When I send a question about a chemistry problem to a chatbot or want a paragraph reordered, it’s as though the chatbot is in my laptop and the answer pops out in seconds.
But maybe it’s popping out hundreds of miles away in a data center full of chips designed for that job. The chips are performing an enormous number of calculations. The results are then sent back to the user. The servers are drawing electricity to power themselves. Their heat is being dissipated by cooling equipment.
So how much electricity and/or water does it take for a single AI prompt?
It’s not something that’s easy to calculate in any honest way. Generating a short text response is not the same as rendering a video. It also depends on the AI model you’re using, the hardware and the cooling equipment in the data center.
For 2025, according to the U.S. Government Accountability Office, technology companies don’t provide enough information to parse out the environmental impact of the generative AI they’re running from the rest of what’s going on in their data centers.
What’s easy to see is how it’s growing.
The U.S. Energy Information Administration said in an Oct. 6 release that electricity usage in the United States is predicted to set records in 2026 and 2027. The EIA forecasts electricity demand will grow from 4,195 billion kilowatt-hours in 2025 to 4,356 billion in 2027.
Part of that increase is new data center growth.
In a 2026 update, Lawrence Berkeley National Laboratory says that data centers could contribute 11.8% of all U.S. electricity demand in 2030. The estimate ranges from 9.5% to 15.3%, depending on how quickly new data centers are built and how efficient they become.
And all that is more than just AI.
Data centers also house cloud storage and basic websites. But AI is an emerging driver of demand because the advanced AI models require powerful chips to be trained. And then they draw electricity anytime new responses are generated.
The International Energy Agency forecasts that data center electricity demand will go from around 460 terawatt-hours in 2024 to over 1,000 terawatt-hours in 2030. That’s more than double in just six years..
It’s not just electricity. Servers get extremely hot and some data centers rely on evaporative cooling methods which require water.
In 2023, it’s estimated by the 2024 United States Data Center Energy Usage Report that US data centers directly consumed around 66 billion liters of water. The power plants that power them may require more water indirectly.
This depends heavily on location. A data center that reuses water in a cold climate isn’t going to be the same pressure that a data center that uses fresh water might be in a community that suffers from droughts.
This is why just reporting national numbers can’t tell the whole story. A data center that looks efficient might still be straining the same water supplies used by nearby residents.
The electricity source is also important. According to the IEA, about 30% of the electricity used by data centers globally comes from coal. Another 26% comes from natural gas. Renewable power is expanding rapidly, but more solar panels can’t magically fix everything right away.
Data centers require consistent power day and night. That doesn’t mean it’s automatically reckless to use AI.
It can be used to help researchers analyze climate data or optimize electric grids, for example. It’s possible that the same technology that’s driving more demand might help cut other kinds of waste as well.
It’s just that gains in efficiency can get lost when overall demand increases even faster. A model might be more efficient per response. But if that model is generating billions more responses, those efficiency gains vanish. It’s possible to mitigate the impact.
Processors could be designed by chip manufacturers to do the same amount of work with less power. Data centers could use different cooling methods that reduce freshwater usage. Utilities would need new clean power generation, and improved transmission lines to handle demand before it starts, not after local grids are already strained.
Reporting is also essential. Many companies like to talk about purchasing renewable energy. But we still don’t have a way to see the electricity and water that goes into any specific AI service as a user. With consistent reporting, it’d be much easier to compare models and reward those who actually cut their footprint.
The environmental cost of any one question might be low. But the environmental cost of millions of people asking questions all day isn’t.
AI might be forever changing the way we learn and create. But if it’s going to become a part of daily life, the physical footprint can’t be hidden behind the screen.
