CoreWeave

Executive Summary

CoreWeave, Inc. (Nasdaq: CRWV) is a GPU cloud infrastructure company that was founded in 2017 as Atlantic Crypto, a cryptocurrency mining operation in New Jersey. Eight years later, it operates 43 data centers with 850 megawatts of active power, runs 250,000 NVIDIA GPUs, and generated $5.13 billion in revenue in fiscal year 2025 — a 168% increase from the prior year.¹

The company went public in March 2025 at $40 per share, raising $1.5 billion in what was the largest U.S. tech IPO since 2021. The stock has since traded as high as $187 and as low as $33.52, and sits around $81 as of March 2026 — roughly a $43 billion market cap.²

The Numbers

CoreWeave's growth trajectory is steep. FY2025 revenue of $5.13 billion made it, by its own account, the fastest cloud platform to reach $5 billion in annual revenue. Adjusted EBITDA hit $3.09 billion — a 154% increase from $1.22 billion in 2024. But the company posted a GAAP net loss of $1.17 billion for the year, widening from an $863 million loss in 2024, driven by depreciation, interest expense, and stock-based compensation.¹

The revenue backlog tells the more important story: $66.8 billion in remaining performance obligations as of December 31, 2025 — more than four times the backlog at the start of the year. This is anchored by multi-billion-dollar contracts with Microsoft, OpenAI, and Meta, which together account for the vast majority of committed revenue.³

Revenue Trajectory

Q1 2026 shows midpoint of $1.9–2.0B guidance. Source: CoreWeave earnings releases, SEC filings

The Founding Story

Commodities Traders Turn Crypto Miners

CoreWeave's founders came from the commodities trading world, not from technology. Michael Intrator, the CEO, co-founded and ran Hudson Ridge Asset Management, a natural gas hedge fund, from 2013 to 2018. Before that, he spent 16 years at Natsource Asset Management managing global environmental markets and energy products. Brian Venturo, now Chief Strategy Officer (previously CTO), was Intrator's colleague at both Hudson Ridge and Natsource, where he traded energy and emissions portfolios.⁴

In September 2017, Intrator, Venturo, and co-founders Brannin McBee and Peter Salanki launched Atlantic Crypto in New Jersey. The thesis was straightforward: GPU-based cryptocurrency mining (primarily Ethereum) was a compute-intensive activity that rewarded the same skills commodities traders had — sourcing cheap inputs (electricity), managing hardware at scale, and optimizing output. They built GPU mining rigs and ran them in data centers, applying the operational discipline of energy trading to crypto mining economics.

The 2018 Crash and the Pivot

The 2018 crypto crash wiped out mining profitability across the industry. Ethereum fell from over $1,400 in January 2018 to under $100 by the end of the year. Atlantic Crypto's GPU mining rigs — thousands of NVIDIA cards — were suddenly idle hardware.

Rather than shut down, the founders recognized that the same GPUs used for mining could serve a broader market. NVIDIA GPUs were increasingly in demand for machine learning training, visual effects rendering, and scientific simulation — workloads that needed the same parallel processing architecture as crypto mining. In 2019, the company rebranded as CoreWeave and pivoted to offering GPU compute-as-a-service. The mining rigs became cloud infrastructure.

The NVIDIA Relationship

The pivot coincided with NVIDIA's growing interest in expanding its cloud ecosystem beyond the three hyperscalers (AWS, Azure, GCP). CoreWeave became an early NVIDIA Cloud Partner and eventually an Elite Cloud Partner — a designation that gave it preferential access to new GPU generations and co-marketing support. This relationship proved foundational: when the AI boom hit in late 2022 with ChatGPT, CoreWeave was one of the few non-hyperscaler providers with significant NVIDIA GPU inventory already deployed and the software platform to rent it out.⁵

NVIDIA's role went beyond partnership. The chipmaker has invested directly in CoreWeave multiple times, culminating in a $2 billion investment in January 2026 at $87.20 per share — making NVIDIA both CoreWeave's primary hardware supplier, strategic partner, and a significant equity holder.⁶

