CoreWeave’s third quarter looked great on the top line and painful underneath it. The neocloud, which rents out the Nvidia GPU capacity that much of the AI industry runs on, reported $3.66 billion in revenue, up 168% from a year earlier. It also posted a $709 million net loss and paid $846 million in interest in a single quarter.
That second set of numbers is why the more interesting part of Tuesday’s earnings call was not the growth. It was CoreWeave explaining, in some detail, how it plans to stop leaning on expensive debt. The answer is a mix of customer prepayments, loans tied to specific contracts, and short term deals priced far above what long contracts pay. It is a clever plan. It also only works if AI compute stays scarce.
The short version
- Q3 revenue of $3.66B, above the $3.4B to $3.6B CoreWeave guided for in August
- Net loss of $709M and $846M of interest expense
- Revenue backlog now stands at $132.5B
- About $9B of customer prepayments and deposits expected in the next three months
- Short term capacity sells for more than $40M per megawatt a year, versus $20M to $25M on four and five year deals
The quarter in numbers
| Metric | Q3 2026 | Context |
|---|---|---|
| Revenue | $3.66B | Up 168% year over year, from roughly $1.37B |
| Net loss | $709M | About 19 cents lost per dollar of revenue |
| Interest expense | $846M | About 23% of revenue |
| Revenue backlog | $132.5B | Up from about $104B at the end of June |
| Expected prepayments | ~$9B | Over the next three months |
| 2026 capex guidance | $35B to $39B | As raised in August |
Look at the interest line next to the net loss. CoreWeave’s interest bill was larger than its entire loss for the quarter. Put differently, before paying its lenders, the business is roughly in the black. That is the core of the neocloud model: the GPUs earn money, but they are bought with borrowed cash long before customers start paying for them, and the cost of that borrowing eats the profit.
“We will not use the unsecured debt market”
CEO Michael Intrator was blunt about it. “We will not use the unsecured debt market to fund our business at current yields, nor do we need to,” he told analysts. That is a direct response to what bond investors have been saying. Last month, yields on some CoreWeave bonds climbed above 15%, a level usually associated with companies that lenders think might struggle to pay them back.
Instead of issuing more of those bonds, CoreWeave is leaning on three other levers.
1. Make customers pay first
CoreWeave expects roughly $9 billion in customer prepayments and deposits over the next three months. More than $3 billion of that is tied to new contracts for capacity that is already built, which means the cash comes in without CoreWeave buying a single new server. Rivals Nebius and IREN have done something similar, relying on large prepayments from Microsoft to fund their buildouts. The industry is quietly shifting the financing burden back onto the hyperscalers and AI labs that want the compute.
2. Borrow against specific contracts
Rather than general purpose bonds, CoreWeave funds most of its buildout with delayed draw term loans, known as DDTLs, that are secured by particular customer contracts. In August it closed a $2.6 billion facility called DDTL 5.5, its first designed specifically for shorter term contracts. It can draw that money in installments as each project needs it, so it is not paying interest on cash sitting idle.
3. Sell short and charge more
This is the bold part. CoreWeave is setting aside more capacity for short term customers who will pay a premium for compute right now.
“We are also allocating more capacity to short term opportunities, where the pricing justifies it, selling capacity at more than $40 million per megawatt annually, while pricing for new infrastructure on four and five year contracts is moving to $20 million to $25 million per megawatt.”
Michael Intrator, CoreWeave CEO
Prices are already moving. CoreWeave raised prices across its offerings by about 25% in July and by roughly another 10% later in the quarter. Intrator said prices are still rising this quarter in a way that “more than offset[s] cost increases.” Compounded, those two hikes alone add up to about 37.5% in a single quarter, which tells you a lot about how tight the market for GPU capacity still is.
The trade off lenders do not love
Short contracts are great when demand is hot. They let CoreWeave reprice often and capture whatever the market will bear. The catch is that lenders care far more about predictability than about peak pricing. A five year contract at $22 million per megawatt is easy to lend against. A six month deal at $40 million is worth more today but could renew at a much lower price, or not renew at all.
| Strategy | Upside | Risk |
|---|---|---|
| Customer prepayments | Cash now, less borrowing | Locks in today’s price if compute gets pricier |
| Contract backed loans | Avoids 15%+ bond yields | Ties each loan to one customer’s health |
| Short term deals | Roughly 1.8x the price per megawatt | Renewal and repricing risk if demand cools |
| Price hikes | Margins improve on existing hardware | Customers may shift to rivals or their own chips |
Prepayments carry their own twist. They reduce CoreWeave’s immediate need to borrow, but they also commit it to deliver compute at a fixed price. If demand keeps surging and spot prices keep climbing, CoreWeave will have sold some of its capacity too cheaply.
Why this matters beyond CoreWeave
CoreWeave is a useful barometer for the whole AI buildout because its business is so exposed. It does not have a giant advertising or software business to fall back on the way Google or Microsoft do. Its fortunes ride on whether companies keep renting GPUs at rising prices. That is also why its stock and Nvidia’s tend to swing together, as investors saw when AI stocks sold off around Election Day.
The constraints are not only financial. Getting power to new sites is slow enough that some operators are improvising, with data centers running retired jet engines while they wait for the grid. Every month a site sits waiting for power is a month of interest payments with no revenue behind them, which makes the cheaper financing CoreWeave is chasing even more important.
The bottom line: CoreWeave has found ways to grow without selling more expensive bonds, and Wall Street’s equity investors liked what they heard. But nearly every lever it is pulling assumes AI compute stays scarce and pricey. If that assumption holds, the strategy looks smart. If demand cools, short contracts and fixed prepayments are exactly where the pain would show up first.
For now, the $132.5 billion backlog says customers are still lining up. The next few quarters will show whether they are willing to keep paying up front, and at ever higher prices, to stay in that line.
About this article: GeekBlog covers U.S. technology news, AI, phones, smartwatches and gaming. Every story is written and checked under our Editorial Policy. Spotted a mistake or have a story tip? Contact our editors.
