The Data Center Trade Has a Credit Problem

Oracle’s Project Jupiter is supposed to be one of the largest AI infrastructure projects in the country. The New Mexico campus secured roughly $18 billion in loans last year. Oracle is the tenant. OpenAI is the ultimate customer for the computing capacity. Blue Owl has about $3 billion of equity in the project. Last week, the debt was being quoted at 89 to 91 cents on the dollar.

Then Oracle issued a force majeure notice. Oracle says Jupiter remains on schedule. Blue Owl says the parties remain committed. What happens if it isn’t?

Project Jupiter may be the proverbial canary that exposes the financial structure underneath the AI infrastructure boom. The industry has spent the last several years discussing gigawatts, GPUs and capital expenditures. Far less attention has been paid to how all of it is being financed. Increasingly, the answer is debt.

Not necessarily debt sitting neatly on the balance sheets of Microsoft, Meta, Oracle or the other companies ordering the capacity. A growing amount sits inside special-purpose vehicles created to build individual data centers. The hyperscaler signs a long-term lease or capacity agreement. Outside investors contribute equity. Banks, private credit funds, insurers and bond investors provide the rest.

The structure turns a hyperscaler’s enormous upfront capital expenditure into a long-term contractual obligation while allowing much of the associated borrowing to reside somewhere else.

The Bank for International Settlements has a less elegant description for this arrangement: “shadow borrowing.”

The BIS warned earlier this year that these structures create new connections among hyperscalers, private credit, insurers and banks, with potential transmission channels through refinancing pressure, changing private-credit appetite and the activation of guarantees.

Jupiter’s $18 billion financing was supposed to be distributed by its syndicate banks to a broader group of investors. According to the Financial Times, that process has stalled. Banks have been left holding more Oracle-linked project debt than intended. Syndicate banks including Santander and Jefferies are now quoting the loans at substantial discounts to par. Nothing needs to default for that to become a problem.

Banks do not warehouse $18 billion of project loans because they enjoy collecting enormous concentrations of construction risk. They originate the loans expecting to distribute them. If buyers disappear, the banks are left with balance-sheet capacity trapped in yesterday’s deal precisely when tomorrow’s deal needs financing.

That is the first transmission mechanism.

• Jupiter gets delayed.
• Jupiter debt reprices.
• Investors demand more compensation for financing the next data center.
• Developers need more equity.
• Projects that were economical at the old cost of capital stop working at the new one.
• More projects are delayed or canceled.
• Existing data-center debt reprices again.
• The process does not require a wave of defaults. It only requires investors to realize they were underpricing the same risk across an entire asset class.
• Enormous amount of paper to reprice.

Meta’s Hyperion project in Louisiana is particularly interesting. Meta and Blue Owl created an SPV called Beignet Investor to finance the project. Blue Owl contributed billions of dollars of equity. Meta contributed a smaller amount. The vehicle then issued approximately $27 billion of debt to finance construction.

The bonds mature in 2049.

Investors are therefore financing an extraordinarily capital-intensive technological asset more than two decades into the future.

Its financing works because Meta’s lease and contractual protections make the underlying project debt look sufficiently safe to investors. That is the basic logic behind much of this market. The borrower may technically be an SPV, but the creditworthiness of the hyperscaler sitting behind the lease transforms the debt into something that can be sold to institutional investors.

The problem is that the asset still has to be built.

• It still needs power.
• It still needs cooling.
• It still needs permits.
• It still needs billions of dollars of equipment.
• It has to arrive close enough to schedule for the original financial model to work.

Jupiter matters to Hyperion even though Meta has nothing to do with Jupiter. If Jupiter debt continues falling, investors do not need to conclude that Meta is insolvent. They only need to reconsider the spread they require to own a 23-year security financing a gigantic single-purpose data center.

Credit contagion usually begins with repricing, not bankruptcy. Suppose investors originally demanded 200 basis points above Treasuries for this kind of debt. After watching Jupiter encounter power problems before it even opens, they decide the appropriate spread is 300 basis points.

• Nothing happened to Meta.
• Nothing happened to AI demand.
• Nothing happened to Hyperion.

The market value of existing bonds can still fall because investors now require a higher return for accepting the same risk. That mark-to-market loss then appears somewhere else.

Pension funds own this debt. Insurance companies own it. Private credit funds own it. Asset managers own it. Banks finance some of those investors and provide funding to the SPVs themselves.

For example, PIMCO anchored the Hyperion financing. The significance is not that PIMCO made a bad investment. There is no evidence of that. The significance is that data-center credit is no longer confined to Silicon Valley balance sheets. It has been distributed into the financial system.

And the structures are getting larger.

The Financial Times reported this week that Big Tech companies are using residual-value guarantees and related structures to support as much as $300 billion of AI infrastructure exposure while keeping much of the associated borrowing outside their conventional corporate debt.

• Meta used such a structure for Hyperion.
• Nvidia has provided guarantees connected to data-center financing.
• Broadcom has provided guarantees supporting chip purchases.

The reason these arrangements exist is that the infrastructure requirements have become too large for even enormously profitable technology companies to fund exclusively through ordinary capital expenditures without changing their financial profiles.

But moving a liability does not eliminate the economic risk behind it, it simply changes who owns it.

The mistake in 2007 was not simply believing that houses would always appreciate. It was assuming risks that looked independent actually were independent. A mortgage in Florida and a mortgage in Nevada appeared geographically diversified until both depended on the same national credit cycle.

The AI infrastructure market deserves the same examination.

A data center in New Mexico and a data center in Louisiana look like different assets.

