Executive Summary

CoreWeave is the cleanest public-market expression of the AI training capex super-cycle: a $99.4B contracted backlog¹, 8× FY26 revenue, that is 98% take-or-pay, anchored by Microsoft (~67% of FY25 revenue), OpenAI ($22.4B), Meta (~$35.2B), and Anthropic. NVIDIA holds an 11.5% equity stake and has signed a $6.3B capacity backstop⁴ through April 2032, plus granted CoreWeave Preferred NPN status: first-to-GA across GB200 NVL72, GB300 NVL72, and the first Vera Rubin NVL72 deployment on June 1, 2026⁹.

Tactical: CRWV trades at $93.51, ~34% upside to our $125 12-month PT. Blended FV ~$122 (30/45/25 weights). Rating Buy on the AI-infrastructure pure-play with NVIDIA backstop, but elevated balance-sheet risk (640% D/E) warrants smaller-than-conviction sizing.

Investment Thesis Summary

Overall Rating: Buy · 12-month PT $125 (Blended FV $122.50)

We initiate Buy on CoreWeave with a $125 12-month price target. CRWV is the purest public vehicle for AI training infrastructure scaling: an $99.4B contracted backlog (8× FY26 revenue) that's 98% take-or-pay, an 11.5% NVIDIA equity stake plus $6.3B capacity backstop through 2032, and structural first-to-GA status across every NVIDIA generation from GB200 through Vera Rubin. At ~5.8x EV/2026E sales vs Nebius at ~17x and IREN at ~24x, CRWV trades at the lowest multiple in the neocloud universe despite the largest contracted backlog. The discount reflects (justified) leverage and concentration concerns but underweights the contracted-revenue visibility. Risk-adjust sizing to half a conviction position given $24.9B debt against $3.3B book equity (~640% D/E).

1 · Business Overview

What CoreWeave Is

CoreWeave is a "neocloud" / GPU-as-a-Service infrastructure provider: a publicly-traded specialist hyperscaler that rents NVIDIA GPU capacity to AI labs and Big Tech under multi-year take-or-pay contracts. Unlike AWS/Azure/GCP, CoreWeave doesn't offer a managed platform, databases, or SaaS, it offers bare-metal GPU clusters tuned for large-scale training and inference. Founded 2017 (originally an Ethereum miner), pivoted to AI compute in 2019, IPO'd March 28, 2025 at $40/share on Nasdaq.

Customer Mix (FY25 Revenue Share)

Why Hyperscalers Buy From CoreWeave Instead of Building

• Speed-to-deploy: CRWV's NVIDIA-preferred allocation lets hyperscalers ramp GPU capacity 12-18 months faster than building their own clusters.
• Capex offload: Multi-year leases moved off-balance-sheet vs $30B+ capex to build internally.
• Specialized stack: NVL72 racks, liquid cooling, InfiniBand fabric tuned for largest-scale training runs.
• Caveat: Same hyperscalers (MSFT TPU/Maia, Google TPU v7, Amazon Trainium2) are simultaneously building competing in-house silicon, see Risks.

2 · NVIDIA Symbiosis: Preferred Cloud, $6.3B Backstop, 11.5% Owner

No other public company has the depth of NVIDIA relationship CoreWeave enjoys. NVIDIA is simultaneously CoreWeave's largest supplier, second-largest equity owner (~11.5% post-Jan 2026 $2B follow-on at $87.20), customer (via Lepton AI platform reselling CRWV capacity), and a contractual capacity backstop through 2032. This three-way alignment, supplier / investor / customer, is the single most important reason CRWV trades at a defensible premium to other AI-infrastructure pure-plays.

The Three Pillars of the Symbiosis

The Cadence Advantage (NVIDIA Generations)

The $6.3B Capacity Backstop

Under the September 2025 agreement, NVIDIA commits to purchase any GPU capacity CoreWeave cannot sell to third parties through April 2032, up to $6.3B aggregate. This converts CRWV's worst-case demand risk (a hyperscaler renewal failure or AI capex pause) from a cliff into a floor: NVIDIA absorbs the inventory at known economics. The backstop is the single most important reason credit analysts have begun to model CRWV's revenue as more contracted than typical infrastructure SPVs.

