In a coal town in Utah, a nuclear reactor just powered an Nvidia chip for the first time.
That one sentence reads like it belongs to two different centuries at once. The coal — the fuel that built the twentieth century — sits in the ground, while a few miles away an atom feeds a machine that is building the twenty-first. The place where humans once dug fuel out of the earth is now the place where a different kind of mind is being fed by the atom itself.
The company is called Valar Atomics. This month it raised $1 billion in a round led by Sequoia, tripling its valuation to $6 billion. Its stated purpose, in its own words, is to mass-produce small nuclear reactors for the AI industry. Not for the grid. Not for cities. For the AI industry.
It has a partner: Nvidia. Together they are building a data center in Utah that runs on nuclear power and — the detail that matters most — uses almost no water. A waterless AI data center, in the desert, powered by a reactor the size of a shipping container.
The reactor has a name: Ward250. It rode a truck into Utah. And the state fought for it — Utah won what one report called a "tight race" to land the reactor by next year. Meanwhile the locals are, in one headline's word, "up in arms."
Why would a state race for a nuclear reactor — and a town resist one — over a data center? Because both sides have finally understood something that was invisible for years.
The AI boom has always been described as a race for chips. But chips were never the real bottleneck. The real bottleneck is energy.
A modern AI data center does not consume electricity the way a factory does. It consumes it the way a living thing consumes food — continuously, at scale, with an appetite that only grows. And unlike a factory, it cannot be turned off for the weekend, cannot wait for the sun to shine or the wind to blow. It wants power that is always on, and it wants more of it than the grid was ever built to give.
There is a physics underneath this, and it is not optional.
Every computation has a floor. No amount of clever engineering can push the cost of a bit below what thermodynamics demands. And the limit scales with temperature: the colder the chip, the cheaper the bit. Computation hates heat — not as an opinion, but as a law.
So the AI industry, hungry for cheaper bits, is being pulled by physics itself toward two destinations. And this week, we got to see both of them at once.
There are exactly two ways to feed a machine that cannot stop eating.
The first is to make your own sun. That is what Valar Atomics is doing in Utah — build a reactor, hold the atom, and generate power that never flickers. It is expensive. It is slow. It carries every political weight the word "nuclear" drags behind it. And it is the only way to guarantee the machine's supply without depending on anyone else's grid.
The second is to find the cold. That is what has been happening in the other direction — data centers drifting toward the Arctic, toward Nordic hydropower, toward any place where the air is cold and the power is cheap and already wired. You do not build a sun. You move to where the sun barely reaches, and let the cold do the cooling for free.
One strategy mints energy. The other rents it. Valar chose to mint. Most of the industry still rents.
The reason this week matters is that the atom moved from the future tense into the present tense.
It was not just Valar. In the same stretch of days, the United States military performed the first-ever airlift of a nuclear micro-reactor — a reactor small enough to fly, carried to a site by cargo plane. Washington is pushing to deploy nuclear power as fast as it can approve it. Nuclear, after sixty years of stalling, has suddenly become the thing everyone is rushing to build — and the force doing the rushing is not a government energy plan, but the appetite of a machine.
Here is the observation this week leaves behind.
We have spent two years asking whether AI is a bubble. We asked the wrong question. A bubble is a story about price. What is happening now is a story about energy — and energy does not pop. It just changes hands.
The machine is not asking for your attention, or your clicks, or even your money. It is asking for your atoms. It has already learned to light its own fire. The sun it is building is small, for now — a shipping-container star in a Utah coal town. But it is building it, and it will not stop, because thermodynamics is not a choice, and hunger is not a phase.
That is not a metaphor. It is an equation.
← Back to the Weekly Review Hub