The $250B Nvidia Guarantee Talks Shipped as $105B. The Shadow Bank Now Has a Signed Contract.

On Monday, August 17, 2026, the leak became a contract. Nvidia, OpenAI, and SoftBank finalized the residual value guaranty behind the PORTS-Pike Technology Campus in Pike County, Ohio, and disclosed it through an 8-K the same afternoon. The signed number is up to $105 billion of Nvidia backing on the first 4.25 gigawatts of IT load, with an option on the remaining 3.8 gigawatts at Nvidia's sole discretion. SB Energy, the SoftBank subsidiary, builds, owns, and operates the site under a 20 year OpenAI lease. Nvidia writes a $1.5 billion equity check directly into SB Energy and becomes the exclusive AI compute infrastructure provider across the entire 8 gigawatt footprint. First capacity comes online in 2028.

We covered the leak in late July, when the Wall Street Journal and Bloomberg put the talks near $250 billion on a 10 gigawatt project. The signed shape is smaller, tighter, and structurally more specific. It is also the second of two Nvidia disclosures inside eight days that make the same point. This is no longer a talk. This is the operating model.

What Actually Got Signed

The instrument is called a residual value guaranty. Nvidia is not writing a check on day one. Nvidia is promising SB Energy and its lenders that if OpenAI walks away from the lease or files, Nvidia covers a defined portion of the shortfall on lease and power payments, and it stands behind a minimum value for the underlying real estate and shell infrastructure. No cash out the door today. A footnote until it triggers. Uncapped only in the specific sense that the triggered obligation is whatever the contract says, capped in the general sense that the ceiling on the initial phase is the disclosed $105 billion.

What Dropped From the July Talks

Three things moved. The topline came in at $105 billion for the initial phase instead of the $250 billion Bloomberg had floated across the full 10 gigawatt basket. The scope tightened from 10 gigawatts to 8, with the guaranty attaching to 4.25 up front and the rest sitting behind an Nvidia option. The exposure is now explicitly phased, tied to specific gigawatts, and gated on OpenAI actually taking delivery.

The second move is what got added. The July talks were about credit support. The signed deal packages the credit support with equity in the counterparty (the $1.5 billion Nvidia check into SB Energy), exclusivity on the compute side (only Nvidia silicon lands in the racks), and a 20 year OpenAI lease shape that makes SB Energy's revenue schedule roughly co-terminous with Nvidia's contingent obligation. Nvidia is not just underwriting the tail risk. Nvidia is the customer of the customer, the shareholder of the landlord, and the sole silicon vendor on the site. Every dollar SB Energy collects from OpenAI passes through a schedule Nvidia is on both sides of.

The third move is what did not get added, and it matters. The July talks had a separate line for up to $350 billion of chip purchase financing sitting on top of the Ohio number. That line is not in the 8-K. The market read of the signed deal versus the leak should not be that Nvidia backed off. It should be that Nvidia disaggregated. Chip financing lives in its own instrument now, presumably routed through the $500 billion third party platform Nvidia announced with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR on August 10. The Ohio guaranty covers the real estate and power stack. The silicon gets financed on its own paper by outside investors, off Nvidia's balance sheet.

The Residual Value Guaranty as a Product

A residual value guaranty is a specific object in corporate finance and it is worth pinning down. When a company leases something (an aircraft, a fleet of trucks, a data center shell), the lender behind the lease worries about two things. One, will the tenant pay every month for the term. Two, will there be anything worth foreclosing on if the tenant does not. A residual guaranty from a third party addresses the second worry directly. The guarantor promises that if the tenant defaults and the lender has to seize and sell the asset, the sale will clear a minimum defined value, and the guarantor makes up the difference.

The lender then rates the debt as if that minimum were the actual recovery value. That lets the debt price at rates a hyperscaler campus with a single startup tenant would not otherwise clear. The tenant gets cheaper capital. The landlord gets a bankable project. The guarantor gets exclusivity, board seats, and a chair at the negotiating table for every future phase. In this deal, the guarantor is a company with a $4 trillion market cap and a 75 percent gross margin, and the tenant is a private AI lab that lost roughly $12 billion in the first half of 2026 and is filing to go public into a rate environment that has not blessed a $500 billion IPO in living memory.

Read that sentence twice. This is Nvidia lending its balance sheet to make OpenAI's pre-IPO capital stack financeable. The Bloomberg framing, of a chip vendor becoming a shadow bank, is not a metaphor any more. The 8-K is the license.

The August 10 Frame

Eight days before Ohio signed, Nvidia announced memorandums with six of the largest infrastructure investors on the planet to establish separate compute financing platforms sized to mobilize over $500 billion of third party capital. The six are Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR. Each runs its own platform under its own underwriting, and Nvidia sits alongside rather than on the balance sheet. That is the vehicle. Ohio is the first publicly disclosed use case.

