Imagine a caravan maker. It sells caravans to a caravan park that only buys one type of caravan. The caravan park leases much of its land from another caravan park. The first caravan park has two big customers. One of the big customers is the caravan maker. The other big customer is the caravan maker’s biggest customer. The biggest customer of the second caravan park is the first caravan park.
Sorry, not caravans. GPUs.
As Tabby Kinder and Rob Smith wrote last week for MainFT:
CoreWeave . . . which leases computing capacity to tech groups building artificial intelligence models, is gearing up for the largest stock market debut of the year.
This week it revealed it was seeking to raise as much as $2.7bn in the share sale, valuing the business at $32bn. As the New Jersey-based group prepares to start an investor roadshow, it is attracting scrutiny for its huge debt burden, borrowing at high interest rates, and forthcoming maturities on billions of dollars of loans.
CoreWeave started out in 2017 as the side hustle of some traders at Hudson Ridge Asset Management, a defunct gas futures hedge fund. First it was an Ethereum miner that pivoted during the 2019 crypto crash to pay-per-hour 3D video rendering. The phrase “machine learning and AI” was added to CoreWeave’s blurb in November 2022, the same month OpenAI launched, and soon grew to consume the whole. Shortly after CoreWeave’s Series C funding round in May 2024, its website title changed from “The GPU Cloud” to “The AI Hyperscaler”.
MainFT’s coverage focuses on the Blackstone and Magnetar Capital-backed company’s $8bn of debt. The crux of the story is the WeWork-style mismatch between its assets and liabilities, along with some apparent carelessness around debt covenants:
CoreWeave . . . violated several key terms of a $7.6bn loan last year, triggering a series of so-called technical defaults.
[The company] disclosed in the exhibits to its IPO document that it had to ask its biggest lender Blackstone to amend the terms of the loan and “waive” these defaults in December.
While CoreWeave did not miss any payments under the loan facility, it made a slew of serious administrative errors, which stemmed from beginning to use the financing to expand into western Europe. This clashed with key terms that in effect restricted the debt’s collateral to the US.
But with CoreWeave due to price its IPO later today, there’s plenty more in the S-1 filing that deserves attention. Here’s a quick tour of other notable items.
The Nvidia thing:
Tim Bradshaw last year asked CoreWeave CEO Michael Intrator about the company’s reliance on Nvidia, its 5.97 per cent shareholder, key supplier and key customer. Intrator . . .
. . . batted off questions about whether prospective investors were concerned about backing a business that had raised capital from Nvidia, only to spend a significant portion of those funds on that company’s products.
“It’s such a crap narrative,” he said. “Nvidia invested $100mn. We’ve [raised] $12bn in debt and equity. It’s an inconsequential amount of money in the relative scale of the amount of infrastructure we’re buying.”
Crap narrative it may be, but let’s take a look at what the S-1 says about customer concentration:
We recognized an aggregate of approximately 77 per cent of our revenue from our top two customers for the year ended December 31, 2024.
And . . .
Our largest customer accounted for 16%, 35%, and 62% of our revenue for the years ended December 31, 2022, 2023, and 2024, respectively.
CoreWeave’s revenue was $1.9bn in 2024. Sixty-two per cent of $1.9bn is $1.18bn. That squares with its Microsoft Master Services Agreement (our bold):
In February 2023, we entered into a Master Services Agreement (the “Microsoft Master Services Agreement”) with Microsoft, pursuant to which we provide Microsoft with access to our infrastructure and platform services through fulfillment of reserved capacity orders submitted to us by Microsoft and as may be amended upon our and Microsoft’s mutual agreement. We have recognized revenue of $81 million and $1.2 billion for the years ended December 31, 2023 and 2024, respectively, pursuant to the Microsoft Master Services Agreement.
That leaves 15 per cent of $1.9bn, or $285mn. That’s not far off the 20-month number CoreWeave gives for its Nvidia contract:
In April 2023, we entered into a Master Services Agreement (the “Master Services Agreement”) with NVIDIA, a beneficial owner of more than 5% of our outstanding capital stock, pursuant to which we provide NVIDIA with our infrastructure and platform services through fulfillment of order forms submitted to us by NVIDIA. As of December 31, 2024, NVIDIA has paid us an aggregate of approximately $320 million pursuant to the Master Services Agreement and related order forms.
