Does Character.AI Use Water? The Real Numbers for a Night of Roleplay (2026)

Yes, Character.AI uses water: every reply is generated in a data centre, and data centres run on electricity and are cooled with water. The real question is how much, and Character.AI publishes no figures at all. We checked all 116 articles in its help centre and found no mention of energy, water or carbon.

That does not leave the question unanswerable. Two AI companies have disclosed per-prompt figures for their own services, Character.AI has published unusually detailed information about how it serves chat, and one of its own numbers puts a ceiling on the answer. Put together, they give a range narrow enough to be useful, and they point to something the general “AI uses water” articles miss. The unit that matters for roleplay is not the message. It is the evening.

This is not an environmental essay. It is the arithmetic for one kind of use, long character roleplay, using figures anyone can check. All sources were checked on 19 September 2026, and where we calculate something ourselves, the working is shown.

The Short Answer

Per message: the best disclosed figures are a quarter to a third of a watt-hour of electricity and a quarter to a third of a millilitre of cooling water, a few drops.
Per heavy hour of roleplay (sixty generations, rerolls included): roughly 15 to 20 watt-hours and about a tablespoon of cooling water. That is about what a 10 watt LED bulb uses in two hours.
The high estimates, a bottle of water for every few dozen replies, come from counting more things (the water used at power stations) for an older model in an earlier year. They are not wrong, but they answer a different question.
What none of these figures include: training the models and manufacturing the chips. Those costs are real, and no one divides them up per message.

The rest of this page shows where each of those numbers comes from.

What Has Actually Been Disclosed

Only two companies running large chat services have published a per-prompt figure. They agree with each other more closely than the headlines suggest.

Google, August 2025. In a technical paper and an accompanying blog post, Google reported that the median Gemini Apps text prompt uses:

Measure	Comprehensive method	Narrow method (chips only)
Energy	0.24 Wh	0.10 Wh
Carbon	0.03 g CO₂e	0.02 g CO₂e
Water	0.26 mL	0.12 mL

The comprehensive figure is the one to use, and it is to Google’s credit that it published both. It counts not only the AI chips while they are working, but also the host machines, the idle capacity kept ready for spikes, and the data centre’s overhead for cooling and power distribution. Google’s own comparisons: the energy is less than watching nine seconds of television, and the water is about five drops. The paper also reports that the energy per median prompt fell 33 times over one year.

OpenAI, June 2025. In a post titled The Gentle Singularity, OpenAI’s chief executive, Sam Altman, gave the average ChatGPT query as about 0.34 watt-hours and about 0.000085 gallons of water, “roughly one fifteenth of a teaspoon”, which is about 0.32 millilitres. No methodology was published alongside it, so it carries less weight than Google’s paper. It lands in the same place, though: a third of a watt-hour and a third of a millilitre.

Two separately disclosed figures, from different companies with different hardware, within about 40% of each other. That is as solid as public numbers in this field get.

What Character.AI Has Said, and the Ceiling It Implies

Character.AI has never published an energy or water figure. It has published a lot about how it serves chat, and that turns out to matter more for roleplayers than any headline number.

In June 2024 the company’s research blog described its serving system in detail. Three statements stand out:

“Today we serve more than 20,000 inference queries per second.”

“On Character.AI, the majority of chats are long dialogues; the average message has a dialogue history of 180 messages.”

and, in the company’s announcement of the same work, that it could serve that volume “at a cost of less than one cent per hour of conversation”.

That last figure puts a ceiling on the energy, and the arithmetic is simple. Suppose every bit of that cent went on electricity, at 10 cents per kilowatt-hour. An hour of conversation could then use no more than 100 watt-hours. In reality most of what it costs to serve an AI model is the hardware, not the power, so the true figure sits well below that ceiling. Cheaper electricity would raise the ceiling, but not by enough to change the order of magnitude. This is a ceiling, not an estimate. What it rules out is the idea that an hour of Character.AI chat costs the planet something on the scale of a household appliance running all day.

Two cautions about dates. That cost figure and the architecture behind it are from 2024. In January 2026, a Character.AI post about new hardware said it “leverages multiple models like Qwen, Mistral and more”, and described serving Qwen3-235B, an open model, on AMD graphics processors. The service has changed since the 2024 post. What has not changed is the shape of the workload.

