CoreWeave’s S-1 Is Finally Here: AI’s Golden Child Under the Hood

Five Hours In—Financial Gymnastics, GPU Bets, and Microsoft’s Grip

I often get questions about accessing the monthly data center construction info—it comes from the U.S. Census Bureau. I extract and visualize the data myself, and I update it monthly so people can check it directly instead of messaging me for details. Here is the link.

---

CoreWeave just dropped their S‑1 a few hours ago, finally. Originally known as Atlantic Crypto, the company launched in 2017 mining Ethereum before rebranding as CoreWeave to focus on GPU-based graphics rendering and AI after crypto prices fell. In 2024, revenue soared more than 700% to $1.92 billion and debt approached $8 billion. CoreWeave has filed to trade on Nasdaq as “CRWV” and aims to raise more than $3 billion, targeting a valuation above $35 billion.

It’s an intricate time: we’re seemingly at an inflection point of the AI data hype. On one hand, you see the latest optimization techniques taking hold and narrative change about what is the next stage of AI models. And then there’s the stock drop in NVIDIA and Microsoft canceling its data center lease. Meanwhile, our own data center growth tracker shows year‑on‑year growth slowing down. Of course, you can interpret these signals in different ways—anyway here’s my first impression of CoreWeave’s data points.

Revenue, CapEx, and Net Losses

Revenue Growth:

• 2022: $16 million

• 2023: $229 million

• 2024: $1.9 billion

CoreWeave’s revenue has soared from a modest $16 million to nearly $2 billion in just two years. This meteoric rise hints at the explosive demand for AI compute, even as the company continues to invest heavily in infrastructure.

Capital Expenditures (CapEx):

• 2023: ~$3.1 billion

• 2024: ~$8.7 billion

Billions are being poured into next‑gen data centers packed with cutting‑edge NVIDIA GPUs. These massive investments are financed partly through over $8 billion in debt, leading to significant interest expenses and a corresponding drag on net income.

Net Losses:

• 2022: –$31 million

• 2023: –$593.7 million

• 2024: –$863.4 million

Despite the revenue surge, CoreWeave is burning cash—illustrating a classic high‑risk, high‑reward scenario where aggressive expansion today bets on dominating tomorrow’s AI landscape.

6‑Year GPU Depreciation: A Calculated Long Game

One of the more intriguing aspects of CoreWeave’s financials is its decision to depreciate its GPUs over six years, using a straight‑line (linear) method. While traditional industry standards once pegged depreciation at three years—and many public cloud providers have since extended that to around four or even 4.5 years—CoreWeave’s six‑year schedule is on the higher side. These GPUs are expected to remain useful for secondary workloads well past their “cutting‑edge” phase. It’s a calculated bet: lower annual depreciation boosts short‑term operational metrics, though it carries the risk that technological obsolescence might outpace the accounting schedule.

From -863M Net Loss to +1.2B Adjusted EBITDA

One very interesting aspect of CoreWeave’s financials is how they transform an $863 million net loss into an adjusted EBITDA of roughly $1.2 billion. Their Remaining Performance Obligations (RPO) are much more indicative of CoreWeave’s business development and future trajectory—showing the long-term revenue commitments that back their massive infrastructure investments. At the same time, the financial engineering is fascinating. This swing is driven by three non-cash adjustments.

• Loss on Fair Value Adjustments: This non-cash expense arises from the periodic revaluation of CoreWeave’s derivative instruments, such as warrants or other embedded options. These instruments often give holders the right to purchase shares—either as part of employee compensation or in financing arrangements. As CoreWeave’s valuation increases, the fair value of these warrants (or similar instruments) also rises, meaning that, on paper, the company’s liability increases. Even though no cash is actually spent, this increase is recorded as a loss on the income statement. However, since it’s a non-cash charge, it’s added back when calculating adjusted EBITDA, to better reflect the company’s underlying operating performance.

• Depreciation & Amortization: Driven by CoreWeave’s enormous capital expenditures—which jumped from about $3.1 billion in 2023 to roughly $8.7 billion in 2024—the annual depreciation expense amounts to nearly $863 million. This non-cash expense reflects the gradual write-off of these investments over time.

• Interest Expense: With CoreWeave carrying over $8 billion in debt to finance its rapid expansion, significant interest expenses are recorded. These costs, while real cash outflows, are excluded from adjusted EBITDA to better reflect the company’s core operating performance.

