CoreWeaveThin moat

CoreWeave rents out artificial intelligence's raw material. It buys Nvidia's most advanced accelerators by the tens of thousands, wires them into clusters engineered for the single job of training and running large models, and sells that capacity by contract to the companies building frontier AI. It is a cloud provider in the way a specialist hospital is a hospital: far narrower than Amazon or Microsoft, and considerably better at the one thing it does.

The revenue arithmetic is unlike anything else in this collection. Fiscal 2025 revenue was $5.13 billion, up 168%¹; trailing revenue is about $6.23 billion — grown from roughly $16 million in 2022. There is no segment chart on the following pages because there is nothing to split: one product, sold to a very short list of customers, of whom Microsoft alone was about 67% of 2025 revenue.

Where the money goes is the more revealing half. CoreWeave has never earned a profit — a net loss of about $1.2 billion in 2025, widening to roughly $1.9 billion on a trailing basis — because the model consumes capital at a rate the income statement cannot yet absorb: roughly $1.15 billion a quarter of depreciation on the GPUs and about $536 million a quarter of interest on the debt that bought them, against 2026 capital spending guided at $31-35 billion². Nvidia is simultaneously a supplier, an investor holding roughly 11%, and a backer of CoreWeave's own customers — a circularity the company does not dispute.

What the market is buying, at roughly $48 billion of value and about 6.4 times sales with no earnings to price³, is the order book: a contracted revenue backlog of $103.7 billion at June 2026, up from $60.7 billion six months earlier. That backlog is the entire bull case, and converting it requires building more capacity, faster, with borrowed money.

The pages that follow separate the two questions the price conflates. The Moat asks what CoreWeave actually owns — specialized scale, the Nvidia relationship, the contracted backlog, the capital machine — and rates it thin, because nearly all of it is rented. The Future Bets follow the attempts to change that: buying its landlord, building five more gigawatts, signing every major AI lab, and putting a software layer above the machines.