And since . . .
None of our other customers represented 10% or more of our revenue for the year ended December 31, 2024.
. . . it seems fair to conclude that CoreWeave’s second-biggest customer in 2024 was Nvidia — which makes the following line feel a bit incestuous:
[O]ur current customers have contractually specified our use of NVIDIA GPUs.
Much more on Nvidia later, but first . . .
The Core Scientific thing:
“We relentlessly and creatively explore additional opportunities to add power capacity”, says CoreWeave’s S-1. Nowhere is that better demonstrated than with Core Scientific.
Core Scientific is a publicly traded crypto miner that collapsed into bankruptcy protection in 2022 alongside its main customer, Celsius Network, whose founder/CEO Alexander Mashinsky last year pleaded guilty to fraud and market manipulation. (Core Scientific’s co-founder, the Viper Room nightclub co-owner and Fatburger promoter Darin Feinstein, stepped down as group co-chair in 2023.)
CoreWeave last year tried to buy Core Scientific. After its takeover proposal was rejected, CoreWeave announced several contracts to rent and modify Core Scientific’s rack space. At the 2024 year-end, Core Scientific’s data centres accounted for “more than 500MW” of CoreWeave’s approximately 1300MW of total capacity (72 per cent of which was not yet switched on).
Core Scientific, in a 2024 results presentation, says its contracts with CoreWeave last 12 years.
Core Scientific’s disclosures also reveal who funds the conversion. A footnote to the above graphic says CoreWeave is paying Core Scientific “up to $1.5mn per HPC [high-performance computing] MW of data centre build-out costs” to a value of about $750mn. In exchange, CoreWeave gets an up-to-50 per cent rebate on its hosting costs. A follow-on deal involves CoreWeave funding $104mn of capex for no hosting rebate; more on that below.
It’s hard to shake the impression that CoreWeave is sinking a lot of capital and wearing most of the risk. Core Scientific’s data centres will still be there long after CoreWeave’s chips are fried.
Yet the market gives Core Scientific an enterprise value of approximately 10 times EBIT — a deep discount to conventional real estate investment trusts, which probably reflects some uncertainty about its anchor tenant.
Meanwhile, CoreWeave’s syndicate of 14 IPO advisers had reportedly been aiming for approximately 15 times forward EBIT. So even if recent speculation proves accurate that the price range has moved down by around 20 per cent, it still looks pretty punchy.
The depreciation thing:
GPUs are quickly depreciating assets. Not only do they burn out, they’re constantly being superseded by new models. Massed Compute estimates value loss of 20 to 30 per cent a year. The investment case for an AI data centre hinges on the rental market growing fast enough to cover sunk costs before their hardware is obsolete.
Here’s how CoreWeave’s S-1 estimates the useful life of its property and equipment:
Technology equipment: 6 years
Software: 3-6 years
Data center equipment: 8-12 years
Furniture, fixtures, and other assets: 3-5 years
Six years in AI is an eternity. Nvidia’s server-grade V100 GPU cost around $10,000 in 2019 and can now be picked up for a few hundred dollars. It’s already four generations behind the times, with another generation due to arrive next year.
CoreWeave’s rapid expansion last year makes it a big bet on Hopper, Nvidia’s last-but-one architecture, which debuted in 2022. The S-1 doesn’t give a detailed breakdown of assets but says a majority of its GPUs use Hopper.
It’s expansion has been mirrored across the data centre industry, which has moved inside a year from a critical shortage of AI compute to a glut.
“I start to see the beginning of some kind of bubble,” Alibaba chair Joe Tsai told a conference this week. “I start to get worried when people are building data centres on spec. There are a number of people coming up, funds coming out, to raise billions or millions of capital.”
