Whose Credit Sits Behind the Rent? — when a household name stands behind AI data-centre rent, "backed" is not one structure

Last edition closed on a question. Greater Seoul's power is scarce, scarcity lifts rents — but a scarcity premium only becomes a return if it survives the trip to investor IRR, and even when the lease economics work, another question remains: whose credit actually sits behind the rent?

This edition answers it. And the answer is rarely just the tenant.

Several of 2026's largest AI infrastructure financings shared one feature: a stronger balance sheet absorbed a defined slice of downside. But the slice was different in every deal — and that difference is the whole story.

Three household names, three different objects: property-lease obligations, compute-equipment lease obligations, and defined minimum-value protection around a long-term data-centre lease. Different beneficiaries, different triggers, different recovery paths. "Backed" is not one structure — and reading it as one is the mistake this edition is about.

A guarantee is not a shield over the whole asset. It is a contract with edges — a scope, a start date, an end date, and a party on the hook. Underwrite the edges, not the logo.

What is covered — and what is not

Nvidia's instrument is not a simple rent guarantee. Per the 8-K, it protects a guaranteed minimum value tied to the OpenAI leases, triggered by specified OpenAI payment or insolvency events — and on trigger Nvidia has a menu of remedies (assume the lease, require reletting, initiate a sale, allow termination, or fund specified costs), with recoveries from reletting or sale reducing what Nvidia owes, up to the cap. The Fluidstack property backstop runs differently again: Google supports specified lease payments, not an asset's residual value. Different instruments, different risks transferred. None covers the whole.

When it starts

In the Nvidia structure and the Google-backed property examples reviewed here, the operating support activates at or around lease commencement / ready-for-service rather than at signing. Nvidia's own words: the guarantee becomes effective as data centres reach ready-for-service, expected to begin from 2028. The operating backstop does not cover the pre-commencement period — construction, completion, grid connection — so those risks sit elsewhere in the project stack, absorbed by whatever sponsor support, collateral or construction financing is separately in place. That is the same window Edition 19 flagged: in Greater Seoul, of 522 power-system applications, only a handful had cleared final approval. That pre-commencement period is exactly where scarcity and execution risk bite.

But outside the operating backstop does not mean unprotected — it means protected differently. At Cipher's Barber Lake, during construction Google warrants are pledged as collateral for the project's noteholders, and Cipher itself carries a completion guarantee to fund the amounts needed to finish. Only at lease commencement does Google's operating backstop take over.

TeraWulf's Lake Mariner runs the same playbook, building by building — US$600m of Google support became effective the moment one hall, CB-3, was delivered. Underwrite which protection is live at the stage your capital is exposed, not the label on the whole deal.

When it ends

The Ohio guarantee can terminate if OpenAI achieves the specified credit-rating threshold. Protection that amortises or terminates as the tenant matures is protection you cannot assume will remain at exit. Duration is not a footnote; it is an underwriting variable.

Who is on the hook — and how far down

Strategic supplier support can strengthen an ecosystem without creating creditor recourse. CoreWeave, for example, disclosed substantial revenue concentration in a single customer — a reminder that even heavily financed AI operators can carry downstream economics that no supplier guarantee reaches. Alignment is not recourse; keep them separate.

Demand is not cash flow

A contracted AI tenant is the start of a chain, not the end of one: contracted demand → site and power → construction → operational capacity → revenue commencement → payment obligation → credit support → recovery. Each arrow is a separate risk, and a strong tenant name collapses none of them. Theseus shows this clearly — at launch it had the tenant and the equity in place, but had not yet identified its first US sites. Contracted demand plus institutional equity is not bankability; it is the raw material bankability is built from.

Who benefits if these support structures persist? The beneficiaries can include developers and lenders whose contracted cash flows become more financeable when a stronger counterparty assumes a defined downside. Strategic support providers can benefit too — Nvidia and Google, for example, gain economic alignment through hardware demand or warrants and equity stakes — but the exact consideration differs by deal, and we should not assume a uniform fee or pricing across them. The loser is the investor who reads "hyperscaler-backed" as a risk grade and ends up carrying the risks the support doesn't cover — the construction period before it begins, the stretch after it ends, or an equity exposure whose protection architecture is undisclosed — without pricing any of them.

There is a tell worth noting. McKinsey's cost-of-compute work sorts everyone funding the AI buildout into five groups — builders, energy providers, hardware makers, operators (hyperscalers and colocation providers), and AI architects (the model developers). It puts hard numbers on the first three, but not on operators or AI architects — because their compute spending is hard to isolate, overlapping with their broader R&D and corporate activity. That is not itself a credit conclusion. But it makes the edition's point: there is a difference between what a contract shows you and the harder-to-see economics underneath it. When even a tier-one firm notes that a whole layer is difficult to measure, that difficulty is a reason for caution, not comfort — hard to see is not the same as safe.

The structures above are outside APAC. Among the APAC transactions reviewed for this edition, we did not find a publicly disclosed AI-tenant support structure directly comparable with them — a bounded finding, not an exhaustive market negative. But the exposure is not hypothetical: Singapore sovereign capital is already participating at the equity layer of a US AI-infrastructure platform, and Korean commentary is already engaging the same risk from two angles — institutional analysis treating AI data-centre credit as a distinct alternative-asset risk, and market coverage describing the "circular financing" created by interlocking strategic support in the same US ecosystem. The audience for this thesis is here before the domestic structures are.

Which raises a broader question: once you look beyond the credit support, is the cheapest asset to build still the best asset to own?