The old adage goes that during a boom, the companies that profit most are the ones selling the picks and shovels. This January, even as consensus had largely settled around the idea that an AI bubble exists, something even Sam Altman acknowledged last August, Blackstone called investing in the “picks and shovels” of AI a “generational” opportunity. The safer bet, we’re told, lies not in the models themselves but in AI’s physical infrastructure: data centers, chips, and electricity. “The Real AI Talent War Is for Plumbers and Electricians,” declared a January headline in Wired.
Of the major players in artificial intelligence, a few might reasonably be considered picks-and-shovels companies. Nvidia, led by Jensen Huang, is one. Another is Oracle, which under Larry Ellison has spent the past year building some of the country’s largest AI data centers to provide computing power for companies like OpenAI.
But insofar as Oracle has been selling picks and shovels, enormous ones at that, it has also, over the past few months, come to be seen as a canary in the AI-bubble coal mine. In the roughly 10 months since September 2025, when Oracle signed a $300 billion deal with OpenAI that sent its stock soaring 36% in a single day, briefly making Ellison the world’s richest man, the company’s shares have fallen more than 43%, wiping out those gains. Meanwhile, the market for Oracle’s credit default swaps, which allow investors to bet on the possibility that the company could miss bond payments, has surged as its debt rating hovers just above junk status.
Since Oracle began building out Stargate, its sprawling data center campus in Abilene, Texas, the company has faced growing scrutiny over the highly leveraged financing behind both the project itself and its broader AI data center buildout.
The scale of the risk Oracle is taking is increasingly visible on its balance sheet. The company now carries more than $160 billion in outstanding liabilities—including $133 billion tied to the AI buildout, according to JPMorgan research cited by Barron’s—while holding less than $40 billion in cash and burning through money, according to its latest SEC filing.
Much of that gamble is tied directly to OpenAI. More than $300 billion of Oracle’s $553 billion in remaining performance obligations, or contracted revenue it has yet to collect, comes from OpenAI, a company that itself is reportedly losing billions. The dynamic has effectively turned Oracle into a public market stand-in for OpenAI. Investors unable to buy shares in the private company have, at least historically, treated Oracle as a proxy bet on OpenAI’s future success (and eventual IPO).
It should be said that among hyperscalers, the giant companies that own and operate data centers—like Meta and Amazon—this balance of cash, debt, and remaining performance obligations is not the norm. Oracle’s debt-to-equity ratio hovers around 415%, while none of the other hyperscalers top 80%.
This disparity makes a certain amount of sense: Oracle doesn’t have the same reserves of non-AI cash as companies like Amazon or Meta. While it has been a major player in tech for decades, its ascent to a nearly $900 billion market capitalization last September was driven almost entirely by hype, or rather its market equivalent: an OpenAI deal.
The past few months of Oracle’s downward swing have been marked by a few notable episodes. The company has reported weak quarterly financial results and faced a class-action lawsuit from bondholders alleging that it had misled them by claiming it would not need a “significant” amount of additional financing for its AI infrastructure buildout. We also got news that Blue Owl Capital, a private lender, had pulled out of financing a $10 billion data center project in Michigan. (Eventually Pimco and Blackstone stepped in.)
It was the Blue Owl retreat that was maybe the most worrying of all. Private lenders specialize in structuring bespoke deals that insulate them from risk; that flexibility is central to their business model. So if even they were stepping back from Oracle, despite the enormous fees attached to financing AI infrastructure, it raised a sharper question: What were they seeing that made the downside look too large to hedge against?
The rise of the shadow bank
The fact that so much of today’s financial activity takes place in the private markets has led some observers to speculate that the data center boom, the very “picks and shovels” of AI, is, at least for the moment, propping up another industry caught in the midst of its own bubble: private credit.
The private credit (or lending) bubble has received far less attention in the nonfinancial press than the AI one, though speculation about its eventual bursting has circulated for years. More recently, a selloff in software companies once considered safe bets—the so-called “SaaSpocalypse,” driven by fears that AI could make enterprise software obsolete—brought the issue back into focus.
For context, private credit refers to nonbank loans made to private companies. “Alternative asset managers,” as firms in the industry call themselves, specialize in this kind of lending. Think Apollo Global Management, Blackstone, or Blue Owl Capital: massive financial institutions that operate, in many ways, like banks. They’re frequently referred to as “shadow banks.”
