Venture capitalists are pouring money into nuclear energy startups, wagering that reactors can satisfy the enormous power appetite of the data centres running artificial intelligence.
It is a striking shift for an industry built on software that scales in months, and the risks are unusually stark.
The appeal is obvious: Data centres need vast, steady power that does not drain local supply or wreck the environment, and nuclear is seen as cleaner than gas turbines with the potential to turn cost-effective over time.
The catch is that almost none of it works commercially yet.
Experimental
Current startups are largely experimental and are not powering operations at scale, a reality that sits awkwardly with venture capital's usual appetite for speed.
This is hard tech, demanding heavy capital and offering no proven business model for years.
Oklo, backed by OpenAI's Sam Altman and now public via a shell-company merger, still generates no revenue, which captures the sector's stage of development.
The field splits into two camps.
Fission, the established technology behind existing plants, is increasingly pursued through small modular reactors and microreactors, compact units meant to be built on an assembly line and clustered together to power a data centre.
Fusion still way off
Fusion, which fuses atoms rather than splitting them, remains largely unproven commercially.
Helion Energy, a fusion startup near Seattle, plans to begin generating power in 2028 from a 50 megawatt plant in Chelan County, Washington, with Microsoft already lined up as a customer.
The investors are not only specialists.
Generalist firms including Valor Equity Partners and Thrive Capital have written large cheques, emboldened by capital-heavy, long-horizon successes such as SpaceX and Anduril that proved moonshots can pay.
Valor, an early SpaceX backer, has seen returns from that bet running into tens of billions of dollars.
The technology companies are moving in parallel.
Small modular option
Amazon has agreed a small modular reactor plant in Washington state with X-energy, Google has partnered with Kairos Power, and Meta has deals with both Bill Gates-backed TerraPower and Oklo.
Microsoft, beyond Helion, has struck an agreement with Constellation Energy to reopen a Three Mile Island reactor shuttered since the 1970s.
History counsels caution.
The current rush mirrors earlier enthusiasm for solar, wind and batteries, cycles that produced heavy losses before yielding a few survivors, and a similar shakeout looks likely within a few years.
Bloom Energy offers the cautionary and hopeful template, enduring wild swings before finding relevance in the AI data centre market it was never designed to serve.
The deeper pattern is investors chasing AI down the stack, from models to chips and now to the electricity itself.
The bet rests on a single conviction, that demand for artificial intelligence shows no sign of easing.