250K GPUs

In its S-1, CoreWeave disclosed staggering GPU procurement numbers. Unlike hyperscalers like AWS or Google, CoreWeave must prove to NVIDIA that it has both customer commitments and data center capacity before receiving GPU allocations. To achieve this, the company aggressively locks in long-term, take-or-pay contracts with AI firms, ensuring demand is guaranteed before GPUs are purchased. Microsoft plays a crucial role in this strategy—not only is it CoreWeave’s largest customer, but it has also helped expand CoreWeave’s data center footprint in North America and Europe. One of the biggest questions around CoreWeave’s meteoric rise is how it managed to secure NVIDIA GPUs ahead of major hyperscalers like Microsoft, Google, and AWS. The answer lies in a mix of supply agreements, customer commitments, and strategic positioning. Unlike hyperscalers, which can shift workloads across their own infrastructure, CoreWeave must demonstrate to NVIDIA that it has both a customer and data center capacity lined up to even qualify for GPU allocations. This creates a different urgency—CoreWeave locks in power and capacity early, which hyperscalers don’t always need to do at the same scale. Additionally, Microsoft, one of CoreWeave’s largest customers, relies on it for surge capacity, meaning CoreWeave plays a key role in filling AI compute gaps when demand spikes. NVIDIA’s allocation strategy also appears to favor third-party cloud providers like CoreWeave, Lambda, and Paperspace in some cases before prioritizing hyperscalers. This has allowed CoreWeave to secure GPUs more predictably than even Microsoft, which sometimes depends on CoreWeave as a backup supply source. Furthermore, CoreWeave has aggressively financed these GPU purchases with billions in asset-backed debt, betting on sustained AI demand while securing early access to NVIDIA’s latest GPUs, including Blackwell, before larger cloud providers can ramp up supply.

Long‑Term Contracts & RPO: Concentrated Revenue in a Risky Market

CoreWeave follows a strict “don’t contract unless you have a contract” mantra—it won’t invest in new infrastructure unless a long‑term, take‑or‑pay agreement is in place. This means that every expensive new GPU cluster is effectively pre‑sold, locking in revenue and justifying massive CapEx. This model works when you have a highly concentrated customer base, and in 2024, CoreWeave’s largest customer—Microsoft—accounted for 62% of its revenue.

Microsoft and CoreWeave have forged an uneasy yet strategic alliance in the AI cloud arena. Microsoft depends on CoreWeave’s nimble, GPU-powered data centers to fill capacity gaps, particularly for burst-demand and niche workloads that its own hyperscale infrastructure can’t immediately cover. At the same time, Microsoft has underwritten a significant portion of CoreWeave’s expansion by signing multi‑year, take‑or‑pay contracts. These agreements guarantee a steady revenue stream for CoreWeave, reducing financing risk and subsidizing the rapid scale-up of their specialized, liquid‑cooled facilities.

However, Microsoft’s internal capacity planning also indicates that once workloads become predictable and long-term usage patterns emerge, many customers may eventually shift to Microsoft’s own reserved capacity. This reserved capacity, integrated within Microsoft’s hyperscale data centers, not only offers more favorable pricing but also delivers a fully bundled, optimized experience through a suite of cloud services. In essence, while third-party GPU clouds like CoreWeave are essential for short-term agility and meeting volatile demand, the long-term stability and cost efficiency of reserved capacity could gradually cannibalize that external reliance, as enterprises lean toward the consistency and long-term savings of a directly managed solution.

CoreWeave’s Dual-Class Share and Ownership

There’s been a lot of confusion over CoreWeave’s stock structure, particularly around what Class A and Class B shares actually mean. While Class B shares don’t trade publicly and exist primarily for control, they still hold economic value because they can be converted into Class A shares at any time. This means that even though IPO valuation is based on the market price of Class A shares, Class B shares are still included in the total shares outstanding—so they factor into overall economic ownership. When founders sell stock, they first convert Class B into Class A, reducing both their economic stake and, eventually, their voting power if they continue selling.

Dual-class structures like this aren’t unusual—Meta, Lyft, and Snap have similar setups—but they create tension between public investors and insiders. Founders can cash out significant sums while still maintaining control, which raises governance concerns. In CoreWeave’s case, the founders collectively sold nearly $500 million pre-IPO, but because they retained Class B shares, they still hold supermajority voting power. So while Class B doesn’t directly impact valuation, it absolutely affects ownership and control, which is why this structure often leads to debate.

