What is the environmental impact of AI?

The environmental impact of AI is two very different stories that usually get told as one. The first story is about individual use — the electricity and water behind each chatbot prompt — and measured honestly, it is tiny. The second is about infrastructure — the worldwide buildout of data centers and the energy system that powers them — and it is enormous, fast-growing, and decided in places most of us never look: siting deals, water permits, and utility contracts. Keeping those two stories separate is the most useful thing you can do in this conversation, because each one asks something different of you.

This page is the reference version: the measured numbers as of mid-2026, in question-and-answer form, with sources. For the full argument about what our sector should do with these numbers, read Our Sector Is Getting AI and Energy Wrong — that essay is the canonical statement of where we stand.

How much electricity does one AI prompt use?

About 0.3 watt-hours for a typical text prompt. That estimate comes from independent analysis by the climate-data researcher Hannah Ritchie, and the first measured disclosures bracket it: Google reports a measured median of 0.24 watt-hours for a Gemini text prompt, and OpenAI's Sam Altman has cited 0.34 for ChatGPT. When the writer Andy Masley amortized in everything upstream — model training, chip manufacturing, the data center itself — a single prompt came to about 0.28 grams of CO2.

Numbers that small are hard to feel, so translate. One prompt is the emissions of driving a sedan about four feet. It is a space heater running for under a second, or a fifth of one printed page of a paperback book. To match the energy you spend on one transatlantic flight, you would need to ask a chatbot 400 questions a day for 80 years.

Doesn't AI use ten times as much energy as a Google search?

That famous ratio compared an early estimate for a chatbot prompt (about 3 watt-hours) with the last per-search figure Google published, which dates to 2009. Today's measured prompt figures sit near 0.3 watt-hours, so the gap has largely collapsed as chips and models became more efficient.

But suppose the tenfold ratio still held. A digital clock uses about a million times as much power as an analog watch — and nobody cares, because the absolute amount is still trivial. A big ratio on a tiny baseline is how this debate keeps fooling people. Even with billions of people prompting, chatbots draw a small slice of data-center electricity, which is itself a small slice of the world's energy. On any honest list of things to cut back for the climate, your prompts do not make the first page.

How much energy does it take to train an AI model?

Training is the biggest single line item in AI computing. The model behind the original ChatGPT took roughly as much electricity to train as 130 U.S. households use in a year, and frontier training runs have grown substantially since. But a training run is a one-time cost spread across the billions of prompts the model then serves. Amortized per prompt, it is already inside the 0.28-gram figure above.

How much water does AI use?

Far less per prompt than the viral posts claimed. The "bottle of water per prompt" meme inverted its own source — the study behind it estimated roughly one 500-milliliter bottle per 10 to 50 prompts, depending on where and when the model runs. Google's measured median is 0.26 milliliters consumed on site per text prompt — about five drops. Fuller accounting that includes the water behind electricity generation raises that by roughly an order of magnitude, which still lands well under a teaspoon.

The real water story is not in your chat window. Two-thirds of the data centers built since 2022 sit in water-stressed basins, because reaching power quickly beats water prudence in siting decisions. In one Georgia county, a data center drew 29 million gallons through an unmetered line during a drought while residents sat under watering bans. Those harms are real — and they were decided by water permits and siting agreements, not by anyone's prompt count.

How much electricity do data centers use overall?

All the world's data centers together — the whole internet, not just AI — draw about 1.5 percent of global electricity, some 485 terawatt-hours a year. The International Energy Agency projects that roughly doubling to about 945 terawatt-hours by 2030, with AI the fastest-growing slice. The corporate receipts point the same direction: Google's emissions are up 51 percent since 2019, Microsoft's rose 25 percent in a single year, and the journalists at MIT Technology Review who did the most careful independent math summarized the trajectory in one line: our AI footprint today is "the smallest it will ever be."

Two counterweights belong next to that. In 2025, for the first time, clean generation covered all of the world's net new electricity demand — the AI boom included — with solar alone covering about three quarters. And the costs that do reach ordinary households arrive through utility economics, not chat windows: data centers drove nearly two-thirds of one U.S. grid operator's capacity-price spike — 9.3 billion dollars spread across the electric bills of thirteen states. Whether the buildout lands clean or dirty on your grid, and who pays for it, is a policy fight. That is exactly why it deserves your attention more than your prompt count does.

Is using AI bad for the environment?

Your personal use is a rounding error, in either direction. Skipping prompts saves an amount of energy too small to measure against a single car trip, and one peer-reviewed comparison in Scientific Reports found that AI systems generating text or images can emit far less CO2 than a person doing the same work at a running computer. Neither fact is a reason to be wasteful. Both are reasons to stop treating individual restraint as climate action.

There is history behind that instinct. The "carbon footprint" was popularized by BP's ad agency to keep public attention on personal behavior instead of on the energy system. Per-prompt guilt reruns the same play on AI — sincere worry pointed at the smallest lever available. The worry is right. The scale is wrong. The levers are elsewhere. We make that argument in full in Our Sector Is Getting AI and Energy Wrong.

What can nonprofits and grant professionals actually do?

Aim concern at the decisions that set the outcomes:

• Back the well-understood climate agenda. Greening the grid, electrifying transportation, better batteries, cleaner building, fixed food systems — the mundane work that is actually bending the curve.
• Show up where data centers are decided. Siting criteria, water permits, and utility rate cases are public processes. Communities negotiating them deserve nonprofit expertise on their side of the table — lawyers, hydrogeologists, rate experts, and a shared record of what other counties actually got.
• Demand real disclosure. Totals and distributions from AI providers, not curated medians, so the numbers in this article keep getting better.
• Use the tool for the mission. The prompts behind an entire organizing campaign cost less energy than one volunteer's drive to the hearing. Draft the ordinance comments, build the town-hall deck, coordinate the volunteers.

The environmental impact of AI is real. It lives in infrastructure, and it responds to governance. Get the numbers right, keep your credibility, and spend your concern where it moves something.

This article was substantially revised in July 2026: early estimates were replaced with measured figures from provider disclosures and the International Energy Agency, and the framing was updated to match our current position.