Nvidia Just Guaranteed $105 Billion of OpenAI's Rent. OpenAI Promised to Pay It Back.

Jensen Huang published about a thousand words on LinkedIn this morning explaining why Nvidia has decided to buy its way into the land, power and shell business. The post, Securing the Infrastructure of Intelligence, sets out the logic behind the Ohio campus announced alongside it, works through an FAQ, and disposes of the circular financing question in six words: "No. OpenAI will pay the lease."

It does not mention the number.

The number sits in the Form 8-K that CFO Colette Kress signed the same day. Nvidia has entered into multiple residual value guaranties with SB Energy covering roughly 4.25 gigawatts of IT load at the Portsmouth site in Pike County, Ohio, plus an option, exercisable at Nvidia's sole discretion, to extend credit support to a further 3.8 gigawatts. Its aggregate payment obligation on the initial commitment is capped cumulatively at $105 billion. The company filed the disclosure under Item 2.03, which covers the creation of a direct financial obligation or an obligation under an off-balance-sheet arrangement.

The cap deserves scrutiny on its own terms. But the provision worth reading twice appears near the end of the disclosure, and several outlets have already reproduced it today as grounds for calm.

OpenAI, the filing says, has agreed to "reimburse and indemnify NVIDIA for any and all amounts actually paid."

Now check what makes Nvidia pay in the first place. The filing defines two trigger events: OpenAI's insolvency resulting in a default under a lease, or OpenAI's failure to make payments under a lease.

Read together, the two provisions describe a loop. If OpenAI cannot cover the rent, Nvidia covers it. OpenAI, having just failed to cover the rent, then reimburses Nvidia. The indemnity only ever comes due in the scenario that establishes it cannot be collected.

What the instrument actually is

The loop is not the interesting part. Three clauses further down is a better tell.

Nvidia's obligations terminate, per the filing, on the earliest of four events: the twentieth anniversary of each lease commencement, OpenAI terminating the lease under its own terms, other customary termination events, and OpenAI achieving a satisfactory credit rating.

A guarantee that lapses the moment your tenant becomes creditworthy is not primarily a hedge on real estate. It functions as a substitute credit rating with a twenty year outside date, and Nvidia has written it in Ohio square footage.

Huang argues something close to this himself, in what is easily the most candid passage of the post. Frontier labs, he writes, have extraordinary demand for training and inference compute but are growing faster than their balance sheets and long-term credit profiles can support. They have customers and revenue without the investment-grade financing capacity to secure AI factory infrastructure on their own. That is a precise diagnosis of the situation, and it is the situation. OpenAI remains unprofitable at a valuation of roughly $852 billion, according to Wall Street Journal reporting summarised by Reuters, and carries no investment-grade rating. SB Energy needs to raise debt against somebody's credit to build eight gigawatts of capacity on a decommissioned uranium enrichment site.

Nvidia is that somebody. Which is why the FAQ answer, true as it is, lands slightly off target. Nobody was confused about whose name goes on the rent cheque. The circularity question has always been about whose credit makes the lease financeable at all, and on that the 8-K is unambiguous.

The exposure line moved roughly thirtyfold in one filing

Six days ago, Morgan Stanley's Joseph Moore told clients that for all the noise about circularity, Nvidia's actual direct credit exposure was "mostly confined to credit backstops with a couple of smaller neoclouds," as reported by CNBC. It was a fair reading of what had been disclosed. Nvidia's Q1 FY2027 10-Q capped maximum gross exposure across every partner facility lease guarantee the company had ever signed at $3.5 billion, with $712 million held in escrow behind it, per Tom's Hardware. Nvidia took those guarantees in exchange for warrants, carried them as credit derivatives, and described their fair value as immaterial. The first of them, disclosed in Q3 FY2026, was capped at $860 million.

That note had a six day shelf life. Every model of Nvidia's contingent liabilities is now stale, nine days before earnings.

The recovery assumption is the whole structure

Huang does not dodge the difficult question. He raises it himself: what happens to PORTS-Pike if OpenAI stops using it?

His answer is the intellectual core of the deal and a genuinely strong argument. The chain runs from CUDA to versatility, from versatility to fungibility, and from fungibility to durable utilization, ending at compute as a productive asset that is rentable and financeable. Capacity could be re-let across Nvidia's ecosystem of clouds, enterprises, labs and startups. An exceptional site, on this reading, outlives any single tenant or generation of silicon.

Each link is plausible, and the whole chain is load bearing. Under the agreements, once a trigger event occurs Nvidia pays the shortfall between the lease's guaranteed minimum value and whatever gets recovered through a replacement lease or a sale. The $105 billion is a ceiling on what Nvidia owes after the market has had its chance to absorb the site. If re-letting works, exposure collapses toward zero. If it does not, it does not.

