---
title: "How Much Water Does AI Use? 5 Surprising Facts"
url: "https://learnaitodayonline.com/how-much-water-does-ai-use/"
description: "How much water does AI use? The honest answer is “it depends.” Here are the real AI water consumption statistics, per-prompt estimates and 2030 forecasts."
author: "Robert Waithaka"
published: "2026-08-22"
last_reviewed: "2026-08-22"
categories: ["AI News"]
tags: ["level-beginner", "start-here"]
site: "Learn Artificial Intelligence"
approx_tokens: 3485
---

# How Much Water Does AI Use? 5 Surprising Facts

How much water does AI use? Search and you'll find answers from “five drops” to “half a bottle per email” — a gap of roughly 2,000 times. The truth is more interesting than any single number. AI water consumption is real and growing fast, but most viral statistics measure different things. Here is what the research actually says, in plain English, with figures dated and sourced.

## How Much Water Does AI Use?

Start with per-prompt estimates, where the confusion begins. Google says a typical AI query uses about five drops; OpenAI’s Sam Altman compares one to a fifteenth of a teaspoon. A widely shared 2024 Washington Post analysis, built with UC Riverside researchers, estimated a 100-word AI email could use 519 millilitres — more than a water bottle.

The researcher behind that figure, Shaolei Ren, has since revised it to roughly 15 millilitres per prompt for today’s systems and calls the original estimate outdated. A 2025 benchmark called “How Hungry is AI?” found efficient models using less than two millilitres per query, while some reasoning models exceeded 150 millilitres.

| Claim you may have seen | Number | What it actually measures | Source |
| --- | --- | --- | --- |
| “A ChatGPT query uses five drops of water” | ~0.26 ml | Median Gemini text prompt, full data-centre measurement | Google via EcoFlow |
| “A query is one-fifteenth of a teaspoon” | ~0.33 ml | Sam Altman's description of a typical ChatGPT query | CBS News |
| “A 100-word AI email uses a bottle of water” | 519 ml | GPT-4 email at an average US data centre (cooling + electricity) | Washington Post & UC Riverside (2024) |
| “The real number is a few drops” | ~15 ml | Ren's revised estimate for today's GPT-4 systems (~5 ml on-site) | Forbes (2026) |
| Efficient vs reasoning models | <2 ml to >150 ml | Range across models and tasks (“How Hungry is AI?”, 2025) | Forbes |

_Table: the viral per-prompt water numbers, decoded. The largest figure is roughly 2,000 times the smallest because they measure different things._

None of these numbers is simply wrong. They measure different models, tasks and accounting boundaries: some count only cooling water, others add the water behind the electricity. The honest answer to how much water AI uses per prompt is a few drops to a small bottle, depending on the model and where it runs.

![How much water does one AI prompt use — published estimates](https://cdn.sanity.io/images/gfihpee1/production/521524dc4fd7b69a4bfa80c44b6dc31599b49886-1748x770.png)

_Figure: published estimates of water used per AI prompt, in millilitres (log scale). Sources: Google via EcoFlow; Forbes; Washington Post & UC Riverside; CBS News._

## AI Water Consumption: On-Site Cooling vs Off-Site Power

AI’s water footprint has two main parts, and conflating them drives most of the confusion. On-site water cools the servers. In evaporative cooling — used by most large facilities — water absorbs heat and evaporates, and 70–80% of what a data centre withdraws is consumed this way, gone from the local watershed until it rains again. Google’s facilities evaporate about 80% of what they withdraw.

The second part is often larger: water used to generate the electricity data centres consume. Coal plants use around 19,000 gallons per megawatt-hour and gas about 2,800, so a fossil-fuelled grid carries a hidden water bill. Because most data centre electricity still comes from fossil fuels, this indirect use often exceeds direct cooling — which is why the UN University and Lawrence Berkeley National Laboratory treat them as one footprint.

## AI Water Usage Statistics: What the Research Says

At the aggregate level, the numbers are substantial. US data centres used an estimated 228 billion gallons of water in 2023 — about 17 billion gallons for cooling and 211 billion gallons tied to electricity generation, according to Lawrence Berkeley National Laboratory. Berkeley projects that total could reach 469–844 billion gallons by 2028.

