Choosing a data centre site used to be arithmetic. You compared electricity prices, land costs and distance to customers. Power was a line in the operating budget.

It is now the question of whether the project happens at all.

What does "time to power" mean?

The interval between committing to a site and being able to energise equipment in it. It has replaced price per kilowatt hour as the first question customers ask and the last one that gets negotiated.

Why did power become the binding constraint?

Because demand outran the system that delivers it. Hyperscalers have committed more than 700 billion dollars to AI infrastructure, while roughly half of the United States projects planned for 2026 were delayed or cancelled. Those projects are not short of capital, land or customers. They are short of power that is already connected and running.

Four things hold the shortage in place.

Queues restack. Each connection study assumes a particular queue ahead of it, so when a project withdraws, the studies behind it have to be redone. Queues lengthen faster than they clear. In Virginia, Texas and Ohio they now run three to five years.

Transmission outlasts generation. A power station takes two to three years. The high-voltage lines to move its output take seven to ten, because they cross jurisdictions and attract objection at every crossing.

Transformers are rationed. A handful of manufacturers hold order books years deep, and lead times have stretched from two years to as much as four. This is a manufacturing constraint, so it does not respond to paying more.

The forecasts were overtaken. Utility planning written before 2023 did not contemplate single sites requesting hundreds of megawatts, which is why the other three constraints tightened at once.

Why does this hurt AI more than earlier workloads?

Because of how fast the equipment loses value. A GPU fleet is worth a fraction of its purchase price after four to five years. If the building arrives three years late, most of the asset's economic life was spent waiting: the capital was committed, the hardware existed, and the return did not happen.

In conventional development a two-year delay is a scheduling problem. In AI infrastructure it can consume the investment case.

What actually shortens time to power?

Not building new transmission. Connecting to transmission that is already there. The seven-to-ten-year wait for new lines is the hardest constraint in the list, and a site whose connection requires them inherits that timeline no matter how the commercial terms are structured.

At Arrow Partners. The campus connects as a general consumer on the national 230 kV transmission system through two substations that already exist and already adjoin the site, with energy under bilateral supply contracts. Panama's generation capacity sits well above current demand, so the system has room for the load.

How is a grid connection made resilient?

An AI campus is an unusually demanding load, because training clusters move tens of megawatts within seconds. A connection built for it has to be redundant at the utility level and has to contain that behaviour on site.

At Arrow Partners. Two independent points of connection at two separate substations, on four consumer circuits, two from each. Any two carry the entire load, so losing a whole substation leaves the campus at full capacity. Behind the connection sits 300 MW / 300 MWh of energy storage across every medium-voltage section, holding the campus's own load swing inside the boundary rather than presenting it to the network, with diesel generation covering the core loads at N+1.

For most of the industry's history, power was a price. It has become a schedule.