Comment CoreWeave this week said it would plow between $20 and $23 billion into GPU bit barns by year’s end in order to meet growing demand from model builders and hyperscalers.
Those figures, along with new commitments from OpenAI and yet unnamed hyperscale customers, were no doubt intended to instill confidence in the rent-a-GPU outfit’s long-term growth prospects.
In reality, the company’s capex outlay illustrates just how fragile the Nvidia-backed AI infrastructure vendor’s business model really is, despite what its triple-digit revenue gains this quarter might suggest.
In its first earnings call since its March IPO, CoreWeave revealed Q1 revenues surged 420 percent year over year to $982 million, far exceeding expectations. Despite this growth, it posted a net loss of $314.6 million. Still, the GPU-slinger expects to continue the momentum over the next quarter, forecasting revenue of between $1.06 billion and $1.1 billion in Q2 and $4.9 to $5.1 billion for the full year.
Even if this revenue comes in as expected, it won’t be enough to sustain the level of investment CoreWeave has in mind. So, how are they going to make it work?
By gambling with someone else’s money.
Unlike more traditional hyperscalers, like Microsoft, which can actually afford to invest tens of billions of dollars a year in AI, CoreWeave’s business model requires juggling a mountain of debt.
A debt-driven gamble on fast-depreciating assets
On its face, the whole affair looks a bit like your typical real estate investment trust (REIT). CoreWeave uses capital raised through funding rounds, debt financing, and now its IPO to build massive datacenters packed with tens of thousands of Nvidia’s Hopper and now Blackwell GPUs.
Over the past few years, CoreWeave has raised a combined $17.2 billion in debt financing and equity, including IPO proceeds, with major contributions from backers like Blackstone, BlackRock, and Magnetar Capital to fuel its expansion. Its current balance sheet shows nearly $9 billion in current and non-current debt.
The difference is that real estate generally appreciates in value and can be amortized over a period of decades, making it a relatively safe investment. GPUs, by comparison, are depreciating assets with a finite shelf life.
“I said before that when Blackwell starts shipping in volume, you couldn’t give Hoppers away,” Nvidia CEO Jensen Huang said during his GTC keynote back in March. “There are circumstances where Hopper is fine. That’s the best thing I could say about Hopper. There are circumstances where you’re fine. Not many.”
Jensen was mostly joking there. Nvidia’s Hopper generation isn’t worthless quite yet, but the point remains. The march of technology is unrelenting and has only accelerated as chipmakers have transitioned to yearly release cadences for new GPUs.
So, while CoreWeave’s cash flow needs only to cover its operating expenses and loan payments, it only has a short period of time to extract all the value it can from these chips before the next generation arrives.
“Our capital expenditures are success-based,” CoreWeave CFO Nitin Agrawal explained on this week’s earnings call. “We enter into compute capex programs when we sign multiyear contracted revenue that more than covers the cost of the capex within the contract terms.”
In other words, CoreWeave’s capital expenditures are matched to long-term compute contracts with hyperscalers and AI labs, like OpenAI, which agree to lease a certain number of GPUs for so many years at a discounted rate over spot instances.
So long as CoreWeave’s cash coming in from long-term GPU leases is greater than its operating costs and loan payments, it should break even.
The risk, of course, is if a customer runs out of cash before it has fulfilled its contract — something that reportedly happened to British model builder Stability AI prior to its restructuring last year.
In the current market, that’s not really a problem. If one customer folds up shop, there’s probably another already waiting in line to take their place.
But if enough customers stop paying their GPU bills, CoreWeave might not be able to keep up with its loan payments, and getting another round of debt financing to pay for the next generation of Nvidia compute might not be so easy.
As it stands, CoreWeave claims to have a revenue backlog of $25.9 billion as of March 31, but it’s only contractually obligated to receive $14.7 billion of that.
CoreWeave’s biggest customers are also its biggest competitors
This wouldn’t be inherently problematic, especially if that risk were distributed across a diverse portfolio of customers, but this doesn’t appear to be the case. The reality is there just aren’t that many companies that can afford the multi-billion dollar leases necessary to maintain CoreWeave’s momentum. Making matters worse, its largest customers also happen to be its biggest competitors.
According to JP Morgan analyst Mark Murphy, just two companies drove 77 percent of CoreWeave’s revenues in 2024, with Microsoft accounting for 62 percent of that.
Along with Microsoft, Google, and IBM have also signed contracts with the neo-cloud, and during its Q1 earnings call, CoreWeave CEO Michael Intrator touted a new commitment from OpenAI that will add $11.2 billion to its revenue backlog, along with a separate deal from a yet-unnamed hyperscaler.
The rub is that all of these companies are working to expand their own datacenter footprints. Microsoft plans to invest $80 billion into AI this year, Google will spend $75 billion, while OpenAI is working with Oracle and SoftBank to plunk $100 billion into a network of AI supercomputers as part of its Star Gate initiative.
We’ve already seen some evidence of Microsoft pulling back on datacenter leases in favor of its own datacenter designs.
CoreWeave can probably count on these hyperscale customers to see out their contracts, but what happens three, four, or five years from now, when those contracts are up? Will CoreWeave’s biggest customers still need them, or will they have ample capacity in their own datacenters?
The answer to that will likely depend on whether demand for generative AI continues to grow at its current pace. For now, enterprise adoption of AI remains in its infancy. A recent IBM CEO survey found that just a quarter of AI initiatives were actually paying off.
Proponents of the tech predict the AI market opportunity will surge to $400-$500 billion by 2028, but cracks are starting to show. This week, it was reported that Meta had delayed the release of its Llama 4 Behemoth model, which has apparently failed to live up to expectations despite its massive 2 trillion parameter count. ®
