CoreWeave
Capital-intensive asset-leasing arbitrage: borrow against Nvidia GPUs, deploy them, and rent compute on multi-year take-or-pay contracts. The economic engine is the spread between rental rates and the combined debt-service + depreciation cost of the fleet. Effectively a leveraged real-asset landlord whose collateral (GPUs) depreciates fast, not a software/SaaS business.
The thesis on this name
State of AI Compute
Short CRWV over 9–18mo as the most-leveraged, least-moated, most circular-financing-exposed neocloud — sized as a HEDGE/canary against the durable-toll longs, with the edge in structurally under-modeled GPU-collateral depreciation, refinancing-cliff cash dynamics, and a book-to-bill/utilization signal that pre-dates a…
State of AI Compute
The intended portfolio hedge/canary — most-leveraged, least-moated, circular-financing-exposed neocloud whose bear case ($30, -70%) fires exactly when the durable-toll longs hurt most. Options-ONLY (put spread): live sh…
State of Data-Center Power
The neocloud GPU-landlord layer is the most leveraged, least-moated, most circular-financing-exposed link sitting UNDER the entire power-and-cooling long basket — and the cleanest canary for the demand-side air-pocket. A CoreWeave debt/refinancing event or utilization pullback would show up in datacenter book-to-bill (Vertiv, Eaton, Comfort) and power lead-times one-to-two quarters before it hits any toll-booth income statement. AVOID as a long; if expressing the demand-quality risk, do it as a defined-risk paired hedge against the durable power/cooling tolls — sized as a hedge, not a core short, because a momentum tape can run a richly-valued, capex-fed name far before fundamentals bite.
State of the AI Cloud
The undisputed neocloud scale leader (first to sign all four top AI labs + Microsoft, ~$99.4B backlog, ~56% adj-EBITDA margin, 1GW+ live / 3.5GW+ contracted power) — but debt-fueled ($31-35B capex, ~$740M Q1'26 GAAP loss) and concentration-exposed (Microsoft ~67% of 2025 revenue, OpenAI ~a third). Own it SMALL as the sector bellwether and the cleanest canary for a demand-side air-pocket, not as an anchor.
State of the AI Cloud
Undisputed neocloud scale leader — first to sign all four top AI labs plus Microsoft, ~$99B backlog, 56% adj-EBITDA margin — but debt-fueled and concentration-exposed.
State of the AI Cloud
Scale leader (all four labs + Microsoft, ~$99B backlog) owned SMALL as the sector bellwether + demand-air-pocket canary, not an anchor.
State of Frontier AI
Pair-short the most leveraged link in the chain vs the NVDA long — $4.2B 2026 maturity, GPU-collateralized debt, customer concentration, circular financing — expressing 'compute demand is real but the financing is fragile' without shorting the theme.
State of Frontier AI
Short the most leveraged, debt-financed link in the chain ($4.2B 2026 maturity, GPU-collateralized loans, customer concentration) against the NVDA/arms longs. Expresses 'compute demand is real but the financing structure is fragile' without shorting the theme outright.
Earnings, margins, COGS & capex
CoreWeave is growing revenue triple-digits (Q1 FY26 $2.08B, +112% YoY) on a $99.4B contracted backlog, and is adj.-EBITDA positive (56% margin), but it is GAAP operating- and net-loss-making (-$144M / -$740M in Q1) and burns multi-billion FCF every quarter because capex (~$7.7B/qtr) and interest (~$0.5B+/qtr) outrun cash rents. The model is a leveraged depreciation/refinancing bet: it funds GPU buys with ~$25B of GPU-collateralized debt and customer-prepay financing, betting the six-year accounting life of the fleet holds long enough to repay the debt. The bear edge sits in under-modeled GPU depreciation (true economic life may be 2-4 years), refinancing cliffs, and extreme customer concentration.
Income statement — where each revenue dollar goes
% of revenueOf every $1 of revenue, ~35¢ is cost of goods and ~66¢ operating expense, and the remainder is an operating loss.
