Nvidia offers up to $105 billion in residual-value guarantees for OpenAI's Ohio data centre campus
A chip vendor is underwriting the second-hand value of its own hardware to get an 8 GW campus financed, which changes how AI capacity gets built.
Nvidia has agreed to provide up to $105 billion in residual-value guarantees to support a large data centre campus in Pike County, Ohio, that will house OpenAI's compute, Data Center Knowledge reported on 19 August 2026. Rather than a direct investment in the site, Nvidia is backstopping the value of the infrastructure and the hardware inside it, alongside a separate $1.5 billion equity investment in SB Energy, which will build, own and operate the campus.
The site, known as the PORTS-Pike Technology Campus, is planned for an initial 4.25 GW of IT capacity, with an option to add a further 3.75 GW for a total of 8 GW. OpenAI has signed a 20-year lease, and Nvidia is the exclusive provider of AI compute infrastructure for the campus.
The guarantee works as a safety net for the owner. If specified default or insolvency events occur, SB Energy would seek to recover value by re-leasing the facility or selling the hardware. Nvidia would then cover any shortfall between what is recovered and a guaranteed minimum lease value. OpenAI has agreed to reimburse Nvidia for any payments actually made under the guarantee, so the chip company's exposure is intended as a bridge rather than a permanent liability.
Nvidia chief executive Jensen Huang said the company was securing long-lived infrastructure for its compute so that OpenAI could deploy what he called the most productive AI factories. Steven Dickens of HyperFrame Research observed that Nvidia is recognising that GPUs retain value well beyond a typical three-year horizon, which makes such a guarantee feasible.
Why it matters
The structure ties the financing of gigawatt-scale campuses to the resale value of accelerators, making a GPU's useful life a balance-sheet question as much as an engineering one. It also shows suppliers moving from selling chips to underwriting the projects that consume them. Lenders and regulators will be reading the fine print of deals like this closely.
The structure ties the financing of gigawatt-scale campuses to the resale value of accelerators, making a GPU's useful life a balance-sheet question as much as an engineering one.
SB Energy builds, owns and operates the campus
PORTS-Pike Technology Campus, Pike County, Ohio: 4.25 GW initially, up to 8 GW.
OpenAI leases the site for 20 years
Nvidia is the exclusive supplier of AI compute infrastructure.
A specified default or insolvency event occurs
SB Energy recovers what it can by re-leasing the facility or selling the hardware.
Nvidia covers the shortfall
Up to $105 billion against the guaranteed minimum lease value.
OpenAI reimburses Nvidia
Any payments actually made under the guarantee are owed back to Nvidia.
What you can learn from this
- IT capacity in gigawatts describes power, not compute. When a campus is quoted at 4.25 GW of IT capacity, the number is the electrical power available to servers and networking, excluding cooling and building overhead. It is used because power is the scarce input; the number of accelerators that fit is derived from it by dividing by the power per rack. A gigawatt is roughly the output of a large power station, which puts the scale in context and explains why grid connection, not construction, often sets the timeline.
- Residual value is what an asset is worth after its planned life. Accounting commonly depreciates computing hardware over a few years, after which its book value approaches zero. If the equipment can actually be re-leased or sold for a meaningful sum, the real residual value is higher than the books show, and someone willing to guarantee that value can lower the cost of borrowing against it. The guarantee in this deal is a bet on exactly that gap, and the bet only pays if a second-hand market for accelerators continues to exist.
- A lease separates who owns the building from who uses it. SB Energy owns and operates the campus, OpenAI rents it for 20 years, and Nvidia supplies the hardware inside. This division lets each party carry the risk it understands best: construction and operation, workload demand, and equipment value. It also means the failure of one party does not automatically strand the physical asset, which is why re-leasing is the first recovery step before any guarantee is called.
- Credit enhancement lowers the cost of capital. Lenders price a loan according to the chance they are not repaid. A guarantee from a company with strong finances reduces that chance, so the project can borrow more cheaply or at all. The economic effect is that the guarantor's balance sheet is lent to the project without cash changing hands unless a default happens, which is why the headline figure is a ceiling on exposure rather than money spent.
- Long-lived infrastructure outlives several hardware generations. Buildings, substations and cooling plants are planned for decades, whereas accelerators are replaced every few years. Designing a campus to accept several generations of equipment is what turns a hardware-specific site into a durable asset, and it is why the deal language stresses long-lived infrastructure rather than any particular chip. The 20-year lease term makes sense only under that assumption.
We teach this
How to use this in practice
- Build a depreciation-versus-resale sheet. In a spreadsheet, model an accelerator with an assumed purchase price depreciated straight-line over three years, then add a second row with a hypothetical resale value for each of the next six years and calculate the gap between the two. Add a third row that sums the gap across a fleet of a thousand units. Done is a chart that shows how much value a guarantee would need to cover in each year, which makes the deal's logic concrete.
- Convert gigawatts to racks. Take 4.25 GW, assume a rack density of 100 kW, and calculate how many racks that supports; then repeat at 140 kW and at the 20 kW typical of older halls. Multiply the 100 kW result by an assumed number of accelerators per rack to get a rough chip count. Done is a note with all three numbers and a sentence on how much the answer changes with density, which is a useful habit when reading any campus announcement.
- Draw the parties and the money flows. Sketch four boxes for SB Energy, OpenAI, Nvidia and the lenders, then draw arrows for lease payments, hardware supply, equity, the guarantee and the reimbursement obligation. Number the arrows in the order they would fire if the tenant stopped paying, from re-leasing through to reimbursement. Done is a diagram where every arrow has a direction and a label and you can trace what happens in a default.
- Read one public power-capacity filing. Find a utility or grid operator's published interconnection queue or large-load report for any region and locate one data centre entry, noting its requested megawatts, its target energisation date and how long it has been in the queue, then compare the requested figure with the 4.25 GW in this story for a sense of scale. Done is a short paragraph describing the entry in your own words, which connects announcement numbers to the grid processes that gate them.
Sources
- Nvidia Backs OpenAI's Ohio Data Center Buildout With $105B Guarantee — Data Center Knowledge
Our reporting is an original summary; full coverage is at the links above.
Don't just read about it — build it.
Square 1 teaches the skills behind the headlines, with every line of your work graded by AI. Find your starting point in 3 minutes.
Get your free skill report