Why AI Data Centers Are Turning to Nuclear Power

The largest technology companies in the world are making billion-dollar bets on nuclear power. Microsoft has partnered with Constellation Energy to restart Three Mile Island. Amazon acquired a data center physically adjacent to a Pennsylvania nuclear plant and signed a deal to purchase its entire output. Meta has contracted with Oklo, TerraPower, and Vistra Nuclear for future nuclear capacity. Google has signed a power purchase agreement with Kairos Power for their molten-salt reactor design.

This is a striking turn. A decade ago, tech companies competed to tout how much wind and solar they were buying. Today, they’re licensing reactors, bidding on decommissioned plants, and funding next-generation nuclear startups. What changed?

The power math stopped working

AI training clusters consume electricity at a scale that breaks conventional supply planning. A cluster of 100,000 NVIDIA H100 GPUs — a mid-sized training run by 2025 standards — draws approximately 70–100 megawatts around the clock. Meta’s Hyperion campus in Louisiana is designed for 5 gigawatts. Microsoft’s Project Fairwater in Wisconsin targets 3.3 GW. Amazon’s Project Rainier in Indiana is 2.4 GW.

These are not data centers in any conventional sense. They are power plants with servers inside. And they need power that is reliable, dispatchable, and available 24 hours a day.

Wind and solar are cheap and getting cheaper. But they are intermittent. A training job running on 100,000 GPUs cannot pause every time clouds cover the solar array. Backup power systems (batteries, gas peakers) add cost and carbon. Large-scale battery storage can smooth intraday variation, but multi-day weather events require something that generates power continuously regardless of conditions.

Nuclear produces electricity 24/7 at capacity factors above 90% — meaning it runs at full output more than 90% of the time, compared to roughly 25–35% for solar and 35–45% for wind. A single large reactor producing 1 GW provides more reliable AI training hours per year than three times as much solar capacity.

The deals that have actually been signed

Microsoft and Constellation — Three Mile Island: The most prominent deal in this space. Microsoft agreed in September 2024 to purchase 100% of the output from Unit 1 of Three Mile Island, which Constellation rebranded as Crane Clean Energy Center. The reactor produces approximately 835 MW and had been shut down since 2019 for economic reasons. With Microsoft as an anchor power purchaser, Constellation restarted the reactor in September 2025. The deal runs for 20 years.

Amazon and Talen Energy — Susquehanna: Amazon Web Services acquired the Cumulus Data data center campus in Pennsylvania, which sits on the same site as Talen Energy’s Susquehanna nuclear plant. Amazon struck a deal to receive 960 MW of nuclear power directly — avoiding the grid entirely — for its nuclear-adjacent data center campus. The Federal Energy Regulatory Commission (FERC) initially blocked direct connections to nuclear plants as potentially destabilizing grid pricing, but Amazon and Talen reached a revised agreement.

Meta and Oklo: Meta has signed a letter of intent with Oklo for 750 MW of capacity from Oklo’s Aurora microreactor design over a 20-year period. Oklo’s 15 MW modular reactors use metallic fuel and can run for multiple years without refueling, targeting deployment near high-power-demand sites. Meta’s deal would represent Oklo’s largest order by far — the startup received NRC design certification approval in 2025.

Meta and TerraPower — Wyoming: Meta has partnered with TerraPower’s Natrium sodium-cooled fast reactor project in Kemmerer, Wyoming, targeting 345 MW of capacity. Natrium uses a molten-salt thermal storage system that can buffer output, providing flexible generation that can ramp up during peak AI training demand.

Meta and Vistra: Meta signed an agreement with Vistra for the uprate of existing nuclear units in Ohio and Pennsylvania — adding generation capacity at operating plants through efficiency improvements rather than new construction.

Google and Kairos Power: Google signed a power purchase agreement with Kairos Power for seven small modular reactors totaling 500 MW by 2035, using their molten fluoride salt reactor design. This was the first commercial nuclear PPA signed by Google, and the first for Kairos.

Why now — and why not sooner?

Nuclear fell out of favor in the 2010s as natural gas prices collapsed and the Fukushima disaster triggered shutdowns across Europe and Japan. New plant construction became economically uncompetitive against combined-cycle gas and rapidly falling renewable costs. Several US plants closed before their licenses expired.

What changed is the coincidence of three factors:

AI demand created a buyer willing to pay for reliability. Nuclear power is expensive per megawatt-hour, but AI companies need reliability more than cheapness. A training job interrupted by grid instability or fuel constraints costs far more than the power savings from cheaper-but-unreliable alternatives.

The regulatory path for SMRs opened. The NRC approved its first small modular reactor design in 2023 (NuScale, since cancelled on cost grounds), and issued design approval to Kairos and Oklo in 2024-2025. The pathway from design to approved nuclear facility has shortened considerably, though execution timelines remain long.

Existing nuclear became cheap relative to replacement cost. Recommissioning a mothballed reactor — or signing a 20-year PPA with a plant that needs an anchor customer — costs far less than building new nuclear capacity from scratch. The Three Mile Island and Susquehanna deals are essentially existing nuclear capacity finding new economic purpose.

What’s actually built vs what’s promised

The gap between nuclear announcements and operating nuclear capacity is significant. The deals worth monitoring:

• Crane Clean Energy Center (Three Mile Island) — 835 MW, operational as of late 2025. The only major nuclear-AI data center deal where the power is actually flowing.
• Amazon-Susquehanna — 960 MW, operational supply contract.
• NextEra / Duane Arnold Iowa — reactivation in progress, no confirmed AI customer.
• Vistra Comanche Peak uprate — in progress.
• TerraPower Natrium Wyoming — construction ongoing, target mid-2030s.
• Oklo Aurora — production scaling, commercial deployments 2027 onward.
• Kairos Power — 500 MW target by 2035, pre-construction.
• Rolls-Royce SMR — targeting UK data center deployments, 2030s.

The practical reality: existing nuclear (restarts and PPAs with operating plants) provides near-term capacity. New nuclear designs (SMRs, advanced reactors) are genuinely interesting but will mostly deliver capacity in the 2030s, not the 2025-2027 window when AI compute demand is compounding fastest.

The grid impact question

Nuclear-AI deals raise a genuine policy question: when a hyperscaler contracts for the full output of a large nuclear plant, what does that mean for everyone else on the grid?

FERC’s initial blockage of the Amazon-Susquehanna direct connection reflected this concern: removing large baseload generation from the grid raises prices for residential and commercial customers who depend on it. The nuclear capacity that had been holding grid prices down would instead flow exclusively to Amazon servers.

The industry’s response — supported in FERC’s revised guidance — is that direct connections are acceptable if the data center can demonstrate grid reliability benefits, such as load flexibility or investment in transmission capacity. But this tension between large AI operators and grid users will intensify as demand grows.

This index tracks 344 AI data centers, including all nuclear-powered and nuclear-contracted facilities. Browse the nuclear energy view for the full list, or download the dataset for research.