AI Is Running Out Of Power

For seventy years the question that governed computing was always the same. How many chips can you afford. Faster processors, bigger budgets, larger clusters, and the only ceiling was money, which could always be raised. That era ended quietly sometime in 2024, and almost nobody announced it. The limit is no longer the chip. It is the electricity to run the chip. The most valuable companies in the world have discovered that they can buy every accelerator a factory can make and still not turn them all on, because there is not enough power coming out of the grid to do it.

In 2018, data centers used about 76 terawatt-hours of electricity in the United States, roughly 1.9 percent of the country. By 2023 it was 176 terawatt-hours, or 4.4 percent, according to the Department of Energy report from Lawrence Berkeley National Laboratory. By 2028 that same report puts it somewhere between 6.7 and 12 percent of all the electricity the nation produces. The International Energy Agency counts the same surge worldwide: about 415 terawatt-hours in 2024, more than double that by 2030, an amount of power larger than everything Japan consumes in a year. The five largest technology companies spent more than 400 billion dollars on this buildout in 2025, and they intend to spend three-quarters again as much in 2026. By the end of the decade the United States will burn more electricity running data centers than it does making aluminum, steel, cement, chemicals, and every other heavy industry combined.

That is not growth. That is a wall. It is not made of money, and it is not made of silicon. It is made of physics.

The Wall

A data center is not a website. It is a furnace for electricity, a single building that pulls the power of a city and never sleeps. The largest now under construction will draw twenty times the electricity of a town of a hundred thousand people, and that load does not spread out politely across a map. Nearly half of all American data center capacity sits inside five regional clusters. The demand lands in a handful of places, all at once, and it never lets up.

The grid was not built for that, and it cannot be rebuilt on the schedule the demand requires. A new high-voltage transmission line takes four to eight years to permit and build. The wait for the giant transformers that connect new load has doubled in the past three years. The natural gas turbines that developers want to bolt onto their sites are sold out, their order books stretching past 2030. Even the memory inside the chips is scarce, with a shortage of high-bandwidth memory expected to run through at least 2027. The IEA estimates that one in five planned data centers is now at risk of delay for a single reason. The electricity will not arrive in time.

And the demand will not wait for it. Data center power use rose 17 percent in 2025 alone. At the AI-specific sites it rose 50 percent. The demand curve points almost straight up. The supply curve is a line of permits, environmental reviews, lawsuits, and copper, and it moves at the speed of poured concrete. A frontier model is trained in a few months. A transmission line takes the better part of a decade. The most impatient industry ever built has run headlong into the most patient machine humans make, and so far the machine is winning.

The Scramble

When an industry sitting on a trillion dollars runs short of a physical input, it does not wait politely in line. It buys its way out. And what the software industry is buying should stop you cold, because it is the oldest and heaviest infrastructure on earth. It is buying nuclear reactors.

On September 20, 2024, Constellation Energy signed a twenty-year agreement with Microsoft to restart Three Mile Island, the reactor beside the site of the worst nuclear accident in American history. The plant had been dead since 2019, shut down for poor economics. Microsoft will take every watt of its 835 megawatts, and Constellation will spend roughly 1.6 billion dollars raising the corpse by 2028. Six months before that, in March 2024, Amazon paid 650 million dollars for a data center wired straight into Talen Energy's 2.5-gigawatt nuclear plant in Pennsylvania, bypassing the public grid entirely.

Then the reactors shrank. On October 14, 2024, Google signed the first corporate agreement in history to buy power from small modular reactors, up to 500 megawatts from Kairos Power, the first unit promised for 2030. Two days later, Amazon put 500 million dollars into the reactor startup X-energy and signed development deals for more than a gigawatt beyond that. By December, the operator Switch had committed to as much as twelve gigawatts of Oklo reactors. By January 2026, Meta had agreed to fund an Oklo power campus of its own. The companies that spent fifteen years telling you software would eat the world are now betting the next fifteen on uranium.

But the restart arrives in 2028. The first small reactors arrive around 2030. The transmission lines, if they were approved this morning, arrive near the end of the decade. Every one of these answers is real, and not one of them arrives in time. So while they wait, the same companies are bolting natural gas turbines straight onto the ground beside their buildings, the oldest fire humans burn grafted onto the newest thing they have built, and even that gas has to be overbuilt by 30 to 70 percent to absorb the violent swings of an AI load. The turbines, too, are back-ordered for years. There is no graceful version of this.

The Power Is Already Distributed

The industry took a workload that is global, constant, and infinitely divisible, the act of answering billions of small questions for billions of people, and it chose to pile that workload into five enormous warehouses. Then it discovered that no five points on the grid can be fed that fast, and so it is reviving nuclear plants to power buildings that never had to be that large. The power crisis is not a law of nature. It is the consequence of a decision. You only need 835 megawatts in one Pennsylvania county if you decided to put the entire computer in one Pennsylvania county.

There is another way to arrange the same arithmetic, and the electricity it requires is already flowing. Not into warehouses. Into hands and onto desks. It runs into roughly three billion smartphones, more than a billion personal computers, and the consoles and tablets and routers humming in nearly every building on earth, each one already plugged into its own local grid, each pulling a trickle, almost all of them sitting idle almost all of the time. The power that looks impossible when you stack it into five buildings becomes ordinary the moment you leave it spread across the planet, because the spreading already happened. Every one of those machines can already run these models, and every one is plugged into a socket that somebody already paid to wire.

The reactors make headlines. The electricity bills that data center towns are now sending to the families who live beside them rarely do. The centralized model is not beating physics. It is losing to physics, slowly, in four-to-eight-year increments, one transformer at a time. The decentralized model never has the fight at all, because it never builds the concentration that starts it. Said plainly, AI is not running out of power. The grid has plenty. AI is running out of power in the five places the industry insisted on putting all of it at once. The next architecture is the one that runs on the current already humming in your pocket and on your desk, and what that architecture looks like is the question this publication exists to answer.