Data-center power deals reshape where AI gets built

Long-term electricity contracts, not compute chips, are increasingly deciding where new AI infrastructure is located.

The location of new AI infrastructure is increasingly being decided by long-term electricity contracts rather than by the availability of compute chips. Power has moved from an operating cost to a siting constraint.

Why it matters

When the binding input is megawatts, the map of where models get trained and served looks different from the map of where semiconductors are made. Regions with cheap, stable baseload power gain a structural advantage regardless of their chip supply.

The numbers

A single large training campus can draw on the order of hundreds of megawatts continuously. Securing that at a predictable price for a decade or more is now a multi-year negotiation in many regions, not a utility call.

Context

Grid interconnection queues are long enough that the timing of a power deal can determine whether a project comes online this year or three years from now. That uncertainty is feeding into where operators commit capital.

Background

The constraint mirrors earlier infrastructure booms in which energy, not equipment, determined where industry clustered. The AI buildout is reproducing that dynamic at a scale the grid has not yet fully absorbed.