CoreWeave (CRWV) · Deep analysis

Report previewCRWV.USReport date Last updated
← Back to reports
Report contents
01

Business model

Business model: contracted compute with concentrated buyers

CoreWeave rents GPU capacity, largely under multiyear commitments, to a small set of AI developers and cloud providers, mostly based in the United States. It buys NVIDIA accelerators, leases data-center space and power, and sells access to clusters that customers use to train and run models. The buyers are large and sophisticated: AI labs and hyperscalers ordering capacity in tens of thousands of GPUs, not enterprises buying a few instances. That makes CoreWeave a capital-intensive compute landlord in an expansion stage. The multiyear contract is the business. It sets revenue, supports the debt that funds the hardware, and decides whether each cluster pays back its cost before the GPUs age out.

Customers: a few buyers set the clearing price

Concentration is structural here, not a phase of growth. Microsoft supplied 62% of 2024 revenue, CoreWeave disclosed in its fiscal 2025 10-K, and Meta and OpenAI together accounted for roughly two-thirds of the backlog as of April 2026. Only a handful of organizations can absorb a dedicated cluster of that scale, and each can build its own data center or buy capacity from a hyperscaler. CoreWeave therefore negotiates with counterparties that have credible alternatives and a clear view of the hardware costs underneath the contract.

Its early advantage was speed: getting NVIDIA's newest chips energized and usable ahead of larger cloud rivals. That advantage is real and perishable. It shows up as time-to-cluster and as allocation during GPU shortages. It does not translate into pricing power, because once a cluster is running the service is close to a commodity sold by the GPU-hour. Whether contract prices hold as those buyers grow larger is the single best test of the model.

Unit economics: capital arrives before revenue

A new cluster consumes cash before it bills. CoreWeave buys the GPUs, secures power and shell space, installs, and ramps. Depreciation is the dominant cost line: it absorbed 48% of FY2025 revenue and about 54% of revenue in the June 2026 quarter, when gross margin fell to 65.9% from 74.2% a year earlier. That is what the model looks like when hardware costs are fixed and the revenue side is negotiated.

The company sits between NVIDIA, which holds pricing power on the input, and a few buyers who hold it on the output. CoreWeave's own leverage comes from utilization and from the speed of turning signed contracts into billable capacity. Software that helps customers schedule and manage clusters supports that, but it is not yet a separate profit pool the company reports.

Contract length matters more than most line items. If a commitment runs shorter than the GPU's depreciable life, the residual value depends on re-renting aging hardware at whatever the spot market then pays.

Demand: training is lumpy, inference is price-sensitive

The spending behind those contracts is concentrated in frontier-model training, a small number of very large programs. Training workloads move to whoever has the newest hardware, so each chip generation carries a shorter commercial window than the last. Inference is steadier and would extend the earning life of older GPUs, but it is easier to spread across providers and more sensitive to cost per token. A pause in frontier training would leave CoreWeave holding capacity its buyers may ask to delay.

Expansion: the schedule is the constraint

Demand has not been the binding constraint; power and data-center delivery have been. CoreWeave leases much of its capacity from third-party developers and waits for those sites to be energized before revenue starts. Every quarter of slippage leaves capital spent and idle. The company also commits to GPU purchases on NVIDIA's timetable, which limits how far it can slow spending when a site runs late.

That gives the business a particular failure mode: orders and backlog can grow while revenue lags, and the gap appears as capital spending without matching sales. Scale is the counterweight. Larger clusters spread fixed site and software costs and improve supplier terms, which is why the company keeps prioritizing growth over near-term profitability.

What would change the judgment

CoreWeave's account of itself rests on contracts that oblige buyers to pay for reserved capacity whether or not they use it, signed years ahead. The judgment that concentration is the main structural weakness would weaken if disclosed terms showed average contract duration at or beyond GPU useful life, with per-GPU pricing holding across successive contract vintages. A cohort of large customers outside the three anchors, signing at comparable prices, would point the same way.

Public disclosure limits how far this can be tested. CoreWeave does not publish per-contract pricing, utilization, renewal rates or customer-level margins. The evidence has to be assembled from concentration percentages, backlog figures, announced awards and the depreciation schedule, and each of those is partial.

