Nvidia Backs $105B for OpenAI's Ohio Data Center Campus

TL;DR: Nvidia has agreed to guarantee up to $105 billion in lease and power obligations for a 10-gigawatt OpenAI data center campus in Ohio. That turns Nvidia from chip vendor into financial backstop, and puts real numbers on how expensive frontier AI infrastructure is becoming.

Financing structure

The headline number is not a normal venture round or capex line item. Nvidia is guaranteeing up to $105 billion tied to lease and power obligations for a 20-year OpenAI lease at the former PORTS-Pike Technology Campus in Pike County, Ohio. Separately, Nvidia is investing $1.5 billion in SB Energy, the SoftBank subsidiary that will build, own, and operate the site.

That matters because it changes the role Nvidia plays in the stack. This is not just selling GPUs into demand that already exists. Nvidia is helping underwrite the physical and contractual basis for that demand to exist at all.

For OpenAI, the benefit is obvious: lower friction in securing a giant site and the electricity needed to make it useful. For Nvidia, the upside is also clear: if OpenAI scales as expected, the chips, networking, and surrounding infrastructure likely flow back toward Nvidia anyway. The risk is that Nvidia is now exposed not only to hardware cycles, but also to occupancy, utilization, and power economics.

Campus scale

The planned campus is enormous even by current hyperscale standards. The site is designed for 10 gigawatts. The first 800 megawatts are expected online by 2028.

A 10-gigawatt figure is large enough that the usual "data center expansion" framing starts to break down. This is closer to an industrial infrastructure project with an AI tenant attached. The total project cost including chips, transmission, cooling, and land is expected to exceed $500 billion, which gives a better sense of the total system cost.

The site will be built on a decommissioned uranium-enrichment property roughly 50 miles south of Columbus. That detail is notable mainly because projects of this size need unusual sites: lots of land, heavy power access, and a path through permitting. SB Energy says the project will support 35,000 construction jobs through 2032 and 2,500 long-term positions.

Strategic interdependence

This deal deepens a relationship that was already central to the AI market. OpenAI needs huge amounts of compute to train and serve newer models. Nvidia benefits when the most compute-hungry labs commit to capacity on a massive time horizon.

The difference here is the level of mutual dependence. If Nvidia guarantees the obligations that underpin OpenAI's campus, it is no longer just a supplier waiting for orders. It has a direct financial interest in OpenAI's infrastructure actually getting built and used. That creates tighter alignment, but also tighter coupling.

For the broader market, this is another sign that frontier AI is moving away from a pure software story. Model capability now depends on financing structures, utility negotiations, land control, long-term leases, and political approval. The winners may be the companies that can coordinate all five, not just train the best model.

Power constraints

The project also sharpens the main bottleneck in US AI infrastructure: electricity.

Ten gigawatts is not a side note. It is roughly equivalent to the output of ten large nuclear reactors, or about 7–8 million homes worth of continuous demand. Connecting that to the grid requires multi-year transmission upgrades and interconnection studies that run on utility timelines, not startup timelines. Even with on-site generation (SB Energy is a solar and battery developer), the power economics will dominate the project's actual cost schedule.

For data center operators, the bottleneck is now physical infrastructure, not silicon. Power and grid access are the binding constraints. The chips exist; the electrons to run them are the scarce input.

What this means for builders

1. Compute will tighten before 2028. The first 800 MW is years away, but the commitment will draw capital to the AI infrastructure story and lock up long-dated capacity. Expect hyperscaler pricing to firm in 2027.
2. The Nvidia-as-financier model is the new normal. Balance-sheet support tied to a customer's lease and power obligations is a new structure. Other chip vendors (AMD, custom silicon from Google/Amazon) will be pressured to match.
3. Power is the strategic moat now. Frontier AI is no longer a software race. It is a power, land, and permitting race. Builders should plan around compute cost volatility into 2028.
4. Concentration risk is real. OpenAI occupies a single huge site. Single-site concentration is a real failure mode for any one tenant — diversification matters.
5. Builder action: optimize architectures now. If you are training or serving models, lock in compute pricing where you can. The 2027–2028 window will look very different from 2026.