A $103.7 billion backlog, three customers deep

revenue. Three customers paid 36%, 26%, and 10% of it. No one else cleared 10%. The 10-Q calls them Customer A, B and C and does not name them.

We think the 10-K names the contracts behind the shift. OpenAI committed up to approximately $6.5 billion through May 31, 2031. Meta committed up to approximately $14.2 billion through December 2031.

One giant became three. That is a seesaw with more seats, not a crowd.

The backlog is real. So is the bill.

As of June 30, CoreWeave had $103.7 billion of remaining performance obligations. Six months earlier it was $60.7 billion. About 41% is due to turn into revenue over the next 24 months. And 98% of quarterly revenue came from committed, take-or-pay contracts.

That is the bull case in one line: the buyers signed.

Now the other side. As of June 30, CoreWeave owed $35,551 million of debt principal. In the June quarter it booked $640 million of net interest expense on an operating loss of $49 million. Net loss: $626 million. In the first half it spent $14,874 million on investing.

It keeps borrowing. IIn August it signed a $2.6 billion delayed-draw loan facility "primarily to finance capital expenditures required to perform certain customer contracts," at SOFR plus 5.50%. In September it sold $4.2 billion of 2.875% convertible notes due 2033.

Read the chain. A few buyers sign. CoreWeave borrows against the signatures. It buys GPUs. We think the lenders are the first in line to be paid.

Who pays, who wins

Who pays: a handful of AI labs and hyperscalers. Their contracts are the collateral for everything else.

Who wins today, we think: the lenders and the GPU sellers. They get paid whether or not CoreWeave turns a profit.

CoreWeave wins later, and only if those few buyers keep using what they ordered. The company says so itself: it expects "customer concentration with a limited number of top customers is likely to continue in future years."

"Microsoft won't go broke," you're thinking. True. Default is not the risk. Demand is. Hypothesis: if one of the three buyers slows, the interest bill does not slow with it.

What changes for you

If you rent GPUs from a neocloud, your capacity sits in a building financed for someone else's contract. Ask who the anchor tenant is.

If you sell into AI infrastructure, the real buyer list is shorter than the headlines. Know the three names before you build a sales plan.

If you invest, a backlog is only as strong as the few signatures behind it.

Watch: CoreWeave's next 10-Q. Does Customer A drop below 36%? Does a fourth customer clear 10%?

At CoreWeave, the AI boom is three customers and a stack of loans.