SIGNAL
The AI buildout has its own accounting problem, and it lives in the gap between a signed power contract and a plugged-in rack. Nebius told investors it has 2.5 gigawatts of contracted power by year-end, but only 800 megawatts to a gigawatt of that will actually be connected and usable for compute. The split is the story, not the headline figure.
A power contract is a reservation with a utility or developer. Connected power is the subset that has cleared substations, transformers, permitting, and GPU delivery. In the reporter's interpretation, most of the gap reflects execution risk: long-lead equipment, grid interconnection queues, and the chips themselves have to arrive on the same schedule as the electrons.
The mechanism is portable. Any AI infrastructure promise can be read on a contracted-versus-connected axis. The roughly 1.5-gigawatt difference Nebius discloses is, in the reporter's rough analogy, on the order of a large nuclear reactor's output, sitting in the queue between contract and switch. Hyperscalers race on connected capacity. Mid-tier clouds like Nebius are betting that capital markets will fund the build before the grid and the supply chain catch up. Plugged in, Nebius joins the hyperscaler-adjacent tier. Stalled, in the reporter's view, the company is a power-reservation ledger with debt.