Energy

Reliable baseload power has become a hard constraint on AI infrastructure economics. Hyperscaler capex through 2030 is multi-hundred-billion dollars, and that capex assumes 24/7 firm power that intermittent renewables alone cannot supply. The PPAs being signed between Constellation, Vistra, and the major cloud platforms are the public evidence of a private decision the industry has already made: nuclear baseload is back in the long-duration plan. This is a multi-decade re-pricing of nuclear utilisation, not a cycle.

Inside that envelope, the value splits three ways. Existing fleet operators have already re-rated on PPA news but are still pricing in only a partial extension of the duration of advantage. Small modular reactors (NuScale public, X-energy and Holtec private) remain pre-commercial and binary; we treat these as option value with concrete licensing milestones to underwrite. The fuel cycle (Cameco at the front end, conversion and enrichment further down) is structurally tight after fifteen years of underinvestment, with Russian enrichment capacity sidelined by sanctions and HALEU supply for advanced reactors a near-term bottleneck.

Edge sources we lean on: (1) PPA disclosures between hyperscalers and nuclear operators, often pre-announced in sustainability filings; (2) NRC licensing flow and grid-interconnect queue data for SMR candidates; (3) uranium spot and term prices versus marginal cost of new mine supply; (4) US DOE HALEU programme awards and conversion-capacity announcements. Grid build-out (transmission, interconnects) is the unsexy compounder under the entire AI-capex thesis, and the layer we still owe the deepest writeup on.