Data-center operator CoreWeave is a stock-market darling. Bears see its finances as emblematic of an AI infrastructure bubble

A vast data center in Plano, Texas, is a symbol of the enormous AI infrastructure boom that has boosted stock markets and driven U.S. economic growth over the past year. The data center occupies more than 450,000 square feet and cost $1.6 billion to construct and equip. It supplies 30 megawatts of computing power to train and run AI models. Yet the company that runs it is a leading candidate to be ground-zero for a future AI financial meltdown.

The data center is one of dozens around the world operated by CoreWeave, a company that develops and manages data centers and sells their computing capacity to technology companies. Its business is at the center of the AI economy—providing computing power to meet the voracious demand of the likes of Microsoft and OpenAI. But CoreWeave doesn’t own the Plano facility, nor does it own most of the data hubs it’s operating. And that is a part of the problem.

The company is built, by its own admission, on a mountain of debt—obligations it has piled up as it races to build out a network of server farms for its customers. And that mountain looms far larger than the piles of cash that CoreWeave has brought in the door so far. When the company announces earnings on Monday, bulls and bears alike will be watching to see how its revenue is growing, and whether it has been able to pare its losses. CoreWeave’s earnings are likely to be a bellwether for the state of the entire AI boom, and for the industry’s massive and expensive infrastructure buildout in particular.

CoreWeave has $7.6 billion in current liabilities—bills that fall due within 12 months—on its balance sheet, and $11 billion in debt overall, according to its most recent quarterly earnings report, filed in August. Coming from a tech giant like Google or Microsoft with tens of billions in free cash flow, such numbers wouldn’t raise an eyebrow. But CoreWeave’s revenues were only $1.9 billion in 2024. On its Q2 earnings call, CEO Michael Intrator told analysts that full year 2025 revenue would land between $5.15 billion to $5.35 billion.  On the same call, the CEO said he expected CoreWeave’s capex for the year would total between $20 billion and $23 billion.

Those short-term figures pale beside a bigger and potentially more onerous obligation that isn’t on its balance sheet: the $34 billion in scheduled lease payments that will start kicking in between now and 2028. Many of those payments are stretched over relatively long terms, of 10 years or more. Still, some of this is for data centers and office buildings that have not yet begun to operate or bring in revenue—representing a vulnerability if any of the as-yet-unprofitable startups CoreWeave sells computing services to are unable to meet their contractual obligations, or if construction delays mean CoreWeave is not able to provide capacity on time, allowing customers to cancel contracts.

Story Continues

In a sense, Coreweave is a metaphor for the broader AI industry at the present moment, as top companies commit to enormous capex spending today in the confidence that it’ll be justified by future revenue from AI platforms and services. Investors appear to be broadly convinced by the company’s narrative: CoreWeave’s stock price is up 160% since the company’s IPO in March.

But Fortune’s analysis of CoreWeave’s filings with the Securities and Exchange Commission, which are laced with warnings and caveats, show how risky the company’s business model could be. In interviews with several analysts, bulls and bears agreed that CoreWeave’s fundamentals, as reflected in its filings, don’t currently add up. “A lot has to go right,” says Thomas Blakely, a managing director of software equity research at Cantor Fitzgerald, who rates CoreWeave “overweight.” (Of 26 equity analysts who currently cover the stock, 14 had the equivalent of buy or outperform ratings on the shares, while nine had “hold” ratings, and three had “underperform” or “sell” recommendations, according to data from S&P Market Intelligence.)

The bears see CoreWeave as a strong candidate to find itself underwater with its mounting liabilities, making it potentially the first domino to fall in the AI ecosystem. “To say they’ll scale out of this is questionable,” says Gil Luria, the head of technology research at the investment firm D.A. Davidson. “I don’t see how it becomes more profitable.” He believes that the likeliest outcome for CoreWeave on a five-year time horizon is bankruptcy—either because its current customers will be able to rely on their own infrastructure by then, or because an increasingly stretched CoreWeave will no longer be able to borrow.

In a statement to Fortune, a company spokesperson said that “CoreWeave’s capital structure and financial performance are strong and underpinned by long-term take-or-pay contracts signed with the world’s leading enterprises and AI labs who partner with CoreWeave because we deliver the best AI cloud.” The statement went on to say that the company structured its contracts to “support and repay any related debt obligations while generating additional free cash flow. CoreWeave operates in a supply-constrained market where demand far exceeds capacity, and our hyper-growth is evidence of the trust leading companies place in us to power their most critical AI workloads.”

Betting big on tomorrow’s revenue

With the company continuing to make huge new spending commitments even as it books new future revenue, AI investors’ attention will be glued to its upcoming quarterly earnings report. One number that everyone will be watching is CoreWeave’s “remaining performance obligations,” or RPOs—essentially revenues that CoreWeave has booked but that have not yet been paid. (Like its scheduled lease payments, CoreWeave’s RPOs are excluded from its balance sheet.)

