The weight of a thought

When you ask an AI a question, nothing seems to happen anywhere but on your screen. In reality, somewhere — often several countries away — a warehouse of chips flickers awake, draws a gulp of electricity, and warms up enough that water has to evaporate to cool it down. A thought has a weight: grams of carbon, sips of water, a flash of heat. We feel none of it, which is exactly why it is worth looking at.

I build with this technology and mostly love it. This is not a doom essay. It is an honest accounting — because you cannot balance something you refuse to measure.

The footprint, made real

Numbers this abstract only land when you can picture them. Pick an activity and see what it actually costs — from one reply to training a whole model:

The headline figures, from published estimates: a single AI chat reply uses on the order of 2.9 watt-hours — roughly ten times a plain web search (per the International Energy Agency's 2024 estimate). Training GPT-3 once was estimated at about 1,287 megawatt-hours of electricity and 552 tonnes of CO₂ (Patterson and colleagues, 2021) — and, by one analysis, some 700,000 litres of freshwater for cooling (Li and colleagues, Making AI Less Thirsty, 2023). Zoom all the way out and the world's data centres drew roughly 460 terawatt-hours in 2022 — more than many entire countries — a figure the IEA projects could approach 1,000 TWh by 2026, with AI a fast-growing slice.

The twist: training is the small story

Here is the thing most headlines miss. Training a giant model is a spectacular one-off — a fireworks display of energy that happens once, then it's done. The part that actually scales is the opposite: inference — the billions of everyday queries, running forever.

One reply is tiny. But a tiny number times billions a day, every day, for years, is not tiny at all. The dangerous figure is never the dramatic one; it's the small one you multiply endlessly. And there is a trap waiting here with a name: the Jevons paradox (from 1865, when better steam engines increased coal use). Make AI cheaper and more efficient, and people don't use the saved energy — they use far more AI. Efficiency, left alone, can grow the total instead of shrinking it.

So how often will we retrain these things?

A fair worry: if a frontier model costs a hundred million dollars and a mountain of carbon to train, and we retrain constantly, is that sustainable? Two forces pull against each other.

Pulling up: frontier labs do retrain and refresh often, and each new generation has been bigger than the last, with training compute roughly doubling every several months for years.

Pulling down, and quietly winning in places: efficiency is improving astonishingly fast. The compute needed to reach a given capability has been falling — by some measures the algorithms alone get about twice as efficient every year or so (work from Epoch AI and others) — on top of steadily better chips. And you rarely need the giant. A small, distilled model can do most everyday tasks at a fraction of the energy; the frontier model is a sledgehammer you should only pick up for the genuinely hard nail.

The future footprint is a race between these two — appetite versus efficiency — and it is genuinely not yet decided.

Two futures, and honest odds

• The runaway. Demand outruns efficiency. Models and usage balloon, data centres strain local grids and water tables, and AI becomes a serious chunk of global electricity — built, too often, on whatever power is cheapest rather than cleanest. Plausible if we sleepwalk.
• The efficient path. Intelligence-per-watt improves faster than demand grows. Data centres are sited where power is clean and cooling is cheap, models are right-sized to the task, and AI even pays down its debt by accelerating the things that decarbonise everything else — better batteries, grid optimisation, materials discovery, fusion research. Plausible if we're deliberate.
• My honest read: somewhere between, and — this is the important part — it is a choice, not a fate. The same industry that can make a model 175 billion parameters large can make it ten times more efficient; which one it optimises for is a decision, and decisions can be pushed.

Striking the balance

Not slogans — the actual levers, roughly in order of leverage:

• Right-size the model. The biggest single win is not using a frontier model for a task a small one handles. Most queries don't need the sledgehammer. This is a builder's responsibility, and it's mostly free.
• Make efficiency the target, not a side effect. Distillation, better algorithms, better silicon — measured in intelligence per watt, published, and competed on.
• Move the compute to clean power. Where and when a data centre runs matters as much as how much it draws; siting near renewables and running flexibly with the grid is a lever most people never see.
• Measure and disclose. We are arguing over estimates because the real numbers are rarely published. You cannot manage what you won't reveal.
• Aim AI at the problem. The optimistic case isn't that AI is carbon-free; it's that a tool this powerful, pointed at clean energy and efficiency, could save far more than it spends. That is a bet worth making on purpose.

The false binary

"Technological advancement versus the Earth" is the wrong frame — a choice you only have to make if you're careless. The real question is whether we build deliberately: using the smallest tool that works, powering it with the cleanest energy we can, and being honest about the bill.

A thought has weight. The point is not to stop thinking — it's to build machines that know their own weight, and to use them like they cost something, because they do. The billions of numbers inside these models (the embeddings I find so beautiful) are not made of nothing. They are made of energy, water, and heat — and how gracefully we carry that is, in the end, up to us.

Sources & further reading: International Energy Agency, Electricity 2024 (data-centre demand and per-query estimates); Patterson et al., “Carbon Emissions and Large Neural Network Training,” 2021 (GPT-3 energy/CO₂); Li et al., “Making AI Less Thirsty,” 2023 (water); Epoch AI on algorithmic efficiency trends; and W.S. Jevons, The Coal Question (1865), for the paradox that still haunts us. Figures are best-available estimates and vary by source and method.