There is a marginal note in an old edition that says it all. In a used book, you sometimes find a reader’s annotation — a single underlined line, a terse “no” in the margin, a date scrawled beside a passage the owner could not let pass. The annotation is not the book; it is the reader’s whole position compressed into a few words. This month, the AI era acquired its marginal note, and it was not written by a reader. It was written by a power company and a technology giant, in the form of a twenty-year contract.
The contract, read carefully
Let me tell you why this particular document stopped me. Vistra announced that Meta will buy, under a twenty-year power purchase agreement, the output of two nuclear plants in Ohio — Perry and Davis-Besse — totaling 2,176 megawatts of existing generation, and will also purchase 433 megawatts of additional capacity that comes online by 2034. The industry press calls it the largest single corporate nuclear commitment on record, and I will not argue with that framing, but I want to read the numbers the way you would read a margin note: slowly, and for what they imply.
The headline number, curiously, is the 2,176 megawatts, and the number that matters more is the 433. The first is the anchor — steady, existing, already on the grid, generation that was running before AI was a line item. The second is the signal — new capacity, committed years in advance, bought before it exists. A company does not sign a contract for capacity that does not yet exist unless it has made a decision about its own future. That 433 megawatts is the sentence underlined in the margin.
What two thousand megawatts actually buys
Let me translate the scale into a language the rest of us can hold. A single AI data center consumes somewhere between 300 and 500 megawatts. Do the arithmetic and the shape emerges: 2,176 megawatts is roughly four to seven large AI facilities running at full load. That is not a hedge position. That is a fleet decision, signed for two decades, priced in nuclear — the most stubborn, least flexible, most reliable form of generation a utility can offer.
I find myself pausing on the word “reliable”. Nuclear is the base-load answer to the thing that every AI expansion keeps tripping over: you cannot schedule the sun, you cannot order the wind, but you can count a reactor. When the largest private buyer of compute in the world chooses to bind itself to a reactor’s output for twenty years, it is saying something precise about what it values — not speed, not novelty, but a load factor that never blinks.
And it is worth noting what kind of a reader signs a twenty-year term. This is not the impatient hand that buys on a rumor or the speculative pencil that annotates every trend. This is the signature of someone who has decided that the industry’s glamorous questions — which architecture, which cluster, which algorithm — matter less than the unglamorous one: where will the watts come from, and can I count on them in 2046? The margin note is blunt precisely because the thinking behind it is long.
Let me picture the room where the signature went down, because every document has a scene behind it. Two legal teams at a table, one representing a power utility with reactor licenses older than half the people in the room, the other representing a technology company whose product line will be unrecognizable by the time the contract matures. Somewhere on that table, a calendar open to 2046. Someone initials the page-count, someone initials the capacity schedules, and the twenty-year term is fixed between two organizations that will each be materially different before it expires. That image — the twenty-year calendar on a table in a room that will not look the same next decade — is the whole story in a single frame.
The grid queue, the real protagonist
And here is where the story turns, because the contract is only half of the margin note. The other half is the queue. In the United States, the backlog of interconnection requests waiting for grid approval has passed 2,600 gigawatts, with an average wait around five years. Nearly half of the AI data centers planned for 2026 are facing delays. The shortfall is around 7 gigawatts. The capital at risk underneath those delayed builds is estimated at roughly 650 billion dollars.
Let me make sure that lands. Five years is not a seasonal delay; it is an epoch in compute time. A data center planned in 2026 and connected in 2031 will open into a technology landscape that its own design cannot have predicted. That is the true cost of the queue — not just the interest on idle capital, but the obsolescence built into waiting. A machine built to serve one generation of models will come online to serve the next one, if it comes online at all.
This is the part of the story that the press release will not tell you. The Meta-Vistra agreement is news because it is the largest, but it is also the exception. It works because the reactors already exist. The 2,176 megawatts is pre-provisioned. The hard, crowded, five-year-waiting part of the grid is exactly what this deal sidesteps — and every other buyer who cannot point to an existing plant is still standing in that queue.
The smaller signature
There is a second, quieter signature in the same week, and it belongs in the same margin. On August 24, NANO Nuclear signed a framework agreement with Tillman Digital Gateway, becoming the preferred nuclear technology supplier for the latter’s planned artificial intelligence industrial park in the United States, with non-binding targets of more than two gigawatts by the mid-2030s and more than six gigawatts by 2040.
