AI's appetite for electricity is now so large that Microsoft, Google, Amazon, and Meta are personally reviving nuclear power — restarting shuttered reactors and pre-buying next-generation small modular reactors. The catch: nuclear operates on a decade-long clock, and the grid bottleneck is already here.

AI-generated (NotebookLM) audio overview. Source: HexLocal in-house research — Research - AI's Energy Reckoning and the Nuclear Comeback (Dr. Priya Nair). Primary external sources include the IEA "Energy and AI" report (April 2025), Goldman Sachs, the Carnegie analysis of announced nuclear capacity, and a UN/ITU report restating and extending the IEA demand projections.

- Data-center electricity demand is projected to roughly double by 2030 to ~945 TWh — about Japan's entire annual consumption — driven primarily by AI inference, not training

- Hyperscalers are making concrete nuclear bets: the Three Mile Island/Crane restart (Microsoft), Google-Kairos, Amazon-X-energy, Meta-TerraPower, and Oracle SMR commitments are all verified deals

- Palisades sets the precedent as the first US reactor restart; TMI/Crane is in progress with a 2027 target — both schedule-dependent

- Reality check from Carnegie: ~13 GW of announced nuclear covers less than 20% of projected 2035 demand — new nuclear is a hedge for the 2030s, not a near-term fix

- The true near-term bottleneck is the grid itself: a 2.6 TW interconnection backlog means time-to-power, not reactor announcements, is the binding constraint

- The calibrated environmental verdict: the power demand is a genuine issue (gas, not nuclear, fills the near-term gap), but "AI is boiling the planet" overstates AI's current global share while the efficiency-and-nuclear response is real but slower than the marketing suggests