Power Is the New Region

The site of America's most famous nuclear accident is being switched back on for one customer. Under a 20-year power purchase agreement signed with Microsoft, Constellation is restarting Three Mile Island's Unit 1, rebranded the Crane Clean Energy Center, with 835 megawatts targeted for 2028 and every one of them earmarked for AI datacenters. A retired reactor, resurrected, for a software company. Whatever else 2024 is remembered for in this industry, that deal is the moment the constraint changed in public.

Here's the thesis: for fifteen years, cloud capacity planning assumed the scarce inputs were racks, then chips. Both eras are over. The binding constraint on cloud buildout is now electricity, the industry's site selection has started following energy instead of users, and that quietly breaks assumptions about regions, latency, and pricing that everyone's architecture diagrams still encode.

The shopping spree

Microsoft's reactor wasn't an eccentric one-off; it opened a genre. The deals since read like a utility's M&A desk got hold of big tech's checkbook.

As of May 2026, tracker sites count 13 announced projects committing roughly 9.8 GW of nuclear capacity to AI infrastructure.

Read the table as a confession. Companies whose competence is software are becoming counterparties to reactor restarts and first-of-a-kind SMR deployments, timelines measured in half-decades, because they've concluded the grid won't sell them what they need on any faster schedule. Nobody signs a 20-year PPA for a temporary problem.

The queue is the moat

The mechanics behind the confession are unglamorous. Getting a new datacenter connected to the grid means joining an interconnection queue, and those queues now run years in the good cases, which turns energy procurement into the longest-lead item in the entire capacity supply chain, longer than chips, longer than construction. Money can compress most shortages. It cannot much compress permitting, transmission builds, or turbine order books, which is why capital has started chasing anything that bypasses the queue: retired reactors, on-site generation, campuses bought specifically because they sit next to existing plants.

The scale of the collision is public arithmetic. The hyperscalers plan around $700 billion of combined capital expenditure in 2026, overwhelmingly for AI infrastructure, and infrastructure at that scale is measured in gigawatts. Microsoft has been unusually candid about where that collides with reality: an $80 billion Azure backlog attributed to power constraints, with purchased GPUs sitting idle because there's no electricity to install them under. Sit with that image. The most valuable chips on earth, in warehouses, waiting for a substation.

Which makes power contracts the new moat. Two years ago the competitive question between clouds was who had the best silicon roadmap; the harder question now is who locked in generation, and queue positions, interconnect agreements, and PPAs signed in 2024 are assets rivals cannot replicate at any price on the same timeline. It's the unbundling era's resource layer: the neoclouds proved compute could be bought outside the bundle, and now everyone discovers that what actually gates compute is a commodity older than computing.

What it quietly breaks

Regions used to be a demand-side concept: put capacity where users, data residency, and enterprise customers are, and price it roughly uniformly. Energy-first site selection inverts that. New capacity lands where megawatts are available, Brandenburg or Pennsylvania or wherever a reactor has spare output, which is not necessarily where anyone's users are, and the decade-old assumption that your provider will simply have capacity near your market when you need it stops being safe. Latency budgets meet geology.

Pricing assumptions crack next. Electricity costs now diverge sharply by location and contract vintage, and uniform-ish regional pricing papers over an input cost that stopped being uniform; my bet is the paper doesn't hold, and region-differentiated compute pricing, or scarcity surcharges wearing another name, arrive within a few years. Timeline assumptions were always the most fragile: demand is compounding now, while the table above delivers its megawatts in 2028 and 2030. The gap between those dates is the era we're in, and it's the era in which capacity allocations, waitlists, and quota negotiations became a normal part of buying cloud, a sentence that would have sounded absurd in 2020.

The steelman: constraints attract solutions

The case for calm is respectable. Efficiency is improving fast, inference is being squeezed onto cheaper silicon, and the industry has a long record of demand forecasts embarrassing themselves; some of those SMR deals are best understood as long-dated hedges and press releases, with first-of-a-kind reactors carrying first-of-a-kind risk, and the grid does eventually build out. If AI demand plateaus, today's power panic will look like the fiber glut of 2001, and contrarians buying distressed capacity will feast.

Concede all of it as possible, and note what the calm case requires: believing simultaneously that the companies spending $700 billion are wrong about demand, and that the constraint their own executives call binding will dissolve before it reshapes the market. Even on optimistic timelines, the operative decade runs on scarce power, and market structure formed during scarcity, the contracts, the queue positions, the siting, outlives the scarcity that formed it. That's the actual lesson of every infrastructure cycle, fiber included: the glut ended, the ownership map it created didn't.

There's a small demand-side moral we can't resist, since our whole product exists on the other end of this telescope: when the industry's binding constraint is electricity, workloads that scale to zero stop being a pricing gimmick and start being a grid courtesy. Idle compute burning watts is now everyone's problem. Yours too: do you know which region your next deployment lands in, and do you know what's powering it?

Related: Neoclouds and the Great Unbundling, the market layer above this resource layer. More about what we're building at light-cloud.com.