Neoclouds like CoreWeave borrow billions against chips that lose value fast, and the financing structures are unlike anything in traditional credit
SC STUDIO / Getty Images
In August 2023, a company most people had never heard of borrowed $2.3 billion using Nvidia $NVDA chips as collateral. As Reuters reported at the time, CoreWeave raised the debt facility, led by Magnetar Capital and Blackstone, with the unusual use of Nvidia H100 processors as security for the loans. The deal marked the first time H100-based hardware had been used as collateral.
Less than three years later, CoreWeave's total debt exceeds $21 billion, a notable increase from under $8 billion in 2024. The financing model it pioneered has become the template for a class of companies known as neoclouds, and the embedded assumptions in these debt structures are now a focal point for credit analysts trying to understand where risk sits in the broader AI debt buildout. Here's what to know.
GPU-collateralized debt is, at its core, asset-backed lending. A borrower pledges physical hardware (GPU servers and related infrastructure) plus the revenue streams from customer contracts as security for a loan. The mechanic resembles how airlines finance aircraft or how utilities fund power plants. But the assets in question have no established track record at this scale, and the technology cycle that governs their value moves far faster than in aviation or energy.
CoreWeave has executed this model through a series of delayed draw term loan facilities, each housed in a separate special purpose vehicle. The borrower in CoreWeave's most recent $8.5 billion facility is a bankruptcy-remote SPV established to own the GPUs and serve as the borrower for the loan. The facility is secured by substantially all assets of CoreWeave Compute Acquisition Co. VIII, LLC, a subsidiary, not the parent company itself. The facility allows multiple draws until June 2027 and matures in March 2032, with borrowings at SOFR plus 2.25% for floating-rate loans or about 5.9% for a fixed-rate tranche, and the facility is secured by substantially all assets of the borrower group.
The SPV structure is designed to ring-fence the debt. The SPV pledges GPUs as collateral for the loans, along with services contracts with customers. The terms of the credit agreement require CoreWeave to have contracts with large and creditworthy companies that cover future debt repayments. In this way, what makes the debt "investment-grade" is not the neocloud's own credit profile, which carries a speculative-grade rating, but the creditworthiness of the customer on the other side. CoreWeave's $8.5 billion DDTL 4.0 Facility received ratings of A3 by Moody's $MCO and A (low) by DBRS, representing what the company called the first investment-grade rated financing secured by high-performance computing infrastructure and an associated customer contract. Bloomberg reported that the loan was backed by CoreWeave's deal with Meta $META, valued at $14.2 billion.
Two assumptions are embedded in every GPU-collateralized loan: that the hardware retains enough value over the loan's life to justify the collateral, and that utilization rates remain high enough to generate the cash flows needed to service the debt. Both are contested. On residual value, the data is thin.
The current AI boom is only three years old, so estimating depreciation remains a challenge. Major hyperscalers have set six-year useful life assumptions for their server infrastructure, but Amazon $AMZN shortened the useful life for a subset of its servers and networking equipment from six years back to five years, effective January 2025, citing the "increased pace of technology development, particularly in the area of artificial intelligence."
Meanwhile, Nvidia has released new data center GPU architectures about every two years: Ampere in 2020, Hopper in 2022, and Blackwell in 2024. CEO Jensen Huang has declared the company is now on "a one-year rhythm" for new chips, which compresses the window in which any given GPU generation holds its competitive edge.
CoreWeave's own GPU rental rates have fallen 50-70% from their peaks, according to analysis by Level-Headed Investing. On utilization, the math depends on customer contracts holding. DBRS noted that the $8.5 billion facility has "highly stable cash flows predicated on the take-or-pay contract with Meta with no exposure to volume risk," and the contract with Meta requires payment whether or not Meta uses the capacity. That provides insulation. But it also means the credit quality of the structure is only as strong as the counterparty's willingness and ability to keep paying.
When a company like Amazon or Meta borrows in the investment-grade bond market, lenders are underwriting the entirety of a massive, diversified cash-flow stream. Independent analyst Dave Friedman has noted that "the bear case isn't 'AI is a bubble' or 'the technology doesn't work.' The bear case is narrower and more technical: The financing structures being used to build the infrastructure embed assumptions about residual values, utilization rates, and counterparty stability that are not well-supported by the data we have."
GPU-collateralized debt operates on different plumbing. The lender's recovery in a default depends on seizing and selling or re-leasing specific hardware whose value is governed by a technology cycle unlike anything in traditional asset-backed finance. Friedman has described CoreWeave as "a leveraged infrastructure vehicle that finances GPUs like power plants, collateralizes them like aircraft, and backstops them with customer prepayments that behave like short-term loans."
CoreWeave's debt-to-equity ratio exceeds 4.8x, according to research from MLQ.ai. More recent data suggests the figure is higher still: total debt-to-equity sits near 8.9, with a current ratio of about 0.5, according to StocksToTrade. The company is paying about 25% of revenue just to cover interest on its debt load, the Motley Fool has reported. Customer concentration adds another layer. Microsoft $MSFT was 62% of CoreWeave's 2024 revenue. The company has since diversified somewhat, signing contracts with Meta, OpenAI, and Anthropic.
But the SPV model means each facility is tied to a specific customer contract, so the credit story for each tranche rests on a single counterparty.
CoreWeave is not the only company using these structures. Oracle $ORCL, Meta, xAI, and CoreWeave have moved roughly $120 billion of AI infrastructure debt off their balance sheets using SPVs funded by Wall Street, according to Cryptopolitan. The model is spreading because it works for both sides: tech companies get capital without dilution or balance-sheet bloat, and lenders get assets they can point to plus contracted cash flows.
The question credit analysts are asking is what happens when these instruments exist at scale and the technology cycle turns. S&P Global $SPGI Ratings analysts wrote in a December 2025 report that "CoreWeave is rapidly evolving its complex debt capital structure as it seeks to finance its significant capital expenditure needs." That complexity is itself a risk factor. The structures are new. The secondary market for seized GPU collateral is thin. And the pace at which Nvidia releases new architectures means today's collateral could be two generations old by the time a five-year loan matures.
Friedman put it this way to Quartz: "Some of this paper ends up in broad-based corporate bond funds that sit in target-date retirement products. The retail holder has no idea why they own it, no framework for evaluating it, and no ability to exit it independently of the fund."
Whether that amounts to a systemic concern or a manageable niche risk depends on how large these structures grow, and how long the AI demand cycle holds.
Join 500,000+ readers who start their day with Quartz.
By subscribing, you agree to our Terms of Service and Privacy Policy.
