Soundtrack: EL-P (ft. Aesop Rock) - Run The Numbers
In my years writing this newsletter I have come across few companies as rotten as CoreWeave — an "AI cloud provider" that sells GPU compute to AI companies looking to run or train their models.
CoreWeave had intended to go public last week, with an initial valuation of $35bn. While it’s hardly a recognizable name — like, say, OpenAI, or Microsoft, or Nvidia — this company is worth observing, if not for the fact that it’s arguably the first major IPO that we’ve seen from the current generative AI hype bubble, and undoubtedly the biggest. Moreover, it’s a company that deals in the infrastructure aspect of AI, where one would naturally assume is where all the money really is — putting up the servers for hyperscalers to run their hallucination-prone, unprofitable models.
You’d assume that such a company would be a thriving, healthy business. And yet, a cursory glance at its financial disclosure documents reveals a business that’s precarious at best, and, in my most uncharitable opinion, utterly rancid. If this company was in any other industry, it would be seen as such. Except, it’s one of the standard bearers of the generative AI boom, and so, it exists within its own reality distortion field.
Regardless, CoreWeave’s IPO plans appear to have been delayed, and it’s unclear when it’ll eventually make its debut on the public markets. I assume the reasons for the delay are as follows.
First, (and we’ll talk about this later), on March 10, OpenAI announced the completion of a deal with Coreweave valued at $11.9bn that would see it procure AI compute from the company, while also taking a $350m stake in the business. This arrangement has, undoubtedly altered some of the calculus behind things like valuations, and so on.
Additionally, Coreweave has now released an amended version of its S-1 — the document that all companies must file before going public, and that acts as a prospectus for would-be investors, revealing the strengths and weaknesses of the business. The new partnership with OpenAI does complicate some things, including when it comes to risk (as we’ll discuss later), and so it naturally makes sense that CoreWeave would have to release an updated version of its prospectus.
I’ve spent far too long reading CoreWeave’s S-1. For the uninitiated, S-1 documents are, as a matter of rule, often brutal. The SEC — and federal law — demands total, frank honesty. It’s a kind of hazing for would-be public companies, where they reveal all their dirty secrets to anyone with a web browser, thereby ensuring those who invest in the company on the first day are able to make informed decisions.
The revelations contained in S-1 documents are, quite often, damning, as we saw in the case of WeWork. It laid bare the company’s mounting losses, its debt burden, and its insane cash burn, and raised questions about the sustainability of a company that had signed hundreds of expensive long-term leases in the pursuit of growth, despite having never made a profit. Within a matter of weeks, WeWork cancelled the IPO and its CEO and founder, Adam Neumann, had left the company.
Sidenote: WeWork, incidentally, would later go public by merging with a SPAC (special purpose acquisition company, which is essentially a shell company that’s already listed on the open markets). SPACs exist for one reason, and that’s to allow shitty companies to go public and raise money from investors without having to go through the scrutiny of a full IPO. At least, that was the case prior to 2024, when the SEC began demanding increased disclosures from companies that sought to merge with SPACs and enter the public markets via the back door.
Unsurprisingly, many of the companies that used SPACs (like failed EV makers Fisker and Lordstown Motors, and Virgin Orbit) ultimately ended up in liquidation, or winding up petitioning a court for Chapter 11 bankruptcy protection. WeWork, for what it’s worth, filed for Chapter 11 in 2023, exiting bankruptcy the following year, albeit as a much smaller company, and one that was no longer listed on the vaunted New York Stock Exchange.
CoreWeave’s S-1 tells the tale of a company that appears to be built for collapse, with over 60% of its revenue dependent on one customer, Microsoft. In early March, the Financial Times reported that Microsoft has dropped "some services" with CoreWeave, citing delivery issues and delays, although Coreweave would later deny this.
The timing, however, is suspicious. It came a mere week after TD Cowen's explosive report that claimed Microsoft had walked away from over a gigawatt of data center operations, and likely much, much more. For context, a gigawatt is about the same as the cumulative data center capacity in London or Tokyo — each city being the largest data center market in their respective regions.
CoreWeave is burdened by $8 billion of debt (with its most recent debt raise bringing in $7.6bn, although that line of credit has not been fully tapped) that it may not be able to service. This figure does not include other commitments which are listed on the balance sheet as liabilities, like its various lease agreements for hardware and data center facilities.
Worse, despite making $1.9 billion in revenue during the 2024 financial year, the company lost $863 million in 2024, with its entire future riding on "explosive growth" that may not materialize, and even if it does, would require CoreWeave to spend unfathomable amounts of money on the necessary capex investments.
CoreWeave is, on many levels, indicative of the larger success (or failure) of the so-called AI revolution. The company's core business involves selling the unavoidable fuel for generative AI — access to the latest (and most powerful) GPUs and the infrastructure to run them, a result of its cozy relationship with (and investment from) NVIDIA, which has given CoreWeave priority access to its chips. As CoreWeave’s own S-1 notes, it was “the first cloud provider to make NVIDIA GB200 NVL72-based instances generally available,” and “among the first cloud providers to deploy high-performance infrastructure with NVIDIA H100, H200, and GH200.”
