Diligence Watchlist A live database of the situations we're tracking across public and private markets — ahead of client demand. Each entry sets out the company, why it's interesting, the questions we'd test, and the experts we'd put on it. Built for the hedge funds and private equity firms that move first.
CoreWeave has built a business that turns the AI compute shortage into a credit product: it borrows, buys Nvidia GPUs, signs multi-year take-or-pay contracts with the largest AI buyers in the world, and rents the compute back out at a margin. The market reads the contracted backlog — now near $99bn — as durable, bankable collateral. We think that reading rests on three assumptions no filing can confirm, and all three are answerable only through primary research with the people around the business. Here's how we'd test them.
The situation
The structure is elegant on paper. Contracted revenue from creditworthy customers backs billions in debt, some of it now investment-grade rated, which funds the next wave of GPUs, which carry the next round of contracts. During a shortage, the spread is enormous and the model compounds.
The fragility sits in what the structure quietly assumes. A frontier GPU fleet is not ordinary infrastructure throwing off predictable cash for a decade. It is a generation-specific asset depreciating against the fastest hardware cycle in modern computing, financed over a horizon far longer than the window in which it holds its economic edge. The contracts amortise on one clock; the hardware loses relevance on another, faster one.
The diligence thesis
The central question is simple to state and hard to answer from public data: the market is pricing customer quality as if it were asset quality. A contract with a strong customer tells you revenue is likely to arrive for the term of that contract. It tells you nothing about what the hardware is worth when the term ends, whether the customer renews rather than building its own capacity, or whether the data centres even come online on schedule.
Three things have to hold for the thesis to work. The hardware has to retain enough value that a two- or three-year-old cluster still justifies its capital recovery. The largest customers — who are also the companies most capable of building their own silicon — have to renew rather than insource. And the buildout has to deliver, against real constraints in power and construction that don't show up until the plan meets the ground. The financials describe the symptom. The cause lives in the operations.
The primary research questions that matter
- Rental economics by vintage. What are GPU rental rates actually doing, split by hardware generation, contract tenor, and customer tier — intensifying scarcity, or quiet compression? The headline numbers can't resolve it because the answer depends entirely on mix.
- Economic useful life under load. How long does a cluster run at high utilisation before it's displaced — not when it stops working, but when it stops being worth the power and rack space it occupies.
- Renewal versus insourcing. For the anchor customers, what does owned compute cost against renewing rented capacity, and where does build-versus-buy tip toward build?
- Buildout and power reality. Where are the genuine bottlenecks — interconnection queues, power availability, construction timelines — and how do delivery milestones compare to the schedule the revenue assumes?
- Secondary-market clearing price. What would a large used-GPU fleet actually clear at if it had to sell, and how deep is the buyer base — an orderly market, or a forced one?
Why this can't be answered from the filings
Backlog, depreciation schedules and contract terms tell you what the company has reported. They don't tell you whether the hardware holds its value, whether the customers stay, or whether the data centres come online on time. Those answers live with the people who transact with the business, compete with it, supply it, or used to run it. That is what primary research is for: expert calls, surveys and channel checks that turn a balance-sheet question into a set of first-hand answers from the operators who actually know.
The experts we've recruited
- Former neocloud infrastructure operators — engineers and fleet-operations leaders who have run dense GPU estates at 90%-plus utilisation at CoreWeave-comparable operators, and can speak first-hand to degradation, refresh cycles, and real unit economics.
- Ex-hyperscaler infrastructure and energy leaders — former data-centre and energy executives from Google, Microsoft and Meta-scale operators, now independent or in advisory seats, who understand the build-versus-buy calculus and the power constraint from the customer side.
- GPU brokers and secondary-market dealers — founders and principals at used-hardware marketplaces who transact H100 and H200 inventory weekly, and can put a real clearing price on residual value and tell us how deep the buyer base actually runs.
- AI-lab and startup compute buyers — heads of compute and infrastructure-procurement leaders at AI-native firms who rent capacity from neoclouds, and can speak to provider selection, cost per token, and how concentration and renewal decisions are actually made.
- Data-centre developers and power partners — executives at the colocation and power-developer firms that build and energise these facilities, who can speak to interconnection timelines, build slippage, and where projects realistically fall behind.
- Competitor neocloud operators — leaders at rival GPU-cloud providers who price against CoreWeave in live deals and can read its competitive position from the outside.
Commission this research
Woozle Research runs done-for-you primary research for hedge funds and private equity firms — expert calls, surveys and channel checks, delivered as finished intelligence rather than raw notes. For CoreWeave, the experts are recruited and the questionnaire is ready; we can move fast and have findings back in your hands inside the window that matters. Commission us to run this primary research for you — or build your own variant of the thesis — and request a quote.