Product & Platform

Kubernetes-Native GPU Cloud

CoreWeave's core product is a GPU cloud platform built natively on Kubernetes — the open-source container orchestration system that has become the standard for managing large-scale compute workloads. Unlike hyperscalers that adapted existing general-purpose cloud platforms to support GPU workloads, CoreWeave built its stack from the ground up around GPU compute, with Kubernetes as the orchestration layer.⁵

The platform supports the full spectrum of AI workloads:

• Training: Multi-thousand-GPU clusters for training large language models and other foundation models, connected via InfiniBand networking for the all-to-all communication patterns that distributed training requires.
• Inference: Production serving of trained models at scale, including the dedicated NVIDIA GB200 NVL72-powered clusters deployed for customers like Perplexity.
• Fine-tuning and post-training: Reinforcement learning, RLHF, and other techniques used to refine pre-trained models for specific applications.

Mission Control and SUNK

CoreWeave has developed two proprietary software systems that differentiate it from colocation-style competitors:

Mission Control is CoreWeave's infrastructure management platform — the software layer that handles provisioning, monitoring, scheduling, and lifecycle management of GPU clusters across its 43 data centers. For enterprise customers, it provides visibility into utilization, cost, and performance without requiring them to manage the underlying hardware.

SUNK (CoreWeave's orchestration software) integrates with NVIDIA's reference architectures, allowing customers and partners to deploy and manage GPU infrastructure using CoreWeave's tooling. In January 2026, NVIDIA announced that SUNK and Mission Control would be integrated into NVIDIA's reference architectures for cloud partners and enterprises — effectively making CoreWeave's software part of NVIDIA's recommended stack.⁶

Why Not Just Use a Hyperscaler?

The value proposition mirrors what other neoclouds offer, with CoreWeave's specific advantages being:

1. Scale of dedicated GPU capacity. 250,000 GPUs across 43 data centers gives CoreWeave one of the largest non-hyperscaler GPU fleets. Customers needing thousands of interconnected GPUs for frontier model training have limited alternatives outside the Big Three clouds.
2. Price. CoreWeave's GPU-hour pricing has historically been lower than hyperscaler equivalents. As of late 2025, H100 instances on CoreWeave were priced starting at approximately $2.49/hour — competitive with other neoclouds and significantly below Azure and GCP list pricing for equivalent configurations.⁷
3. Software maturity. Years of operating Kubernetes-native GPU infrastructure at scale has produced tooling (Mission Control, SUNK) that enterprise customers value for managing complex multi-node training jobs — the same workloads that are notoriously difficult to manage on general-purpose cloud platforms.

Infrastructure

Data Center Footprint

CoreWeave ended 2025 with 43 data centers and 850 megawatts of active power. The facilities are spread across the United States, with concentrations in New Jersey, North Dakota, Pennsylvania, Wyoming, West Texas, and Nevada, plus a facility in Ontario, Canada. The company targets exiting 2026 with over 1.7 GW of active power and has a longer-term goal of 5+ GW by 2030.⁸

Source: CoreWeave Q4 2025 earnings, company presentations¹

Known Facility Locations

Not exhaustive — CoreWeave operates 43 facilities total. Additional sites in Norway and Sweden announced but not yet detailed. Source: Baxtel, CoreWeave docs, press releases⁸

Active vs. Contracted Power: Neocloud Comparison

Active = live capacity (end of 2025). Secured/Pipeline = contracted + in-development; definitions vary by company. Nebius 3 GW is 2026 target. Source: Q4 2025 filings, company presentations

The Expansion Plan

CoreWeave plans to roughly double its active power capacity in 2026, from 850 MW to over 1.7 GW. The company has 3.1 GW of contracted power to support this expansion. The 2026 capital expenditure budget of $30–35 billion — up from $10.3 billion in 2025 — gives a sense of the scale of the buildout. Most of this goes to GPU procurement, data center construction, and power infrastructure.¹

