The aggressive expansion of artificial intelligence infrastructure has pushed frontier labs far beyond the capital limits of their corporate balance sheets. Because foundational model training demands astronomical physical resources, securing long-term grid capacity now requires financial acrobatics that bypass standard disclosure. In the largest infrastructure transaction in AI history, OpenAI has signed a 20-year lease for approximately 8 gigawatts of capacity at the PORTS-Pike campus in Ohio, developed by SoftBank subsidiary SB Energy.
The Anatomy of the Ohio Megaproject
The physical scale of the Ohio buildout underscores how drastically datacenter economics have shifted. As reported by The Wall Street Journal, gross capacity reaches 10 gigawatts once cooling and transmission overhead are included. Situated on land that encompasses a former US Department of Energy uranium enrichment plant, the site draws baseload electricity from an existing 9.2-gigawatt gas facility owned by the US government and financed by Japan under a bilateral trade pact. Commercial deployment begins in 2028 with an initial 800 megawatts, structured so OpenAI only pays as operational capacity goes live.
To bridge the solvency gap between OpenAI's startup balance sheet and this titanic capital outlay, Nvidia stepped in as the financial linchpin. Rather than underwriting OpenAI's lease payments directly, Nvidia guarantees the residual asset value of the completed facilities across the initial 4.25-gigawatt buildout phase, capping its downside exposure at $105 billion. If OpenAI defaults, SB Energy must first attempt to secure an alternative tenant and liquidate the physical assets before Nvidia covers any remaining valuation deficit. In return for absorbing this catastrophic downside risk, Nvidia locks in exclusive silicon supply rights for the first half of the campus and is injecting $1.5 billion in equity directly into SB Energy.
Nvidia CEO Jensen Huang captured the raw operational calculus behind this triangular architecture during the project announcement.
"LPS stands for land, power, and shell, meaning the site, the electricity supply, and the building itself."
According to Huang, these physical constraints have officially superseded silicon and networking hardware as the primary bottleneck in the AI race. Huang estimates that every system generation deployed at this scale will consume roughly 1.5 million GPUs, generating between $150 billion and $200 billion in direct hardware revenue. Across all active sites, Huang projects OpenAI's cumulative commitments will demand roughly 12 gigawatts of Nvidia compute by 2030, swelling to 16 gigawatts worth an estimated $600 billion if Nvidia exercises its option over the remaining 3.75 gigawatts in Ohio.
The Three-Trillion-Dollar Off-Balance-Sheet Expansion
This triangular pact illustrates a systemic accounting shift across the tech sector. A Wall Street Journal investigation revealed that major tech incumbents now harbor roughly $3 trillion in off-balance-sheet AI commitments, creating an opaque financial web where institutional investors can no longer accurately measure corporate debt burdens. By monopolizing land, power, and shell capacity for decades, a cartel of interconnected hyperscalers is actively bifurcating the market: elite players control closed megawatt clusters, while the broader enterprise tier is forced into running compact, efficiency-tuned small models on rented margins.