Somewhere in the last two years, "AI company signs nuclear power deal" went from a strange-sounding headline to a routine one. This isn't a publicity stunt — it's a direct, practical response to a genuine bottleneck that's now constraining AI infrastructure growth more than chip supply is.
Much of the public conversation around AI infrastructure focuses on GPU shortages, but by 2026 the more binding constraint for many hyperscalers has become electricity. Microsoft has disclosed an $80 billion backlog of Azure orders it cannot fulfill specifically due to power availability — not chip supply, not construction timelines, power. Data center electricity demand grew faster than the International Energy Agency's overall 17% data center growth estimate for 2025, and rack power density has climbed 3-5x since 2022 as GPU-dense server racks replaced conventional compute.
AI data centers need power that is both carbon-free (to meet corporate sustainability commitments) and available continuously at massive, predictable scale — a specification that solar and wind, with their intermittency, don't meet on their own without substantial storage buildout. Nuclear is essentially the only current technology that meets both requirements simultaneously at gigawatt scale. Microsoft's agreement to restart a reactor at the former Three Mile Island site (rebranded Crane Clean Energy Center) is the most publicly visible example, but it's part of a broader pattern of hyperscalers signing long-term (often 20-year) power purchase agreements directly with nuclear operators.
Companies don't sign 20-year power contracts for infrastructure they expect to be temporary. These deals are a strong signal that hyperscalers are planning AI infrastructure capacity on a multi-decade timeline, not a hype-cycle one — regardless of how individual AI product cycles or model releases play out in the shorter term. It also reflects a broader shift: AI infrastructure planning has started to resemble utility-scale industrial planning more than traditional tech company capital allocation, which is part of why hyperscaler capital intensity ratios (45-57% of revenue in some cases) now look more like an energy company's balance sheet than a software company's.
For anyone working in or around cloud infrastructure, this is useful context for understanding *why* certain regions have better cloud availability and pricing than others, why some hyperscaler capacity announcements come tied to specific power agreements, and why "just spin up more compute" has genuine physical-world constraints behind it that occasionally show up as availability limits or pricing changes in ways that aren't always obvious from the API layer alone.