The data-center capital stack: four assets at one address

In most real-asset contexts, capital stack means seniority: senior debt at the bottom, then mezzanine, then preferred, then common equity, ranked by who gets paid first. That reading still applies to a data center, and the instruments that populate it are covered separately in data-center capital structuring.

This page uses the other axis, and for an AI facility it is the more consequential one. Before you can decide what ranks where, you have to decide what is being financed — and the honest answer is that a data center is not one asset at all. It is four, assembled at one address by four different processes, on four different schedules.

The conventional reading treats the building as the asset and everything else as either the land under it or the equipment in it. That works when the equipment is a minor share of the cost. In an AI facility it is not a minor share, and the framing stops describing what is actually there.

Each layer is a real asset with its own life, its own failure mode, and its own class of buyer. The table is the framework the rest of this page — and most of the rest of this site — is built on.

Land and interconnection. The parcel and the right to draw power to it — a queue position, an executed interconnection agreement, transmission access, or the permits to generate on site. This is the layer with the longest life and the least mobility, and it is the one that determines whether anything above it can exist at all. Its risks are entitlement risks: rezoning, permitting, queue standing, community position. See powered land for how the layer is qualified and speed-to-power for why its timing dominates.

Shell and fit-out. The building, the structural capacity, the mechanical and electrical infrastructure, the cooling. It is built once, over a defined period, against a defined budget, and then it stands for decades. Its dominant risk is concentrated almost entirely in the build itself — completion, cost, schedule — and falls away sharply once the facility is energised and occupied.

Generation and power equipment. Turbines, substations, switchgear, transformers, backup and behind-the-meter generation. This layer looks like part of the building and behaves like an energy asset. Its risk is delivery and availability: whether the equipment arrives on the timeline the project was underwritten against, and whether it runs at the availability the offtake assumes.

Compute. The accelerators, networking and the racks they sit in. It is the only layer that can be physically moved, the only one with a genuine secondary market, and the only one whose value can fall materially while it is still working perfectly. Its risks are obsolescence and utilisation, and it turns over several times within the life of the building around it. What capital tests before advancing against it is set out in what lenders underwrite on a GPU cluster.

Once you stop treating the address as one asset, the joints between the layers become the thing that has to be documented — because common ownership is no longer doing that work silently.

A seam is the contractual relationship that binds one layer to the next, and each one has to carry weight independently:

• Land to shell. Site control has to outlast the construction period with margin, and the parcel under the building has to be the parcel the interconnection was studied against. These come apart more often than they should. The site financeability gates are largely a test of this seam.
• Shell to power. Whether the generation is inside the facility asset or beside it, and whether the facility's right to that power survives a change of ownership on either side. A building with no durable claim on its own electricity is not a complete asset.
• Power to compute. Contracted capacity for at least as long as the compute is contracted for. Power that expires first turns a performing asset into a relocation problem.
• Shell to compute. The hosting or colocation agreement: term, termination rights, access on dispute, and what happens to the equipment at the end. Access is the precondition for every remedy, so a seam that lets one party deny it is an impairment of the layer above.

The distinction between a site with power and a building with power — and what each is worth to which buyer — is worked through in powered land vs powered shell.

The framework is not an analytical convenience. It has one consequence, and the consequence is the reason it exists.

A single instrument spanning all four layers is priced by whoever is least comfortable with the worst layer in it. The shortest-lived asset sets the term for the longest-lived one, the longest-lived one makes the shortest look better secured than it is, and every refinancing of the compute becomes an event for the entire structure. Financed as one asset, the stack misprices every layer in it — and a project that is perfectly sound once separated can be unattractive to all of the capital that would otherwise fund it.

That argument, and how the separation is actually executed, is set out in full at tenor mismatch: financing compute inside infrastructure. It is the consequence of this page and the position the rest of the site is organised around.

Continuum works the stack layer by layer: qualifying the site, establishing what each layer is and what binds it to the next, and structuring the deal so that each layer can be arranged with capital matched to its life and its risk.