OpenAI’s 10GW Ohio Lease Comes Before the Power

OpenAI has signed a 20-year agreement for a 10GW Ohio campus with SB Energy. The campus still needs the power supply that would make those gigawatts usable. A customer can commit to decades of computing before the site can deliver its first unit of capacity.

That is the contradiction at the center of AI data centers. Developers cannot wait for demonstrated utilization before building power systems, cooling plants, and server halls; customers cannot wait for all of those assets to exist before reserving capacity. Both sides have to act against a future neither can fully verify.

AI data center project finance converts that expected demand into construction capital through long leases, supplier support, and strategic equity stakes. It brings the buildout forward, but it also changes where the risk sits: from replaceable computing equipment toward power access, specialized campuses, and the institutions financing assets that cannot easily move when demand changes.

Power procurement became the product, not the accessory

In 2015, coverage described a 43MW wind arrangement for Google’s California headquarters. Apple partnered with First Solar on a 130MW California solar farm during a period when corporate electricity purchases were commonly presented as a way to reduce an operating footprint. Those arrangements supplied or offset power for businesses whose basic premises already existed.

OpenAI’s 20-year, 10GW Ohio agreement covers planned data-center capacity rather than an electricity purchase, so the two numbers do not measure the same thing. Google sought electricity for an existing headquarters; OpenAI needs a credible power path to create its campus.

The Department of Energy announced a SoftBank–SB Energy partnership to develop the Ohio campus with its own power supply. That announcement identifies an approach, not a demonstrated operating configuration. In the available coverage, neither the DOE nor SB Energy specifies the site’s generation mix, storage duration, interconnection rights, or when power could serve the contracted load.

SB Energy can sell the pipeline before it can sell its utilization

SB Energy’s IPO filing put 8.8GW of data-center capacity in Texas and Ohio under the combined description “contracted or under construction” and reported a $430B data-center-segment backlog. The filing does not, in the available coverage, separate every project’s contracted capacity from construction-stage capacity. Neither figure establishes how many servers are operating or how much of the backlog is unconditional revenue.

A long customer commitment gives a developer something a prospective site alone cannot offer: expected payments against which lenders and equity investors can assess construction risk. SB Energy also sought $5B to $7B in IPO proceeds and awarded OpenAI warrants valued at roughly $5.5B as it pursued the buildout. The warrants tie the anchor customer to the developer’s potential upside; the contemplated offering would bring in capital from investors outside the operating relationship.

SB Energy disclosed that it is “substantially dependent” on OpenAI. That is the other side of an anchor tenant’s value. One customer can make a very large campus easier to present to financiers while making its economics unusually sensitive to that customer’s needs and credit. SB Energy has not disclosed project-level termination rights, deposits, or a revenue schedule in the available coverage, leaving investors unable to measure how much protection the headline backlog affords.

Nvidia supports the finished asset, not every step toward it

Nvidia agreed to provide up to $105B of support for the Ohio campus and to invest $1.5B in SB Energy, according to coverage of the filing. Reporting described the larger support as a backstop for part of the completed campus’s value. It was not described as an unconditional promise to pay every construction bill or absorb every delay.

Nvidia’s position is distinct from OpenAI’s lease and SB Energy’s development work. The chip supplier has a commercial interest in a campus that can house a large compute deployment, and its support can make the finished property easier to finance. But a backstop keyed to completed value leaves lenders and the developer with questions about what happens before completion: whether power arrives, whether costs change, and what security supports each stage. The public terms do not settle how those risks are divided.

This is circular AI financing in a concrete form: a prospective equipment supplier helps support the property into which its equipment may go. Calling it circular does not make the lease fictitious or the support worthless. It identifies why financiers must distinguish independent customer demand from commitments made by parties that also benefit from the buildout.

Grid operators can reject a contract’s timetable

PJM Interconnection proposed requiring large data centers either to bring their own generation or to curtail consumption to prevent a large-scale outage. PJM’s proposal puts an operating condition beside the customer’s commercial promise: the grid must remain reliable when the campus draws power, not merely when its developer signs a lease.

Texas halted approvals of data-center grid connections for audits while more than 474GW of projects sat in the queue—over five times the state grid’s peak demand. A queue entry is not dispatchable electricity. Nor can a developer solve every problem by putting generation beside a server hall: the site still needs dependable supply, cooling, and a way to operate when equipment or a power source fails. Reporting on data centers switching to local generators has also described grid risks from unannounced disconnections.

Barclays counted 46GW of announced hyperscaler AI data-center capacity in 2025, with electricity use at full utilization comparable to roughly 44.2 million US households. “Announced” and “at full utilization” do important work in that estimate. Grid operators have to plan for enormous potential loads; financiers have to decide which of those loads can actually be delivered and used. This is where the AI infrastructure bottleneck acquires an address: a connection approval, a source of firm power, and a campus able to receive it.

Other developers use the template, then encounter the site

Anthropic signed a 20-year lease worth about $19B for a planned TeraWulf data center in Kentucky. The site is planned at roughly 400MW, with power delivery expected in the second half of 2027—not yet an accomplished delivery. Google took a different position in Cipher Mining’s arrangement with Fluidstack: it guaranteed $1.4B of a $3B, 10-year agreement and obtained the right to purchase a 5.4% stake in Cipher. The tenant lease and the third-party guarantee are different instruments, but each gives a developer a stronger counterparty to take to capital providers.

Amazon applied the same financing logic to power when it signed a 20-year agreement to buy 690MW from Constellation Energy, supporting more than $3B in investments at the nuclear operator’s Maryland site. A power producer can use a durable buyer commitment to support plant investment, just as a campus developer can use a durable compute commitment to support construction. In each case, the contract improves the case for spending; the plant or campus must still be delivered.

Crusoe raised $3.9B at an approximately $30.9B post-money valuation while pursuing factory-built data centers. Factory construction addresses part of the delivery timetable by moving repeatable work away from individual sites. It does not remove customer selection: Crusoe paused a Wyoming plan after failing to win customers including Google amid concerns about cost and timing. A developer can raise substantial capital and still decline to build a particular campus when buyers will not accept its terms.

Financiers can stop a deal faster than builders can stop a campus

AI infrastructure companies borrowed more than $100B in 2025, according to Refinitiv, while smaller borrowers faced higher interest rates as debt investors questioned unproven AI businesses. Lenders, insurers, and underwriters have also had to work out how much exposure to retain on a novel class of specialized property. For the wider shift from self-funded cloud expansion to contracted capacity and special-purpose financing, the related AI infrastructure project-finance picture matters: a campus lease is one part of a larger financing stack, not a substitute for examining it.

Investors reportedly questioned SB Energy’s sought valuation above $50B, and the company delayed its planned IPO. SoftBank reportedly sold $11.1B of high-yield junk bonds to replace a bridge loan connected to its OpenAI investment. Neither development proves that the Ohio campus will fail. Both show that the cost and availability of capital remain negotiable even when the customers and strategic backers are prominent.

Construction has a different clock. Developers order equipment and arrange power against demand they expect years from now; customers may reassess compute prices and model economics sooner than a specialized campus can be repurposed. A lender can decline the next financing, and a prospective customer can reject a Wyoming site. Once a developer has committed to a particular power path and poured the foundations for a particular load, it has fewer such choices. In Ohio, the lease is already measured in decades; the feeder still has to carry the load.