CoreWeave’s net interest expense was $846 million in the third quarter. A year earlier it was $310 million.
The company reported that figure after the market closed on Nov. 10, alongside revenue of $3.66 billion for the three months ended September, up 168% from a year earlier and ahead of the $3.55 billion analysts had expected. Its adjusted loss was $1.27 a share, narrower than the $1.46 loss forecasters had modeled. The net loss came to $709 million, and the shares rose as much as 15% in after-hours trading.
Those lines, read together, describe the AI rental business on a single page. Demand for graphics processors is strong enough to nearly triple revenue in twelve months. The cost of the debt taken on to buy those processors is rising almost as quickly, up 173% over the same period.
CoreWeave lifted its full-year revenue forecast to $13 billion to $13.3 billion, from a previous range that topped out at $13.2 billion. The increase is close to a rounding error for a company of this size. The second number management attached to it carries more weight: capital spending of $37 billion to $39 billion in 2027, a level the company said it can hold rather than escalate.
That is a statement about financing as much as about construction. A GPU cloud buys chips, racks them in leased buildings and sells the resulting compute on contracts that usually run several years. The chips lose value on a schedule the company sets. The debt comes due on a schedule its lenders set.
When the two drift apart, a growing business can run short of cash while its income statement still looks presentable.
Management said a new financing arrangement had eased concern about the debt. Investors appeared to agree. The 15% move was a reaction to the funding news as much as to the revenue beat, and the two arrived in the same release, which made them hard to tell apart.
CoreWeave is the largest of the so-called neoclouds, a group of operators that rent Nvidia accelerators by the hour to AI developers. Several smaller competitors have listed or are preparing to, and the big cloud providers have built capacity of their own. Price at the margin is set by whichever operator happens to have idle machines that week.
Most of the revenue comes from a small number of customers. Contracts run for years, and some carry take-or-pay terms, which is why lenders have been willing to fund construction against them. The risk is not that those contracts vanish overnight. It is that one gets renegotiated at renewal, or that a customer finishes building its own capacity and lets the agreement lapse.
The depreciation schedule carries a similar kind of risk, quieter and harder to see. CoreWeave writes its chips down over six years, a period some analysts consider generous for hardware that is superseded every generation. A shorter life would push reported profit lower without changing a dollar of cash, which is why the choice attracts attention in a business where the asset is also the collateral.
What the 2027 guidance really says is that management expects the buildout to plateau. Capital spending of $37 billion to $39 billion is enormous in absolute terms and flat relative to the current pace. A company that expected demand to keep compounding would not describe a ceiling. A company that has learned how expensive it is to finance one might.
The demand side of the ledger has been the easier half to explain. Enterprises are running inference on rented capacity because building it themselves takes eighteen months and a supply agreement they cannot get. The labs rent because their own clusters are committed to training. Neither group has many alternatives at the scale CoreWeave sells, which is the argument for the revenue line holding.
The cost side is where the argument gets difficult. Interest is a contractual obligation. Revenue is a contract with a customer who may renegotiate. Depreciation is an accounting choice.
Only one of those three is fixed, and it is not the one that produces the growth.
A quarter in which revenue rose 40% and interest rose 20% would be the first real evidence that the model is maturing. The reverse, another period in which the financing line outruns the top line, would revive the older complaint that this is a bet financed with debt on the price of compute rather than a cloud business.
The company has answered that complaint by pointing at contracted backlog and at utilization. Both are real, and neither is audited the way a bond covenant is. Lenders have accepted the trade, which is the strongest evidence available that the contracts are worth what the company says. Lenders also get paid before shareholders do.
The shares have been volatile since the listing, and the after-hours pop shows how much of the equity case rests on the financing line rather than the operating one. A company growing 168% should not need a debt announcement to move 15%. That it did tells you which number the market was actually afraid of.
The question left open is not whether demand for rented compute survives the next two years. It is who absorbs the loss if the price of that compute falls while the debt stays where it is.
