A company that most people have never heard of is among the year’s best-performing technology firms—and a symbol of the complex, interconnected, and potentially catastrophic ways in which AI companies do business these days.

CoreWeave’s IPO in March was the largest of any tech start-up since 2021, and the company’s share price has subsequently more than doubled, outperforming even the “Magnificent Seven” tech stocks. On Wall Street, CoreWeave is regularly referred to as one of the most important companies powering the AI revolution. In the past few months, it has announced a $22 billion partnership with OpenAI, a $14 billion deal with Meta, and a $6 billion arrangement with Nvidia.

Not bad for a former crypto-mining firm turned data-center operator with zero profits and billions of dollars in debt on its books.

CoreWeave’s business model consists of buying up lots of high-end computer chips, and building or leasing data centers to house those chips. It then rents out those assets to AI companies that need computing power but prefer not to take on the huge up-front costs themselves. If this is straightforward enough, CoreWeave’s financial situation is anything but. The company expects to bring in $5 billion in revenue this year while spending roughly $20 billion. To cover that gap, the company has taken on $14 billion in debt, more than half of which comes due in the next year. Many of these loans were issued by private-equity firms at high interest rates, and several use complex forms of financial engineering, such as giving the money to newly formed legal entities created for the explicit purpose of borrowing on CoreWeave’s behalf (more on that later). CoreWeave also faces $34 billion in scheduled lease payments that will start kicking in between now and 2028.

From the May 2025 issue: The new king of tech

The money that CoreWeave is making, meanwhile, comes from just a few intimately connected sources. A single customer, Microsoft, is responsible for as much as 70 percent of its revenue; its next biggest customers, Nvidia and OpenAI, might make up another 20 percent, though exact numbers are hard to find. Nvidia is also CoreWeave’s exclusive supplier of chips and one of its major investors, meaning CoreWeave is using Nvidia’s money to buy Nvidia’s chips and then renting them right back to Nvidia. OpenAI is also a major CoreWeave investor and has close financial partnerships with both Nvidia and Microsoft.

All of this might make CoreWeave the purest distillation of a trend sweeping through the AI sector. In recent months, tech giants including Amazon, Google, Meta, Microsoft, and Oracle have been making gargantuan investments in new data centers, tying together their fortunes through circular financing deals, and borrowing huge piles of debt from lightly regulated lenders. The companies and their most ardent backers argue that these deals will set them up to capture the limitless profits of the coming AI revolution. But the last time the economy saw so much wealth tied up in such obscure overlapping arrangements was just before the 2008 financial crisis. If the AI revolution fails to materialize on the scale or the timeline that the industry expects, the economic consequences could be very ugly indeed.

The extreme financialization of the AI sector reflects a simple reality: The infrastructure required to train and run AI systems is so expensive that not even the largest companies have enough cash to pay for it all. Spending on data centers is conservatively projected to exceed $400 billion this year, roughly the size of the economy of Denmark; McKinsey estimates that it will reach nearly $7 trillion by 2030. Creative measures are necessary to pay for all of this investment.

At the center of the action is Nvidia, the world’s most valuable company. Companies that train and run AI systems, such as Anthropic and OpenAI, need Nvidia’s chips but don’t have the cash on hand to pay for them. Nvidia, meanwhile, has plenty of cash but needs customers to keep buying its chips. So the parties have made a series of deals in which the AI companies are effectively paying Nvidia by handing over a share of their future profits in the form of equity. The chipmaker has struck more than 50 deals this year, including a $100 billion investment in OpenAI and (with Microsoft) a $15 billion investment in Anthropic. Formally, these transactions don’t obligate the AI companies to spend money on Nvidia’s chips—an Nvidia spokesperson told Bloomberg that the company “does not require any of the companies we invest in to use Nvidia technology”—but in practice, that’s where the money goes.

OpenAI has made its own series of deals, including agreements to purchase $300 billion of computing power from Oracle, $38 billion from Amazon, and $22 billion from CoreWeave. Those cloud providers, in turn, are an important market for Nvidia’s chips. OpenAI has also invested in several smaller AI start-ups, which in exchange have agreed to pay for ChatGPT enterprise accounts. Even when represented visually, the resulting web of interlocking relationships is almost impossible to track.

