Apparently, I could whip out my phone, right now, and ask ChatGPT 38,000 questions before using as much water as it takes to grow a single California almond! At least, that’s the claim OpenAI boss Sam Altman made on the Sources podcast last week – and it sounds reassuringly precise for anyone concerned about the impact of their AI usage.
It’s also the sort of comparison that makes me immediately suspicious. There’s no single, satisfying answer to how much water or electricity “AI” uses, because asking a chatbot to fix a typo is very different from feeding a reasoning model a 200-page document. Images need more work again, and video is in another league entirely.
Still, we’ve got enough real-world measurements to get a decent idea. The TL;DR version? An ordinary text prompt probably uses less power and water than you think. The billions of prompts being processed every day, though, are where things start to get a little uncomfortable…
One prompt is surprisingly small
Altman said in 2025 that an average ChatGPT query uses around 0.34 watt-hours (Wh) of electricity and 0.000085 gallons of water – roughly 0.32ml. OpenAI hasn’t published the workings behind those numbers, so there’s no way to check how it defined an average query or what was included. Boo.
Google,at least, has been rather more forthcoming. Its 2025 study of Gemini Apps counted the AI accelerator, host computer, idle capacity, and data-centre overhead. It found that a median text prompt used 0.24Wh of electricity and 0.26ml of water for cooling – around five drops!
Google also compared the electricity use with watching nine seconds of television. Put another way, 1000 median Gemini prompts would use about 0.24kWh. At an illustrative electricity price of $0.25/25p per kWh, that’s around six cents/6p of electricity.
That doesn’t mean an AI service costs pennies to run, mind – chips, buildings, networking, and staff have a habit of wanting to be paid for. But the raw energy used by a quick text exchange can be surprisingly small. Google also says that improvements to its models, software, and hardware cut the energy required for its median prompt by a factor of 33 in a year.
Bigger jobs change the maths
The problem with “one prompt” is that it could mean almost anything. Epoch AI estimated that a typical GPT-4o query used about 0.3Wh, but a 10,000-token input – roughly a short paper or long magazine feature – could raise that to around 2.5Wh. Processing 100,000 tokens could need almost 40Wh before the answer is written.
Long answers take more work too, while reasoning models can carry out extra computation before replying. Agents might call several models and tools to finish what looks like one task from your side of the screen. One tap, in other words, doesn’t necessarily mean one job.
Pictures are heavier still. A study of 88 open AI models found that image generation used around 60 times more energy than text generation on average in its tests. It used older models running on one type of GPU, so that’s not a price tag you can slap onto a current ChatGPT or Midjourney image. It’s a useful clue, though.
There’s even less public data for consumer AI video. Generating moving footage means creating many frames, keeping objects consistent, and often trying several versions before one looks right. And (I know you won’t want to hear this), asking for another take because somebody’s acquired an extra finger isn’t environmentally free.
AI Water usage is harder to count
Servers produce heat, and data centres have to get rid of it somehow. Some cooling systems evaporate water, while closed-loop liquid systems reuse it. Air cooling can reduce direct water consumption, but could require more electricity instead. Climate and location make a difference too.
Then there’s the water used elsewhere. Power stations can consume it while generating electricity, and manufacturing chips and servers has its own footprint. That’s where apparently precise comparisons get particularly messy.
Google’s 0.26ml figure covers water consumed while cooling its infrastructure. A lifecycle assessment published by Mistral AI estimated 45ml for a 400-token response from Mistral Large 2, but included upstream impacts such as server manufacturing. Different model, different system, different boundaries.
That doesn’t make Mistral 173 times thirstier than Gemini though. It means the two figures are measuring different things. Until companies report the same information in the same way, ranking chatbots by the nearest millilitre is fairly pointless.
Where that water is consumed can matter more than the total, anyway. A litre used in a cool, water-rich region isn’t equivalent to a litre consumed during a drought – and users rarely know which data centre has handled their request.
Scale is the real problem
A few drops multiplied by a few questions is still very little. A few drops multiplied by billions of daily requests – with images, videos, and agents thrown in – is rather more significant.
The International Energy Agency estimates that all data centres used 415 terawatt-hours of electricity in 2024, or around 1.5% of global consumption. It expects that to more than double to roughly 945TWh by 2030, slightly more than Japan uses today, with AI the biggest driver of the growth.
Those figures cover all data-centre activity, including cloud storage, streaming, websites, and business systems. They aren’t an AI-only total. Even so, the IEA expects accelerated servers – mainly driven by AI – to account for almost half of the increase.
The pressure won’t be spread evenly, either. According to the IEA, a typical AI-focused data centre can use as much electricity as 100,000 households. The largest facilities being built could use 20 times as much. That’s enough to cause very local problems, even if the global percentage doesn’t sound terrifying.
Use AI where it counts
Feeling guilty about every useful chatbot question won’t fix much. But it does makes more sense to match the tool to the job.
Smaller or faster models are normally enough for summaries, rewrites, and simple questions. Deeper reasoning modes make more sense when the work benefits from them. Clear instructions can also avoid several rounds of regeneration, without turning prompt writing into a second job.
Images and video are worth being choosier about as well. Generating ten near-identical pictures for fun is likely to matter more than polishing one useful text response. Running AI locally isn’t automatically cleaner, either – it simply moves some of the electricity use to your own computer.
Most of the responsibility still sits with the companies building and buying these systems. Comparable reporting, efficient hardware, sensible locations, lower-carbon electricity, and cooling that suits the local water supply will make a far bigger difference than one person shaving a sentence off a prompt.
So, no, asking ChatGPT for a dinner idea isn’t the environmental equivalent of leaving the shower running for six hours. But while there’s absolutely no need for self flagellation, AI isn’t somehow free because one ordinary response can be measured in drops and fractions of a watt-hour, either. Use it when it earns its keep, go easy on the heavier stuff – and congratulations for even thinking about your consumption in the first place.
Read the full article here
