CoreWeave has the order book every AI-infrastructure company wants: a $103.7B backlog at the end of June 2026, revenue up 112% to $2.58B in the quarter, and a customer list that now runs beyond Microsoft to Meta and OpenAI. The demand is not the question. The question is who pays for the machines before the revenue arrives. In the second quarter CoreWeave spent $6.4B on capital equipment and generated $0.68B of cash from operations, so its own business funded about a tenth of its build. The rest came from debt of about $35B, convertible notes and share sales, against equity of $5.0B on $77.1B of assets. Net interest was about a quarter of revenue, and the adjusted operating margin was 5%. Its newer loan facilities are rated below investment grade, and on 4 November its 2031 bonds traded as low as 79.5 cents on the dollar after elections that put Texas datacentre construction in doubt. Guidance for 2026 capex is $35–39B. The call is BEARISH on that structure, not on AI demand: a company whose assets depreciate in a few years, funded largely with debt and new shares, has to keep refinancing on good terms to stay ahead of its own build, and the terms are moving against it. Conviction is MEDIUM because the backlog is real and the customers are large: it may refinance its way through for longer than the numbers suggest. BEARISH, conviction MEDIUM, one year. A bearish call is a stated view, never an instruction to short.
Nebius: sold out and prepaid, but the build has to earn its keep
Watching: the neocloud whose customers prepay its build, spending $20–25B on about $3B of revenue.
Q2 2026 revenue $582M (+454%), adjusted EBITDA ~40%.
Customers had prepaid $6.0B; cash $8.0B against $8.5B of debt.
ai-infrastructuredatacentersneocloud