Nuclear Power Is the New A.I. Trade. What Could Possibly Go Wrong?
Want to make a wager on the future of artificial intelligence? Nvidia is the obvious bet. It designs the chips and software that make A.I. run, is the stock market favorite and has gained 150 percent this year alone.
But some stock traders have found a less obvious, backdoor choice: utility shares, specifically those of companies that own nuclear power plants.
Two such companies — Constellation Energy and Vistra — were the top performers in the S&P 500 in September through Thursday, with returns of more than 30 percent. Shares of Constellation Energy, the biggest nuclear plant operator in the United States, surged when it signed a deal on Sept. 20 to supply Microsoft’s burgeoning A.I. data centers with energy from the Three Mile Island nuclear power plant near Harrisburg, Pa.
If you were a consumer of news in 1979, or are aware of the history of nuclear power, you will know that Three Mile Island was the site of the worst nuclear disaster in U.S. history. Now, because of A.I. energy demand, the utility intends to reopen an undamaged part of Three Mile Island that was mothballed five years ago and change its name, to the less evocative Crane Clean Energy Center.
Vistra says it, too, has been in talks with several A.I. operators. And the Constellation deal with Microsoft followed an A.I.-nuclear power agreement in March involving Talen Energy and Amazon. Talen agreed to sell Amazon large quantities of electricity from the Susquehanna Steam Electric Station — a nuclear plant near Berwick, Pa., in which the utility has a 90 percent ownership stake.
Consider these stock returns for the year, including dividends, through Thursday:
-
The S&P 500 stock index, 21 percent.
-
Constellation Energy, 121 percent.
-
Vistra, 199 percent.
-
Talen Energy, 178 percent.
Other utilities have gotten a boost from A.I., too, but generally not as spectacularly as these companies.
Additional inactive nuclear plants, like the Palisades plant in Michigan and the Duane Arnold plant in Iowa, are under consideration for reactivation, an arduous and expensive process that requires major investment and regulatory approval. Nuclear plants that are operating and supplying power for other purposes may be shifted to more lucrative operations with A.I. data centers down the road.
In a conference call last month, Robert Blue, chief executive of Dominion Energy, said the company “is certainly open to the idea” of such a deal at its Millstone Power Station, which began operations in Waterford, Conn., in 1970.
It’s not just nuclear power plants that may have their lives extended. Old coal-fired power plants scheduled for retirement are being given a new lease on life. Heightened demand for power from A.I. data centers is outstripping the ability of energy producers to build and operate cleaner power generators using wind or solar power. That comes at a cost in air pollution, including carbon emissions that contribute to global warming.
Nuclear power plants, needless to say, have drawbacks of their own. For one thing, even the best of them produce radioactive waste that can’t be eliminated in a human lifetime. But nuclear plants don’t burn carbon or spew tiny, lung-damaging particles into the atmosphere. While it’s hard and costly to build a new one and get it approved by regulators, old nuclear power plants are very much in demand.
Yet for anyone steeped in the environmental issues associated with nuclear power, it is mind-boggling to contemplate a new bonanza in nuclear energy stocks, propelled by the A.I. boom.
An Insatiable Need
As a reporter for Newsday on Long Island, I covered nuclear power in the early 1980s, as it entered a period of steep decline from which it never fully recovered.
In the wake of Three Mile Island, many people around the world feared nuclear catastrophe. Bipartisan opposition led to the demise of a completed but never fully operational commercial nuclear plant at Shoreham on Long Island’s picturesque north shore. The state took over the Long Island Lighting Company, the formerly blue-chip utility that built Shoreham.
I never dreamed back then that in 2024, artificial intelligence would cause a stock market boom for utilities with nuclear power assets. That was beyond my forecasting ability, and, I think, that of any human. A.I may be able to help with that kind of forecasting one day, though I don’t think it’s ready now. But A.I.’s hunger for energy is already insatiable.
It needs energy in two main ways. First, advanced A.I. runs on supercomputers, filled largely with Nvidia equipment, that suck up enormous amounts of electricity to crunch the data that “trains” the A.I. systems and gives them something to say, embellish and transform.
Second, these supercomputers need vast power so A.I. chatbots can respond to questions and conduct searches. When you ask a chatbot for help with your Spanish homework or with a cooking recipe or for information about a stock, even that simple search on a bot like ChatGPT devours electricity.
How much electricity, exactly? I spoke with Jesse Dodge, a senior researcher at the Allen Institute for A.I., a Seattle nonprofit, who has worked extensively on this subject. He and his colleagues estimated the kilowatt-hours consumed by advanced Nvidia hardware on a simple search, and he summarized it this way: “We estimate that one query to ChatGPT could use as much electricity as could light one lightbulb for about 20 minutes.”
What’s more, he said, even A.I. searches done automatically and unintentionally — when, say, you run a traditional search on Google and it also offers you A.I. answers — require significant amounts of energy.
A.I. does some things well, but many searches produce trivial or nonsensical answers. “A.I. hallucinates, and its answers often aren’t reliable,” Dr. Dodge said.
Widespread use of A.I. by consumers is burning up energy — and using other precious environmental resources, like water to cool the power plants and data centers — that would be difficult to justify if people just thought about it, he said. “This is a problem that is getting bigger and bigger.”
It’s a global problem, but also a regional one, because consumers prize quick answers, and there’s less delay when A.I. supercomputers and their power sources are close to the major metropolitan areas where most consumers live.
We’re way beyond the experimental phase of artificial intelligence, with a handful of researchers performing occasional, energy-intensive searches. The field is mushrooming with millions of people conducting A.I. searches and new applications coming all the time. That’s why many researchers estimate that the total energy expenditure of A.I. is moving into globally significant totals — consuming somewhere in the realm of the energy use of a country like Sweden or Argentina now, and maybe, by 2030, more than India. Who knows where A.I. will end up?
I wouldn’t say the sky’s the limit. There’s considerable skepticism in some investing circles. No doubt many of the new A.I. applications will be splendid, but it’s not clear how useful or profitable the technology will be. Goldman Sachs issued a report in June with the provocative title “Gen AI: Too Much Spend, Too Little Benefit?”
As an investor, I avoid wagers on the future of any particular innovation and put my money into broad, low-cost index funds that track the entire market. I’m no more capable of judging A.I.’s ultimate potential than I am of assessing nuclear power’s long-term future in the marketplace. Is the current stock surge the start of a long-term revival? I don’t pretend to know.
But enthusiasm for artificial intelligence is still growing. A.I. needs energy, even if it has to come from Three Mile Island. Backdoor bets on A.I. through nuclear power utilities are, at this moment, an odd and creative alternative to buying Nvidia.