CoreWeave, Inc. (CRWV) - Deep Dive Research Report

Report Date: May 5, 2026

1. What the Company Does

CoreWeave rents GPU computing power to companies that need it to train and run artificial intelligence models. That is the business in one sentence. The company owns tens of thousands of NVIDIA GPUs, houses them in data centers it leases or builds, connects them with high-speed networking, wraps them in Kubernetes-based software, and sells access to that compute on multi-year contracts.

Founding Story

CoreWeave was born from cryptocurrency mining. In 2016, Michael Intrator and Brian Venturo - both working at their hedge fund, Hudson Ridge Asset Management - bought their first GPU to experiment with Ethereum mining. They turned a hobby into a business, incorporating as Atlantic Crypto in 2017 alongside co-founders Brannin McBee and Peter Salanki. The company ran GPU rigs out of a New Jersey data center, mining Ethereum for transaction verification fees.

The 2018 crypto crash was the pivot point. Rather than shut down, the founders recognized that their fleet of GPUs had value beyond cryptocurrency. In 2019, they renamed the company CoreWeave and began offering cloud computing services. The insight was simple but well-timed: GPUs were becoming the essential hardware for machine learning, and the hyperscale cloud providers (AWS, Azure, Google Cloud) were not purpose-built for GPU-heavy workloads. CoreWeave would be.

The timing proved prescient. When the generative AI explosion arrived in late 2022, CoreWeave had something almost nobody else had: large quantities of NVIDIA GPUs already deployed, operational expertise in running GPU clusters, and a direct procurement relationship with NVIDIA. The company went from $16 million in revenue in 2022 to $1.9 billion in 2024 to $5.1 billion in 2025.

Core Value Proposition

CoreWeave solves a specific problem: AI companies need thousands of GPUs working together as a unified cluster, and traditional cloud providers are not optimized for this. Training a frontier AI model requires linking thousands of GPUs via ultra-fast interconnects, managing checkpointing and fault tolerance, providing massive parallel storage, and keeping utilization rates near 100% for months at a time. Running inference at production scale requires different but equally demanding infrastructure - low latency, auto-scaling, and efficient batching.

CoreWeave's pitch is that it built its entire stack - from the physical hardware layer to the orchestration software - specifically for these workloads. There is no legacy virtualization layer, no shared-tenancy overhead, no general-purpose compromise. GPUs run on bare metal. Networking is purpose-built. Storage is co-designed for AI data patterns.

How It Works in Practice

A customer like Anthropic or Meta signs a multi-year contract specifying how many GPUs they need, for how long, and with what networking and storage requirements. CoreWeave then procures the GPUs from NVIDIA, leases or builds data center space with sufficient power, racks and connects the hardware, deploys its Kubernetes-based management software, and hands the customer a dedicated cluster accessible via API. The customer pays a fixed monthly fee for their reserved capacity, typically for 3-7 years.

The company also offers on-demand compute for smaller customers who do not need dedicated clusters, but the bulk of revenue comes from large reserved-instance contracts with hyperscalers and AI labs.

2. Business Segments

CoreWeave operates as a single reportable segment: cloud services revenue from GPU compute infrastructure. However, its business functionally breaks into distinct service layers that serve different customer needs:

Reserved Instance Compute (Core Business)

This is the revenue engine. Large customers (Meta, Microsoft, OpenAI, Anthropic) sign multi-year contracts for dedicated GPU clusters. These are not shared resources - the customer gets exclusive access to specific hardware, configured to their specifications. Contracts typically run 3-7 years with take-or-pay structures, meaning the customer pays whether they use the capacity or not.

Reserved instances account for the vast majority of CoreWeave's $66.8 billion contracted backlog. The average contract duration and take-or-pay structure give revenue visibility that is unusual for a cloud company - more akin to a regulated utility or a long-term lease business. The trade-off is massive upfront capital expenditure: CoreWeave must build the infrastructure before the revenue begins.

On-Demand and Spot Compute

Smaller AI companies, research labs, and developers access CoreWeave's platform without long-term commitments. They pay hourly rates for GPU time, scaling up and down as needed. This is a smaller portion of revenue but serves as a customer acquisition funnel - companies start on-demand, prove out their workloads, and graduate to reserved instances as their needs grow.