CoreWeave owns over 250,000 NVIDIA GPUs across 32 data centers, supported by more than 260MW of active power, making it competitive with many of the familiar hyperscalers I’ve mentioned in previous newsletters, despite being a company few have ever heard of. By comparison, Microsoft bought 485,000 GPUs in 2024 and aimed to have as many as 1.8 million GPUs by the end of that year, though it's unclear how many it has. Meta likely has somewhere in the region of 600,000 GPUs, and according to The Information's AI Data Center Database, Amazon has hundreds of thousands of its own.
In short, CoreWeave's position is one that at the very least competes with the hyperscalers, and is both a fascinating and disturbing window into the actual money that these companies do (or don't) make, and the answer is "not very much at all."
Furthermore, CoreWeave's underlying financials are so dramatically unstable that it's unclear how this company will last the next six months. As I'll get into, CoreWeave's founders are finance guys that have already cashed out nearly $500 million before the IPO, but did so in a way that means that despite only retaining 30% of the company's ownership, they retain 82% of the voting power, allowing them to steer a leaky, poorly-built ship in whatever direction they see fit, even if doing so might send CoreWeave into the abyss.
If this sounds familiar, it’s pretty much the same arrangement that Mark Zuckerberg has with Facebook. Despite only holding a small percentage of the company’s equity (around 13%), he holds the majority of voting shares, as well as the role of chairman of Facebook’s board, ensuring his position as CEO can never be challenged, regardless of any pressure from shareholders.
CoreWeave is a company that is continually hungry for more capital, and its S-1 cites potential difficulties in obtaining new cash as a potential risk factor. It intends to raise around $4 billion at IPO, which presumably will go towards servicing its debt and fuelling future expansion, as well as funding the day-to-day operations of the business. However, as I'll walk through in this newsletter, that will not be enough for this company to survive.
Sidenote: Remember when I said that companies have to lay out all their dirty laundry in the S-1, including potential risk factors? One of the factors cited is the questionable future of AI, and the failure of its customers “to support Al use cases in their systems” when those AI use cases are deployed on CoreWeave’s iron.
Again, Microsoft is Coreweave’s biggest customer. Essentially, it’s saying that Microsoft might not actually do a good job of getting people to use Copilot, or the OpenAI models it licenses through its own ecosystem, and that would, in turn, hurt CoreWeave.
The same document also mentions the usual stuff: the reputational harm that generative AI poses to its creators and those linked to them, regulatory scrutiny, and the uncertain trajectory of AI and its commercialization.
And, while we’re on the subject of risk factors, a few other things caught my eye. CoreWeave cited “material weaknesses in [its] internal control over financial reporting” as a risk factor. As a public company, CoreWeave will be forced to prepare and publish regular (and accurate) financial reports. While building the S-1, CoreWeave said it “identified material weaknesses in our internal control over financial reporting” which means that “there is a reasonable possibility that a material misstatement of our annual or interim financial statements will not be prevented or detected on a timely basis”
The good news: It’ll be able to fix them. The bad news? Doing so likely won’t be completed into 2026, and it’ll be “time consuming and costly.”
CoreWeave says that “negative publicity” could harm the company’s prospects, “regardless of whether the negative publicity is true.” This is a fairly generic statement that could apply to any business, and you’ll see similar generic warnings in most S-1 prospectuses, as they’re supposed to be a comprehensive representation of the risks that business faces. One line, however, did stand out. “Harm to our reputation can also arise from many other sources, including employee misconduct, which we have experienced in the past.” Interesting!
Anyway, I have a great deal of problems with this company, but let’s start somewhere simple.
Paging Doctor Zitron…
Number One — CoreWeave Does Not Have A Stable Business, And Is A Bad Omen For Generative AI Writ Large
To properly understand CoreWeave, we have to look at its origin story. Founded in 2017, CoreWeave was previously known as Atlantic Crypto, a cryptocurrency mining operation started by three guys that worked at a natural gas fund. When the crypto markets crashed in 2019, they renamed the company and bought up tens of thousands of GPUs, which CoreWeave offered to the (at the time) much smaller group of companies that used them for things like 3D modelling and data analytics. This was a much smaller business, and far less capital-intensive, with CoreWeave making $12m in 2022 with losses of $31m.
When ChatGPT's launch in late 2022 activated the management consultant sleeper cells that decide what the tech industry's next hypergrowth fixation is going to be, Coreweave pivoted again, this time towards providing the computational muscle for generative AI. CoreWeave became what WIRED would call "the Multibillion-dollar Backbone of the AI boom," a comment that would suggest that CoreWeave was far more successful than it really is.