So the question is not whether AI data centres are fungible in the abstract. It is whether this site is fungible in the specific scenario that triggers the guarantee, and that scenario is OpenAI's insolvency.

Consider what the world looks like in that state. The most prominent frontier lab has failed to pay rent on its flagship campus. Demand from the frontier lab segment is, at best, an open question. Credit markets that spent 2026 underwriting gigawatt-scale AI construction have repriced. And Nvidia needs a tenant prepared to sign a twenty year lease on 4.25 gigawatts in rural southern Ohio at something near the guaranteed minimum value.

The recovery assumption, in other words, is negatively correlated with the event that triggers the payout. That matters more than the headline cap, and nobody can price it, because no purpose-built multi-gigawatt AI campus has ever come up for re-letting. There is no comparable transaction anywhere in the market.

$105 billion, received as good news

The market reaction may be the most revealing thing about today.

Three weeks ago the Journal reported that Nvidia was weighing a guarantee of roughly $250 billion for this site, with separate talks on as much as $350 billion to finance the accelerators going inside it, again per Tom's Hardware. On August 10, Nvidia announced memoranda with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilise more than $500 billion of third-party capital, an explicit effort to shift the financing burden off its own balance sheet. The shares fell 2.9% that day anyway. Last Friday the Journal reported the guarantee had been cut to under $120 billion, covering only the project's first phase, after investors raised concerns about the risk Nvidia was absorbing.

Today it landed at $105 billion and the stock is trading near a record high, around $225.

A nine figure off-balance-sheet obligation was received as de-risking, because the anchor had been set at $250 billion. In under a year, the market's reference point for tolerable vendor credit exposure has travelled from $860 million to a $105 billion relief rally.

The case for the deal

It is a strong one and deserves stating properly.

Nvidia has attached roughly $105 billion of capped, phased, conditional exposure, offset by recovery proceeds, to a stated $600 billion revenue opportunity through 2030. Huang puts OpenAI's existing and planned commitments at about 12 gigawatts of Nvidia compute, rising to roughly 16 if Nvidia exercises the Portsmouth option, and estimates each generation of systems at the site at approximately 1.5 million GPUs, or $150 billion to $200 billion of revenue. The obligation does not begin until leases commence from 2028. It declines as OpenAI pays rent. It dies outright if OpenAI earns a rating. It triggers only on insolvency or non-payment. And on default Nvidia retains real optionality: assume the lease, force a re-let, run a sale process, allow termination, or defer for up to a year while covering specified project costs.

If compute demand compounds and OpenAI survives, this is an excellent trade, and Huang's supply chain analogy holds up. Nvidia has spent a decade pre-buying scarce inputs on visibility its competitors lacked and has been right almost every time. Bank of America's Vivek Arya read the August 10 consortium as a pivot away from vendor financing, with the capital burden sitting with the banks rather than Nvidia's balance sheet. Wells Fargo's traders read the same announcement as Nvidia remaining part of the financing regardless. Both readings appear in the same CNBC piece. Today's filing is evidence for the second.

Four things still undisclosed

The guaranteed minimum value schedule is not in the filing, and that omission matters more than anything else here. Those figures, not the $105 billion cap, determine Nvidia's real economic exposure. The cap describes the worst case; the schedule would describe the expected one.

Whether the cap sits gross or net of re-let and sale proceeds is also unstated, as is how Nvidia intends to account for the obligation. The earlier backstops were carried as credit derivatives at immaterial fair value, a treatment a $105 billion commitment may not survive.

What Nvidia received in return is unclear. Previous guarantees came attached to warrants. The 8-K says nothing on the point, though Nvidia is separately investing $1.5 billion of equity into SB Energy alongside SoftBank and OpenAI, according to the press release.

And the exclusivity has holes in it. OpenAI will deploy Nvidia's full-stack DSX platform at the site "subject to limited exceptions," per the 8-K, which does not say what those exceptions are.

One timing note should shape coverage over the next ten days. Nvidia states that the form of the agreements will be filed as an exhibit to the 10-Q for the quarter ended July 26. That quarter's earnings call is scheduled for August 26. Analysts get their questions in before anyone outside the parties has read the contracts.

The argument that has to carry it

The deal may well work out. Compute demand may compound exactly as Huang expects, the Portsmouth campus may host four generations of Nvidia systems on land the Department of Energy spent decades decommissioning, and the guarantee may expire unexercised the day OpenAI prices an IPO and earns a rating.

But the case has to rest on the fungibility argument, which is at least testable and which Huang clearly believes. It cannot rest on a promise of repayment from a company that would have to be insolvent to owe it.