In the tradition of previous quick S-1 teardowns (Snowflake, Palantir, Confluent, Klaviyo, Cerebras, etc), some quick notes on the CoreWeave S-1 from my colleague Aman Kabeer and I. As in prior efforts, this is not meant to be 100% comprehensive (and it’s certainly not investment advice!).
The CoreWeave IPO is going to be fascinating to watch: partly because it is undeniably exciting, and partly because it is not going your standard tech IPO. It presents a profile that in some ways is typical of a hyper-growth tech unicorn (explosive growth, large losses, dual-class stock structure with Class A/B shares, etc.), but in other ways, it is very unlike most tech IPOs of the past. Its specialized business model, heavy infrastructure focus, heavy customer concentration (Microsoft), reliance on big partners (NVIDIA), financial structure ($7.9B in debt) and unusual risk factors make it a unique case that blends the characteristics of a cloud provider, a hardware company, and a startup riding an exciting but also sudden and unproven market wave.
SOME KEY TAKEAWAYS
Frivolity: Crypto pivot, NJ in da house: For all the jokes on social media about founders, startups and VCs pivoting from Web3 to AI, CoreWeave is an example of a business that started as a crypto mining operation, stockpiled GPUs, and pivoted to AI with spectacular success.
In the same (frivolous) vein, for all the “all AI is in SF” mantra, CoreWeave is headquartered in… New Jersey.
The First Generative AI IPO: Depending on pricing, there’s likely going to be tremendous interest in the IPO. This is in part because of the well-documented dearth of tech IPOs, but most importantly because it’s the first IPO of the Generative AI era (Cerebras, as far as we know, is still stuck in CFIUS review of their relationship with G42).
There are very few “AI pure plays” on public markets – Palantir is one example (cue in the never ending discussion as to whether they are truly an AI company), as is (even more arguably) C3 AI. Other than that, the way to “play” Generative AI has been to invest in Mag 7 companies.
It is no accident that the first companies to file S-1s in the Generative AI era are infrastructure companies. The market has been forming supply-first (chips, data centers, foundation models), with the major hope that the demand side surfaces equally meaningfully in years to come.
Not a Real Estate Play: A negative take on CoreWeave and comparable companies one would often hear in tech circles is that the company is a “real estate play”, with limited technology and software. The argument seemed to be supported by the fact that the co-founders of the company come from a financial, rather than technological, background.
In its S-1, CoreWeave does a strong job dispelling that notion, positioning itself as the “AI Hyperscaler”. The current hyperscalers were built “to host websites, databases, and SaaS apps that have fundamentally different needs than the high performance requirements of AI”, hence a “massive need” for a purpose-built AI cloud platform, including the infrastructure and integrated software.
Its potential acquisition of Weights & Biases seems to confirm its major push towards being a software player.
Explosive Revenue Growth: The Generative AI era certainly has seen its share of incredible growth stories (OpenAI, Cursor, etc), but they actually pale in comparison to Coreweave’s growth – from about $15.8 M in 2022 to $228.9M in 2023, and then $1.92B in 2024, a remarkable +737% Y/Y growth last year,, reflecting surging demand for AI compute services
Large (and Growing) Losses: At the same time, CoreWeave remains deeply unprofitable. It incurred a net loss of $31M in 2022, which widened to $593.7 million in 2023 and $863.4 million in 2024 (-45% margin). In other words, expenses grew almost as fast as revenue – the 2024 net loss was about 45% of revenue (an improvement from 2023, when the net loss exceeded 2× annual revenue).
Margins and Cost Structure: CoreWeave’s gross margin is high (74% in FY’24), indicating that providing GPU compute at scale can be quite profitable before overhead. However, operating expenses and other costs are massive – including depreciation of expensive hardware, data center operations, and interest on debt – leading to negative operating margins. In short, the core service is high-margin, but the heavy infrastructure investments and financing costs currently outweigh those gains.
As an example of massive investment required: the spectacular amount of fiber buildout to support even a single 32K GPU cluster (~600 mi of fiber cables, ~80K fiber connections). In many ways, this resembles the physical buildout required to support the dot com era
That said, given capital intensity of the business, with high non-cash expenses (D&A, Loss on Fair Value Adjustments) driving GAAP efficiency down, the company is actually highly efficient on a non-GAAP basis, with ~$1.2B in Adj. EBITDA in FY’24 (~64% Margin)
The Coreweave IPO will be an interesting bellwether for where public markets lean on Rev. Growth vs. [GAAP] Efficiency Question
Customer Concentration: CoreWeave’s revenue is highly concentrated in a few large customers. Notably, Microsoft accounted for ~62% of CoreWeave’s revenue in 2024
– an extremely large single-client share. A handful of other tech and finance firms (such as Meta, IBM, and hedge fund Jane Street) make up much of the remaining revenue, meaning the company depends on a small number of big spenders. This concentration presents a risk: the loss of any major customer (or even a cutback in their usage) would have a material impact on revenue. The IPO filing highlights this reliance as a key risk, as such heavy dependence on one or two customers is unusual for a company of this scale.
Long-Term Commitments and Expansion Motion
Current customer base of AI Labs / Model Developers & AI Enterprises (MSFT, NVIDIA, IBM, Meta, etc) are not making short-term bets, with multi-year agreements with CoreWeave as evidence.
$15.1B in Remaining Performance Obligations (RPO) as of 12/31/24 & 4-year weighted average contract length
Committed contracts comprised an astonishingly high 96% of FY’24 Revenue
In addition, the expansion motion is already working.