Together, these arrangements amount to an entire industry making a double-or-nothing bet on a product that is nowhere near profitable. A single company, OpenAI, is simultaneously a major source of revenue and investment for several cloud companies and chipmakers; a close financial partner to Microsoft, Oracle, and Amazon; a significant customer for Nvidia; and a leading investor in AI start-ups. And yet the company is projected to generate only $10 billion this year in revenue—less than a fifth of what it needs annually just to fund its deal with Oracle. It is on track to lose at least $15 billion this year, and doesn’t expect to be profitable until at least 2029. By one estimate, AI companies collectively will generate $60 billion in revenue against $400 billion in spending this year. The one company that* is* making a lot of money from the AI boom, Nvidia, is doing so only because everyone else is buying its chips in the hopes of obtaining future profits.

The AI companies and their boosters see this as a gamble worth taking. Demand for AI services, they point out, is growing at an exponential rate. According to calculations by Azeem Azhar, a widely cited AI-industry analyst, the direct revenues from AI services have increased nearly ninefold over the past two years. If that pace continues, then it’s only a matter of time before AI companies will begin making record-shattering profits. “I think people who fixate on exactly how these investments are being financed are stuck in an outdated way of thinking,” Azhar told me. “Everyone is assuming that this technology will improve at a linear pace. But AI is an exponential technology. It’s a whole different paradigm.”

If, however, AI does not produce the short-term profits its proponents envision—if its technical advances slow down and its productivity-enhancing effects underwhelm, as a mounting body of evidence suggests may be the case—then the financial ties that bind the sector together could become everyone’s collective downfall. The extreme concentration of stock-market wealth in a handful of tech companies with deep financial links to one another could make an AI crash even more severe than the dot-com crash of the 2000s.

And a stock-market correction might be the least of America’s worries. When equity investments go bad, investors might lose their shirts, but the damage to the real economy is typically contained. (The dot-com crash, for example, didn’t cause mass unemployment.) But the AI build-out is so expensive that it can’t be funded by equity investments alone. To finance their investments, AI companies have taken on hundreds of billions of dollars in debt, a number that Morgan Stanley expects to rise to $1.5 trillion by 2028. When a bunch of highly leveraged loans go bad at the same time, the fallout can spread throughout the financial system and trigger a major recession.

The AI sector’s debt is, of course, not guaranteed to go bad. But the complex way in which it is arranged and packaged isn’t reassuring. For instance, earlier this year, Meta decided to build a new data center in Louisiana that will cost $27 billion. Instead of applying for a loan from a traditional lender, the company partnered with Blue Owl Capital, a private-equity firm, to set up a separate legal entity, known as a special-purpose vehicle, or SPV, that will borrow the money on Meta’s behalf, build the data center according to Meta’s instructions, and then lease it back to Meta. Because Blue Owl is technically the majority owner of the project, this setup keeps the debt off of Meta’s balance sheet, enabling the company to keep borrowing at low interest rates without worrying about a hit to its credit rating. Other companies, including xAI, CoreWeave, and Google, have borrowed or plan to borrow huge sums through similar kinds of arrangements.

Meta has described its arrangement with Blue Owl as an “innovative partnership” that is “designed to support the speed and flexibility required for Meta’s data center projects.” But the reason the credit-rating system exists is to give lenders and investors a clear sense of the risk they are taking on when they issue a loan. A long history exists of companies trying to circumvent that system. In the run-up to the 2008 financial crisis, several major financial institutions used SPVs to keep billions of dollars in household debt off of their balance sheets. Enron, the energy corporation that famously collapsed in 2001 after a massive accounting scandal, used SPVs to mask its shady accounting practices. “When I see arrangements like this, it’s a huge red flag,” Paul Kedrosky, a managing partner at SK Ventures and research fellow at MIT who has written extensively about financial-engineering techniques, told me. “It sends the signal that these companies really don’t want the credit-rating agencies to look too closely at their spending.”

