Notes
The Nuclear Renaissance
On this page
For decades nuclear power was the energy source the developed world had quietly decided to stop building. Then artificial intelligence changed the arithmetic. The data centers training and serving frontier models need electricity that is not merely abundant but firm — available every hour, at enormous scale, with carbon commitments attached — and that demand has done what climate policy alone never quite managed: it has made new reactors bankable again.
Why the demand came back
Global data-center electricity use reached roughly 415 TWh in 2024, about 1.5 percent of world consumption, and is projected to roughly double to around 945 TWh by 2030 as AI workloads scale. Renewables are cheap but intermittent; diesel backup is dirty and increasingly unacceptable. So the hyperscalers — Microsoft, Google, Amazon, Meta — began signing long-term power purchase agreements and exploring on-site generation directly.
The signature event was the Three Mile Island restart: in 2024 Constellation Energy announced it would reopen the undamaged Unit 1 reactor, dormant since 2019, under a twenty-year deal with Microsoft. A plant whose name was a byword for nuclear failure became a symbol of nuclear revival. Other operators have moved to relicense and uprate existing fleets on the same logic.
Small reactors and the factory model
The more durable shift is architectural. Conventional gigawatt plants are bespoke, slow, and prone to cost overruns. The renaissance is betting instead on small modular reactors (SMRs) and micro-reactors — units built on an assembly line and shipped to site, trading peak output for the cost discipline of mass production. The wager is that the learning curve that drove down solar and batteries can be applied to fission if reactors become products rather than megaprojects.
Beyond the near term sit older ideas getting a fresh look: thorium fuel cycles and molten-salt designs that run hot at low pressure, promising better safety margins and less long-lived waste. Speculative timelines of the AI era lean hard on this combination, imagining a 2030s buildout powered by factory-produced micro-nuclear and thorium systems.
Fusion as the long horizon
Further out is fusion — the reaction that powers the sun, pursued for half a century and still not delivering net energy to a grid. Private capital has nonetheless poured in, and the honest framing is one of patience: fusion is the plausible endgame of energy abundance, not a 2030 solution. The constraint that pushes some futurists toward orbital compute and lunar helium-3 is precisely that terrestrial fusion arrives later than terrestrial demand.
Complementary, not rivals
The instinct to ask whether nuclear beats batteries or solar misreads the system. They do different jobs. Renewables supply the cheapest marginal electrons; batteries firm the short timescales — sub-second frequency response through multi-hour shifting; nuclear supplies the always-on baseload that neither does economically alone. A grid carrying both AI loads and electrified transport needs all three layers at once. The renaissance is real, but it is best understood as the slow-moving firm-power tier of a portfolio, not a victor in a contest.
Related concepts
Battery Technology & Grid Storage
Appears in
Details
- Section:
- Notes
- Updated:
- 2026-06-26