SB Energy’s 8.8GW Mixes Contracts and Construction

In September 2026, as SB Energy filed for an IPO, it placed 8.8GW of data-center capacity in one category: “contracted or under construction.” The filing did not divide the two. At that scale, the missing split matters more than another decimal place.

SB Energy reported the figure alongside a $430 billion data-center backlog. Coverage followed, increasing from two articles in 2026Q2 to 10 in 2026Q3. The jump does not validate the projects. It marks the moment when the company stopped being covered principally as a renewable developer and started being evaluated as an infrastructure system.

AI campuses make power a design constraint for the entire site. Developers that can assemble land, generation, storage, interconnection, financing and an anchor tenant can become strategic AI-infrastructure providers. They also concentrate their exposure in a few unproven, capital-intensive projects. SB Energy is testing that role at public-market scale.

A megawatt without a deadline is the wrong product

Utility-scale solar developers traditionally optimized generation cost, construction execution and long-term power sales. An AI campus imposes a different constraint. The customer needs electricity at a specified site, on a specified schedule, with enough reliability to support hardware that cannot wait for an interconnection queue.

The U.S. Department of Energy made that coupling explicit in March. Its partnership with SoftBank and SB Energy paired a 10GW Ohio data center with its own power supply at project inception. The department treated power provision as part of the building rather than a contract to procure after selecting the site.

PJM Interconnection applied the same logic from the grid operator’s side. PJM proposed that large data centers either bring generation or accept curtailment to avoid a large-scale outage. Google approached the constraint from the demand side, signing agreements with five utilities to reduce data-center consumption during peaks by as much as 1GW. Other developers have pursued on-site plants while grid access, permitting and equipment supply delay conventional service.

Each actor attacks a different hour of the same load curve. The developer brings supply. The customer shapes demand. The grid operator protects system stability. A power-purchase agreement covers only one edge of that triangle.

A campus operator therefore values deliverability more than nominal generation. A solar project may produce inexpensive megawatt-hours and still fail the campus if it cannot serve the right node at the right time. A battery may shift output and still fail if its duration, controls or interconnection rights do not match the load. The scarce asset is the coordinated right to consume power at the required node and hour.

OpenAI makes the campus one indivisible contract

OpenAI signed a 20-year, 10GW Ohio agreement with SB Energy. SB Energy is set to build, own and operate the site for the compute tenant. Nvidia supplies the AI infrastructure and supports part of the completed campus’s value.

The parties made those roles mutually dependent. Nvidia agreed to invest $1.5 billion in SB Energy and provide support of up to $105 billion for the Ohio campus. Nvidia also secured an exclusive AI-compute-infrastructure position. Its conditional backstop applies if OpenAI defaults and the completed campus cannot be re-leased or sold at a similar valuation.

OpenAI received another instrument. SB Energy awarded the company warrants valued at about $5.5 billion while preparing an IPO that targeted $5 billion to $7 billion in proceeds. OpenAI is simultaneously customer and prospective owner. Nvidia is equipment vendor, shareholder and conditional credit support. SB Energy is developer, landlord and power coordinator.

Together, the three parties price several risks. SB Energy needs the tenant to justify construction. OpenAI needs the campus to secure future compute. Nvidia needs the campus to absorb hardware. Lenders need all three relationships to survive long enough for the asset to produce cash.

A conventional lease lets an underwriter evaluate rent, term and tenant credit. The Ohio structure adds hardware exclusivity, residual facility value, power delivery and supplier support. The data center resembles one large computer whose power, cooling, networking and financing must boot together.

Nvidia moved inside the capital stack

Nvidia’s backstop matters because the chip supplier absorbs a risk that an arm’s-length vendor would normally leave with the developer and its lenders. The protection is conditional rather than absolute. It depends on OpenAI defaulting and on SB Energy failing to re-lease or resell the campus at a comparable value. Those conditions force financiers to estimate a market that barely exists: the residual value of a hyperscale AI campus built around a particular tenant and hardware ecosystem.

