Nvidia Backstops OpenAI's Ohio Lease With Up to $105B

The most interesting term in OpenAI's new Ohio lease has nothing to do with capacity. It concerns who absorbs the loss if the site ever empties out.

The Wall Street Journal reported on August 17 that OpenAI signed a 20-year lease for the PORTS-Pike campus in Ohio, and that Nvidia is guaranteeing the residual value of the finished data centers up to $105 billion.

The shape of the deal

The campus is going up on the grounds of a former US Department of Energy uranium enrichment site. OpenAI gets 8 gigawatts of IT capacity; with cooling and supporting infrastructure the gross figure reaches 10 gigawatts. Power comes from a 9.2-gigawatt gas plant owned by the US government and financed by Japan.

Phase one delivers 4.25 gigawatts of IT capacity by 2028, starting at 800 megawatts. The developer and landlord is SB Energy, a SoftBank subsidiary. Nvidia is the exclusive chip supplier for the first half of the site and is separately investing $1.5 billion in SB Energy.

Why Nvidia is signing as guarantor

The guarantee is narrower than the headline number suggests. It covers phase one only, and Nvidia pays out solely if OpenAI walks away and SB Energy cannot find a replacement tenant or sell the facilities. It functions as an insurance policy rather than a cheque.

Jensen Huang's stated reasoning is that the constraint has moved. Chips and networking gear are no longer the bottleneck, he argues; "LPS" is, meaning land, power, and shell. Before Nvidia can sell more chips, somebody has to finance the buildings those chips go into, so Nvidia is using its own balance sheet to clear the blockage in front of its own demand.

The $3 trillion nobody's balance sheet shows

The quieter half of the same reporting deserves more attention than it got. By the Journal's count, nine major technology companies carry roughly $3 trillion in AI-related commitments off balance sheet: leases and purchase agreements that are not recorded until payment or delivery begins.

None of this breaks an accounting rule. It does mean that anyone trying to size the industry's real obligations has to look well past the investment figures in the press releases.

How gigawatts reach your invoice

Gigawatts and hundred-billion-dollar guarantees feel remote from a company running a support chatbot or a document pipeline. The connection is direct. Token prices are built on the depreciation and power costs of exactly these facilities, and today's prices are a stage in an investment cycle rather than a settled market rate.

Two practical consequences follow. First, do not model your business on today's unit price holding. Measure cost per transaction and know now which features stop making money if that price doubles. Second, capacity of this scale arriving in 2028 improves the odds that latency-tolerant work gets cheaper. Anything that can run overnight, such as batch document processing or archive analysis, is worth architecting so it can move to a cheaper tier later.

And the unavoidable question: is this a bubble? The honest answer is that nobody knows, and the size of the capital commitment proves neither that the technology works nor that it doesn't. Your own measure should be the return on the automation you have running, not the gigawatt count in Ohio.

Sources: The Decoder, TechCrunch, Wall Street Journal