CoreWeave's leverage strain

CoreWeave grew so fast that the usual cloud story no longer fits. The company started as a specialist renter of Nvidia GPUs, then rode the generative AI boom into hyperscale territory. In 2024 it posted $1.9 billion in revenue, up from $229 million in 2023, and in 2025 it reached $5.1 billion. That is real growth. It is also the kind of growth that can hide a more basic fact: CoreWeave has had to spend and borrow at a pace that would be alarming in almost any other corner of tech (sec.gov, sec.gov). The reason is simple. CoreWeave is not selling software with near-zero marginal cost. It is trying to assemble one of the world’s largest AI compute fleets before its rivals do. That means buying or financing enormous numbers of GPUs, securing power, and fitting out data centers for dense, liquid-cooled clusters. By early 2026 the company was advertising 40-plus data centers, gigawatts of contracted power, and ultra-low-latency fiber. In 2024 it had told investors Microsoft was its largest customer and that customer concentration was a material risk. Hypergrowth did not remove that fragility. It made the machine bigger (coreweave.com, sec.gov). That is why the balance sheet matters more than the revenue chart. CoreWeave’s 2025 results showed $5.177 billion in operating expenses on $5.131 billion in revenue, leaving it with an operating loss even before the debt burden did its work. Net interest expense for 2025 hit $1.229 billion. Net loss was $1.167 billion. The company’s own annual report says total indebtedness had reached $21.6 billion by the end of 2025, with another $3.7 billion of untapped borrowing capacity. This is not a company leisurely funding expansion out of profits. It is a company racing the clock with lenders alongside it (investors.coreweave.com, sec.gov, bloomberg.com). The financing structure shows how far that race has gone. In May 2025, just weeks after its IPO, CoreWeave sold $2 billion of senior unsecured notes with a 9.25% coupon. That was expensive money. By March 2026 it had moved to something more elaborate: an $8.5 billion GPU-backed loan that Bloomberg described as secured not just by chips but by a customer contract, with Meta-linked offtake at the center of the package. In other words, lenders were not simply betting that GPUs hold value. They were betting those GPUs already had a buyer on the other side (sec.gov, bloomberg.com). That matters because CoreWeave’s product is not generic cloud capacity. The pitch is tightly packed, latency-sensitive AI infrastructure. The company says its cloud is built for large GPU clusters, private high-performance connections, and production inference where delays and data movement matter. But that kind of capacity is hard to move around if the financing behind it assumes specific utilization and specific counterparties. A warehouse full of GPUs is one thing. A debt stack underwritten by named customer demand is another. The more specialized the fleet becomes, the less comforting the old idea of “hard asset backing” looks on its own (coreweave.com, bloomberg.com). CoreWeave does have one powerful defense against all of this: demand has been strong enough to keep pulling the company forward. In February 2026 it reported a revenue backlog of $66.8 billion, more than four times where it began 2025. That is the number bulls reach for, and not without reason. But backlog is not cash in the bank, and it does not make the capital intensity disappear. The company’s infrastructure page still promises mega-clusters of more than 100,000 GPUs, while its 2025 financials show the cost of building that promise in plain view: billions of dollars of expense, over a billion dollars of interest, and a business that can grow at breathtaking speed while still depending on ever more intricate forms of debt to keep the racks filling up (investors.coreweave.com, coreweave.com, sec.gov).