Most of "AI's" electricity isn't AI.

Data centers used about 440 terawatt-hours of electricity in 2025 — more than Pakistan, Bangladesh, and Nigeria use put together. That's around 650 million people, nearly twice the population of the United States, and it's projected to more than double by 2030. The number is real, and it's worth taking seriously.

So let's take it apart — because two things about it get misunderstood.

It isn't. A single AI answer runs about 0.24 watt-hours — by Google's own measure, less than nine seconds of watching TV, or roughly 1–2% of one phone charge. (That's not even the smallest figure available; companies publish a "just the chip" number too, and I'm not using it.) Even a long, heavy answer is a minute or two of a microwave.

So which is it — nine seconds of TV, or 440 terawatt-hours? Both. The footprint was never personal; it's industrial — billions of queries, plus training, plus video, running around the clock. Trimming your own prompts to save the grid is like draining a reservoir with an eyedropper.

This is the one almost nobody catches. That 440 terawatt-hours isn't AI's bill — it's every data center on Earth. Netflix and Spotify, every email, your bank, your cloud photos, this newsletter's own host. The power-hungry AI everyone pictures is only about 5 to 15% of it.

The number people wave around to prove AI is boiling the planet is mostly… not AI.

That's the real reason the worry misfires — usually not hypocrisy, just an honest blur. "AI" is a marketing umbrella, not a technology. Apple Photos finding your kid's face is AI. So are Netflix and Spotify recommendations, autocorrect, your spam filter, Maps reading traffic. The person sure they "don't use AI" has been opted in since their phone first sorted their photos.

It's small today but climbing fast, and it helps to know why. An AI chip is far more power-dense than an ordinary server: a single AI rack can draw as much electricity as 50 to 100 homes use — not by taking power away from anyone, just by consuming that much itself. And the work itself is getting dramatically heavier. A text answer is about 0.24 watt-hours; a single five-second AI video can burn through as much energy as running a microwave for a full hour — thousands of times more per use. That leap is the real engine behind the curve, and it's the part to take seriously: as AI shifts from text to images to video, the cost per use climbs into a different league entirely. The one thing pushing back the other way is efficiency — the same gains that made a text prompt 33× cheaper in a year, now being turned on video. Which makes this a race on two fronts: how fast the tech gets leaner, and — just as much — how fast the rest of us show up to decide where and how it all gets built. The first is on the engineers. The second is on us, in the rooms where it's approved.

Not your conscience — the shared grid. A data center plugs into the same wires your house does, so the power plant feeding it has to make more, and the bill for that new capacity gets split across everyone on the grid. Which is why the fight was never about your usage — it's about who pays. PJM, the largest US grid operator, traced billions in higher capacity costs to data-center demand last year. In response, seven AI companies — Amazon, Google, Meta, Microsoft, OpenAI, Oracle, and xAI — signed a White House Ratepayer Protection Pledge in March 2026: a voluntary promise to fund their own grid upgrades rather than pass the cost to households. But it covers only those seven companies' own builds, and household bills have kept climbing well above inflation — a voluntary pledge won't fix that on its own. What can: making it binding. Several states are already writing separate "large-load" rate classes so data centers pay their own way instead of the neighborhood. That's a rate-case decision — and rate cases are public.

Your friend's instinct to do something is a good one — it's just better aimed at the rate case and the zoning board than at the prompt box. Trimming your prompts won't cancel a data center or keep your neighbors' rates down. Showing up to the hearing might.

And it helps to see how small the slice really is — all the world's data centers are about 1.5% of global electricity, a sliver next to industry and homes:

One thing that genuinely surprised me: most of what gets counted as AI's water isn't the water cooling the servers — it's the water evaporated upstream at the power plant, making their electricity. So a chunk of "AI's water problem" was really its electricity problem in disguise. Which is oddly hopeful: clean up the power and you drain both footprints at once. You were never being asked to choose which crisis to care about.

There's no single switch — because not all electricity is the same. Two questions decide everything: is it clean, and is it always-on? A data center runs 24/7, so a source has to be both — and almost none are. Worse, one popular "green" option, bioenergy (burning crops, wood, or plant waste for fuel), only looks clean: the UN found that swapping to it cuts carbon ~70% but multiplies water ~30× and land ~100×. Reduce one harm, inflate two others.

Bioenergy is the worst case, and not what the industry is choosing. What it's actually building: efficiency (Google's median prompt got 33× more energy-efficient in a single year, while the answers got better), and firm clean power — over 9.8 GW of nuclear committed (Microsoft is restarting Three Mile Island), with geothermal coming online this year in Nevada and Iceland. The trade-off is real, and the better choices exist, and they're being installed right now.

Even a perfectly clean data center can be a bad neighbor. The loudest complaints in these towns aren't about carbon — they're about the constant roar of the cooling fans, an industrial block dropped beside homes drinking the local aquifer, and public hearings scheduled for 10 a.m. on a Tuesday, when no one with a job can show up. The winnable asks are concrete: disclose the water and power deals, hold hearings people can actually attend, keep these off strained grids and away from homes, and make new centers bring their own clean power. Ireland did exactly that — data centers reached about 21% of its electricity, so it froze new Dublin-area builds until 2028. None of those were inevitable; each was a decision someone showed up for.

So here's where it leaves you. The scariest number is mostly not even AI; the part that is isn't your prompt history; and there's no green button — just a menu, with the good options getting cheaper every month and real decisions being made in public rooms about which ones get built. The worry that survives a fact-check usually leaves you somewhere more useful than the panic did: not guilt at your keyboard, but a hearing you could attend.