NVIDIA Put a $105B Backstop Behind OpenAI’s Ohio Campus

NVIDIA has put its balance sheet behind part of OpenAI’s next infrastructure build. The headline number is enormous: a cumulative cap of $105 billion. The contract is more precise—and more useful—than the headline. NVIDIA is not paying that amount today, lending it to OpenAI or guaranteeing every cost at an eight-gigawatt campus. It has agreed to cover a defined residual-value shortfall on leases for the first 4.25 gigawatts if specified defaults occur after completed premises are ready for service.

In brief, for the window from 17 August 09:04 to 18 August 09:04 in Tehran:

• NVIDIA signed residual-value guarantees tied to about 4.25 gigawatts of IT load at SB Energy’s PORTS-Pike campus in Ohio.
• OpenAI is the tenant; NVIDIA’s initial aggregate payment obligation is cumulatively capped at $105 billion and begins only as lease and ready-for-service conditions are met.
• The arrangement moves a material part of tenant and residual-value risk onto the chip supplier, while OpenAI has agreed to reimburse NVIDIA for amounts it actually pays.
• A separate venture-capital tape reported $1.23 billion across four AI and robotics rounds, but accessible primary transaction notices were not available by the cutoff, so those figures remain secondary-source reports.

The day’s central signal is not simply that another very large data centre was announced. It is that securing AI demand now requires an explicit allocation of credit, completion, power and remarketing risk across the customer, developer and hardware vendor.

The evidence ledger

NVIDIA’s Form 8-K filed on 17 August 2026 is the controlling source for the guarantee, with timing and filing classification shown in the EDGAR filing index. The company filed it under entry into a material definitive agreement and creation of a direct financial obligation or obligation under an off-balance-sheet arrangement. That classification matters more than promotional descriptions of a partnership.

What NVIDIA actually signed

SB Energy will build, own and operate the PORTS-Pike data-centre campus. OpenAI will be the tenant under a 20-year lease, and the site is intended to host NVIDIA compute. The first signed support covers leases representing approximately 4.25 gigawatts of IT load. NVIDIA can choose, in its sole discretion, to support roughly another 3.8 gigawatts.

The SB Energy announcement dated 17 August describes land, power and shell credit support for an initial 4.25 gigawatts, with the remaining capacity as an option. It also says the campus will come online in phases beginning in 2028, that at least 10 gigawatts of new generation is planned, and that SB Energy and AEP Ohio expect at least $4.2 billion of regional grid investment. These are plans and commitments, not delivered capacity.

Each guarantee generally becomes effective when its corresponding lease starts. NVIDIA’s payment duty is also conditional on the lessor satisfying ready-for-service requirements for the premises. The filing says those conditions are expected to begin in 2028. Capacity that is late or not ready does not create the same exposure as an operating lease.

That phasing is operationally important. A single $105 billion headline compresses several premises, commencement dates and declining obligations into one number. Finance teams evaluating similar deals should model exposure by phase rather than treating the cap as day-one debt.

The cap is not a cheque

The trigger is narrower than “OpenAI no longer wants the capacity”. NVIDIA’s filing identifies two trigger events: an OpenAI insolvency that causes a lease default, or OpenAI’s failure to make lease payments. If one occurs, NVIDIA generally pays the shortfall between the guaranteed minimum lease value and proceeds recovered through a replacement lease or sale.

NVIDIA then has choices. It may assume the lease, require SB Energy to seek a replacement tenant, initiate a sale, permit termination, or defer a remedy for as long as one year while paying specified project costs. The structure therefore resembles a residual-value backstop, not an unconditional promise to pay all future rent immediately.

The obligation ends at the earliest of several events, including the twentieth anniversary of lease commencement, a permitted lease termination, OpenAI reaching a satisfactory credit rating, or other customary termination events. OpenAI has agreed to reimburse and indemnify NVIDIA for amounts NVIDIA actually pays to SB Energy.

That reimbursement is valuable but does not erase credit risk. A reimbursement claim is least useful precisely when a tenant is insolvent or unable to pay. It changes NVIDIA’s legal recovery path; it does not guarantee cash recovery.

Why a chip supplier became a guarantor

NVIDIA is using credit support to reserve scarce land, power and building capacity for systems built around its stack. Its 17 August newsroom release says it will also invest $1.5 billion in SB Energy. The commercial exchange is clear: NVIDIA accepts contingent exposure in return for a long-lived, exclusive venue for NVIDIA compute.