The two announcements read as one architecture. The $500 billion platform finances the silicon. The Ohio guaranty backstops the real estate and power stack. Nvidia takes the equity in the operating counterparty, locks exclusivity across the site, and keeps the cash side of the balance sheet clean. This is not a scrambling response to OpenAI's capital needs. This is a purpose-built rail for financing AI infrastructure at frontier lab scale, designed to route around the constraint that no single hyperscaler is willing to put a half trillion dollar buildout on its own books.

The other frontier labs are not on this rail yet. Anthropic pre-bought its compute inside Google's balance sheet on the $200 billion TPU deal we covered in May, which is a fundamentally different instrument (a supply agreement, not a guaranty). xAI runs its capex through Elon's own leverage. Meta funds out of Meta's operating cash flow. What is emerging in the Nvidia model is the first vertically integrated compute financing stack that operates independently of any single hyperscaler's treasury, which is exactly what you would build if you thought the next $2 trillion of AI capex was going to be bigger than any one hyperscaler is willing to write.

What Repriced

Nvidia's five year CDS was already at a record 82 basis points after the July leak. It moved another few basis points wider on the Monday session, though well inside the July jump. The stock closed roughly flat on the announcement, in contrast to the 6 percent two session drop that followed the leak. The disclosure at $105 billion instead of $250 billion, and the explicit contingent language (payments trigger only on OpenAI default or insolvency), gave the equity desks something concrete to plug into a model, which is usually what calms an equity move faster than a credit move.

But the credit market's repricing of Nvidia is now a settled fact rather than a one day event. Nvidia five year protection costs more than Alphabet's again on the print. A company with $215.9 billion of trailing revenue and 75 percent gross margins is now priced by lenders as if its default probability were meaningfully higher than a company that runs the search monopoly. Nothing about the operating business changed this week. What changed is the set of promises attached to it, and the credit market is telling you that the promises are big enough to matter.

Our Take

The signed number is smaller than the leak. That is a headline. The mechanism is the story. Nvidia has now paired a $500 billion off balance sheet financing platform with a $105 billion contingent guaranty and $1.5 billion of direct equity into the operating counterparty on the first publicly disclosed use case, all inside eight days. The company is not just selling chips into the AI buildout. It is underwriting the buildout's counterparty risk, financing its silicon through third party capital, and taking equity in the landlord that houses the racks. That is a fully vertically integrated compute finance stack, run by the vendor with the biggest gross margin in the industry, and it looks a lot more like how a national infrastructure bank operates than how a chip company usually does.

Second order for OpenAI: the IPO math just got easier. A 20 year lease on 8 gigawatts of Nvidia backed capacity is exactly the kind of contracted supply number a $500 billion filing needs to justify its revenue trajectory, and the contingent structure means OpenAI itself does not have to raise the capital that the campus consumes. We covered the confidential S-1 dynamics last month. The Ohio signing is the compute footnote that turns those dynamics into a plausible story an underwriter can pitch.

Second order for the rest of the industry: every other frontier lab now has to decide whether to compete for a chair on Nvidia's rail or build a rival one. Google has an in-house rail (TPU capex on Alphabet's balance sheet). Amazon has Trainium and a hyperscaler treasury. The interesting question is Anthropic, which sits on Google's TPU compute today but takes AWS Trainium and Nvidia GPUs from multiple sources on the side. If Anthropic ever wants a purpose built campus with Nvidia silicon at the scale OpenAI just secured, the rail exists. The price of admission is exclusivity on the compute side, which Anthropic has historically declined to give any single vendor.

Three signposts to watch. One, whether the option on the remaining 3.8 gigawatts gets exercised inside 12 months (fast exercise means Nvidia is confident in OpenAI's revenue ramp, slow exercise means Nvidia wants a second data point on default probability first). Two, whether the second publicly disclosed use of the $500 billion platform surfaces a non OpenAI lab, and if so which one, because that will tell you whether this rail is a purpose-built OpenAI vehicle or an actual industry-wide instrument. Three, whether Nvidia's next 10-Q discloses the guaranty as a specific line item under contingent liabilities or buries it in footnote aggregate, because the SEC has been asking pointed questions about off balance sheet AI exposure since the spring.

We are tracking the deal cadence on our OpenAI provider page and the Nvidia side on the Nvidia page. Next data point to watch: whether the SB Energy 8-K adds any covenant language around minimum utilization or take-or-pay floors that turn the OpenAI lease from a soft commitment into a hard one. The shape of that language will decide whether the Ohio deal is a lease or, in economic substance, a debt instrument Nvidia just co-signed.