‘Little room for error’: what a debt-laden cloud computing company says about the AI boom

‘Poster child of AI bubble’ CoreWeave symbolises increasingly risky business model of datacentre buildout The Egyptians had pyramids. The early 21st-century tech bros have massive buildings full of expensive glass and metal blocks providing the substratum for a powerful inhuman intelligence.
‘Poster child of AI bubble’ CoreWeave symbolises increasingly risky business model of datacentre buildout
The Egyptians had pyramids. The Roman empire had roads. The early 21st-century tech bros have massive buildings full of expensive glass and metal blocks providing the substratum for a powerful inhuman intelligence.
Building all those datacentres is taking a while, locals everywhere are angry about them and it is unclear whether the inhuman intelligence inside them is actually going to make their backers rich.
This could be a problem for the people building those buildings – and therefore everyone else.
Almost no company is more exposed to the downsides of the AI boom than CoreWeave, a band of former crypto miners turned datacentre merchants, which has been called the poster child of the AI bubble.
The company sells the raw power for AI, known as compute, out of what it says is a network of 52 datacentres around the world, from Texas to Scotland.
Many of these are under construction; some are little more than announcements. CoreWeave is building them by taking on staggering quantities of debt. The company now owes $47bn (£36bn) and is borrowing more, with $10bn of this figure due next year. It bled $709m last quarter, on revenues of $3.6bn.
CoreWeave already looked dangerously exposed this autumn, when creditors put its risk of default – being unable to pay its debts – at over 60% at some point over the next five years.
Its earnings, released earlier in November, provided a brief respite from some of their fears – its margins and revenues “knocked it out of the park”, said Jordan Chaflin, head of technology at CreditSights, a financial research firm.
That was before scandal hit Nvidia, the most valuable business in AI and the company that makes CoreWeave’s entire business model possible. Nvidia’s chief executive, Jensen Huang, admitted to having an affair with an alleged Chinese spy. He’s now out. This will have consequences, and no one knows what they are yet.
Huang ran his business on close relationships and, as Nvidia grew, these relationships became the scaffolding of the AI economy. Few companies will be more affected by his removal than CoreWeave.
Its vulnerabilities – and they are many – are a blueprint for the wobbly structure of the AI bubble itself, on which the fortunes of the world’s stock market and the global economy now depend.
CoreWeave is what is called a “neocloud”. Its basic business model is this: the world’s biggest AI companies, like OpenAI, want billions of people to start using their products but don’t want to sink a fortune into building the large datacentres it takes to run them. Nvidia, its primary backer, wants to sell lots of computer chips to the world’s biggest AI companies.
So Nvidia provided the finance for CoreWeave and its peers to buy its chips, stuff them into datacentres, and rent that computational power to the world’s biggest AI companies.
Nvidia owns approximately 11% of CoreWeave. There are other arrangements, too, including a $6.3bn deal in which Nvidia agreed to buy “compute” that CoreWeave cannot sell – a sort of guarantee.
That’s not the only side of the equation Nvidia is propping up. It also has invested in CoreWeave’s clients – companies like OpenAI, Anthropic and nearly a dozen others that rent these chips. A dollar that Nvidia invests in one of these can end up as income on CoreWeave’s balance sheet, and then revenue for Nvidia as CoreWeave purchases chips.
Circular arrangements like this are the scaffolding of the AI economy. CoreWeave – and companies like it – are the vehicle through which Nvidia can make its balance sheet look spectacular, and OpenAI and Anthropic can provide ever-more-powerful models to their customers without paying workmen to break earth in Texas.
Think of it as a sandwich where neither slice of bread – Nvidia or a frontier AI company using its chips – wants to finance or build, say, $1tn worth of sheds stuffed with semiconductors, so the peanut butter in the middle takes on the risk.
Analysts expect CoreWeave to be a capital-intensive – meaning big-spending – business. Someone has to be. Most appear to have accepted its considerable risk.
“Their business model is essentially … you’re doing this massive buildout of infrastructure,” said Chaflin. “It is really quite painful from a free cashflow perspective, during the buildout.”
That infrastructure “needs to get funded with debt” Chaflin said. He emphasises that CoreWeave is not the only company doing this: it is the business model of “hyperscalers” and “neoclouds” – the companies building AI datacentres to support a future, still-theoretical AI economy.
In a US Securities and Exchange Commission (SEC) filing this summer, CoreWeave listed 18 different varieties of debt it is now paying back, amounting to roughly $35bn. In its most recent earnings, which do not break down this debt, that number is up by nearly $12bn, to $47bn.
It expects its interest payments to crest over $1bn next quarter. More than $10bn of its debt comes due next year.
What is this debt? Is it sustainable? And what happens if CoreWeave can’t pay it?
CoreWeave has been a test case in creating a new kind of debt, which has allowed it to finance its operations. This is GPU-backed debt – where a GPU is a specialised computer chip.
CoreWeave borrows money against its main asset, which are those computer chips made by Nvidia.
Since this has looked like a shaky proposition to most lenders – AI is a new, frothy technology – it couples those computer chips with customer contracts. In other words, CoreWeave has borrowed money by showing banks and private credit agencies that it has the goods, and it has people paying money to use them.
This has allowed it to borrow nearly $25bn.
