AI's power hunger: why chatbots need nuclear plants
Data centers, Three Mile Island restarts, household bills near server farms, and the uncomfortable physics behind every AI answer.
In September 2024, Microsoft signed a deal that would have sounded like satire a decade earlier: it agreed to buy twenty years of electricity from Three Mile Island, the Pennsylvania nuclear plant whose 1979 partial meltdown froze American nuclear power for a generation. The reactor being restarted was not the damaged one — but the symbolism was lost on no one. The company needed the power for one thing: data centers running artificial intelligence.
Google ordered small modular reactors. Amazon bought a nuclear-adjacent data center campus. Utilities that had spent twenty years planning for flat demand suddenly redrew their forecasts upward. The AI boom, it turns out, is not made of cleverness alone. It is made of electricity, and the bill has arrived.
Why does AI need so much power?
Because intelligence, as currently manufactured, is heat. Training a frontier model means tens of thousands of specialized chips running flat out for months, each drawing more power than a space heater, in halls where the electricity bill runs like a small city's. And training is the cheap part in aggregate — the real load is inference, the everyday answering. Every chatbot reply, every generated image, every AI summary in a search result is computed somewhere, and researchers estimate an AI answer costs several times the electricity of an old-fashioned search.
Multiply by billions of queries a day, then add the arms race: every major lab building bigger clusters simultaneously, each new data center campus measured in hundreds of megawatts, some planned in gigawatts — the scale of a nuclear plant per campus. Data centers already consume around four percent of US electricity, and credible forecasts see that share more than doubling within a few years. Grid operators who spent two decades bored are now the most popular people in tech.
What is it doing to the grid, and to your bill?
Creating a queue. In the data center corridors — northern Virginia above all, where a huge share of the world's internet traffic touches ground — new projects wait years for grid connections. Utilities have delayed coal plant retirements to keep up, which is the uncomfortable footnote to the industry's climate pledges. Household electricity prices in data center regions have risen noticeably, and the fights over who pays for new transmission lines — the tech giants or everyone's monthly bill — are now regular features of state utility hearings.
Water is part of the ledger too: many data centers evaporate millions of liters on hot days to keep the chips cool, a sore point in dry regions that competed to attract them.
Is nuclear actually the answer?
It is the answer the tech companies like, because AI's appetite has a specific shape: enormous, constant, around the clock. Solar and wind are cheap but intermittent; a training run does not pause for a cloudy week. Nuclear delivers steady carbon-free power, which is why the boom resurrected it — restarts, life extensions, and a wave of orders for small modular reactors that had spent years as slideware.
The catch is the calendar. A restarted reactor helps in a couple of years; new ones take five to fifteen. The AI buildout is happening now, which means the near-term gap gets filled by whatever is available — including a lot of natural gas. The honest picture is a race between the industry's genuine clean-energy spending and its even faster demand growth, with efficiency gains on the chips themselves as the wildcard. Each generation of AI hardware does more work per watt; so far, appetite has outrun efficiency every year.
Should you feel guilty about asking a chatbot things?
Not personally — your individual queries are a rounding error next to a single training run, and hand-wringing over one prompt is like fighting climate change by whispering. The scale problem is industrial and will be solved or not at the level of chips, siting, and grids.
But the story is worth following, because it is where AI's abstractions hit physical reality. The limiting factor on artificial intelligence turned out not to be ideas, data, or even money — it is transformers, turbines, land, water, and the patience of neighbors. The cloud, it turns out, is a building; the building is on your grid; and the future of the smartest technology ever built now depends on the oldest kind there is: the ability to generate power and move it through a wire.