Dek: CoreWeave grew 168% and still lost $709 million in a quarter. How it pays for the next data centre has quietly changed, and that change tells you who is carrying the risk in AI capex now.
CoreWeave prepayments: the customer pays before the GPU arrives
CoreWeave reported third-quarter results on November 10, and the financing detail is more interesting than the headline number. The company now asks customers for money up front, borrows against individual contracts instead of its own name, and sells more of its capacity on short terms.
CEO Michael Intrator said on the call that CoreWeave won't use the unsecured debt market at current yields, and added that it doesn't need to. That's a striking sentence from a company that has raised billions in debt.
The replacement is customer cash. CoreWeave expects about $9 billion in prepayments and deposits over the next three months, and more than $3 billion of that arrives with new contracts for capacity that's already built. No new hardware required. The customer signs, the money lands, the revenue starts.
The CoreWeave loss sits next to 168% growth
The quarter itself was strange in the way AI infrastructure quarters tend to be. Revenue hit $3.66 billion, up 168% year on year. The net loss was $709 million.
Interest expense for the quarter was $846 million, which is larger than the loss.
Read that again.
The company spent more servicing debt than it lost on operations, and the operations are growing at a rate most software companies never see.
Growth and debt in the same sentence isn't automatically bad. A business building data centres has to spend before it earns, and debt is how that gap gets bridged. The question is what happens to the bridge when the interest bill outruns the profit for another four quarters.
CoreWeave contracts got shorter and far more expensive
Here's the part that changes how you read the whole story. CoreWeave raised prices about 25% in July and another 10% later in the quarter, and said prices were still climbing.
On short-term deals, capacity is now selling for more than $40 million per megawatt per year. Four and five year contracts are moving at $20 million to $25 million per megawatt. That's roughly half the short-term price. Same rack, same power draw, different signature on the page.
From CoreWeave's side that's obvious arithmetic. Short deals capture today's scarcity, and scarcity is real. From the customer's side it's a different deal: you pay a premium for the right to leave.
How GPU cloud financing moved off the bond market
The third piece is the borrowing itself. In August the company closed a $2.6 billion facility, DDTL 5.5, built for shorter-term customer contracts and drawn in installments as those contracts come in.
That structure exists because lenders look at different things than equity investors do. A bank lending against a five-year contract with a creditworthy customer can model the cash flows, while a three-month deal at a high price leaves the lender guessing about where the market sits next quarter.
So CoreWeave built a facility that matches the contracts it's signing. Draw as you go, against paper the lender can read.
The bond market gave its own verdict last month, when some CoreWeave bonds traded above 15%. That's a yield that says the lenders want to be paid for doubt, whatever the equity story is doing.
The renewal risk inside CoreWeave contracts
Put the three pieces together and a trade-off appears.
Prepayments fund construction with somebody else's money, which is cheap and fast to arrange. Short contracts let you charge double for the same megawatt. Both help cash flow right now.
Both also put the risk somewhere specific. If demand cools in two years, the customer on a nine-month deal walks, and the capacity goes back on the market at whatever the market pays then. The four-year contract at half the price would have looked better in that world.
Lenders understand this, which is why long contracts carry more weight in a credit review than short ones do. A company shifting toward short deals is trading balance-sheet safety for present income. That's a legitimate choice in a boom. It's the wrong one if the boom ends mid-contract.
CoreWeave still signs four and five year deals, and the $132.5 billion backlog covers both shapes. The mix is what moved, not the model.
What the AI capex shift means if you rent GPUs
If you buy compute rather than sell it, three things follow from this quarter, and none of them are abstract.
Prices are up, and the increases came twice in one quarter. Should you take the short deal or the long one? That depends on how well you can forecast your own demand, and most teams can't forecast it twelve months out.
Prepayment is becoming normal, which means your finance team needs to fund compute before it shows up on an invoice. And short terms carry a premium. That's only worth paying if you genuinely expect your needs to change.
The same pattern shows up elsewhere. Nebius and IREN have both leaned on large Microsoft prepayments to fund capacity, which suggests the prepayment model isn't a CoreWeave quirk. It's how the second tier of GPU clouds stays solvent while it builds.
That's the actual state of AI capex. The money is still flowing. It's just arriving earlier, from customers, against narrower contracts, and at prices that assume the shortage lasts.
Watch the renewal rate when the short deals start expiring. That number will say more about this business than any quarter of 168% growth.