Revenue trend
Margins
down/flat - compressing as new low-utilization capacity ramps
flat-to-down - down from ~60%+ as scaling costs hit
negative - heavy depreciation + SBC keep GAAP operating loss
negative/worsening - interest expense (~$536M, doubled YoY) drives the gap below operating
negative - structural while building; capex >> rents
COGS structure
COGS is dominated by data-center operating cost: power/electricity, colocation and facility lease expense, and the operations/staffing to run the GPU fleet. The far larger economic cost sits BELOW gross profit as (1) GPU depreciation (six-year straight-line, the contested assumption) and (2) interest on the GPU-collateralized debt (~$536M in Q1, doubled YoY). So a 65% GAAP gross margin masks a GAAP operating loss once depreciation and SBC are loaded, and a net loss once interest is loaded.
Capex
Q1 FY26 capex $7.70B; FY26 capex guide $31-35B. It funds buying Nvidia GPUs (Hopper/Blackwell GB200/GB300) plus data-center build-out and power to hit >1.7 GW active power by end-2026 (>3.5 GW contracted). Capex is the company's defining feature: ~2.5x revenue, funded by debt + customer prepayments, and the source of both the growth and the refinancing/depreciation risk.
Latest earnings
Mixed: revenue BEAT (~$2.08B vs ~$1.97B LSEG consensus) but EPS MISS (adj. loss -$1.12 vs ~-$0.90 expected); Q2 revenue guide midpoint (~$2.525B) below ~$2.69B consensus [fact/estimate]
FY26 revenue reaffirmed $12-13B; Q2 FY26 $2.45-2.60B; FY26 capex raised to $31-35B; exit-2026 run-rate raised to $18-19B [guidance]
- Revenue backlog / RPO
- $99.4B (Q1 FY26, +~50% QoQ; FY25 RPO $60.7B)
- Contracted / active power
- >3.5 GW contracted; >1 GW active (target >1.7 GW active by end-2026)
- Total debt / interest
- $24.86B total debt; interest expense ~$536M in Q1 (doubled YoY)
- Net loss / FCF
- Net loss -$740M; FCF -$4.71B (capex $7.70B)
Growth drivers
- Long-term take-or-pay contracts — $99.4B revenue backlog (Q1 FY26), up ~50% QoQ, with >$40B new commitments booked in the quarter [fact]
- Mega-customer ramps — OpenAI commitments up to ~$22.4B and Meta up to ~$35B (through 2031-2032) layering on as Microsoft's share falls [fact]
- Capacity expansion — contracted power >3.5 GW, targeting >1.7 GW active by end-2026; largely 'sold out' of 2026 capacity [fact]
- Blackwell generation (GB200/GB300) deployment commanding premium rental rates [fact]
- Privileged Nvidia access (Nvidia owns ~7%, gives early-allocation + a backstop capacity purchase agreement) [fact]
- Exit-2026 annualized run-rate guided up to $18-19B, implying continued contract conversion [guidance]
Reported financials — SEC EDGAR
Audited GAAP figures pulled from SEC filings · latest filing 2026-03-02. The audited primary-source spine — not financial advice.
Revenue — annual (GAAP)
Margins & balance sheet — FY’25
Bull & bear
CoreWeave is the category-defining neocloud riding the largest infrastructure build-out of the decade, with a $99.4B contracted backlog, privileged Nvidia access, and 56% adj.-EBITDA margins that turn into real cash as capacity matures and capex intensity normalizes.
- $99.4B backlog (+~50% QoQ) and exit-2026 run-rate guided to $18-19B give rare multi-year revenue visibility for a hyper-growth name [fact]
- Take-or-pay contracts with Microsoft/OpenAI/Meta de-risk demand; Nvidia's ~7% stake + capacity backstop align the most important supplier [fact]
- 56% adj.-EBITDA margin shows the unit economics work at the asset level; FCF turns positive once the capex super-cycle moderates [fact]
- First-mover scale, deployment speed, and Nvidia allocation are hard for new entrants to replicate quickly [estimate]
- If GPU demand stays tight through 2-3 more generations, the six-year depreciation assumption proves conservative and equity compounds [estimate]
CRWV is the most-leveraged, least-moated, most circular-financing-exposed neocloud: a leveraged bet that fast-depreciating GPU collateral and a few mega-contracts will service ~$25B of debt long enough to deleverage. We hold it SHORT over 9-18 months as a hedge/canary, with the edge in structurally under-modeled GPU-collateral depreciation, refinancing-cliff cash dynamics, and customer/financing circularity.
