How it works
The problem it solves is a familiar one. A lender is asked to fund a very large, very specific asset — a gigawatt-scale hall full of GPUs, say — on a long lease to a tenant whose credit the lender does not like. The lender's question is not "will the rent get paid" so much as "what do I own if it doesn't". For a purpose-built asset the answer is usually: something worth much less than it cost.
A closes that gap by having a third party promise a floor. If the asset, at the trigger date or after a default, fetches less than the agreed number, the guarantor writes a cheque for the shortfall. The lender can now size the loan against the floor rather than against a guess. Structurally it is a written put: the guarantor has sold someone the right to hand them a disappointing asset at a fixed price.
An illustration, with invented round numbers. A lender puts up $10bn against a campus. The backstop sets a floor of $6bn on the completed asset. The tenant walks in year six, the site is eventually re-leased or sold for $4bn, and the guarantor pays the $2bn difference. The lender's exposure was never really to the tenant at all.
Two details do most of the work in real deals, and both cut against anyone reading the headline number.
The first is the cap. Backstops may not cover the whole asset or the whole lease. They can cover a stated slice, such as only the first phase of a multi-phase build, and cover value rather than rent. A twenty-year lease with a headline contract value in the tens of billions can sit on a backstop that is a fraction of it.
The second is the waterfall. A backstop can require the beneficiary to exhaust mitigation first: try to re-lease the site, and only then claim the residual shortfall. Where it does, the guarantor's real exposure is smaller and much later than the headline suggests. It also means nobody finds out what the guarantee is worth until years after the thing has gone wrong.
In return for writing the put, the guarantor gets something. This might include exclusive supply into the project or an equity stake in the developer. That is the tell for what the instrument really is: not charity, and not insurance either, but a subsidy paid in contingent liability to manufacture demand for your own product.
Why it matters to this crash
Our thesis is that the credit of the buildout is being manufactured on one balance sheet. The residual-value backstop is the mechanism by which that happens.
When we looked at the structure of the Ohio deal on 19 August 2026, the shape was exactly this. OpenAI signed a 20-year lease on a 10GW campus with SoftBank's SB Energy; Nvidia agreed to backstop a portion of the value of the completed data centre, starting with phase one of around 5GW; Nvidia became exclusive chip supplier for the first half of the site and took equity in SB Energy. The commitment, per the Wall Street Journal, "is structured to require a series of mitigation steps before the chip maker would be obligated to make any payments". We wrote at the time that this "is not a guarantee of $105bn of rent. It is a residual-value floor with a waterfall in front of it", and that our estimate of Nvidia's direct exposure went down on that reporting.
That is the honest read of one deal. The systemic read is less comfortable. Every backstop of this kind rests on an untested assumption: that a purpose-built GPU hall retains substantial with a tenant other than the one it was designed for. Nobody knows whether that is true, because no such asset has yet been re-leased to anyone. And the risk is wrong-way by construction. The world in which the floor gets tested — AI demand disappoints, the anchor tenant cannot pay, second-hand compute is abundant — is precisely the world in which the guarantor's own shares and credit are worst placed to absorb it.
What would make this dangerous
A first real re-leasing or sale of a large purpose-built AI campus, at any price, because it would replace assumption with a comparable. If that price prints well below cost, every floor in the complex is instantly worth arguing about.
Aggregate backstop notionals appearing in a guarantor's contingent-liabilities footnote and turning out to be a large multiple of what individual deal reporting implied.
Backstops layered on the same tenant across multiple developers, so that a single counterparty failure triggers several floors at once rather than one.
Lenders pricing loans as if the floor were a payment obligation rather than a claim at the end of a multi-year mitigation waterfall. If credit committees have modelled prompt payment on a guarantee that pays slowly, contingently and capped, the mispricing sits in the loan book, not the guarantee.
And the plainest signal: the guarantor's own widening while it continues writing new floors. A backstop is worth exactly the credit of the entity behind it, and no more.