A single AI chat uses a small amount of energy and water — Google and OpenAI's own disclosed figures put one typical prompt at well under half a watt-hour and a few drops of water, roughly the impact of a few seconds of TV. The real environmental question isn't any one chat, though — it's the fast-growing number of large data centers being built to run AI at a massive scale, which is measurably adding to electricity demand and, in some regions, to local water use.
If you've seen headlines saying a single ChatGPT question "drinks" a bottle of water, or that AI is quietly draining the power grid, it's fair to wonder whether you should feel bad about using it. The honest answer is more reassuring than the scariest headlines, but also more complicated than "don't worry about it." Here's what's actually been measured, and by whom.
Where the Scary Numbers Came From
Much of the alarm traces back to a 2024 estimate from a UC Riverside researcher, widely reported as: writing a short, roughly 100-word response with ChatGPT could consume around 500 milliliters of water — about a standard bottle. That number spread fast because it's vivid and easy to picture.
There's an important update to that story that gets shared far less often. In 2026, the same researcher, Shaolei Ren, revised his own estimate down significantly — to roughly 15 milliliters for a typical modern prompt — pointing to more efficient chips and cooling systems, and noting that the original figure was based on specific 2024 assumptions that don't reflect how today's systems run. That's a more than 30-fold drop in the estimate, from the same source.
What AI Companies Themselves Have Disclosed
The most solid numbers available are the ones companies have published about their own systems, since they can measure their actual hardware directly instead of estimating it from outside.
In 2025, Google released a technical report on the environmental cost of a typical Gemini Apps prompt: about 0.24 watt-hours of electricity and about 0.26 milliliters of water — roughly five drops. For comparison, Google says that's about what a TV uses in under nine seconds. OpenAI CEO Sam Altman published a similar breakdown for ChatGPT the same year: an average query uses about 0.34 watt-hours of electricity and a tiny fraction of a gallon of water, which he likened to about one second of an oven running.
Those figures aren't independently audited, and both companies chose how to define an "average" query. But they're a real data point from the source, and they land in the same small ballpark as each other — nowhere near a bottle of water per question.
So Why Do the Numbers Vary So Much?
Because "how much water does AI use" isn't one question — it's several, and each study answers a different one. Some estimates count only the water evaporated to cool the computer chips right there in the data center. Others add the water used miles away, at the power plant generating the electricity the data center runs on — which can be a much bigger number depending on how that electricity is produced. Prompt length, which AI model answers it, and how efficient a given data center is all shift the result further. None of the studies are necessarily "wrong" — they're measuring different slices of the same system, which is exactly why the same underlying question can produce numbers 30 times apart.
The Part That Actually Matters More: Scale
Here's the more useful way to think about it. A single question to an AI chatbot genuinely is small — even the higher estimates amount to a few sips, not a bottle, per typical use. But AI is being asked billions of times a day, and more importantly, tech companies are building enormous new data centers specifically to run it.
The International Energy Agency, which tracks global energy use, estimates data centers currently account for roughly 1.5 to 2% of global electricity consumption — and that AI-focused data centers saw their electricity use jump about 50% in 2025 alone, far outpacing data center growth overall. That share is projected to approach 3% of global electricity by 2030. In some places, new data centers are also drawing on local water supplies in regions that are already short on water, which is a legitimate concern for the communities nearby, separate from what any single chat "costs."
In other words: the honest environmental story about AI isn't really about you asking a question. It's an infrastructure story — about how fast these data centers are being built, where their power comes from, and how transparent companies are about both.
What You Can Actually Do With This
If this is something you care about, the individual-chat-by-chat approach isn't where your effort matters most — the per-query numbers are small enough that trimming your own usage won't move the needle much. What does help, if you want to act on it: support and favor companies that publish real efficiency data rather than staying silent, and keep an eye on local reporting if a data center is proposed near you, since that's the level where water and power use genuinely add up to something worth weighing in on.
What to try next: For a broader look at how these tools actually work under the hood, What Does 'AI' Actually Mean? is a good next stop. And if you want to sort AI facts from AI hype more generally, Common AI Myths, Debunked covers more of the claims worth double-checking.
Sources
- Measuring the environmental impact of AI inference — Google Cloud Blog
- In a first, Google has released data on how much energy an AI prompt uses — MIT Technology Review
- Sam Altman: ChatGPT queries consume 0.34 watt-hours of electricity and 0.000085 gallons of water — Data Center Dynamics
- Shaolei Ren revises widely cited AI water estimate from a full bottle to 15 milliliters per prompt — Digg
- Data centre electricity use surged in 2025, even with tightening bottlenecks driving a scramble for solutions — International Energy Agency
