NVIDIA releases new graphics processing unit (GPU) architectures almost every year, but companies purchasing these GPUs depreciate them over 5 to 6 years. This discrepancy raises concerns about the risks associated with GPU-backed loans, as the actual value of the collateralized chips may decline faster than their book value.
According to the IT industry on September 28, a generation of GPUs is said to maintain market competitiveness for only 2 to 3 years. Financial researcher Sasha Steffen analyzed that the residual value of the H100 after three years is estimated to be between 50% and 70%, but critics argue it could drop by more than 70% during that period. Michael Burry, a real-life figure from the film 'The Big Short,' claims that hyperscalers will understate depreciation by $176 billion from 2026 to 2028, promoting a short-selling narrative.
These concerns are amplified by the rapid growth of GPU-backed loans. AI-focused cloud companies, which have less financial power than big tech firms, are increasingly using GPUs as collateral or employing multiple guarantees and asset-backed financing techniques to secure funding for data center investments.
The structure is as follows: Neocloud establishes a special purpose vehicle (SPV) to transfer GPUs and customer contracts, which the SPV then uses to borrow money. The principal and interest are repaid through fees paid by customers. Even if the parent company goes bankrupt, the collateral remains tied to the SPV, allowing creditors to recover their investments. Loans are primarily disbursed through a delayed draw term loan (DDTL) structure, aligned with the schedule for purchasing and installing GPUs.
The trend began with CoreWeave in the United States. In August 2023, CoreWeave secured $2.3 billion using NVIDIA's H100 as collateral, marking the first instance of H100-based hardware being used for loan collateral. Since then, CoreWeave's debt has ballooned to approximately $35 billion by the end of the second quarter of this year. The company reported that it has raised over $30 billion in debt and equity this year alone. As debt has increased, its second-quarter interest expenses also surged to $640 million, more than double the $267 million reported a year earlier.
Interest rates reveal that the safety net for these loans is not the GPUs themselves, but the customers. CoreWeave's initial GPU-backed loan interest rate was around 15%, but a March loan of $8.5 billion, based on a contract with Meta, received an investment-grade rating, reducing the fixed interest rate to about 5.9%. In contrast, loans secured by contracts with non-investment-grade clients in May and August had interest rates that added 4.50 and 5.50 percentage points to the SOFR, respectively. This structure indicates that customer creditworthiness determines interest rates; if customers falter, the remaining collateral is only the depreciating GPUs.
According to telecommunications media Capacity, the loan-to-value (LTV) ratio for GPU-backed loans is around 60% to 70%. This means that if chip values drop by more than 30% to 40%, the collateral could fall below the loan principal.
The revenue structure is also tight. GPU cloud provider Acer analyzed that the depreciation costs of the top two publicly traded Neocloud companies are about half of their revenues, and none of the major players are reporting profits under U.S. Generally Accepted Accounting Principles (GAAP).
Given this situation, concerns about an AI-driven 'subprime' crisis are emerging. U.S. law firm Quinn Emanuel has analyzed that a decline in GPU values could lead to a chain reaction of defaults by Neocloud, failures of data center SPVs, and impairments of asset-backed securities (ABS). The firm stated, 'This risk is fundamentally a valuation battle,' predicting that disputes over the actual value of the collateral will be most intense among creditors, borrowers, and investors.
The structure where chip suppliers support customer debts is also seen as a risk factor. Semiconductor analysis firm SemiAnalysis noted that most Neocloud companies would struggle to secure sufficient loans without direct leasing agreements with hyperscalers, stating, 'NVIDIA is acting as a central bank.'
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.