In the years following the Great Recession of 2007-09, after which the big banks were told they had to hold more money on their books via the Dodd-Frank Act, lending shifted hands. Private nonbank financial institutions (NBFIs), as they’re often called, had suddenly hit the jackpot: They weren’t subject to the same capital requirements as banks, and they could structure deals with far greater flexibility.
That trend has only picked up steam. Over the past five years, private lending has come fully into its own, with the amount of private credit in circulation more than tripling to roughly $3 trillion. (Unfortunately for those involved, the transformation of firms like Apollo has come with some added bureaucracy. “We are becoming a bank. It truly sucks,” an Apollo executive told the Financial Times in late 2025.)
The worry in the financial press has been that, absent regulation, private lenders would become too willing to forgo necessary due diligence on deals—that they’d take borrowers at their word rather than look closely under the hood. Recently, those fears have begun to materialize, with two recent high-profile bankruptcies sending shockwaves through the private credit world. One involved the auto lender Tricolor; the other, the auto parts company First Brands. In both cases, lenders—private and public alike—failed to properly sound the alarm. Billions of dollars in investor money vanished seemingly overnight. Those worries carried on into the winter and spring: Blackrock, Blackstone, Apollo, and Blue Owl have all halted redemptions on various funds as investor anxiety has spread.
But even as fears surrounding private credit have further materialized—stock in Blue Owl has dropped nearly 50% over the past year—industry executives have argued that they are still finding highly sought-after returns, particularly in data center development. And when critics reflexively invoke “the bubble,” their response is often a patronizing reassurance: that these financial products are structured specifically to avoid risk. Appearing on the Odd Lots podcast in January, Michael Zawadzki, global chief investment officer at Blackstone Credit & Insurance, insisted that lenders are financing “15- or 20-year take-or-pay contracts” in which they receive “a fixed sum every single month” from tenants—“no matter usage . . . operating costs.”
In search of cockroaches
The notion that private credit firms are particularly adept at structuring leases, and that they primarily target the most cash-rich companies in the world, might paint Oracle as something of an outlier. In that case, Blue Owl Capital’s decision to pull out of the data center deal is not a sign of things going wrong, but of things going right—of the prudence of the private markets, of the virtues of financialization. Such has been the narrative of our favorite neoliberal financiers for decades. “Democratizing capital,” the junk-bond king and private-credit forebear Michael Milken wrote in 2000 for California Lawyer, “has encouraged the growth of such new financial entities . . . to challenge the dominance of the few large banks and insurance companies that used to decide who received financing.”
Along similar lines is another story you’ll hear quite often: that there will be some winners and some losers, and that the job of lenders is to pick them. Bad apples, it follows, are a natural consequence of the market economy. This is the story frequently told about the dot-com bubble (e.g., Google won, Pets.com lost) and one expressed by many financial commentators today vis-à-vis the perceived AI bubble. Yes, Oracle might have a cash problem, but Google or Anthropic will reap enough benefits that further macroeconomic growth will be possible and progress can be maintained. Hence we shouldn’t give up on AI. (The same sort of logic holds true for lenders; Lehman Brothers was a bad apple, worth making an example of, but Wells Fargo was an honest Main Street bank.)
And much hay is made about the AI ouroboros, the image commentators like to use to describe how ginormous companies like OpenAI, Nvidia, AMD, CoreWeave, Broadcom, Oracle, and even the other hyperscalers are paying each other to keep the hype going while not turning a substantial profit. It’s reminiscent of earlier speculative loops, like the telecom or railway bubbles. It’s why it’s hard to tell who the good apples are, per se, or whether there are any at all. The unprofitable OpenAI, for one, seems to undergird nearly everything: In November 2025, Sam Altman revealed on X that OpenAI has about $1.4 trillion in multiyear infrastructure commitments through 2033. (Although, CNBC reported in February that OpenAI has told investors it’s targeting $600 billion in compute spend by 2030; whether those numbers are incongruous is to be determined.)
That OpenAI is looking for new ways to find profit—some it once claimed it would avoid, like advertising—hasn’t yet seemed to scare off sophisticated investors who are more worried about missing the final growth stretch than getting mauled in a crash.
Over the past year, some investors have announced that they are hedging their AI bets. In November, for example, Deutsche Bank said it was exploring ways to manage its data center exposure—a relatively simple statement that nonetheless sent Main Street market-watchers into a millenarian frenzy. But what this means for private credit is more opaque: The concern is that AI exposure has delayed, and perhaps magnified, a future private-credit crash.
