OpenAI’s 20-Year Ohio Deal Has an Nvidia Backstop

In August 2026, OpenAI signed a 20-year agreement with SB Energy for a 10GW Ohio data center campus while offering enterprise customers capacity commitments of one to three years. Nvidia will sell the accelerators and backstop part of the facility’s value at completion. Customers see adjustable capacity; SB Energy must assemble it like a power station.

Developers and financiers now treat AI infrastructure as project finance. They use long campus agreements, dedicated power, public-sector partnerships and supplier guarantees to turn uncertain future compute demand into assets they can build and fund now. By linking the same model builders, chip vendors, developers and capital providers, those contracts also allow one changed demand assumption to travel through an entire construction cycle.

Elastic cloud was designed to hide fixed assets

Traditional cloud computing separated the customer’s decision from the provider’s physical commitment. A business rented instances or storage without selecting a substation, financing a building or choosing the useful life of the hardware inside it. The provider pooled demand across customers, allowing customers to treat capacity as a service while absorbing the fixed assets behind the interface.

Large-scale AI changed the unit being financed. Cloud providers were already confronting that mismatch in 2023, when infrastructure designed for ordinary cloud workloads struggled to accommodate large AI systems. A modern AI campus operates as one machine: power supply, facility, network, accelerators and workload scheduling constrain one another. Developers finance the shell against expected tenant demand; tenants secure compute only after someone commits years earlier to land and electricity.

Cloud providers could pool standardized infrastructure among many customers. Frontier-model developers need someone to commit capital, power and credit before a concentrated workload arrives. Customers still buy elastic capacity even though the system beneath it depends on long contracts.

SB Energy assembled the contract before the campus

SB Energy’s Ohio sequence shows the developer’s new job: assemble equity, public participation, power, compute supply and an anchor customer into a package that lenders can evaluate as a whole.

• January 2026: OpenAI and SoftBank each invested $500 million in SB Energy, giving the developer and operator $1 billion to support its expansion.
• March 2026: The U.S. Department of Energy announced a partnership with SoftBank and SB Energy to develop a 10GW Ohio data center with its own power supply.
• August 2026: OpenAI signed a 20-year, 10GW agreement, while Nvidia agreed to backstop a portion of the completed facility’s value.

OpenAI and SoftBank strengthened the developer with equity. The DOE partnership addressed a site too large for an ordinary grid connection; OpenAI supplied long-duration demand, and Nvidia stood behind part of the residual asset value.

To underwrite AI infrastructure finance, lenders examine a named customer’s obligation, SB Energy’s construction role, the route to power, Nvidia’s commitment and the asset’s value if the original plan changes. Folding those promises into one package lets financiers price them as a system, the model examined in AI infrastructure project finance.

Nvidia has become seller, guarantor and buyer of last resort

Nvidia’s position in Ohio belongs to a broader change in the chip supplier’s role. The company disclosed $3.5 billion in guarantees to businesses leasing land, power and data-center facilities, four times the level disclosed in its third quarter. Nvidia has also promised to rent back unused GPUs from young cloud providers including Firmus and Sharon AI in exchange for a share of their revenues.

With a facility guarantee, Nvidia supports a developer or lessee that lacks sufficient credit. With a GPU rent-back promise, it supports utilization after the equipment has been installed. In both cases, Nvidia reaches beyond the equipment sale and assumes exposure to the market that must eventually use the equipment.

By supporting the sale, Nvidia also makes a credit decision. The company benefits when developers can order systems sooner, but the guarantee connects its financial outcome more directly to occupancy, rental rates and the residual value of completed capacity. Nvidia gains information and influence that an ordinary component vendor would lack while accepting part of the downside that vendor would leave with the buyer.

Google has taken a related route around Anthropic, assembling an approximately $200 billion financing program with more than $150 billion tied to TPUs and participation from Broadcom, Blackstone, Apollo and others. Google’s program differs from Nvidia’s guarantees, but both use a technology supplier’s position to bring outside capital into a compute buildout. The accelerator now helps organize the financing around it.

Twenty-year liabilities meet one-to-three-year products

The Ohio agreement places a long-lived campus obligation beside a much shorter commercial horizon. OpenAI’s Guaranteed Capacity product gives customers discounted compute through commitments lasting one to three years, based on spending levels. OpenAI committed to Ohio for two decades while its customers can reconsider usage every few years.