The longer-term target, supported by the January 2026 NVIDIA partnership, is 5+ GW of AI computing capacity by 2030. That would require roughly 6x the current active power — a massive construction and procurement program that will test the company's execution capacity and financing ability simultaneously.⁶

Key Contracts

The Big Three

CoreWeave's revenue is concentrated in three customers that together account for the vast majority of its backlog:

Source: CoreWeave filings, Reuters^{9, 10, 11}

Revenue Backlog Growth

Revenue backlog (RPO + other committed contracts) as of quarter end. Q4 2024 implied from "more than 4x" growth cited in Q4 2025 earnings. Source: CoreWeave earnings releases, 10-K

Microsoft

Microsoft has been CoreWeave's anchor customer since before the IPO, accounting for 62% of 2024 revenue and 67% of 2025 revenue.¹² The relationship reflects Microsoft's need for GPU capacity beyond what it can build internally — even at a capital expenditure run rate exceeding $100 billion annually, Microsoft cannot build data centers fast enough to meet demand from Azure, OpenAI, and its own Copilot products. CoreWeave provides overflow capacity.

CEO Michael Intrator has described the concentration as a natural result of "mind-bendingly large deals" rather than a failure to diversify.¹² Management expects Microsoft's share to decline below 50% of revenue as the OpenAI and Meta contracts ramp through 2026.

OpenAI

The OpenAI relationship is CoreWeave's most strategically significant. In March 2025, just ahead of the IPO, CoreWeave announced an $11.9 billion, five-year contract with OpenAI for dedicated compute capacity. OpenAI simultaneously invested $350 million in CoreWeave stock via a private placement.⁹ The deal was expanded by approximately $4 billion in May 2025¹⁰ and by a further $6.5 billion in September 2025¹⁸, bringing the total commitment to approximately $22.4 billion through May 2031.

The structure is notable: CoreWeave financed a dedicated $2.6 billion debt facility specifically to build infrastructure for the OpenAI contract, with the contract itself serving as collateral. OpenAI gets dedicated GPU capacity for training and serving its models; CoreWeave gets a revenue-backed debt structure that funds the buildout.¹³

Meta

In September 2025, CoreWeave signed a deal with Meta worth approximately $14.2 billion through December 2031, with options to extend through 2032. Under the deal, CoreWeave will provide Meta access to NVIDIA's latest GB300 systems for AI compute capacity powering Meta's consumer products and AI applications. The Meta deal marked a significant diversification beyond the Microsoft/OpenAI axis.¹¹

Perplexity and the Inference Opportunity

In March 2026, CoreWeave announced a multi-year strategic partnership with Perplexity to power its AI inference workloads using dedicated NVIDIA GB200 NVL72-powered clusters. This deal is smaller than the Big Three contracts but signals CoreWeave's push into inference — a market segment that is growing faster than training as more companies deploy AI models in production.¹⁴

Market Opportunity

The GPU Supply-Demand Gap

The market CoreWeave operates in is defined by a structural imbalance: demand for GPU compute — from AI model training, inference, and fine-tuning — has grown faster than any single provider can build capacity. Hyperscalers (Microsoft, Google, Amazon, Meta) collectively spent over $400 billion on data center capex in 2025 alone, with 2026 projections approaching $700 billion, and the constraint is not chips or capital — it's physical infrastructure. New data center sites take 3–5 years to permit and build from scratch.

This gap created the opening for GPU cloud providers like CoreWeave, Nebius, and IREN. They offer hyperscalers and AI labs a way to access GPU capacity faster than building it themselves, through long-term contracts that de-risk both sides: the customer gets guaranteed capacity, the provider gets revenue visibility to support debt financing.

Training vs. Inference

The AI compute market is shifting. The initial wave (2023–2025) was dominated by training — building frontier models required massive GPU clusters running for months. The next wave is inference — serving those models to end users at scale. Inference workloads are more distributed, more latency-sensitive, and growing faster in aggregate compute demand than training.