- Depreciation mismatch: six-year GAAP life vs likely 2-4yr economic life (Burry/Chanos thesis) overstates earnings and overstates the collateral backing ~$25B of debt [estimate]
- Refinancing cliff + cash burn: FCF -$4.71B in one quarter, interest ~$536M/qtr (doubled YoY); a risk-off market makes rolling GPU-collateralized debt expensive or impossible [fact/estimate]
- Concentration: Microsoft ~67% of FY25 revenue; loss/renegotiation of one mega-counterparty would impair the backlog's value [fact]
- Circular financing: Nvidia funds the buyer of Nvidia chips and backstops unsold capacity; OpenAI is a customer Nvidia also funds - a confidence shock re-rates the whole chain [estimate]
- No durable moat: reselling commodity Nvidia GPUs; hyperscaler custom silicon (TPU/Trainium/Maia) and rival neoclouds (Nebius, Crusoe, Lambda) erode pricing power [fact/estimate]
What it is worth
Reverse-DCF / EV-multiple sanity check, framed as a short/hedge. EV ~$89B vs ~$6.2B LTM revenue (~14x EV/sales) and ~$1.16B Q1 adj. EBITDA (~$5-6B annualizing => ~15-18x EV/adj.EBITDA, well above mature-infra multiples). The price implies high-teens revenue CAGR for years AND that the six-year GPU depreciation life holds AND that ~$25B of debt refinances smoothly - three contestable assumptions stacked.
$35-65
a depreciation re-rating, a refinancing scare, an AI-capex air-pocket, or a mega-customer renegotiation compresses the multiple and forces the leverage into focus - the 9-18mo short payoff [estimate]
$80-110 (near current ~$101)
triple-digit growth persists but FCF burn + leverage + concentration cap the multiple; range-bound on each capex/print [estimate]
$130-175+ (sell-side ~$132 mean target / prior 52-wk high ~$173): backlog converts, FCF inflects positive, depreciation proves conservative, Microsoft concentration falls cleanly [estimate]
Adj. EBITDA flatters the picture: it excludes the depreciation and interest that turn the business into a -$740M net loss; valuing on EBITDA ignores that the asset depreciates and the debt accrues. The reverse-DCF only works if economic GPU life ~ accounting life - the crux of the short.
SWOT
Strengths
- Scale + speed — largest pure-play neocloud, fastest cloud platform to $5B ARR; $99.4B contracted backlog gives multi-year revenue visibility [fact]
- Privileged Nvidia relationship — ~7% Nvidia ownership, early GPU allocation, and a Nvidia capacity-purchase backstop reduce demand-side tail risk [fact]
- Take-or-pay contract structure with blue-chip counterparties (Microsoft, OpenAI, Meta) underpins the backlog [fact]
- Operationally adj.-EBITDA positive (56% margin) and best-in-class GPU utilization/time-to-deploy vs general-purpose clouds [fact]
Weaknesses
- Extreme leverage — ~$25B debt on ~$4.8B equity, ~$536M/qtr interest (doubled YoY), and persistent multi-billion FCF burn [fact]
- Severe customer concentration: Microsoft ~67% of FY25 revenue; top handful of customers ~all revenue [fact]
- GAAP loss-making (-$740M Q1) with net margin -36% — profitability depends on the contested six-year GPU depreciation assumption [fact]
- Asset is depreciating collateral — GPUs may have 2-4yr economic life vs 6yr accounting life, mismatching debt amortization [estimate]
- Thin moat: it resells a commodity (Nvidia GPUs) that hyperscalers and rivals can also buy; switching is contract-, not technology-locked [estimate]
Opportunities
- Continued AI-training/inference demand could keep capacity 'sold out' and rates firm through the next GPU generations [estimate]
- Diversification away from Microsoft as OpenAI/Meta ramp lowers single-customer risk if it executes [fact]
- Software/orchestration layer (CoreWeave fleet management, Weights & Biases acquisition) could add a stickier, higher-margin layer [fact]
- Falling DDTL cost of carry (from ~15% to ~5.9%) and unsecured-notes refinancing could lower the interest drag [fact]
