On Wednesday morning, before the opening bell, shares of CoreWeave ($CRWV) jumped roughly 11 percent. The company had just said that third-quarter revenue rose 168 percent to $3.7 billion, and it lifted its forecast for 2026 to between $13 billion and $13.3 billion. Its closest rivals moved with it: Nebius gained about 10 percent, while Applied Digital and Iren each added a little over 6 percent.
CoreWeave belongs to a group of companies, often called neoclouds, that buy huge numbers of AI chips, install them in data centers, and rent the computing power to companies that train and run AI models. The business needs enormous sums of money up front, and it earns that money back slowly, one month of rent at a time.
A jump like Wednesday’s usually gets read as a verdict on demand. The story people tell is simple: investors worried that the AI boom was slowing, a company at the center of the boom posted big numbers, and the worry eased. And yet the thing that scared investors this month was never revenue. It was debt, and a forecast doesn’t pay debt.
So what does an 11 percent jump actually settle about a buildout that has been financed but not yet proven?
What the Jump Re-Rates
The results themselves were strong, and it’s fair to say so plainly. As CNBC laid out on Wednesday morning, CoreWeave now expects $1.275 billion to $1.375 billion in adjusted operating income for 2026. Analysts polled by Visible Alpha had expected about $1.07 billion. The revenue forecast also cleared the LSEG consensus of $12.86 billion.
A company whose quarterly revenue grew 168 percent from a year earlier is selling something customers want. Nobody rents that much computing power by accident, and nobody signs up for it twice unless the first batch was useful. On the narrow question of whether the machines are being rented, the quarter gave a clear answer.
The worry that pushed the stock down earlier in the month came from a different market. CNBC noted that spreads on CoreWeave’s credit default swaps had widened past 11 percent, a level that traders generally treat as distress. A credit default swap is insurance against a borrower failing to pay its debts. When that insurance costs more than 11 percent a year, the people selling it are saying the risk of default is serious.
Those two numbers describe the same company in different languages. The revenue line says people are renting the machines. The swap spread says lenders doubt the rent will cover the loans that paid for the machines. Both can be true at once, because a data center is paid for years before it earns back its cost, and the loan comes due on its own schedule.
That’s why the most telling line in the article came from Michael Intrator, the chief executive, rather than from the forecast. He said growth was supported by “multiple sources of funding, including financing backed by customer contracts and around $9 billion in customer prepayments and deposits.” On a slow reading, the sentence describes who pays for the buildout rather than how big it will get.
The Finance Side of the Same Buildout
Any large bet made in advance runs into a problem that Marty Cagan describes in Inspired, his book on building software products. Every business case, he writes, rests on two inputs: how much money the project will make and how much it will cost. At the start, he argues, nobody can know either one with any honesty. “We can’t know how much money we’ll make because that depends entirely on how good the solution turns out to be.”
Cagan wrote that about product teams who fake precision to win a spot on next year’s plan. A data center is the same business case, poured in concrete. The cost side is mostly fixed on the day the chips are ordered. The revenue side depends on whether the companies renting the capacity build AI products good enough to justify the rent, years from now. No one can see that on the day the loan is signed.
Consider the order of events in a single data center. The chips are ordered, the building is leased or built, and the power is contracted, all before the first customer runs a single job. The loan starts charging interest on day one. Rent starts when the tenant shows up, and it continues only as long as the tenant’s own business justifies the bill.
Cagan’s answer for product teams is to stop pretending and find a cheaper way to learn before committing. CoreWeave can’t run a cheap experiment on a data center. What it can do is change who carries the unknown, and the funding mix Intrator described does exactly that.
A prepayment is a customer agreeing to pay before the capacity exists. Financing backed by customer contracts is a loan whose safety rests on those customers’ promises rather than on CoreWeave’s own balance sheet. In both cases, the customer has put money or a signature behind the claim that it will need the machines.
The market has noticed the change. John Blank, chief equity strategist at Zacks, said CoreWeave “has done well to allay concerns by moving away from unsecured bond financing.” An unsecured bond is a bet on the borrower alone, while a contract-backed loan is a bet on the borrower’s customers. Wednesday’s jump rewarded that shift in who stands behind the debt.
When the Worst Case Comes First
Investors facing a long, uncertain bet tend to protect against the worst outcome before they argue about the best one. That’s why the swap spread moved first and loudest this month. Before anyone could settle whether AI demand would last, lenders wanted to know what happens to CoreWeave’s loans if it doesn’t.