What’s probably happening is that hyperscalers are no longer compelled to sprint and establish competitive moats by training the biggest models, so don’t need to rent as much emergency capacity, while smaller operators are waiting to see where things land before committing funds. Here’s what Goldman Sachs (a CoreWeave IPO lead underwriter) says in a note dated March 24 that downgraded ratings and forecasts for several Taiwanese AI server makers:
‘Training’ server will remain the growth driver given the increasing need for computing power to upgrade advanced AI models, but the volume ramp up is slower than we previously expected due to the combined reasons of product transitioning and uncertainties of demand and supply. As the GPU platform is transiting to next generation in 2H25, shipment can potentially slow during the transition period. Uncertainties remain in production ramp up, given the complexity of full rack systems and there remains debates on the demand for intense computing power after the release of more efficient AI models like DeepSeek.
The near-term result is a rental price war.
Nvidia doesn’t give list price for high-end data centre hardware like the H100, its 2022 flagship, but resellers tend to quote a price of around $30k. Renting one for an hour used to cost between $4 and $8. Now it costs as little as $1.
It’s difficult to know how aggressively CoreWeave is competing with rivals on price because of . . .
The contract length thing:
Over the past couple of years, CoreWeave has all but abandoned pay-as-you-go and moved its customer base on to take-or-pay contracts, billed monthly. The per-hour prices agreed are unlikely to match those quoted on its website, which haven’t come down much since the boom times.
The company’s S-1 gives a “weighted average” customer contract length of around four years.
Several of the S-1’s risk factors are about how contract pricing is unproven. It’s also, from the perspective of logic, all a bit challenging. Why would a company commit to a rental that’s not much shorter than the predicted useful life of the asset being rented? What do they get out of hiring a rapidly-depreciating GPU other than the ability to renegotiate mid-contract or walk away? Isn’t flexibility in the face of uncertainty the whole point of being asset-light?
And in the context of CoreWeave’s 12-year site leases and the recent switch of pricing model, what does a four-year “weighted average contract duration” mean in practice? Most of last year’s revenue came from Microsoft, which has commitments to 2030, and last year’s only other customer of note was Nvidia, whose biggest customer last year was Microsoft. As the FT reported last month, Microsoft has already pulled some business from CoreWeave “over delivery issues and missed deadlines”. (CoreWeave denied that contracts were cancelled.)
Assumptions of weak pricing power and high customer churn are premised on data centre compute being commoditised. Going by a recent RBC Capital Markets client survey, that seems to be the consensus view on the buy-side:
But is that fair?
In terms of financials, CoreWeave’s operating expenses last year were 50 per cent tech/infrastructure (ie, buying stuff) plus 26 per cent for power, etc (ditto). Interest costs and writedowns were what turned its $324mn of operating income into a $863mn net loss. From those figures, it might be argued that customers have been making more use of its balance sheet than its cloud computing expertise.
The SPE thing:
You say “special-purpose entity” and people will automatically think Enron. They have no reason to here. CoreWeave’s S-1 makes clear it doesn’t use off-balance-sheet vehicles, as you’d expect. The company’s talk of monetising AI compute has only a superficial similarity to Enron’s pitch to make broadband a new asset class.
True, CoreWeave has raised most of its debt through a wholly owned special purpose vehicle, CoreWeave Compute Acquisition Co. IV LLC, which uses an undisclosed number of its parent company’s GPUs and services contracts as collateral. But it’s all relatively transparent. Even the technical defaults are disclosed, albeit it takes a dig through the ancillary docs and a trained eye to spot them:
We note the Enron echo only because analysts at DA Davidson have heard it too. Here’s an extract from their recent note:
The key for CoreWeave was the ability to secure $12B worth of loans in order to purchase $12B worth of data center capacity. CoreWeave took a $100M investment from NVIDIA, a $320M contract from NVIDIA to buy its capacity, and a multi-year deal with Microsoft to raise $1.6B of equity and $12.9B of debt commitments, mostly at 10-14% interest but up to 17%. This allowed CoreWeave to purchase 250,000 GPUs from NVIDIA (about $10B worth). We believe the ~$8B it spent on GPUs made it a 6-7% customer for NVIDIA.
How is this different from Enron’s Special Purpose Entities?
The previous description may have sounded familiar for investors in the early 2000s. Enron used Special Purpose Entities it created in order to offload assets and liabilities off its balance sheet and inflate its profits by generating revenue from these entities. In Enron’s case these SPEs were controlled by executives and were hidden from the public, where in CoreWeave’s case there are 3rd party investors and more transparency, though the impact to the balance sheet and profitability are reminiscent.
