What would it actually take for AI to run green?

Last time I said the “100% renewable” badge on your favorite AI tool is a lie. Fair question back at me: fine, smart guy, what would real look like? So let's build it. Here's what it actually takes to run an AI data center green, and why it's a much higher bar than a press release. The good news: the tools exist. The honest news: almost nobody's clearing the whole bar yet.

Can an AI data center ever actually be green?

Yes. But “green” has to mean two hard things at once, not one easy one.

The trick companies pull is treating “green” as an energy story only. Buy some clean power on paper, wave the badge, done. Real green is two problems solved together: where the power comes from, and how much fresh water gets destroyed to stay cool. Solve one and brag about it while ignoring the other, and you're still part of the problem. Clear both, with receipts, and now we're talking. Here's what each one actually requires.

What does real clean energy look like, not the paper kind?

It means matching your power to clean sources every hour, in the place you're actually running, not once a year on a spreadsheet.

This is the big one. Most “renewable” claims are annual accounting: a company adds up a year of dirty power, buys an equal amount of clean power somewhere else, and calls it even. Real is harder. Google's target is what's called 24/7 carbon-free energy, matching its electricity use with local clean power every hour of every day by 2030. It's tough, and even Google isn't all the way there yet. But that's the honest scoreboard. Powered around the clock by clean electrons, not just canceled out on paper.

Getting there needs new kinds of always-on clean power, because the sun sets and AI doesn't. That's why the serious money is moving toward firm clean sources. The IEA reports the pipeline of nuclear deals with next-gen small reactors nearly doubled to 45 gigawatts, and that AI demand is now pushing forward geothermal and long-duration batteries too. That's the real version. Not a certificate. New clean power that runs at 2 a.m.

How do you stop it from drinking so much water?

You stop evaporating it. You cool the chips in a sealed loop that reuses the same water over and over instead of boiling it off into the sky.

This is the fix almost nobody talks about, and it's real. Microsoft's newest data center design uses chip-level cooling in a closed loop. You fill it once when you build it, and it just keeps circulating, no fresh water needed to run. Microsoft says the design avoids more than 125 million liters of water a year per data center. That's the difference between a building that drinks a town's water and one that basically doesn't.

There's a number to watch here, and it's your best tool to cut through the spin: Water Usage Effectiveness, or WUE. It's just how many liters a data center burns per unit of computing. Microsoft got its average down to 0.30 liters per kilowatt-hour, a 39% drop from a few years earlier, partly by using recycled and reclaimed water instead of drinking water. A company serious about water can tell you its WUE. A company that's greenwashing will show you a logo instead.

Wait, doesn't saving water just burn more energy?

Sometimes, yes. And this is exactly the trap that makes “green” so slippery.

Here's the honest part almost no marketing team will say out loud. When you stop evaporating water to cool a building, you usually have to work the machines a little harder, which uses a bit more power. Microsoft admits it flat out: switching away from evaporation causes “a nominal increase in our annual energy usage.” Save the water, spend the watts.

Water and energy are a seesaw. Cheaters push down one side, point at it, and hope you don't look at the other.

Real green means holding both down at the same time: clean, around-the-clock power AND closed-loop water. If a company fixes one by making the other worse and only shows you the good half, that's not a solution. That's a magic trick.

Is there a way to just use less of everything?

Yes, and it's the least glamorous and most important lever: efficiency and smart siting.

Two things move the needle. First, the machines keep getting more efficient. The IEA says the energy used per AI task is dropping faster than almost anything in energy history. More work per watt, more work per drop. Second, where you build matters enormously. Put a data center in a cool, water-rich region and it needs far less cooling than one baking in a desert that's already fighting over water. Some operators are also capturing the servers' waste heat to warm nearby buildings instead of dumping it. Same heat, second job.

None of this is exotic. It's boring engineering and honest planning. Which is probably why it doesn't make the press release.

So how do I tell a real one from a fake one?

Ask for the receipts. Three of them. If a company can't produce all three, the “green” is decoration.

• One: what percent of your power is matched to clean sources hour by hour, in the region you run, right now? Not the annual certificate. The hourly number.
• Two: what's your WUE, and are you cooling with recycled water or the same water people drink?
• Three: are your new builds closed-loop, and when do they come online? Microsoft's zero-water sites are being piloted in 2026 and start opening in late 2027. Real plans have real dates.

A company doing the work will answer all three without flinching, because they're proud of the numbers. A company hiding behind a badge will change the subject. Now you know which is which.

The bar, in one line

Green AI isn't a sticker. It's clean power every hour, water that circulates instead of evaporates, both held down at once, and the receipts to prove it. That's the bar. It's high, it's real, and a handful of players are actually starting to clear it.

So here's my ask, to the companies and to you. Stop clapping for the logo. Start asking for the three numbers. The technology to run AI clean already exists. What's missing isn't the engineering. It's the honesty to be measured by it.