Let me think about what that agreement is, precisely. It is not a signed guarantee of electrons. It is a reservation — a statement of intent between a developer and a reactor vendor that both sides intend to build nuclear-powered AI infrastructure together, on a scale that dwarfs even the Meta deal. The targets are explicitly non-binding. That is the honest way to read it: a first edition, not a final edition. The author is still drafting; the margin note is in pencil, not ink.
But the direction of the pencil stroke is unmistakable. In the same week, one of the world’s largest technology companies signed the largest corporate nuclear contract on record, and a nuclear developer and an AI industrial-park builder agreed on multi-gigawatt ambitions for the 2030s. Two signatures, one sentence in the margin: the energy question has moved from the back of the book to the front.
Why electricity became the first constraint
Let me step back and ask the question the numbers invite. Why did power become the binding constraint at all? The answer is arithmetic that the industry has been slow to accept. Compute capacity is, in principle, nearly infinitely expandable — add chips, add racks, add facilities. Power is not. A single AI campus draws as much electricity as a small city, and every megawatt must be generated, transmitted and guaranteed before the first chip can be switched on. The bottleneck was always going to migrate from the chip fab to the grid; the only question was when the migration would become visible.
I will correct my own phrasing there, because “migrate” is too gentle. It did not migrate; it announced itself. A 650-billion-dollar capital pipeline facing a seven-gigawatt shortfall is not a constraint being discovered. It is a constraint being met. The market’s answer is not to make the queue shorter — nothing moves fast in transmission — but to buy certainty wherever it already exists. That is why the 2,176 megawatts mattered, and why the 433 new megawatts mattered more. Certainty is the scarce good, and nuclear is the only generation type that sells it wholesale.
The arithmetic nobody wants to do
Let me do the arithmetic out loud, the way one does when a sum refuses to stay in the margin. Roughly 650 billion dollars of hyperscale capital is waiting on a power system that cannot deliver it on schedule. Seven gigawatts of planned capacity is already at risk of delay. And the queue, which is the only road into the grid, is over 2,600 gigawatts long and roughly five years deep. Put those figures next to each other and the uncomfortable conclusion writes itself: the shortage is not in the reactors, not in the data centers, not even in the capital. The shortage is in the connective tissue — the lines, the approvals, the years.
That is the detail that changes the tone of every other number in this story. The technology is not the obstacle. The finance is not the obstacle. The obstacle is a queue, which is to say an administrative process wearing a hard hat. I have read a great many documents in my time, and I have rarely seen one where the binding constraint was so plainly a form in triplicate rather than a machine.
What a twenty-year commitment forces you to think
There is something else in the term length that deserves a moment of unhurried attention. A twenty-year power purchase agreement is not an operating decision; it is a worldview. It forces the buyer to think about the year 2046 — about what electricity demand looks like when the current generation of infrastructure is obsolete, when the current corporate leadership is retired, when the current policy debates have been settled and superseded. You do not sign for twenty years because you expect certainty. You sign for twenty years because you have decided that uncertainty is the alternative, and that a fixed, predictable, almost boring source of power is the most rational thing you can hold.
Curiously, that is what makes nuclear such an appropriate chapter in this particular story. It is the opposite of fashionable. It takes a decade to build and runs for half a century. Everything about it is slow, deliberate and unglamorous — which is precisely why it suits a problem that is fundamentally about patience. The AI industry moves at the speed of a refresh cycle; the grid moves at the speed of a construction cycle. The contract is the handshake between those two clocks.
The long shelf life
There is a bookish way to end this, and it is the honest one. A power purchase agreement is a document with a remarkable property: it outlives nearly everything in the industry that signs it. Twenty-year terms span multiple model generations, multiple chip architectures, multiple corporate reorganizations. The reactors in Ohio will still be generating after the current generation of AI infrastructure is retired. In that sense, the contract is not a bet on any particular technology. It is a bet on the one thing that does not change: that computation, whatever form it takes, must be powered.
Curiously, that is what makes this the truest marginal note of the AI era so far. The technology companies have spent years annotating their own futures in terms of models, parameters and capabilities. This document annotates the future in megawatts. It is less romantic and more durable — an edition that will still be accurate when the glossier ones have gone out of print.
Let me close where I began. There is a marginal note in an old edition that says it all, and this month the AI era finally got its own — its marginalia, written not by a reader but by a signing room. The underlined sentence is a twenty-year contract; the terse “no” in the margin is a grid queue that cannot move faster than five years. Read together, they say what no press release will: the first hard constraint on artificial intelligence is not intelligence at all. It is electricity. And the readers of this particular book — the companies, the utilities, the grid — have already started writing their replies in the margins, in megawatts.