Nevertheless, CoreWeave has — through its relationship with NVIDIA, which holds a reported 5% stake in the company — an extremely large amount of GPUs, and it makes money by renting them out on a per-GPU-per-hour basis. Its competition includes companies like Lambda, as well as hyperscalers like Amazon, Google and — believe it or not — Microsoft, all of whom sell the same services.
What's important to recognize about CoreWeave's revenue is that, despite whatever WIRED might have said, the majority of its revenue does not come from "being the backbone of the AI boom," but as auxiliary cloud compute provider for hyperscalers. When a hyperscaler needs more capacity than it actually owns, it’ll turn to a company like CoreWeave to pick up the slack, because building a new datacenter is — as noted in the previous newsletter — something that can take between three and six years to complete.
CoreWeave's customers include AI startup Cohere, Meta, NVIDIA, IBM, and Microsoft, the latter of which is its largest customer, accounting for 62% of all revenue during the 2024 financial year. It’s worth noting the speed in which CoreWeave became highly reliant on a single customer to exist. By contrast, in 2022 its largest customer accounted for 16% of its revenue, suggesting a far more diversified — and healthy — revenue base.
Although CoreWeave says its reliance on Microsoft will decrease to 50% of revenue as (if?) OpenAI starts shifting workloads to its servers, the current reality remains unchanged. Broadly speaking, CoreWeave is dependent on a few big-spending “whales” to stay afloat.
Per the S-1, 77% of CoreWeave's revenue comes from two of its customers, the latter of which remains unnamed, and is only referred to as “Customer C” in the document. However, based on reporting from The Information, it’s reasonable to assume it’s NVIDIA, which agreed in 2023 to spend $1.3 billion over four years “to rent its own chips from CoreWeave.”
Once you remove these two big contracts, CoreWeave only made $440 million in 2024.
These numbers deeply concern me, and I'll explain why.
- CoreWeave is, other than the hyperscalers, one of if not the largest holder of GPUs in the cloud space.
- CoreWeave sells, other than the GPUs themselves, the most "valuable" service in generative AI — compute.
- As a result, CoreWeave — as an independent company with the scale of a hyperscaler — is indicative of demand for generative AI services.
- CoreWeave's revenue, outside of its large contracts, amounts to about $440m, or $1.9 billion with these contracts.
- As a reminder, CoreWeave lost $863 million servicing these contracts.
- These numbers suggest one (or all) of the following:
- CoreWeave's business model does not make enough revenue to cover its costs.
- Outside of auxiliary capacity for hyperscalers, CoreWeave does not have a fundamentally sound or scalable business model.
- There is a fundamental lack of demand for compute for generative AI.
In short, CoreWeave is existentially tied to the idea that generative AI will become both a massive, revenue-generating industry and one that's incredibly compute-intensive. CoreWeave's future is one that requires an industry that has yet to show any meaningful product-market fit to grow so significantly that compute companies turn into oil companies at a time when Microsoft — the largest provider of GPU compute and the hyperscaler with the highest amount of proposed capex spending — has pulled back from both over a gigawatt of compute capacity and (reportedly) some of its contracts with CoreWeave.
CoreWeave’s three largest customers have, according to its S-1, increased their committed spend by around $7.8 billion during the 2024 financial year, representing a fourfold increase in the initial contract value. For the sake of clarity, this reflects future spending commitments — not actualized revenue from providing services to these companies.
While this might seem like good news, that's also nearly four times its current revenue from three customers, and as Microsoft has reportedly proven with its other compute contracts, big customers can simply cancel contracts on a whim.
Put simply, this dependence on a handful of hyperscalers represents a fundamental — and potentially fatal — vulnerability.
Sidenote: On the subject of vulnerabilities, the updated S-1 prospectus talks about a theoretical “counterparty credit risk.” What does that mean? Essentially, it’s when one party defaults on paying for services that the other party has paid for. If you don’t pay your mortgage, that’s counterparty credit risk.
CoreWeave is saying that, should a customer fail to pay its bills for infrastructure built on their behalf, or for services rendered, it could have a material risk to the business. The S-1 gives the example of its arrangement with OpenAI, where CoreWeave has agreed to provide certain services (and build certain infrastructure) in exchange for $11.9bn over the course of the next five years.
Although CoreWeave talks generally about the risk of counterparty credit risk, and only cites OpenAI in hypothetical terms, it’s also the only company named in this section. Which makes sense, because the chances of Microsoft or Meta becoming insolvent in the immediate future are slim, whereas OpenAI’s entire existence depends on its ability to raise more money than any startup in history, indefinitely, while also never making a profit.
And, as readers of this newsletter will know, I don’t rate its chances.
One last note on risks: Perhaps the biggest, in my view, is the fact that there’s really nothing inherently special about CoreWeave besides its existing infrastructure. Cloud GPUs are incredibly commoditized, and the core factors of differentiation between the various players are price, availability, and the exact hardware available. In fairness to CoreWeave, it has some strength in the latter point, with a close relationship with Nvidia that’s afforded it access to the latest and greatest hardware as it becomes available.
