Power Generation
Whoever can deliver electrons to a rack wins. The 2026 bottleneck of the AI build-out.
What this layer does
A single 1 GW data center campus draws roughly the same electricity as a city of 500,000 people. The IEA projects data centers to consume around 4–5% of US electricity by 2030, doubling current levels. That growth lands on a grid that has spent decades flat — coal retiring, renewables intermittent, nuclear under-built. The result: every form of dispatchable generation (gas turbines, existing nuclear, new SMRs) is suddenly back in fashion, and electricity contracts are being signed at prices unimaginable five years ago.
This layer offers some of the cleanest AI exposure in the public market because the buyer is contractually visible: a hyperscaler PPA signed today shows up in IPP earnings for 15+ years.
Sub-categories
Merchant generators selling into wholesale markets. The cleanest beta to AI power demand — price-takers in tightening markets.
Utilities with material AI/DC load growth in their service territory. Get to rate-base the transmission upgrades.
Solar, wind, BESS. Hyperscalers are still the largest corporate PPA buyer of renewables, especially for “24/7 carbon-free” commitments.
The bridge fuel of the AI era. Lead times have stretched to 4–7 years for new gas turbines.
Utilities and IPPs with operating nuclear fleets — the cleanest baseload power. CEG/Microsoft deal at Three Mile Island re-defined the sub-category.
Small modular reactors. Still pre-revenue at most names, but the long-cycle AI nuclear thesis.
Fuel for the nuclear thesis. Two-way market: SMRs require HALEU fuel, restart fleet requires conventional fuel.
Gas suppliers with proximity to DC alleys (Appalachia, Haynesville, Permian). Many running new DC-tied agreements.
Grid-scale storage to firm renewables and shave peaks for DC sites.
Fuel cells, microgrids, gas-engine peakers placed directly at the data center to bypass grid wait times.