Platform Software and Services

CoreWeave has been building a software layer on top of raw compute through acquisitions:

• Weights & Biases (acquired May 2025): MLOps platform for experiment tracking, model management, and collaboration. Used by tens of thousands of AI teams worldwide before the acquisition. Gives CoreWeave stickiness beyond raw compute.
• OpenPipe (acquired September 2025): Reinforcement learning platform for training AI agents. Adds fine-tuning and model customization capabilities.
• CoreWeave Kubernetes Service (CKS): Proprietary orchestration layer that runs Kubernetes directly on bare metal without a hypervisor, delivering 20% higher GPU cluster performance than alternatives.
• Mission Control: Unified security, observability, and orchestration across the full stack.
• AI Object Storage: S3-compatible storage optimized for AI workloads, reaching $100 million ARR in Q3 2025 with 75% cost reduction versus alternatives and zero egress fees.

CoreWeave Federal

Launched October 2025, this division targets U.S. government agencies and the defense industrial base. Pursuing FedRAMP authorization to compete for programs like the Department of Defense's $9 billion Joint Warfighting Cloud Capability (JWCC). NASA's Jet Propulsion Lab is an early customer. CoreWeave joined the Department of Energy's Genesis Mission in December 2025.

3. Products and Business Detail

Product Catalogue

Compute Products:

• NVIDIA H100 GPU Clusters: The workhorse for current-generation AI training and inference
• NVIDIA GB200 NVL72 Systems: Next-generation Blackwell architecture clusters (CoreWeave achieved NVIDIA Exemplar Cloud status for these)
• NVIDIA Vera Rubin Platform: Announced for second-half 2026 deployment - CoreWeave will be among the first cloud providers to offer this
• CPU Instances: AMD EPYC-based compute for non-GPU workloads
• Bare Metal Servers: Direct hardware access without virtualization overhead

Networking:

• InfiniBand and high-speed Ethernet interconnects
• Arista Networks 1.6-terabit Ethernet switches for 100,000+ GPU clusters
• NVIDIA BlueField DPUs for network offloading

Storage:

• CoreWeave Object Storage: S3-compatible, AI-optimized, exascale
• VAST Data Clusters: Single-tenant, multi-protocol access at petabyte scale
• GPU-local caching: High-throughput data access co-located with compute

Software:

• CoreWeave Kubernetes Service (CKS): Bare-metal Kubernetes with 5x faster model downloads and 10x faster inference spin-up
• SUNK: Combines Kubernetes agility with Slurm orchestration for HPC workloads
• Weights & Biases: MLOps experiment tracking and model registry
• OpenPipe: Reinforcement learning and model fine-tuning
• Mission Control: Security, observability, and orchestration suite
• Cross-Cloud Capabilities: Portability to AWS, Azure, or GCP without rebuilding workflows

Infrastructure Footprint

As of December 31, 2025:

• 43 active data centers across North America and Europe
• 850+ megawatts of active power (target: 1.7 GW by end of 2026)
• 3.1 gigawatts contracted (nearly all coming online by 2027)
• 250,000+ NVIDIA GPUs deployed
• 5 GW target by 2030 (per NVIDIA partnership announcement)

Key Locations:

• United States: Multiple sites across New Jersey, Pennsylvania, Texas, and other states
• Lancaster, Pennsylvania: $6 billion purpose-built campus, 250+ MW initial capacity under 15-year agreement
• Kenilworth, New Jersey: Joint venture with Blue Owl Capital
• Denton, Texas: 260 MW campus built by Core Scientific (delayed 60 days by summer storms in 2025)
• United Kingdom: Crawley (live October 2024), London Docklands (live December 2024), additional sites
• Continental Europe: Norway, Sweden, and Spain ($2.2 billion investment committed)

Capacity Expansion

CoreWeave added approximately 260 megawatts of active power in Q4 2025 alone - nearly matching what it took years to build previously. The company went from 32 data centers at the start of 2025 to 43 by year-end, adding 8 new domestic sites in Q3 alone. With $30-35 billion in planned 2026 CapEx (more than double the $14.9 billion spent in 2025), the build-out is accelerating.

