Tenor mismatch: financing compute inside infrastructure

What gets described as "a data center" is a stack of assets that have almost nothing in common except location. Each has its own economic life, its own risk, and its own natural holder.

Treating the stack as a single asset produces a structure that is wrong for every layer in it.

The short asset drags the long one. Capital sized around equipment that will be obsolete in a few years cannot be advanced on the terms that thirty-year infrastructure supports. Fold the compute into the same facility and the land and shell are financed on compute's timetable — expensively, and shorter than they need to be.

The long asset flatters the short one. Run the logic the other way and it is worse. Durable infrastructure supporting the whole structure makes the compute layer look better secured than it is, because the recovery analysis is quietly leaning on assets that are not really available to cover an accelerator that has lost most of its value.

Refinancing arrives at the wrong moment. A single facility matures once. A stack whose layers turn over on different schedules needs the compute layer to refinance or roll several times over the life of the building. Bundled, every one of those events becomes an event for the entire structure, including the parts that were performing perfectly well.

Nobody can price it. Different capital wants different layers. A single instrument spanning all four is priced by whoever is most uncomfortable with the worst layer in it, which is how a good site ends up funded on terms set by its GPUs.

The alternative is not complicated in principle. Hold each layer in a way that lets capital matched to that layer fund it, and define the relationships between them contractually rather than by common ownership.

In practice that means deciding four things:

• Which entity holds which layer. Land, facility and equipment do not have to sit together, and usually should not.
• What each layer's cash flow actually is. A facility earns rent or a hosting fee; a compute position earns contracted compute revenue. Once separated, each has a stream that can be underwritten on its own terms.
• How the layers are bound. Leases, hosting agreements and access rights are what replace common ownership. They have to be durable enough that a party financing one layer is not exposed to a dispute in another.
• Where the residual risk sits. Someone carries the risk that the compute is worth little at the end. Making that explicit is what allows the rest of the structure to be priced properly.

The result is that each layer is funded by capital that understands it, on a term that matches it, and a problem in one does not automatically become a problem in all four.

The mismatch has always existed in infrastructure. What makes it acute now is the ratio.

In a conventional data center the equipment inside the building is a modest share of total cost, and treating it as part of the facility is a rounding error. In an AI facility the accelerators can dominate the capital budget outright — the shortest-lived layer in the stack becomes the largest line in it.

At that ratio the mismatch stops being an accounting nuance and becomes the thing that determines whether a project can be financed at all. A structure that works when the fast-depreciating layer is a small share fails when it is the majority.

This is also why so much capital that is genuinely interested in the sector has been slow to deploy into it. The infrastructure investors are comfortable with the land, the power and the building, and are being asked to take obsolescence risk they do not price. The equipment financiers are comfortable with the accelerators and are being asked to take development and construction risk they do not price. Presented as one asset, a project can be unattractive to both while being perfectly sound once separated.

Before the layers can be separated, four things have to be established, and they are usually established in this order:

1. How long is the compute contracted for, and by whom? This sets the term the compute layer can support and is the first thing to establish, because everything else is arranged around it. 2. What is assumed about the equipment at the end of that term? A residual assumption is a position someone is taking. It should be identified and allocated deliberately. 3. Does the site position outlast the building, and does the building outlast the compute contract? Where it does not, the mismatch is inverted and the whole structure is exposed to the layer that should have been the most stable. 4. Who holds each layer, and what binds them? The answer determines what can be financed separately and what is trapped together.

Getting these wrong is rarely fatal at the outset and frequently fatal later, because the structure is set early and the mismatch only becomes visible when the first layer needs to refinance.