The constraint on compute is now electrical

The compute buildout has hit a limit that is not semiconductors. In a growing number of markets, the binding constraint is getting power to the site.

The interconnection queue

Connecting a large new load to a transmission network requires a study process, and the queues in several major grids are measured in years. A datacentre with chips ordered and a site acquired can still be waiting on an interconnection agreement.

This is why siting decisions have moved toward places with existing spare capacity — retired industrial sites, regions with stranded generation, and anywhere with a direct line to a plant.

The second-order effects

Behind-the-meter generation is being built specifically to avoid the queue. That changes who the counterparties are: operators are now negotiating with generators rather than utilities.

Water follows power. Cooling at this scale is a water question in exactly the regions where power is cheapest and water is not abundant. Several municipalities have started refusing on that basis.

Latency-tolerant workloads move. Training does not care where it runs. Inference serving users does. The split is pushing training toward remote cheap power and inference toward expensive metro sites, which is a structural change in how the estate gets planned.

What to watch

Interconnection queue lengths in the main grids, and power purchase agreements signed directly with generators. Those two say more about compute capacity eighteen months out than any chip announcement.

Reported across the trade press; analysis ours.