“Three of our top five committed contract customers by TCV as of December 31, 2024 signed agreements for additional capacity within 12 months of their respective initial purchase dates. These agreements, measured during each respective 12-month period from the initial date of signing, represent a cumulative increase of approximately $7.8 billion in committed spend and a multiple of approximately 4x on initial contract value.”
True Enterprise AI is not here yet
What does the CoreWeave S-1 tell us about the reality of the enterprise AI market?
As noted above, the key customers are other infrastructure and technology providers – MSFT, Meta, NVIDIA, IBM
For the rest, it’s still early. From the S-1:
“Extend into broader enterprise customers across new industries and verticals, including regulated industries like banks, high-frequency trading, and pharmaceutical companies, as they begin to develop and build their own dedicated AI solutions. We anticipate that new industries and use cases will arise to take advantage of developing AI capabilities as AI models become more accessible and cheaper, presenting additional growth opportunities for our platform.”
“As AI continues to find product market fit, we expect these enterprises to grow their indirect consumption of our platform through AI labs”
Supplier Dependence (NVIDIA): CoreWeave is almost entirely dependent on Nvidia for current and future GPUs. Any disruption in Nvidia’s supply chain, a decision by Nvidia to prioritize other buyers, or Nvidia’s pricing power could hurt CoreWeave. Additionally, if Nvidia’s GPU technology were to fall behind or alternative AI chips rise, CoreWeave’s single-vendor strategy could leave it vulnerable. This “all eggs in one basket” supply approach is a key risk (especially given how critical GPUs are to the business).
ADDITIONAL NOTES
Company / Background
- CoreWeave was founded in 2017, originally as a crypto mining company ‘Atlantic Crypto’, by Michael Intrator (CEO), Brian Venturo (Chief Strategy Officer) and Brannin McBee (CDO)
- Rebranded in 2019 to CoreWeave and pivoted into providing compute resources ahead of the AI boom, given existing stockpile of GPUs for their mining operations
- CoreWeave is building the ‘AI Hyperscaler’, providing cloud-based GPU computing resources, storage & purpose-built software services for AI-native workloads & applications. Three key components underpin CoreWeave’s full stack GPU cloud platform:
- Infrastructure Services – access to GPU & CPU compute, with high-performance networking purpose-built for AI workloads, and storage
- Managed Software Services – Managed Kubernetes environments, VPC offering & bare metal offering
- Application Software Services – SUNK (ability to intelligently schedule jobs on top of Kubernetes on a single cluster), CoreWeave’s proprietary Tensorizer (efficiency & model loading optimizations) & inference optimization services
- CoreWeave’s data center network, with 27 locations in the US & 4 locations in Europe (expansion in 2024), is optimized for AI-specific workloads. Company advertises up to 20% improvement on MFU vs. existing hyperscalers
Financials
GAAP
- Revenue
- FY’24: $1.9B (+737% Y/Y)
- FY’23: $229M (+1346% Y/Y)
- Gross Margin
- FY’24: $1.4B (~74% Gross Margin)
- FY’23: $160M (~70% Gross Margin)
- Operating Income
- FY’24: $324M (~17% Margin)
- FY’23: $(14.5)M (~(6)% Margin)
- Net Loss
- FY’24: $(863)M (~(45)% Margin)
- FY’23: $(594)M (~(259)% Margin)
Non-GAAP & Cash
- As highlighted above – company has several large non-cash expenses owing to capital intensive & asset-heavy nature of the business (high D&A) as well as capital structure and valuation history (high Loss on Fair Value Adjustments) contributing to significant net loss on GAAP basis
- Adj EBITDA
- FY’24: $1.2B (+64% Margin)
- FY’23: $104M (+45% Margin)
- CFO
- FY’24: $2.7B
- FY’23: $1.8B
- Operating FCF (CFO – Capex)
- FY’24: -$6B
- FY’23: -$1.1B
- Change in Cash (including Debt & Equity raised to finance capex)
- FY’24: $1.6B
- FY’23: $473M
Balance Sheet
- At first glance, the summary BS upfront paints a difficult picture, with two key observations standing out:
- Just ~$1.4B Cash vs. ~$8B Total Debt (~$2.5B current Debt)
- Negative Working Capital (i.e. Current Assets < Current Liabilities → Current Ratio <1x)
- That doesn’t paint the full picture of the business, however, given the unique way they’ve constructed their ‘just-in-time’ debt financing model
- How does their debt financing work?
- CoreWeave’s Debt Financing, specifically their Delayed Draw Term Loans (DDTLs), are collateralized against contracted customer obligations. They are only taking on debt to match the CapEx needed to build towards exact obligations specified in signed contracts
- CoreWeave only purchases equipment after the customer is contracted (i.e. they have purchase orders to fulfill customer contracts / obligations ready to go, but don’t actually make the purchase until demand is firmed up)
- Customer Contracts / Outlines Obligations → Buy & Build Towards Obligations
- Timing Matters – Interest expense concerns are mitigated by the fact that it’s a DDTL, meaning you only pay interest on the drawn amount at any given time
- You can draw the DDTL at will to build towards contracted obligations, but as you do that & meet each milestone you also recognize requisite Revenue
- RPO (Remaining Performance Obligations), & RPO Growth specifically, is the key metric to measure CoreWeave’s financial health
- $15.1B of RPOs ending FY’24 (+53% Y/Y)
Capital Structure
- Co-Founders sold ~$500M in Class A shares in a secondary in late 2024, however given Class A / Class B structure (Class B – 10x voting power) still retain majority control of the business
- Magnetar is the largest single shareholder in the business, with ~35% ownership
- Fidelity (which was a large buyer in the secondary) is the second largest institutional shareholder, with ~8% ownership
