Straight answer: yes, every ChatGPT query uses some water, mostly to cool the data centres running the servers, though the exact amount is disputed and depends heavily on which data centre handled your request. For most small and mid-sized businesses this is not the sustainability issue it’s being sold as, but for anyone in public sector tendering, ESG reporting, or client-facing sustainability claims, you need a straight answer ready, not a shrug.
Related reading: What Is a WhatsApp Business API Account and How It Differs From the Ap.
The number everyone quotes, and why it’s messier than it sounds
The figure that did the rounds in 2023 came from a paper by researchers at the University of California, Riverside and the University of Texas at Arlington, titled “Making AI Less Thirsty.” They estimated that ChatGPT needs to “drink” roughly a 500ml bottle of water for a short conversation of about 20 to 50 questions and answers. That number covers both the water evaporated directly in cooling towers and the water used indirectly to generate the electricity that powers the servers.
The same team estimated that training GPT-3 in Microsoft’s US data centres could have directly evaporated around 700,000 litres of clean freshwater, which they compared to the water needed to manufacture several hundred BMW cars.
Then in June 2025, OpenAI pushed back with its own number. Sam Altman wrote that the average ChatGPT query uses about 0.000085 gallons of water, which he described as roughly one fifteenth of a teaspoon. That’s a wildly different figure from the university researchers, and both can be technically true at the same time, because they’re measuring different things: training versus a single query, direct cooling water versus the full electricity supply chain, and one specific data centre versus an industry average.
This is the bit most coverage skips over: there is no single, agreed number for “how much water does ChatGPT use,” because it depends on which data centre answered your prompt, what cooling system that facility uses, and whether the local electricity grid is powered by hydro, gas, or coal. A prompt answered by a data centre in Iowa on a hot July afternoon costs more water than the same prompt answered by a data centre in a cooler climate using closed loop cooling.
Where the water goes
Two things use water in AI, and they’re often lumped together, which muddies every headline you’ll read.
- Direct cooling. Many large data centres use evaporative cooling towers, similar in principle to how sweating cools you down. Water evaporates and takes heat with it. This water is mostly lost to the local water table, not returned.
- Indirect water via electricity. Generating electricity, especially from coal, gas, or nuclear plants, uses water for cooling turbines. So even a data centre using air cooling instead of water still has an indirect water footprint through its power supply.
The Associated Press reported in 2023 that Microsoft’s data centre cluster in West Des Moines, Iowa, which was used to help train GPT-4, drew an estimated 11.5 million gallons of water in July 2022 alone, the same month training was reportedly underway. Iowa has had drought conditions in several of the last five years. That’s the part of this story that gets missed: it’s not the total volume that matters most, it’s where that volume is drawn from, and whether the local community is already short of water. A million litres in a region with a river running through it is a non-event. The same volume in a drought-stressed county is a genuine local problem, and it’s one of the reasons Microsoft, Google, and Amazon have all pledged to become “water positive” by around 2030, meaning they aim to replenish more water than their data centres consume.
A real conversation I had about this
A client of mine, an agency owner in Bristol with eleven staff, came to me last spring because she’d lost a local authority tender partly on an ESG questionnaire. One question asked her to describe the environmental impact of any AI tools used in service delivery, including energy and water use. She’d been using ChatGPT daily for months, for drafting, research, and content, and had never once thought about it in those terms. She had no answer, so she wrote “not applicable,” and the assessor marked her down for it, because “not applicable” reads as “hasn’t thought about it,” which is exactly what happened.
We rebuilt that section together. Not with guilt, and not with a made-up green claim, but with three honest sentences: which AI tools she uses, an acknowledgement that all cloud-based AI tools have an energy and water footprint tied to data centre operations, and a note that she chooses providers who’ve published sustainability commitments (OpenAI, Microsoft, and Google have all published water and carbon targets, even if the detail is patchy). That’s it. She won the next tender she went for. The lesson wasn’t “stop using ChatGPT,” it was “have an answer ready before someone asks.”
The comparison nobody running this argument wants you to sit with
Here’s the part that tends to get left out of both the doom pieces and the tech company rebuttals. A single Google search uses a tiny fraction of a teaspoon of water too, and nobody’s cancelling their Google account over it. Streaming an hour of Netflix, running a Zoom call, or storing your email in the cloud all draw on the exact same data centre infrastructure, cooled the exact same way, and almost none of your business’s software vendors publish per-use water figures for any of it. ChatGPT gets singled out not because its footprint is uniquely bad, but because OpenAI is the one company that got specific enough with a number for researchers and journalists to argue about.
Put it against the water footprint of things your business already does without a second thought: a single cotton t-shirt takes roughly 2,700 litres of water to produce, one pint of beer takes around 150 litres, and a kilogram of beef can take over 15,000 litres depending on how the cattle were raised. If your business serves lunch at events, ships physical products, or runs a fleet of vans, your actual water and carbon footprint almost certainly dwarfs your ChatGPT use by a wide margin. Worrying loudly about a fifteenth of a teaspoon per prompt while doing nothing about the rest is optics, not sustainability, and most businesses asking the ChatGPT water question have never once audited the water footprint of anything else they buy.
That’s not an argument for ignoring it. It’s an argument for proportion. If water use matters to your business, measure the whole picture, not just the one input that happens to have a viral statistic attached to it.
