Why Power — Not Chips — Is the Real AI Constraint

For two years, the AI conversation was about chips: who could get them, who could make them, who was allocated what. That era is ending. Chips are increasingly available to those willing to pay. What is not available — at any price, on any timeline a CFO would recognize — is the power to run them.

The constraint has moved down the stack, from silicon to substations.

The arithmetic of the wall

Data-center capital spending is now on the order of $750 billion a year globally — a figure that has crossed above what the world invests in oil and gas production. That capital is buying buildings and hardware at extraordinary speed. What it cannot buy at the same speed is firm electricity.

A large AI campus can require as much power as a mid-sized city. Grid interconnection queues in the major markets now run to years, not quarters. Transformers and high-voltage equipment carry multi-year lead times of their own. Transmission — the least glamorous asset class in the entire economy — has become the gating item on the most glamorous one.

The result is a strange inversion. Companies that spent decades treating electricity as a utility line-item are discovering it is now their strategic constraint. The binding question for AI growth in 2026 is not "how many accelerators can you get?" It is "where is your power coming from, when does it arrive, and how firm is it?"

Why this is an energy-market problem, not a tech problem

The instinct of technology companies is to solve constraints with engineering and purchasing power. Both help less here than they expect, because power is not bought the way hardware is bought.

Securing firm power at scale means power purchase agreements negotiated with generators who have their own queues of suitors. It means siting decisions that trade off land, water, fiber, climate, and local politics. It means fuel questions — gas supply for new generation, and increasingly, arrangements that would have been familiar to an LNG desk long before they became interesting to a cloud provider. It means navigating regulatory regimes that differ not just by country but by utility territory.

These are not procurement exercises. They are negotiated, relationship-driven, multi-year commitments between counterparties who need to trust each other's ability to perform. The skills involved — structuring, counterparty diligence, cross-border execution — come from energy markets, not from software.

What the next phase looks like

Three shifts are already visible.

Generation is following compute. The fastest-moving developers are no longer waiting for the grid; they are contracting for, and in some cases building, dedicated generation — gas, renewables paired with storage, and early nuclear commitments. "Behind-the-meter" has gone from an industry term to a board-level strategy.

Geography is back. For twenty years, the internet flattened location. Power has un-flattened it. Regions with surplus generation, fast interconnection, and cooperative regulators are pulling in capital that would never have looked at them before. The map of AI is being drawn by the map of electricity.

Energy players hold new cards. Utilities, independent power producers, midstream operators, and fuel suppliers now sit on the critical path of the technology industry. Many of them are only beginning to realize what their position is worth — and what kind of counterparties they are now dealing with.

The convergence point

None of this is a temporary bottleneck that better chips will dissolve. Every improvement in compute efficiency has so far been consumed by greater ambition. The demand curve for intelligence, it turns out, is a demand curve for energy.

That is the thesis behind our work at the intersection of these markets: the decisive deals of this cycle are being made where power, capital, and compute meet — by people fluent in all three languages. The wall is real. So is the opportunity on both sides of it.