![How much water does AI use — US data centre water use in 2023, cooling versus electricity](https://cdn.sanity.io/images/gfihpee1/production/b385f23723fe3d38050dbc62a9afb5d583d77d47-1383x688.png)

_Figure: of the 228 billion gallons US data centres used in 2023, only about 17 billion went to on-site cooling; 211 billion were tied to electricity generation. Source: Lawrence Berkeley National Laboratory via CBS News._

| Statistic | Figure | What it measures | Source |
| --- | --- | --- | --- |
| US data-centre water use, 2023 | 228 billion gallons (17B cooling + 211B electricity) | Total direct and indirect water | LBNL via CBS News |
| US projection, 2028 | 469–844 billion gallons | Range depending on growth and grid mix | LBNL |
| Global data-centre water, 2025 | 4.5 trillion litres | All data centres worldwide | UNU-INWEH |
| Global projection, 2030 | 9.3 trillion litres | All data centres worldwide | UNU-INWEH |
| Training GPT-3 (direct use) | 700,000 litres | Fresh water evaporated during training in Microsoft's US data centres | “Making AI Less Thirsty” |
| AI water footprint, 2025 | 312.5–764.6 billion litres | Global AI workloads, estimated from data-centre metrics | Patterns (2026) |
| US AI servers, 2024–2030 | 731M–1.125B m³ per year | Annual water footprint of US AI servers | Nature Sustainability (2025) |
| Share of withdrawn water that evaporates | ~70–80% | On-site cooling water consumed rather than returned | EESI / Google data |

_Table: the key AI water consumption statistics, each with its source and what it actually counts._

Globally, UN University researchers found data centres consumed 4.5 trillion litres of water in 2025, and project 9.3 trillion litres by 2030 — roughly double — with AI’s share of data centre demand growing from about a fifth to 40%. For AI specifically, a 2025 Nature Sustainability study estimated US AI servers alone could carry a water footprint of 731 million to 1.125 billion cubic metres per year between 2024 and 2030, most of it indirect. A 2026 Patterns paper puts the global AI water footprint at 312.5–764.6 billion litres in 2025 — comparable to the world’s annual bottled-water consumption.

![Global data centre water consumption projected to double by 2030](https://cdn.sanity.io/images/gfihpee1/production/923d5333655cf988d813bc214cd0593531dbafdb-1137x730.png)

_Figure: global data-centre water consumption is projected to grow from 4.5 trillion litres in 2025 to 9.3 trillion litres by 2030. Source: UNU-INWEH via Reuters and UN News._

Training matters too: “Making AI Less Thirsty” estimated that [training GPT-3](https://learnaitodayonline.com/how-does-ai-learn/) in Microsoft’s US data centres directly evaporated 700,000 litres of fresh water.

## Where the Water Comes From: Location Matters

Aggregate numbers understate the real issue, which is local. Roughly two-thirds of data centres built since 2022 sit in water-stressed regions, and the Guardian found 517 of 809 planned US data centres in areas that spent the past year in drought. Around 40% of US data centres are in places with high or extreme water stress. A single large facility can draw up to five million gallons a day — as much as a town of 50,000 people.

| Local pressure | Figure | Source |
| --- | --- | --- |
| Planned US data centres in drought areas | 517 of 809 | Guardian analysis (Cleanview data) |
| US data centres in high or extreme water stress | ~40% | KETOS / Florida Water & Pollution Control Operators Association |
| Phoenix cooling water today → planned build-out | 385M → 3.7B gallons/year | Ceres via Consumer Reports |
| Peak-day cooling demand vs annual average | 6–30× | UC Riverside & Caltech via Forbes |
| New US water infrastructure needed by 2030 | $10–58 billion | UC Riverside & Caltech via Forbes |
| A single large facility on a hot day | up to 5 million gallons/day | EESI / EPA |

_Table: why location matters more than any per-prompt number._

In Phoenix, direct cooling already uses about 385 million gallons a year, and Ceres projects 3.7 billion — an 870% increase — once planned facilities open. Peak days matter more than annual averages: cooling demand can spike to 6–30 times the average in hot weather, which is why US water systems may need $10–58 billion in new infrastructure by 2030. In The Dalles, Oregon, Google takes more than a quarter of the town’s water. That is the hidden environmental impact of AI: global benefits, concentrated local costs.

