Four five-year leases are doing the work of a twenty-year AI commitment

A take-or-pay contract that the sponsor does not carry as debt still gets a building financed, and the lender ends up underwriting a name against which it holds no recognised claim ^[2]. That is the whole mechanism, and it sits one party away from the cash flows everybody models, because in this structure it is the frontier labs that contract for, lease and build the data centres ^[3], while the companies whose demand justifies the concrete are multi-line public issuers whose AI activity is one segment among several and disclosed accordingly ^[9].

So the arithmetic that matters is not annual capex but committed term. Five years contractually binding against fifteen years of renewals is seventy-five per cent of a twenty-year headline sitting at the tenant's discretion ^[14], and the post is dated September 2026 ^[15], which puts the near edge of the reset window roughly four months out and the far edge roughly twenty-seven ^[16]. Naked Capitalism's evidence for how thin the disclosure is happens to be circumstantial and persuasive: Ed Zitron has to write many multi-thousand-word posts to establish that the main players ex Nvidia are showing paltry revenue against huge and growing capex ^[10].

Underneath the leases is the funding, and the funding is where the sizing fails. Even the cash-rich incumbents lack the internal resources and do not expect equity issuance to cover it, so they borrow ^[11]; the bonds are issued for general corporate purposes rather than earmarked ^[5]; and the private credit layer that carries the rest contains funds buying private equity investee debt, borrowing through subscription lines, and taking money from hedge funds and individuals who themselves borrowed to invest ^[7]. Which layer count you are in decides the outcome: one layer produced Japan's deep downturn and what Richard Koo called a balance sheet recession, while levered borrowings on top of borrowings produced 1929 and 2007-2008 ^[12]. Subprime, for all of it, was documented, with loans pooled into securitisations and title transfers and sale prices on public record, costly to aggregate but there ^[8].

The window can resolve three ways. Revenue arrives, the labs renew, and the resets are administrative. The labs decline the second term, and the loss lands on developers and their lenders, which hurts and stops there if those lenders are unlevered. Or those lenders are levered in the manner the Financial Times and others could not quantify ^[6], in which case containment is an assumption rather than a finding. My reading, and it is probably wrong in emphasis rather than direction, is that the reset date is the more informative number than the capex line, since capex can be slowed in a quarter and a lease expiry cannot be moved. Naked Capitalism's reference point for how quickly a covenant-grade name can decay is General Electric, once AAA and since fallen from investor grace, which is the question it then puts to Meta ^[13]. Two things would kill the thesis: if the take-or-pay obligations turn out to be parent-guaranteed rather than lab-level, the one-party-away problem disappears, and if the twenty-year structure is a recollection rather than a document ^[4], as the author concedes it is, the wall loses most of its height.