The AI build has a power problem

The thing slowing the AI build isn't silicon. It's getting power to the site.

Everyone’s counting GPUs. The harder number to find is megawatts.

Talk to anyone actually trying to stand up a data center right now and you hear the same complaint: the chips aren’t the holdup, the grid connection is. Lead times on the big transformers that tie new capacity into the network have reportedly blown out to something like 128 weeks. Before the pandemic it was a couple of months. On top of that there aren’t enough electricians, a lot of the proposed sites sit where water is already tight, and a growing list of states is floating moratoria.

So you get a strange mismatch. A company can place its chip order this quarter and still wait two-and-a-half years for the power to run it. Software moves fast. Substations don’t.

That matters more than it sounds. Whatever you believe about AI (next electricity, or next dot-com), the build waits in the same physical line either way. That probably makes it slower, more expensive, and more political than the clean forecasts assume.

If you’ve read any capital-cycle history, this rhymes. Money piles into the exciting part (models, chips, IPOs) and ignores the boring part everything depends on (the wires). The boring part is usually where the return quietly ends up, precisely because nobody wanted it.

This isn’t a ‘sell AI’ take. It’s a timing one. A build gated by power is a slower build, and a slower build running into a hot inflation print and a nervous Fed is a different animal than ‘just add more chips.’

Free weekly letter and explainers like this; the daily research, levels and the Daily Compass are at moatpeak.com.

Educational research only — not investment advice. MoatPeak Group, UAB.