Analysts had modeled Microsoft’s revenue for weeks. The number that moved other companies’ shares was capital spending, which came in about $5.5 billion below expectations.
Within an hour of Microsoft’s disclosure, CoreWeave and Nebius were each down nearly 5% in after-hours trading. IREN fell about 3%. None of those companies reported anything. They fell because a customer said it would spend less than expected on the thing they sell.
The link is direct. The new cloud providers, the ones the industry calls neoclouds, exist largely to hold capacity that the largest technology companies need but do not want to build themselves. Nebius signed a contract with Microsoft late last year worth close to $20 billion and running into the 2030s. CoreWeave counts Microsoft among its largest customers. When Microsoft trims a capex plan, the effect lands on the suppliers first, because a supplier’s revenue is a customer’s budget line.
Investors read the miss as a delay rather than a cancellation, which is the charitable interpretation and also the one the numbers support. Microsoft has been leasing capacity faster than it can energize it, and a company that cannot get substations built does not need to sign for more racks. The less charitable reading is that the largest buyer of AI compute is pausing to see whether the demand it has already contracted for turns into revenue.
Either way, the reaction says something about how these companies are financed. CoreWeave listed on Nasdaq in March 2025 at $40 a share, raising $1.5 billion, and has borrowed heavily against GPUs ever since. Nebius came out of the wreckage of Yandex’s international assets and rebuilt itself as a European AI cloud. Both are spending ahead of revenue, and both are paying for it in credit markets rather than out of cash flow.
That is the structure that makes a $5.5 billion revision matter. A contract for $20 billion of compute is an asset. The debt raised to build the data center that serves it is a fixed obligation with a floating rate attached. If delivery slips by two quarters, the revenue slips by two quarters and the interest does not slip at all.
Elon Musk put the bear case in its bluntest form over the weekend, writing on X that more than 200 neocloud companies would fail. His argument was about access to capital: the debt market has effectively closed to companies whose only revenue comes from AI data centers, which leaves them dependent on the equity market at exactly the moment investors have started asking about returns.
Gartner gave CNBC a version of the same forecast with a longer horizon. The research firm expects 80% of the neoclouds operating today to be acquired or shut down by 2030. That is not a prediction about demand for computing. It is a prediction about who survives the financing.
The distinction matters, because the two outcomes look identical for a quarter and diverge afterwards. Demand for AI capacity is not in question at Microsoft, which is still spending vastly more than it did two years ago and told investors the shortfall was about timing and mix. What is in question is how many balance sheets can carry the gap between signing a contract and collecting on it.
The collateral behind much of that borrowing is the hardware itself. A lender will advance against GPUs because the GPUs can be seized and leased to someone else, but that trade only works while the equipment is current. Each new generation resets what the previous one is worth, and a company that financed a 2025 build with 2027 assumptions discovers the gap on its balance sheet rather than in a press release.
The history of infrastructure buildouts offers a rough map. Telecom carriers in the late 1990s signed long-term capacity deals with equipment makers, financed the build with debt, and discovered within three years that the contracts were worth less than the fiber. The survivors were the ones with other sources of revenue, which in this cycle means the hyperscalers and the chip makers, not the companies renting out racks.
There is a counterargument worth taking seriously. Neocloud contracts are longer and more specific than the telecom deals were, often take-or-pay, often with prepayment. A customer that has committed to a decade of capacity cannot walk away cheaply. That gives lenders something to underwrite, and it is why CoreWeave and Nebius can still borrow at all.
What it does not give them is a cushion on execution. Building a data center on schedule requires power, cooling, networking and chips to arrive in the same quarter. Any one of those slipping pushes revenue into the next period while the depreciation clock starts on the equipment already installed.
The next set of answers arrives with the companies’ own earnings, when CoreWeave and Nebius report and are asked, in effect, how much of their contracted backlog Microsoft has already told them to slow down. Microsoft, for its part, will be asked whether the $5.5 billion is a deferral or a decision. For the companies that fell on the news, the difference between those two words is worth more than any forecast they could publish.