Why Roleplay Is Different From Asking a Question

The general articles on AI and water are about someone asking ChatGPT a question: a short prompt and a single answer. Character roleplay is a different workload, in two ways.

Every reply has to take in the whole conversation so far. A model has no memory between messages. To write reply number 181, it has to process the character’s definition and the previous 180 messages again, or as many of them as fit. That is why the 180-message figure matters: Character.AI was telling you, in 2024, that its typical request carries a long history, not a single question. Processing that history takes compute, and compute takes power.

Character.AI’s engineering post is frank about the cost:

“As dialogues grow longer, continuously refilling KV caches on each turn would be prohibitively expensive.”

Its answer was a system that keeps each conversation’s processed history in memory between turns, sends every message in the same chat to the same server, and reuses the stored work instead of redoing it. The post reports a 95% cache rate. In plain terms, the company built its infrastructure around the fact that its users hold long conversations. That is good for the footprint, and it is also the best evidence that a roleplay reply, served naively, is heavier than a one-line question.

Roleplay happens in sessions, not queries. Nobody asks Character.AI one thing and leaves. The per-message figure is tiny. The per-evening figure is that tiny number multiplied by however many replies you read, and every reroll counts as a full reply, whether or not you keep it.

The Arithmetic for One Evening

Here is a worked example with the assumptions stated, so you can substitute your own.

Assumption: one hour of active roleplay, with a reply every minute or so including rerolls, gives 60 generations. A slower player might produce 30. A player who rerolls a lot might produce more.

On the disclosed per-prompt figures (Google’s 0.24 Wh and 0.26 mL; OpenAI’s 0.34 Wh and about 0.32 mL):

	Low (Google)	High (OpenAI)
Energy per hour	60 × 0.24 = 14.4 Wh	60 × 0.34 = 20.4 Wh
Cooling water per hour	60 × 0.26 = 15.6 mL	60 × 0.32 = 19.2 mL

A tablespoon is about 15 millilitres, so a heavy hour of roleplay comes to roughly a tablespoon of cooling water.

To make the energy concrete, with our own arithmetic:

A 10 watt LED bulb left on for an hour uses 10 Wh, so a heavy roleplay hour is about one and a half to two hours of one light bulb.
Google equates one median prompt with under nine seconds of television, so sixty prompts is under nine minutes of TV.
A typical phone battery holds roughly 15 to 20 Wh, so a heavy roleplay hour is in the region of one phone charge. That counts only the data centre’s side, not the phone you are reading on.
Heating a litre of water from 20°C to boiling takes about 93 Wh of heat, so a heavy roleplay hour is about a fifth of one kettle.

Two hours every night for a year comes to about 730 hours. At 14 to 20 Wh an hour, that is roughly 10 to 15 kilowatt-hours a year, and roughly 11 to 14 litres of cooling water.

Is the per-prompt figure the right one for a long-context roleplay reply? It is the best available, and the direction of the error is not obvious. A long history makes each reply heavier to process. Caching, which Character.AI documents at 95%, takes most of that back. If you want a cautious version, double every number above. The conclusion survives it.

Why the Scary Numbers Are So Much Bigger

If you have read that AI “drinks a bottle of water” every few dozen replies, that claim has a real source, and it is not the same claim as the figures above.

It comes from academic research by Pengfei Li, Jianyi Yang, Mohammad A. Islam and Shaolei Ren, published in its final form in Communications of the ACM in 2025 as “Making AI Less ‘Thirsty’”. Its best-known finding is that GPT-3 needed “a 500ml bottle of water for roughly 10–50 medium-length responses, depending on when and where it is deployed”. That works out to 10 to 50 millilitres per response, between forty and two hundred times Google’s figure. Three differences account for the gap:

What is counted. The academic estimate includes water consumed at power stations to generate the electricity (which the paper calls scope-2 water), not only the water the data centre uses for its own cooling. Google’s 0.26 mL counts the data centre’s cooling water. Thermal power generation consumes a lot of water, so including it raises the figure a great deal. Both approaches are legitimate. They measure different things.
When. The academic figure was calculated for GPT-3, a 2020 model, on the infrastructure of the time. Google reports a 33-fold fall in energy per prompt in a single year. Serving has become far more efficient since the GPT-3 era, and Character.AI’s own post describes a 33-fold fall in its serving costs between late 2022 and mid-2024.
Where. The paper’s central point is that water use varies with location and even time of day, because climate and the local electricity mix change how much water a kilowatt-hour costs. The same reply can have very different water costs in two different data centres.

