What AI Costs, and What It Asks of Us

You've heard the headline by now: AI is straining the power grid. Data centers are going up everywhere, utilities are nervous, and the electricity numbers keep climbing. All true. But the headline skips the more interesting question, and the more useful one: what, exactly, is eating all that power?

I went looking for real measurements rather than vibes, and the short answer surprised me. It's not spread evenly. Some things AI does cost almost nothing. Others cost thousands of times more. Once you can see the difference, a lot of the hand-wringing turns into something you can actually reason about.

Some AI is thousands of times hungrier than the rest

Start with a single task. A team from MIT Technology Review did the unglamorous work of measuring the energy of different AI jobs the same way, so they're actually comparable. Here's what one task costs, end to end:

Read that carefully, because the shape matters more than any single number. A short text answer is a rounding error — a few feet on an e-bike. Generating one image costs meaningfully more. And a five-second AI video costs 3.4 million joules — more than seven hundred times a single image, enough to run a microwave for over an hour. For one clip.

Notice something else: it isn't a tidy "text is cheap, images are expensive" story. A long answer from a big model costs more than a standard image. The real driver underneath is how much computing the task demands — and generating pixels and frames demands enormously more than generating words. Video sits in a category of its own.

It's not just the model — it's how much you ask it to chew

The task type is only half of it. The same model, on the same kind of request, swings wildly depending on how much you feed it. Epoch AI dug into this and found the cost scales sharply with the length of the input:

A quick question is about a third of a watt-hour. Hand that same model a long document to read, and you're at eight times that. Point it at a big research task — the kind an AI "agent" runs when it works through a problem in steps, reading and writing along the way — and a single job can cost more than a hundred times a simple question. Nothing about the model changed. You just asked it to hold more in its head at once, and the cost climbs faster than the input does.

This is the part most people miss. The AI features quietly showing up in your search bar, your inbox, and your business apps aren't all the cheap kind. The moment a tool starts summarizing long threads, generating images, or running multi-step agents on your behalf, it moves up this curve.

And it's the running, not the training, that adds up

There's a common assumption that the huge energy cost of AI is training — the one-time, months-long process of building a model. Training is genuinely enormous. But it happens once. What happens billions of times a day is inference: the model actually answering, generating, and doing. By current estimates, inference now accounts for 80 to 90 percent of the computing power spent on AI.

That reframes everything above. Each individual query might be small. But small, multiplied by everyone, running all day, every day, forever, is the real load — and it's tilted heavier every month as more of that everyday use shifts toward the expensive end of the first two charts.

Which is why the curve bends up

Put those together — heavy tasks, longer inputs, and constant everyday running — and you get the number the utilities are worried about. The International Energy Agency projects that electricity used by data centers will roughly double by 2030:

From about 415 terawatt-hours in 2024 to around 945 by 2030 — from one and a half percent of the world's electricity to nearly three. The servers driving that growth, the ones built for AI, are the fastest-rising slice of all. This isn't a spike that settles back down. It's a new baseline forming.

And it's already landing on the power bill

Here's where this stops being an abstract industry story and reaches you. That demand doesn't stay politely inside the tech sector. A data center bids for the same electricity your business and your home draw from, and when a campus that pulls as much power as a small city moves into a region, everyone on that grid feels the squeeze.

It's not hypothetical anymore. Consumer Reports found that residential electricity prices jumped 7.1% in 2025 — and topped 20% in some states; in Virginia's "Data Center Alley," prices have climbed 267% over five years, and one resident watched a monthly bill go from about $100 to $281. Researchers at Harvard's Belfer Center describe this as a watershed moment for the U.S. grid, warning that the opaque deals utilities strike to power these campuses can quietly "shift power costs onto other consumers" — which is to say, onto you.

And it doesn't stop at the electric meter. The same rising cost of power and compute flows into what you pay for hosting, storage, and every cloud subscription your business rents by the month. When the electricity underneath gets more expensive, so does running the infrastructure your software lives on — which is part of why cloud providers have been trimming discounts and nudging prices upward on services that used to only get cheaper.

So what do you actually do with this?

Not panic, and not swear off AI — that would be the wrong lesson from a genuinely useful technology. You don't need a carbon calculation running in your head every time you open a chatbot. You just need a feel for the shape of the thing: a quick question costs almost nothing, while images, video, and the long, autonomous tasks are where the real weight lands.

Hold that lightly, and the rest takes care of itself — because you already know how to do this. The same everyday conscience that makes most of us turn off a light we're not using, or not leave the tap running, works just as well here. You don't have to be an engineer or an environmentalist to use a powerful tool with some care. You just have to be a person who'd rather be deliberate than wasteful, and let that ordinary decency reach one more corner of how you work. Whatever it is that makes you care how your choices add up — for your business, your kids, the world you're handing on — it applies to this too. It always did.

One last measurement: what this article cost

It seemed only fair — after two thousand words on AI's energy — to turn the meter on this very article. Here's the rough tally for making it: the drafts, the research, the charts, the several rounds of revision, and yes, the energy spent working out these numbers and drawing the chart below. It came together the way a lot of things do now, with a human deciding and an AI assisting, so the bill splits in two.

The AI did the generating — drafting, researching, rendering the charts — and drew on the order of two million joules, roughly half a kilowatt-hour, or a few cents of electricity. Still less than one five-second AI video from the first chart. The human did the slower, smaller work of reading, judging, and deciding what stayed, running the whole time on a single brain at about twenty watts — a couple of almonds' worth of calories across the session. The thinking itself, the effort above simply sitting and breathing, cost maybe half a calorie. A crumb. (And every revision to get this figure honest nudged the AI's side up a little further. The meter counts itself.)

It's a clarifying picture rather than a discouraging one. The machine's cost was real but bounded, and nearly all of it was reasoning-heavy work run over and over — the very pattern the charts above describe. The human's cost was almost nothing, and it carried the one thing no meter can price: judgment, care, and knowing when a draft has lost the plot.

If you'd ever want a second set of eyes on where AI actually sits in your operations — what's running, what it's costing, and what's worth keeping — that's the kind of thing we help small businesses think through, across the New York metro area and the Puget Sound alike.