Calculating Economic Ownership

Economic ownership represents the percentage of the total company that a person or entity owns. It’s calculated as follows:

Economic Ownership (%) = (Total Shares Owned (Class A + Class B) / Total Shares Outstanding) × 100

Pre-IPO

• Total shares outstanding: ~825 million

• Founders’ stakes:

• Michael Intrator: (2.39% of Class A + 47.33% of Class B) ≈ 12.15%

• Brian Venturo: (0.53% of Class A + 31.61% of Class B) ≈ 8.88%

• Brannin McBee: (0.24% of Class A + 24.22% of Class B) ≈ 6.60%

Collectively, the founders owned roughly 27.63% of the company pre‑IPO.

Post-IPO

• Total shares outstanding: ~1.1 billion

• Founders’ stakes:

• Michael Intrator: (1.9% of Class A + 38.4% of Class B) ≈ 10.87%

• Brian Venturo: (0.5% of Class A + 25.6% of Class B) ≈ 7.99%

• Brannin McBee: (0.2% of Class A + 19.6% of Class B) ≈ 5.93%

After dilution, the founders’ total economic ownership drops to roughly 24.79%. Yet, thanks to their Class B shares, they still control over 80% of the voting power.

Pre-IPO Tender Offers & Secondary Sales: Cashing Out While Retaining Control

Before the IPO, CoreWeave’s insiders monetized their stakes through two major tender offers, allowing them to secure hundreds of millions in cash without relinquishing control.

2023 Tender Offer

Price: $309.86 per share

Total Shares Sold: ~2.07 million for ~$643 million

Key Details: Fidelity snapped up roughly 1 million shares for $310 million.

Founders’ Sales:

• Michael Intrator sold 354,931 shares for $110 million.

• Brian Venturo sold 231,126 shares for $72 million.

• Brannin McBee sold 172,368 shares for $53 million.

2024 Tender Offer

Price: $939.85 per share

Total Shares Sold: ~692,562 shares for ~$651 million

Key Details:

• Magnetar acquired 96,749 shares for $90.9 million.

• Fidelity purchased 70,817 shares for $66.5 million.

• Glenn Hutchins’ funds bought 10,639 shares for $9.9 million.

Additional Founder Sales:

• Venturo sold about $105 million worth.

• McBee sold about $97 million worth.

• Intrator sold about $50 million worth.

Total Founder Share Sales Pre‑IPO: Approximately $487 million+

Even though the founders have sold a significant amount through these tender offers, they’ve maintained control via their Class B shares, ensuring that while their economic stake might shrink, their voting power remains dominant.

CoreWeave’s Highly Unconventional Debt Structure

CoreWeave’s capital structure is one of the most unusual in the cloud computing industry, and it’s drawing intense scrutiny from investors and financial analysts alike. Unlike traditional tech companies, which finance expansion through a mix of equity raises and corporate bonds, CoreWeave has structured its debt using a web of asset-backed debt, customer-tied interest rates, and high-yield financing from private credit giants like Blackstone and Magnetar.

Here’s what makes CoreWeave’s debt so unique—and why it could either supercharge its expansion or weigh down its future profits.

1. How CoreWeave Finances GPUs & Data Centers: Asset-Backed SPVs

CoreWeave doesn’t hold most of its debt directly on its balance sheet. Instead, it finances its GPUs, networking infrastructure, and data centers through Special Purpose Vehicles (SPVs)—separate legal entities that issue debt backed by specific revenue-generating assets.

This structure has two key benefits:

• Lower risk for lenders → Since the debt is backed by tangible assets (like GPU clusters) and contracted customer revenue, lenders have stronger collateral and less risk.

• Better financing terms for CoreWeave → Instead of issuing expensive corporate debt or diluting equity shareholders, CoreWeave can raise billions in debt secured by predictable revenue streams.

This type of financing is common in real estate or infrastructure projects, but highly unusual in tech—making CoreWeave one of the few companies applying asset-backed securitization to cloud computing.

2.The Private Credit Play

$324M operation income vs $361M interest expense

Rather than borrowing from traditional banks or issuing corporate bonds, CoreWeave’s largest lenders are Blackstone and Magnetar—two of the biggest names in private credit and structured finance. These firms specialize in high-yield, complex financing deals and typically lend to companies that need flexible but expensive capital. Unlike traditional tech lenders, they are in this to extract high returns.

CoreWeave’s debt structure prioritizes its lenders in cash flow distribution. According to its S-1, a substantial portion of its cash flow must go toward servicing debt, including both principal and interest payments, which directly reduces its ability to reinvest in operations, fund growth initiatives, or pursue capital expenditures .