Over $450 billion has been poured into the tech sector through private credit as of late 2025, per Bloomberg—which is a large number, yes, but likely still understates their exposure. Robert Dodd, an analyst at the investment bank Raymond James, told Bloomberg in February that private creditors classify their loans to software companies based on their end markets. “If your software business is in healthcare, the fund classifies it as healthcare exposure,” he said. “The software exposure is meaningfully higher than it looks.”
The sheer scale of the AI buildout has required such staggering amounts of capital that hyperscalers, even the cash-rich ones, have turned to more unconventional fundraising methods. One of the most prominent is the special purpose vehicle, or SPV, which is, at its core, fairly simple (though modifications can be made): These financial vehicles are used to raise money for data center projects while keeping debt off companies’ balance sheets.
First, a holding company, the SPV, is created by a hyperscaler (or by a developer working on its behalf), then banks and private creditors lend to the SPV—not the hyperscaler—in order to finance the development of the data centers. The hyperscaler then leases the finished facility back. Oracle has leaned on SPVs more heavily than perhaps anyone else, for an obvious reason: the debt sits on the SPV’s books rather than the hyperscaler’s.
For lenders, these arrangements offer a way to earn attractive returns from blue-chip tenants on long-term leases. A company like Meta is a particularly appealing customer, one willing to pay higher private-market rates while also sitting on enormous cash reserves. That contradiction is reflected in the disconnect between debt ratings and yields. The debt Meta used to finance its planned flagship hyperscale data center, Hyperion, received an A+ rating from S&P Global, despite the fact that its yield looked something more like a high-yield or “junk” bond. (Also of note: According to the Financial Times, some of these SPV deals are being cut up, pooled together and resold to investors as new asset-backed securities.)
One of the catches for lenders concerns the collateral. If a hyperscaler defaults, lenders’ claims are limited to the SPV’s underlying assets, usually the data center itself, including its real estate and chips. That is why many of these deals include residual value guarantees (or RVGs), under which the hyperscaler agrees to compensate lenders if the value of the property drops significantly. Meta’s RVG on Hyperion, for example, is worth roughly $28 billion, according to the company. While the RVG shifts some risk back onto the hyperscaler, it too remains off the balance sheet as a liability, unless the probability of the property value collapsing becomes significant—generally understood as greater than 50% odds, as determined by an auditor hired by the hyperscaler.
Even though this exposure doesn’t appear on hyperscalers’ balance sheets, they’re still ultimately on the hook. That was much of the basis for the proposed class-action lawsuit filed in January by a group of Oracle bondholders led by the Ohio Carpenters’ Pension Fund, who allege that the company misled investors about the scale of the debt it was preparing to take on. Documents tied to Oracle’s $18 billion bond sale in September 2025 said the company “may” need to borrow more. Seven weeks later, Oracle returned for an additional $38 billion in loans to finance data centers tied to its OpenAI contract. The spike in Oracle’s credit default swaps followed soon after.
The collateral itself is another problem. If AI demand has been even a bit overestimated, then the value of these assets could plummet far enough to wipe out many of the parties involved, even with RVGs in place. A February report from the ratings agency Moody’s flagged that hyperscalers—not OpenAI, but the public companies Amazon, Meta, Alphabet, Microsoft, and Oracle—have amassed more than $662 billion in off-balance-sheet commitments, more than all the debt on their balance sheets.
And irrespective of demand collapse, the collateral is itself unusual and worth taking stock of. There’s, for one, the fact that top-of-the-line chips might last only a few years considering they’re being run 24/7 for training and that new chips are constantly being developed, making old ones obsolete. Those chips are themselves now being collateralized. CoreWeave, for example, has built a multibillion-dollar business in part by turning these relatively short-lived chips into financial products. The Information, as part of a series of predictions for 2026, suggested that Oracle may soon do the same. (That the financing and construction boom is racing ahead of the physical infrastructure needed to support it only makes the situation more tenuous.)
Forecasts are as dramatic in the world of private credit. Morgan Stanley research, published in 2025 and cited by Apollo on its own website, predicted that private credit would contribute $800 billion in financing AI infrastructure over the next three years. Presumably, much of that has already been deployed.
To assume that the private credit bigwigs were performing up to snuff, and not chasing one more rush, or perhaps offloading the risk to pension funds and insurers—would be naïve. In February, in an effort to find cash to return to weary investors, Blue Owl sold $1.4 billion of its debt to an insurer it owned at near par. “They made an arm’s-length economic decision,” the investment firm’s copresident, Craig Packer, said over an earnings call. “There is nothing behind the scenes that would in any way undermine that conclusion.”