Threats
- AI-capex air-pocket / 'bubble' unwind — if hyperscaler GPU demand softens, rental rates and renewals fall while debt stays fixed [estimate]
- GPU obsolescence (Blackwell -> Rubin -> next) compresses residual value of the collateralized fleet faster than modeled [estimate]
- Circular-financing scrutiny (Nvidia funds buyer of Nvidia chips — OpenAI a customer Nvidia also funds) invites a confidence shock / re-rating [estimate]
- Refinancing cliff — large debt maturities into a higher-rate or risk-off market could create a liquidity squeeze [estimate]
- Hyperscalers in-sourcing (custom silicon — Google TPU, Amazon Trainium, Microsoft Maia) shrinks the merchant-GPU rental TAM [fact]
Moats, dependencies & bottlenecks
Moats
~7% Nvidia ownership, capacity backstop) real today but Nvidia spreads allocation across many neoclouds and hyperscalers; not exclusive The single most important edge, but it is granted, not owned - Nvidia can dilute it at will [fact/estimate]
fastest to $5B ARR, time-to-deploy clusters) scale helps win mega-contracts but is capital-replicable by well-funded rivals Advantage is operational execution, not a structural barrier [estimate]
contractual switching cost, not technological lock-in; renewals are the risk Backlog is only as good as counterparty willingness/ability to pay and renew [fact]
early; could become stickier but not yet a proven lock-in The only path to a genuine moat beyond commodity GPU rental [estimate]
a liability, not a moat; high leverage is a vulnerability vs cash-rich hyperscalers Hyperscalers self-fund GPUs from operating cash flow; CoreWeave rents capital expensively [fact]
Dependencies
Sole effective supplier of the core asset, a ~7% owner, and a backstop customer - existential single-point dependency and the heart of the circularity critique [fact]
~67% of FY25 revenue; any reduction/renegotiation impairs revenue and backlog value before OpenAI/Meta fully ramp [fact]
~$25B debt, multi-billion quarterly FCF burn; relies on continuous access to GPU-collateralized + unsecured debt and customer prepayments [fact]
Up to ~$22.4B (OpenAI) and ~$35B (Meta) commitments; concentration shifts but does not disappear [fact]
Largely leased facilities + power contracts; gating constraint on activating contracted GW and a fixed cost if demand softens [fact]
Advantages
- Largest pure-play neocloud with first-mover scale and the fastest path to $5B ARR in cloud history [fact]
- Privileged Nvidia relationship — early allocation of scarce latest-gen GPUs + ~7% Nvidia ownership + capacity backstop [fact]
- $99.4B contracted backlog with blue-chip counterparties gives multi-year revenue visibility [fact]
- Operational excellence: high GPU utilization and rapid cluster deployment vs general-purpose clouds [estimate]
- 56% adj.-EBITDA margin demonstrates the asset-level unit economics work when capacity is utilized [fact]
Weaknesses
- Extreme leverage (~$25B debt, ~$536M/qtr interest doubled YoY) on a thin ~$4.8B equity base [fact]
- Persistent multi-billion FCF burn (-$4.71B in Q1 FY26) with no clear near-term self-funding [fact]
- Customer concentration: Microsoft ~67% of FY25 revenue [fact]
- Profitability hinges on a six-year GPU depreciation assumption many investors believe is too generous (true life 2-4yr) [estimate]
- No durable technology moat — reselling a commodity (Nvidia GPUs) any well-capitalized rival or hyperscaler can also buy [estimate]
- Circular-financing optics (Nvidia funds the buyer of Nvidia chips, also backstops demand) create confidence-shock fragility [estimate]
Bottlenecks
- Power and data-center capacity — activating >1.7 GW of active power by end-2026 is gated by electricity, grid interconnect, and colocation availability [fact]
- Capital access — the model cannot grow faster than it can raise GPU-collateralized debt + customer prepayments; rate/risk-off shocks throttle growth [fact/estimate]
