OpenAI’s Ohio Data Center: 8 GW Announced, 800 MW First

OpenAI says it has agreed to secure approximately 8 gigawatts of information-technology capacity at the PORTS-Pike Technology Campus in Ohio. That is the long-range envelope. The first capacity the company expects to become available is 800 megawatts in 2028—one tenth of the headline figure.

The distinction is the story. An 8 GW agreement is not 8 GW of AI compute operating today, one fully financed construction phase, or one power plant ready to switch on. It sits above a chain of leases, credit support, permits, generation, transmission, data-center buildings, and installed computing systems. Each link has its own conditions and timetable.

OpenAI has a 20-year agreement for capacity delivered over time. NVIDIA has filed conditional credit support for leases covering about 4.25 GW of IT load, while support for roughly another 3.8 GW remains optional. The first 800 MW is expected in 2028. The announcements and filings reviewed for this article do not establish that any OpenAI compute is operating at the site now.

This is how to read the OpenAI Ohio data-center deal—and other giant AI-infrastructure announcements—without confusing contracted ambition with delivered capacity.

The OpenAI Ohio data-center commitment ladder

OpenAI’s August 17 announcement says SB Energy will build, own, and operate the data center under a 20-year lease. OpenAI will be the customer and pay as completed capacity becomes available. NVIDIA will supply the site’s AI compute infrastructure.

Those facts are meaningful commitments, but they do not all sit at the same level.

The announced, supported, first-delivery, and operating stages of the OpenAI Ohio data center

Layer	Current evidence	What it does not prove
Long-range agreement	OpenAI says it secured approximately 8 IT-GW under a 20-year lease.	That all 8 GW is financed, permitted, built, equipped, or available.
Initial NVIDIA credit support	NVIDIA’s filing covers leases for approximately 4.25 GW of IT load, subject to conditions.	That NVIDIA has paid \$105 billion or that 4.25 GW is ready.
Possible expansion	NVIDIA may support roughly another 3.8 GW at its sole discretion.	A binding commitment to the expansion.
First delivery target	OpenAI expects the first 800 MW in 2028, largely using existing AEP infrastructure.	A guaranteed completion date.
Operating compute	No cited announcement or public record shows OpenAI capacity running at PORTS-Pike now.	Nothing about what may already be under construction or ordered.

The table is not a probability forecast. It is a way to keep different kinds of evidence separate. A signed commercial agreement matters more than an aspiration; a ready-for-service data hall matters more than a lease; and measured operating load matters more than either.

What “8 IT-gigawatts” actually measures

A watt measures power: the rate at which energy is used or supplied. One gigawatt is 1,000 megawatts. Energy accumulated over time is measured in watt-hours, so a facility drawing 800 MW steadily for one hour would use 800 megawatt-hours. The deal’s GW figures describe capacity, not annual electricity consumption.

The IT load qualifier is equally important. It refers to electrical capacity for the computing equipment itself. A complete campus also needs power for cooling, conversion, networking, lighting, and other facilities. Generation capacity is a different measurement again. A plan for 10 GW of new generation is not automatically 10 GW delivered continuously to servers, and it is not interchangeable with 8 GW of IT load.

Nor does 8 IT-GW specify a number of graphics processing units, or GPUs. The count depends on the accelerator generation, rack design, utilization, networking, cooling, and power available at each phase. A 20-year site can host several hardware refreshes. NVIDIA’s filing says the initial 4.25 GW is intended for its full-stack DSX platform, subject to limited exceptions, but it does not freeze one chip model for the lease term.

That is why Axios described the arrangement as a “land, power and shell” structure. The durable asset is a serviced place to install changing generations of compute. The hardware inside it can change much faster than the lease.

NVIDIA’s $105 billion is a conditional backstop

The financing headline also needs unpacking. NVIDIA’s August 17 Form 8-K describes multiple residual value guaranties related to leases for approximately 4.25 GW of IT load. Its aggregate payment obligation for that initial commitment is capped at $105 billion.

That does not mean NVIDIA transferred $105 billion to the project. The filing says an agreement generally becomes effective when the relevant lease begins. NVIDIA’s payment obligations require specified ready-for-service conditions and are triggered if OpenAI becomes insolvent and defaults, or if it fails to make lease payments.

After a trigger, NVIDIA would generally cover a shortfall between a lease’s guaranteed minimum value and the amount recovered through reletting or a sale. It could assume the lease, require SB Energy to seek a replacement tenant, start a sale, allow termination, or defer a remedy for up to a year while paying specified project costs. OpenAI has agreed to reimburse and indemnify NVIDIA for amounts NVIDIA actually pays under the agreements.

