The Economics: Costs, PPAs, and Why AI Changed Everything

Nuclear economics is a story of two numbers pulling in opposite directions. Capital cost: enormous — a new gigawatt-class plant is a $10–35 billion, decade-long civil-engineering campaign, and Vogtle's overruns are the cautionary tale that froze Western orders for a generation. Operating cost: tiny — fuel is cents per million BTU equivalent, staff and maintenance are stable, and the marginal cost of the next megawatt-hour from a running plant is among the lowest in the energy system. The consequence: an existing nuclear plant is one of the best assets in electricity, while a new one is one of the hardest things in capitalism to finance. Nearly everything interesting in nuclear economics — lifetime extensions, restarts, the SMR bet, government cost-sharing — is an attempt to keep the first fact while escaping the second.

The standard yardstick, levelized cost of electricity (LCOE) — total lifetime cost divided by total lifetime output — makes new nuclear look expensive against wind, solar, and gas, and for years that ended the conversation. But LCOE prices every megawatt-hour the same regardless of when it arrives, and the grid does not work that way. A megawatt-hour at 3 a.m. in a January cold snap is not the same product as one at noon in April. Nuclear's product — firm, clean, around-the-clock power at a 90%+ capacity factor — is precisely the product LCOE undervalues and, it turns out, precisely the product the 2020s suddenly demanded.

The demand shock has a name: data centers. After two flat decades, US electricity demand is growing again, with AI training and inference the marginal driver — hyperscale campuses of hundreds of megawatts to gigawatts that must run every hour of every day, owned by companies with trillion-dollar valuations and public 24/7 carbon-free-energy pledges. Wind and solar can't serve that load alone at any reasonable storage cost; gas serves it but breaks the pledges. Nuclear is the only proven option that satisfies both constraints — and uniquely, the customer can simply buy it for decades up front, converting nuclear's financing problem into a creditworthy revenue stream.

The instrument is the power purchase agreement (PPA) — a long-term contract for energy at agreed prices — and the deals layer of this map is a gallery of its new variants. Restart economics: Microsoft's 20-year PPA resurrecting TMI Unit 1 as the Crane Clean Energy Center; the Palisades restart, whose 200+-day-and-counting return arc is visible in this site's own history data. Uprates and existing fleets: Meta's 20-year deal at Clinton — a plant that nearly closed for lack of revenue a decade ago, now anchored by a tech offtaker. Co-location / behind-the-meter: Amazon's campus literally adjacent to the Susquehanna plant (the violet marker wired to the red one on the demand layer), buying power before it ever touches the public grid — a structure so novel it triggered federal proceedings over who pays for the grid everyone else still uses. New-build offtake: Google–Kairos, Amazon–X-energy, Meta–Oklo — contracts signed for reactors that don't exist yet, the purest expression of the thesis that demand certainty, not technology, was the binding constraint.

Stack on top the policy floor — production tax credits for existing plants (the Inflation Reduction Act's insurance against more premature closures), DOE cost-share for demonstrations, the ADVANCE Act's licensing reforms — and the economic picture as of mid-2026 is something nuclear hasn't enjoyed in fifty years: both ends of the ledger moving in its favor at once. The honest caveats: none of the announced SMR fleets has yet delivered a commercial megawatt-hour; tech demand forecasts could deflate; and the industry's history of cost overruns has earned every skeptic. The next few years of this map — Natrium's construction, the restarts' return to the live layer, whether the deal flow keeps converting into concrete — are, rather precisely, the experiment running.