Questions worth following

  • Will CoreWeave's capacity delivery keep pace with contracted megawatts?

    Delivering 3.7 gigawatts of contracted capacity from 1.5 gigawatts active requires building data centers, securing power, and installing GPUs. Delays push revenue and cash out while debt service continues. The answer determines whether the backlog converts into cash before maturities.

  • Will CoreWeave's take-or-pay contracts survive a customer pullback?

    The backlog supports both the debt and the revenue forecast. Firm commitments would delay the revenue effect of a spending pause by years; contracts customers can cut without penalty would mean reported backlog overstates future cash, leaving lenders exposed to the shortfall.

02

Financial quality

Financial quality: cash still trails the build

CoreWeave's gross margin is real; the earnings below it are not yet. The company earned a 67.7% gross margin in the September 2026 quarter and still lost $0.71 billion, because depreciation on a fast-expanding GPU fleet and interest on the debt that bought it absorb nearly everything the equipment brings in. Free cash flow has been negative in every fiscal year on record and reached −$17.9 billion in the first nine months of 2026. The operating line has started to improve. The funding structure has not.

The record: growth financed on the liability side
USD bn unless notedFY2022FY2023FY2024FY2025Q3 FY2026Assessment
Revenue0.020.231.915.133.66Strong
Revenue growth y/y (%)n/mn/m731169169Exceptional
Gross margin (%)23.469.974.371.767.7Strong, eroding
Operating income-0.02-0.010.32-0.050.21Weak
Net income-0.03-0.59-0.86-1.17-0.71Very weak
Net margin (%)-196.2-259.4-45.1-22.7-19.4Very weak
Interest expense0.010.030.361.230.85Weak
Operating cash flow0.001.832.753.061.98Adequate
Capex, cash basis0.072.948.7010.319.45Very weak
Free cash flow-0.07-1.11-5.95-7.25-7.47Very weak
Total assetsn/d4.9817.8349.3091.48—
Total liabilities/assets (%)n/d1121029396Very weak
Total debtn/d1.568.2021.84~47Very weak
Current ration/d0.500.390.460.44Very weak
ROE (%)n/d99.6170.9-79.9-15.7Not meaningful
ROA (%)n/d-11.9-7.6-3.5-0.8Weak

*Single quarter; flows are not comparable with annual columns. Q3 debt from Counterpoint Research.

Revenue went from $1.91 billion in FY2024 to $5.13 billion in FY2025, and company guidance calls for $13.0–13.3 billion in 2026, implying roughly $4.7–5.0 billion in the December quarter. Gross margin moved the other way, from 74.3% to 71.7% to 67.7%, a cumulative decline of more than 6 percentage points. The cost side explains it: capacity entering service carries a heavier depreciation charge per dollar of revenue than the fleet already billing.

Equity cannot carry this balance sheet. It was negative at the end of FY2023 and FY2024, $3.33 billion at FY2025, and $4.03 billion at September 30, 2026 against $91.5 billion of assets. The 171% ROE reported for FY2024 is arithmetic on a near-zero denominator, not a return. ROA gives the usable version: −7.6% in FY2024, −3.5% in FY2025, −0.8% in the September 2026 quarter.

Quarterly: operating income turns, cash does not
USD bnRevenueOper. incomeNet incomeTotal assetsTotal liabilitiesCashReceivablesOCFCapexFCFAssessmentTrend
Q1 FY20250.98-0.03-0.3221.8618.791.281.060.061.41-1.35Weak—
Q2 FY20251.210.02-0.2926.2422.421.151.93-0.252.45-2.70WeakStable
Q3 FY20251.360.05-0.1132.9129.031.891.661.692.39-0.70WeakImproving
Q4 FY20251.57-0.09-0.4549.3045.973.953.171.564.06-2.50WeakMixed
Q1 FY20262.08-0.14-0.7455.5750.813.022.122.987.69-4.71WeakMixed
Q2 FY20262.58-0.05-0.6377.0772.056.402.540.686.42-5.74WeakDeteriorating
Q3 FY20263.660.21-0.7191.4887.465.534.281.989.45-7.47WeakMixed

Net PP&E is not broken out in the summary balance sheet; the capex line is the best available read on the asset base.

