How the Lights Stay On · The Atom Comes Back

The Atom Comes Back, part 3

The trouble is that bigger plants also become enormous custom construction projects that take fifteen years to build and routinely run over budget. Each one is a first-of-a-kind. Each one requires assembling thousands of skilled workers in one place for a decade.

So a small group of nuclear engineers, starting around the late 2000s, had a different idea. What if the reactor were small enough to be built in a factory? What if instead of pouring concrete on a site for a decade, you trucked a reactor in?

These are called small modular reactors, or SMRs. They range from a few megawatts (small enough to power a remote town or an industrial facility) up to about 300 megawatts (small enough that you can make several of them in parallel on an assembly line).

The bet is that what nuclear lost over the last fifty years was not the technology — the technology is fine — but the manufacturing experience. By making reactors smaller, simpler, and more standardized, you might be able to relearn how to build them, the way the aerospace industry learned to build jets.

Four American companies are leading the bet, each with a different approach.

NuScale, based in Oregon, has the most mature design. Each NuScale module is 77 megawatts and uses standard light-water reactor technology — the same kind that has powered submarines and nuclear plants for sixty years, just made smaller. NuScale is the only US SMR design that has received NRC design certification. In September 2025, the Tennessee Valley Authority signed an agreement to deploy up to 6 gigawatts of NuScale capacity across its seven-state region — the largest SMR commitment in the country.

Kairos Power, based in California with reactor development in Tennessee, uses a more radical design. Its reactors are cooled by molten fluoride salt instead of water. The fuel is in the form of small ceramic-coated pebbles — TRISO fuel, the design that can withstand extreme temperatures without releasing radiation. Kairos has a partnership with Google to provide up to 500 megawatts of nuclear power by 2035. The first test reactor, called Hermes, is expected to operate at low power in 2026 in Oak Ridge, Tennessee.

TerraPower, founded by Bill Gates, is building a 345-megawatt sodium-cooled reactor called Natrium. The site is in Kemmerer, Wyoming — a former coal plant. Natrium pairs the reactor with a molten salt heat-storage system, so it can ramp output up and down to follow demand, which conventional reactors cannot easily do. The NRC is expected to issue a construction permit decision in the first half of 2026. Meta has signed an agreement with TerraPower for up to eight Natrium plants. NVIDIA has invested.

X-energy, based in Maryland, is building a high-temperature, gas-cooled reactor called the Xe-100. Its first commercial deployment will be at a Dow Chemical facility in Texas. Amazon has invested in X-energy and signed for several reactors.

None of these reactors has yet produced commercial electricity. The earliest realistic dates for any of them are 2028 to 2030, with most deployments clustering in the early-to-mid 2030s. SMRs are a real bet, but they are still a bet. China and Russia, both ahead of the United States in deployment, have working SMRs. The United States does not, yet.

The most striking thing about the new American nuclear story is who is paying for it.

It is not, primarily, the federal government. It is not, primarily, the traditional utilities. It is — overwhelmingly — the technology companies.

In addition to Microsoft and the Crane Clean Energy Center, here are the other major deals as of 2026.

Amazon Web Services bought a data center campus directly next to the Susquehanna nuclear plant in Pennsylvania in March 2024 — the first deal in which a data center co-locates with a nuclear plant. Amazon also invested in X-energy.

Google signed a 500-megawatt deal with Kairos Power for advanced reactors by 2035.

Meta signed for up to eight Natrium reactors from TerraPower.

Oracle announced plans for data centers powered by three small modular reactors.

NVIDIA, the chipmaker whose GPUs power most of the AI industry, has invested in TerraPower.

OpenAI joined Amazon, Google, Meta, Microsoft, Oracle, and xAI in signing the Trump administration’s “Ratepayer Protection Pledge” in March 2026 — committing to build, bring, or buy their own power and pay for the grid infrastructure upgrades their data centers require, rather than passing the cost to ordinary electric customers. The pledge is voluntary and not federally enforceable (electricity regulation lives mostly at the state level), but the public commitment is real.

By the spring of 2026, more than 10 gigawatts of nuclear capacity has been committed by hyperscalers and AI companies. Most of it will not come online for several more years. Some of it may never be built, if the technology fails to deliver or the projects fall apart. But the financial commitment is real, and it is the largest American investment in nuclear power since the 1970s.

Here is the part to sit with. For four decades, the case for new American nuclear power was made by environmentalists, scientists, and a small fraction of utilities, mostly without success. The case is now being made, and paid for, by the companies building artificial intelligence.

This is a genuinely strange alliance. It is also one of the most consequential energy stories of the decade.

It would be tempting to call this a “nuclear renaissance” and move on. The phrase has been used so many times over the last twenty years that it has lost most of its meaning.

The honest version is more careful.

American nuclear power is reopening some closed plants. It is building one type of large reactor — the AP1000 design at Vogtle — very slowly and very expensively. It is making serious progress on at least four different small modular reactor designs, none of which has yet produced commercial electricity. It is signing big-name deals that may or may not be built. And it is doing all of this against the backdrop of a still-unsolved waste problem, a public memory of three major accidents, and a track record on cost that does not, in the American context, inspire confidence.

It is also doing it at a moment when the world genuinely needs more firm, clean electricity than anyone knows how to build. The demand from data centers alone, by 2030, will be larger than the entire current output of every nuclear plant in the United States combined. Renewable energy is being built faster than anyone predicted, but renewable energy alone cannot solve the problem if the goal is also reliable, around-the-clock, low-carbon electricity. Some new firm clean power is going to be built. Some of it will probably be nuclear. How much is the open question.

If you are reading this in middle school or high school, you will live through the answer. You will see the SMRs come online or fail to come online. You will see the Vogtle experience repeated or improved. You will see whether the Microsoft-Constellation kind of deal becomes standard or stays unusual. You will see, in your lifetime, whether the United States can build large clean infrastructure or whether the Vogtle pattern repeats.

A few years from now, if all goes as planned, the cooling towers at Three Mile Island will begin to vent steam again.

Most days, the steam will be a thin white plume rising off the towers’ tops, dissipating into the Pennsylvania sky. The reactor inside Unit 1 will be running quietly, producing about 835 megawatts of electricity — enough to power something like 800,000 American homes, or, in the new arrangement, a chunk of the AI infrastructure that will help train and run the artificial intelligence used by hundreds of millions of people.

The plant will produce essentially no carbon emissions in operation. It will run 24 hours a day, 7 days a week, 365 days a year, with brief outages every eighteen months for refueling.

It will also be sitting next to Unit 2, the reactor that suffered a partial meltdown in 1979 and has been silent ever since. The two reactors will be on the same island in the Susquehanna River. The same towers will be visible from the same highway. The plant will have, in some ways, gone full circle.

Whether you think this is the right way to power the future, or the wrong one, or one of several reasonable ways, is a question this book is not going to answer for you. There are intelligent, honest people who believe nuclear power is the only realistic path to a clean grid. There are intelligent, honest people who believe it is dangerous and uneconomic and we should not build any more of it. The arguments are real on both sides. The trade-offs are real. The choices are being made right now.

The cooling towers, in any case, will be back. Whether they will be the start of a real American nuclear comeback, or a one-time symbol of a moment when tech companies briefly needed something only nuclear could deliver, will be one of the most important stories of the 2030s.

It is not yet decided. The thing about being twelve or fourteen years old when you read this book is that you may be the one who decides.

All 18 chapters of How the Lights Stay On · Library · All courses