- Nvidia GPU allocation — supply of the latest accelerators (Blackwell GB300, then Rubin) caps deployable capacity [fact]
- Skilled operations talent to stand up and run large GPU clusters at high utilization [estimate]
- Customer-renewal timing — converting backlog to cash depends on counterparties taking and renewing capacity on schedule [estimate]
Top signals & trends
Top signals
Successful diversification is bullish; a stall or a Microsoft renegotiation is the bear trigger [fact]
FCF -$4.71B in Q1; the bull case requires capex intensity to fall and FCF to turn - watch each print [fact]
Falling DDTL carry (15%->5.9%) is a positive tell, but the size of upcoming maturities into any risk-off window is the core short catalyst [fact/estimate]
Evidence that economic life is <6yr (re-pricing of expired contracts, write-downs) would validate the depreciation-mismatch thesis [estimate]
$99.4B backlog only matters if it converts and renews at price; the 95%-of-original-price H100 re-book is the bull data point to track [fact]
CRWV is the highest-beta expression of the AI-infra trade; any hyperscaler capex cut hits it first and hardest [estimate]
Trends
Drives the backlog and 'sold out' 2026 capacity; the entire bull thesis rests on this persisting [fact]
New gens command premium rates (positive) but accelerate obsolescence of the in-place collateralized fleet (negative for depreciation/residuals) [estimate]
Bloomberg/short-seller attention raises the odds of a confidence-driven re-rating across the chain [estimate]
In-sourcing erodes the merchant-GPU rental TAM over time, capping CoreWeave's long-run pricing power [fact]
DDTL carry fell from ~15% to ~5.9%; cheaper refinancing reduces the interest drag if markets stay open [fact]
Scarcity supports rates for those who have power, but gates CoreWeave's ability to activate contracted GW [fact]
Ecosystem & competitor graph
Suppliers feed the company; customers pull from it. Line thickness shows the strength of each tie (supply-chain dependency, customer earnings contribution). Hover to isolate a tie.
Core GPU supplier (Hopper/Blackwell GB200/GB300), ~7% owner, and capacity backstop - the dominant single dependency
Server/rack integrator assembling GPU systems for neocloud build-outs
GPU server/rack supplier to AI data centers
Power, cooling, and thermal-management infrastructure for high-density GPU data centers
Data-center host/colocation partner (the $9B all-stock acquisition was terminated Oct 2025 after a shareholder vote; ongoing hosting relationship)
High-speed networking/switching for GPU cluster fabric (Ethernet); peer to Nvidia's InfiniBand
Largest customer, ~67% of FY2025 revenue; rents CoreWeave capacity to serve Azure/OpenAI workloads
Up to ~$22.4B in multi-year commitments (private); a customer Nvidia also funds - the circularity node
Up to ~$35B in commitments through 2031-2032 to support Llama training/inference
Backstop 'customer' via a capacity-purchase agreement (~$6.3B through 2032) for unsold capacity - both supplier and buyer
Enterprise AI customer of CoreWeave compute
Closest listed pure-play neocloud peer; spun out of Yandex, ~3.5 GW contracted, won large Microsoft/Meta deals - the cleanest comp and competitor for the same contracts [fact]
Both CoreWeave's largest customer AND a competitor that can in-source GPU capacity; the dependency cuts both ways [fact]
Hyperscaler renting GPU capacity + custom Trainium/Inferentia silicon shrinks the merchant-GPU TAM [fact]
TPU custom silicon + GPU cloud; self-funds capex from cash flow, structurally lower cost of capital [fact]
Aggressively building GPU cloud capacity and winning large AI-training contracts (incl. OpenAI), competing for the same mega-deals [fact]
Private neoclouds (Crusoe, Lambda, Together) plus listed IREN compete for GPU-rental workloads; fragment pricing at the margin [fact]