Is restarting Three Mile Island really a good idea?

By 2027, AI might use as much power globally as the entire country of Sweden uses now. As tech companies try to figure out how to run their data centers without trashing their climate goals, some are turning to a power source that would have seemed unlikely a few years ago: restarting old nuclear reactors.

In September, Microsoft announced a new deal to buy energy from Three Mile Island in Pennsylvania, the site of a partial meltdown in 1979—the worst nuclear accident in U.S. history. The damaged reactor was shut down decades ago; another reactor kept running until 2019, when it closed for financial reasons. Microsoft’s insatiable need for energy is bringing it back. In Michigan, energy demand from AI is one of the reasons that a recently-closed nuclear plant now plans to reopen next year. Another utility in Iowa is considering restarting a shuttered nuclear plant. The Biden administration hopes to reopen more as part of the fight against climate change.

Separately, some tech companies are considering building new nuclear power plants to supply their data centers. Other nuclear plants will stay open longer than expected, like one in California that was scheduled to close next year will now stay open while it seeks a 20-year extension.

From an emissions perspective, nuclear’s resurgence is a good thing: burning nuclear fuel doesn’t create greenhouse gas emissions. Global leaders now hope to triple nuclear energy production to meet climate goals. But given the obvious risks—and the fact that there’s still no good way to deal with nuclear waste from spent fuel—does a focus on nuclear power really make sense?

We already needed more energy—but AI sped up the timeline

AI isn’t the only reason that electric grids are strained. Transportation is electrifying, from cars to buses and trains. Buildings are adding heat pumps and induction stoves. In the U.S., new clean energy factories, making solar panels, batteries, and EVs, are also adding to energy demand. But the growth of AI is adding pressure faster.

“We basically have to decarbonize the entire economy and the only way we can do that is by moving away from the use of fossil fuels and relying on electricity supplied predominantly by zero carbon resources,” says Jackson Morris, director of state power sector policy at the nonprofit NRDC. “That’s not a new trend. However, this recent uptick from data centers largely means that some of those projections that maybe were 10 years out or even 15 years out are happening in the next three or five years.”

It’s hard to predict the exact demand, both because AI’s growth is new and because it’s not clear how quickly data centers and AI itself may be able to adapt to reduce energy use. But energy demand is clearly growing faster in some regions than others. “We already see it in utility demand forecasts, which are telling us that these large customers are coming quickly, and in geographically concentrated ways,” says Michael O’Boyle, senior director of energy at the think tank Energy Innovation. In Virginia, for example, Dominion Energy expects data centers to add a massive 20% to its peak demand by 2030.

The sense of urgency could potentially push risky projects forward. “The industry is trying to create the pressure on the regulator to really cut back on even the most fundamental safety issues or considerations,” says Edwin Lyman, director of nuclear power safety at the nonprofit Union of Concerned Scientists. “That dynamic could lead to mistakes that could be really costly in terms of increasing risk.” (Of course, some utilities may also rush to build new polluting fossil fuel power plants because they think that’s the fastest way to meet demand.)

Uncertain risk

The nuclear power industry has always grappled with understanding risk. From the beginning, companies like GE and Westinghouse (both of which had been part of the Manhattan Project) refused to take on the liability for accidents at potential new nuclear power plants. Congress eventually passed a law limiting private company liability, and putting most of the liability on the public. “Without that, the industry would not have developed,” Stephanie Cooke, a journalist who’s covered the nuclear power industry for decades, explained on the Energy Transition Show podcast.

Government documents show that the agencies in charge—the Atomic Energy Commission, and then the Nuclear Regulatory Commission—weren’t really sure about the precise risk. The calculations were “expert guesswork,” Thomas Wellock, a historian for the NRC, wrote in a detailed report a few years ago. In a nuclear power plant, with around 20,000 safety components, “there were innumerable unlikely combinations of failures that could cause an accident.”

