On October 25, Elon Musk said SpaceX has closed $40 billion of debt for GPUs, and that the 200-plus neoclouds (companies that rent out GPUs) will see “guppy-level death and consolidation. Some will make it through, but not many.” A shakeout in a young market is normal. The better question is who survives, and his own post points to the answer: companies with revenue that doesn't come from AI data centers.
Not shut off, repriced
The strongest part of the post is the contrast inside it: “we're investment grade and have Starlink & launch revenue.” Reuters reports SpaceX's third-quarter revenue rose 161% to $14.58 billion, and its CFO expects fourth-quarter capital spending of $35 billion to $40 billion. A company that size, with revenue that doesn't depend on AI, can borrow for GPUs. Compare that with what we covered last week: CoreWeave's bonds touched a 15% yield with a $104 billion backlog behind them, and its interest costs were close to a quarter of revenue. Nebius fell with the group but was still up more than 100% for the year as of October 16.
That isn't a market that has closed. It's a market discriminating between borrowers. The same debt is available at one price to a company with a second source of revenue and at another price, or not at all, to a company whose only business is renting GPUs.
Consolidation is the normal ending
Young infrastructure markets tend to attract far more entrants than they can support, and they end with a few large survivors and a lot of acquisitions. We can't verify that there are 200 neoclouds, but whatever the real number, expecting a small share of them to make it is reasonable. What decides who is in that share isn't how many GPUs they bought. It's who can keep those GPUs busy and keep refinancing the debt.
What a real cloud looks like
Renting GPUs by the hour is a thin business when token prices keep falling and the hardware ages quickly. The hyperscalers aren't in that position because AI is one of many things their customers buy. They have storage, networking, databases, managed services and a base of customers who already pay them, and that revenue is what funds the buildout. Musk's point about non-AI revenue is the same point from the lender's side.
Neoclouds are moving that way. Nebius's reported acquisition of Inferize was a purchase of software for starting GPU workloads faster, and companies like Modal, which sells software for serverless GPU compute, say they have crossed $1 billion in run rate. Whether a neocloud can grow into a full-stack cloud is open. The hyperscalers took many years and a lot of capital, and they aren't standing still. Our view, which is an opinion and not something we can prove, is that the market is big enough for more than a few winners and is unlikely to end up with a single one. But the survivors will look less like GPU landlords and more like clouds.
What this means, in plain English
expect fewer, bigger GPU rental companies. The ones that last are likely to have other ways to make money, not just renting out chips. That's our view, not a forecast.
What we're watching
Whether the terms of SpaceX's deal become public, since a closed deal at a stated price would test Musk's claim. Whether other neoclouds that can't borrow cheaply add revenue beyond GPU rental, or sell to someone who can. And actual issuance data on AI-related debt, which would show whether credit is tightening for everyone or only for the most leveraged.
Sources: Reuters via Yahoo Finance, SiliconANGLE, Forbes, 24/7 Wall St., Modal.