The Hidden Cost of a Single Answer
You type a question into an AI tool and an answer appears in a second or two. It feels like nothing, like magic. The answer did not come from your phone. It came from a building the size of a warehouse, packed with computers, humming day and night, drinking power and water to stay cool.
One question is tiny. The trouble is there are billions of them a day, and the buildings that answer them are now big enough to strain the power grid of a whole town. Every answer has a cost. You just never see where it gets paid.
What One Answer Costs
Start with a single query. In 2025, OpenAI CEO Sam Altman put numbers on it. By his account, an average ChatGPT question uses about 0.34 watt-hours of electricity and about 0.000085 gallons of water. Put another way, that is roughly what a kitchen oven burns in a second, and about a fifteenth of a teaspoon of water.
On its own, that is not a lot, which was Altman’s whole point. Take the figures with a grain of salt. It was OpenAI’s math, and likely not independently checked. Researchers at UC Riverside who studied the question on their own put the water cost far higher, closer to a 500ml bottle for every ten to fifty questions. OpenAI also never defined an average question, and its numbers leave out the huge one-time cost of training the model.
The electricity runs the computers. The water is less obvious. Data centers run hot, and many are cooled with water, evaporated away to carry off the heat, the same way sweat cools you down. So every answer spends a little power and a little water, out of sight.
Not Every Answer Costs the Same
The calculation was for a plain text reply. Ask for something heavier and the cost climbs fast. Researchers at Hugging Face and Carnegie Mellon measured it and found that creating a single AI image can take around sixty times the energy of a text answer. The most demanding image models burn through roughly half a phone charge for one picture, while a text reply is a sliver of that.
Then there is the cost before you ever type a word. Training a model happens once, at a scale that dwarfs any single query. One study estimated that training GPT-3 evaporated about 700,000 liters of fresh water, close to 1.4 million bottles, just to teach the model before a single person used it. The quick text answer is the cheap end of the range. The images, the video, and the training behind all of it sit at the expensive end.
Why this all Adds Up
A fifteenth of a teaspoon sounds like a rounding error. Multiply it by the traffic, and it stops being so small.
Hundreds of millions of people now use these tools, firing off billions of questions a day, and that is before you count AI writing code, generating images, and running inside other apps. All of it lands in the same place: data centers, and a lot more of them. Building and running them has pushed up electricity demand so fast that in some regions it is straining the grid and pushing utilities to lean on gas and other fossil fuels to keep up. The International Energy Agency expects data centers to roughly double their electricity use by 2030, climbing from about one percent of the world’s power to three.
This is where it stops being abstract and lands on your street. A data center moving into your area can raise local power demand, compete for your region's water, and show up on the public utility bills everyone shares. The free answer on your screen is paid for in part by the grid your home runs on.
None of this is a reason to feel guilty for asking AI a question, or to expect the lights to flicker. Data centers are infrastructure, the same kind that already sits behind streaming and online shopping, and the companies building them face growing pressure to use cleaner power and pay their own way. The point is not alarm. It is to see the real price behind an answer that feels free, and to keep it in mind.
Next week: Everyone keeps saying AGI is coming. What does that even mean?
