Nvidia's $105B OpenAI backstop is a residual-value guaranty, not rent insurance
Nvidia's credit support for OpenAI's Ohio data center is a residual-value guaranty capped at $105 billion, triggered only after a default and a failed relet or sale. The cash Nvidia is actually committing is $1.5 billion, into landlord SB Energy.

Nvidia is providing credit support worth up to $105 billion on OpenAI's Ohio data center campus, reported by Semafor and Implicator. The instrument is a residual-value guaranty, and that is not the same thing as guaranteeing OpenAI's rent.

What the instrument actually is
The widely repeated version of this deal is that if OpenAI cannot cover the payments, Nvidia is on the hook. That describes a lease guaranty or a surety bond, where the guarantor steps in and pays the missed rent. It is not what Nvidia signed.
A residual-value guaranty covers the shortfall in the asset's value, and only at the end of a sequence:
- OpenAI defaults, through insolvency or missed rent.
- The landlord tries to re-let or sell the buildings, and fails to recover the expected value.
- Only then does Nvidia cover the residual shortfall, up to the $105 billion cap.
Nvidia sits in a third-loss position behind a default and a failed remarketing. The $105 billion is a ceiling on a contingent obligation, not a number anyone expects to move. The distinction matters for how you read Nvidia's balance sheet: this is not $105 billion of debt or spending, and it is not $105 billion of expected loss.
What it does do is make the asset financeable. Lenders and the landlord underwrite the buildings knowing that a floor exists under their resale value, which lowers the cost of capital for the whole campus.
The counterparties and the asset
The site is PORTS-Pike, in Pike County, Ohio, on the grounds of the former Portsmouth uranium enrichment complex.
- SB Energy, the SoftBank-backed developer, builds, owns and operates the campus. It is the landlord.
- OpenAI is the tenant, on a 20-year lease.
- Nvidia provides the credit support and is the exclusive chip supplier to the site.
That last pairing is the part worth sitting with. Nvidia is underwriting the residual value of buildings that will be filled exclusively with Nvidia silicon. If the tenant fails and the campus has to be remarketed, the asset Nvidia is guaranteeing the value of is a facility purpose-built for its own hardware. The guaranty and the supply agreement are not independent bets.
The guaranteed asset is described as finished land, power and shell buildings, not the chips inside them. Capacity is roughly 4.25 IT-GW initially, with an option for about 3.8 GW more, so close to 8 GW if fully exercised. IT-GW refers to critical IT load, which is smaller than total facility draw once cooling and overhead are counted. Ready for service from 2028.
The number nobody printed
Nvidia is also investing $1.5 billion in SB Energy, the landlord. That is the actual cash leaving the building.
Set the two figures side by side: $1.5 billion committed against a $105 billion contingent cap. The cash is about 1.4% of the headline number. Nearly every writeup led with $105 billion, which is the least likely of the two to ever be paid, and buried or omitted the $1.5 billion, which is certain. Nvidia is buying equity exposure to its customer's landlord for the price of a mid-size acquisition, and providing a contingent floor that costs nothing unless a chain of failures completes.
The primary source is the filing describing the arrangement, and the operative language on trigger conditions and the cap sits there rather than in the coverage. Nvidia's filings are on SEC EDGAR.
Why the circularity question survives the correction
Getting the instrument right does not dissolve the concern, it sharpens it. Fortune notes the arrangement came in roughly $145 billion below earlier accounts, which it reads as a signal about artificial chip demand.
The real circularity is not that Nvidia is paying OpenAI's rent, because it is not. It is that Nvidia is using its balance sheet to lower the financing cost of buildings that exist to house Nvidia GPUs, bought by a customer whose ability to pay depends on capital markets that are themselves reading Nvidia's involvement as validation. The support is contingent and cheap. The demand it helps underwrite is booked as real.
The takeaway
Do not model $105 billion as an Nvidia liability. Model $1.5 billion of equity in SB Energy and a contingent residual-value floor that only pays after an OpenAI default and a failed relet, and read the filing for the exact trigger language before assuming where in the waterfall it sits. The number to watch from here is not the cap. It is whether the roughly 3.8 GW option gets exercised, and whether the next campus of this size also needs a chip vendor to put a floor under the real estate before anyone will finance it.
Subscribe to Techpresso
Free daily newsletter, read in 5 minutes.
Subscribe free