Milestones

• 2017: Founded as Atlantic Crypto
• 2019: Renamed CoreWeave, pivoted to GPU cloud
• 2022: $16M revenue; early NVIDIA partnership established
• 2023: Raised $2.3 billion in debt; revenue approached $500M
• March 2025: IPO on Nasdaq at $40/share, raising $1.5 billion (largest U.S. tech IPO since 2021)
• May 2025: Acquired Weights & Biases
• July 2025: Proposed $9B acquisition of Core Scientific (later terminated October 2025)
• September 2025: Acquired OpenPipe
• January 2026: NVIDIA invested $2 billion at $87.20/share
• March 2026: Closed $8.5 billion investment-grade financing facility (first ever for GPU-backed infrastructure)
• April 2026: $21 billion Meta expansion + $6.8 billion Anthropic deal announced within 48 hours

4. Customers

Who Buys

CoreWeave's customer base spans three tiers:

Tier 1 - Hyperscalers and AI Model Labs:

• Microsoft: 67% of 2025 revenue; anchor customer since pre-IPO; primarily training infrastructure for Azure AI services
• Meta: $35 billion total contracted (original $14.2B in September 2025 + $21B expansion in April 2026); capacity for Llama model training and inference through December 2032
• OpenAI: $11.9 billion five-year contract signed March 2025 for dedicated compute
• Anthropic: $6.8 billion multi-year deal signed April 2026 for NVIDIA Vera Rubin GPU capacity to power Claude inference

Tier 2 - AI-Native Companies:

• Cognition, Cursor, Midjourney, Runway, Moon Valley (AI video generation)
• CrowdStrike (security), Hippocratic AI (healthcare)
• These companies build AI products as their core business and need dedicated compute

Tier 3 - Enterprise and Financial Services:

• Goldman Sachs, Morgan Stanley, Jane Street (financial services)
• Rakuten, Nissan (global enterprise)
• MercadoLibre (e-commerce/fintech)
• NASA Jet Propulsion Lab (federal/research)

Buying Decision

For Tier 1 customers, the buying decision is made at the executive or board level. The evaluation involves: total cost of ownership, GPU availability timeline, networking performance, cluster reliability, and contract flexibility. Sales cycles for these deals stretch 3-6 months and involve extensive technical due diligence.

For Tier 2 customers, the decision-maker is typically a VP of Infrastructure or CTO. They evaluate CoreWeave against hyperscaler equivalents (AWS, Azure, GCP) on price-performance, GPU availability, and platform features. Sales cycles run 1-3 months.

Why They Choose CoreWeave

1. GPU availability: CoreWeave can deliver new NVIDIA architectures faster than hyperscalers because its entire supply chain is purpose-built for this. It was the only provider submitting MLPerf inference results for GB300 GPUs.
2. Performance: SemiAnalysis awarded CoreWeave its Platinum Cluster Max ranking for two consecutive years - no other cloud provider has achieved this even once.
3. Price: 50-70% cost savings on training workloads versus hyperscalers, per industry benchmarks.
4. Bare-metal architecture: No hypervisor overhead means customers get full hardware performance.
5. NVIDIA relationship: NVIDIA invested $2 billion and committed to co-developing the platform, making CoreWeave the reference implementation for new architectures.

Switching Costs

Switching costs are moderate to high:

• Multi-year take-or-pay contracts lock in revenue but also lock in customers
• Customers who integrate Weights & Biases for MLOps have workflow lock-in
• Large-scale training runs cannot easily be interrupted or migrated mid-job
• However, the underlying workloads (PyTorch, Kubernetes) are portable - customers could theoretically move at contract renewal

Customer Concentration

This has been CoreWeave's most scrutinized risk metric. The trajectory:

• Early 2025: ~85% of backlog from one customer (Microsoft)
• Q2 2025: ~50% from top customer
• Q3 2025: ~35% from top customer
• Q4 2025: No single customer >35% of revenue backlog

The Meta expansion ($35B total) and Anthropic deal ($6.8B) in April 2026 further diversified the base. Nine of ten largest customers executed multiple agreements, and the company now counts nine of the ten leading AI model providers as customers.

Contract Structure

Contracts are predominantly multi-year reserved instances with take-or-pay provisions. Revenue is recognized ratably over the contract period. The $66.8 billion backlog (as of Q4 2025) represents extraordinary forward visibility. The weighted average remaining contract life extends through the late 2020s, with the Meta deal running through 2032.

