The AI build-out just told you how much power it needs. The nuclear plants it points to won't arrive in time to supply it.

The most honest number in this week's tech earnings was not a profit. It was a power bill in disguise. Meta told investors it would spend between a hundred and thirty and a hundred forty-five billion dollars this year, most of it on data centers; Microsoft booked forty-one billion in a single quarter; Amazon and Google are each running toward roughly two hundred billion a year. Strip away the language of models and agents, and what those figures describe is an enormous, sudden demand for electricity — United States data centers already drew about three hundred and thirteen terawatt-hours in 2025, and that figure has been climbing more than twelve percent a year. So the reasonable question, the one the renderings skip, is simple: where does the power come from? The companies have an answer ready, and it is nuclear. The answer is mostly a press release.

Look at what has actually been signed, and then look at when it arrives. Meta's headline nuclear commitment — deals with Oklo, Vistra and TerraPower — adds up to as much as 6.6 gigawatts, which sounds like a wall of clean power until you read the dates. The Oklo reactors are targeted to begin coming online as soon as 2030. TerraPower's first two Natrium units are penciled for 2032, with the remaining capacity stretching to 2035. The uprates and existing-plant purchases from Vistra fill in through 2034. The full 6.6 gigawatts is a 2035 figure. Set that against a demand curve rising twelve percent a year starting now, and the mismatch is the whole story: the power is real, the plants may even get built, but almost none of it shows up in the years the electricity is actually needed.

Price it per kilowatt-hour, then per year

Now do the part I always do, which is convert the announcement into the two numbers that decide whether it matters — cost per kilowatt-hour, and megawatts delivered this decade. On cost, new nuclear is not close. Electricity from a new reactor runs about three times the cost of the same energy from a solar or wind plant, and the small modular reactors everyone is excited about are, unit for unit, more expensive still. The last large reactors the United States actually finished, at Vogtle, came in above thirty-six billion dollars, roughly double their estimate. A single Natrium reactor is estimated at over nine billion dollars for three hundred and forty-five megawatts. Against those figures, the money the tech companies have committed is startlingly small: the largest single nuclear investment announced is on the order of six hundred and fifty million dollars, and the sector's total commitments amount to a fraction of one percent of what these firms spend on data centers every year. A company spending a hundred and forty billion dollars a year and pointing at a few hundred million in reactors is not powering its build-out with nuclear. It is buying a story about nuclear.

A company spending a hundred and forty billion dollars a year and pointing at a few hundred million in reactors is not powering its build-out with nuclear. It is buying a story about nuclear.

There is a reason the commitments are structured the way they are. Most of these deals are options, not purchases — the right to buy power if and when the reactor exists, which does not obligate anyone to buy electricity that turns out too expensive to use. And even the option runs into a supply problem the announcements rarely mention: the advanced reactors depend on a specialized fuel, high-assay low-enriched uranium, and the entire licensed United States supply of it is around nine hundred kilograms a year against a need that would have to scale fifty to a hundred times over. No commercial small modular reactor is under construction in the country today. A technology with no units being built, a fuel supply two orders of magnitude short, and a first-of-a-kind timeline that has already slipped — Natrium's commercial date moved from the early 2020s to 2031 — is not a 2026 power source. It is a 2030s hope, and the grid does not run on hope this quarter.

What actually keeps the servers on

So what is powering the data centers going up right now? The boring, unglamorous answer, which is usually the true one: natural gas and whatever the existing grid can spare. The largest air permit for gas-fired generation ever granted in the United States — 7.65 gigawatts, from a single project in Texas — was issued to feed exactly this kind of load. Coal burn in the United States rose about thirteen percent to meet data-center demand, reversing years of decline. Renewables, for their part, have quietly grown faster than nuclear and now generate more; nuclear's share of global electricity has drifted down to under nine percent, roughly half what it was thirty years ago. Even Japan's much-publicized pledge tells the story if you read both halves: forty billion dollars for small modular reactors, and thirty-three billion for natural gas. The reactors are the announcement. The gas is the plan.

And this is where the cost stops being an abstraction and lands on a bill, because the megawatts have to come from somewhere and someone pays for the wires. When a hyperscaler plugs a multi-gigawatt campus into a regional grid, the near-term supply is gas and existing capacity, and the cost of firming it up — new transmission, upgraded substations, higher wholesale prices in a tighter market — does not stay neatly inside the data center's fence. It spreads across every ratepayer on that grid, which is one reason forty-four percent of Americans in recent polling oppose new data-center construction, rising to fifty-seven percent when the project is in their own community. The nuclear announcements are, among other things, a way of not talking about that invoice — of pointing at a clean reactor arriving in 2035 while the actual electrons this year come from a gas plant that someone else's utility bill is quietly paying to run. I report from a place where the gap between the promised power plant and the one that actually turns on is a daily fact of life, and the lesson travels: the energy source that matters is the one delivering electricity on the date you need it, at a price someone can pay. By that test, the nuclear renaissance powering artificial intelligence is, for this decade, a gas expansion wearing a press release.

None of this means the reactors will never come, or shouldn't. Firm, carbon-free power is worth building, and some of these projects may well deliver in the 2030s and matter enormously when they do. But an energy claim has to clear a bar that has nothing to do with the ribbon-cutting: it has to supply real power, on the date it is needed, at a cost that competes. Measured that way, the honest column for the AI build-out's electricity between now and 2032 reads gas, grid, and a rising coal number no one wants to print — with nuclear as a line further down the page, dated later than the demand and priced above the alternatives. The megawatts decide, not the announcements. Watch which plants actually switch on, and when, and read the utility bill underneath. That is where the real energy story of artificial intelligence is being written, and it is not the one on the press release.