AI's power issue could be nuclear's big break

Here's the part of the AI story that doesn't get enough attention. Everyone argues about chips, but the chips are useless without something far more basic: electricity, and an enormous, constant amount of it. AI data centers run flat out, 24 hours a day, and the grid was not built for what's coming. That single problem is doing something almost nobody predicted a few years ago. It's reviving nuclear power, and it's pulling a whole group of energy stocks into the AI trade.

Just how much power are we talking about?

A lot, and it's climbing fast. The International Energy Agency expects electricity demand from data centers to roughly double, from about 485 terawatt-hours in 2025 to around 950 by 2030, which would put them at roughly 3% of all electricity used on the planet. Some estimates are even more aggressive, pushing past 1,000 terawatt-hours as soon as this year. AI is the main reason.

The catch isn't just the size of the demand, it's the shape of it. A data center training or running AI models wants power that is always on, steady, and predictable. Solar and wind are cheap and growing fast, but they're intermittent. You can't tell a hyperscaler their AI cluster only runs when the wind blows. That's why "baseload" power, the kind that runs around the clock, suddenly matters again. And the cleanest source of always-on baseload is nuclear.

Why nuclear, and why now

Nuclear checks the exact boxes AI needs: it runs 24/7, it produces enormous output from a single site, and it's carbon-free, which matters to tech companies that have spent years making climate promises. After decades of being treated as a relic, that combination has made it the surprise winner of the AI buildout.

Here's the honest nuance most headlines skip, though. Through about 2030, nuclear is not actually the main source filling the gap. The IEA's numbers show natural gas and renewables doing most of the near-term heavy lifting, with gas the single largest new source. Nuclear's biggest contribution, especially from new small reactors, mostly shows up after 2030. So nuclear's "big break" is really two separate stories: restarting plants that already exist, which is happening right now, and building new reactors, which is a 2030s bet. Keep those two apart and the whole sector makes a lot more sense.

And don't count out solar, it's just playing a different position. Solar is the fastest power source to build, which is why hyperscalers are pairing it with batteries in behind-the-meter "energy parks" that cover daytime load while gas and nuclear handle the around-the-clock base. The US is expected to add a record 43 gigawatts of utility-scale solar in 2026, up about 60% from the year before, and First Solar (FSLR), the largest US panel maker, just posted record sales on AI-driven demand with a backlog approaching 48 gigawatts. The catch is the same one nuclear solves: the sun isn't always up, so solar only fully answers the AI power problem when it's bolted to enough battery storage from the likes of Tesla and Fluence. For a deeper look at the solar side, that's its own story, but it belongs in the same conversation.

The deals that flipped the script

What turned this from a thesis into a real trade was a wave of deals where Big Tech went straight to nuclear operators and signed long-term contracts. As of mid-2026, every major hyperscaler has at least one:

• Microsoft and Constellation are restarting Three Mile Island. The dormant Unit 1, rebranded the Crane Clean Energy Center, is coming back under a 20-year power purchase agreement, with all 835 megawatts going to Microsoft. Constellation expects to spend roughly $1.6 billion to revive it, with the plant due back online around 2028. It's the first US nuclear plant being brought back specifically for a single customer. (Michigan's Palisades plant is actually the first retired US reactor to return to service at all, but it feeds the grid rather than one buyer.)
• Amazon and Talen Energy signed a 17-year deal in 2025 for 1.92 gigawatts from the Susquehanna plant in Pennsylvania, running through 2042, after Amazon bought an adjacent data center campus for $650 million. Talen expects the contract to bring in roughly $18 billion over its life.
• Google committed to 500 megawatts from small modular reactor startup Kairos Power.
• Amazon also put $700 million into X-energy for up to 12 small reactors.
• Meta went biggest of all, lining up as much as 6.6 gigawatts across multiple partners to feed its Prometheus AI supercluster.

When the richest companies in the world start signing 20-year power contracts, the market notices.

How to think about the stocks

This is not one trade, it's three, and they carry very different risk. Lumping them together is how people get hurt.

1. The companies that own reactors today. These are the real, cash-generating businesses, and they're the ones signing the hyperscaler deals. Constellation Energy (CEG) runs the largest US nuclear fleet, has been landing contracts steadily, and in early 2026 closed its Calpine acquisition, bolting roughly 21,000 megawatts of gas and geothermal onto that nuclear base. As of late June 2026 it traded in the low $300s, with analyst targets up in the high $300s. Vistra (VST), which owns the second-largest competitive nuclear fleet after its Energy Harbor deal, traded around $148 and has been one of the standout power stocks of the AI era. Talen Energy (TLN) is the Amazon partner. These are the lower-risk way to play the theme, because the plants already exist and the revenue is real.

2. The small-reactor moonshots. Oklo (OKLO) and NuScale (SMR) are bets on technology that mostly isn't generating commercial power yet. The upside is huge, the execution risk is just as big, and the volatility is brutal. NuScale has swung between roughly $9 and $57 over the past year and, as of late June 2026, trades around $11, down sharply on the year. Oklo has also pulled back hard in 2026 after a massive prior run, even with Nvidia's Jensen Huang publicly backing nuclear for AI. NuScale is the only US company with an NRC-certified SMR design, and it now has two approved. Just remember what you're buying: a 2030s story priced today, which is exactly why these names whip around so violently.

3. The fuel. More reactors means more uranium, and the supply math is tight. Uranium has been holding around $86 a pound, and Cameco (CCJ), the big Western miner, has delivery commitments running well above its planned production through 2030. That gap is the bull case for the fuel side in one sentence. Nuclear and uranium funds have been among the hottest trades of the AI-power theme over the past couple of years, though the most speculative names have cooled off in 2026.

The risks worth respecting

This is a genuinely exciting theme, which is exactly why it's worth staying clear-eyed. A few things to keep in mind:

• Timelines are long. New reactors take years to permit and build. The SMR names could deliver and still not produce meaningful power until the 2030s. The stock can get ahead of the actual electrons.
• A lot is already priced in. After moves like Oklo's, a chunk of the good news is in these stocks. It's a "bubble within a bubble" worry that rhymes with the broader AI valuation debate.
• The near-term reality is gas, not nuclear. If you're betting on nuclear to solve AI's power crunch in 2026, the data says you're early. The immediate fix is natural gas and renewables; nuclear is the longer game plus the plant restarts.
• Regulation and politics can speed this up or slow it down in a hurry.

Bottom line

AI's power problem is real, it's huge, and it isn't going away, which makes nuclear one of the most interesting long-term beneficiaries of the entire AI boom. The smart way to look at it is to separate the buckets: the utilities that own reactors and are banking hyperscaler money today, the small-reactor startups that are thrilling but speculative and years out, and the uranium underneath all of it. The chips get the headlines. Increasingly, the power behind them is the story worth watching.

Sources

• International Energy Agency, Data centres and data transmission networks

This is analysis and opinion for information only, not investment advice, and I'm not a financial advisor. Figures come from market reporting, the IEA, and analyst estimates as of mid-2026 and vary by source, so do your own research and confirm current numbers before making any decisions.