SVPs aren’t the only 2008-era financing tool making a comeback. Data-center debt totaling billions of dollars is being sliced up into “asset-backed securities,” which are then bundled and sold to investors. This is not an inherently problematic way for companies to fund their borrowing. But Kedrosky argues that during periods of heightened speculation, these vehicles turn debt into a financial product whose worth is disconnected from the value of the underlying asset it represents—which can encourage reckless behavior. “Investors see these complex financial products and they say, I don’t care what’s happening inside—I just care that it’s highly rated and promises a big return,” Kedrosky said. “That’s what happened in ’08. And once that kind of thinking takes off, it becomes really dangerous.”

Rogé Karma: Just how bad would an AI bubble be?

Then there are the so-called GPU-backed loans. Several data-center builders and cloud providers, including CoreWeave, have obtained multibillion-dollar loans to purchase chips by posting their existing chips as collateral, just as many homeowners used their homes as collateral to take out loans for second and third homes in the 2000s. But, as Advait Arun, an analyst at the Center for Public Enterprise, notes in a recent report on the AI sector’s finances, whether that collateral will hold its value is far from clear. When new chip models are released, the value of older models tends to fall. According to Arun, if the collapse in chip prices were steep enough, a vicious cycle could ensue. As older chips fall in value, any loan using those chips as collateral suddenly becomes at risk of default. Lenders might respond by calling in their loans early, before companies have the revenue to pay them back. At that point, the lender might try to sell the chips to recoup their investment, but that will only flood the market with even more chips, driving down the values of existing chips even further, causing other lenders to call in their loans and so on. “A few months ago I would have told you that this was building toward a repeat of the dot-com crash,” Mark Zandi, the chief economist at Moody’s Analytics, told me. “But all of this debt and financial engineering is making me increasingly worried about a 2008-like scenario.”

The federal government responded to the 2008 crisis by limiting the ability of traditional banks to take on big, risky loans. Since then, however, private-equity firms, which aren’t subject to the same regulatory scrutiny as banks, have gotten more heavily into the lending business. As of early this year, these firms had lent about $450 billion in so-called private credit to the tech sector, including financing several of the deals discussed above. And, according to one estimate, they will lend it another $800 billion over the next two years. “If the AI bubble goes bust, they are the ones that will be left holding the bag,” Arun told me.

A private-credit bust is almost certainly preferable to a banking bust. Unlike banks, private-equity firms don’t have ordinary depositors. In theory, if their loans fail, the groups that will be hurt the most are institutional investors, such as pension funds, university endowments, and hedge funds, limiting the damage to the broader economy. The problem is that nobody knows for certain that this is the case. Private credit is functionally a black box. Unlike banks, these entities don’t have to disclose who they are getting their money from, how much they’re lending, how much capital they’re holding, and how their loans are performing. This makes it impossible for regulators to know what risks exist in the system or how tied they are to the real economy.

Evidence is growing that the links between private credit and the rest of the financial system are stronger than once believed. Careful studies from the Federal Reserve estimate that up to a quarter of bank loans to nonbank financial institutions are now made to private-credit firms (up from just 1 percent in 2013) and that major life-insurance companies have nearly $1 trillion tied up in private credit. These connections raise the prospect that a big AI crash could lead to a wave of private-credit failures, which could in turn bring down major banks and insurers, Natasha Sarin, a Yale Law School professor who specializes in financial regulation, told me. “Unfortunately, it usually isn’t until after a crisis that we realize just how interconnected the different parts of the financial system were all along,” she said.

An AI-induced financial disaster is far from inevitable. Still, given the warning signs, one would hope for the federal government to be doing what it can to reduce the risk of a crisis. Instead, the Trump administration is doing the opposite. In August, the president signed an executive order that instructs federal agencies to loosen regulations so that ordinary 401(k) holders can invest directly in “alternative assets” such as, yes, private credit, a change that could expose a far broader swath of the public to the fallout if AI loans go bad. Perhaps that is the key difference between 2008 and 2025. Back then, the federal government was caught off guard by the crash; this time, it appears to be courting one.