• Different developers.
• Different lenders.
• Different power markets.
• Different SPVs.
• Different bonds.

But trace the cash flows far enough and the same names keep appearing.

• Oracle.
• Meta.
• Microsoft.
• Amazon.
• OpenAI.
• Blue Owl.
• Nvidia.

The same handful of hyperscalers and AI companies are simultaneously customers, tenants, guarantors, equipment purchasers, investors and counterparties across hundreds of billions of dollars of infrastructure.

The diversification may therefore be considerably smaller than the number of projects suggests.

The same is true of the physical bottleneck. Every model assumes that the facility eventually gets electricity. Texas just halted new state-issued data-center permits while it audits the industry’s effects on the grid.

Utilities elsewhere are becoming more skeptical of enormous power requests. Exelon recently imposed stricter collateral requirements on prospective data-center customers. Its high-probability pipeline subsequently fell roughly 40%. AEP Ohio tightened its requirements and saw its pipeline fall by more than half.

Some of that is probably healthy. Developers routinely submit multiple power requests while evaluating competing locations. Requiring meaningful deposits removes speculative demand.

But it also tells us something important about the numbers used to justify the buildout.

• A gigawatt requested is not a gigawatt financed.
• A gigawatt financed is not a gigawatt constructed.
• A gigawatt constructed is not a gigawatt energized.
• An energized data center is not necessarily a profitable one.

There are trillions of dollars of projected AI infrastructure investment sitting between those distinctions. Now consider what happens if Jupiter is not isolated. One large project misses its power date. Its debt falls. Comparable project debt reprices. The next SPV has to pay another 100 basis points. Its developer needs another billion dollars of equity to make the financing work. The equity investor reduces its expected return or walks. The project is postponed. Equipment orders are pushed out. Revenue expected by the infrastructure supplier moves into another quarter. Its earnings estimates fall. Its stock falls. The market begins questioning the customer’s remaining commitments. The customer’s credit spread widens.

But that customer’s credit supports leases and guarantees sitting behind still more data-center financings.

Those bonds reprice. Suddenly the market is no longer asking whether AI needs more compute. It is asking who financed the compute already ordered. That is how an infrastructure problem becomes a credit problem.

And once it becomes a credit problem, the feedback loop runs in both directions.

Oracle is a useful example because its financial position is already becoming more dependent on capital markets. The company reported approximately $5 billion of negative free cash flow last quarter while simultaneously raising $20 billion through a stock sale. S&P downgraded Oracle in July to one notch above junk.

At the same time, Oracle reported $664 billion of remaining performance obligations. Those numbers are usually presented as opposing evidence. They shouldn’t be. The enormous backlog is precisely why Oracle requires enormous amounts of infrastructure. The enormous infrastructure requirement is precisely why Oracle requires enormous amounts of capital.

The question is whether the cash generated by fulfilling those contracts arrives quickly enough to support the capital required to fulfill them.

That is a duration problem. And duration problems become dangerous when the cost of money rises.

Imagine the market begins treating Oracle-linked data-center obligations as materially riskier. Oracle’s financing costs rise. Its equity falls. Lenders demand better terms. Projects become more expensive. Cash flow deteriorates further.

Now repeat the exercise across every company funding AI infrastructure.

The hyperscalers have strong balance sheets, but the scale matters. The Bank of England noted in July that hyperscaler bond issuance during the first half of 2026 had already exceeded issuance for all of 2025. It specifically warned that declining free cash flow was increasing dependence on future capital-market access while off-balance-sheet structures were spreading AI exposure throughout private credit, structured finance and asset-backed markets.

That is the part of the story I think the equity market is underestimating.

The bear case does not require AI to fail.

• It does not require OpenAI to disappear.
• It does not require people to stop using ChatGPT.
• It does not even require most data centers to be canceled.

It requires capital markets to decide that financing them should cost more.

AI infrastructure is being constructed under assumptions about completion dates, electricity availability, equipment values, customer credit and refinancing conditions. Change several of those assumptions simultaneously and the return on the underlying investment changes very quickly.

The first losses would probably appear exactly where they are appearing now: project debt trading below par. Then equity in highly leveraged developers. Then suppliers whose revenue depends on the projects. Then hyperscalers whose contractual commitments support the financing. Then the lenders and asset managers holding the credit.

At some point the distinction between an AI trade and a credit trade disappears.

The market has spent several years valuing companies partly on the assumption that enormous AI capital expenditures demonstrate enormous future demand. Those expenditures themselves support semiconductor revenue, utility investment, construction activity and data-center valuations.

If tighter credit causes the infrastructure buildout to slow, every participant begins losing someone else’s projected revenue.

A canceled data center reduces the developer’s investment. It also removes a utility load forecast, a construction contract, a transformer order, cooling equipment, networking equipment and potentially billions of dollars of accelerators. Those suppliers then revise their own investment plans. That is how a capital-expenditure boom works in reverse.

None of this establishes that such a cycle has begun, but Jupiter provides something the market did not have six months ago: a price. Eighty-nine to ninety-one cents on the dollar.

Watch the bonds.

Specifically, watch whether Meta’s Hyperion debt and other recently issued data-center securities begin widening in sympathy with Jupiter despite having no project-specific problem.

If they do, the market will be telling us that investors no longer view Jupiter as an isolated construction problem.

They will be repricing the asset class.

And if that happens while the industry still needs hundreds of billions of dollars of additional financing, the problem will no longer be whether America can build enough data centers.

It will be whether Wall Street is still willing to finance them at the price on which the boom was built.