How much electricity do AI data centres use?

The IEA estimates data centres consumed around 415 terawatt hours in 2024, about 1.5% of global electricity. It projects that roughly doubles to around 945 TWh by 2030, just under 3% of world consumption, growing about 15% a year — more than four times faster than electricity demand from all other sectors combined.

Are Big Tech's emissions rising because of AI?

Yes, according to their own reports. Microsoft's emissions rose from 13 million tons of CO2 equivalent in 2020 to 20 million tons in 2025, a 54% increase, including a 25% rise in a single year, while the company is publicly pledged to be carbon negative by 2030. Google reported a 37% annual increase in electricity demand in 2025 and a 25% year-over-year rise in supply chain emissions.

How much water does a data centre use?

A large data centre can consume up to 5 million gallons per day — comparable to a town of 10,000 to 50,000 people. US data centres consumed about 449 million gallons per day as of 2021. Much of it is evaporated for cooling rather than returned.

Is AI destroying the planet?

Not by itself. The IEA puts data centres at about 1.5% of the world's electricity in 2024, rising to just under 3% by 2030, and says their growth is less than 10% of the rise in global electricity demand to 2030. Planes, cement, steel and farming are all much bigger. The real worries are how fast it is growing, the water and land it takes from local towns, and the fact that the companies building it are missing their own climate promises.

Does running AI locally use less energy?

No. Your own computer actually uses more electricity to answer a question than a data centre does, because their chips are purpose-built for it. Researchers found cloud chips get roughly 1.6 to 7.4 times more work out of the same electricity than a high-end home machine. What running at home does avoid is the fresh water used to cool those buildings, and it does not require any new building to be put up.

GreenCube

ES Get GreenCube

AI & the planet

The cloud is not a cloud. It's a warehouse full of computers.

Ask a cloud AI a question and it doesn't get answered in the sky. It gets answered in a giant building full of computers. They get hot, so they're cooled with water. They run on electricity. Thousands more are going up right now. Here are the real numbers — mostly from the companies themselves — including the ones that make us look bad.

The trajectory

Nothing else is growing this fast.

These come from the International Energy Agency — the careful people whose job is counting the world's electricity. Not campaigners.

Twice

as much power by 2030

These buildings will use double the electricity in 2030 that they used in 2024.

1.5% → 3%

of all the world's electricity

Every plug, every light, every factory on Earth. These buildings take 1.5% of it today. Nearly 3% by 2030.

4× faster

than everything else

Their power use grows about 15% a year. Everything else on the grid grows four times slower.

Source: IEA, Energy and AI — “around 415 terawatt hours (TWh), or about 1.5% of global electricity consumption in 2024”, projected to “double to reach around 945 TWh by 2030”.

Their own reports

The companies promising to fix it are the ones building it.

We didn't dig this up. They printed it themselves, in their own yearly reports, about themselves.

Microsoft

Their promise: by 2030, take more pollution out of the air than they put in

13M → 20Mtons of greenhouse gas a year, 2020 to 2025. That's 54% more in five years.

+25%in one year alone — because of all the AI buildings they're putting up.

2% → 13%how much of their pollution comes from the electricity they buy — that jumped in one year.

Four years left on that promise, and they're going the wrong way.

Google

Their promise: stop adding pollution by 2030

+37%more electricity used in a single year (2025).

+25%more pollution from building things — the concrete, steel and chips that new AI buildings are made of.

78%of the fresh water they used, they put back. So about a fifth of it never came back.

Running them got 2% cleaner. Building them got a lot dirtier.

Sources: Microsoft's 2026 sustainability disclosure · Google's 2026 Environmental Report

The part with a postcode

One building drinks
like a small town.

Electricity is hard to picture. Water isn't. It comes out of a real reservoir, next to real people — and cooling the computers boils it away, so it doesn't come back.

5M gallons

A day. For one big building. That's as much water as a town of 10,000 to 50,000 people uses.

449M gallons

A day, for all the US ones together. And that was back in 2021, before the AI rush even started.

519 ml

Ask a cloud AI for a 100-word answer and that's roughly one small bottle of water, boiled away.

700,000 L

Boiled away just teaching GPT-3 to talk — the same fresh water it takes to build 370 BMWs.

Sources: EESI · UC Riverside · “Making AI Less Thirsty”. Disclosure is largely voluntary, so communities often can't find out what's being used next door.

What we're not going to tell you

AI is not why the planet is in trouble.

We sell AI that runs on your computer. So it would suit us to finish this page with "…and that's why the cloud is killing the planet." But that isn't true. And if we lied to you here, you'd be right to stop believing everything above it.

The same report says these buildings will still only be about 3% of the world's electricity in 2030. Planes, cement, steel and cows are all far bigger problems. Anyone telling you AI is the main villain here is selling you something.

And one more, about us. Your laptop is not the greener choice. It actually burns more electricity to answer you than their building would — their chips are built for this and yours aren't. Researchers measured it: theirs get between 1.6 and 7.4 times more work out of the same electricity.

So here is all we'll claim, and it's true: nobody boils off a town's drinking water to answer your question, and nobody puts up a new warehouse because you asked one. Your computer is already here, and it's already switched on.

Sources: IEA · “Intelligence per Watt”, arXiv 2511.07885.

Go deeper

The rest of the receipts.

Deep dive

How much water does AI use?

The full numbers, why credible researchers disagree about them, and what actually reduces it.

Read the answer → The other cost

What happens to your data

Water and power are one price you pay. Your private chats are the other one — and you pay that every time you type.

See what stays on your machine →

You can't fix all this from a text box. But you do get to pick where your own AI runs.

Get GreenCube — €9 One-time payment. Runs on the computer you already own.