If you had told an investor in 2023 that “nuclear stocks are AI theme stocks,” they would have thought you were talking nonsense. But since 2024, the market has given a very direct answer with stock prices:
Constellation Energy has multiplied several times over in two years, and names like Vistra, Talen Energy, and Public Service Enterprise Group have all been collectively re-rated.
The market is telling you that existing nuclear assets have suddenly become extremely scarce in this AI wave.
Why has nuclear power been re-rated?
In one sentence: the electricity AI data centers need, and the electricity existing nuclear power can provide, happen to be exactly the same kind of electricity.
AI data centers have three characteristics when it comes to power consumption.
First, the scale is huge; a GW-scale campus is no longer news.
Second, the load is stable. Once GPU training starts, it can run for weeks or months, which is completely different from residential electricity use with daily and nightly peaks and troughs, and it requires baseload power.
Third, it is extremely sensitive to outages. An interruption of just a few hours can ruin a training run, so high availability is also required.
Wind and solar plus storage can solve scale and cleanliness, but they are inherently a tier weaker when it comes to baseload stability. Natural gas can provide baseload, but carbon emissions and long-term fuel prices are problems. The only remaining option that can simultaneously satisfy “large scale, baseload, low carbon, and high availability” is nuclear power.
The problem is that nuclear supply is fixed.
The United States currently has about 94 commercial reactors, with nearly 100 GW of installed capacity in total, and there has been basically no net new capacity added over the past twenty years. It takes more than ten years from approval to grid connection to build a new reactor, and commercial SMR small modular reactors will not arrive until around 2030.
So in the five-to-seven-year window ahead, the nuclear power that AI can immediately use is only this batch of existing units. With fixed existing supply and explosive incremental demand, prices are naturally driven up.
The hyperscale cloud companies have already voted with real money.
The most symbolic deal is Constellation restarting Unit 1 at Three Mile Island and signing a 20-year, 835-megawatt power purchase agreement with Microsoft, with the contract value estimated at about $16 billion.
Amazon directly bought the campus next to Talen’s Susquehanna nuclear plant in Pennsylvania and signed a 1,920-megawatt PPA.
Google and Meta have also successively announced their own nuclear power agreements.
By 2026, the entire AI industry has already locked in more than 9.8 GW of nuclear capacity for AI infrastructure. These are not expressions of intent; they are 10- to 20-year contracts that lock in the future cash flows of existing nuclear power.
This is a very typical example of a “transition-period winner.”
Existing nuclear power was originally valued by the market as a sunset asset, with stable cash flow but no growth story, and valuation multiples that had long stayed in the single digits on P/E.
When AI came along, it suddenly became one of the scarcest clean baseload power sources over the next decade. Contract prices jumped from the past three or four cents per kilowatt-hour to seven or eight cents, or even higher, and together with the certainty brought by ultra-long contracts, the entire valuation framework was replaced.
Bloomberg’s estimate is that by 2050, U.S. nuclear installed capacity may grow by 63% to 159 GW due to AI demand, corresponding to more than $350 billion in construction investment.
When looking at this line of logic, there is one thing to pay attention to:
The revaluation of nuclear power is, in essence, pricing the fact that “the pace of physical-world construction cannot keep up with the expansion of computing power.” Whoever already has connected clean baseload power that can start selling electricity immediately is the one who gets a cash-flow ticket in the AI era.
This is also why the secondary-market AI main theme can extend all the way from GPUs to a nuclear plant built half a century ago.