Power is the ceiling

On paper, compute is a manufacturing problem: build more capacity, ship more accelerators. In practice, the accelerators are arriving faster than the power to run them. A single data-centre campus now asks for load measured in hundreds of megawatts, on a schedule the grid was never designed to meet.

The constraint is physical and slow to clear. Generation, transmission, and the transformers and switchgear in between all carry multi-year lead times, and much of that equipment is already spoken for. You cannot order a substation the way you order servers.

So the question behind any AI-infrastructure view is not only who makes the chips, but who can energise them, and when. Sites with firm power and a grid connection become scarce assets. The equipment that delivers that power becomes a quieter, longer-duration way to own the same trend.

We watch generation, grid interconnection and electrical equipment as closely as we watch silicon. The build-out is gated there. This note is informational only, and not investment advice.