CoreWeave, Inc.

CoreWeave, Inc. (NASDAQ: CRWV) is a specialized AI hyperscaler — a 'neocloud' — that buys NVIDIA GPUs at enormous scale, racks them in data centers, and rents that compute to AI labs and enterprises as a managed cloud service. Founded in 2017 as Atlas Mining, an Ethereum-mining operation, the company pivoted to GPU cloud in 2019 and rode the generative-AI boom to one of the fastest revenue ramps in infrastructure history, IPO-ing in March 2025. Revenue scaled from roughly $1.9B in 2024 to ~$5.1B in 2025, and Q1 2026 revenue more than doubled again to $2.08B.

The defining structural fact is that CoreWeave is asset-light on power. Unlike vertically integrated peers (Nebius) or ex-Bitcoin miners that own their substations and land (IREN, Cipher), CoreWeave does not own its electrons. It leases powered, built-out data-center capacity — from third-party landlords like Core Scientific, Galaxy/Helios, Chase/Lancium, Switch and others — and brings the GPUs. Its disclosed power footprint is on the order of ~3.5 GW of contracted/leased capacity across dozens of sites, but verified owned power is effectively zero: nothing online, contracted, or announced as owned. Sourcing is classified as Leased.

The economic model is therefore a spread business at massive scale: borrow against GPUs and signed customer contracts, lease the power and shells, buy chips, and capture the gap between contracted compute revenue and the stacked cost of debt, depreciation, and lease payments. That makes CoreWeave the purest expression of the neocloud trade — closest to compute, furthest from power — and the single most contract-rich, capital-intensive name in the AI-infrastructure complex, with a ~$99.4B revenue backlog (RPO) against ~$24.9B of debt as of Q1 2026.

The bull case is that CoreWeave is the default merchant supplier of frontier AI compute, and being asset-light on power is a feature, not a bug. By leasing capacity rather than building it, CoreWeave avoids multi-year grid-interconnection queues and concrete-pouring timelines, and instead spends its capital on the scarce, revenue-generating asset — GPUs — turning them on faster than anyone. The result is a ~$99.4B backlog (RPO) anchored by the highest-credit counterparties in technology: Microsoft (~67% of FY2025 revenue), OpenAI (up to ~$22.4B through May 2031), Meta (up to ~$14.2B through 2031), and a multi-year Anthropic agreement signed in April 2026. Adjusted EBITDA of $1.16B at a 56% margin shows the core unit economics work; bulls argue that as contracts ramp and the GPU fleet seasons, operating leverage and a maturing debt stack convert that backlog into durable cash flow.

The bear case starts with the same asset-light structure: CoreWeave rents its power and owns none of it. Every megawatt it sells sits on someone else's land, behind someone else's interconnect, under a long-dated lease — so it captures the thinnest slice of the value chain (the GPU spread) while bearing the fattest cost (the GPUs themselves, financed with ~$24.9B of debt). It has no power arbitrage, no land bank, no merchant-generation optionality, and limited control over its most important input: electricity cost and availability. The Core Scientific episode crystallized the vulnerability — CoreWeave tried to buy its largest landlord in a ~$9B all-stock deal to internalize ~1.3 GW and erase ~$10B of future lease overhead, but Core Scientific's shareholders rejected the merger in October 2025, leaving CoreWeave a 12-year tenant rather than an owner.

The other half of the bear case is financial fragility under that structure. Despite positive EBITDA, CoreWeave posted a $740M net loss in Q1 2026, with ~$536M of quarterly net interest expense and $7.7B of single-quarter capex funded almost entirely by debt secured against GPUs and individual customer contracts. Revenue is dangerously concentrated (Microsoft ~67%), counterparty credit (OpenAI's ability to pay through 2031) is unproven, and the model is circular: NVIDIA is supplier, investor, and customer. At a ~$52-55B market cap on a leased power base with ~0 owned GW, the equity is priced for flawless conversion of contracted compute into paid, profitable revenue — with none of the hard-asset floor that owning the power would provide.

The purest neocloud in the complex: CoreWeave owns the GPUs and sells compute as a service, but owns ~0 GW of power — it leases ~3.5 GW of built-out capacity from third-party landlords. Maximal exposure to compute, minimal control over electrons, which is why it sits at the far compute end of the spectrum.