When this should change what you do
For most small businesses using ChatGPT to draft emails, summarise meetings, or brainstorm content, the honest answer is: it shouldn’t change much day to day. The per-prompt footprint is small next to almost everything else your business consumes. But there are three situations where it matters.
- You’re bidding for public sector or enterprise contracts. ESG and procurement questionnaires increasingly ask about digital tools, not just physical operations. Have a one paragraph answer ready, as my Bristol client learned the hard way.
- You’re running AI at real volume. A solo founder asking ChatGPT forty questions a week is a rounding error. A customer service team running thousands of AI-generated replies daily, or a content operation generating hundreds of images or videos a day, is a different scale of consumption and worth a second look, especially if you’re also making public sustainability claims.
- You’re choosing between providers on other grounds anyway. If you’re comparing AI vendors for cost or capability, sustainability commitments are a reasonable tiebreaker. Google’s data centres reported using 5.6 billion gallons of water across its global operations in 2022, up about 20% on the year before, and Microsoft’s water use rose 34% the same year, largely attributed to AI expansion. Both companies have since published water positive targets and started disclosing more detail, which is more transparency than most competitors offer.
A short, practical checklist
If you want to be able to answer this rather than guess, here’s what I’d do, in order:
- List every AI tool your business uses regularly, not just ChatGPT: image generators, transcription tools, AI-powered CRM features, all of it. Our roundup of AI image tools for business is a good place to see how varied this list can get once you count everything.
- Check whether each provider has published a sustainability or environmental report. OpenAI, Microsoft, Google, and Anthropic all have public statements, even if the level of detail differs.
- Write one short, honest paragraph about your AI use and its environmental footprint, ready to drop into any tender or client questionnaire, before you’re asked for it under time pressure.
- If your AI use is high volume, ask your provider directly about their water sourcing and whether their data centres are in water-stressed regions. Most sales teams won’t have the answer immediately, but asking puts it on record.
- Keep proportion in view: audit your other water and carbon costs (travel, shipping, catering, office energy) at the same time, so the AI line item sits in context rather than standing alone as the scary number.
If building that kind of AI policy from scratch feels like more than you have time for, this is exactly the sort of groundwork an AI implementation coach can help you put together in a few sessions rather than months of trial and error.
What this means for how you talk about AI publicly
If you’re a marketing or consulting business, this question is going to keep coming up, because AI adoption is rising fast and scrutiny is rising with it. Our own AI Search Demand Report 2026 found a steady climb in searches around “is AI bad for the environment” and similar questions, which tells me clients and prospects are starting to ask their suppliers about it, not just journalists asking tech companies. Get ahead of that by having a plain answer written down, the same way you’d have a plain answer for GDPR or data security. You don’t need a green claim you can’t back up. You need three honest sentences and a source.
It’s also worth keeping half an eye on how fast the underlying numbers move. We track this kind of shift in our weekly AI news roundup, because provider commitments, efficiency improvements, and public figures on this topic change every few months as newer, more efficient models roll out. The water-per-query figure for GPT-5 class models is already lower than the older estimates that keep getting recirculated online, simply because newer chips and cooling systems are more efficient. Quoting a 2023 study as if it describes 2026 models is one of the most common mistakes I see in this whole debate.
And if the sustainability angle is a selling point for your business, don’t just talk about AI. Look at your whole toolkit. Our list of free business tools is a decent starting point if you’re trying to cut costs and consumption across the board, not just around AI.
The bit I want you to take away
ChatGPT uses water. So does your kettle, your office aircon, the server hosting your website, and the coffee shop where you hold client meetings. The question worth asking isn’t “does AI use water,” because the answer is obviously yes and always was going to be. The question worth asking is whether your business has an honest, specific answer ready when someone asks about it, and whether you’re applying the same scrutiny to AI that you apply to everything else you buy. Most businesses aren’t, which is exactly why the ones who do this stand out in tenders, in client conversations, and in how seriously people take their sustainability claims generally.
Related reading: fun chatgpt prompts.
Related reading: how to use whatsapp business on laptop.
Frequently asked questions
How much water does one ChatGPT question use?
Estimates range from OpenAI’s own figure of about one fifteenth of a teaspoon per query to independent university research suggesting a short 20 to 50 question conversation could use around 500ml, roughly a bottle’s worth. The gap exists because the two figures measure different things: a single query versus a full conversation, and direct cooling water versus the full electricity supply chain behind it.
Is ChatGPT worse for water use than other cloud software my business already uses?
Not clearly, no. Google searches, Zoom calls, cloud email, and video streaming all run on the same type of data centre infrastructure and use water the same way. ChatGPT gets singled out mainly because OpenAI and independent researchers have published specific numbers that other software vendors generally haven’t bothered to.
Do I need to mention AI water use in ESG reports or client tenders?
Increasingly, yes, if you’re bidding for public sector or larger enterprise contracts. A short, honest paragraph naming the AI tools you use and noting your providers’ published sustainability commitments is enough; leaving the question blank or writing “not applicable” tends to read as not having thought about it at all, which can cost you the tender.
Are AI companies doing anything to reduce water use?
Yes. Microsoft, Google, and OpenAI have all published water positive or efficiency targets, generally aiming to replenish more water than their operations consume by around 2030, and newer chips and cooling systems are already reducing the water needed per query compared with the models used in the 2023 studies that most headlines still quote.