So the fair summary is this. Counting only the data centre, a roleplay reply costs a few drops. Counting the power station as well, it costs more, and how much more depends on where the data centre is and what its grid runs on. Neither you nor Character.AI’s users as a whole get to choose that grid, which is one reason the disclosure matters.

What the Per-Message Figures Leave Out

None of the numbers above, including Google’s, cover two costs.

Training. A model has to be trained before it can answer anything, and that is a large one-off expense. The same academic paper estimates that training GPT-3 in Microsoft’s U.S. data centres could have consumed 5.4 million litres of water in total, 700,000 of them on site. Training costs are shared across every reply the model ever gives, so they shrink per message as usage grows, but they are not zero, and no company publishes them in a per-message form. Character.AI now serves open models trained by other companies, which moves the training cost somewhere else but does not remove it.

Hardware. Manufacturing the chips and servers has its own energy and water cost, which the academic paper calls embodied water. Again, it is real and it is not divided up per reply in any published figure.

Images and video. Everything on this page is about text. Character.AI has been moving into image generation, animated avatars, comics and short-form video, a shift we cover in why Character.AI feels worse than it used to. Those are separate workloads, generally more compute-intensive per item than a text reply, and none of the figures above cover them.

Character.AI Compared With ChatGPT and Other Platforms

The search results show this is the question many people actually want answered, and the honest answer comes in two parts.

Per reply, there is no reason to expect Character.AI to be heavier than ChatGPT, and some reason to expect it to be lighter. Neither company publishes a per-message figure for its own service, so nobody outside them can settle it. But Character.AI has put a great deal of public effort into making replies cheap to serve, and it now describes running open models such as Qwen3-235B. That model’s full name, Qwen3-235B-A22B, records that it is a mixture-of-experts design which uses only about 22 billion of its 235 billion parameters for each token. Cheaper to serve generally means less power per reply.

Per person, a heavy roleplayer probably uses more than a typical ChatGPT user, simply because of volume. Someone who asks ChatGPT five questions a day and someone who roleplays for two hours a night are not doing the same thing. On this question, what you do with the platform matters more than which platform you use.

The same logic applies to every other platform in our directory. A platform that runs a large frontier model for every turn, with a long memory system feeding each reply, is doing more work per turn than a small chat model. That is often exactly why its stories are better. No roleplay platform we cover publishes environmental figures, and we do not estimate what they have not disclosed.

What You Can Actually Change

If the footprint matters to you, these are the levers that actually move the numbers, from biggest to smallest:

Reroll less. Every swipe or regenerate is a complete new reply, whether you keep it or not. Rerolling ten times to find the right line costs ten replies. This is the only lever that can change your total several times over.
Prefer shorter replies. Producing text is the expensive part of a reply, so a model asked for three paragraphs does more work than one asked for one.
Be deliberate with images and video. They are heavier workloads than text and sit outside every figure on this page.
Do not assume running a model at home is greener. A data centre serves thousands of conversations at once on the same chips, which is a large part of why its per-reply figure is so low. A home graphics card serves one. As an illustration with our own arithmetic: a 300 watt card taking 20 seconds to write a reply uses about 1.7 Wh, several times the data-centre figure. Local models have real advantages, covered in our local models guide, but efficiency per reply is not usually one of them. Where the power comes from matters too: a home on a clean electricity supply changes the picture.
Weigh it against the evening it replaces. An hour of text roleplay costs about a fifth of a kettle and under ten minutes of television. Whether that is a lot depends on what else you would have done with the hour.

