The GPU Economy Is Becoming a Power Economy

For two years, the defining scarcity of the AI buildout was silicon. Allocation of H100s decided which labs shipped models and which ones wrote apology threads. That era is ending — not because chips became abundant, but because a harder constraint moved to the front of the line.

Ask anyone actually siting a data center in 2026 what gates the project. It is not GPUs. It is interconnection: the years-long queue to attach new load to a transmission system that was never designed for single customers asking for a gigawatt.

The unit of account is changing

You can hear the shift in how the industry talks. Nobody brags about chip counts anymore; they announce campuses denominated in gigawatts. OpenAI's Stargate program was unveiled not as "N million GPUs" but as a capital program — up to $500 billion — whose real currency is contracted power. Meta describes Prometheus and Hyperion in gigawatts. xAI's Colossus expansion in Memphis is discussed in terms of turbines, not tensor cores.

This is not marketing drift. It reflects the actual bill of materials. A modern NVL72-class rack draws on the order of 120–140 kW — a single rack now consumes what a small office building did. Multiply across a frontier training campus and the electrical plant, substations, and transmission upgrades rival the silicon as a share of project cost, and exceed it as a share of project risk.

Compute is fungible. Interconnection is not. You can resell a GPU; you cannot resell a place in the queue.

Power as balance-sheet strategy

The hyperscalers understood this before the market did. Microsoft's 20-year agreement with Constellation to restart Three Mile Island Unit 1 — 835 MW of firm, carbon-free baseload — was the tell: a software company underwriting the reopening of a nuclear plant because the grid could not otherwise promise it the electrons. Google's deal for small modular reactors with Kairos and Amazon's investment in X-energy follow the same logic. These are not sustainability gestures. They are supply-chain vertical integration, aimed one layer below the chip.

The financial system is repricing accordingly. Power purchase agreements are becoming the collateral of the AI economy — the thing lenders can underwrite when the depreciation curve of the GPU itself is a matter of theological debate. A signed PPA with a creditworthy counterparty increasingly does the work that a chip allocation letter did in 2023.

What this reorders

Three consequences follow from power becoming the binding constraint:

1. Geography flips. The map of AI infrastructure stops looking like the map of fiber and starts looking like the map of stranded generation: West Texas wind, Gulf Coast gas, decommissioned nuclear, hydro in the Nordics and Quebec.
2. Time horizons stretch. Chips are a quarterly procurement problem. Generation is a decade-long one. Companies whose planning cadence was software release cycles now negotiate 20-year offtakes.
3. The political surface expands. Electricity is regulated, rate-based, and local. Every gigawatt an AI campus contracts is a gigawatt someone's utility commission has to explain to ratepayers. The backlash cycle has already begun.

The GPU economy is not going away. But it is being annexed by an older, slower, more physical economy — one where the scarce asset is measured in megawatts and the queue is enforced by transmission engineers, not sales reps. The companies that win the next phase will be the ones that treated electrons as a first-class strategic input while their competitors were still counting chips.