Someone told me last week that one AI email drinks half a litre of water. He wasn't wrong. Researchers at UC Riverside ran it and a 100-word email can cost around 500 millilitres once you count the water it takes to generate the electricity behind the server. Half a bottle, gone, for a few sentences. Real number. It should sit uncomfortably.
That same email, measured another way, costs about 0.3 millilitres. Cooling water only, off the vendor's own sheet. Same email. Two figures a thousand times apart and both of them honest, because they count completely different things. One counts the power station out the back. One counts the pipe running past the chip. So whenever somebody waves a scary AI water number at you, the first question is always what did they actually measure.
Power's the same. For years the line was that every question you ask costs about 3 watt-hours. Epoch AI went and measured a typical query in 2025 and got roughly 0.3 watt-hours, about ten times below the number everyone was repeating. Reasoning-heavy stuff, the models that sit and think, run a fair bit higher. But the everyday query is a tenth of the myth.
Now the part i won't dress up, because the honest version matters. The total is climbing and it's climbing fast. The IEA reckons data centres pulled about 415 terawatt-hours in 2024, near enough to 1.5 percent of the world's electricity, and they expect that to roughly double to around 945 terawatt-hours by 2030, about 3 percent. That's a genuine load on the grid. Anyone telling you the energy problem is solved is selling something. It isn't solved. It's real and growing, and the water sits underneath the power the whole way down.
Here's what i keep coming back to though. The same thing AI is actually good at, spotting a pattern in a mess too big for a person to hold, is the exact thing you aim at a cooling system or the way a chip's laid out. And the curve's already bending. Epoch found the compute needed to hit a fixed level of capability has been halving roughly every eight months. That's faster than Moore's Law ever moved silicon. So the job that cost a bucket of energy last year costs a cup this year for the same result. Total use keeps rising while the cost per task keeps dropping, both true at once, and the second one is the lever.
I've stood in a commercial kitchen watching an old cool room chew through power because nobody had touched the seals in five years. Same machine, same job, half the draw once you fix what it can't see itself. That's the shape of what's coming for the big rooms full of chips. Better cooling and placement, hardware built for the actual work instead of brute force. The efficiency gains are the boring engineering that never makes the headline, and theyre moving quicker than the fear is.
None of that lets us off the hook. We run G'dai Mate careful with tokens on purpose. Every job we build has to earn the compute it burns, do real work for a real business, save someone hours they'd otherwise grind through by hand. If a thing can't clear that bar it doesnt ship. Waste is waste, whether it's water or a bloke's afternoon.
So both things are true. The cost is real and the fear is fair, and the machine that worries you is the same one bending the curve back down. Stay honest about the number. Make it earn its keep.