When CoreWeave first started doing this, many people thought GPU-backed debt was a bad idea – not least because GPUs, unlike real estate, say, lose their value over time. In October 2025, the news website the Verge called the company “an Eldritch horror” with “no obvious path towards profitability”.
But then other companies – Crusoe, Nscale, Lambda – also started to take out loans backed by GPUs. Slowly, markets started to buy into the idea. In 2024, when CoreWeave took out its first GPU-backed loan, it had to pay a high interest rate of 15%. This has decreased to 9% on some subsequent loans.
This March, CoreWeave even managed to class $8.5bn of this debt as “investment grade” – suggesting it is trustworthy – and sell it to institutional investors, including insurance companies and pension funds.
Short-sellers, who bet on a fall in a company’s share price, are nervous about this idea. A major point of unease is the timeline for how fast the chips will lose their value, or depreciate.
CoreWeave has suggested this could happen over six years. There are market watchers who believe the real timeline could be far shorter. This could mean that at some point its loans might be backed by collateral that has plummeted in value.
No one knows if this will actually be a problem. Chaflin is optimistic. “We’re in an extremely tight environment for AI infrastructure,” he said, and there is demand for secondhand chips.
CoreWeave also has long-term contracts, he said, meaning that its customers are committed to pay for most of the lifespan of the chips they are renting.
This will work as long as those customers – such as OpenAI – stay solvent. For some of them, that might come down to: will Nvidia still finance them?
The GPU-backed debt is not CoreWeave’s only debt. Beginning last summer, the company also started issuing a different kind of debt called senior notes.
While the GPU-backed debt is backed by CoreWeave’s expensive computer chips, these senior notes are unsecured. This means there is no specific asset tied to them that a creditor could repossess if the loan becomes unpayable.
Should CoreWeave go belly-up, the owners of the senior notes are in theory entitled to be repaid first. How that repayment will be made is unclear. The company has few assets outside GPUs. It does not even own many of its datacentres outright. Instead it leases the buildings from other companies.
A spokesperson for CoreWeave said: “Customers are signing multiyear commitments for earlier-generation GPUs extending beyond their depreciable lives, with pricing above their original contracts. And we are financing our growth with an increasingly diverse and flexible funding strategy, lowering our weighted average cost of debt, and in most cases eliminating any reliance on renewals to repay that debt.”
CoreWeave had little choice but to take out these loans, and to issue this debt. It needs them to finance its expansion. The capital intensity of this – the sheer, huge sums it needs to keep going – are all part of its business model.
No one knows if CoreWeave has a long-term future. For analysts and short-sellers closely watching CoreWeave’s – and others’ – financials, the answer lies somewhere in the thousand moving pieces that compose its commitments and its dependencies.
Will it have the cash to repay its vast debts? If it doesn’t, who might step in to save it? If it defaults, who might it take down with it?
CoreWeave is bleeding hundreds of millions of dollars each quarter paying the interest on its debt. There is a real possibility that its revenues will not be as strong this year. CoreWeave makes money by renting chips out to paying customers. It says it has more than $100bn of contracts in the pipeline, a reassuring number.
But, unlike most of its competitors, CoreWeave accepts large sums upfront for its multiyear contracts. In other words, its annual revenues largely reflect years of future income for services that it has yet to carry out.
“Roughly two-thirds of the cash its business ‘generated’ was customers paying in advance, not profit from service already delivered,” said Brian Hopkins, at the research firm Forrester.
This suggests CoreWeave’s existing customers are unlikely to be enough to keep its income – $3.4bn last quarter – steady. It will have to keep finding new customers because it has already booked much of the value of its current contracts.
But there are also question marks over the solvency of its customers. Leading AI companies, including OpenAI and Anthropic, are heavily loss-making. They have so far managed to keep paying for “compute” – chips in CoreWeave’s datacentres – by borrowing money, and from companies like Nvidia.
Will they keep raising this money? Possibly not. Recent reporting has indicated that Nvidia, in a post-Jensen Huang world, might pull back from its role as the “central bank of AI”.
A spokesperson for CoreWeave said: “Demand for CoreWeave’s AI cloud platform is not an assumption, it’s backed by $132.5bn in contractual commitments from the world’s leading AI labs and enterprises. We grew to 1.7GW of active capacity in the last quarter and are benefiting from a very strong pricing environment.”
Finally there is the question of if political headwinds could delay or freeze datacentre projects. This is an unknown – and its interest payments rely on the answer.
“CoreWeave pays interest whether or not a datacentre comes online on time but it only earns revenue once the capacity is delivered. Local restrictions or permitting fights stretch that gap, and the company is already guiding to more than $1bn of interest expense a quarter,” said Hopkins.
What happens if it misses a payment, or if it defaults? For the pension and insurance firms that hold its debt, that is a critical question.
“The only asset on the balance sheet that isn’t pledged to someone is Michael Intrator’s kidneys,” said Amir Zabeti, the managing director of Mancala Capital, which is currently shorting CoreWeave. (Michael Intrator is the chief executive of CoreWeave.)
CoreWeave’s debt payments over the next three years are already equivalent to two years of income, said Hopkins. And the company still needs more money to build datacentres to service its customers. It is about to another round of debt.
“That leaves very little room if anything slips,” said Hopkins.
- room
- error
- cloud
- company
- boom
- datacentre
- intelligence
- debt




