How it actually works
Anyone financing a long-lived physical asset is underwriting two separate things. The first is the contracted cash flow: rent under a lease, payments under a power contract, instalments on a loan. That part is a credit question, and it is answerable. You look at the tenant, read the lease, and price the chance they stop paying.
The second is the residual: whatever the thing is worth on the day the contract ends. That is not a credit question. It is a question about a market that may not exist yet.
Some illustrative arithmetic, with invented round numbers. Suppose a data centre costs $1bn to build and is let for fifteen years to a tenant whose rent, discounted, covers $700m of it. The remaining $300m is not covered by anything anybody has signed. It is covered by an assumption: that in year sixteen the building can be re-let, or sold, for at least that much. Underwrite the residual at $300m and the deal works. Underwrite it at zero and the sponsor needs $300m more equity, or the lender needs a much bigger coupon. The number is a judgement, and a judgement is a thing that can be wrong quietly for years.
What makes purpose-built AI capacity awkward is that almost every component of the residual is contested. The shell and the power interconnect might be durable. The cooling is specified for a particular density of accelerator. The accelerators themselves depreciate on a schedule nobody agrees about. And the pool of potential replacement tenants for a gigawatt-scale hall is roughly the same handful of firms that are currently building their own.
The other residual
Our dispatches use "the residual" in a second, related sense: the bottom slice of a securitisation, the equity piece the sponsor keeps, which only absorbs losses once the below the bondholders is exhausted. Same idea, different asset. What is left over after the contracted claims are satisfied, applied to a pool of loans rather than a building. We used it that way on 19 August 2026 to explain why lenders' shares can sit near record highs while sixty-day delinquencies run at 6.9% on the Fitch and Equifax data: rising losses eat the retained residual long before they reach the equity, which is also the subject of our entry on .
Why it matters to this crash
The residual is where the financing structure keeps its unfalsifiable assumption, and the has noticed.
When QTS marketed the Microsoft-linked "Project Odyssey" notes in Georgia, the paper was expected to be rated Baa3 by Moody's and BBB− by Fitch, and talked at a yield around 7.63%, which is where single-B junk trades. We wrote on 19 August 2026 that with the investment-grade index at 81bp, this was roughly four times the average with an investment-grade stamp attached. The agencies were rating the lease. The buyers were pricing the residual, and the gap between the two is the price of not knowing.
The same assumption is load-bearing everywhere else in the complex. Nvidia's commitments, up to $125bn across its financing platforms as we noted on 19 August 2026, are not guarantees of rent. The Ohio arrangement with SB Energy and OpenAI is explicitly a residual-value floor sitting behind a waterfall of mitigation steps, including an obligation to try re-letting the site first (WSJ). Meta and BlackRock's roughly $14bn El Paso campus carries $450mn of all-risk cover once operating, meaning the tail beyond that sits with bondholders rather than with an insurer that would have to reserve against it. And the SEC's clarification that a data centre is not a "financial asset" switches off risk retention and Regulation AB disclosure precisely where tenant concentration and residual value live (Telegraph).
Goldman analysts put lease commitments at about $1.5tn, up from roughly $200bn five years ago. A very large share of that number rests on a residual assumption that has never been tested against a transaction.
What would make this dangerous
The first observable print is the one to wait for: a purpose-built AI hall re-let to a tenant other than the one it was designed for, at a disclosed rent. If that rent comes in materially below the original, every residual assumption in every deal behind it is wrong at once, and the follow.
Short of that, watch for a actually being called, with the mitigation waterfall exhausted and Nvidia writing a cheque. Watch for the same sponsor pricing the same asset type wider a year apart, as the El Paso bonds did against their 2025 comparable. Watch for a ratings agency changing its residual assumption on an existing deal rather than a new one. And watch for any hyperscaler shortening the depreciation life it assigns to accelerators, which is the same admission made in accounting rather than in the bond market.