Nvidia said on Monday it will guarantee up to $105 billion in financing for an OpenAI data center in southern Ohio, which is a sentence worth reading twice. The world’s most valuable chipmaker is not selling GPUs into this deal so much as underwriting the credit of the tenant who will buy them. That is a different business, with a different risk profile, and Wall Street has spent three weeks arguing about whether it is a moat or a warning.
What Nvidia Signed
The project sits at the PORTS-Pike Technology Campus in Pike County, Ohio, on the grounds of a decommissioned uranium enrichment site. CNBC reported the financing commitment alongside the operating structure, and UPI laid out the counterparties. The terms that matter:
- Nvidia guarantees up to $105 billion, covering certain lease and power payments and backstopping the value of the infrastructure if OpenAI defaults.
- SB Energy builds and operates it. SB Energy is a SoftBank subsidiary, which puts Masayoshi Son on both the construction side and, through SoftBank’s OpenAI stake, the demand side.
- OpenAI signs a 20-year lease as anchor tenant, the longest-dated commitment the company has made to anything.
- Capacity phases in from 2028, funding an initial 4.25 gigawatts with an option for 3.75 more, reaching eight gigawatts.
- Nvidia puts $1.5 billion of equity into SB Energy, and SB Energy plus SoftBank commit at least $4.2 billion to regional grid infrastructure supporting 10 gigawatts.
Read that list as a capital stack rather than a press release and the shape is clear. Nvidia is not the vendor here. It is the credit enhancement that makes the whole structure financeable, the way a monoline insurer once made municipal paper investment grade. Take Nvidia’s guarantee out and a 20-year lease to a company that has never posted an annual profit does not attract project finance at any sane rate.
The Number Came Down, and That Is the Story
The interesting figure is not $105 billion. It is the distance between $105 billion and what was originally floated.
In late July, CNBC reported the two companies in talks over a backstop worth as much as $250 billion. Nvidia stock fell about 4.5% intraday on the news. Investors were not celebrating a bigger deal; they were pricing the risk that Nvidia had quietly become a financing company. Axios described the reaction as reigniting the circular-financing concern that has shadowed the AI trade all year.
Nvidia listened. The guarantee was cut to under $120 billion, then settled at $105 billion by signing. Fortune framed the final structure as landing $145 billion below what had been reported, and read the shrinkage as an admission that some portion of AI chip demand is manufactured by the chipmaker’s own balance sheet.
That is the structural why, and it is worth stating plainly: the size of the guarantee got negotiated down not by the counterparty, but by Nvidia’s own shareholders. The equity market set the ceiling on how much vendor financing it would tolerate before it started marking down the multiple. Jensen Huang did not discover restraint. He got a price signal.
When the Supplier Becomes the Lender
Circular financing is a loaded phrase, so define it precisely. The concern is not that Nvidia invests in customers. It is that revenue booked as chip sales may be partly funded by Nvidia’s own capital, which makes demand look more organic than it is and flatters the growth rate that supports the valuation. Jim Cramer warned in July that the pattern echoes the vendor-financing arrangements that preceded the dot-com crash, when Lucent and Nortel lent telecom startups the money to buy Lucent and Nortel equipment, then wrote off the receivables when the startups folded.
The comparison is imperfect in Nvidia’s favor, and honesty requires saying so. A guarantee is not a loan. Nvidia is not booking a receivable from OpenAI; it is promising to make lease and power payments whole if OpenAI cannot, and it holds a claim on physical infrastructure at a site with grid interconnection that takes years to replicate. If OpenAI fails, Nvidia inherits eight gigawatts of powered data center in a market where power, not silicon, is the binding constraint. That is a materially better collateral position than Lucent ever had.
But the accounting cleanliness is exactly what makes the exposure easy to underrate. We covered Nvidia’s broader $500 billion push to organize Wall Street financing around AI infrastructure earlier this month, and this Ohio commitment is one node in that network rather than a standalone bet. Stack the guarantees, the equity stakes, the debut $20 billion bond sale that funded some of it, and Nvidia’s real balance sheet risk is distributed across a dozen agreements that are individually defensible and collectively unmapped. CNBC’s read that the company’s moat is shifting from chips to capital is right, and a capital moat is a leverage moat.
There is a timing problem underneath all of it. This campus energizes in 2028 against a 20-year lease, which means the deal was underwritten on assumptions about the cost of capital and the useful life of GPUs that are both moving right now. The useful-life question is not academic: it is the same depreciation argument that ran through CoreWeave’s second-quarter backlog disclosure, where how long a GPU earns determines whether the contract is profitable or a writedown with a long fuse.
Nvidia has built something genuinely hard to copy: not a chip advantage, which Google and Amazon are chipping at with their own silicon, but the ability to make an eight-gigawatt project bankable by signing its name. The question nobody has answered is what that signature is worth in a year when the market decides AI capex was financed on assumptions that no longer hold. Guarantees are cheap to issue and expensive to honor, and they tend to come due together.