It’s still hard to calculate risk. “Despite decades of development, the art of probabilistic risk assessment for nuclear plants is still plagued with large uncertainties and a great deal of guesswork, which is sometimes called ‘expert elicitation’—essentially getting a bunch of experts in a room and tabulating their guesses,” says Lyman. “Some of the most questionable aspects of these calculations are the estimated frequencies of natural disasters such as severe earthquakes and floods—the latter becoming increasingly difficult to predict in the face of climate change—and the success rates assumed for personnel when they are required to carry out certain actions in the course of an accident.”

The nuclear accident at Fukushima in 2011, which damaged four reactors and displaced more than 160,000 people, was unlikely, but it happened. Even if the risk is small, it’s possible that aging nuclear plants in the U.S. could also have accidents, with the potential for radiation sickness or increased risk of cancer or other disease for people who are exposed to high doses. Nuclear waste is another challenge. More than a quarter million metric tons of waste that could be dangerously radioactive for thousands of years currently sits near nuclear power plants and weapons facilities, waiting for geological storage facilities that aren’t yet open.

Challenges at decommissioned plants

At the nuclear plant in Michigan that plans to restart, there’s now evidence of corrosion on heat exchangers—thousands of thin metal tubes that have always been a safety issue for this type of reactor, according to Lyman. “These are very old plants to begin with, but it wasn’t originally put into a state where they expected to start it again,” he says. “That means that the environmental conditions for a lot of components weren’t optimal, and they’re starting to see major problems.”

Holtec, the company running the plant, said in a statement that “we expect to encounter conditions that may have been an unknown initially but as we work towards a restart we will continue to address issues as they arise,” and that it expects to be able to repair the problem within the timeline it originally planned.

The Michigan plant was only closed for two years. The plant at Three Mile Island, which was closed longer, could potentially have more problems of the same kind. Before either plant can come back online, they’ll have to pass safety inspections by the Nuclear Regulatory Commission. It’s not certain yet that will happen. In the meantime, the federal government is helping boost the projects. The Biden administration just finalized a $1.52 billion loan guarantee to the project in Michigan. The project at Three Mile Island is seeking $1.6 billion loan guarantee. Both will also get tax credits for clean energy production.

Even if a plant is managed perfectly and doesn’t face disasters like earthquakes or flooding, sabotage is also possible. Despite the risks, environmental groups often support nuclear power; NRDC, for example, says that it evaluates each project on a case-by-case basis. Others, like Greenpeace, argue that nuclear power can never be a “green” energy source.

‘Advanced’ nuclear power also has challenges

There are only a few decommissioned nuclear power plants that could feasibly be restarted. If more data centers rely on nuclear power, they’re more likely to turn to newer nuclear tech that’s billed as safer and more sustainable. But a 2021 report from the Union of Concerned Scientists found that those claims aren’t necessarily true. One new reactor from a startup called TerraPower, for example, is supposed to produce less nuclear waste. The UCS report says that this type of “sodium-cooled” reactor is likely to use uranium less efficiently, and wouldn’t reduce waste. It also could have new safety issues because sodium coolant can burn when exposed to air or water.

“The debate in recent years has really been heavily distorted by the entrant of really a few deep-pocketed Silicon Valley billionaires who have gotten interested in nuclear power,” says Lyman. “I think it wasn’t an accident. I think there’s a concerted effort by the elements of the nuclear industry to try to get private capital on board with nuclear power.”

As investors like Bill Gates and Sam Altman have backed nuclear startups, “their lobbying and public relations machines have gone into overdrive to promote the idea that nuclear power is great, the safety issues have been overblown, and that there’s this new generation of plants on the horizon that’s going to disrupt the whole industry,” Lyman says. “The hard reality is that it just isn’t a technology that’s really ripe for that kind of disruption.”

Renewable energy can scale up

Solar and wind power plants, with battery storage, are safe and cheaper to build and operate; other clean energy sources, like next-generation geothermal systems, can also help power data centers. Technically, these sources could “absolutely” scale up enough to meet new energy demand from data centers, says Michael O’Boyle, senior director of energy at the think tank Energy Innovation, but it will require overcoming some challenges that have held renewable energy back. It “will require better interconnection processes, better utility planning processes, and addressing some of the local restrictions on building wind and solar,” he says.