The filing adds two boundaries that shorter reports can lose:

• credit support for approximately another 3.8 GW is exercisable at NVIDIA’s sole discretion; and
• the complete agreement form had not yet been filed. NVIDIA said it would include it with the quarterly report for the period ending July 26, 2026.

NVIDIA’s separate $1.5 billion investment in SB Energy is an investment. The $105 billion cap is a conditional contractual exposure. Combining them into one “NVIDIA is spending” number would misstate the deal.

The structure still ties together a tenant, a developer, and the exclusive compute supplier. Axios reported that this renewed concern about circular financing in the AI industry. The filed conditions do not prove the project is unsafe or self-financing; they identify who absorbs specified losses if OpenAI cannot keep paying. Useful follow-up evidence will be actual lease commencements, the full guaranty text, project financing, and any NVIDIA payments—not the cap alone.

Power, permits, and buildings each have a clock

The 2028 first-delivery target and the broader buildout depend on different infrastructure.

OpenAI says the first 800 MW should rely largely on existing AEP infrastructure. Later phases require new power plants, transmission lines, permits, environmental review, and financing. The developer’s PORTS-Pike overview describes a broader plan for 10 GW of new generation, including 9.2 GW of natural-gas generation, plus $4.2 billion in high-voltage transmission investment.

The Department of Energy’s March fact sheet attributes those figures to the public-private project and says the new generation includes Japanese funding under the U.S.-Japan Strategic Trade and Investment Agreement. These are announced investment and capacity plans, not completed plants or measured output.

AEP’s own timeline is narrower. Its May transmission announcement says a customer-funded project would add a substation and about 50 miles of 765-kilovolt line. AEP expected construction to begin in early 2027, finish in late 2028, and power to begin flowing in 2029, pending regulatory approvals.

That does not necessarily contradict OpenAI’s 2028 target. OpenAI explicitly associates the first 800 MW with existing infrastructure; AEP associates 2029 with new transmission. They are separate milestones and should stay separate until the companies publish a more detailed energization schedule.

Permitting is also unfinished. As checked on August 19, the federal Permitting Dashboard showed environmental review and permitting in progress, with 0 of 4 listed processes complete and an estimated completion date of December 23, 2026. The dashboard warns that planned and paused dates can change. It also does not replace every state, local, or utility approval.

WOUB’s local reporting adds a useful land constraint. State permit filings cover more than 1,000 acres of private property around the former Portsmouth Gaseous Diffusion Plant, while the federal site is still undergoing cleanup. SB Energy’s final layout across private and federal land was not yet clear.

Jobs, water, and ratepayer benefits remain claims to measure

OpenAI projects 35,000 construction jobs during a six-year buildout through 2032 and 2,500 long-term operating jobs. It also says SB Energy will pay the project-specific grid and transmission costs rather than shifting them to Ohio ratepayers. Those are company projections and commitments, not measured outcomes yet.

Water has the same evidence boundary. OpenAI says the data centers will use closed-loop, air-cooled systems that recirculate water instead of continuously consuming it through cooling towers. It compares expected ongoing use after the system is filled with an office building serving a similar number of people—but says it will publish the actual expected water use only after the site design is final.

The comparison therefore cannot substitute for a quantified site estimate. Useful reporting will need the final design, water source, initial fill, maintenance and washdown demand, landscaping, peak-day use, and any water needed indirectly by power generation.

OpenAI promises an annual report covering local hiring, community investment, water use, and project energy use. That creates a future accountability mechanism. The first report should be compared with permit records and utility milestones, not only with the original promises.

What would turn 8 GW from a headline into infrastructure

The next decisive evidence is not another total-cost estimate. It is a sequence of narrower events:

• federal and state approvals completing for defined parcels and phases;
• a final site and water design with quantified demand;
• AEP and generation projects receiving approvals and meeting energization dates;
• specific data halls satisfying ready-for-service conditions;
• the relevant leases beginning and OpenAI payments starting;
• compute systems being installed, accepted, and used; and
• annual reports showing actual jobs, energy, water, and community investment.

Until those milestones arrive, the accurate description is substantial but conditional: OpenAI has contracted for a very large Ohio capacity pathway, NVIDIA has agreed to backstop an initial set of future leases under specified conditions, and the partners are targeting 800 MW first. The 8 GW number is the destination on the agreement—not the site’s present operating state.

Sources

1. OpenAI announcement for the PORTS-Pike project
2. NVIDIA Form 8-K on PORTS credit support
3. Federal permitting dashboard for PORTS Technology Campus
4. Department of Energy fact sheet for the Portsmouth AI project
5. AEP Ohio announcement for the Piketon-area transmission project
6. PORTS-Pike Technology Campus project overview
7. Axios report on the OpenAI, NVIDIA, and SB Energy agreement
8. WOUB report on OpenAI’s Pike County lease plan