The Real Reason Nuclear Is Back: AI’s Electricity Bill

The hottest trade in deep tech isn’t a chatbot — it’s a power plant. Why every major AI company is suddenly buying nuclear, and what the rush reveals about where the real bottleneck now sits.

The most valuable thing in artificial intelligence right now isn’t a model or a chip. It’s a megawatt. The companies racing to build superhuman AI have hit a wall nobody drew on the pitch deck: there isn’t enough electricity to run it.
So they’ve gone shopping for the most reliable, round-the-clock, carbon-free power on the planet — nuclear. In the past year every major AI company has signed a nuclear deal, and a technology the world had written off as a relic has become the hottest corner of deep tech.

The demand curve broke the grid

The numbers are hard to overstate. Goldman Sachs projects data-center electricity demand could climb around 160% by 2030, and global data-center consumption is on track to nearly triple — from roughly 460 terawatt-hours in 2024 toward 1,300 by 2035. In 2025 alone, an estimated $580 billion went into AI data-center infrastructure. The grid was simply never built for a load like this.

AI compute needs power that never blinks: 24/7, carbon-free, and ideally right next to the servers. Solar and wind are cheap but intermittent, and batteries only stretch so far. That single requirement — always-on clean baseload — is what dragged nuclear back from the dead.

Big Tech is buying reactors, not renting them

This isn’t a thought experiment. Microsoft signed a roughly 20-year, ~$16 billion deal to restart the Three Mile Island plant (about 835 MW) to feed its AI load. Amazon committed over $20 billion around Talen’s Susquehanna plant and led a $500 million round into the reactor developer X-energy. Google contracted 500 MW of small modular reactors from Kairos; Meta signed Oklo for a 1.2 GW campus in Ohio.

The pattern is unmistakable. The companies selling intelligence are vertically integrating into power generation. When your product is capped by electrons, you stop being a software company and quietly become an energy company that happens to write code.

Believe the demand, not the timeline

Now the sober part the hype skips. The supposed star of the story — the small modular reactor — barely exists yet. Most SMR designs won’t deliver commercial power until the 2030s; leading names carry heavy losses, thin revenue, and share prices well off their highs. First real power is years out and the economics are unproven at scale. (China’s Linglong One, the first commercial onshore SMR, entering service in 2026, is the notable exception.)

So read it the way a deep-tech investor should: the demand signal is real and durable, but the supply is a decade-long build carrying binary technical and regulatory risk. The opportunity isn’t the reactor stock of the week — it’s everywhere the power actually flows, from transmission and cooling to storage and the picks-and-shovels around them