AI Cloud Firms Face Debt Risks Backed by Nvidia GPUs

Article Summary

• Emerging AI cloud companies are using GPUs as collateral for substantial loans.
• Jim Chanos warns of the risks associated with this debt due to a lack of clear profitability.
• Rapid obsolescence of GPUs could increase the risk of debt defaults.

Mounting Risks in GPU-Backed AI Debt

Emerging AI cloud computing companies, seeking to challenge tech giants in the artificial intelligence market, have secured substantial loans collateralized by graphics processing units (GPUs) from a $5 trillion chip manufacturer. These firms are using the funds to purchase more AI chips to scale their operations. They lease data centers, fill them with AI hardware, and rent out the computing power to tech companies for training and running AI models. According to a report by The Information in July, four companies in this growing sector—three of which have backing from Nvidia—have amassed over $20 billion in debt backed by Nvidia AI chips (GPUs). Jim Chanos, famed for accurately predicting the Enron collapse, sees red flags in this surge of GPU-backed debt, as the "neoclouds" taking on these loans lack a clear path to profitability, making debt repayment challenging. "The business model of neoclouds, and many AI companies themselves…are currently loss-making enterprises," Chanos stated in a recent interview. "You're simply hoping that changes, otherwise, these debts will face default." This financing trend was pioneered by CoreWeave and is a form of asset-backed lending: loss-making companies can leverage their assets as collateral to obtain funding they wouldn't otherwise qualify for, typically at higher interest rates. Because these so-called neoclouds are much smaller than the tech giants they're trying to challenge, GPU-backed financing offers an efficient means of rapid expansion. Nvidia has invested heavily in this space, aiming to expand its customer base beyond the tech behemoths. According to The Information, Nvidia-backed CoreWeave, which went public earlier this year, and its competitor Fluidstack, each have approximately $10 billion in debt collateralized by their GPU reserves. Lambda and Crusoe, also supported by Nvidia, have taken on $500 million and $425 million in GPU-backed debt, respectively. CoreWeave reported losses of roughly $65 million in 2024, while Fluidstack reported losses of under $1 million for the same year. Lambda and Crusoe's financials are not publicly available. Another significant risk in AI chip-backed debt is a topic of heated debate in the tech investing community: Nvidia’s GPUs may depreciate, or lose value, faster than companies are accounting for. Most cloud providers (including CoreWeave) expect their AI chips to generate revenue for approximately six years. Amazon reduced the depreciation lifespan of its AI chips—their effective economic life—from six years to five years in January. But Chanos argues that AI chips could depreciate even faster as Nvidia releases new AI products roughly every 18 months. First, this would make it harder for companies to generate enough revenue from the chips to repay debts. Second, the market value of the chips could fall below the value of the loans they secure, increasing the risk of default. Chanos points out that even assuming a long lifespan for the chips, the neoclouds are not currently profitable: "Many of the business models we look at, some of the new data centers and neoclouds like CoreWeave and Nebius, don't make money even on an 8 or 10-year depreciation," he said. "The capital costs are still too high, higher than the rate of return on AI data center deals." "If these things (Nvidia's chips) only have a three-year economic life…then the whole economic rationale of many of these deals completely collapses," Chanos said.