As I’ve written before, the growing number of huge data centers is quickly ramping up the demand for more electricity. AI data centers typically use 200 megawatts to one gigawatt, as opposed to older data centers that operate in the range of 10 to 30 megawatts. This raises the immediate question of where that energy is coming from. And it’s not just data centers, there are also EVs and other large electrification efforts. So, the idea of using nuclear reactors is seen in some quarters as the ideal solution. They produce steady reliable power 24/7, create no emissions, and have exceedingly long lifetimes.
How Much Power Is Needed?
According to a December 2024 article in the Bulletin of the Atomic Scientists , “U.S. electric utilities have nearly doubled their estimates of how much electricity they’ll need in another five years. Electric vehicles, cryptocurrency, and a resurgence of American manufacturing are sucking up a lot of electrons, but AI is growing faster and is driving the rapid expansion of data centers. A recent report by the global investment bank Goldman Sachs forecasts that data centers will consume about eight percent of all U.S. electricity in 2030, up from about three percent today.”
The leaders of the major AI companies feel they need to pursue AI’s rapid growth while also meeting their climate commitments. So, they have all come to the same conclusion: Nuclear energy, whatever it costs, is the only viable solution. I have therefore decided to delve into the pros and cons of using nuclear energy to feed the AI data centers.
Nuclear Pros
Electrical energy from nuclear reactors is generated when fissionable material — uranium 235 — sustains a chain reaction, which heats a coolant to about 300 °C. The coolant, usually water, heats up water in a separate loop turning it to steam, which spins a turbine to generate the electricity.
Traditional nuclear power plants produce about 1000 megawatts, but currently there is an emphasis on smaller reactors that produce about 300 megawatts. According to the International Atomic Energy Agency (IAEA), these small modular reactors (SMRs) “work particularly well for data centers because they are designed to be built in segmental units, making phased deployment possible. As an AI cluster expands, so can its nuclear power source.”
“They can operate close to industrial zones, including data center campuses. With SMRs, tech companies can avoid dependence on constrained regional electricity grids and reduce transmission losses.” SMRs are designed to deliver up to 300 megawatts per module so they are a perfect match for the AI centers.
Not only are SMRs much smaller than a full-scale reactor — about 15 feet across compared to about 140 feet, but depending upon the technology, they can be much safer. One of the dangerous features of traditional full-scale reactors is that since they are cooled by water at more than 300 °C, extremely high pressure — about 2,250 pounds per square inch — is needed to keep the water in a liquid state. Special piping and welding procedures are needed for those conditions. If one of those pipes ruptures, there is the possibility of wide-spread radiation. For years, there have been a number of these events, with burn injuries and some deaths among on-site workers, but fortunately, so far there has been little dangerous radiation leakage. So, another advantage for using SMRs is that they do not need high-pressure water for cooling, rather they can use different coolants such as molten salt that don’t require high pressure.
Also, once regulatory authorities approve the designs, SMRs can be built relatively quickly and can even be sited next to places like chemical plants that can directly use their heat.
Nuclear Cons
The first commercial SMR was invented by researchers at Oregon State University in 2007, and it wasn’t until 2013 that their technology was used by the private company NuScale Power, Corvallis, OR, to develop the first full-scale prototype. Their SMR was finally approved by the U.S. Nuclear Regulatory Commission (NRC) in 2022. Two more models were approved in 2025. But, at present, in January 2026, there are still no SMRs in actual commercial operation. After a number of delays, it is expected that they will be operational by 2030. But the mammoth AI buildout is not waiting, they will need lots of power a lot sooner — in the next three to five years. So, while SMRs seem like a good idea, several of the tech companies are talking about restarting some of the old decommissioned full-scale reactors.
According to an article in the MIT Technology Review , Microsoft signed a long-term power purchase agreement with Constellation, the owner of the Three Mile Island Unit 1 nuclear plant in Pennsylvania, which was not damaged in the 1979 meltdown. It was shut down in 2019 and is targeted to restart in 2027.
The other restart project is the Palisades Nuclear Power Plant in Covert, MI. It was initially scheduled to restart in 2025, then “early 2026” — the current estimate is for late March of this year.
Looking at the kinds of work needed to recommission these plants frightens me because it speaks to all the possible ways things can go wrong. According to an article in the Engineering News Record , “Steam generator upgrades [at the Palisades plant] are a critical area of work after an inspection last year found that 1,400 cooling tubes were cracked and needed repair,” and there are similar issues at Three Mile Island. What does that imply for the future? There are a great many critical systems in nuclear power plants and to my mind that means there are a great many things that can go wrong.
What Can Possibly Go Wrong?
The Bulletin of the Atomic Scientists article delves into some of the specifics of what can go wrong with nuclear plants, large and small. “... both the government and the tech industry are largely ignoring the known and significant downsides of nuclear power —including high costs, long construction times, accidents, nuclear weapons proliferation risks, and environmental contamination from uranium mining and radioactive waste disposal.”
I was amazed to learn that the U.S. has no uniform plan for storing spent nuclear fuel, which has a typical half-life of 30 years. According to the Carnegie Endowment for International Peace , “It is the policy of the NRC to store spent nuclear fuel in specially designed pools at reactor sites for a certain period before placing it into dry casks. The United States does not have a program for consolidated interim storage, let alone a permanent repository, for used power reactor fuel, meaning that nuclear power states will be de facto nuclear storage sites. Over time, this could pose increased risks to surrounding communities through groundwater contamination from radioactive materials such as tritium resulting from system leaks, condensation, and evaporation.”
For SMRs there is the additional risk that since they are widely dispersed, they will require more transportation of nuclear fuel, and they will also increase the amount of spent fuel that needs to be dealt with.
And there is little consideration of the risk that fissile material produced by nuclear plants, which according to the Bulletin of Atomic Scientists, “bad actors could use to make nuclear weapons. [Some of the nuclear startups] propose to reprocess their waste to keep costs down. But that reprocessing produces plutonium that could be diverted for use in nuclear weapons.”
Another significant problem is the need to recruit a skilled workforce, who need to be certified by the NRC for each specific plant.
And, not least, there is a danger that pressure from private industry to speed up the nuclear buildout could influence the government to weaken the NRC’s regulatory capacities, which could prove very dangerous.
My Bottom Line
The Carnegie Endowment article best sums up my thoughts about the prospects of nuclear power as the solution for AI’s energy needs. “...for tech entrepreneurs who are pouring billions of dollars into nuclear energy to run energy-intensive AI operations in their data centers, timing is key. They might lose interest if the nuclear energy fleet does not join the grid when they need it the most — in the next two to five years. Not only could this lead to a plethora of unfinished nuclear reactor projects across the country that serve no economic purpose, but the push for faster results itself might lead to cutting corners on safety, exacerbating risks for host communities.”