The problem is that, for the most part, with enough money you could make a company as equally capable as CoreWeave. And, indeed, CoreWeave does effectively the same thing as other hyperscalers like Google Cloud and Azure and AWS, as well as upstarts like Lambda Labs.
So, tell me, why is this business worth $35bn?
CoreWeave simply doesn't have meaningful demand or revenue resulting from its services. $440 million — with some of that revenue likely coming from other hyperscalers, albeit those who haven’t spent as much as Microsoft — is a pathetic sum that suggests either not enough people want to use CoreWeave’s services, or the services themselves are not actually that lucrative to provide, likely due to the ruinously-expensive costs of running hundreds of thousands of GPUs at full tilt.
Regardless of the reason, the company selling the literal fuel of the supposed AI revolution is losing hundreds of millions of dollars doing so.
Worse still, CoreWeave is entirely dependent on its largest customers, to the point that their entire business would collapse without them...and frankly, given the precarious nature of its financials, might even collapse with them.
That's because servicing this revenue is also incredibly costly. According to The Information, CoreWeave spent over $8.5 billion in capital expenditures in 2024, and funding said expenditures required CoreWeave to take on onerous amounts of debt, to the point that it's unclear how this business survives.
Number Two — CoreWeave Has Taken On A Fatal Amount of Debt
Forgive me, as the following is going to be a little dense.
In simple terms, CoreWeave's operations requires it to be in a near-constant state of capital expenditure, both to build the data centers it needs to serve customers from, to purchasing massive amounts of power, to acquiring the specialized GPUs necessary to run AI workloads.
CoreWeave has raised a total of $14.5 billion in equity funding (selling stock in the company) and debt financing (loans). Many of these loans are collateralized not by money or real estate, with “the Company’s property, equipment and other assets,” which includes the value of the GPUs used to power its operations. This is a new kind of asset-backed lending model created specifically to fund compute-sellers like CoreWeave, Crusoe, and Lambda, similar to how a mortgage is backed by the value of the property.
The problem is that, yes, GPUs are depreciating assets, a fact that will eventually become problematic for these companies. They eventually slide into obsolescence, as new chips and new manufacturing processes come out. With constant, intensive use, they wear down and fail, and thus require replacing. As noted later in this piece, as these assets lose value, CoreWeave is forced to increase its monthly payments as the collateral is (presumably) no longer sufficient to satisfy the outstanding debt.
In CoreWeave's case, the majority of its raised capital was raised in debt, with the majority of that coming in two Delayed Draw Term Loan facilities (DDTL). This usually means that while you have access to a certain amount of money, said money is only disbursed in predefined installments. These installments may come after a certain period of time, or when the company reaches a certain milestone. DDLTs work unlike personal loans, where you typically get the cash upfront, and start repayments immediately. The contracts are custom-written for each loan, and reflect the needs (and risks) of the business.
These loans can have wildly different terms based on their collateral and the incoming revenue of the company in question. All numbers below are an estimate based on the terms of the loans in question, and I do not attempt to factor in the actual cost of interest. For the most part, I've worked out the terms of the loans and the repayment schedule, and given you what I believe will be the lowest amount CoreWeave will owe.
For the sake of both your and my sanity, I'm going to focus on just these loans, as I believe they are, on their own, enough to potentially destroy CoreWeave.
Problem Loan Number 1: DDTL 1.0
- Size: $2.3 billion (fully drawn as of S-1 filing)
- 14.11% Annual interest in 2024
- At least $892 million a year in interest payments
CoreWeave's first Delayed Draw Term Loan (DDTL 1.0, as it calls it), came from titanic alternative asset management firm Blackstone (not to be confused with Blackrock), and Magnetar Capital, an Illinois-based hedge fund most famous for its involvement in the creation of Collateralized Debt Obligations, the security product that created the Global Financial Crisis of the late 2000s, raising $2.3 billion. This loan has now been fully drawn.
The effective annual interest rate on this loan is a little bit more than 14%, averaging at 14.11% in 2024 and 14.12% in 2023. The reason for the variance is because of how the interest rate is actually calculated. It combines the Term SOFR of the period (which is an average of the interest rate on treasury bond activity outside normal trading hours), plus 9.62%, or an unspecified “alternative base rate” plus 8.62%. At the time of writing, the 180-day average of SOFR is around 4.6%, although this can fluctuate depending on market conditions.
The terms of the loan require quarterly payments based on the company's cash flow and, starting in January 2025, the depreciated value of the GPUs that are used as collateral to provide the loan, and CoreWeave has until March 2028 to fully pay it off. Interest accrues monthly, and there is a final (though unspecified) balloon payment. Per the amended S-1 document, CoreWeave has paid $288 million in principal and $255 million in interest since the inception of the loan.
