SB Energy's net loss widened to roughly $3.2 billion in the first half of 2026, up from about $250 million a year earlier. The company earned $140 million in renewables revenue over the same period. It operates zero data centers.
The roughly $3 billion hole in the middle of that income statement is, reportedly, not an operating loss. It is the remeasurement of a warrant liability — the accounting cost of a promise SB Energy made to its own tenant. In January it granted OpenAI warrants worth an estimated $3.6 billion. By June 30 those warrants were worth $5.5 billion. The landlord paid the tenant to sign, and the price kept climbing.
Set that $5.5 billion against the $5 billion to $7 billion SB Energy hopes to raise in an IPO expected as soon as this month, and the tenant-acquisition cost swallows 92% of the offering. Then add the third party. Nvidia, the chip supplier whose hardware will fill the building, agreed to guarantee up to $105 billion tied to OpenAI's leases on the first 4.25 gigawatts.
The tenant got paid to move in. The supplier co-signed the loan. Read plainly, the money runs backwards down the entire supply chain.
The tenant got paid to move in. The supplier co-signed the loan. Read plainly, the money runs backwards down the entire supply chain.
The easy verdict is that this is circular financing gone insane — a dot-com echo where everyone buys from everyone and demand is a hall of mirrors. That read is wrong, or at least lazy. This is not absurdity. It is engineering.
Last week these were reportedly talks: Nvidia reportedly weighing a $250 billion guarantee, SoftBank hunting for permanent financing to replace a $40 billion bridge. This week the structure is resolved and on paper — a $105 billion capped residual-value guarantee, a concrete warrant block, an exclusive Nvidia site, and explicit federal backing underneath. The abstraction has hardened into an instrument.
Here is what the instrument is for. The scarce asset in the AI buildout was never gigawatts. Every operator says power is short, land near transmission is short, turbines are sold out to 2029 — but if capacity were the binding constraint, SB Energy would auction 4.25 gigawatts in Pike County to the highest bidder and pocket the premium. Instead it paid the bidder.
The real scarcity is a signature a lender will underwrite. SB Energy has no operating data centers, no data-center revenue, and a renewables unit earning $140 million a half. No credit committee writes a construction loan against a speculative shell. What makes the debt bankable is a 20-year lease with a counterparty the committee will accept.
And there is the problem: OpenAI is, reportedly, the counterparty everyone wants and nobody can price. It reportedly generates real revenue at a scale no other lab matches, and it loses money at a scale no tenant in the history of commercial real estate approaches. It is unbankable on its own credit.
So the four parties assembled a synthetic credit rating out of parts. Walk the capital stack in order and each contributes a different piece.
Nvidia contributed the balance sheet. Its $105 billion residual-value guarantee means that if OpenAI cannot pay the rent on the first 4.25 gigawatts, Nvidia covers the loss. That is the guarantee a lender actually underwrites — not OpenAI's cash flow, but Nvidia's.
SoftBank contributed the equity and the political weight. It has reportedly committed more than $64 billion across OpenAI, the SB Energy investments, and Stargate, and its strategy requires owning both the model layer and the physical layer, which means both must keep working.
SB Energy contributed the only thing it had left: upside in itself. The warrants cost nothing today. If the company fails, they expire worthless and the incentive cost nothing. If it succeeds, OpenAI collects — but only at market-cap milestones that make the founders and SoftBank rich first, with the dilution landing on whoever buys the IPO.
Each piece plugs a hole in OpenAI's creditworthiness that OpenAI could not plug itself. The lease becomes an asset SB Energy can borrow against because Nvidia stands behind it, SoftBank funds around it, and the tenant is paid in equity to make its own signature more valuable.
Now note the tell. If this were a comfortable market, none of this scaffolding would be necessary — you would lend against the lease and move on.
Instead the financing is getting harder, not easier. Oracle's $16.3 billion data-center package reportedly needed PIMCO to anchor $10 billion after US banks stepped back. That is not what a comfortable market looks like. When investment-grade Oracle needs a private-credit anchor because banks won't hold the paper, the marginal AI data-center loan is being priced as something closer to venture risk.
The warrant thrives in exactly that environment because it costs nothing now and defers the reckoning to the equity market. But deferral is not elimination. The structure only holds if every party keeps committing — Nvidia keeps guaranteeing, SoftBank keeps funding, IPO buyers keep valuing SB Energy above its warrant milestones, and OpenAI keeps generating enough revenue to make a 20-year lease plausible.
Pull any one thread and the synthetic rating unravels, because the rating was never a property of OpenAI. It was a property of the pool.
That is the systemic consequence. The four parties did not distribute risk across independent balance sheets; they pooled their solvency into a single object. Nvidia's guarantee is, reportedly backed by revenue Nvidia expects from selling chips into this same campus — as much as $150 to $200 billion per hardware generation. SoftBank's equity depends on both OpenAI and SB Energy succeeding. SB Energy's stock is the collateral for the warrants and the basis for the IPO. There is no isolated failure available here. If demand stalls, everyone fails together.
Which is where the politics enters, and why it is load-bearing rather than decorative.
Underneath the campus sits at least 9.2 gigawatts of natural gas, with $33.3 billion of Japanese funding committed to the generation and DOE blessing on federal land — the decommissioned Portsmouth Gaseous Diffusion Plant. This is not a private real-estate deal. It is a public-private partnership with the Department of Energy and the Department of Commerce inside it.
Then the framing. Trump has said communities that reject data centers "want to end up being backwards and poor," and that China "could not be happier" with the anti-data-center movement. Set that against the ground reality: more than 500 local bans, moratoria in New York, power-grid audits in Texas and Ohio, a moratorium bill in the House.
Read the two together and the political rhetoric stops looking like cheerleading and starts looking like a repricing shield. The pooled risk is enormous and shared. If it ever had to be repriced — if a lender or a public-market buyer decided OpenAI's synthetic rating was a fiction — the loss would cascade through Nvidia, SoftBank, SB Energy, and a federally backed gas plant simultaneously.
So the state is being enlisted to make that repricing politically impossible. When the president frames refusal as national decline and hands China the role of beneficiary, he is converting a credit question into a loyalty test. Backwardness or growth. That framing exists precisely because the honest financial question — is this tenant bankable? — has an uncomfortable answer that the structure was built to avoid asking out loud.
The warrant is the elegant piece of a machine designed to defer that question. The federal backing is the piece designed to make sure no one is allowed to ask it.
Both work only as long as demand keeps rising. The moment it stops, there is no comfortable market to sell into, no isolated party to absorb the hit, and no politically safe way to admit the pool was underwater. There is only the shared failure the state has already pre-committed to defend.