Nobody outside OpenAI knows exactly how much water one ChatGPT message uses, and the published numbers sit further apart than you’d expect. OpenAI’s own figure is about a fifteenth of a teaspoon per message. Research accepted by a major computing publication put an older model at a 500 ml bottle for every 10 to 50 replies. Neither side is making anything up. They are counting different things.

Here is the short version:

• OpenAI’s number: about 0.000085 gallons per average query, which works out to roughly 0.32 ml.
• The research number: GPT-3 consumed a 500 ml bottle per 10 to 50 medium-length responses, so somewhere between 10 and 50 ml each.
• Why they differ: definitions, mostly. What counts as water being “used”, and where the data center happens to sit.

Why are the published numbers so far apart?

Because “water used” is four questions wearing one coat, and different researchers answer different ones.

The first is whether you mean water taken or water used up. Scientists separate water withdrawal, the freshwater taken from the ground or a river, from water consumption, defined as withdrawal minus whatever gets discharged back. A cooling tower takes water, evaporates some and returns the rest. With good water quality, roughly 80% of what it takes is evaporated.

The second is whether you count the power plant. Servers need electricity, and generating electricity uses its own water. The figures below are per kilowatt-hour, the same unit that turns up on a household electricity bill. Those two numbers are nowhere near each other:

Where the water goes	Taken (withdrawal)	Used up (consumption)
Cooling the data center	What the tower adds, including water it later discharges	1 to 9 liters per kWh of server energy
Generating the electricity	43.8 liters per kWh (US average)	3.1 liters per kWh (US average)

Look at the bottom row. A writer who counts the water a power plant takes gets 43.8 liters per kilowatt-hour. A writer who counts what it uses up gets 3.1. That is more than ten times the difference, and both of them can say “water used by AI” with a straight face.

The third question is which model answered you, and how long the answer was. A quick reply from a small model and a long one from a heavyweight are not the same amount of work.

The fourth is the one almost nobody mentions. Where is the building, and what month is it? The same study behind the viral bottle claim estimates that data centers evaporate somewhere between 1 and 9 liters per kilowatt-hour of server energy. The bottom of that range is Google’s average across a whole year, worldwide. The top is a large commercial data center during an Arizona summer. Nine times the water for an identical message, decided by weather and a map.

Where does the “bottle of water per email” claim come from?

It comes from a Washington Post article published on 18 September 2024, headlined “A bottle of water per email: the hidden environmental costs of using AI chatbots”. The article is behind a paywall. Its summary line is not, and it says what was measured: the Post worked with researchers at the University of California, Riverside to work out how much water and power ChatGPT, running GPT-4, uses to write an average 100-word email.

So the famous number came from analysis the paper commissioned for itself, not from a study it was quoting. We could not read the figure it arrived at, so we are not repeating it here.

What we can read is the published research from the same group: Making AI Less “Thirsty” by Pengfei Li, Jianyi Yang, Mohammad A. Islam and Shaolei Ren, accepted by Communications of the ACM. On the cost of running the model, its words are: GPT-3 “needs to ‘drink’ (i.e., consume) a 500ml bottle of water for roughly 10 – 50 medium-length responses, depending on when and where it is deployed.”

Look at how much moves between those two. The newspaper measured GPT-4 writing a fixed 100-word email. The study measured GPT-3 across replies of no fixed length, and says outright that the answer depends on when and where the data center runs. Neither one is the other’s headline, and the range inside the study is already fivefold before the weather gets a say.

The same study is where the training figures come from, and those get mangled too. It says training GPT-3 in Microsoft’s US data centers can consume 5.4 million liters in total, including 700,000 liters of on-site water. The 700,000 travels well on its own, quoted as though it were the whole cost of training. It is not, and both numbers sit in the same sentence of the same study.

What has OpenAI actually said?

One sentence, in a blog post about something else. In his 2025 essay “The Gentle Singularity”, Sam Altman wrote that the average query uses about 0.34 watt-hours of electricity, a watt-hour being one thousandth of the kilowatt-hour above, and “about 0.000085 gallons of water; roughly one fifteenth of a teaspoon”.

Convert the gallons and you get about 0.32 ml, which matches his teaspoon comparison. That is the arithmetic checking out, not the claim.

The claim itself cannot be checked, because it arrives without a method. The essay does not say what “average” means, which models it covers, or whether the electricity behind the answer is included. It is a parenthesis in a piece about superintelligence. That doesn’t make it wrong, and it is the only figure OpenAI has given. It does mean nobody outside the company can reproduce it.

What does Google say about Gemini?

Google published a number too, and more usefully, it published two.

In its write-up of how it measures AI inference, Google estimates the median, or middle-of-the-road, Gemini Apps text prompt at 0.24 watt-hours of energy and 0.26 ml of water, which it describes as about five drops, using data from May 2025. Then it shows what happens under a narrower method that counts only the chips actively working: 0.12 ml.

Same company. Same prompt. Half the water. The only thing that changed is where the boundary was drawn.

That single comparison explains the entire mess better than any argument about who is exaggerating. If one company can honestly publish two numbers that differ by more than half, two organizations measuring different systems can differ by far more without either one lying. It is also the sort of detail you only get when a company shows its work, which is what makes the Gemini figure easier to trust than the ChatGPT one.