3 · The $99.4B Backlog: Hyperscaler Take-or-Pay Mechanics

The $99.4B remaining performance obligations (RPO) balance disclosed at Q1 2026 is the single most important number on CoreWeave's balance sheet. It represents ~8× FY26 revenue under guidance and provides multi-year visibility unusual for any pre-profit infrastructure company. The contracts are structured as take-or-pay: customer pays for committed capacity whether or not it's used, with ~98% of FY25 revenue running through this structure.

Backlog Building Blocks (Mega-Contracts)

Recognition Schedule

Per the Q1 2026 disclosure, the $99.4B is scheduled to recognize as revenue:

• ~36% within 2 years (through Q1 2028), anchors the FY26-FY27 guidance
• ~75% within 4 years (through Q1 2030), visibility runway
• ~25% beyond 48 months: long-dated commitments dependent on power buildout

The pace of conversion is the single biggest swing factor in our model, a 10% acceleration of the 0-2-year bucket lifts FY26 revenue ~$1.3B and pulls forward the EV/Sales compression by ~6 months.

4 · Operating Drivers

GPU Fleet Composition (Illustrative)

Power-to-Revenue Bridge

The 2.5 GW gap between current contracted capacity (3.5 GW) and active capacity (1.0 GW) is the execution bridge for the 2026-2027 ramp. Each 100 MW that converts from contracted-to-active translates to roughly $500M of incremental annual revenue at current rates.

Data Center Footprint

• Active sites: ~49 data centers across US (NJ, IL, TX, NV), UK (London), Europe (Norway, Sweden) per coreweave.com/ai-data-centers.
• Major capacity deals: Applied Digital "Delta Forge 2" lease (Apr 2026)⁶; Core Scientific merger terminated Oct 30, 2025¹¹ (stockholder vote failed; CoreWeave confirmed in Dec 2025 it would not re-attempt the acquisition).
• 2030 build plan: ~8 GW of capacity across 50+ sites, ~70% liquid-cooled.

5 · Financial Health

Quarterly & Annual Summary

The Depreciation Problem

D&A was $1.14B in Q1 2026, 55% of revenue. CoreWeave depreciates GPUs over 4 years (industry-standard), but NVIDIA's 18-month platform cadence (Hopper → Blackwell → Rubin) means H100s deployed 2023 are economically obsolete by 2026 even though they're 50% depreciated on the books. The accounting works; the cash economics are tighter. If useful life proves to be 3 years (not 4), reported margins compress further and a bear-case scenario crystallizes. The result: Q1 26 adjusted operating income compressed to $21M (1.0% margin) vs $163M (17.0%) in Q1 25, management called this the cyclical low point as committed capacity outpaces revenue recognition.

5b · Capital Allocation: Debt Stack & Capex

Debt Profile

2026 CapEx Funding Mix

Management guides $31-35B of 2026 capex against $12-13B of revenue. Funding mix: ~50% from new debt issuance (DDTL 5.0 + planned 2026 senior notes), ~30% from operating cash flow + adjusted EBITDA, ~15% from NVIDIA-related vendor financing, ~5% from existing cash. FCF is structurally negative until the contracted-backlog converts faster than incremental capex, modeled to occur in late 2027 / early 2028 in the base case.

6 · Valuation & Comps

Peer Comparison

Consensus

6b · EV/Sales Scenario Model

CRWV has structurally negative FCF for the visible forecast horizon (until late 2027 / early 2028 base case), so a DCF would force terminal-value assumptions that swamp the explicit forecast. We use a 2026E EV/Sales scenario framework: assign a multiple × 2026E revenue × net debt adjustment ÷ shares = implied price per share.

Sensitivity: EV/Sales × 2026E Revenue

7 · Catalysts (Next 12 Months)

8 · Scenario Targets

9 · Bull vs Bear Debate

Our weighted view: Bull thesis carries 30% probability (backlog is real, NVIDIA partnership is structural), bear thesis 25% (leverage tail is non-trivial), base case 45%. Net: Buy with discipline: own CRWV at smaller-than-conviction sizing into Q3-Q4 2026 catalysts, but maintain stops below $80.

10 · Risk Factors

11 · Technicals

Key Levels

★ Interactive Tools

Ask the Thesis AI-assisted checking…

Describe a CRWV scenario in natural language; the assistant returns a structured impact analysis against this dashboard's EV/Sales model, $99.4B backlog, and NVIDIA-relationship framework. Powered by Claude via a Cloudflare Worker proxy.