If CoreWeave is booking the kinds of contracts that will pull it out of debt sooner rather than later, the RPOs level—and the forecast for how quickly those future bookings are likely to turn into actual cash—are where they would show up. The company has announced several major new deals since its last quarterly earnings announcement, including a $14.2 billion agreement to supply Meta with computing capacity, and an pact with AI startup Poolside for a data center stuffed with 40,000 Nvidia GPUs. So it is likely its RPO total will climb significantly.

Bulls argue that this is exactly how the boom will play out in CoreWeave’s favor: The revenue will come through, in great quantity, and its scale will solve the company’s problems by catching up with and then outpacing its capital expenditures. In that scenario, the company becomes the next Levi Strauss or Amazon Web Services, providing the “picks and shovels” of the AI boom, and getting filthy rich. “The potential is beyond the scope of our imagination at this point,” says Kevin Dede, a senior technology analyst at the financial services firm H.C. Wainwright, which rates CoreWeave a “buy.”

But for now CoreWeave is miles away from being profitable and is bleeding cash, absent its ability to issue debt. The RPOs it reported in its most recent earnings that are likely to be realized in the next 12 months are not, on their own, sufficient to cover its current obligations and announced growth plans (more on that shortly). The company has razor-thin operating margins—1.6% in the past quarter. After accounting for its large interest expenses, those margins turn sharply negative. The company lost more than $600 million on $2.2 billion in revenue in the first six months of 2025. “That’s not great,” Luria says. “Is there any way that gap closes?”

Barring an enormous surge in revenue over the next 12 months or so, the company will likely need to borrow more money, or renegotiate with creditors, in order to cover the obligations already on its books. To be sure, the AI boom could deliver that revenue surge—but even slight weakening in spending growth across the industry could hit CoreWeave disproportionately. Kerrisdale Capital, an investment management firm that is shorting the stock, is pithy in its conclusions: CoreWeave, it wrote in a September report, is  “the poster child of the AI infrastructure bubble.”

A pivot to AI, fueled by debt

CoreWeave began as a crypto-mining company, a side project of a few friends who were hedge-fund traders. Crypto mining, like AI, relies heavily on graphic processing units, or GPUs, with the chips racing to solve complicated algorithms that spit out currency rewards for correctly verifying blockchain transactions, and CoreWeave was a steady buyer.

Over time, Brian Venturo, one of the company’s founders, realized that the rise in AI would be a major factor fueling the surge in demand for the computing power of the GPUs that CoreWeave was already accumulating. Beginning in October 2021, CoreWeave entered into two deals with asset management firm Magnetar Capital, raising first $50 million in convertible notes, and then a year later, an additional $125 million, also in convertible notes. The company used nearly all of it to buy GPUs from Nvidia. Over the next few years, CoreWeave would secure billions of dollars in a combination of debt and equity, building out a sprawling array of data centers across the U.S., and eventually expanding to the U.K.

This March, CoreWeave’s IPO made it one of the closest things to a pure-play AI company to debut on public markets. Initially, fears about the company’s debt load restrained its stock’s performance. But its shares took off in May after it reported its first quarterly earnings, including soaring revenue growth of 420% quarter-on-quarter. While the price has declined since a peak in June, CoreWeave shares closed on Friday at $104, up from a debut of $40.

But as Luria puts it, the bear case for CoreWeave is simple math. The company’s business model is to borrow capital and then use that capital to build data centers filled with GPUs and then sell time on those GPUs to AI companies. “The question is, are they getting a sufficient return … on their investment to justify the interest they’re paying on their debt?” Luria says.

Its most recent filings show how hefty that debt has become. The problem isn’t just the amount of CoreWeave’s debt. It’s the structure—most of it is more expensive than average for corporate debt, and much of it comes due in the next nine months. Of CoreWeave’s current liabilities, $3.6 billion is debt payable by June 30, 2026, just part of $11 billion in overall debt the company carries on its balance sheet. Much of that debt carries hefty interest rates of between 9% and 15%, according to the financial statements, with 11% being the weighted average rate overall. (This fall, rates on newly issued investment-grade corporate debt have hovered between 5.5% and 6%, according to Moody’s. The rate at which CoreWeave is able to borrow has come down over time, with most of its newer debt issued at closer to 9%. )

The majority of CoreWeave’s outstanding debt, its statements show, is in the form of two loans, called Delayed Draw Term Loan (DDTL)1.0 and DDTL 2.0. There is $1.8 billion outstanding on the DDTL 1.0, at a 15% interest rate, and $5 billion in DDTL 2.0, at an 11% interest rate. Both of those loans require quarterly principal payments starting in January 2026. (The interest rates on both these loans are floating.)

This is where the company’s RPOs come in. CoreWeave says that as of June 30 it had a little over $30 billion, the majority of which should turn into actual sales over the next four years. The company says 50% of that amount, or $15 billion, will be recognized in the next two years. Assuming half of that will in turn be recognized in the next year, that means the company should have $7.5 billion coming in. But if its operating margins remain at just 1.6%, the company will only generate $120 million in income from this $7.5 billion—not enough to cover its interest expenses or make the principal repayments on its debt.  That implies that CoreWeave’s returns remain far below the cost of its capital.