It’s tricky to actually calculate the monthly payments on this. The previous balance repayments aren’t a useful guide, as the loan wasn’t fully utilized in 2023, with CoreWeave carrying a balance of $1.3bn. There are 13 quarters between December 31, 2024 and March 31, 2028. If we divide the outstanding debt of $2bn by thirteen, we have around $150m. That only covers the principal, and not the interest — which, I remind you, stands at an arse-clenchingly steep 14.11%.
Nor, for that matter, do the previous repayments include the increase in principal payments starting from January 2025 to reflect the depreciating value of the collateral.
As Reuters reported in 2003, CoreWeave used its H100s as collateral for this loan. Those chips are, at this point, nearly two years old. While it’s unclear how the resale value of these chips has changed over time, it’s worth noting that the cost to rent a H100 in the cloud has dropped from between $4.70 to $8 an hour in late 2023, to just $1.90 at the time of writing. That will, undoubtedly, affect the value of CoreWeave’s collateral.
As the S-1 notes, this loan has a final balloon payment. It’s unclear how big this payment will be, as the filing doesn’t provide any detail. Still, regardless of the balloon payment size, it’s not unreasonable to expect that, from this year, CoreWeave will be spending around $250m each quarter to service this loan, or $1bn annually.
DDTL 1.0 also imposes liquidity requirements, although CoreWeave is only required to have $56 million in cash on hand to keep this loan, and that amount goes down as the principal reduces through repayments.
Problem Loan Number 2: DDTL 2.0
- At least 10.53% annual interest
- At least $760 million a year in interest payments
- $7.6 Billion ($3.8 billion drawn, $3.8 billion remaining)
CoreWeave's biggest, sexiest loan was also co-led by Blackstone and Magnetar, and allows it to draw up to $7.6 billion by June 2025 (with the option to extend for a further three months), with several fees (both upfront, at closing, and annually). These are calculated in rather esoteric ways.
Take, for example, the yearly fee. This is equal to 0.5% of the difference between $7.6 billion and the average outstanding debt on the loan, or $6.1 billion (at least $75 million), whichever is greater. In essence, CoreWeave pays a fee if it uses the loan and pays interest on whatever amount it chooses to borrow. The loan must be fully repaid in 60 months from whenever the money was drawn.
Per DDTL 1.0, a series of bizarre interest calculations based on standard/prime interest rates are involved, but DDTL 2.0 can also calculate further interest rate increases based on CoreWeave's credit rating. And there’s a huge scope for variation, with the acceptable range starting at 5% and ending at 12%. That eye-watering 10.53% interest rate I mentioned earlier — although far higher than what a consumer with decent credit could pay on a mortgage or car loan — isn’t necessarily the highest it could reach. It could get much, much worse.
These mob-like terms suggest Blackstone and Magnetar don't necessarily trust CoreWeave to survive, and intend to rip out whatever guts are left if it doesn't. CoreWeave's S-1 says that the actual average interest rate being charged on the amounts borrowed was 10.53%, but again, this could go up.
The number I've put above could as much as double to $1.52 billion a year (again, with no consideration of accrued interest) in the event that CoreWeave chooses to draw on the remaining $3.8 billion, something that I'm fairly confident will happen based on its aggressive capital expenditures.
DDTL 2.0 also has a brutal covenant — that if CoreWeave raises any other debt, it must use that debt to pay off this debt. This raises the question as to how it’ll manage to pay the DDTL 1.0 balloon payment, should any future debt raised be used to satisfy the DDTL 2.0 loan.
On the subject of future debt, the updated S-1 prospectus says that, in order to meet the requirements of OpenAI’s $11.9bn deal, CoreWeave will have to take on additional financing. How will it accomplish this while also remaining compliant with the terms of DDTL 2.0, particularly when it comes to how it’ll use the proceeds of any future borrowing?
The S-1 sheds some light here. The company has created a “special purpose vehicle” — essentially, a separate company owned and controlled by Coreweave, though technically distinct — that will “incur indebtedness to finance the obligations under the OpenAI Master Services Agreement.”
Number Three: CoreWeave Does Not Have Access To The Capital Necessary To Meet Its Obligations
All of this might seem a little dense, but it's actually pretty simple. CoreWeave made slightly under $2 billion in revenue in 2024, but somehow ended up losing $863 million. In effect, CoreWeave spends $1.43 to make $1.
As of January 2025, CoreWeave's obligations under DDTL 1.0 will likely reach $1bn a year, if not more. Starting from October 2025, it’ll need to start repaying the DDTL 2.0 loan, and these repayments will depend on whether it draws more capital from the loan, and whether its interest rate increases or decreases based on its perceived risk. Regardless, it’s not hard to imagine a scenario where its debt repayments surpass its entire 2024 revenue.