Meta’s Ohio Project Walleye points in the same direction. The project sought $3 billion in loans whose lenders would fund both the data-center building and the associated power assets. Meta’s proposal gave creditors one underwriting perimeter for compute and electricity instead of asking separate capital pools to assume that the pieces would meet on schedule.

Nine major technology companies had about $3 trillion of AI-related off-balance-sheet commitments, five times their roughly $600 billion in reported capital expenditure.

Reported capex records assets that cross a company’s accounting boundary. Lease commitments, special-purpose vehicles, supplier guarantees and future purchases can finance the same physical buildout from outside that boundary. The concrete does not care which balance sheet recognizes it.

This is why AI data-center financing increasingly depends on contract architecture. Nvidia can lower one part of the campus’s financing risk by attaching its own capital and support. It also deepens the campus’s dependence on the same commercial ecosystem. Nvidia’s support lowers one risk while concentrating another.

Solar earns a premium only when it behaves like capacity

SB Energy’s project descriptions do not establish that its AI campuses will be powered principally by solar plus storage. The Ohio project has been described as a data center with its own power supply, leaving the generation mix open.

Other companies are buying different forms of control. Microsoft has discussed a roughly $7 billion Texas plant initially capable of producing 2.5GW. AI labs have deployed on-site gas generators while grid capacity lags demand. OpenAI is hiring a power-trading lead to execute commodity hedging across its data-center portfolio. Google has contracted for demand response.

Solar and storage can provide part of that control. Solar fixes a portion of long-term energy cost. Storage shifts production into higher-value hours. Software coordinates charging, discharge and curtailment. Lunar Energy’s $232 million financing for batteries and virtual-power-plant software illustrates the distinction: isolated batteries store energy; coordinated batteries respond to a system.

A campus must stay online during the hour when grid supply tightens, the battery empties and the training run continues. Dispatchable generation, hedging and flexible load remain in the procurement set because annual averages cannot guarantee that hour.

A renewable developer earns a higher role by coordinating those intervals. It must connect generation, storage, grid service and load controls under operating rules that lenders and tenants can enforce. A developer that can provide that control, explored more broadly in data-center power resilience, offers more strategic value than any battery or turbine in isolation.

$430 billion needs a denominator

SB Energy’s two headline figures measure different things. Backlog expresses future contract value. Gigawatts express planned electrical scale. Neither figure identifies how much physical capacity has passed construction, interconnection and financing gates.

SB Energy also grouped two different states under the 8.8GW label. “Under construction” describes physical progress. “Contracted” describes a commercial relationship. A project can satisfy both descriptions, and the disclosure does not provide the split. Investors cannot derive installed capacity by adding the words together.

OpenAI’s role adds concentration risk. SB Energy says it is substantially dependent on OpenAI. The same anchor tenant that makes a large campus financeable also makes the developer sensitive to one customer’s performance, construction schedule and future appetite for compute.

In January, PJM reduced its summer 2027 peak-demand forecast from about 164GW to about 160GW because some projects, including data centers, lacked firm service or construction commitments. The four-gigawatt revision shows how quickly announced load disappears when a system operator applies readiness criteria.

Lenders must apply a similar filter. They have to price interconnection, construction, insurance, equipment, residual value and tenant durability inside a novel asset class. A signed commitment may solve one variable while leaving five others unfunded. A large backlog can therefore coexist with substantial compute execution risk.

The filing’s silence around the 8.8GW split matters more than another decimal place in the backlog. Contracted capacity demonstrates demand. Construction demonstrates expenditure. Operating capacity demonstrates delivery. Only the last category closes the loop.

SB Energy presents 8.8GW as scale. For investors, it is a diligence question: how much already has a site, power plan, financing, hardware support and durable tenant, and how much remains a set of promises waiting for one another? Until the company supplies that denominator, the figure mixes physical progress with commercial intent.