Someone at dinner last night was cross about AI and water. Fair enough — the numbers are real. A large data centre can drink a few million gallons a day, and the new AI ones are bigger than most. The figure that gets quoted — Nvidia's, the baseline it uses to sell the liquid cooling that avoids it — is roughly 2.6 million gallons per megawatt per year for the cooling alone, before you count the power station feeding it. Multiply that across a build-out the size of a small country's electricity demand and you have a number worth being cross about.
So they were right. They were just cross at the tenant when the problem is the plumbing.
Here is what the cooling water actually does, because it surprised me too. The dominant method, for now, is evaporative — hot air off the servers passes through a tower, and the heat leaves by turning part of the water into vapour. That portion is gone. Not borrowed, not warmed and handed back: evaporated into the sky. Google's data centres withdrew 7.8 billion gallons in 2024 and consumed 78 percent of it that way. The water isn't cooled and returned. It's distilled and released.
That word is the whole article, so let me sit on it. Evaporative cooling is the first half of a still. You put energy in, the pure water leaves as vapour, the dissolved salts stay behind. What a data centre doesn't do is the second half — catch the vapour and condense it. It vents the clean water to the atmosphere and keeps the concentrate, a saltier stream called blowdown that gets dumped back into the system with its calcium and silica and chloride concentrated up. Distillation at scale, with nobody collecting the clean side and everybody downstream getting the brine. The reason operators love it is that the energy doing the evaporating is waste heat they wanted gone anyway. Two problems cancel out. The heat leaves, the water phase-changes, and it's nearly free — to them.
The power station does the same thing again, one step upstream. A thermal plant boils water to spin a turbine, condenses it, and reuses it in a closed loop — that part is fine. But condensing the steam needs its own cooling water, and in a recirculating plant most of what it consumes leaves the same way the data centre's does — evaporated off cooling towers, gone. Counted honestly, consumption against consumption, a tower-cooled coal plant evaporates something like seven hundred gallons per megawatt-hour. Run a data centre on that grid around the clock and the power station can consume as much water as the building's own cooling does, sometimes more — a second bill it never sees on its own meter. The same data centre on wind or solar PV carries almost none, because a turbine spun by wind, or a panel sitting in the sun, doesn't boil anything — solar's steam-raising cousin does, but that's a different machine from the panels going onto the build-out. The thing that decides whether AI is a water disaster isn't the AI. It's the cooling choice, the climate, and what's making the electricity.
It's worth being exact here, because this is where both sides reach for the wrong number. A data centre's cooling tower and a thermal plant's cooling tower evaporate the same way and consume about the same share of what they draw — the data centre's seventy-eight percent and a wet-cooled plant's eighty are the same physics, not a difference in appetite. Per unit of heat shed, AI isn't thirstier than power generation; it's running the identical machine. What it can't do is what the old once-through plants did — pull a river through the condenser and hand nearly all of it back a few degrees warmer, consuming maybe one percent. That reads as frugal until you see the withdrawal it hides: thirty to fifty times the water, drawn and warmed and returned to do its own kind of harm. So the percentage that flatters the power plant and damns the data centre is itself a thin metric — consumed-share tells you how the heat leaves, and nothing about how much was taken, where from, or whether the basin could spare it. The argument is loud about the one number that was never the point.
Now — does the evaporated water come back? Yes. It rains somewhere. Almost nothing leaves the planet. I had the same thought, and it's correct, and it doesn't help as much as you'd want. Water sits in the atmosphere about eight to ten days and travels with the weather, often out to sea, rarely back to the aquifer it left. A lot of what gets evaporated was drawn from groundwater that recharges over centuries, or from the potable municipal supply — so you're trading slow, local, drinkable water for a marginal contribution to rain over the ocean. Fresh water is only useful in the right place, in the right form, at the right time, at drinkable quality. Evaporation conserves the molecules and scrambles all four. The global cycle balances. The basin doesn't.
Which is the actual problem, and it isn't really an AI problem at all. It's that the price of water carries almost none of the information the decision needs. A gallon from a stressed summer aquifer and a gallon of recycled greywater in a wet climate cost the operator about the same flat municipal rate. So the price can't steer anyone toward the second. The expensive parts — the depleted aquifer, the concentrated brine handed downstream, the heat vented instead of warming homes — sit outside the operator's ledger. They're real. They're just billed to the basin, the downstream user, and the future, none of whom are at the table. The free lunch is only free because someone else gets the bill.
We have watched this exact failure run at national scale, and it's worth looking at before anyone assumes a government would do better. The Murray-Darling Basin Plan was meant to fix precisely this — cap the over-extraction, issue tradable water entitlements, buy some back for the rivers. Make the scarce thing carry a price and let trade route it. Thirteen billion dollars and a 3,200-gigalitre recovery target later, the market did route the water. But an entitlement is a volume, a number of megalitres, and so the system optimised brilliantly for the one variable it could price and stayed blind to the ones it couldn't. Timing — a spring flood pulse does ecological work the same volume dribbled out in February simply doesn't. Place — water in the main channel isn't water on the floodplain where the trees and the birds actually live. Connectivity — fish need the river to join up end to end, which no volume captures. The numbers bear it out: across five years of releases, seventy-nine percent of the environmental water stayed in the channel and never reached the floodplain at all. The 2026 review still rates the Coorong and the Murray Mouth as vulnerable, the south lagoon still too salty. The water was delivered. The regime was still wrong.
A vibrant river is not a stock of water. It's a pattern — variability, pulses, the right water arriving in the right form at the right time. Hand a market a thin metric and it will optimise the thinness with terrifying efficiency, and worse, it'll launder the loss as a solved problem. Environmental water allocated, box ticked, river still dying in the dimensions nobody counted. That's not a market failure or a government failure specifically. It's that any instrument — a price or a rule — inherits the blindness of its measure. Give it volume and it goes blind to everything that isn't volume.
So back to the dinner table. The people cross about AI are not wrong about the harm. They're one level too shallow on the cause. AI didn't create the bad arrangement — it walked into a building with century-old plumbing and turned on a very large tap, and the noise finally made the leak audible. Nobody marches about cooling towers at gas plants. Too old, too abstract, too embedded. AI is new and corporate and easy to resent, which makes it the perfect thing to be angry at, which makes it genuinely useful — but only if the anger gets redirected one level down to the arrangement. Stop at "AI bad" and you've spent the public's attention on a scapegoat and left the plumbing exactly as it was.
The honest version isn't that AI doesn't use much water — that's the vendor's line, and it leans on a real but narrow trick: run the cooling liquid hot enough to shed heat through dry radiators instead of evaporation, and the facility water drops to near zero. True, in the right cold climate, on the right grid. A stack of conditions, not a done deal, and it quietly moves the water question from the building to the power station rather than answering it.
And the honest version isn't that AI is draining the rivers, either. The honest version is that AI is the stress test that revealed we never learned to price water for place, time, and quality — the same lesson the Murray-Darling has been teaching for a decade to anyone watching. The tenant is loud. The plumbing is the story. The argument worth having at dinner is not whether to evict the tenant. It's whether we're finally going to look at the pipes.