5. Competitive Landscape

Industry Structure

The GPU cloud market has three tiers:

Tier 1 - Hyperscalers (60%+ of total cloud market):

• AWS, Microsoft Azure, Google Cloud Platform
• Massive scale, integrated ecosystems, established enterprise relationships
• But: GPU capacity constrained, multi-tenant architecture not optimized for AI, slower to deploy new hardware

Tier 2 - Neoclouds (fastest-growing segment):

• CoreWeave (dominant player by revenue and scale)
• Crusoe (focused on sustainable/stranded energy)
• Lambda Labs (developer-focused, smaller scale)
• Nebius (Yandex spin-off, European focus)
• Together AI, Fireworks AI (inference-focused platforms)

Tier 3 - Specialty/Spot Providers:

• RunPod, Vast.ai (marketplace models targeting individual developers)
• Smaller scale, lower price, less reliable

Why CoreWeave Wins

1. Scale: CoreWeave generated more revenue in H1 2025 ($2.1B) than Lambda ($250M) - an 8:1 gap. No other neocloud approaches CoreWeave's infrastructure footprint.
2. NVIDIA preferential access: The $2 billion NVIDIA investment and Exemplar Cloud status mean CoreWeave gets new chips (GB200, Vera Rubin) before competitors.
3. Financial capacity: $28 billion in financing commitments over the past 12 months enables CapEx at a scale no other neocloud can match.
4. Enterprise credibility: Investment-grade customers (Meta, Microsoft) willing to sign multi-billion dollar contracts validates operational quality.
5. Full-stack software: Weights & Biases, CKS, Mission Control create stickiness that raw compute cannot.

Where CoreWeave Loses

1. Against hyperscalers for integrated workloads: Enterprises that want AI plus storage plus databases plus CDN in one ecosystem will choose AWS or Azure.
2. Against Lambda/RunPod for small-scale access: Developers wanting quick, cheap, no-commitment GPU time often find Lambda or RunPod simpler.
3. Cost of capital disadvantage vs. hyperscalers: Amazon and Microsoft can self-fund data centers from operating cash flow. CoreWeave must borrow at 5-6%+ interest, compressing margins.
4. Lock-in concerns: Some customers fear depending on a single-product company for critical infrastructure.

Barriers to Entry

• GPU supply: NVIDIA allocates finite production across buyers. New entrants cannot simply order 250,000 GPUs.
• Capital: Building a competitive fleet requires $10B+ in upfront investment with negative cash flow for years.
• Operational expertise: Managing thousands of GPUs at near-100% uptime requires deep systems engineering knowledge.
• Customer trust: AI companies will not risk their training runs on unproven infrastructure. Track record matters.
• Financing access: Lenders now accept GPU-backed financing because CoreWeave proved the model. A new entrant would face higher rates and less availability.

Structural Shifts

The competitive landscape is consolidating. CoreWeave's failed acquisition of Core Scientific (shareholders voted it down in October 2025) showed appetite for vertical integration. Hyperscalers are building custom AI chips (Google TPUs, Amazon Trainium) to reduce NVIDIA dependence. Meanwhile, every major AI lab is signing long-term neocloud contracts as insurance against hyperscaler capacity constraints.

6. Industry

Demand Drivers

AI infrastructure demand is driven by three forces:

1. Training compute: Each generation of frontier AI models requires 3-10x more compute than the last. GPT-5, Claude 4, Gemini Ultra - each pushes GPU demand upward.
2. Inference at scale: As AI models move from research to production (customer service bots, code assistants, image generators), inference compute demand grows with user adoption.
3. Enterprise AI adoption: Traditional companies (banks, pharma, manufacturing) are building internal AI capabilities, creating a new demand layer beyond AI labs.

Market Size

The neocloud market is growing explosively:

• 2025: $24-25 billion (Mordor Intelligence / Synergy Research Group)
• 2026 estimate: $35 billion (Mordor Intelligence)
• 2030 forecast: $65-180 billion depending on research firm (ABI Research: $65B in GPU-as-a-Service; Synergy: $180B total neocloud)
• 2031 forecast: $237-400 billion (Mordor: $237B; Synergy: approaching $400B)
• Growth rate: 46-69% CAGR through 2030-2031

The broader cloud computing market exceeds $600 billion. Neoclouds are carving out a specialized, high-growth niche within it.

Supply Chain Position

CoreWeave sits between NVIDIA (chip supplier) and AI companies (compute consumers). It is essentially a capital-intensive intermediary that:

• Buys GPUs from NVIDIA
• Procures power and data center space
• Adds software and operational value
• Sells compute access on long-term contracts

This position creates a paradox: CoreWeave depends heavily on NVIDIA for supply, but NVIDIA also depends on CoreWeave as a major distribution channel for its chips. The $2 billion NVIDIA investment cements this mutual dependence.

