Is AI Bad for the Environment? Energy and Water, Explained

The short answer

One AI prompt uses very little energy and water, about as much as a few seconds of watching TV. The problem is scale. Billions of prompts, plus training ever-larger models, are driving fast growth in data center electricity use, and that demand is concentrated in particular places, on particular power grids and water supplies.

How much energy does one AI prompt use?

Two AI companies have published figures:

• Google (Gemini): a median Gemini text prompt uses about 0.24 watt-hours, which Google compares to watching TV for less than nine seconds (Google, August 2025).
• OpenAI (ChatGPT): Sam Altman wrote that an average query uses about 0.34 watt-hours, about what an oven uses in a little over a second (reported by DCD, June 2025). OpenAI didn't publish its method, and the figure wasn't peer reviewed.

Both are for text. Generating images and video generally takes more, and neither figure includes the energy used to train the model in the first place.

How much water does AI use per prompt?

Data centers use water mainly to keep servers cool, and power plants use water to generate the electricity they run on.

• Google: about 0.26 milliliters, roughly five drops, for a median Gemini text prompt.
• OpenAI: about 0.000085 gallons, roughly one fifteenth of a teaspoon, for an average ChatGPT query.
• The "bottle of water" figure: a widely shared 2023 estimate from UC Riverside researchers put GPT-3 at about a 500 ml bottle for every 10 to 50 responses. It's much higher because it described older models and hardware and counted the water used to generate electricity, which the company figures largely leave out.

Where and when a model runs matters a lot, too. A data center in a hot, dry place during summer uses far more water for cooling than one in a cool climate.

Where the real impact is: scale

The per-prompt numbers are small. The totals are not, and they're rising fast:

• Worldwide: data centers used about 415 terawatt-hours of electricity in 2024, around 1.5% of global electricity. The International Energy Agency expects that to roughly double to about 945 terawatt-hours by 2030, with AI the main driver (IEA).
• In the US: data centers used about 4.4% of the country's electricity in 2023 (176 terawatt-hours), and could use 6.7% to 12% by 2028 (US Department of Energy and Berkeley Lab).
• Training: the UC Riverside team estimated that training GPT-3 alone directly evaporated about 700,000 liters of fresh water ("Making AI Less Thirsty").

That demand isn't spread evenly. Data centers cluster where land, power, and fiber are available, so a few regions carry most of the load. Local grids may keep older gas plants running longer to meet it, and in dry regions data center cooling competes with homes and farms for water. That local strain, more than any single prompt, is why communities push back on new data centers.

Is it getting better?

Efficiency is improving quickly. Google reported that the energy used by its median Gemini text prompt fell 33 times over a single year, thanks to better models, chips, and data centers. But usage is growing even faster. Cheaper, more efficient AI tends to get used more, so total demand keeps climbing. Cleaner power and smarter cooling matter as much as efficiency. I've sketched two speculative ideas for this in the Lab: cooling a data center with rainwater and burying one under power lines, cooled by the soil.

What you can do

• Use AI where it saves real work. A prompt that saves you an hour is a good trade. Regenerating the same answer twenty times isn't.
• Pick the right size of model. Many simple tasks don't need the biggest model available.
• Go easy on images and video. They generally take far more computing than text.
• Keep it in proportion. Your daily AI use is small next to driving, heating, and flying. The big decisions sit with the companies building data centers and the grids supplying them.