SoftBank owns a company that plans to sell $5 billion to $7 billion of stock before it has switched on a single server. SB Energy filed its S-1 with the SEC on September 1, and the filing reads less like a business plan than a map of every bottleneck in the AI buildout, priced in trillions.
NVIDIA is putting $1.5 billion into the company at the IPO price and guaranteeing up to $105 billion of obligations at one Ohio campus, so its chips keep getting a home.
OpenAI received warrants worth an estimated $5.5 billion to sign a 20-year lease. The tenant became the shareholder; the landlord is going public to fund its own construction.
Power first, compute second
The company describes itself as "power-first." The phrase matters. In 2025, the standard playbook was to find a site, sign a grid connection, and build the shell. Then beg a utility for enough megawatts. That model is dead. Interconnection queues now stretch past a decade in some regions, and AI campuses need gigawatts, not megawatts.
So the company reverses the order. It originates the power source first, often co-developing dedicated generation alongside the digital infrastructure, then builds the data center around the megawatts it already controls. The company already operates about 2.2 GW of solar and battery assets, with another 2.5 GW under construction. The data centers feed off that base.
The result is a vertically integrated platform covering land, grid capacity, generation, design, construction, and operation. The entire lifecycle sits on one balance sheet. Hyperscalers and frontier labs get turnkey, triple-net-leased facilities with delivery timelines measured in the difference between signed contract and switched-on.
It is not a small gap. Of the 8.8 GW-IT the company has contracted, only about 0.8 GW is under construction. The other 8.0 GW is signed but unbuilt. The company expects its first data center revenue in the fourth quarter of 2026, when the initial phase of its Cosmos Technology Campus in Texas reaches rent commencement.
The numbers behind the filing
Contracted but unbuilt AI infrastructure
SB Energy's backlog is dominated by data center leases that become revenue only as facilities reach rent commencement · EBC Financial, Sep 2026
H1 2026 revenue came to $138.7 million, almost all from the legacy power business. Up 66.4% year over year. The headline $3.21 billion net loss is mostly non-cash: about $2.57 billion from warrant revaluation and $589.5 million from stock-based compensation. The cash story is less dramatic than the accounting one, but the scale of the construction bill ahead is not.
The company has raised $19 billion in project capital to date and closed roughly $2.4 billion in equity across four rounds, including $1 billion split between SoftBank and OpenAI in January 2026. It also issued $999 million of senior secured notes due 2031 to finance the Cosmos campus, where a SoftBank affiliate is the tenant under a 15-year triple-net lease.
The listing itself is unusual in structure. NVIDIA committed to buy $1.5 billion of a new class of non-voting stock, Class N, at the IPO price in a concurrent private placement. Separately, a prepaid forward gives NVIDIA the right to shares at 90% of the IPO price. Together the two arrangements give NVIDIA roughly $3 billion of IPO-linked exposure.
The bigger number is the guarantee. NVIDIA's $105 billion figure is a cap on guarantees supporting about 4.25 GW of IT capacity at PORTS-Pike, not an equity check. The obligations cover land, power, and building-shell costs, and they shrink as OpenAI fulfils its 20-year lease. NVIDIA also gets the strategic prize: the campus is designed around its compute, with each generation of hardware potentially representing $150 billion to $200 billion of NVIDIA revenue.
The tenant wrote the landlord a check
OpenAI sits on both sides of the table. It is the anchor tenant at PORTS-Pike under a 20-year lease covering about 8 GW of the campus's 8.0 GW capacity. It is also an equity investor, having put $500 million into the company alongside SoftBank in January. And it holds warrants worth an estimated $5.5 billion, issued at $0.01 exercise price, which were the inducement to sign.
We are substantially dependent on OpenAI as a tenant and strategic partner, and any adverse change in its financial condition or willingness to perform its contractual obligations could materially and adversely affect our business.— SB Energy, Form S-1
That is the S-1's own risk language, and it is the crux of the story. An AI lab being paid billions in equity to sign a lease flips the tenant-landlord relationship in AI infrastructure. The frontier lab gets upside in the very balance sheet IPO investors are being asked to price. The landlord's entire contracted pipeline leans on the financial health of two anchor tenants: SoftBank and OpenAI. Losing either, through insolvency, breach, or a reduced lease scope, could eliminate related revenue entirely.
This is the tension at the heart of the model. The vertical integration solves the power bottleneck, which is real. But it concentrates risk in a way no pure data center REIT does, because the tenants are also the controlling shareholder's partners and the guarantor's customers.
What it means for AI infrastructure investors
The IPO is a pricing event for the whole power-first thesis. Renaissance Capital pegs the offering at roughly $5 billion, the WSJ-reported range of $5–7 billion. The valuation has been floated at more than $50 billion. A figure that, if it holds, prices 8.8 GW of contracted capacity almost entirely on the strength of the backlog and the partners behind it.
The comparison to the internet buildout is instructive. PwC's Global Data Centre Outlook, published a day after the filing, projects $31.6 trillion of data center capex through 2050, with ICT equipment growing from 70% to 93% of the total. That recurring refresh cycle, chips replaced every four to six years, is precisely why a company like NVIDIA is willing to guarantee construction financing: the campus is a permanent home for its silicon, not a one-off order.
The company is the most explicit example so far of an AI lab receiving equity as inducement for a compute commitment. As we wrote in August, institutional capital is reorganizing around AI data centers. SK Telecom spun off its SK Horizon unit as a $2.2 billion bet on the same thesis. The company takes it further: the power generator, the landlord, and the compute buyer are all becoming the same thing.
Where does the power-first model take the market by 2030?
Probability: 65%. Grid interconnection bottlenecks are structural, not cyclical, and every frontier lab now wants the same equity kicker its rival received.
✅ Arguments for
Confirmation criteria: a second power-first developer files to go public, or a hyperscaler signs a build-own-operate deal on this model.
❌ Arguments against
Disconfirmation criteria: an anchor tenant defaults or restructures, or regulators cap equity stakes for tenants in infrastructure IPOs.
Development scenarios
🟢 Optimistic scenario (30%)
Implications: power-first developers attract institutional capital on repeat, and the AI infrastructure asset class starts to trade like a utility-growth hybrid.
🟡 Base-case scenario (50%)
Implications: it becomes the reference case for AI infrastructure IPOs, and its quarterly disclosures are read as a proxy for the whole power-first model.
🔴 Pessimistic scenario (20%)
Implications: equity-for-lease structures draw regulatory scrutiny, and the AI infrastructure IPO window tightens for the next wave of power-first developers.
First data center revenue at Cosmos, scheduled for Q4 2026 rent commencement
NVIDIA's $105 billion guarantee cap drawn or extended, signaling PORTS-Pike construction pace
A second equity-for-lease deal by OpenAI with another developer, confirming the model generalizes
The amended S-1 price range, converting the $5–7 billion estimate into official terms