooled systems consume little to no water, while evaporative or adiabatic cooling uses more water during hotter and drier periods. A data center’s water demand can therefore differ by site and season, even when the computing task is similar. The report cautions that those differences can distort comparisons and recommends trailing-twelve-month or annual averages when comparing fleets.
Location matters beyond the facility’s cooling equipment. Electricity sources differ in their water use, and a facility’s effect on local water resources depends on the place and the surrounding demand. A national or company-wide total can obscure pressure in a water-scarce area.
What company water totals do—and do not—show
Company reports provide context about infrastructure, but their totals are not automatically AI-only measurements. These published figures have different scopes and metrics:
| Figure | What it measures | What it does not establish |
|---|---|---|
| Google data centers consumed 6.1 billion gallons of water in 2023, 17% more than in 2022. | Google’s data-center water consumption, as reported in its 2024 Environmental Report. | It is not an AI-only total. The report said the increase mirrored growth in electricity use; that does not show that AI alone caused it. Google’s 2024 Environmental Report. |
| Google replenished 1 billion gallons in 2023, equal to 18% of its freshwater consumption that year. | A reported replenishment figure and its relationship to Google’s freshwater consumption. | Replenishment is not the same as reducing consumption at a particular data-center site. Google’s 2024 Environmental Report. |
| Microsoft reports 14.2 million cubic meters replenished in FY25. | A company-level replenishment figure in Microsoft’s sustainability reporting. | It is not an AI-attributed water-use total. Microsoft’s Environmental Sustainability Report. |
Earlier reporting offers another example of why attribution needs care. The Associated Press reported that Microsoft’s company-wide water consumption rose 34% from 2021 to 2022. Researcher Shaolei Ren attributed most of that growth to AI, but that was his assessment—not a company-published AI-specific measurement. The report also described Microsoft data-center operations in Iowa and their local water context. Associated Press, September 9, 2023.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Check the metric before comparing water claims
“Water use” can refer to different things. Withdrawal is water taken from a source; consumption is water that is not returned to the same source in the same condition or within the relevant timeframe. Replenishment describes efforts to restore water or support water availability; it is a separate measure, not a synonym for lower consumption at a facility. The OECD review discusses these distinctions and warns that reporting remains incomplete.
Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteWindows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallBefore comparing two AI water figures, check:
- Workload: Is it model training, batch processing, or interactive serving? What model and task are included?
- Boundary: Does the figure cover direct cooling, water associated with electricity, chip manufacturing, or some combination?
- Metric: Is it withdrawal, consumption, or replenishment? Does it distinguish freshwater from other sources?
- Place and period: Which facility or region, climate, season and reporting year does it cover?
- Method: Is it a production measurement, company-wide report, model-based estimate, or scenario projection?
- Infrastructure: What cooling system and data-center overhead are included?
These details are essential because a site-specific estimate, a fleet total and a per-prompt measurement answer different questions.
How much of AI’s water use is known?
There is no complete, consistent public accounting that turns all AI workloads into one comparable water total. In its 2022 review, the OECD cited earlier sources estimating that 33–50% of data-center operators compiled and reported water-use metrics. That is a historical estimate from the report, not a current census of operators. The OECD also cautioned that water reporting was less developed than reporting on energy and greenhouse-gas emissions.
The practical implication is not that AI’s water use is unknowable, but that individual figures need their scope attached. Better disclosure would make it easier to distinguish an AI workload from other data-center activity and compare facilities on consistent terms.
Why the footprint matters—and what can change it
Water impact is local as well as global. As the OECD notes, a data center can compete with other users for water in a water-scarce location even if data centers account for a small share of a country’s total water use. A modest national percentage does not, by itself, show that a facility has little local impact.
Providers can address the footprint through more efficient computing workloads, informed facility siting, cooling choices suited to local conditions, and clearer water measurement and reporting. No single per-prompt figure captures those decisions or their effects across every site.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