The terms need not match because OpenAI can pool customers and workloads. They run on different clocks. The campus, power system and financing must amortize over years, while model architectures, accelerator performance and the cost of AI reasoning can change within months. A capacity choice that appears economical at signing can become expensive before the physical asset has recovered its cost.

CME Group and Silicon Data made that price risk explicit by announcing futures contracts tied to daily benchmarks for on-demand GPU rental rates. The contracts make compute-price uncertainty tradable, placing a daily reference price beside a 20-year facility agreement.

For developers, the contracted megawatt permits construction that a collection of revocable cloud orders would not support. It also forces OpenAI to manage capacity optionality internally. Overbuying leaves capital idle, while underbuying limits product growth. OpenAI’s shorter customer contracts preserve flexibility for users by placing more of that balancing task on the model company.

OpenAI is both SB Energy investor and Ohio tenant

SoftBank holds stakes in both OpenAI and SB Energy. In Ohio, the group therefore has exposure to the customer expected to consume the capacity and the developer expected to deliver it. Nvidia participates as equipment supplier and partial value backstop. OpenAI participates as investor and tenant. The parties occupy different legal roles, but several depend on the same utilization assumptions.

CoreWeave moved $2.6 billion of AI-data-center construction debt off its balance sheet through special-purpose vehicles. Banks also marketed more than $56 billion of investment-grade construction loans tied to Oracle’s future leases.

An SPV gives lenders a defined set of assets and contracts to claim. Tenant demand still determines cash flow. If the tenant pays, the vehicle can isolate a project from unrelated corporate risks. If utilization economics weaken across several large tenants at once, separate vehicles can transmit the same market shock to lenders, developers and suppliers holding similar collateral.

In 2025, U.S. data-center capacity that was built, underway, planned or stalled topped 80GW. Much of that total represented future delivery rather than operating facilities, so financing decisions depended on forecasts made before customers could demonstrate actual usage at the completed sites.

The contracts assign the first loss at each stage. They cannot ensure that construction, power, utilization and technology fail for different reasons.

Power scarcity leaves revenue forecasts exposed

Hyperscalers had announced 46GW of AI data-center capacity, enough at full utilization to consume power comparable to roughly 44.2 million U.S. households. Electricians, substations, generation and firm grid service cannot expand at software speed. Those limits prevent every announced campus from arriving simultaneously.

Ohio has protections that an ordinary speculative development lacks. The Department of Energy participates in the project, and SB Energy’s plan includes its own power supply. Those features can reduce permitting and grid-availability risk relative to a campus waiting in a conventional interconnection queue. Scarcity can therefore increase the value of sites that already possess credible power and public support.

A powered campus still faces technology refreshes, tenant renegotiations and lower compute prices. A project with strong customer demand can also stall when its power or construction timetable becomes uneconomic. Firm megawatts separate viable sites from paper proposals while leaving workload forecasts unresolved.

Revenue, backlog and powered land settle different claims

Operating campuses already have substantial demand. CoreWeave reported $2.58 billion in second-quarter revenue, up 112% year over year, alongside a $104 billion revenue backlog and 1.5GW of contracted power. Meta was also reported to be under contract for roughly 1.6GW of Crusoe capacity across facilities in Texas and Missouri. Customers are spending money, signing contracts and reserving power at industrial scale.

Underwriters read each measure differently. Quarterly revenue shows that operating capacity has found paying workloads. Backlog records contracted future business, subject to delivery and customer credit. Contracted power establishes access to a scarce input. The three measures sit at different points between land acquisition and durable cash flow.

Crusoe demonstrated the distinction inside a single developer. Meta reportedly contracted for its Texas and Missouri capacity, while Crusoe paused a proposed Wyoming data center after failing to secure customers including Google amid concerns about cost and timetable. At Applied Digital, delayed lease renewals contributed to quarterly revenue of $53 million, below the $63 million estimate.

Lenders and developers therefore have to examine the asset beneath the headline. They ask whether construction can be phased and power can serve another tenant. They also test whether the facility can accept new accelerator generations, whether scheduling can keep installed systems occupied and whether the anchor customer remains creditworthy through a technology refresh. A signed agreement answers the first financing question; operations answer the rest.

To the user, a token stream still appears from nowhere. In Pike County, it begins with a dedicated-power plan, a developer-owned campus, a 20-year obligation and Nvidia’s promise over part of the completed facility. The closing binder gives every ton an address.