If you use both tools, our comparison of ChatGPT and Gemini covers the everyday differences.

Why are the daily totals so enormous?

Because they are a small number multiplied by two guesses.

Take a widely circulated graphic from Business Energy UK, which reports ChatGPT using around 39.16 million gallons of water a day. Their methodology section is open about how they got there: they took a reported 400 million weekly users, divided down to a daily figure, assumed five prompts per person per day, and multiplied by a per-response water estimate.

Credit where it’s due, they publish those assumptions. The trouble is that the headline number travels and the assumptions stay home. Change the five prompts to two, or swap the per-response estimate for OpenAI’s, and the graphic looks nothing like it does now.

That user count has also aged. OpenAI said in February 2026 that ChatGPT had passed 900 million weekly users. Run the same arithmetic with today’s figure and the headline more than doubles, without anyone measuring a single extra drop.

The projection with real work behind it is deliberately broader. The same Communications of the ACM study estimates that global AI demand could account for 4.2 to 6.6 billion cubic meters of combined water withdrawal in 2027, which it puts at more than the total annual withdrawal of four to six Denmarks, or half of the United Kingdom. That is the whole industry, not one chatbot.

So does your own use matter?

By every published estimate, the water behind your individual messages is small. A fifteenth of a teaspoon and 50 ml are both small when the thing you are comparing them to is a shower.

The concern the researchers actually raise is not about your message count. It is about concentration: data centers cluster in particular places, some of them already short of water, and they draw hardest at the hottest times, which is exactly when a town has least to spare. That is a planning argument rather than a personal one, and it doesn’t resolve into advice about whether to send one more prompt.

What we would push back on is the confident number, from either direction. If you see a single figure for AI’s water use with no mention of what it counts or where the building is, it isn’t a measurement. It is one of several defensible numbers, picked. For more claims like this one, see our look at the AI myths that keep going around, and if you are curious what else happens after you press Enter, we traced where your messages actually go. Both sit in our AI safety section, along with the rest of the questions people actually worry about.

What do we still not know?

OpenAI has not published a methodology, so its figure cannot be independently checked. The Washington Post’s per-email number sits behind a paywall, so we have not verified it and have not repeated it here. And these figures move fast: Google reported that the energy behind its median text prompt fell 33-fold over a recent twelve-month period, which tells you how quickly any of this ages.

Anyone quoting a precise water cost for one ChatGPT message, to two decimal places, with no caveats, is telling you more about their confidence than about the water.

Sources

• Making AI Less “Thirsty”: Uncovering and Addressing the Secret Water Footprint of AI Models — Li, Yang, Islam and Ren; accepted by Communications of the ACM. Source of the 500 ml claim, the 1 to 9 L/kWh range, the training figures and the 2027 projection
• The Gentle Singularity — Sam Altman’s essay containing OpenAI’s 0.34 Wh and 0.000085 gallon figures
• Measuring the environmental impact of AI inference — Google’s methodology and its 0.26 ml and 0.12 ml Gemini figures
• ChatGPT Energy Consumption Visualized — Business Energy UK, including the methodology section behind its daily total
• A bottle of water per email — The Washington Post, 18 September 2024. Paywalled; we cite only its headline, date and public summary line
• Scaling AI for everyone — OpenAI, 27 February 2026, source of the 900 million weekly users figure

Frequently asked questions

So is it a teaspoon or a bottle?

Both, depending on what you count. OpenAI's figure of about a fifteenth of a teaspoon per message and the research figure of a 500 ml bottle per 10 to 50 replies are not answers to the same question. One is a company describing an average message on its current models. The other is a study of GPT-3, an older model, and it says plainly that the figure moves depending on when and where the data center is running. OpenAI hasn't published a method, so nobody outside the company can line the two up.

Does the water disappear?

Some of it does, in the sense that matters for a water bill. Researchers separate water withdrawal, meaning water taken from a river or the ground, from water consumption, meaning the part that evaporates and does not come back to the same place. A cooling tower discharges some of what it takes and evaporates the rest, and with good water quality about 80% of what it takes is evaporated. The water rejoins the water cycle eventually, but not necessarily in the town that supplied it.

Does one ChatGPT message really use a whole bottle of water?

No, and this is the most common way the figure gets garbled. The 500 ml bottle in the research covers 10 to 50 replies, not one, which puts a single reply somewhere between 10 and 50 ml on that estimate. It also describes GPT-3, an older model, rather than whatever answered you today. A bottle per message is roughly ten to fifty times what even the least flattering published study claims.

Does ChatGPT use more water than Google Gemini?

There is no clean way to compare them, because the two companies have not shown the same amount of work. Google puts a middle-of-the-road Gemini text prompt at 0.26 ml using data from May 2025, and it published the method behind that number. OpenAI's figure of roughly 0.32 ml arrived without one. The two are close enough that the gap between them is smaller than the uncertainty in either, and we can't tell whether they count the same things.

Is the water for cooling the computers or for making the electricity?

Both, and that split is where most of the arguing comes from. Some of the water evaporates in the data center's cooling system, and more is used at the power station generating the electricity those servers run on. An estimate that counts only the cooling will always be smaller than one that counts the power plant too. Neither is dishonest, but they are answers to different questions, and headlines rarely say which one they picked.