How Much Water Does ChatGPT Use?

How much water does ChatGPT use? OpenAI’s own figure, stated by CEO Sam Altman, is about 0.000085 gallons per query, roughly 0.32 mL or one fifteenth of a teaspoon, alongside 0.34 watt-hours of electricity. OpenAI has published no newer figure as of November 2026, including for the GPT-6 models now running ChatGPT. Independent estimates that count only data-center cooling land a little higher, around 1.5 to 4 mL per prompt. The scary viral number, about 519 mL (a full bottle) to write a 100-word email with GPT-4, is not a contradiction. It counts the entire electricity and training supply chain, which is why credible figures span more than 150x.

This guide gives you a dated, sourced answer for every version of the question, so per query, per prompt, per question, per day, and per year. You will see why estimates differ so wildly, why long reasoning answers and big documents cost far more than a quick reply, and what the newest data centers change. You will also see how a single prompt compares to a burger or a cotton t-shirt, and the one lever that actually changes your footprint. We lead with the data, not the panic, and we flag what the numbers do not yet show.

The Key Takeaways

• OpenAI’s official figure: ~0.32 mL per query. Sam Altman states an average ChatGPT query uses 0.000085 gallons of water (about one fifteenth of a teaspoon) and 0.34 watt-hours of electricity.
• Independent cooling-only estimates: ~1.5 to 4 mL per query. Only about 15% of the water tied to a request is on-site cooling; the rest is electricity generation and training.
• The viral “bottle per email” number is ~519 mL for a 100-word GPT-4 email, from UC Riverside via the Washington Post. It bundles the full lifecycle, which is why it is over 150x the cooling-only figure.
• Heavier answers use more water. A medium GPT-5 reply was estimated at about 18 watt-hours (up to 40) versus about 2 Wh for GPT-4, and a 100,000-token document input at almost 40 Wh. Nobody has published a figure for the GPT-6 models yet.
• Per prompt it is negligible; at scale it is not. If 16 million Americans wrote one ChatGPT email a week for a year, that alone would use more than 435 million liters of water.
• New data centers are switching to closed-loop cooling. OpenAI says its Stargate Michigan site, announced in June 2026, will use about as much water as a typical office building. That is a company design figure, not an audit.

How Much Water Does ChatGPT Use Per Query, Prompt, or Question?

A single ChatGPT query, prompt, or question uses somewhere between about 0.32 mL and a few milliliters of water for an ordinary text reply, with heavier reasoning answers pushing into the tens of milliliters. The reason there is a range and not one tidy number is that different estimates draw the boundary in different places. The lowest figure is OpenAI’s own, the middle figures count only data-center cooling, and the famous 519 mL “bottle” number counts the electricity and training behind the answer too.

OpenAI’s official figure: about 0.32 mL per ChatGPT query

In his June 2025 essay The Gentle Singularity, Sam Altman wrote that “the average query uses about 0.34 watt-hours” and “about 0.000085 gallons of water; roughly one fifteenth of a teaspoon.” That works out to roughly 0.32 mL per question, the smallest credible figure in circulation and the only one from OpenAI itself. It is a useful anchor because it comes straight from the company, and the energy half lines up with independent work.

The catch is that Altman never defined “average,” never published a methodology, and did not say whether training or electricity generation were included. Independent researchers therefore treat it as a plausible floor for on-site use rather than a full accounting. Epoch AI’s own analysis lands close on energy, estimating a typical GPT-4o query at about 0.3 watt-hours. That is ten times lower than the widely repeated 2023 estimate, so the old “ChatGPT uses ten times a Google search” line is outdated.

The viral “bottle of water per email” claim: about 519 mL

The number behind every alarming headline traces to “Making AI Less Thirsty,” a 2023 study by Pengfei Li, Shaolei Ren and colleagues at UC Riverside. They estimated that GPT-3 “drinks” roughly a 500 mL bottle for every 10 to 50 medium responses, depending on when and where the data center runs. In September 2024 the Washington Post, working with Ren, restated it as about 519 mL of water to write a 100-word email with GPT-4, which is the version that went viral.