Regulatory Environment

• Power permitting: Data center construction requires utility connections, which involve multi-year permitting timelines
• Environmental: Growing scrutiny on data center energy consumption and water usage for cooling
• FedRAMP: Required for U.S. government contracts; CoreWeave is pursuing certification
• Export controls: U.S. restrictions on NVIDIA chip exports to China benefit CoreWeave by keeping supply focused on domestic demand
• Data sovereignty: European customers may prefer CoreWeave's European sites over U.S.-only alternatives

Cyclicality

This is the critical question. AI infrastructure spending is currently in a massive upcycle. The historical parallel is the 1990s telecom build-out, where companies like Global Crossing and Level 3 accumulated enormous debt building fiber networks, then collapsed when demand failed to meet projections.

Bulls argue this cycle is different because: (a) AI models demonstrably generate economic value, (b) demand is growing faster than supply even with massive build-out, and (c) take-or-pay contracts de-risk investment.

Bears argue that: (a) AI spending could decelerate if models plateau, (b) multi-year contracts can be renegotiated or customers can default, and (c) GPU depreciation and technology transitions create obsolescence risk.

7. Growth Triggers

Sources: Q1 FY25 call (May 14, 2025), Q2 FY25 call (Aug 12, 2025), Q3 FY25 call (Nov 10, 2025), Q4 FY25 call (Feb 26, 2026)

• $21 billion Meta expansion: Extended AI infrastructure agreement through December 2032, including initial deployments of NVIDIA Vera Rubin platform. Announced April 9, 2026, post-Q4 call. (Repeated across multiple investor communications)

• $6.8 billion Anthropic deal: Multi-year agreement for NVIDIA Vera Rubin GPU capacity to power Claude inference at production scale. Announced April 10, 2026. Compute to come online "later this year."

• NVIDIA Vera Rubin deployment: CoreWeave among the first cloud providers to deploy Rubin platform in H2 2026, enabling workloads like agentic AI, drug discovery, and climate simulation. (Q4 FY25 concall, Feb 26, 2026)

• 1.7 GW active power by year-end 2026: More than doubling from 850 MW at end of 2025. CapEx of $30-35 billion planned. (Q4 FY25 concall, Feb 26, 2026)

• Revenue run-rate $17-19 billion exiting 2026, growing to $30+ billion by end of 2027: Management's explicit multi-year trajectory. (Q4 FY25 concall, Feb 26, 2026)

• Inference workload shift: CEO emphasized "inference is the monetization of artificial intelligence" - the transition from training-dominated to inference-dominated compute is expanding CoreWeave's addressable market. (Q2 FY25 concall, Aug 12, 2025; repeated Q3, Q4)

• Financial services vertical expansion: Goldman Sachs, Morgan Stanley, Jane Street signed as customers. New vertical with high willingness to pay and long-term needs. (Q2 FY25 concall, Aug 12, 2025)

• CoreWeave Federal and DOE Genesis Mission: Pursuing FedRAMP certification to access $9 billion JWCC program and broader federal AI spend. NASA JPL already a customer. (Q3 FY25 concall, Nov 10, 2025)

• European expansion: $3.5 billion committed across UK, Norway, Sweden, and Spain. UK sites already operational. Continental European sites coming online through 2026. (Q2 FY25 concall, Aug 12, 2025; repeated Q3)

• 5 GW of AI factories by 2030: NVIDIA partnership to co-develop infrastructure. NVIDIA helping CoreWeave source land and power. (Announced January 2026)

• Customer base doubling: Customers with $1M+ annual commitments grew nearly 150% in FY25; customers with $100M+ annual revenue tripled YoY. Q4 added approximately twice as many new reserved instance customers as any prior quarter. (Q4 FY25 concall, Feb 26, 2026)

• AI Object Storage reaching scale: Hit $100M ARR in Q3 2025 with 75% cost reduction versus alternatives and no egress fees. Adds revenue stream beyond pure compute. (Q3 FY25 concall, Nov 10, 2025)

8. Key Risks

1. Debt Overhang and Refinancing Risk

CoreWeave carries approximately $29.8 billion in total debt. Interest expense was $388 million in Q4 2025 alone - roughly 24% of that quarter's revenue. The company's plan requires continued access to debt markets at favorable rates.

