Nuclear power plant
Feature

Why AI Data Centers Are Turning to Nuclear Power

5 minute read
Sharon Fisher avatar
By
SAVED
AI’s energy appetite is exploding. Now data centers are turning to nuclear power — reviving old reactors and betting big on small modular tech.

Facing increasing power requirements to meet artificial intelligence demand, a number of AI data center operators are turning to nuclear energy. But not all nuclear-powered AI data center projects are the same. Here are some of the distinctions.

The Case for Nuclear-Powered AI

It’s no secret that AI is increasing the load data centers put on the electrical grid. At the same time, many data center operators are looking for ways to reduce the amount of carbon they put into the environment. 

That’s why they’re looking at nuclear energy — seen by many as cleaner than carbon-based energy, such as natural gas — as a way to meet the data center industry’s power needs. AI data center companies such as Amazon, Google, Microsoft and others have announced partnerships with various nuclear power companies.

“Data centers need a lot of electricity,” said Jacopo Buongiorno, professor of nuclear engineering at the Massachusetts Institute of Technology, noting that data center projections call for 50 gigawatts of energy. “That’s a huge growth in demand. They want it to be clean, carbon-free and reliable — available at the time and stable in terms of voltage and current and frequency. Nuclear plants are ideally suited to meet these requirements.”

But most of the power produced by existing nuclear plants is already in use. That means coming up with new sources of nuclear power, which takes time and money. In particular, regulatory requirements for nuclear power take years.

Related Article: AI’s Energy Bills Are Soaring. Here’s How Businesses Can Cut Costs

Bringing Old Nuclear Plants Back Online

Due to these demands, several AI data center operators are looking at restarting nuclear power projects from the 1970s and 1980s that were shut down in past years.

“There were plants that got shut down because they were not economical,” said Timothy Crook, a member of the American Nuclear Society (ANS) with prior experience in AI for nuclear, who is now working on advanced nuclear safety. “Instead of getting decommissioned, they might be restarted. They couldn’t compete with natural gas, but now there is a need, and a lot of push for that energy to be clean.”

For example, in September 2024, Microsoft entered a 20-year agreement with Constellation Energy to power its data centers using nuclear energy from the Three Mile Island Unit 1 reactor — renamed the Crane Clear Energy Center. Similarly, in March 2024, Amazon purchased a data center from Talen Energy, powered by the adjacent Susquehanna Steam Electric Station, which, according to the ANS, is the sixth-largest nuclear power plant in the US.

The Next Generation of Nuclear Power

Longer term, AI data center operators are looking at new types of nuclear technology, often using small modular reactors (SMRs) that won’t require the familiar concrete dome associated with nuclear reactors of the past. Several such projects are under development, but unlikely to be available before 2030.

One of the centers for these developments is Idaho National Laboratory (INL) based in Idaho Falls, Idaho, which hosted a workshop for data center operators last fall, said Brian Smith, director of nuclear reactor development at the organization. Participants got to see test bed facilities at INL and hear about programs such as Gateway for Advanced Innovation in Nuclear (GAIN) and the National Reactor Innovation Center (NRIC). INL already has regulatory authority, radiological controls and the physical protection necessary to help developers of new nuclear technologies go through Nuclear Regulatory Commission requirements, Smith explained.

Advanced nuclear technology projects underway include:

  • TerraPower, founded by Microsoft's Bill Gates, broke ground in June 2024 on a 345 MW sodium-cooled fast reactor with a molten salt-based energy storage system in Kemmerer, Wyoming. In January, the company signed a memorandum of understanding with Sabey Data Centers to develop a collaboration agreement.
  • Kairos, also developing a molten salt SMR, signed an agreement in October to provide power for Google data centers by 2030. In summer 2024, the company broke ground on a plant in Oak Ridge, Tennessee.
  • Xenergy, which is developing an SMR that uses a bed of graphite pebbles, has partnered with Amazon.

Smith — who’s based in the data center corridor in Ashburn, Virginia and has spoken at many data center conferences in Washington, D.C. — said that operators are eager to hear about timelines. “You don’t necessarily care that I turned one on in Idaho,” he said. “You want to know when you can plug it in.”

There will probably be a deployment by 2030, Smith added, but he wouldn’t commit to which company he thought it would be. “There are more than a handful of highly viable technology solutions out there. It may come down to being a horse race on who’s the first commercially deployed advanced reactor."

Related Article: The Billion-Dollar Data Center Boom No One Can Ignore

Why 'Behind the Meter' Matters for AI

Generating power is just part of the problem. The other part is getting the power to where it’s needed. With deregulation, many electrical utilities have decoupled generation and transmission. So, even if nuclear power generation ends up being low cost, the enormous amounts of power that AI data centers demand can result in a hefty transmission bill.

That’s why some AI data center operators are also looking at co-locating their facilities with the nuclear reactor on-site. Because they’d have access to the power before it entered the grid — a concept termed “behind the meter” — the data centers wouldn’t need to pay to transmit it.

But that solution doesn’t sit well with other users of the power grid who feel that they’d be unfairly subsidizing AI data center operators.

In November 2024, the Federal Energy Regulatory Commission (FERC) agreed with them in one such case, saying that Talen Energy — which was working with Amazon on a behind-the-meter project that was expanding existing facilities at the Susquehanna nuclear plant in Pennsylvania — had not sufficiently proved its case. However, later that month, Talen filed for a reconsideration, which is still being debated.

MIT is also researching ways to deliver new nuclear projects on schedule and on budget, Buongiorno said, including new reactor and plant designs with lower cost and simpler construction, materials that can run the reactor longer, business models that make them more profitable and advanced control algorithms that reduce the number of expensive human operators. “There’s a whole suite of projects and activities we do that aim at improving economic sustainability,” he said.

Bipartisan Backing for Nuclear Data Centers

All of this is happening against the current changing US political landscape. Former President Joe Biden issued an executive order in January directing the Department of Energy to work on nuclear powered data centers, and while President Donald Trump’s administration hasn’t said a lot about them yet, US Secretary of Energy Chris Wright had served before his appointment on the board of directors of Oklo Inc., which is working on a couple of nuclear projects, Crook said.

In fact, nuclear-powered data centers may be one issue that all sides can agree on. “It’s clear to me that nuclear is one of those energy technologies that has widespread bipartisan support,” said Smith, who served as acting deputy assistant secretary, nuclear reactors, for the Department of Energy until January.

Learning Opportunities

And once nuclear energy starts to power AI, AI may be able to return the favor, Crook said. “Models are only as good as the data they’re trained on,” he said. “If you want fusion to happen, the path for investment is to get data on advanced nuclear technologies. What gets me excited is that AI data will discover fusion for us.”

About the Author
Sharon Fisher

Sharon Fisher has written for magazines, newspapers and websites throughout the computer and business industry for more than 40 years and is also the author of "Riding the Internet Highway" as well as chapters in several other books. She holds a bachelor’s degree in computer science from Rensselaer Polytechnic Institute and a master’s degree in public administration from Boise State University. She has been a digital nomad since 2020 and lived in 18 countries so far. Connect with Sharon Fisher:

Main image: ververidis on Adobe Stock
Featured Research