Power is the binding constraint on intelligence
Interconnect queues, transformers and cooling water gate the buildout long before chips do.
The scarce asset in compute is not silicon. It is a site with firm power and a grid connection on a timeline that matters. That single sentence reprices a set of deeply unglamorous businesses: behind-the-meter generation, grid equipment, thermal management, water rights, and developers who can carry a permit end to end.
We have a campus in the file that had chips secured, site secured, and an interconnect position putting firm power roughly three years out. The developer paid a large premium for on-site generation to bridge the gap, and that premium — not the hardware — determined the project's return. The comparable project without on-site generation is still waiting, and its tenant has already renegotiated once.
Underwriting the capital stack
These are infrastructure returns financed with infrastructure instruments, increasingly underwritten against single-tenant offtake from a handful of counterparties. The risk to price is concentration, not demand.
A twenty-year amortisation against a five-year tenant is a credit question wearing a technology costume. It is fine while demand holds and nobody asks. It stops being fine the first quarter the tenant's own efficiency work reduces draw.
Contracted versus utilised
The cleanest early tell on overbuild is the gap between contracted and utilised capacity at the largest buyers. Contracted capacity is a commitment; utilised capacity is a fact. When the two diverge for more than a couple of quarters, someone is holding stranded cost and has not written it down yet.
We track that gap monthly alongside interconnect queue positions and energisation dates in the major corridors.
The efficiency risk
If model efficiency outruns demand growth, contracted power becomes stranded cost and the equipment supply chain absorbs a demand shock it has spent three years building capacity for.
We do not think this is the base case, but it is the case that would hurt most, because the financing structures assume it cannot happen.
Where we would put money
Boring links in the chain with real lead-time moats: transformers, switchgear, thermal management, and developers with a permitting track record. None of it is exciting. All of it is gating.
This post sits inside the Energy & compute stream. The standing note holds the full thesis, bear case, and signal list.