Blue Owl has about $3 billion of equity in Project Jupiter, 20 banks lent $18 billion, and the loans were quoted near 90 cents a week before Oracle's force-majeure notice. The financing holds and the data halls are still going up. The desk expects the cost of a one-year slip to show up in wider spreads and tighter power clauses on the next wave of AI data-center deals.
Ninety cents. That was roughly where Santander and Jefferies were offering pieces of the $18 billion construction loan behind Oracle's Project Jupiter campus in New Mexico on September 18, the Financial Times reported, after efforts to sell the debt beyond the roughly 20 original banks stalled. Six days later Oracle sent the developer, Blue Owl's Stack Infrastructure, a force-majeure notice tied to power delays. Oracle shares closed down 3.5% and Blue Owl down 3.6%, Reuters reported, citing people familiar with the matter who put the likely slip at about a year.
Jupiter will probably get built. Oracle says it "remains on our planned schedule", and Blue Owl says the notice "does not change the financial commitments to this multiyear project". What matters is who carries the cost of a year's wait, and what that does to the price of the next deal. On the evidence so far, that cost falls mostly on the equity and the banks, and the market has already started charging for it.
The trouble is a gas line. The 2.45-gigawatt campus in Doña Ana County was designed to run on an on-site microgrid of Bloom Energy fuel cells fed by Energy Transfer's roughly 17-mile Green Chile pipeline. The New Mexico State Land Office twice refused the pipeline a crossing of state trust land, and its in-service date has moved from August 2026 to February 1, 2027, El Paso Matters reported. The fuel-cell air permit is still outstanding, and the New Mexico Supreme Court has stayed that proceeding. Oracle's own September 14 statement says the data halls are going up under county permits. Buildings without power earn nothing, though.
According to the people Reuters spoke to, Oracle cannot terminate the lease under any circumstances, pays the debt costs, and is contractually responsible for securing power. Blue Owl earns about 9% on its equity while the campus is under construction and a levered yield of about 11% once it is complete. The notice lets Oracle stretch the development-stage period and push back the start of the higher completion rent. Oracle still services the debt; Blue Owl simply waits longer for its step-up.
On those figures the direct cost to Blue Owl is small. Two percentage points on $3 billion is about $60 million a year. The bank loss is larger on paper: a 10-point discount on $18 billion is $1.8 billion of mark-to-market pain for any lender forced to sell, and the FT said the syndicate is holding more of the loan than planned. The exposure is duration. A lender underwriting a 2028 delivery is now underwriting 2029 against a tenant that S&P cut to BBB-, one notch above junk, in July.
Compare Meta's Hyperion deal with Blue Owl in October 2025. Blue Owl funds took 80% of a joint venture, Meta kept 20%, and a bankruptcy-remote vehicle, Beignet Investor, sold $27.3 billion of 6.581% amortizing bonds due 2049 at Treasuries plus 225 basis points, IFR reported. The bonds were rated A+, one notch below Meta, because S&P judged that the structure passed "substantial credit risk to Meta". Meta signed only a four-year initial lease, backstopped by a 16-year residual value guarantee. Oracle's Jupiter lease is harder to walk away from than Meta's, yet its debt sits in bank books at a discount instead of an investment-grade bond. The tenant's rating does most of the work in these structures. Hyperion's investors bought Meta's balance sheet with a building attached. Jupiter's lenders bought Oracle's.
That balance sheet is stretched. In the quarter to August 31, Oracle reported operating cash flow of $23.1 billion against capital spending of $28.5 billion, a free-cash outflow of $5.4 billion, with $125.3 billion of notes payable and $36.4 billion of cash. It raised $20 billion of equity through an at-the-market programme in the same quarter. It guides to $90 billion to $95 billion of gross capex for the fiscal year, about $70 billion net of customer prepayments. Off the balance sheet sit $260 billion of data-center lease commitments that start between now and fiscal 2029 and run 15 to 19 years, per its annual filing. Spread evenly, that is roughly $14 billion to $17 billion of rent a year.
Against that, remaining performance obligations of $664 billion look like cover. About 13% of that, or roughly $86 billion, is due to be recognised in the next 12 months. The rest depends on halls like Jupiter's being powered. OpenAI's contract, reported in 2025 as $300 billion over five years from 2027 for 4.5 gigawatts, works out to about $13 billion of revenue per gigawatt-year. If Jupiter's 2.45 gigawatts were sold at that average, a year's delay would push about $33 billion of revenue to the right. That is a rough scaling of reported terms, not a disclosed Jupiter figure, but it shows why Oracle wants the option. Deferring rent on a campus that cannot yet bill OpenAI is sensible cash management for a company burning $5 billion a quarter.
Oracle's own statement supports the case against alarm. Michael Egbert, an Oracle spokesman, said force-majeure notices "are commonplace in developments of this scale" and "do not by themselves establish a project delay". Lawyers told Reuters such clauses are spreading through data-center development agreements. All true. But a tenant confident of a 2028 delivery has little reason to serve notice in September 2026, and the loan was trading at 90 before the notice existed. The loan market had priced the delay before Oracle put it in writing.
Grid data says Jupiter's problem is common. Lawrence Berkeley National Laboratory's Queued Up 2026 report found the median project reaching commercial operation in 2025 took 61 months from interconnection request, against 36 months for 2015 projects. About 2,060 gigawatts of generation and storage were waiting at the end of 2025. Behind-the-meter gas was supposed to be the shortcut around that queue. Jupiter shows the shortcut still needs a pipeline permit.
Blue Owl's pitch to insurers and wealth clients rests on contracted, investment-grade rent, and this makes the "contracted" part slower. The syndicate banks own more of a sub-par loan than they wanted. Bloom Energy's largest deployment, part of a partnership for up to 2.8 gigawatts, now waits on a state land office. SoftBank's SB Energy, building an Ohio campus leased to OpenAI, postponed its IPO this week after investors balked at a valuation above $50 billion, and Reuters said the Jupiter episode is running through that discussion. OpenAI loses time. The winners are developers with energized land and firm utility service, who can now charge for certainty, and new lenders, who get to write the tighter documents.
The desk expects the next large AI build-to-suit financing to carry an explicit power-delivery covenant or a spread at least 50 basis points wider than comparable 2025 deals, and expects Jupiter's loan to still be quoted below 95 cents when Oracle reports fiscal second-quarter results in December. We also expect Oracle to hold its $90 billion to $95 billion capex guide at that report, which would tell investors Jupiter is a timing issue. The checkpoints are SB Energy's rescheduled pricing, now pushed to mid-to-late October at the earliest; the New Mexico Supreme Court's ruling on the air permit; Blue Owl's third-quarter results in early November; and February 1, 2027, the pipeline's target date. A pipeline in service on time, or Jupiter debt quoted back above 95 cents, would mean the desk was too pessimistic. A second tenant force-majeure notice on another OpenAI site would mean it was not pessimistic enough.