CoreWeave: The Economics Are Starting To Change

CoreWeave is realizing roughly $40 million of annualized revenue per MW, signaling strong scarcity pricing, while newer contracts are delivering improved contribution margins. Its expansion beyond GPU rentals is gaining traction, with non-GPU products above $400 million in ARR and managed inference exceeding $100 million. However, leverage and investment needs remain substantial risks: debt stands at $35.6 billion and 2026 capex is projected at $35 billion to $39 billion.

Analysis

CRWV's equity is becoming a leveraged claim on sustained AI-compute scarcity rather than a conventional software multiple story. Incremental margin improvement matters more than top-line growth because it can validate that new capacity is being contracted at returns above the firm's unusually high funding cost; however, the revenue-per-MW figure is not directly comparable with data-center peers without utilization, contract duration, power pass-throughs, and GPU-generation mix. The near-term upside catalyst is evidence that newer deployments retain pricing as supply expands, which could drive a sharp re-rating in both equity and unsecured credit over the next 1-3 months.

The key second-order issue is financing capacity. A capex program roughly equal to the existing debt stack creates refinancing and dilution sensitivity if GPU lease rates normalize, hyperscaler demand concentrates, or debt markets demand a wider spread. Better inference and non-GPU mix may diversify revenue, but it also introduces execution risk: these offerings must demonstrate materially higher gross profit dollars, not merely ARR, before they offset capital intensity. Beneficiaries of a sustained scarcity regime include NVIDIA (NVDA), power-dense data-center infrastructure providers such as VRT and ETN, and potentially GPU-cloud peers; the losers are smaller, uncontracted GPU lessors whose financing costs rise before their utilization catches up.

Contrarian view: the market may be underestimating the value of contracted capacity but overestimating the duration of scarcity. If large cloud platforms internalize more compute or next-generation GPU supply materially improves availability, CRWV faces simultaneous pressure on renewal pricing and asset residual values—the adverse combination for a debt-funded lessor. Falsify the constructive case if management shows declining revenue per deployed MW, weaker contribution margins on newer cohorts, customer concentration increases, or credit spreads widen despite reported ARR growth.

Key Decisions for Investors

• Maintain CRWV as a tactical long only after quarterly disclosure confirms stable-to-rising revenue per deployed MW and expanding contribution margin; target a 1-3 month re-rating, but size as high-volatility equity given financing dependence. Exit on a material reduction in 2026 capex funding visibility or a guidance cut tied to utilization/pricing.
• Prefer a relative-value expression: long NVDA or VRT versus short CRWV in equal beta-adjusted notional for investors seeking AI-compute exposure with lower balance-sheet risk. The trade should work over 6-12 months if infrastructure demand persists but GPU rental economics normalize; risk is CRWV securing long-duration contracts at unexpectedly attractive economics.
• Monitor CRWV debt and credit-default-swap spreads as the leading signal rather than relying on ARR headlines. A sustained spread widening concurrent with equity strength is an alert to reduce longs, as the capital structure requires continued debt-market access to execute the buildout.
• Do not underwrite a standalone long on the non-GPU and inference businesses until management provides retention, gross-margin, customer-concentration, and capex-per-dollar-of-ARR metrics. Positive disclosure on those measures would be a catalyst for a higher-quality revenue mix and could justify adding exposure.