CoreWeave's backlog is heavily weighted toward training (the Microsoft, OpenAI, and Meta contracts are primarily training infrastructure). The Perplexity deal signals an attempt to capture inference demand, which requires different infrastructure characteristics — lower latency, more geographic distribution, and different GPU configurations.

The TAM Question

IDC projects global AI infrastructure spending will reach $758 billion by 2029, with accelerated servers (GPU-based) accounting for over 94% of total market spending.¹⁵ The relevant question for CoreWeave is what share of this spending flows to independent GPU cloud providers versus hyperscaler-owned infrastructure. Currently, neoclouds capture a single-digit percentage of total AI compute spend — but that percentage on a rapidly growing base still produces enormous absolute numbers.

Competitive Landscape

The Neocloud Peers

CoreWeave operates in a field of GPU cloud and AI infrastructure providers, each with a distinct origin story and strategic positioning:

CoreWeave vs. Nebius

Nebius is probably the closest direct competitor in terms of product vision — both offer full-stack GPU cloud platforms with managed software layers. The key differences:

• Scale: CoreWeave generated roughly 10x Nebius's 2025 revenue ($5.13B vs. $529.8M) and has a correspondingly larger GPU fleet and data center footprint.
• Balance sheet: Nebius's capital structure is significantly cleaner. As of year-end 2025, Nebius held ~$3.7 billion in cash against ~$4.9 billion in total debt — a D/E of ~1.1x. The company has since raised an additional $4.3 billion in convertible notes and $2 billion from NVIDIA in early 2026. CoreWeave's D/E is ~4.5x, with far heavier reliance on GPU-collateralized debt and equipment financing.
• Origin: Nebius inherited deep AI engineering talent from Yandex, giving it software capabilities (Token Factory, Aether) that are more tightly integrated than CoreWeave's platform. CoreWeave's engineering DNA is infrastructure operations — running GPU clusters reliably at scale.
• Geography: Nebius has a more global footprint (Finland, Iceland, France, Israel, U.K., plus U.S.), while CoreWeave is primarily U.S.-based.

CoreWeave vs. IREN

IREN comes from the other direction — an infrastructure-first company (former Bitcoin miner) that controls 4.5 GW of secured, grid-connected power capacity (partially renewable) and is building the GPU cloud layer on top. The comparison:

• Software vs. power: CoreWeave's advantage is its software platform and years of Kubernetes-native GPU orchestration experience. IREN's advantage is physical infrastructure — cheap, renewable power at a scale CoreWeave has to lease or build.
• Customer base: CoreWeave has multiple hyperscaler customers (Microsoft, OpenAI, Meta). IREN's AI cloud revenue is anchored by a single $9.7 billion Microsoft contract.
• Revenue stage: CoreWeave is at $5 billion in revenue and scaling. IREN's AI cloud business generated $17.3 million in Q2 FY26 — growing rapidly, but from a much smaller base.

The Hyperscaler Question

The long-term competitive risk for all neoclouds — CoreWeave included — is that the hyperscalers are building their own GPU capacity at massive scale and, increasingly, designing their own AI chips (Microsoft Maia, Google TPUs, Amazon Trainium). If hyperscalers bring enough captive capacity online, or if AI compute demand moderates, the need for third-party GPU cloud providers contracts. CoreWeave's contracts provide multi-year protection, but the structural question remains: is the neocloud model a permanent layer of the AI stack, or a temporary response to a supply shortage that will eventually resolve?

Business Model & Financials

How CoreWeave Makes Money

CoreWeave's revenue model is conceptually simple: it buys NVIDIA GPUs, deploys them in data centers, and rents the compute capacity to customers under long-term contracts (typically 3–5 years). Revenue is recognized over the contract term as compute is delivered. The gross margin depends on the spread between what customers pay per GPU-hour and CoreWeave's cost of GPU depreciation, power, cooling, networking, and facility costs.

