The first cracks are showing up where the money was supposed to flow most freely. Around $18 billion of loans tied to Oracle’s Project Jupiter data center in New Mexico are under pressure, according to the Financial Times, with banks in the syndicate, including Santander and Jefferies, marking the debt at 89 to 91 cents on the dollar and struggling to sell more of the paper to outside investors.
The loans were meant to be the easy part. Project Jupiter is part of Oracle’s $300 billion contract to supply computing power to OpenAI, which in turn anchors the $500 billion Stargate program, the largest single bet on artificial-intelligence infrastructure ever assembled. Lending against a campus with a signed customer like that was supposed to be a straightforward trade. Instead, the debt is piling up on bank balance sheets.
The problem is not that the customer has vanished. It is that the arithmetic of the whole enterprise has started to strain. Oracle has guided to as much as $95 billion of capital spending in its fiscal 2027, while telling investors it expects to recover no more than $25 billion from customers over the same horizon. The gap between what the company is spending and what it can demonstrate coming back is the number the credit market is now staring at.
The rating agencies noticed before the banks did. S&P cut Oracle’s credit rating in July, leaving it one notch above junk, and the markdowns on the Jupiter loans are the market arriving at the same conclusion the agency reached. A company that must keep spending to defend its position, with an uncertain path to getting that spending back, is a harder credit to hold than a tech giant with cash pouring in.
OpenAI’s own numbers sharpen the picture. In a private presentation to investors in July, according to the Financial Times, the company projected negative free cash flow of $278 billion between 2026 and 2030, with cumulative spending on computing and infrastructure reaching roughly $856 billion by the end of 2030. Revenue was projected to climb from $36 billion to $350 billion over the same period, a trajectory that would be extraordinary and, if it fell short, catastrophic for the financing built on it.
The cash on hand does not stretch to the plan. OpenAI raised $122 billion in March 2026, an amount that once looked like a cushion. At the current rate of spending, the presentation implied, that capital would be exhausted by 2028, leaving the company to return to markets that are already showing signs of fatigue.
The fatigue is measurable across the sector. The five largest U.S. cloud companies have guided to more than $750 billion of combined capital spending in 2026, a figure with no precedent. JPMorgan estimates that investment-grade credit markets will need to supply more than $2.1 trillion of data center financing over five years, a demand that assumes the banks keep absorbing loans the way they did when the boom began.
What has changed is the willingness to hold the risk. The syndicated loan market exists to distribute debt, and the Jupiter episode shows the distribution seizing up: the paper is marked down, the resale to a wider investor base has stalled, and more of it is stuck on the books of the banks that originated it. That is the market sending a signal that the revenue assumptions underneath the debt have started to look optimistic.
None of this means the buildout stops. The contracts are signed, the construction is under way, and the strategic logic that drives the spending, that whoever controls the computing power controls the next era of software, has not changed. But the financing is no longer an afterthought to the engineering, and it was once.
The question the credit market is asking is whether the revenue curves the companies have drawn will arrive on schedule. If they do, the debt will be repaid and the markdowns will reverse. If they do not, the loans on the books today are the first installment of a much larger reckoning, and the banks holding them will be the first to pay it.
The Stargate program, announced in 2025, was always a financing experiment as much as a construction one. Its backers said the money would come from a rotating cast of investors, lenders, and partners, with the computing power sold forward to pay them back. That model worked while the demand for AI capacity was still a promise. It is now a quarterly number, and the number has not yet grown fast enough to make the debt behind it look safe.
Other borrowers are watching. The data center sector has become one of the largest consumers of corporate debt in the country, and if the biggest, best-connected project in the sector cannot distribute its loans at par, smaller and less proven projects will face a harder market. The banks that arranged the Jupiter financing have an interest in the loans finding a home, but their markdowns suggest they are no longer willing to pretend the risk has not changed.
For now the boom continues on paper and on the ground, but its financing has begun to slow, and that is the earliest and most reliable warning the market has. The spending was always going to be enormous. The open question was whether the world’s banks would keep funding it, and the first answer is arriving.