That figure is not wrong, it is just wide. It measures cooling water plus the water used to generate the electricity, and in some framings the chip-manufacturing and training water too. The same study estimated that training GPT-3 in Microsoft’s US data centers directly evaporated about 700,000 liters of clean freshwater. You can check the math in the original UC Riverside paper, but the headline takeaway is simpler, so a per-email “bottle” claim is a lifecycle number, not a cooling number.

Ren himself now warns against reading it as today’s number. In June 2026 he told the Deseret News that the figure should not be treated as a measure of current AI water demand: “We cannot just use a number from two years ago to describe today’s system.” The same report notes that about half of the bottle is water evaporating off hydropower reservoirs, not data-center cooling.

Do newer ChatGPT models use more water?

Probably more per answer, but nobody has measured the current models. ChatGPT now runs on OpenAI’s GPT-6 family, and OpenAI has not published an energy or water figure for any of them. The best public evidence is one generation old. Epoch AI puts a typical GPT-4o reply near 0.3 watt-hours. In August 2025 the University of Rhode Island’s AI lab estimated that a medium GPT-5 response averages about 18 watt-hours and can reach 40, against roughly 2 watt-hours for GPT-4, as DatacenterDynamics reported. Per token, each generation gets more efficient. Per answer, reasoning models do far more work, so the net direction has been up.

Because cooling water tracks energy closely, an answer that costs eight to nine times the electricity costs roughly that much more water. So the honest framing is that a quick reply from a light model is a few drops, while a long reasoning answer from a frontier model is a meaningful multiple of that. For comparison, Google’s official figure for a median Gemini text prompt is 0.26 mL, in the same ballpark as ChatGPT’s official number. The model and task you pick matter far more than the brand on the logo.

Long documents are a hidden multiplier

The length of what you paste in matters too. Epoch AI estimates that a 10,000-token input, about a long magazine article, raises the cost of a query to around 2.5 watt-hours. A 100,000-token input, roughly 200 pages, takes almost 40 watt-hours. That is an upfront cost: once the document is in the chat, follow-up questions do not pay it again. For the electricity side in full, see how much energy ChatGPT uses.

Why Do ChatGPT Water Estimates Vary So Wildly?

Same prompt, two honest answers, a gap of more than 150x. Here is why that happens.

If you have seen “a teaspoon” and “a whole bottle” for the same ChatGPT prompt, neither side is lying. They measure different systems, and once you see the boundary the figures stop fighting each other. On-site water is what a data center evaporates to cool its chips. Indirect water is what power plants consume to generate the electricity those chips run on, which is often larger than the cooling itself.

Analysis by writer Andy Masley, drawing on Ren’s data, estimates that only about 15% of the water attributable to a request is actual on-site cooling, with the other 85% sitting in electricity generation and training. Apply that split and the cooling-only cost of a query falls to roughly 1.5 to 4 mL, which is about 150 to 300 times smaller than the viral 500 mL-per-message version. Neither number is fake; they answer different questions.

The practical rule is to always ask what is being counted, where, and when before repeating a figure. A cooling-only number from a cool, water-rich region in winter and a full-lifecycle number from a hot, fossil-powered grid in summer can differ by an order of magnitude for the exact same prompt. Understanding how AI systems actually work under the hood makes that spread a lot less mysterious. The same scope problem runs through every AI carbon footprint figure.

How Much Water Does ChatGPT Use Per Day and Per Year?

The short version is that nobody has an exact total, and that is the point.

There is no official per-day or per-year water total for ChatGPT alone, and anyone quoting one to the liter is guessing. What we do have is a credible scale projection. The Washington Post and Shaolei Ren estimated that if 16 million Americans used ChatGPT once a week to write an email for a year, that single habit would consume more than 435 million liters of water, roughly what every household in Rhode Island uses in a day and a half.

Zoom out to AI water use overall and the trajectory is steep. The UC Riverside team projects global AI water withdrawal of 4.2 to 6.6 billion cubic meters by 2027, more than the total annual withdrawal of several small countries. ChatGPT is only part of that, but it is one of the largest single drivers, and inference, the everyday business of answering prompts, is now the dominant and fastest-growing share rather than one-time training. The story is the curve, not any one year’s figure.