Financial Summary

*2026 CapEx guidance per CFO includes finance lease payments ($14.9B on same basis for FY2025). Source: CoreWeave 10-K, earnings releases^{1, 3}

The GAAP vs. Adjusted EBITDA Gap

The gap between CoreWeave's $3.09 billion adjusted EBITDA and its $1.17 billion GAAP net loss is driven by three factors:

1. Depreciation. CoreWeave depreciates billions in GPU hardware over 3–5 year useful lives. This is a real economic cost — GPUs do lose value as newer generations arrive — but it's a non-cash charge that adjusted EBITDA excludes.
2. Interest expense. On $21.6 billion in debt, interest expense is substantial. Q3 2025 alone saw $311 million in interest expense — tripling year-over-year. This is a cash cost that adjusted EBITDA also excludes.
3. Stock-based compensation. SBC is significant for a recently-IPO'd company competing for engineering talent in the AI infrastructure space.

The adjusted EBITDA metric shows the business generates cash at the operating level. Whether that cash is sufficient to cover depreciation, interest, and reinvestment is the central financial question.

Quarterly Trajectory

The Q1 2026 guidance of $1.9–2.0 billion came in below the $2.29 billion consensus estimate — a dynamic that echoes across the neocloud sector, where the gap between contracted backlog and near-term revenue recognition creates frequent expectation mismatches. Contracts take time to ramp as infrastructure is built and deployed.¹

Capital Structure & Valuation

Balance Sheet

Source: CoreWeave balance sheet, SEC filings¹⁶

The Debt Architecture

CoreWeave pioneered GPU-collateralized debt — using its NVIDIA GPU inventory and customer contracts as collateral to secure large debt facilities. The capital stack includes:

The total capital raised over the past 18 months exceeds $25 billion across equity and debt.¹³

The 2026 Debt Wall

A $4.2 billion principal debt repayment is due later in 2026 — roughly equal to CoreWeave's current cash position plus one quarter of revenue. The company's ability to manage this maturity, either through refinancing, revenue-funded repayment, or additional capital raises, is a near-term focal point for investors and analysts.¹⁷

Valuation

The 52-week range of $33.52 to $187.00 — a 5.6x spread — captures the market's uncertainty about CoreWeave's trajectory. At $81, the stock is roughly 2x the IPO price and 57% below the all-time high.

Neocloud Leverage Comparison

Debt-to-equity ratio as of Q4 2025 (Dec 31, 2025). Source: Financial Modeling Prep API, company filings

Key Opportunities

Key Risks

Key Takeaways

The factors that will determine CoreWeave's trajectory over the next 12–18 months:

1. Debt management. The 2026 maturity wall is the near-term test. Whether CoreWeave refinances, repays from cash flow, or raises additional capital — and at what cost — sets the tone for the balance sheet going forward.
2. Revenue ramp vs. guidance. The Q1 2026 guidance of $1.9–2.0B was below consensus. If revenue accelerates through 2026 toward the $12–13B guide, the backlog-to-revenue conversion thesis holds. If it disappoints, the debt burden becomes harder to justify.
3. Customer diversification. Microsoft at 67% of revenue needs to decline materially. The pace at which OpenAI and Meta contracts ramp — and whether new customers emerge beyond the Big Three — determines whether CoreWeave is a three-customer business or a platform.
4. CapEx execution. Spending $30–35 billion in a single year while operating 43+ data centers and deploying hundreds of thousands of GPUs is an operational challenge at the limits of what any company has attempted. Delays or cost overruns hit both revenue timing and capital structure.
5. Inference traction. The shift from training to inference is the next growth vector. CoreWeave's Perplexity deal is a start, but capturing material inference market share requires a different go-to-market and potentially different infrastructure than the large-cluster training contracts that dominate the current backlog.

CoreWeave's story is one of scale achieved at extraordinary speed — from a crypto mining startup to a $5 billion revenue business in under nine years, funded by the most aggressive debt structure in the AI infrastructure sector. The contracts are signed, the GPUs are deployed, and the customers are paying. What remains to be seen is whether the capital structure built to fund this growth can sustain the weight of its own ambition, or whether the gap between adjusted EBITDA and GAAP profitability becomes the defining constraint of the next chapter.