The footprint

Every answer a chatbot gives you is produced somewhere physical: a building full of computers, drawing electricity, sipping water and running on chips made from mined materials. AI feels weightless on your screen. It is not. Understanding the footprint, without exaggerating it or waving it away, is part of being literate about the technology.

Electricity, water and metal

Training a large model means running thousands of specialised chips flat out for weeks or months, which consumes serious amounts of electricity. Then comes the part people forget: serving the model to millions of users every day, which over time can use far more energy than the training did. Data centres also need cooling, and many use significant amounts of water to do it, which matters most in the dry regions where they are often built. And the chips themselves depend on mined and refined materials, with hardware turning over quickly as newer generations arrive. Exact figures are genuinely disputed, partly because companies publish so little, but the direction is not in dispute: AI demand has become one of the drivers of new data centre construction and rising electricity use around the world.

Why do these particular computers use so much power? Because a modern AI chip packs an extraordinary number of circuits doing arithmetic at once, and every calculation releases a little heat. Put tens of thousands of those chips into one building and you get two bills: the electricity to run the chips, and then more electricity, and often water, to remove the heat they give off, because chips fail when they cook. Cooling is why a data centre's location matters so much. Cool climates and plentiful water make the job cheaper, and a data centre landing in a hot, dry region raises harder questions than the same building somewhere cold and wet.

The arithmetic of small times billions

The footprint argument confuses people because both sides are right about different halves of one multiplication. Work through it with deliberately illustrative numbers. Suppose a single chatbot query used one watt hour of electricity, roughly what a laptop draws in a minute or two. On its own that is nothing worth a headline. Now suppose a popular service handles a billion queries a day. That is a gigawatt hour of electricity every single day, enough on Australian household averages to power tens of thousands of homes, consumed by one service, continuously, and growing. Neither half of the sum lies. The per query cost really is tiny, and the total really is enormous, because tiny times billions is how the internet works. So when one person tells you your prompt is wrecking the planet and another calls the whole thing a rounding error, they are usually both quoting one half of the same multiplication.

The honest tension

Here is what makes this a real question rather than a simple scandal. The same technology that consumes all this energy is also being used on climate problems: improving weather and flood forecasting, helping design better batteries and materials, optimising electricity grids and making buildings and logistics less wasteful. Some of that promise is real, some is marketing, and nobody can yet total up both sides of the ledger honestly. It is also true that AI is not the biggest thing on the ledger: data centres as a whole use a small share of global electricity, though a fast growing one in some regions. Both the cost and the benefit are real, uncertain and growing at the same time.

One of the questions in the list below, whether efficiency saves anything overall, has a name and a long history. In the 1860s the English economist William Stanley Jevons noticed that as steam engines became more efficient, Britain burned more coal, not less, because efficient engines made steam power cheap enough to use everywhere. The pattern is now called the Jevons paradox, and AI may be living it in real time: each generation of chips and models does more per unit of electricity than the last, and total consumption keeps climbing anyway, because falling cost per task invites vastly more tasks. Efficiency is still worth having, but it is not the same thing as restraint, and anyone offering efficiency figures as proof the problem is solving itself is skipping the second half of the story.

Two claims to be suspicious of

You will meet two confident claims about AI's footprint, and both deserve suspicion. The first is the viral guilt claim, the post telling you a single prompt consumes some shocking amount of water or power. These figures usually come from rough estimates divided up in questionable ways and stretched far beyond what the underlying research measured, and the honest answer is that your individual query costs little and nobody outside the companies knows the exact figure. The second is the corporate comfort claim, that the issue is negligible or already solved by renewable energy purchases. Buying renewable certificates does not always mean a data centre runs on renewable power at the hour it draws it, and a small share of global electricity can still be an enormous amount of new demand landing on particular grids in particular towns. The literate position sits in the middle: your prompt is not a sin, the industry's total appetite is not nothing, and the useful pressure is for honest reporting rather than personal guilt.

The Australian angle

This is arriving in Australia, not just happening overseas. Data centres are being planned and built around Australian cities, with Sydney in particular among the region's largest data centre markets, and their demand lands on a grid that is part way through its own hard transition, still generating a large share of its electricity from coal while renewable supply and storage are built out. That timing matters. A new data centre on a coal heavy grid adds emissions in a way the same building would not on a grid running on hydro or wind, and big new industrial customers can either help fund the transition, by signing long contracts for new renewable supply, or simply compete with households for what already exists. Which of those happens is a policy and commercial choice, which is why the first question in the list below is the first one worth asking about any Australian project.

Questions worth asking, rather than a verdict

• What is powering it? A data centre running on renewable energy has a very different footprint from one running on coal or gas.
• Is the use worth the cost? Medical research and a joke image generator draw from the same grid but do not carry the same justification.
• Are companies reporting honestly? Real transparency about energy and water use would let the public judge instead of guess.
• Is efficiency saving anything overall? Chips and models keep getting more efficient per task, but total use keeps climbing because usage grows faster. Cheaper often means more, not less.
• Who bears the local cost? The water and power strain lands on the towns next to the data centre, not on the users far away.

So what should you do with all this? Not much guilt, and a little proportion. Your AI use sits alongside your streaming, gaming and scrolling as one more digital habit with a physical cost, and the same person who worries about a prompt will happily stream video for three hours, which also runs through data centres. The case for changing your personal behaviour is weak. The case for paying attention is strong, because the decisions that matter, where data centres are built, what powers them and whether companies report honestly, are being made now, and they respond to public pressure. Treat this the way this topic treats everything: less as a verdict to adopt, more as a question to keep asking.