AI Killed Nuclear Power. Then It Brought It Back.

It's Monday, September twenty-eighth. I use AI tools all day. Every question I type, every prompt, every image I generate, somewhere a data center is burning electricity to answer it. I knew this in the abstract, the same way I know my flight burns jet fuel. When you don't see the smokestack, it's easy not to think about what's keeping the lights on.

What I didn't know until I started looking into it is that AI's hunger for power has done something almost nobody would have predicted five years ago. It has dragged the American nuclear power industry out of early retirement and kicked off the biggest energy buildout in a generation. The same tech companies that spent a decade buying wind and solar are now racing to lock up nuclear reactors, including existing ones, restarting ones, and designs that don't exist yet, because AI doesn't sleep and neither can the grid.

In 2023, data centers consumed four point four percent of total US electricity. That was 176 terawatt hours. One terawatt hour powers about 70,000 homes for a year.

By 2028, according to the Lawrence Berkeley National Laboratory, that number is projected to hit somewhere between six point seven and twelve percent of all US electricity. The high end is 580 terawatt hours, which is more than the entire state of California uses.

Globally, the International Energy Agency estimates data centers consumed 460 terawatt hours in 2022. By 2026 that could more than double to over 1,000 terawatt hours. That is more than a third of the electricity generated by every nuclear power plant on earth combined. Roughly the entire electricity consumption of Japan.

And AI is the biggest driver. A single ChatGPT query uses roughly ten times the electricity of a Google search. Training a large language model consumes somewhere between 50 and 100 gigawatt hours, the equivalent of powering several thousand homes for an entire year. The models keep getting bigger. Meta's next generation Llama model is reportedly being trained on a cluster of over 100,000 GPUs. The power draw from a cluster that size is comparable to a small city.

The five largest tech companies, Alphabet, Microsoft, Amazon, Meta, and Apple, spent a combined 59 billion dollars on capital expenses in a single quarter last year, up 63 percent from the year before. Most of that is data centers. There are more data centers under construction in the US right now than there are active data centers, according to Dan Ives, an analyst at Wedbush.

The North American grid is not ready for this. The North American Electric Reliability Corporation warns that much of the continent is at risk of supply shortages. The PJM grid, which covers 13 states including Virginia, the biggest data center hub in the world, is the most exposed. Forecasted data center expansions alone have pushed its capacity costs to 16 billion dollars.

AI needs power. A lot of it. Every hour of every day. Data centers do not care if the sun is shining or the wind is blowing. They run at full capacity, constantly.

So the tech companies went shopping for energy. For most of the last decade, big tech was the largest corporate buyer of wind and solar in the world. Every major company had a pledge to be carbon neutral by 2030.

But wind and solar are intermittent. They generate power when nature cooperates, not when your data center needs to train a model at 3 AM. Natural gas is reliable and quick to build. Meta is quietly signing gas deals for its largest AI campus in Louisiana. But burning gas clashes directly with those net zero pledges.

Nuclear is different. It produces zero carbon emissions. It runs at more than 92 percent capacity, generating power nearly all the time, every hour of every day. Compare that to solar at roughly 25 percent and wind at about 35 percent. Nuclear is the only carbon free source that matches the 24 hour demand profile of a data center.

Until very recently, the nuclear industry was shrinking. The US had 104 operating reactors at its peak in the 1990s. More than a dozen shut down since 2013. Plants were closing because cheap natural gas made them uneconomical, not because of safety concerns or aging equipment. The industry was in slow retreat, and everyone assumed that retreat would continue.

AI flipped the math completely. Suddenly there is a customer willing to pay above market rate for 24 hour carbon free electricity, locked in for 20 years. That changes the economics of every reactor still standing and every new design ever drawn up.

Here is what has happened in roughly the last two years.

Microsoft. In September 2024, Microsoft signed a 20 year deal with Constellation Energy, the largest nuclear operator in the US, to restart Unit 1 at Three Mile Island in Pennsylvania. Unit 1 was shut down in 2019 because cheap natural gas made it unprofitable. Unit 2 next door is the one that partially melted down in 1979, and it is still closed.

Microsoft is buying the entire 835 megawatt output for two decades. Constellation is spending 1.6 billion dollars to restore it, backed by a 1 billion dollar federal loan. A 262 ton transformer arrived from South Korea last month. The plant is expected online in the second half of 2027. Three Mile Island. The name is synonymous with nuclear disaster. And it is reopening because Microsoft needs power for Copilot and Azure.

Google. The same month, Google signed a contract with Kairos (KYE-ross) Power to buy 500 megawatts from a fleet of small modular reactors, with the first expected online in 2030 at Oak Ridge, Tennessee. No commercial reactor of this design has ever been deployed in the United States.

Amazon. The same week, Amazon announced a 700 million dollar investment in X-energy for up to 12 small modular reactors in Washington State, plus a 650 million dollar data center campus purchase in Pennsylvania that will be directly powered by the existing Susquehanna nuclear plant. Total investment on that campus alone exceeds 20 billion dollars.

Then came Meta. In January 2026, Meta became the single largest corporate buyer of nuclear power in the world, committing to up to 6.6 gigawatts, enough to power roughly five million homes. The biggest piece: up to eight Natrium reactors from TerraPower, the nuclear startup founded by Bill Gates, with the first two online as early as 2032. Meta also signed with Oklo (OH-kloh), the company backed by Sam Altman, for another 1.2 gigawatts, and locked in power from existing nuclear fleets.

In total, as of mid 2026, tech companies have signed 13 nuclear deals committing nearly 10 gigawatts. That is roughly the equivalent of ten large traditional reactors. The entire US fleet is about 95 gigawatts. The Palisades (pal-ih-SAY-deez) nuclear plant in Michigan, shut down in 2022, is also being restarted. Google and NextEra Energy announced a deal to restart a plant in Iowa. Small modular reactor startups are everywhere, funded by tech money, racing to be the first to market.

After four decades of retreat, the American nuclear industry is sprinting.

We talk about AI in terms of intelligence: benchmarks, capabilities, reasoning. We almost never talk about it as a physical thing that consumes concrete, steel, cooling water, and uranium rods. But that is what it is. Every time the models get bigger, so does the meter. Nuclear plants that sat cold for years are being dusted off and refueled because we want smarter chatbots.

Nuclear power spent 40 years as the thing nobody wanted to touch. Its reputation was defined by Three Mile Island, Chernobyl, Fukushima. Cheap natural gas made it a bad investment. And the thing that finally turned it around was not a government program or a climate treaty or a breakthrough in reactor design. It was the fact that you and I keep typing questions into a text box, and the computers that answer never sleep.

Stay informed, stay curious, and we'll see you tomorrow.