Mechanism: If AI spending decelerates, or if credit markets tighten, CoreWeave could face a situation where its debt service costs exceed its cash generation capacity. $3.6 billion in debt matures by June 30, 2026, and much of the older debt carries interest rates between 9-15%.

Calibration: Medium probability, high severity. The recent $8.5 billion investment-grade facility (A3/A-low rated, at 5.9% fixed) shows the market's willingness to lend. But the company remains in negative free cash flow territory due to massive CapEx, and any disruption to the financing cycle would be immediately damaging.

2. Customer Concentration and Counterparty Risk

Microsoft represented 67% of 2025 revenue. OpenAI is projected to lose $14 billion in 2026 and may face cumulative losses through 2029. If any anchor customer reduced commitments, the revenue hole would be enormous.

Mechanism: Take-or-pay contracts provide legal protection, but a customer in financial distress (OpenAI) or strategic shift (Microsoft building internal capacity) could renegotiate or default. CoreWeave would retain the hardware but lose the revenue stream backing its debt.

Calibration: Low-medium probability in the near term (contracts are binding), but a structural risk over 5+ years as the market matures.

3. AI Spending Plateau or Reversal

CoreWeave's entire business depends on AI companies continuing to spend aggressively on compute. If AI model scaling hits diminishing returns, or if a recession forces budget cuts, the demand underpinning $66.8 billion in backlog could weaken.

Mechanism: Unlike a telecom network where the infrastructure has indefinite life, GPUs depreciate rapidly (3-5 year useful life) and face obsolescence with each new architecture. If demand slows, CoreWeave would own depreciating assets financed by long-term debt.

Calibration: Low probability in the next 2 years given signed contracts. Medium probability over 5+ years. The 1990s telecom analogy is imperfect but not irrelevant.

4. NVIDIA Dependency

CoreWeave's competitive advantage rests substantially on preferential access to NVIDIA's latest GPUs. If NVIDIA decides to sell directly to end customers, vertically integrate into cloud services, or prioritize other distribution partners, CoreWeave's position erodes.

Mechanism: NVIDIA has invested $2 billion in CoreWeave, aligning incentives. But NVIDIA also partners with every hyperscaler and sells directly to enterprises. A shift in NVIDIA's distribution strategy - or the emergence of competitive AI chips (AMD, custom silicon) - would undermine CoreWeave's differentiation.

Calibration: Low probability in the near term given the equity investment and partnership. But NVIDIA's strategic interests and CoreWeave's interests are not permanently identical.

5. Construction Execution Risk

CoreWeave is attempting to deploy over $30 billion in CapEx in 2026 while building out to 1.7 GW of active power. The Denton, Texas delay (60 days due to weather) cost the company 62% of its peak stock price and triggered a securities class action lawsuit.

Mechanism: Data center construction involves utility interconnections, permitting, weather, supply chain logistics, and contractor coordination. At CoreWeave's scale, even small delays cascade into missed delivery dates, which delay revenue recognition while debt service continues.

Calibration: High probability of some delays (construction at this pace inevitably encounters issues). Medium severity per incident, but cumulative delays could shift investor sentiment materially. A securities class action (Hagens Berman) is already active over the Denton delays.

6. Technology Transition Risk

Every NVIDIA architecture generation (Hopper to Blackwell to Rubin) renders the prior generation less competitive. CoreWeave must continually replace its fleet to maintain performance leadership. Customers on older hardware may demand upgrades or exit at contract renewal.

Mechanism: A GPU purchased today depreciates over 3-5 years. If inference workloads shift to cheaper hardware (TPUs, AMD MI300X) or if NVIDIA's next-gen chip changes the form factor (requiring different cooling, power, or networking), existing infrastructure loses value.

Calibration: This is a permanent structural feature of the business, not a one-time risk. Management manages it through contract structuring (passing hardware refresh costs into pricing) but it requires perpetual heavy CapEx.

9. Walk the Talk

Concall dates used: Q1 FY25 (May 14, 2025), Q2 FY25 (Aug 12, 2025), Q3 FY25 (Nov 10, 2025), Q4 FY25 (Feb 26, 2026)

Q1 FY25 (May 14, 2025) - First Public Earnings Call

CoreWeave entered the public market with strong momentum. Q1 revenue hit $982 million (420% YoY growth), with $25.9 billion in revenue backlog. Management set the tone by guiding to $4.9-5.1 billion for full-year 2025 and promising "over 900 megawatts of active power before the end of the year." The company had 420 MW of active power and 1.6 GW contracted at the time.