We believe this structure may continue to work as long as demand for AI continues to grow exponentially. As long as demand for AI grows faster than hyperscalers are able to build data centers, CoreWeave may be able to use the proceeds of the IPO, borrow more debt and continue the cycle. However, if Microsoft ceases to need overflow capacity and/or OpenAI is not able to raise the $11.9B it is committed to, CoreWeave’s growth path may not be sustainable.
Hmmmmm.
The Magnetar thing:
CoreWeave’s Nvidia relationship isn’t the only one with Freudian overtones. Here’s what the S-1 reveals about Magnetar, another co-owner and customer:
In August 2024, we entered into an agreement (as amended, the “MagAI Capacity Agreement”) with a fund managed by Magnetar (“MagAI Ventures”). Under the MagAI Capacity Agreement, we will provide certain portfolio companies of MagAI Ventures with a predetermined amount of cloud computing services at a pre-negotiated hourly rate. The specific amount of cloud computing services to be used by each portfolio company, if any, will be negotiated individually with each portfolio company, and will be subject to final approval by MagAI Ventures.
We received a refundable deposit of approximately $230 million in connection with the MagAI Capacity Agreement. Any consumption of cloud services by MagAI Ventures, including by their portfolio companies, under this arrangement is deducted from this deposit amount, with the unused portion refunded back to MagAI Ventures at the end of the term of the arrangement.
A fund operated by CoreWeave’s co-owner paying a $230mn deposit to CoreWeave might look a bit conflicted, but it’s not like CoreWeave’s an investor in Magnetar funds!
Wait, sorry, yes it is:
On June 14, 2024, we [CoreWeave] contributed an aggregate amount of $50 million to a fund managed by Magnetar (“MAIV”) in connection with MAIV’s purchase of shares of preferred stock in a private company.
The escrow thing:
Buried in CoreWeave’s “subsequent events* addendum is this paragraph.
In February 2025, the Company modified multiple lease agreements with a single landlord. The modifications changed the contracted power capacity, term, and contractual payments, and terminated the related escrow agreements. As a result of the modification, the Company will receive an additional 70 MW of contracted power capacity. The Company received a refund of $304 million of unused escrow funds previously included within other non-current assets, and expects to make approximately $1.7 billion of additional rent payments over the 13 year term of these leases.
The *single landlord” is Core Scientific, which refers to the follow-on deal in its results presentation. What’s odd here is the $304mn refund. For any company swimming in liquidity, it seems small beer.
The OEM loan thing:
Page 84 of the S-1 has the following breakdown of debt:
Term loan facility (4) is an interesting one. It’s a $1bn credit line from JPMorgan, mostly unsecured, that CoreWeave agreed in December.
Meanwhile:
The Company entered into various agreements with an OEM between February and December 2024 whereby the Company obtained financing for certain equipment with an aggregate notional balance of $1.3 billion as of December 31, 2024. Related to the financing agreements, the Company granted a security interest for the financed equipment. The agreements are accounted for as financing arrangements, with terms between two to three years. The financing arrangements have a stated repayment schedule over the term with effective interest rates between 9% to 11%. The Company did not incur any debt issuance costs associated with the financing arrangements. Interest expense for the year ended December 31, 2024 was $60 million.
From the above paragraph, only an expert in supplier-finance disclosures will follow who’s paying whose bills. What we can say is that “certain equipment” purchased is highly likely to be Nvidia chips, and that CoreWeave’s S-1 names Dell and Super Micro Computer as among its OEM partners. The term loan’s size and proximity to the financing agreement are further complications. Whatever’s going on, it’s another aspect of the business that might look uncomfortably circular.
Stuff like this doesn’t tend to get picked up because CoreWeave rents GPUs rather than, for example, caravans. The market for generative AI has been growing in a way that the market for towable holiday accommodation has not.
The internal economics of both industries are not dissimilar, however, particularly around mismatches between sunk capex, asset depreciation, contract lengths and uncertain returns. But maybe, if the caravan industry were as insular and interconnected as AI, it would have just as exciting a growth story to tell.
CoreWeave declined to comment.
Further reading:
— CoreWeave
