AI Goes Nuclear
AI’s exponential energy appetite is quietly rebooting America’s nuclear industry.
Before AI can scale minds, it needs to scale matter.
In 2024, a quiet but critical realization took hold across the tech sector: artificial intelligence is not just a software revolution - it’s a thermodynamic one. The infrastructure demands of modern AI are beginning to reshape not only the tech stack, but the entire energy stack underneath it.
Training a GPT-4-class model consumes around 500 megawatt-hours of electricity - enough to power 15 homes for a year. But inference is the real power hog: AI models deployed globally must respond instantly to millions of queries, with racks of GPUs running 24/7. A traditional server rack might draw 10 kilowatts. A dense rack of H100s for AI inference draws over 100 kW. That’s not a marginal increase - it’s an exponential one. And it’s changing what data centers look like, where they get built, and what powers them.
Grid strain isn’t hypothetical. It’s already playing out in Northern Virginia - home to the largest concentration of data centers on Earth with 275+. More than 70% of global internet traffic passes through this corridor. Loudoun County’s data centers now consume more electricity than all of its homes combined.Dominion Energy has begun turning away new grid connection requests. Emergency diesel generators, once backup, are now front-line solutions. It’s a preview of the future: the cloud doesn’t scale unless power does.
In anticipation of this strain, the five largest tech companies spent a staggering $59 billion on capex in Q1 2025 - a 63% jump YoY. Increasingly, that capex is going into power acquisition and generation, not just compute.
For years, tech companies leaned on wind and solar to hit ESG targets. But AI inference isn’t interested in sustainability reports. It’s interested in uptime. The sun sets. The wind stalls. Batteries are expensive, and at this scale, insufficient. Clean power isn’t the same as reliable power. And for 24/7 inference, only one option checks every box: nuclear.
So what do trillion-dollar firms do when they realize their business model runs on electrons? They start buying the grid.
Microsoft: Rebooting Three Mile Island
In September 2024, Microsoft signed a 20-year power purchase agreement with Constellation Energy to revive Unit 1 at Three Mile Island - the same plant that partially melted down in 1979 and brought U.S. nuclear expansion to a halt. Unit 1 was shut down in 2019, deemed uneconomical amid cheap natural gas. Now it’s being brought back online to supply 835 megawatts of baseload power to Microsoft’s AI data centers.
This marks the first time a decommissioned U.S. reactor is being restarted at scale. What was once a national liability is now a strategic asset - a symbol of nuclear’s rebranding from risk to necessity.
Microsoft is also investing in the future beyond fission. Its partnership with Helion Energy, a fusion startup, targets the first commercial fusion prototype by 2028. Ambitious? Sure. But when you have Microsoft money, you can place moonshots on the sun.
Amazon: Buying the Grid
In June 2025, Amazon signed a long-term deal with Talen Energy to secure up to 1,920 MW from the Susquehanna Steam Electric Station - enough to power a small country, or perhaps a few AWS regions.
Earlier last year, it dropped $650 million to acquire a nuclear-powered data center campus outright.
It also backed X-Energy, a leading small modular reactor (SMR) startup, and is reportedly exploring nuclear-powered AWS regions.
Google: Betting on SMRs
Google, true to form, is taking a systems-first approach to nuclear. Its initial move came in October 2024, when it signed a deal with Kairos Power, a startup developing molten-salt small modular reactors. The agreement covered up to 500 MW of projected capacity across six to seven SMRs, with deployment targeted for 2030. The pitch: safer, factory-producible reactors that can eventually be colocated near high-density compute clusters.
Then in May 2025, Google added a second nuclear partner: Elementl Power, a company focused on larger-scale advanced nuclear projects. This newer deal involves siting and funding three plants expected to generate 600 MW each, further diversifying Google's future power mix.
Meta: Late But Present
Meta joined the party in June 2025, signing a 20-year PPA with Constellation Energy to draw 30 MW from the Clinton nuclear plant in Illinois. The capacity is modest, but it signals a strategic shift - away from carbon offsets and toward operational baseload coverage.
Even Meta sees the writing on the grid.
You can’t run a trillion-token multimodal model in real time on a volatile grid. The economics of AI have turned electricity from a cost center into a bottleneck. And that has quietly made nuclear energy - long frozen in political stalemate - essential infrastructure for the AI era.
After decades of stagnation, the U.S. nuclear sector is experiencing its first real demand shock - and it's coming not from Washington, but from hyperscalers. Private companies are reviving old reactors, funding new ones, and placing billion-dollar bets on next-gen fusion.
In 2024, Congress moved to fast-track SMR licensing and opened up federal land for colocated energy and data infrastructure. TerraPower, backed by Bill Gates and Warren Buffett’s PacifiCorp, is building a sodium-cooled reactor in Wyoming. Fusion players like Helion and Commonwealth Fusion Systems have raised billions with backing from Microsoft and Google alumni.
What once required government urgency is now being driven by AI demand curves. What began as a GPU arms race is now an energy land grab. The next big AI breakthrough? It might not be a model. It might be a reactor.