The company’s technical and financial explanation published on 17 August estimates that each generation deployed at the initial site could require about 1.5 million NVIDIA GPUs and represent $150 billion to $200 billion of NVIDIA revenue. Those figures are management estimates, not contracted revenue disclosed in the 8-K. They depend on future system mix, pricing, installation schedules, upgrade cycles and OpenAI demand.

NVIDIA also argues that a standardised site could be remarketed to another tenant if OpenAI failed. That is plausible, but not frictionless. A replacement customer would need the credit, demand, power profile and operating capability to absorb capacity measured in gigawatts. Reletting risk can be reduced by a reusable design; it cannot be assumed away.

This is a more explicit version of a pattern already visible across the sector. Our analysis of Nebius capacity prepayments showed how customer cash can finance buildout before infrastructure is fully deployed. The CoreWeave and Lumentum constraint review showed that contracted demand still depends on power, components and completion. Here, the support instrument is neither a prepayment nor ordinary project debt: it is a vendor-backed guarantee designed to make a very large lease bankable.

Where the risk now sits

The arrangement divides rather than removes risk:

• SB Energy carries delivery risk. It must assemble land, generation, grid connections, shells and operations on the promised schedule.
• OpenAI carries lease and utilisation risk. It pays as completed capacity comes online and must find productive demand for the compute over a long term.
• NVIDIA carries contingent credit and residual-value risk. A qualifying OpenAI default can turn commercial support into payments, lease assumption or remarketing work.
• Utilities and public bodies carry execution dependencies. Transmission, generation, remediation and permits remain prerequisites even where project sponsors say ratepayers will be protected.
• The supply chain carries upgrade risk. A 20-year site will host multiple hardware generations, so cooling, power density and networking must remain adaptable.

The federal Ohio energy project fact sheet provides further background: the wider plan includes at least 9.2 gigawatts of gas generation, new high-voltage lines and four substations. These physical dependencies are a useful counterweight to the financial cap. A guarantee can improve a project’s credit profile; it cannot create completed megawatts.

The smaller capital tape

Axios Pro Rata reported on 17 August that AI video company Higgsfield raised $400 million, inference provider Groq raised $350 million, voice-dictation company Wispr raised $280 million, and construction-autonomy company Gravis Robotics raised $200 million. The reported total is $1.23 billion.

Those rounds show capital continuing to fund applications, inference and physical automation alongside infrastructure. But the brief did not find accessible company announcements or transaction filings for those four rounds by the cutoff. Their amounts, stages and valuations should therefore be read as attributed reporting, not independently confirmed facts. An acquisition report in the same news cycle was excluded because it lacked on-record confirmation.

What remains unknown

The 8-K gives a term-level map, not the complete contracts. NVIDIA says the forms of agreement will be filed with its Form 10-Q for the quarter ended 26 July 2026. Until then, readers do not have the phase-by-phase guaranteed minimum values, detailed ready-for-service tests, credit-rating threshold, project-cost definitions or all termination provisions.

The public material also does not establish how quickly replacement proceeds would be realised after a trigger, what discount a sale could require, or how NVIDIA will measure the obligation in its financial statements. The filing identifies the regulatory category, but the later quarterly report should provide the fuller accounting treatment and quantified exposure.

There is also a concentration question. OpenAI is simultaneously the intended user, lease obligor and source of reimbursement if NVIDIA pays. NVIDIA is the technology supplier, investor in the developer and guarantor. These aligned incentives can accelerate construction, yet they also make demand, credit and supplier economics less independent. Our review of AI contract debt and revenue concentration explains why contract value should be separated from delivered capacity, recognised revenue and cash collection.

What to watch next

• The next NVIDIA 10-Q for the filed agreement forms, accounting classification and any fair-value or maximum-exposure disclosure.
• Lease commencements and ready-for-service certificates, not just construction starts.
• Power-generation, transmission, remediation and permitting milestones for capacity beyond the first 800 megawatts.
• Whether NVIDIA exercises the option for the remaining roughly 3.8 gigawatts.
• Evidence that OpenAI demand and cash generation scale with the lease schedule.
• Any change in the expected 2028–2030 delivery window or the proposed grid funding structure.

The operational lesson is precise. AI infrastructure is no longer constrained only by chips or capital in the abstract. It is constrained by which balance sheet is willing to stand behind a particular site, tenant and completion schedule. NVIDIA’s $105 billion cap makes that allocation visible. The next disclosures must show how much of that contingent promise becomes effective, phase by phase.