Is ChatGPT’s Water Use Actually a Big Deal?

For your personal usage, no. At OpenAI’s official figure, a single beef burger, which takes more than 600 gallons of water to produce, equals well over a million ChatGPT queries, and a cotton t-shirt at more than 700 gallons is in the same territory. Those numbers come from Water Footprint Network figures of 16,000 liters per kilo of beef and 2,700 liters per cotton shirt. By that yardstick, asking a chatbot a question is a rounding error against everyday consumption, and guilt over individual prompts is misplaced.

The legitimate concern is concentration and timing, not the per-prompt cost. Data centers cluster in a handful of regions, often water-stressed ones, and draw heavily at specific hours. A facility pulling millions of gallons a day in Arizona or Texas is a real local problem even while it stays a global rounding error. Both things are true at once, which is why a good answer to “is ChatGPT’s water use bad” is “not for your usage, but yes for where the buildout is concentrated.”

What new AI data centers do differently

The newest AI campuses are designed to use far less water than the evaporative-cooled sites behind the 2023 estimates. OpenAI broke ground on Stargate Michigan, a 1 GW site, in June 2026. It says the closed-loop cooling system needs about as much water as a typical office building. Oracle’s fact sheet for its Stargate sites with OpenAI puts cooling-system maintenance at about 12,625 gallons per building per year.

Two caveats keep this from being the whole answer. These are company design figures, not audited consumption, and closed-loop cooling does nothing about the water power plants use to make the electricity, which is most of the lifecycle total. Microsoft, OpenAI’s longtime cloud partner, says it replenished more water than it withdrew for the first time in its 2025 fiscal year, more than 14 million cubic meters. In the same 2026 sustainability update it concedes that sustainability solutions “are not scaling fast enough to meet demand.”

How to Cut Your ChatGPT Water Footprint

The single biggest lever you control is model choice, because the spread between a light model and a frontier reasoning model is enormous. Sending a simple question to a heavyweight model can use close to ten times the energy and water of a lean model that answers it just as well. Matching the model to the task beats any amount of prompt-trimming, and comparisons like DeepSeek vs ChatGPT and our guide to the most capable AI models show how different the efficiency profiles really are. For the greenest tools to switch to, see our guide to the most eco-friendly AI.

In practice that means defaulting to a lighter model for routine questions and reserving the heaviest reasoning mode for problems that need it. Batch related questions into one session instead of many cold starts, and resist regenerating an answer reflexively, since each retry is a fresh query. If you work with a long document, upload it once and keep asking in that chat, because the cost of reading it is paid up front. None of this requires guilt; it just requires using the heavy option on purpose rather than by default.

Lighter-footprint chatbots

If you want to go a step further, our EcoGPT review covers a tree-planting AI chatbot that runs small open-source models on Google Cloud servers in Finland and funds a tree for every 200 prompts, pairing the lighter-model approach with active environmental offsetting. Another lighter-footprint option is covered in our GreenPT review, a privacy-first chatbot that leans on energy-efficient models. If you search with Ecosia, our Ecosia AI review covers how its built-in chat works and what it says about water.

Run the light and the heavy model from one app

Matching the model to the question only works if switching is easy. With Fello AI you don’t have to pick one chatbot for everything. GPT, Claude, Gemini, DeepSeek and other models sit in one app on Mac, iPhone and iPad. Sending each question to the model that fits it is a click, not a second subscription. Our getting started guide shows how the model picker works.

Conclusion

So how much water does ChatGPT use? OpenAI’s official answer is about 0.32 mL per query, independent cooling-only estimates put it near 1.5 to 4 mL, and the viral 519 mL “bottle per email” figure is a full-lifecycle measure rather than a contradiction. Heavier reasoning answers and long documents lift those numbers. Nobody has yet measured the GPT-6 models now running ChatGPT. The real story is the scale curve, not any single prompt. The most useful next step is not to stop using ChatGPT; it is to be deliberate about which model you reach for, starting with a clear look at how the leading models differ.

FAQ