AI Energy & Water FAQ

Q: Does AI use more energy than a Google search?

It depends which search you mean. A typical AI text prompt now lands near ~0.3 Wh. The often-cited "0.3 Wh per Google search" figure dates to 2009, so against that an AI prompt is in a similar ballpark; but a modern search is far more efficient (~0.04 Wh), so compared with search as it works today an AI answer uses several times more.

Q: Does this include training the model?

No. This calculator estimates inference — the cost of running your queries — not the one-time training cost, which is amortized across billions of uses.

Q: How much energy does a ChatGPT, Claude, or Gemini query use?

A typical short text query to a frontier assistant like ChatGPT, Claude, or Gemini uses about 0.3 watt-hours of electricity — roughly enough to power a 10-watt LED bulb for two minutes. Google measured its median Gemini text prompt at 0.24 Wh (2025) and Epoch AI estimates ~0.3 Wh for ChatGPT; comparable frontier models (including Claude) are in the same ballpark, though providers publish little per-query data. Reasoning or long-context queries use much more (~3.9 Wh and up). The older, widely repeated "3 Wh per query" figure is now considered roughly 10x too high for a typical short query.

Q: How much energy does my AI use add up to over a year?

Per query the amount is tiny, but it scales linearly with how much you use. At ~0.3 Wh per text query, 50 queries a day for a year is about 5.5 kWh — a few hours of a typical air conditioner, or roughly a sixth of a US household’s daily electricity. Image generation, reasoning models, and whole teams multiply that. Use the calculator with a daily/weekly/monthly/yearly toggle to estimate your own total.

Q: Which uses more energy: a text answer or an AI image?

An image. A typical text reply is around 0.3 Wh, while generating one image is on the order of 3 Wh for a frontier model — about ten times more — because image generation runs many denoising steps. Video is heavier still.

Is using AI bad for the environment?

A single AI prompt uses a small amount of energy and water — often comparable to a few seconds of a lightbulb and a fraction of a cup of water. It isn't something to agonize over per question. The footprint becomes meaningful at scale: heavy daily use across millions of people, and especially image or video generation, which costs far more than text.

How much water does one AI prompt use?

A typical text prompt uses only about a millilitre of water — a tiny fraction of a small cup. That combines on-site datacenter cooling and the water embedded in generating the electricity used (about 4 mL consumed per watt-hour, applied to a ~0.3 Wh query; Li et al., arXiv:2304.03271). It varies widely by region — roughly 0.5 to 9 mL/Wh. Longer answers and image generation use proportionally more.

Does AI use more energy than a Google search?

It depends entirely on which "Google search" you mean. A typical AI text prompt now lands near ~0.3 Wh — measured and estimated by Google and Epoch AI in 2025. The widely cited "0.3 Wh per Google search" figure dates to 2009, so against that old number an AI prompt is in a similar ballpark, not the 10–100× heavier some headlines claim. But a modern search is far more efficient — roughly 0.04 Wh, around 7× lower than the 2009 figure — so compared with search as it works today, an AI answer does use several times more. The honest takeaway: an AI prompt is comparable to a 2009-era search and several times a modern one — and search itself now often embeds AI.

Is generating an image worse than asking a question?

Yes — usually substantially. Image generation runs many denoising steps over a large model and typically costs several times more energy and water than a short text reply. Video is heavier still. If you care about minimizing footprint, generating lots of images is where it adds up fastest, not asking text questions.

This is also why the calculator estimates images differently. Text energy scales with tokens — how much the model reads and writes — so text activities let you set input and output token counts. An image has no tokens: its cost is dominated by a roughly fixed number of denoising steps per image, so it's estimated as a flat per-image figure instead (about 3 Wh for a frontier model, less for a smaller one). That's why image activities show a locked per-image value rather than token fields. You can adjust the per-image figure under Advanced → Assumptions.

How much energy does a ChatGPT, Claude, or Gemini query use?

A typical short text query to a frontier assistant — ChatGPT, Claude, or Gemini — uses about 0.3 watt-hours of electricity, roughly enough to power a 10-watt LED bulb for two minutes. That figure is grounded in 2025 production data: Google measured its median Gemini text prompt at 0.24 Wh, and Epoch AI estimates ~0.3 Wh for a typical ChatGPT query. Comparable frontier models (including Claude) sit in the same ballpark, though providers publish little per-query data. Reasoning and long-context queries run much higher (~3.9 Wh and up). The older, widely repeated ~3 Wh figure is now considered roughly 10× too high for a typical short query.

How much energy does my AI use add up to over a year?

Per query it's tiny, but it scales linearly with how much you use. At ~0.3 Wh per text query, 50 queries a day for a year is about 5.5 kWh — a few hours of a typical air conditioner, or roughly a sixth of a US household's daily electricity. Image generation, reasoning models, and adding teammates all multiply that. Use the calculator's day / week / month / year toggle to see your own total.

Which uses more energy: a text answer or an AI image?

An image, by a wide margin. A typical text reply is around 0.3 Wh, while generating one image is on the order of 3 Wh for a frontier model — about ten times more — because image generation runs many denoising steps over a large model. Video is heavier still.

Does this include training the model?

No. This calculator estimates inference — the cost of running your queries — not the one-time cost of training the model. Training is enormous but amortized across billions of subsequent queries, making the per-prompt share tiny, and published training estimates diverge too much to present honestly alongside per-query figures.