SANTA CLARA, 26 AUG 2026 — Nvidia has agreed to provide guarantees of up to US$105bn supporting OpenAI's twenty-year lease of a data centre campus in Pike County, Ohio, being developed by the SoftBank-backed SB Energy.
The figure is being reported as Nvidia backing $105bn of financing, which readers reasonably hear as Nvidia spending $105bn. It is not spending anything like that. The guarantees are tied to the residual value of the leased infrastructure, and Nvidia's actual cash commitment here is US$1.5bn, invested into SB Energy.
What a residual-value guarantee is
A residual-value guarantee is a promise about what an asset will be worth later, not a payment now.
When a lessor finances a twenty-year facility, its exposure is the possibility that the tenant stops paying and the asset is worth less than the outstanding balance. A guarantee from a creditworthy third party removes that risk, and removing it lowers the interest rate on the whole structure. No money changes hands unless the asset underperforms.
So the $105bn is a ceiling on a contingent obligation, not a budget. In the case where everything proceeds as planned, Nvidia pays nothing under it, and the instrument will have done its work by making the financing cheaper than OpenAI could have obtained alone.
That is the difference between a company deploying capital and one lending out its balance sheet.
What Nvidia is actually exposed to
Contingent does not mean harmless. The exposure here is unusual.
A residual-value guarantee on AI infrastructure is a bet that the infrastructure holds its value. Buildings, substations and cooling plant generally do. The accelerators inside them are another matter, and their depreciation curve is famously steep — a top-end training part from three years ago is already well down it, displaced by successors that Nvidia itself shipped.
This creates an odd position. The company guaranteeing the hardware's future value is the same one whose roadmap is the main force reducing it. Nvidia is better placed than anyone to know how fast that happens, which is an argument that it has priced the risk properly. The guarantee is also safe only while Nvidia maintains a liquid secondary market for its older parts — a commitment it has not made in public.
How much of the $105bn attaches to short-lived silicon rather than to long-lived buildings has not been disclosed, and it is the number that would settle whether this is prudent or aggressive.
The circularity, and the precedent nobody enjoys citing
Nvidia will be the exclusive chip provider for the facility.
The pieces fit together neatly. Nvidia's credit standing makes the lease affordable for OpenAI, which then buys Nvidia compute to fill the facility. Those sales become Nvidia revenue, supporting the same credit standing that backed the lease.
This is vendor financing, and it is neither novel nor inherently improper. Capital equipment industries have used it for a century, for the sound reason that a manufacturer often has a better view of an asset's value than a bank does, and can therefore take a risk a bank would price out of reach.
This structure attracts scrutiny because it recalls the telecommunications build-out of 1999 to 2001. Equipment makers then funded their own order books, extending generous financing to carriers who could not otherwise buy, booking the resulting sales as revenue, and taking the loss when the carriers failed. The mechanism was legitimate and the volume was the problem.
Whether this is that is not knowable now, and it should not be asserted. The structure belongs to a family with a well-documented failure mode. It appears in the accounts as demand right up until the counterparty cannot pay.
SoftBank is on more than one side of this
The developer, SB Energy, is SoftBank-backed, and SoftBank's other positions make the diagram busier than a single guarantee suggests.
SoftBank is among OpenAI's largest backers. It also backs the developer building the campus that OpenAI will lease for twenty years. And the earlier US$1bn that went into SB Energy to expand data centre infrastructure came from OpenAI and SoftBank together — the tenant helping capitalise its own landlord.
None of that is hidden and none of it is improper; large investors routinely hold positions across a supply chain, and having an interest on both sides of a lease can genuinely align incentives to get a project built.
This makes the arrangement hard to read as a market price. A twenty-year lease between a tenant and a landlord who share a major shareholder, credit-supported by the vendor who will supply that tenant exclusively, is not an arm's-length transaction. It does not tell an outside observer what such capacity is worth. It is four parties with overlapping interests agreeing terms among themselves.
That matters for anyone using these announcements to calibrate the cost of AI infrastructure, because the headline numbers from deals like this are not comparable to what an unaffiliated operator would pay.
This is not an isolated instrument
The arrangement is one of several converging on the same problem, which is that AI infrastructure now needs more capital than the operating businesses generate.
We reported this month on a separate structure using compute-backed debt at a US$35bn scale, with Apollo and Blackstone on the lending side. Global data centre capital expenditure is on course to pass one trillion dollars.
When an industry uses guarantees, residual-value support and asset-backed lending all at once, it is signalling that ordinary capital has become too expensive. These instruments exist to make financing possible at rates the underlying cash flows would not otherwise support.
The physical numbers are the slow part
Against financial figures moving in tens of billions, the construction timetable is a useful corrective.
The site is planned for as much as 8 gigawatts eventually. The first 800 megawatts is expected online in 2028 — a tenth of the eventual capacity, roughly two years from now. The lease runs twenty years.
Those numbers describe the actual constraint. Capital can be arranged in weeks; grid interconnection, generation and transmission cannot. A twenty-year lease on a facility whose first tranche arrives in 2028 is not only a bet on OpenAI's solvency. It is a bet that demand for this kind of compute persists into the 2040s, in a technology whose leading architectures have not lasted five years.
What it means from here
The reporting context is Nvidia's approaching quarterly results, where revenue is expected near US$92bn with guidance above US$100bn for the following quarter. Those are enormous figures and they are the reason the financing structures deserve attention rather than the reverse.
The question for any of these announcements is simple. Which number is cash, and which is a promise? In this case $1.5bn is cash and $105bn is a contingent obligation whose composition has not been disclosed, and the two have been reported under one headline.
For operators in this region the practical read is about pricing. If frontier compute in the United States is being built on guarantees that lower its cost of capital, then regional capacity competing for the same tenants is competing against a financing advantage rather than an engineering one. That is worth knowing before concluding that a rival campus is simply better run.
