OpenAI Negotiates 20-Year Long-Term Lease to Lock in Prices, Nvidia Bets on Ohio 10GW Data Center with "Lease First, Rent Later" Model

OpenAI is negotiating a 20-year ultra-long-term leasing agreement, aiming to lock in a 10GW AI data center campus in Ohio. NVIDIA is simultaneously providing credit support, forming a "lease-to-loan" closed-loop transaction model. If successful, this will be OpenAI’s largest infrastructure investment commitment to date.
(Background: OpenAI and NVIDIA announced a strategic partnership)
(Additional context: OpenAI has submitted an IPO application! Following Anthropic and SpaceX, making a move on Wall Street)

Table of Contents

Toggle

• What does NVIDIA’s “credit support” really mean?
• From a $10 billion intention to $3 billion equity
• What does a 20-year lease represent?
• Observations from Taiwan’s perspective
• Where are the risks?

OpenAI, a pioneer in artificial intelligence, is negotiating a 20-year ultra-long-term lease agreement, targeting a 10 GW AI data center campus on federal land in Ohio.

According to a report by The Information on June 10, two key insiders involved in the negotiations revealed that OpenAI has entered the “late-stage negotiations.” If the deal is finalized, it will be OpenAI’s largest infrastructure investment commitment ever.

What does NVIDIA’s “credit support” really mean?

This isn’t just about “signing a contract and paying rent.”

NVIDIA has discussed providing credit support for the project — in other words, NVIDIA might front the capital or provide guarantees, allowing OpenAI to lease the data center under more favorable terms. The chips running inside the data center will, of course, be NVIDIA’s own GPUs.

This model is called “financial engineering” by Tom’s Hardware: NVIDIA funds the construction, OpenAI rents NVIDIA chips, forming a “funding → leasing → recoupment” closed loop.

From a $10 billion intention to $3 billion equity

This isn’t the first collaboration between OpenAI and NVIDIA.

In February this year, the two finalized a $3 billion equity investment (shrinking from the initial $10 billion intention), and announced a strategic partnership: NVIDIA will invest up to $10 billion in phases to deploy a 10 GW AI data center, with the first phase expected to go online in the second half of 2026.

OpenAI also signed a $30 billion cloud computing agreement with Oracle, with data centers under construction in Abilene, Tennessee, and Michigan. If Ohio’s 10 GW campus joins the race, OpenAI’s global AI computing power map will expand further.

What does a 20-year lease represent?

20 years is a long time. For comparison, the generation cycle of AI chips is about 2-3 years, and the Vera Rubin series chips have just entered mass production. In 20 years, OpenAI might be running a hybrid cluster of 6-8 different chip generations simultaneously.

But a super-long lease also locks in an advantage: the electricity costs in Ohio.

Ohio is one of the states with the lowest electricity prices in the Midwest, at about $40-50 per megawatt-hour, far below California’s over $80 per MWh. At a scale of 10 GW, the annual electricity cost savings could amount to tens of millions of dollars. A 20-year lease effectively locks in this long-term cost advantage.

Perspectives from Taiwan

NVIDIA’s “investment → leasing” strategy is increasingly similar to Taiwan’s wafer foundry “wafer leasing” model.

Taiwan’s leading AI chip manufacturer TSMC achieved NT$431.8 billion in revenue in Q1 2026, up 73% year-over-year. NVIDIA’s Vera Rubin chips are all produced on TSMC’s 3nm process. Every GPU rented by OpenAI is backed by Taiwan wafer production.

If Ohio’s 10 GW data center fully adopts the latest Vera Rubin Superchip, estimating about 300,000 to 400k GPUs per GW, then a 10 GW scale would mean 3-4 million chips — equivalent to about 3-4 months of TSMC’s AI chip capacity.

Where are the risks?

The cost of a super-long lease is flexibility.

If AI inference demand continues to grow as expected, a 20-year lease is a reasonable bet. But if AI computing power faces a “superbust” adjustment similar to early 2026 (a risk recently warned by Michael Burry), OpenAI might face idle capacity.

More immediate pressure comes from competition. Nebius announced a 310 MW data center project in March, CoreWeave completed a new round of funding in April, and UK startup Nscale secured $2 billion in March. The “arms race” for AI data centers is extending from the East Coast to the Midwest.

This article is sourced from The Information and Tom’s Hardware, translated and compiled by Dongqu editor Flip.