GPU Cloud = Financial Product

The shortage of Nvidia GPUs did not end when more chips started shipping. It changed form. At CoreWeave, one of the biggest specialist AI cloud providers, the bottleneck is now balance-sheet capacity. On March 31, the company said it had closed an $8.5 billion delayed-draw loan backed by GPU infrastructure and a customer contract, which it described as the first investment-grade financing of its kind for high-performance computing assets (coreweave.com). That is the important shift. GPUs are no longer just scarce hardware. They are collateral. That changes how AI capacity gets built. A normal cloud expansion story is about demand forecasts, datacenter leases, and server orders. CoreWeave’s new facility turns those same ingredients into something closer to project finance. The company can draw about $7.5 billion at first, then step up to the full $8.5 billion as the underlying assets are installed and stabilized, with the debt maturing in March 2032 (coreweave.com). Bloomberg reported that the structure is backed by a Meta contract worth more than $19 billion, which is why lenders were willing to treat racks of AI servers as financeable infrastructure instead of speculative tech inventory (bloomberg.com). Once that happens, GPU supply stops being only a manufacturing problem. It becomes a question of who can package chips, contracts, and future cash flows into securities that large lenders want to own. CoreWeave’s loan carried ratings of A3 from Moody’s and A from DBRS, and the company said the deal was oversubscribed, with MUFG and Morgan Stanley leading and Blackstone Credit & Insurance anchoring the facility (coreweave.com). That is not what a hardware shortage looks like. It is what an asset class looks like. The reason this matters is that CoreWeave is already carrying enormous leverage. Bloomberg reported that the company had $21.6 billion of debt at the end of 2025, plus another $3.7 billion of untapped borrowing capacity (bloomberg.com). Its own 2025 results show the cost of that strategy in plain numbers: $5.13 billion in revenue, $1.23 billion in net interest expense, a $1.17 billion net loss, and a revenue backlog of $66.8 billion that management is using to justify more buildout (coreweave.com). Fast growth is real here. So is the financing burden. That burden is easier to ignore when revenue is exploding. CoreWeave’s 2024 revenue was $1.92 billion, up 737% year over year, but 62% of it came from Microsoft, according to the company’s IPO disclosures and CNBC’s reporting on the March 2025 listing (cnbc.com, cnbc.com). That concentration risk did not disappear when Meta and OpenAI added huge commitments. It just moved one layer down. If a few giant customers underpin both revenue and the loans used to buy the machines, then customer concentration becomes funding concentration. This is why the debt structure matters more than the chip count. If GPUs are financed against long-term contracts, then the price and availability of compute start to reflect credit conditions as much as semiconductor output. The winning providers will not just be the ones that can get Nvidia supply. They will be the ones that can convince banks and insurers that a rack of Blackwell servers and a hyperscaler contract deserve investment-grade treatment. CoreWeave says it has secured about $28 billion of debt and equity commitments in the past 12 months (coreweave.com). The AI cloud is starting to look less like a datacenter business and more like structured finance with very hot chips inside.