Furthermore, CoreWeave has made a lot of commitments. It’s planning to invest over a billion dollars to convert a New Jersey lab building into a data center, it’s part of a $5 billion effort with Blue Owl, Chirisa and PowerHouse, it’s committed to invest over a billion pounds sterling in UK-based data centers, it’s committed to invest an additional $2.2 billion in data centers in Europe, it’s committed to a $600 million data center project in Virginia, and allegedly have exercised an option with Core Scientific — a deeply dodgy company I'll describe in a minute — to create "approximately 500 Megawatts of Critical IT Load at Six Core Scientific Sites," with the agreement "increasing potential cumulative revenue to $8.7 billion over 12 Year Contract Terms."
In short, CoreWeave has committed to billions of dollars of data center buildouts, and the only way it can pay for them is with burdensome loans that it, as of right now, does not appear to have the revenue to support.
CoreWeave spent approximately $2.86 billion to make just under $2 billion, with $1.5 billion of that coming from the cost of running its infrastructure and scaling its operations, and the rest coming from hundreds of millions of dollars of interest payments and associated fees.
These numbers do not appear to include capital expenditures, and by its own admission, the vast loans that CoreWeave has pulled are necessary to continue funding them. Worse still, NextPlatform estimates that CoreWeave spent about $15 billion to turn its $7.5 billion of GPUs into around 360 Megawatts of operational computing power.
Per its S-1, CoreWeave has contracted for around 1.3 Gigawatts of capacity, which it expects to roll out over the coming years, and based on NextPlatform's math, CoreWeave will have to spend in excess of $39 billion to build its contracted compute. It is unclear how it will fund doing so, and it's fair to assume that CoreWeave does not currently have the capacity to cover its current commitments.
How does CoreWeave — a company with roughly $1.3 billion in the bank and more than $4.6 billion of debt available to draw and an inability to raise further capital without paying said debt off — actually continue doing business?
Number Four - CoreWeave Is Using A Suspicious and Unproven Partner To Build its Entire Infrastructure
Before We Go Any Further, A Note On "Contracted Power"
"Contracted power" does not necessarily mean that it exists. "Contracted power" is a contract that says "you will provide this much compute." This term is used to deliberately obfuscate the actual compute that a company has.
As of writing this sentence, CoreWeave has "more than 360" megawatts of active power and "approximately 1.3 GW of total contracted power." This means that CoreWeave has committed to building this much.
These figures will become relevant shortly.
In 2017, a company was founded with the goal of mining digital assets like Bitcoin to generate revenue, and to provide mining services for others. Several years later, it pivoted into providing compute for generative AI.
Confusingly, this company is not CoreWeave, but Core Scientific, a totally different company entirely that went public in 2022 in a disastrous SPAC-merger, and later filed for Chapter 11 bankruptcy that same year. It exited bankruptcy court in January 2024, having shed $400m in debt and restructured its obligations to creditors, and once again returned to the public markets, where it trades on the NASDAQ.
In June 2024, CoreWeave made an unsolicited proposal to acquire Core Scientific that it rejected (three days after announcing a 12-year-long deal with CoreWeave to provide 200 megawatts of compute), before signing an extension of an already-existent 12-year-long deal in August 2024 to deliver "an additional 112 megawatts of computing infrastructure to support CoreWeave's operations" according to CNBC.
This capacity, according to CNBC, will be operational by "the second half of 2026," and would involve repurposing existing crypto mining hardware. Here's a quote from CNBC about how easy that'll be:
Needham analysts wrote in a report in May that almost all infrastructure that miners currently have would “need to be bulldozed and built from the ground up to accommodate HPC,” or high-performance computing.
Great!
As of now, Core Scientific holds, according to CEO Adam Sullivan, "the largest operational footprint of Bitcoin mining infrastructure," and per the above analyst quote, it's very obvious that you can't just retrofit a crypto mining rig to start "doing AI," likely because the GPUs are different to the ASICs (Application Specific Integrated Circuits) used in crypto mining, meaning the server hardware is different, which means the entire bloody thing is different.
Nevertheless, CoreWeave's S-1 repeatedly mentions that it’s made an agreement with Core Scientific for "more than 500 MW of capacity."
Right now, however, it's unclear how much capacity Core Scientific actually has, despite both its and CoreWeave's suggestions. Core Scientific, as of February 2025, had approximately 166,000 bitcoin miners — which, I should add, are likely all application-specific chips that only mine bitcoin!, which means that none of that has (or, potentially, any of their data center operations have) anything to do with GPUs or compute for AI.
In fact, I can find little proof that Core Scientific has any meaningful compute capacity at all.
Once you dig into its financial filings, things get weirder. Per its most recent annual report for the year ending December 31, 2024, Core Scientific made $24.3 million in HPC hosting revenue (referring to high performance computing, which includes generative AI workloads).
That isn’t a typo. $24.3 million. By contrast, it generated $408m in revenue from mining and selling cryptocurrencies for itself, and $77m for mining crypto for third-parties.
Sidenote: Assuming it’s possible for Core Scientific to repurpose bitcoin miners for AI workloads, how does that help the business? As noted, mining crypto for resale, and for external partners, provides the overwhelming majority — nearly 95% — of its revenue.