![US data centre water projections and Phoenix cooling demand](https://cdn.sanity.io/images/gfihpee1/production/beb7bc805f14b298d823a6858e3a6568c9b45b1f-1605x598.png)

_Figure: US data-centre water use is projected to reach 469–844 billion gallons by 2028, and Phoenix-area cooling water is projected to jump from 385 million to 3.7 billion gallons per year. Sources: LBNL; Ceres via Consumer Reports._

## Is AI Bad for the Environment Compared to Other Things?

Set beside everyday water use, AI’s footprint looks smaller than the headlines suggest. American lawns use roughly 2,900 billion gallons a year and golf courses more than 500 billion — many times more than all US data centres combined. Fodder crops for animal feed use about 5.5 trillion gallons of irrigation water. AI cooling water is less than what Americans use for golf courses, car washes, pools and restaurant dishwashing combined. And yes, the viral claim is true: one hamburger can use more water than years of daily ChatGPT queries.

But that comparison can lull people into complacency. Data centre demand is growing faster than almost any other sector, it is concentrated in the driest places, and efficiency gains are being swallowed by the rebound effect — cheaper, faster AI simply gets used more. So “is generative AI bad for the environment?” is the wrong question. The useful one is: compared with what, where, and at what growth rate? Even “is Netflix worse than ChatGPT?” misses the point — both run on the same data centres, and AI is roughly 15–20% of data centre electricity today.

## Can the Water Used by AI Be Reused?

Yes — and the industry is moving that way, though slowly. Closed-loop cooling recirculates the same water and can cut freshwater use by up to 70%. Amazon says it cools more than 100 data centres with recycled water, saving hundreds of millions of gallons of drinking water a year, and Google uses reclaimed water at about a quarter of its campuses. Microsoft is piloting zero-water-evaporation cooling designs.

The catch: most existing facilities still draw potable water, much of it evaporates, and closed-loop systems often need more electricity — which can raise indirect water use if the grid runs on fossil fuels. Water quality matters too: cooling-tower discharge can carry concentrated minerals and treatment chemicals that strain local wastewater plants. Reuse helps, but it does not erase the footprint.

## How to Read the Next AI Water Headline

You do not need to memorise statistics. When you see a claim about AI and water, ask three questions: What is being counted — cooling only, or electricity too? Where and when was it measured — a desert data centre in August differs from one in Ireland in March? And how old is the estimate, given how fast this technology changes?

The most honest answer to “how much water does AI use” is a range with assumptions attached, and any source offering a precise single number is probably selling you a simplification. If you want the terms straight, our [AI glossary](https://learnaitodayonline.com/ai-glossary-essential-terms/) is a good place to start, and our explainer on [what an LLM is](https://learnaitodayonline.com/what-is-an-llm/) covers the machinery behind the water bill.

As a user, your choices matter less than the collective pattern, but they are not nothing: a single AI image can use around 1,450 times more energy than basic text, so simpler tasks on default settings help. Running models on your own hardware does not remove the footprint — it moves it — though our [Ollama guide](https://learnaitodayonline.com/ollama-download-install-run-local-ai-models/) covers local setup. The real fix is systemic: transparency, mandatory reporting, better siting, and water planning that treats data centres like the industrial water users they are.

---

## Sources

- [Making AI Less “Thirsty” (arXiv, 2025)](https://arxiv.org/abs/2304.03271)
- [2024 United States Data Center Energy Usage Report (Lawrence Berkeley National Laboratory)](https://eta-publications.lbl.gov/sites/default/files/2024-12/lbnl-2024-united-states-data-center-energy-usage-report.pdf)
- [Environmental impact and net-zero pathways for sustainable artificial intelligence servers in the USA (Nature Sustainability, 2025)](https://www.nature.com/articles/s41893-025-01681-y)
- [The carbon and water footprints of data centers (Patterns, 2026)](https://doi.org/10.1016/j.patter.2025.101430)
- [Data Centers and Water Consumption (EESI, 2025)](https://www.eesi.org/articles/view/data-centers-and-water-consumption)
- [How much water does AI really use? (CBS News, 2026)](https://www.cbsnews.com/news/how-much-water-does-ai-really-use/)
- [Circular water solutions for sustainable data centres (World Economic Forum)](https://www.weforum.org/stories/2024/11/circular-water-solutions-sustainable-data-centres/)

_Last reviewed: 22 August 2026. Water and energy figures move quickly; this article will be re-checked quarterly._