NVIDIA Backs Ohio Site With 20-Year Guarantee, Supplies OpenAI 4.25GW

NVIDIA described this as securing "LPS" — shorthand for Land, Power, and Shell, the three resources that are as hard to come by as semiconductors when building AI data centers. Huang wrote that "in the AI economy, compute is revenue," and said NVIDIA will now stake out LPS the same way it has managed semiconductor, packaging, and memory supply chains.

Why treat this differently from hyperscalers

Large cloud companies like Google and Microsoft secure sites and power and build data centers using their own balance sheets. According to NVIDIA's post, this approach still accounts for most of the company's revenue. The problem lies with frontier AI labs like OpenAI, whose revenue is growing rapidly. NVIDIA explains that while these labs show strong customer demand and revenue growth, many of them lack the investment-grade credit needed to back decades-long infrastructure contracts. As a result, the bottleneck to growth has shifted — not from algorithms or customer demand, but from securing compute itself.

PortsPike, by the numbers

The initial deployment capacity is 4.25 gigawatts. NVIDIA said OpenAI will build and operate its own AI factory at the site, using NVIDIA's "DSX" full-stack platform spanning GPUs, CPUs, networking, and software.

Item	Value
Initial deployment capacity	4.25GW
Additional expansion room	3.75GW (8GW total)
GPUs per generation	~1.5 million units
Estimated revenue per generation	$150–200 billion
OpenAI's total commitment	~12GW through 2030
Total commitment if expanded	Up to 16GW
Total opportunity through 2030	~$600 billion

NVIDIA explained that the site itself is a vessel capable of swapping in multiple generations of compute over 20 years. As the current generation of GPUs ages, it will be replaced with the next generation, and then the one after that. If the site expands by another 3.75 gigawatts, OpenAI's total NVIDIA compute commitment would grow from 12 gigawatts to 16 gigawatts, and the associated opportunity would grow to roughly $600 billion through 2030.

On the question of circular financing

The fact that NVIDIA is guaranteeing part of the lease and power costs, along with a set residual value, for its own customer OpenAI's site immediately raised questions of circular financing. NVIDIA addressed this directly in its blog post: "No. OpenAI will pay the lease." The guarantee takes effect in stages between 2028 and 2030, as the data center actually comes online, and NVIDIA's remaining exposure decreases as OpenAI pays rent and the facility is put into operation. NVIDIA also added that even if OpenAI stops using the site in the future, the CUDA ecosystem means the asset can be resold to other cloud companies or startups, making it a rotatable asset.

Reports emerged on August 15 that NVIDIA had reduced its guarantee for OpenAI's Ohio data center from $250 billion to under $120 billion. Since this PortsPike deal is also in Ohio, it can be read as an extension of a broader effort to redesign guarantee structures under investor pressure. However, this release alone does not clarify exactly how the per-generation revenue estimates and 20-year guarantee terms announced here line up with the figures reported before and after that earlier reduction.

Editor's take

This isn't the first time NVIDIA has ventured into real estate deals, but taking on the role of a 20-year guarantor itself is unusual. It's as if the company selling picks and shovels in a gold rush is now guaranteeing the lease on the mine site itself. The reason this sounds strange is clear — a substantial share of NVIDIA's revenue comes from large spending by customers like OpenAI, and if NVIDIA is also backing the infrastructure that enables that spending, it becomes hard to tell who is really bearing the risk. NVIDIA's answer was clear: OpenAI pays the rent, and NVIDIA is betting that the site and compute can be sold to other customers if needed. For that logic to hold, one precondition is required: the CUDA ecosystem must be strong enough that GPU assets are genuinely "sellable."

Compared with other AI factory news from around the same time, this feels different. Deals like Firebird building roughly 70,000 GPUs' worth of capacity in Armenia, or Together AI bringing in 10,000 B300s in India, were all structured as "customers building with their own money." PortsPike is different. This is the first time NVIDIA has directly taken on part of the site risk itself — and it did so just two days after reports emerged that it had halved its guarantee in that very same state of Ohio. This simultaneous pattern of shrinking one guarantee while rolling out a new one should be read as a sign that NVIDIA is carefully redesigning its exposure to its single largest customer, OpenAI.

For domestic companies, the practical implications of this news are limited. Still, for any company considering AI infrastructure investment, it's worth noting that infrastructure guarantee structures can become this complex when contracting with frontier AI labs rather than major cloud providers. In the coming weeks, the key thing to watch is how credit rating agencies and investors actually respond to this guarantee. How this deal is referenced in NVIDIA's stock performance or in OpenAI's next funding round will reveal whether this is truly a "strategic and prudent" selection — or an unavoidable choice made to prop up its biggest customer.