Core Scientific has run at a loss for the last three quarters, losing $265 million in Q4 2024, $455 million in Q3 2024, and $804 million in June 2024.
Core Scientific has one HPC client: CoreWeave, which is referred to as “Customer J” in the 10-K form — the annual financial report that every publicly-traded company must publish at the close of each financial year.
Core Scientific, according to its 10-K form:
"...was contractually committed for approximately $1.14 billion of capital expenditures, mainly related to infrastructure modifications, equipment procurement, and labor associated with the conversion of a significant portion of its data centers to deliver hosting services for HPC... [with] $899.3 million [being] reimbursable by our customer under our agreements."
That customer — Core Scientific's only HPC customer — being CoreWeave, with the expenses expected to "occur over the next year."
How exactly will Core Scientific, a company that was bankrupt last time this year and lost over $265 million in its last quarter afford the up-front capital expenditures from CoreWeave's expansion? Core Scientific has around $836 million in cash and cash equivalents on hand and is still in the process of cleaning up its already-existent piles of debt, and even then...how does any of this work, exactly?
And given that the company recently exited Chapter 11 bankruptcy protection, it’s unlikely to receive capital on favorable terms.
Hey wait a second...in Core Scientific's latest 10-K, it proudly boasts that it has "approximately 1,317 MW of contracted power capacity to operate and manage one of the largest center infrastructure asset bases." CoreWeave's S-1 says that it has "...total contracted power extends to approximately 1.3 GW as of December 31, 2024, which we expect to roll out over the coming years."
Core Scientific's only customer (CoreWeave) is contracted to build 1.3 gigawatts of capacity, and mysteriously, that's exactly how much CoreWeave, Core Scientific's only customer, has said it’s contracted. While Core Scientific has said a chunk of that capacity is reserved for expanding its cryptocurrency-ming operations, it is still an extremely suspicious coincidence.
Nevertheless, Core Scientific, as of right now, does not appear to have any meaningful HPC infrastructure. While it may have seven data centers, that doesn't mean it’s able to meet the demands of companies like Microsoft and OpenAI, both customers of CoreWeave, as evidenced by the fact that it made $8.5 million in HPC revenue last quarter, and $24.3m for the entire financial year.
Somehow, Core Scientific intends to spend a billion dollars building HPC infrastructure, a thing it has yet to meaningfully do (it has, as of November, broken ground on a site in Oklahoma), and somehow deliver over a gigawatt of capacity to a company that will allegedly reimburse it, at some point, somehow, with the money they do not have.
What the fuck is going on?
CoreWeave Is Both A Time Bomb and a Bad Omen For Generative AI
To summarize:
- CoreWeave is burdened by interest payments that may balloon to more than $2 billion a year, and lost $863 million on $2 billion of revenue.
- 62% of that revenue is from Microsoft, which has materially pulled back on data center buildouts and may have dropped some CoreWeave contracts, though CoreWeave denies this is the case.
- CoreWeave's expansion, which is critical to servicing future revenue and growth, requires it to invest tens of billions of dollars that it does not have.
- CoreWeave's data center expansion is dependent on what is primarily a Bitcoin mining company — Core Scientific — that appears to have no current HPC capacity building out over a gigawatt of capacity at a time where it does not appear to have built any.
- Converting cryptocurrency mining data centers to HPC data centers is effectively starting from scratch. There is no easy or logical way to repurpose a bitcoin miner for AI.
If CoreWeave makes it to IPO — and it may do so as soon as next week — it will raise about $4 billion, which might give it enough runway to continue operations for a year, but by October 2025 it’ll face upwards of $500 million of loan payments a quarter, all while trying to scale up an operation that doesn't appear to have a path to profit.
The reason I've spent thousands of words walking you through CoreWeave's problems is that this is the first meaningful tech IPO in some time, and the first one directly connected to the AI boom.
CoreWeave's financial health and revenue status suggest that there either isn't demand or profit in providing services for generative AI. This company — the so-called backbone of the generative AI boom, and one of the largest holders of NVIDIA GPUs, with a seemingly closer relationship with the company than Meta or Microsoft, based on its early access to the company’s latest hardware — does not appear to be able to get meaningful business for its operations outside of hyperscalers. While it may sell by-the-hour compute to regular companies, it's clear that that market just doesn't exist at a meaningful revenue point.
If NVIDIA is selling the pickaxes for the gold rush, CoreWeave is selling the shovels, and it mostly appears to be turning up dirt. If this were a meaningful growth industry, CoreWeave would be printing money, just like how the automobile created an entire generation of billionaire oil barons, like John D. Rockefeller and Henry Flagler. And yet, it appears that, outside of Microsoft, it can't even scrape a billion dollars of revenue out of being the single-most prominent independent provider of AI compute.
