Nvidia guarantees land, power and shell for 4.25 GW Ohio AI campus with OpenAI as tenant
A chipmaker underwriting buildings and power for 20 years shows how gigawatt-scale AI capacity gets financed.
Nvidia said on August 17 that it has partnered with SB Energy to secure land, power and shell capacity at the PORTS-Pike Technology Campus in Portsmouth, Ohio, with OpenAI to operate the facility as a tenant. The announcement, published as a post by chief executive Jensen Huang on the company's blog, describes an initial deployment of 4.25 gigawatts of AI factory capacity, with a further 3.75 gigawatts of capacity remaining at the site.
Under the arrangement Nvidia is guaranteeing what it calls land, power and shell, or LPS, infrastructure for roughly four gigawatts over a term of about 20 years. The guarantee covers defined portions of lease and power payments along with a specified residual-value commitment, and it phases in as the data centres come online between 2028 and 2030.
The post frames the deal in the context of OpenAI's broader commitments, which it puts at roughly 12 gigawatts of Nvidia infrastructure through 2030 with potential expansion to about 16 gigawatts. Nvidia says each generation deployed at that scale represents about 1.5 million GPUs and between $150 billion and $200 billion in revenue, and it describes the total opportunity as roughly $600 billion of Nvidia compute through 2030.
The facility is to run on Nvidia's DSX platform, which the company describes as a full-stack AI factory design spanning GPUs, CPUs, networking and infrastructure software. The post does not give a construction timeline for the first phase, the identity of the power source, or how the campus will be cooled, and it does not disclose the financial terms of the guarantee itself.
Why it matters
A chip vendor guaranteeing a tenant's lease and power payments is a new way of financing AI capacity, and it ties Nvidia's balance sheet to the buildings its chips will sit in. Deals measured in gigawatts rather than megawatts reset expectations for what a single campus is. Southern Ohio also reflects a search for land and power outside the traditional cloud hubs.
A chip vendor guaranteeing a tenant's lease and power payments is a new way of financing AI capacity, and it ties Nvidia's balance sheet to the buildings its chips will sit in.
- Initial Ohio deployment4.25 GW
- Remaining capacity at the site3.75 GW
- OpenAI commitments through 2030~12 GW
- Potential expansion~16 GW
Figures: Nvidia blog post, August 17, 2026
What you can learn from this
Gigawatts are the unit of AI capacity now. A data centre is sized by the electrical power it can deliver to IT equipment, because power determines how many accelerators can run. One gigawatt is a thousand megawatts, and a large cloud facility a few years ago was typically in the tens of megawatts. A 4.25 GW campus is therefore comparable to the output of several large power plants, which is why siting depends on electricity before anything else and why grid operators are now central to the industry. Cooling and power conversion consume part of that budget, so the IT load is always smaller than the headline figure.
Land, power and shell are the long-lead items. Chips can be ordered and delivered in months, but acquiring land, securing grid or generation capacity and building the shell of a facility takes years. The phrase LPS captures the three things that must exist before any server is installed. That is why capacity announcements now come with dates in 2028 and beyond and why guarantees focus on those elements rather than on the equipment inside. Each of the three has its own permitting process and its own set of counterparties.
A guarantee shifts risk to the guarantor. When a supplier promises to cover portions of lease and power payments if the tenant does not, lenders and developers can finance construction with more confidence. The trade-off is that the guarantor now carries exposure if demand fails to materialise. Understanding who bears the risk in an infrastructure deal tells you who has the most to lose if forecasts miss, and it explains why such commitments are usually phased in rather than granted all at once.
Full-stack platforms bundle chips with the building. DSX-style reference designs specify not just the accelerator but the CPUs, networking, power distribution and software that go around it. Bundling reduces integration risk for the operator and ties the design to one vendor's roadmap. It behaves like a standardised blueprint that can be copied across sites with fewer surprises, at the cost of flexibility to mix suppliers.
Generational replacement drives the revenue maths. Nvidia's per-generation figures assume that a site refills with new GPUs roughly every product cycle. Accelerators are replaced faster than buildings, so a 20-year shell may host several successive fleets of chips, which is why the vendor counts revenue per generation rather than per site. The residual-value commitment in the guarantee is a way of pricing what the building is worth after the first fleet retires.
We teach this
How to use this in practice
Convert a power figure into a rough accelerator count. Take a published campus power number, assume a working figure for power per rack and per accelerator from any vendor specification sheet you can find, and compute an order-of-magnitude count. Compare your answer with the 1.5 million GPUs per generation figure in the post and think about what overhead (cooling, networking, storage) explains the gap. Done looks like a short worked calculation with your assumptions listed and the two assumptions that move the answer most marked. Round to one significant figure; precision is not the point.
Draw the LPS timeline for a hypothetical site. On a single line from today to 2030, mark land acquisition, power agreement, shell construction, fit-out and first GPU generation, then the second generation. Add the approvals and grid studies that sit between steps, and mark the point at which the first payment under a guarantee like this one would start, noting how far it sits from today. Done looks like a diagram that makes clear which items run in parallel and which block the rest.
Read one utility interconnection queue. Most grid operators publish a list of requested connections. Find the queue for a region you know, filter for data centre or large-load entries and note the requested megawatts, requested dates and current status. Done looks like a note with three entries, their sizes and how long each has been waiting. If the operator publishes withdrawal figures, note how many requests dropped out in the last year.
Model the guarantee in a spreadsheet. Set up a simple table with annual lease and power payments over 20 years, a tenant-pays column and a guarantor-pays column, and a toggle for tenant default in a given year. Add a residual-value row at the end of the term. Done looks like a sheet where flipping the toggle shows the guarantor's exposure for the remaining term and how it changes with the default year. Then add a second tenant and see whether spreading the risk changes the picture.
Sources
- Securing the Infrastructure of Intelligence — NVIDIA Blog
Our reporting is an original summary; full coverage is at the links above.
Don't just read about it — build it.
Square 1 teaches the skills behind the headlines, with every line of your work graded by AI. Find your starting point in 3 minutes.
Get your free skill report