Operating income turned positive at $0.21 billion in Q3 2026 after four quarters of small losses, and operating cash flow was positive in six of the last seven quarters. That cash figure is volatile for reasons only loosely tied to the income statement: it hit $2.98 billion in Q1 2026 on $2.08 billion of revenue, then fell to $0.68 billion in Q2 while revenue rose 24%. Prepayments and payable timing move it more than margin does.

Three relationships that set the quality

Revenue and cost per megawatt decide the outcome. Counterpoint Research estimates CoreWeave earned about $9 million per active megawatt annualized in Q3 2026 against roughly $3.5 million of running cost, a 62% cash margin, with depreciation and interest adding about $6.8 million. On those estimates breakeven sits near $10 million per megawatt, leaving the fleet about $1 million short. Gross margin can hold while the business still loses money at the unit level; the gap has to close through price, mix toward newer chips, or cheaper financing.

Capex intensity explains the borrowing. Cash capex ran at 201% of revenue in FY2025 and 283% in the first nine months of 2026, and management guides to $37–39 billion for the full year. Capex recognized in Q3 2026 was $14.5 billion against $9.44 billion actually paid for equipment, so part of the build sits in accrued purchases that will draw cash later. Free cash flow does not overstate the burden, and it may still understate it.

Working capital has been a mild help so far. Receivables rose from $2.54 billion to $4.28 billion in Q3 2026, faster than revenue's 42% sequential gain, but receivable days fell from about 128 in FY2025 to roughly 85, so collection improved even as the balance grew. Whether that holds as enterprise customers with longer payment terms join the mix is unresolved.

What to check, and one offset
  • Refinancing. A current ratio of 0.44, with $5.53 billion of cash against $11.54 billion of short-term debt and $24.39 billion of current liabilities at September 30, 2026. Total debt was about $47 billion. The maturity schedule and undrawn facilities set the real constraint, and neither appears in the summary statements.
  • Interest. $0.85 billion in Q3 2026, about four times operating income and 23% of revenue. Counterpoint expects more than $1 billion in Q4. The fixed-versus-floating mix determines how much of that cost is locked in.
  • Equipment commitments. The $5 billion gap between capex recognized and cash paid in Q3 is a future claim on liquidity; purchase commitments would size it.
  • Backlog credit quality. The $132.5 billion backlog, of which roughly $34 billion falls within 24 months, rests on a narrow set of AI labs and hyperscalers. How those counterparties pay, and when, drives receivables and cash conversion.

The offset worth holding onto: CoreWeave re-signed older Ampere and Hopper GPUs at 30–50% above prior-quarter prices, Counterpoint reported. If aging hardware reprices upward, the depreciation schedule is less punishing than the unit economics imply. The debt is the part that does not reprice in the company's favor.

Questions worth following

  • Can CoreWeave cover debt service from operating cash flow?

    If operating cash flow cannot cover interest and near-term maturities, CoreWeave must keep raising external capital. That would increase dilution or refinancing risk and reduce the value of its contracted revenue backlog.

  • Can CoreWeave sustain pricing per megawatt as contracts renew?

    Revenue per megawatt has slipped from $8.48 million in 2025 to $7.92 million annualized in the first half of 2026. If pricing keeps falling while depreciation and lease costs rise, the spread that supports debt service narrows. Renewal pricing shows whether demand still exceeds supply.

Generated by DeepVal AI. Data and models may contain errors or be outdated. For research only; not personalized investment advice.

Go beyond the preview

Continue reading the full report

  1. Key risksWhat could break the investment case
  2. Valuation and margin of safetyWhat the current price assumes, and how much room for error it leaves
  3. SummaryOur overall judgement and what to watch next

Free to sign up · 20 research points · unlocking this report uses 10

20 points cover 2 full reports. No card needed.

Research reflects its generation date. Preview the first two modules.

More on CoreWeave, Inc. Class A Common Stock

CoreWeave, Inc. Class A Common Stock

Business, moat and key questions

Open company page