The words of JPMorgan CEO Jamie Dimon, delivered after the Tricolor and First Brands crises, have loomed large over Wall Street for the past eight months: “When you see one cockroach, there are probably more.”
Is AI ‘the savior we need’?
A late-January report from TD Cowen found that banks and bank-like entities had begun pulling back from Oracle. More and more Oracle layoffs have followed, partially in an effort to mollify investors—though the cuts themselves have also been used to finance further AI investment. TD Cowen’s research “indicates that multiple Oracle data center leases that were under negotiations with private operations struggled to secure financing, in turn preventing Oracle from securing the data center capacity via lease.” One being Blue Owl Capital, of course.
What’s revealing here, however, is that Oracle is not merely a passive harbinger of a forthcoming bubble burst; it’s itself becoming an agent in the popping. If Oracle is having trouble building data centers, both in terms of financing, but also in terms of literal construction—a separate subject Bloomberg has reported on—and it can’t deliver on what OpenAI needs, it only places more stress on a company that’s already stretched thin. The fact that Oracle has $66 billion in SPV commitments—most of any of the hyperscalers—makes matters worse for everyone involved.
The disconnection between the market performing well (or even merely holding steady) and the widespread sense that some sort of crisis is coming can feel disconcerting, even from a distance. Knowledge that the parabolic growth of the Magnificent Seven stocks—Apple, Alphabet, Amazon, Meta, Microsoft, Nvidia, Tesla—is carrying the American economy barely changes that. Will there be a crisis and a bailout similar to 2008? OpenAI got stuck in a still-ongoing media imbroglio in November when its CFO suggested that the government could backstop their investments in order to help with financing, a statement Altman had to quickly repudiate. But the company’s chief executive has nonetheless refused to promise that the firm won’t pursue such backstops. (Will Oracle, captained by Trump-connected Larry Ellison, be bailed out somehow?)
There’s even the question of whether the AI-leveraged tech giants who’ve dipped into the debt worlds recently, like Meta or Microsoft, can restrict their exposure. On January 29, Microsoft stock had the second-largest single-day dollar decline in U.S. stock market history, trailing only Nvidia’s last year. The drop-off was, in large part, a product of slower-than-expected growth as well as the announcement that Microsoft would be growing its Oracle-sized commitments to OpenAI. That same day, Oracle’s stock cratered, despite Blackstone’s announcing its interest in taking a stake in the Michigan project that Blue Owl pulled out of.
The precarity of the current situation harkens back to decades prior, when we were looking for answers to other macroeconomic—and, importantly, political—crises. The inevitable end of Keynesian capitalism in the ’70s was answered by monetarism and neoliberalism, while the exact nature of their end remains to be seen. During that inflection point—from President Ford, through Carter, to Reagan—despair struck similarly to today. Immortalized in political theorist Melinda Cooper’s Counterrevolution: Austerity and Extravagance in Public Finance, were the words of the famed economist Arthur Laffer, then a Ford adviser, “Perhaps the solution to our doomsday problem is the exact opposite of the solution found at the end of the first millennium . . . We need the appearance of God.”
Calls for deliverance remain today. This January, at Davos, the hedge fund magnate Ken Griffin called for divine intervention too. “Within the private sector,” Griffin said, “the question is will AI create the productivity acceleration . . . to overcome the profligate spending we’re currently engaged in.” He finished, “The world needs a savior, and the hope is that AI is the savior we need.”
People used to joke in Silicon Valley, when Oracle was starting up, that “the difference between God and Larry [Ellison] is that God doesn’t think he’s Larry.” (This is also the title of his biography.) Last year, President Trump even called him “a sort of CEO of everything.” And Larry’s dominion has only further grown; see his high-profile media ventures, like Oracle’s purchase of new stock in TikTok, or his son’s Warner Bros. takeover—which he effectively bankrolled. Yet even those deals carry significant risks: Both the TikTok and Warner Bros. buyouts were financed heavily with Gulf spending, including by Gulf state sovereign wealth funds. With the impact of the Iran war on Gulf economics, concerns have arisen about the debt raised by the Ellisons for the Warner Bros. deal, which has yet to be fully closed.
In looking at Larry Ellison, it almost feels too obvious to compare either the man or his company to Icarus—as easy a metaphor as it is to reach for, given, for one, Oracle’s debt-to-equity ratio. It’s better instead to think of the aptly named company through the frame of Babylon: as an exercise in hubris whose eventual collapse sent wide-ranging, disorienting, and destructive shockwaves through everything around it.