Furthermore, it's unclear how CoreWeave actually intends to expand. Core Scientific is a tiny, unproven party that has yet to build an HPC data center, one that has to front the money for CoreWeave's expansion in the hopes that it’ll be reimbursed. Building data centers isn't easy, and Core Scientific's previous work as a Bitcoin mining firm does not necessarily apply thanks to the massively-different server architecture involved with running superclusters of GPUs.
CoreWeave should have been a positive signal for generative AI, or at least a way for AI boosters to shut me up. If generative AI had this incredible demand — both from companies looking to integrate it and users looking to use it — CoreWeave would be making far, far more money, have a far more diverse customer base, and, if I'm honest, not have to take out more than five times its revenue in burdensome loans with loan shark-level interest rates.
In reality, this company is a dog, and will show the markets exactly how little money or growth is left in generative AI. NVIDIA's remarkable GPU sales have been conflated with the success of generative AI, rather than seen as a sign of desperation, and a signal of how willing big tech is to spend billions of dollars on something if they think their competition is doing so.
Really, the proof is in the use of those GPUs, and CoreWeave gives us a transparent — and terrifying — expression of the lack of excitement or real usage of generative AI. As I hypothesized a few weeks ago, I believe that outside of OpenAI, the generative AI industry is terribly small, a point that CoreWeave only underlines.
Based on its revenue, how much could Amazon Web Services, Google Cloud, or Microsoft Azure really be making? Based on reporting by The Information, OpenAI spends roughly $2 billion on the compute to run its models and a further $3 billion to train its models, paying Microsoft a discounted rate of around 25% the normal cost. Even in the most optimistic figures, given how much bigger and more popular ChatGPT is than literally every other generative AI company, how can any hyperscaler be making more than $3 billion or 4 billion in revenue a year from selling AI compute?
Without conceding that generative AI has a future beyond the frothy present, one also has to question whether there’s even much of a place for massive hyperscaler investment, given the rise of new, more efficient models. I’m not merely talking about DeepSeek. The largest version of Google’s newest Gemma 3 model can run on a single H100 GPU, and according to Sundar Pichai, requires one-tenth the computing power as similar models. Separately, Baidu’s ERNIE 4.5 model reportedly has one-hundredth the computational demands as GPT-4.5, while delivering similar performance, and its X1 reasoning model allegedly outperforms DeepSeek R1 at half the cost.
These numbers also suggest that OpenAI is likely charging way, way less than it should be for its services. If it costs CoreWeave $493 million (yes, this is napkin math) — this is its "cost of revenue," and that amount only includes rentals, power and personnel to run its services — to service 360 megawatt of power, and Microsoft's 7.5 gigawatts of power is, say, 70% OpenAI's compute, it may cost Microsoft over $7 billion. It's already been well-established that OpenAI's costs eat into Microsoft's profits.
Again, these are estimates, as we don't know Microsoft's exact costs, but it's reasonable to believe that its contracted compute with CoreWeave was likely to facilitate OpenAI's growth, which is further ratified by The Information's AI data center database, which reports that the upcoming data center buildout in Denton, Texas is, and I quote, "...[for] Microsoft [to] rent to use by OpenAI."
And you'll never guess who's building it. That's right, Core Scientific, which announced on February 26 2025 that it was partnering with CoreWeave to expand its relationship across its Denton, Texas location.
It's unclear how this data center gets built, or whether OpenAI will actually use it given its new plans for a "Stargate" data center in Abilene Texas, and the general chilling of the relationship with Microsoft. Furthermore, Microsoft's indeterminately-sized cancellations with CoreWeave pair with its own retreat from data center buildouts, which coincides with Microsoft releasing OpenAI from its exclusive cloud compute provider relationship, which coincides with OpenAI's plans to build gigawatts of its own capacity.
How, exactly, does any of this make sense?
While I can only hypothesize, I believe that this move is Microsoft's attempt to disconnect from OpenAI, dumping its exclusive relationship and canceling its own capacity expansion along with contracts with CoreWeave, citing, according to the Financial Times, "delivery issues and missed deadlines," which would make sense, as it appears that CoreWeave's infrastructure partner does not appear to have expertise in building data center capacity, or any actual cloud compute capacity that I can find.
Think about it. It was reported last year that OpenAI was frustrated with Microsoft for not providing servers fast enough, after which Microsoft allowed OpenAI to seek other compute partners, which in turn led to OpenAI shacking up with Oracle and SoftBank to build out the future of OpenAI’s compute infrastructure. Once this happened, Microsoft decided to (or had already been in the process of) massively reduce its future data center capacity at a time when OpenAI’s latest model necessitates bringing hundreds of thousands of GPUs online.
Even if you disagree with my thesis, how is Microsoft going to support OpenAI’s growth any further? OpenAI’s latest models o-3 and GPT 4.5 are more compute-heavy than ever. How, exactly, does canceling over a gigawatt of planned capacity make sense?
It doesn’t. And I think we’re about to see what happens when the world’s biggest startup becomes desperate.