The AI datacenter buildout is being financed with debt, and debt reprices faster than earnings report. If the buildout cracks, credit spreads on the GPU cloud companies should show it before anything shows up in a quarterly. So I spent some time working out which of them to actually watch, and the answer comes from the liability structures. They are not interchangeable.
CoreWeave is the most exposed. When I pulled the numbers in June, it was sitting on more than $20 billion of debt against roughly negative $22.6 billion net cash, funded about five parts debt to one part equity, with a current ratio around 0.5 and a funding gap I put near $17 billion. It is the most debt-dependent name in the group, which means its spreads move first. Just as important, its debt is watchable: straight senior secured term loans plus a liquid credit default swap. That CDS already demonstrated what it does under stress, running from 371 to 773 basis points between September and December 2025 before settling back around 450.
CoreWeave’s own borrowing history tracks the whole buildout’s credit cycle. Its first facility in 2023 priced at 15%. By July 2025 it was borrowing at SOFR plus 4. By 2026 it had a facility at SOFR plus 2.25, investment-grade-style pricing for the same borrower the market wanted 15% from three years earlier. Watching that number reverse would tell you the cycle had turned, without waiting for an analyst to say so.
Applied Digital is the second name to watch. Its debt is straight senior secured notes, $2.15 billion at 9.25% from March 2026 and $1.59 billion at 7% from June, plus private notes on top. High coupons and no conversion feature mean the bond prices are nearly pure credit signal. It is also a landlord to CoreWeave with an Oracle-anchored campus, so its credit and CoreWeave’s are concentration-linked: trouble at one shows up in the other.
Nebius looks like the same trade, but its balance sheet is very different. It raised $4.3 billion in convertibles, but against roughly $4 billion of debt it holds over $9 billion of cash, with a current ratio above 3, and much of its funding came as equity and Nvidia pre-funded warrants rather than leverage. The cash cushion means its credit cracks late, and the convertible structure means its spreads embed equity volatility, so the signal is noisy anyway. It would confirm a break rather than warn of one.
Then there are the ones you cannot watch at all. Crusoe and Lambda are private, and their debt is private too: bank facilities from Brookfield, Goldman, and a syndicate, in Lambda’s case a $1 billion senior secured loan. Private facilities do not trade and do not print anywhere, so there is no continuous price on their credit. The only signal they emit is the press release, and a press release about a facility being priced or upsized is by construction good news. The most leveraged parts of the buildout are the ones you hear from last, and only when they choose to say something.
So the watch order follows from that. CoreWeave and Applied Digital first, because their debt is straight, traded, and large relative to everything else on their balance sheets. Nebius as confirmation. The private companies not at all, until after the fact. None of that required a view on AI demand. It is just a matter of who owes what, in which instrument, and whether that instrument has a price.