Survivors look obvious only afterward, and Ray Dalio shows how rare they are in Principles for Dealing With the Changing World Order. He tabulates the worst investing outcomes in major countries, and the striking thing in his table is who is missing from it. “The US, Canada, and Australia were the only countries that didn’t experience sustained periods of losses.”
Dalio’s table tests the cheerful reading of Wednesday’s jump. A strong year doesn’t keep anyone off that table, because the table measures what came afterward, over stretches long enough for debts to come due in worse conditions. A forecast tells you about the next year. The debt on a data center runs much longer than that.
Blank’s own words follow the same order of thinking. “It might be the bottom for CoreWeave,” he said, adding that the shares are “so beaten up” that there is “more value on the upside than the down.” That’s the language of someone sizing the worst case and finding it smaller than the share price implied, which is a different claim from saying AI demand is safe.
The move in the rest of the group says the same thing. Nebius, Applied Digital, and Iren rose on CoreWeave’s numbers, not their own. Investors are treating the neoclouds as one credit question: if the best known of them can fund itself through its customers instead of through bond buyers, the others probably can too.
Wednesday’s jump doesn’t re-verify the demand case for AI infrastructure. What it rewards is a change in who carries the risk that the demand case is wrong. Prepayments and contract-backed loans move that risk off CoreWeave’s bondholders and onto the balance sheets of the customers who signed. Demand is still the open question; it has simply been handed to the companies most confident that they need the capacity.
Commercial real estate settled this problem long ago. A developer who wants to build a shopping center rarely gets a construction loan on the strength of the plans alone. The lender wants anchor tenants, the department stores and grocers, to sign long leases before the steel goes up. Those signed leases make the loan safe, and they let the developer borrow on the tenants’ credit instead of its own.
CoreWeave’s version lines up part for part. The data center is the mall, and the AI companies renting capacity are the anchor tenants. The prepayments and contract-backed loans are the signed leases, and the bond market is the construction lender deciding whether to fund the steel. A developer’s share price tells you how many leases are signed, not whether shoppers will come, and Wednesday’s 11 percent is a count of signed leases.
That is a well-made piece of financial engineering, and it deserves credit. Moving from unsecured bonds to contract-backed funding is what a careful borrower does when its lenders get nervous. It also means the question of whether all this computing power earns its keep has moved upstream, to the tenants.
From Promises to Spreads
If this financing arrangement lasts, CoreWeave and its rivals could come to look less like technology companies and more like landlords. Investors would probably judge them by the length of their contracts and the credit of their tenants, the way they already judge office towers and the owners of cell towers. A five-year commitment from a well-funded tenant may come to matter more to the stock than any forecast of how fast AI usage grows.
The demand question wouldn’t disappear in that world; it would sit with the tenants. The AI companies that prepaid for capacity would carry the bet that their products justify the rent, and their own lenders and shareholders may well start asking the questions now aimed at CoreWeave. A slow year for AI could show up first in a tenant’s results rather than in a landlord’s swap spread.
The neoclouds themselves would most likely earn a steadier, smaller return, closer to a utility’s than a startup’s. That trade would cost them some of the upside that made the stocks exciting in the first place. It would also make them far harder to kill, which is the property their lenders were asking about all month. It would not be surprising if the debt changed form along with the business. Loans tied to specific contracts, with terms that track each tenant’s credit, could replace the broad bond issues that frightened the market this month. The cost of a neocloud’s next dollar of debt would then depend on who signed the lease behind it. That is a more honest way to fund a building whose revenue no one can know in advance.
The Margin That Decides
Wednesday’s jump settles less than it seems to, and more than the skeptics allow. It doesn’t prove that demand for AI computing will last as long as the loans that fund it. It does show that CoreWeave found a way to borrow against its customers’ commitments, and that investors trust those commitments more than they trusted the company’s own bonds.
What decides the outcome is the gap between what the capacity earns and what the money behind it costs. A forecast can widen that gap on paper for a year, but only paying tenants widen it for the life of the loan. Cagan’s point applies here in its plainest form: no one can know the revenue in advance, so the best defense is to make sure someone else has promised it.
A buildout is only as safe as the leases signed before the concrete sets, and no forecast has ever signed one.
Sources
Cagan, Marty. Inspired.
“CoreWeave stock pops 13% after outlook raise eases AI slowdown concerns.” cnbc.com.
Dalio, Ray. Principles for Dealing With the Changing World Order.
Leave a Reply