Q2 FY25 (Aug 12, 2025) - Raised Guidance

Revenue came in at $1.2 billion (207% growth), beating the implied Q2 run-rate from prior guidance. Management raised full-year guidance to $5.15-5.35 billion. CoreWeave announced Goldman Sachs, Morgan Stanley, and Jane Street as new financial services customers. The $6 billion Lancaster, Pennsylvania data center investment was announced alongside the Weights & Biases acquisition completion.

Notably, management guided Q3 revenue to $1.26-1.30 billion. The Core Scientific acquisition proposal ($9 billion all-stock) signaled ambition to vertically integrate into owned power capacity.

Q3 FY25 (Nov 10, 2025) - Beat on Revenue, Missed Expectations on Guidance

Q3 revenue of $1.4 billion beat the $1.26-1.30 billion guide by a meaningful margin. Backlog nearly doubled to $55.6 billion. Customer concentration improved dramatically (35% largest customer, down from 50% a quarter earlier). Active power grew to 590 MW.

However, management disclosed that one third-party data center provider caused delivery delays affecting Q4 expectations. CapEx guidance was revised downward to $12-14 billion from a prior expectation, and full-year revenue guidance was tightened to $5.05-5.15 billion (the low end was below the prior $5.15-5.35 billion range). The stock initially surged on the beat but later fell 62% from its peak as the Denton delays emerged and construction concerns spread.

Core Scientific shareholders voted down the acquisition on October 30, 2025. A strategic bet that would have given CoreWeave 1.3 GW of owned power was dead.

Q4 FY25 (Feb 26, 2026) - Full Year Delivered, Aggressive 2026 Guidance

Q4 revenue hit $1.6 billion, bringing the full year to $5.1 billion - within the tightened guidance range. Active power reached 850 MW (below the "over 900 MW" original target from Q1, which was later revised). Management acknowledged that deploying infrastructure ahead of schedule led to a temporary margin compression: lease costs and depreciation commence while customer revenue ramps over subsequent months.

2026 guidance was aggressive: $12-13 billion revenue, $30-35 billion CapEx, $900M-1.1B adjusted operating income, and a target of $17-19 billion annualized run-rate exiting the year. The stock fell 18% the next day as investors digested the CapEx requirements and debt narrative.

Assessment

Management's credibility is mixed. They delivered on the big financial promises: revenue came in within guided ranges for FY25 despite construction setbacks. Customer diversification improved dramatically from 85% concentration to 35% in one year. The backlog grew from $25.9 billion to $66.8 billion across four quarters.

Where management stumbled: the "over 900 MW" active power target from Q1 was quietly revised; the Core Scientific acquisition was proposed and rejected within the same fiscal year; and the Denton construction delays were not fully transparent until after the stock had already declined significantly (triggering a securities class action).

The pattern is: management sets ambitious targets, mostly delivers on revenue and customer metrics, but has struggled with infrastructure execution timelines and occasionally moves goalposts on operational metrics. They are credible on demand and commercial execution. They are less credible on construction timelines and cost management. The gap between "vision" and "execution complexity" is real but narrowing as the company matures.

10. Shareholder Friendliness Index

CoreWeave went public in March 2025. It has existed as a public company for approximately 14 months as of this report.

Dividends: None paid. None declared. No dividend policy exists. The company is in a heavy growth and negative free cash flow phase, with all capital directed toward infrastructure expansion.

Share Buybacks: None announced. No share repurchase program has been authorized by the board.

Dilution: CoreWeave issued shares for its IPO ($1.5B raised), received a $2 billion equity investment from NVIDIA at $87.20/share (January 2026), and proposed an all-stock acquisition of Core Scientific (which was rejected). The Weights & Biases and OpenPipe acquisitions likely included stock consideration. Net dilution over the company's brief public history is material but quantifying the exact share count change requires the latest 10-Q.

Capital Allocation Priority: All available capital is directed toward CapEx ($30-35B planned for 2026) and debt service. Management's stated long-term target is 25-30% operating margins, but the path to shareholder returns through dividends or buybacks is years away, if it arrives at all.