What Character.AI Could Publish

Google showed that this can be disclosed, with a published methodology, in a form ordinary users can understand. Character.AI already publishes detailed engineering posts about how it serves chat, and it knows its own cost per hour of conversation, because it has quoted it. A per-message energy and water figure, with the method stated, would answer one of the most-searched questions about the platform in a single paragraph.

Until it does, the defensible answer is the one on this page: a few drops per message, about a tablespoon per heavy evening at the data centre, more once the power station is counted, and most of the difference comes down to how many times you press reroll.

For the rest of what has changed at Character.AI, our troubleshooting guide covers what now looks broken but is not, and the alternatives list is organised by what you would be leaving for.

Frequently Asked Questions

Does Character.AI use water? Yes. Every reply is generated on servers in a data centre, and data centres use electricity and, for cooling, water. Character.AI publishes no figure of its own, but two AI companies have disclosed theirs: Google says its median Gemini text prompt consumes 0.26 millilitres of water, about five drops, and OpenAI’s chief executive has put an average ChatGPT query at roughly one fifteenth of a teaspoon. A Character.AI reply is the same kind of workload, so a figure in the region of a few drops per message is the reasonable expectation, with the caveats this page explains.

Is Character.AI bad for the environment? Per message, the impact is very small: the best disclosed figures are a fraction of a watt-hour of electricity and a few drops of water. What adds up is volume. Roleplay is long, and a heavy evening can run to dozens of replies plus rerolls. On the disclosed per-prompt figures, an hour of heavy roleplay comes to roughly 15 to 20 watt-hours, about what a 10 watt LED bulb uses in two hours, and roughly a tablespoon of cooling water. That is modest next to most household energy use, but it is not zero, and it excludes training the models and building the hardware.

Does Character.AI use more water than ChatGPT? Nobody outside the two companies can say for certain, because neither publishes per-message figures for its own service. Per reply, there is no reason to expect a Character.AI message to be heavier: the company has published how much engineering it put into making replies cheap to serve, and it says it now runs open models such as Qwen and Mistral. Per person, it may well be more, because roleplay sessions are long. Character.AI has said the average message it serves comes with a dialogue history of 180 messages, and people roleplay for hours rather than asking one question.

How much water does Character.AI use a day? There is no published figure for the company as a whole, and any number you see for its total daily water use is an estimate built on assumptions. For one person, the arithmetic is simpler. Using Google’s disclosed figure of 0.26 millilitres of cooling water per text prompt, sixty replies in an evening comes to about 16 millilitres, roughly a tablespoon. Estimates that also count the water used to generate the electricity come out much higher, which is the main reason published numbers disagree so widely.

Why do estimates of AI water use differ so much? Mainly because they count different things. Google’s figure of 0.26 millilitres per prompt counts the water its data centres consume for cooling. The widely quoted academic estimate of a 500 millilitre bottle for every 10 to 50 responses also counts water consumed at power stations to generate the electricity, and it was calculated for GPT-3 in 2023. Google separately reports that the energy per median Gemini prompt fell 33 times in a single year. Different scope, different year and different location easily add up to a hundredfold gap.

How can I reduce the environmental impact of my AI roleplay? The biggest lever is the one you control directly: every reroll or swipe is a complete new generation, so rerolling ten times costs ten replies. Shorter responses also help, because producing text is the expensive part. Image and video generation are separate and heavier workloads than text chat. Running a model on your own computer is not automatically greener either: a home graphics card generating one reply at a time is usually less efficient per reply than a data centre serving thousands of chats at once.

Sources, checked 19 September 2026: Google, “Measuring the environmental impact of AI inference” (Google Cloud blog, 21 August 2025) and the accompanying paper “Measuring the environmental impact of delivering AI at Google Scale” (arXiv 2508.15734); Sam Altman, “The Gentle Singularity” (June 2025); Li, Yang, Islam and Ren, “Making AI Less ‘Thirsty’” (Communications of the ACM, 2025; arXiv 2304.03271); Character.AI, “Optimizing AI Inference at Character.AI” (research and company blogs, 20 June 2024) and its January 2026 post on serving with DigitalOcean and AMD. Worked figures marked as our arithmetic are ours. Character.AI is a trademark of Character Technologies, Inc.; Arcanum is an independent publication with no affiliation to it or to any company named here, and references are nominative.