Why AI's Power Crisis Is Forcing Hyperscalers to Restart Nuclear Plants

Energy, not chips, is now the binding constraint on AI infrastructure—evidenced by ERCOT's frozen 474-gigawatt interconnection queue forcing hyperscalers into multi-decade nuclear power contracts.

Three weeks ago, the AI infrastructure bottleneck was still chips. Today it is power. Microsoft's 20-year Chevron deal, Google's €13 billion Finland commitment with a 22-year nuclear PPA, and NextEra's $1.9 billion Department of Energy loan to restart Duane Arnold for Google now define the competitive frontier. Amazon has secured 1.92 gigawatts from Pennsylvania through 2042. These are not sideline investments—they are the binding constraint. Hyperscalers have capital, they have land, they have compute designs locked in. What they lack is the grid capacity and power supply to run them.

The shift is unmistakable: when Microsoft's $115.95 billion FY2026 capex (35% of revenue) flows to energy co-equal with compute, and Google commits €13 billion to ensure Finland's nuclear supply chain accelerates, the infrastructure bottleneck has rotated 90 degrees. Qualcomm's $60 billion AWS custom-chip deal and OpenAI's dual-sourcing between TSMC and Samsung both enable inference scaling—but neither solves what happens when you plug 5 gigawatts into a grid that froze 474 GW in interconnection requests. ERCOT's queue lock is not transient congestion; it is a structural admission that US grid operators cannot absorb the power density AI now demands.

This shift exposes every neocloud still independent. Nebius needs $250 billion through 2030. CoreWeave burns capital at 201% of FY2025 revenue. Anthropic has $517 billion in compute contracts yet trails OpenAI's $750 billion runway. These are not scaling problems—they are extinction problems. Hyperscalers can sign century-long PPAs because they have the balance sheets for $100+ billion capex programs; neoclouds do not. The market is testing whether anyone outside the hyperscaler four can raise the capital to compete when the price of entry is now energy supply chains, not just data centers.

Nuclear operators and grid integrators are now bottleneck suppliers. Constellation Energy raised EPS outlook to $12.00 on 920 MW of new deals including a 15-year Walmart PPA. Every nuclear plant within 100 miles of a hyperscaler cluster is now in play. The export-control squeeze on G42—forcing Abu Dhabi's sovereign AI fund to consider US restructuring just to access Nvidia chips—signals that geopolitical and infrastructure constraints now intersect and compound. When China supplies 30% of AI data center transformers and 66% of optical modules, and US policy forbids non-US-controlled projects from buying them, the infrastructure war becomes a materials war.

What to watch: whether hyperscalers move capex to behind-the-meter generation when grid delays bite (Tesla's Memphis BESS at xAI's Colossus is the blueprint); whether neoclouds seek acquisition or sovereign backing; and whether custom silicon actually delivers competitive economics that justify supply-chain diversification away from Nvidia. The next twelve months will confirm whether AI infrastructure has become a public utility play—in which case the winners are hyperscalers, utilities, and national governments—or whether private capital can still compete. The power bottleneck suggests public utility.