Additionally, CoreWeave’s credit agreements impose strict financial covenants that limit its ability to take on more debt, issue dividends, or make significant business changes—such as acquisitions or mergers—without lender approval . This means CoreWeave’s ability to maneuver strategically is constrained by the demands of its debt holders.

3. Customer-Tied Interest Rates

One of the most unusual aspects of CoreWeave’s debt structure is its customer-linked variable interest rates—meaning the cost of borrowing fluctuates based on the creditworthiness of CoreWeave’s customers.

CoreWeave’s Delayed Draw Term Loan Facilities (DDTL 1.0 and DDTL 2.0) are structured so that the interest rates on outstanding loans depend on the financial strength of the customers whose contracts back the debt. Specifically:

• Investment-grade (IG) customers, like Microsoft, allow CoreWeave to borrow at lower rates, with spreads of 6.5% over SOFR (Secured Overnight Financing Rate) or 5.5% over base rate loans.

• Non-investment-grade (non-IG) customers, which could include startups or riskier AI firms, push the interest rate significantly higher—up to 13.0% over SOFR or 12.0% over base rate loans.

This structure makes sense because CoreWeave’s business is built on long-term, take-or-pay contracts. The more creditworthy the customer, the less risky the revenue stream, allowing CoreWeave to secure cheaper financing.

Since Microsoft accounts for 62% of CoreWeave’s revenue, this dynamic directly lowers CoreWeave’s borrowing costs. In contrast, if CoreWeave were to shift toward a customer base filled with smaller, riskier AI firms, its overall financing costs would rise significantly—potentially straining profitability.

This debt structure ties CoreWeave’s financial health even more closely to Microsoft—not just as a customer, but as an indirect subsidy for its cost of capital. If Microsoft were to reduce its contracts or shift to in-house capacity, CoreWeave’s cost of borrowing could spike, making its expansion far more expensive.

4. The Imbedded Put Option

CoreWeave’s Series C convertible preferred stock also comes with a put option that allows holders to force the company to repurchase 29,874,066 shares of Class A common stock at $38.95 per share, amounting to roughly $1.2 billion. This right kicks in after either August 15, 2029, or two years post-IPO closing—whichever comes first. It terminates if the stock’s 20-day volume-weighted average price (VWAP) hits $68.16 (175% of $38.95) over any consecutive 30 trading days after the IPO, or if certain events like share sales occur beforehand. This mechanism puts CoreWeave on the hook for a cash payout if exercised.

If triggered, this $1.2 billion repurchase obligation could strain CoreWeave’s finances, especially given its $8 billion debt load and hefty capital expenditures ($8.7 billion in 2024 alone). On the flip side, it doubles as a performance carrot—pushing the company to boost its stock price above $68.16 to dodge the payout entirely. The put option is held by heavy-hitting Series C investors like Coatue, Magnetar, whose influence could sway CoreWeave’s strategy and market perception. For IPO investors, it’s a mixed bag: the put pressures CoreWeave to deliver strong returns, but it also looms as a financial wildcard that could dent long-term value if the company stumbles or if these investors cash out at the expense of broader shareholder interests.

5. HoldCo Guarantees & Debt Risk: What Happens If Something Goes Wrong?

While the SPVs technically hold CoreWeave’s debt, the parent company (HoldCo) guarantees it. This means:

• If the revenue generated from GPU contracts isn’t enough to cover the debt payments, CoreWeave’s broader balance sheet is still responsible for repayment.

• Even though SPVs are designed to provide a legal separation, the HoldCo guarantee limits how much risk CoreWeave is actually shifting away from its core business.

This blurs the line between traditional corporate debt and asset-backed financing, meaning CoreWeave can’t entirely escape its liabilities if demand slows or contracts get canceled.

6. Who Actually Owns CoreWeave’s Future Profits?

Because CoreWeave’s entire AI cloud expansion is being financed through structured debt, many analysts believe debt holders, not equity investors, will extract most of the company’s economic value for years.

This raises two key questions:

1. How much free cash flow does CoreWeave actually get to keep after paying its debt?

2. What happens when these loans mature in five years? Will CoreWeave be able to refinance, or will it need to raise dilutive equity?

If AI demand keeps growing, CoreWeave could continuously refinance its debt and scale, effectively operating on an “evergreen” debt model. But if AI demand slows—or if customers like Microsoft start shifting workloads to their own data centers—CoreWeave could face significant financial stress.

For now, CoreWeave is playing a high-stakes financial game—leveraging asset-backed loans, customer-linked interest rates, and strategic partnerships to fuel its meteoric rise. But whether this structure remains sustainable in the long run is still an open question.