Assessment: This is not a shareholder-return story. It is a growth-at-all-costs story. Investors buying CoreWeave are buying future free cash flow generation, not current distributions. The company has explicitly prioritized scale over profitability and shows no indication of shifting that priority in the foreseeable future.

11. Scenarios

Bull Case

The AI infrastructure supercycle sustains through the decade. CoreWeave delivers on its 2026 CapEx plan without material delays, bringing 1.7 GW online by year-end. The Meta, Anthropic, and OpenAI contracts ramp as scheduled, and Microsoft remains a stable base. Revenue exits 2026 at the $17-19 billion run-rate management guided, and by 2027 reaches $30 billion. The Vera Rubin deployment goes smoothly, reinforcing CoreWeave's "first to market" advantage with each NVIDIA generation.

Critically, inference workloads explode as AI agents go mainstream - every enterprise deploys AI assistants, every consumer app has a model behind it. The inference market dwarfs training, and CoreWeave's auto-scaling Kubernetes infrastructure captures disproportionate share. CoreWeave Federal lands major defense contracts. The cost of capital continues declining as the business proves its durability, reaching investment-grade across the entire capital structure.

In this world, CoreWeave becomes the defining infrastructure company of the AI era - the AWS of GPU compute. Its early mover advantage, NVIDIA relationship, and $66.8 billion backlog compound into a durable franchise.

Base Case

CoreWeave delivers most of what it has guided but with bumps. Some construction projects slip by weeks or months. The 2026 revenue target is met ($12-13 billion) but the run-rate exiting the year comes in at the lower end ($17 billion). Customer diversification continues, with Microsoft dropping below 50% of revenue, but the company remains dependent on a handful of large contracts.

Interest expense remains a significant drag, consuming 20-25% of revenue. The company reaches operating income profitability on an adjusted basis but generates minimal free cash flow because CapEx continues to match or exceed revenue. The stock trades on future potential rather than current economics.

The AI spending environment stays strong but shows signs of rationalization - customers negotiate harder on pricing, contract durations shorten slightly, and a few smaller AI companies (like those backing moonshot projects) pull back or fail. CoreWeave navigates through it because its largest customers (Meta, Microsoft) are financially healthy and contractually committed.

Bear Case

AI spending growth decelerates sharply in late 2026 or 2027. Perhaps a major model lab publishes a paper showing diminishing returns from scale. Perhaps a recession forces enterprises to cut AI budgets. Perhaps inference workloads migrate to cheaper custom silicon (Google TPUs, Amazon Trainium) rather than NVIDIA GPUs.

CoreWeave finds itself with $30+ billion in debt, massive data centers under construction, and customers who begin requesting contract modifications. OpenAI, burning $14 billion in 2026, either raises at punitive terms or seeks to renegotiate its $11.9 billion commitment. Microsoft, having built internal capacity, lets contracts roll off rather than renewing.

The securities class action over the Denton delays expands. Credit markets tighten for GPU-backed financing. CoreWeave's cost of capital rises just as its revenue growth slows. The company is forced to cut CapEx dramatically, ceding the "first to deploy" advantage to better-capitalized hyperscalers. The 1990s telecom parallel that bears have warned about begins to play out - not necessarily bankruptcy, but a painful period of deleveraging, slower growth, and broken promises.

In the worst version, a new NVIDIA architecture requires entirely different data center designs (new cooling, new power density), making CoreWeave's recently built facilities partially obsolete before their debt is repaid.

Report compiled from publicly available earnings call transcripts, SEC filings, press releases, and industry research. CoreWeave's Q1 2026 earnings call is scheduled for May 7, 2026 - two days after this report's date.

Sources:

• CoreWeave Investor Relations
• CoreWeave Q4 2025 Earnings Call Transcript - Motley Fool
• CoreWeave Q3 2025 Earnings Call - Investing.com
• CoreWeave Q2 2025 Earnings Call - Investing.com
• CoreWeave Q1 2025 Results - PR Newswire
• CoreWeave and Meta $21B Deal
• CoreWeave Anthropic Agreement
• CoreWeave $8.5B Financing Facility
• NVIDIA and CoreWeave Partnership
• CoreWeave IPO S-1 Filing
• CoreWeave Wikipedia
• Neocloud Market Forecast - Synergy Research
• CoreWeave Stock Decline - CNBC
• Core Scientific Merger Terminated
• CoreWeave Data Center Locations - DGTL Infra
• CoreWeave Federal Market Entry