Nvidia Is Guaranteeing 105 Billion Dollars For One OpenAI Data Centre

One hundred and five billion dollars. One building. Nvidia has agreed to back up to that much financing for a single data centre campus in Ohio, and OpenAI has signed a twenty year lease on it before the concrete is poured.

What the Nvidia OpenAI Ohio data center deal actually contains

The site is called the PORTS-Pike Technology Campus. SB Energy, the SoftBank unit, is building and will own it. Nvidia is not the landlord and not the tenant; it is the credit behind the whole thing, and it has also put 1.5 billion dollars directly into SB Energy on top of the financing guarantee.

The first phase calls for 4.25 gigawatts of compute, with an option to expand to eight. The first 800 megawatts is due online in 2028.

Eight gigawatts is roughly the output of eight large nuclear reactors. Committed to one tenant, in one state, for one workload.

The part that makes it interesting

OpenAI has committed to using Nvidia chips exclusively on the campus.

So Nvidia guarantees the money that buys the land and the power, and the tenant contractually promises to spend that money on Nvidia silicon. The financing and the revenue are the same dollars travelling in a circle.

Jensen Huang has pushed back hard on that framing. His argument is that the compute opportunity in front of Nvidia is worth around 600 billion dollars, and that underwriting a customer's build is simply how you service demand you can already see. Vendor financing is not new. Lucent did it for telecoms carriers in the late nineties, and Cisco ran a version of it too.

Whether you accept the framing depends on what you think happens if OpenAI's revenue curve flattens. In a normal build the risk sits with a bank, a landlord, a utility and a tenant, spread across four balance sheets. Here it concentrates. We looked at how much of OpenAI's own future is priced into a similar set of assumptions in our breakdown of OpenAI's IPO valuation and S-1 filing.

Why the power number matters more than the dollar number

The 105 billion is the headline, but gigawatts are the real constraint. You cannot buy your way past a grid interconnection queue, and you cannot ship a data centre to a place that has no spare generation.

That is why the campus sits in Ohio next to power rather than in Virginia next to users. It is also why chip roadmaps are increasingly written around performance per watt rather than raw throughput, something visible in Google's Frozen v2 TPU successor and in every Blackwell generation slide Nvidia has shown this year.

What to watch

Three things will tell you whether this was underwriting or over-extension.

First, whether the 2028 date holds, because interconnection delays are the norm rather than the exception. Second, whether other hyperscalers follow with their own vendor-guaranteed builds, which would confirm this is now the standard structure rather than a one-off. Third, whether OpenAI's consumer and enterprise revenue keeps pace with the compute it has just contracted for, a question that also hangs over Gemini's billion-user scale on the other side of the market.

AI infrastructure just stopped being a line on a capex sheet and became a financing product. That is the shift worth tracking, whatever the campus ends up costing.