The next important test of the artificial-intelligence boom may occur not inside a laboratory, but among the lenders financing an unfinished data center.
At Project Jupiter, an Oracle-leased campus in southern New Mexico intended to support computing services for OpenAI, roughly $18 billion in project loans are being quoted at 89–91 cents on the dollar, the Financial Times reported September 18. Efforts to distribute the debt more widely have stalled, leaving banks with more exposure than intended, according to the reporting, subsequently relayed by Reuters.
Those are reported quotations—not independently verified transactions, payment defaults or realized bank losses. Nevertheless, they raise a consequential question: can the institutions financing AI infrastructure sell that risk to investors on the terms they originally expected?
The gap between a building and a business
The original financing was designed to spread exposure. In November 2025, Reuters, citing Bloomberg, described an approximately 20-bank consortium providing project financing, with further distribution to banks and institutional investors planned. Oracle was identified as the intended tenant. This is therefore project debt connected to Oracle—not simply $18 billion of Oracle corporate bonds.
The distinction matters. Who ultimately absorbs a loss depends on the borrowing entities, guarantees, lease terms and other contractual protections—not merely whose name appears in the headline.
Power supply adds another layer of uncertainty. Oracle announced in June that the campus would switch from its earlier gas-turbine and diesel-generator design to Bloom Energy fuel cells. In September, it emphasized that the data-center buildings and proposed microgrid are separate facilities with separate permitting requirements.
But the regulatory story has changed. On September 17, New Mexico’s Supreme Court denied two petitions and lifted the earlier stays affecting air-permit proceedings and construction-water authorization. The air-permit process can resume; that is not the same as receiving the permit or delivering an operational power system. A description of Jupiter as indefinitely frozen by those court orders would be misleading.
For lenders, the remaining question is economic as much as legal: when will the campus become capable of earning revenue? Construction progress alone cannot answer it.
What Oracle’s cash flow reveals
Oracle’s latest filing shows the scale of the financing challenge. In the quarter ended August 31, it generated $23.1 billion of operating cash flow but spent $28.5 billion on capital expenditures, producing approximately negative $5.4 billion in free cash flow. These are companywide figures, not Jupiter’s accounts.
A crucial detail sits inside that operating cash flow: $11.4 billion came from customer prepayments carrying a significant financing component. Those payments are already included in the $23.1 billion—not an additional pool of cash. They help fund delivery, but should not be mistaken for cash generated entirely from services already performed.
Oracle also disclosed $288 billion of additional lease commitments, predominantly for data centers, generally running 15–19 years once commenced. These were not yet recognized on its balance sheet. They are future contractual commitments, not $288 billion of immediately payable debt.
There is substantial evidence on the other side. Oracle reported cloud-infrastructure revenue of $7.4 billion, up 121% year over year, and $664 billion in remaining performance obligations. The business is delivering rapid revenue growth. But contracted future revenue still requires successful delivery and collection; it is not cash available today.
When a project problem becomes a credit cycle
The adverse scenario does not require AI to stop advancing.
It requires financing conditions to deteriorate faster than projects can begin generating sufficient cash. If banks cannot distribute loans without substantial discounts, they may demand more sponsor equity, charge higher rates or finance fewer subsequent projects. Delayed projects could then reduce equipment orders and expected supplier revenue.
That would turn a financing problem into a capital-spending slowdown.
Jupiter alone does not establish that sequence. Its difficulties could remain specific to its power arrangements, permitting history and contractual structure. Successful commissioning and renewed lender demand would strengthen that interpretation.
The unanswered questions are concrete: how much debt has actually been drawn and sold, at what prices, with what guarantees, and how much remains with each lender? The sources reviewed do not resolve them.
The decisive evidence would be repetition across unrelated projects. If other AI financings encounter similar distribution problems, Jupiter would become more than a difficult construction story. It would be evidence that the price—and availability—of capital is beginning to constrain the AI buildout.
AI can succeed technologically while some of the capital financing it earns disappointing returns. Project Jupiter is a case worth watching because it puts that distinction into focus.