And there are far more capital expenses to come. CoreWeave continues to spend heavily to purchase Nvidia GPUs—which make up the great majority of its capital expenditures—and other equipment to outfit its data centers. In the first half of 2025, for example, it invested $4.7 billion in property, plant and equipment while bringing in only $2.2 billion in revenue. The report from Kerrisdale cites similar concerns as Luria in justifying its short position. CoreWeave is “a debt-fueled GPU rental business with no moat, dressed up as innovation,” the firm writes, arguing that the stock faces a 90% downside.

Can CoreWeave count on its customers?

Luria and Kerrisdale both cite CoreWeave’s highly concentrated customer base as another potential peril—a reality that CoreWeave itself acknowledged in its last earnings report. “A substantial portion of our revenue is driven by a limited number of our customers, and the loss of, or a significant reduction in, spend from one or a few of our top customers would adversely affect our business,” the company wrote. Most notably: In the second quarter of 2025, an eye-popping 71% of CoreWeave’s revenue came from Microsoft alone.

To be sure, Microsoft has a better credit rating than many countries, including the United States, and is unlikely to renege on its contract with CoreWeave. And CoreWeave’s contracts with its lessees generally require them to pay to the ends of their leases even if they don’t wind up utilizing them, except in case of “nonperformance.”

Much of CoreWeave’s utility has come from offering readily available compute as those companies raced to scale up their own operations. But with Microsoft set to spend tens of billions of dollars on developing its own data centers, it may not need CoreWeave’s services in the future. “They will pay their obligations, but the likelihood of them renewing at the end of the contract is much less guaranteed,” Luria says.

CoreWeave’s other major customer, OpenAI, is another matter. In March, CoreWeave entered an agreement with Sam Altman’s AI giant, with OpenAI committed to paying $11.9 billion through October 2030, with a $4 billion expansion announced in May. But OpenAI itself has made commitments far beyond its current cash flow—including commitments in the many hundreds of billions to Oracle, Nvidia and other data center providers.

If OpenAI runs into any financial troubles, Luria says, CoreWeave likely wouldn’t be first to receive payments, compared to much larger OpenAI partners like Microsoft, Amazon, or even Oracle. (That’s also the case for some of the other, smaller venture-backed and loss-making AI startups CoreWeave serves, such as poolside, Cohere, and Mistral.) To rely on CoreWeave’s ties to OpenAI, he says, “You have to believe that OpenAI is going to be unbelievably successful, so much so that it can pay everybody that’s ahead in line.”

All eyes on earnings

All of this means there’ll be plenty at stake when CoreWeave reports earnings on Monday. The company will also be digesting a recent setback: In late October, the shareholders of Core Scientific, a crypto miner with a hoard of computing resources that CoreWeave coveted, rejected CoreWeave’s $9 billion all-stock acquisition offer. (CoreWeave will console itself with a $270 million breakup fee—an amount that will help it cover the $360 million it is expected to pay Core Scientific to lease data center capacity from it next year. Overall, CoreWeave is on the hook to pay Core Scientific some $10 billion in future lease payments over the next 12 years, which may have been one reason CoreWeave was eager to try to use its highly-valued shares to purchase Core Scientific in an all-equity deal valued at $9 billion when the offer was initially made.)

The argument for CoreWeave’s success is just as simple as the math against it: AI is going to represent a transformational shift in the global economy, and CoreWeave is powering the technology’s growth. As the development of AI accelerates, so will the demand for computing power from companies aside from the hyperscalers, all hungry for the services of providers like CoreWeave. What’s more, today’s creditors may be willing to wait a little longer for that demand to materialize: The lenders behind DDTL 1.0 and 2.0 recently renegotiated the terms of the loan to delay the start of principal payments.

Blakely, at Cantor Fitzgerald, says that CoreWeave is already diversifying its customer base, noting the recent deal with Meta. “It’s a growth business,” he says, in reference to CoreWeave’s capital expenditures. “If business is growing, you have to invest against it.” He and other bulls see a future where the company is no longer laden with debt. Once CoreWeave sheds its liabilities over the next five years, Blakely says, a growing percent of its revenue can begin coming from infrastructure that’s already paid off, boosting CoreWeave’s margins. Moreover, Blakely says that new models of chips might run more efficiently or be able to fetch a higher premium from customers, allowing CoreWeave to hold more pricing power.

Blakely says that those $30.1 billion in RPOs—the revenue that has been booked but not yet delivered—are likely to increase meaningfully in CoreWeave’s next earnings, based on recently announced agreements. But the number to watch might be the additional obligations in borrowing that the company will take on to service them: If those obligations scale up alongside the booked revenue, bears say, CoreWeave still risks running out of cash unless it takes on yet more debt, raises more equity, or gets existing creditors to extend terms.

Blakely acknowledges that a sustainable path forward for CoreWeave is perilous. “A lot has to go right,” he says. Still, he compares the current moment in AI to the beginning of the smartphone era, where analysts doubted Apple’s claims that it would transform global communication. “CoreWeave is a leader there in terms of this market,” Blakely says. “If they can maintain that lead … they will be able to participate in the spoils.”

This story was originally featured on Fortune.com

View Comments