GPU Securitization
Definition
GPU securitization is the financing of AI compute hardware through a bankruptcy-remote special-purpose vehicle that owns the GPUs and the contracts that rent them out, and issues notes repaid from those rental cash flows and, at the end, from selling or re-leasing the hardware. It is equipment asset-backed finance applied to a new asset: the same architecture used for aircraft, railcars and shipping containers, with two differences that dominate the credit analysis. GPUs lose value far faster than any of those assets, on the order of a third or more per year as new generations launch, and the rental contracts behind them are shorter and more concentrated. The notes are sized to the contracted cash flows, not to the hardware's headline price.
Why it matters
The AI build-out is being financed with debt secured by hardware that may be worth a third less in twelve months, and securitization is the mechanism by which that debt reaches investors who never see a data center. Whether it is safe depends on one design choice: are the notes repaid from contracted rent inside the contract term, or do they need the GPUs to be sold for a good price afterwards? The first is a receivables deal with strong counterparties; the second is a bet on residual value in a market with no downturn history. For anyone whose private credit fund, tokenized compute product or marketplace note is ultimately backed by GPUs, that question, and the identity of the renters, is the whole analysis.
Common misconceptions
- •The collateral is not really the GPUs. It is the rental contracts. Structures are sized to contracted cash flow; the hardware is what you are left with if the contracts fail, and its value on that day is the number every rating analysis stresses.
- •GPU depreciation is not like equipment depreciation. Aircraft and railcars lose value with age and use; GPUs lose value when a faster chip ships. The curve is a step function tied to product cycles of roughly one to two years, which is why residual assumptions are the contested input.
- •A securitization does not make the operator's risk disappear. The operator usually services the assets, may guarantee performance, and often holds the junior notes. Servicer failure and concentration in one operator's contracts remain in the structure.
- •The CoreWeave facilities were not public securitizations. They were large delayed-draw loan facilities secured by GPUs and contracts; they established that institutional lenders would finance GPUs at scale, which is what later securitization structures build on.
Technical details
The structure
Originator: the GPU operator buys hardware and signs rental or capacity contracts with customers (AI labs, hyperscalers, enterprises).
Sale to the SPV: the GPUs, the contracts and the receivables are transferred to a bankruptcy-remote special-purpose vehicle in a true sale, so that the operator's insolvency does not reach them.
Notes: the SPV issues senior notes and often one or more junior classes; the operator or a sponsor retains the equity. Senior notes are sized so that contracted cash flow covers debt service with a cushion after operating costs and servicing fees.
Servicing: the operator continues to run the hardware and bill customers as servicer, paid a fee; a backup servicer is named in case it fails, which for GPUs means another operator able to take over the facility and the customer relationships.
Cash flow waterfall: rental collections pay data center and power costs (if the SPV bears them), servicing fees, senior note interest and principal, reserve account top-ups, junior notes, then residual to the equity. Triggers on collections, utilization or counterparty defaults switch the waterfall to full turbo amortization.
Sizing: cash flow first, hardware second
Contracted cash flow: the base case sums the remaining payments under signed contracts, net of costs, and discounts them; senior notes are typically sized so that contracted cash flow alone amortizes them, with uncontracted capacity treated as upside or ignored.
Advance rate against hardware: the resulting note balance is also expressed as a percentage of the GPUs' appraised value. Reported ranges for current-generation hardware under multi-year contracts run about 50 to 70 percent; lower for older generations or uncontracted inventory.
Overcollateralization and reserves: a cash reserve of several months of debt service and a target overcollateralization level give the notes room for delayed collections and a slow re-leasing period.
Amortization profile: the strongest structures amortize the senior notes to zero before the contracts expire, so residual value is a cushion rather than a source of repayment. Structures with a balloon at the end of the contract term are the ones exposed to the resale market.
The rating and credit questions
Counterparty credit: who pays the rent, for how long, and can they terminate. A three-year take-or-pay contract with a large, well-capitalized customer supports far more debt than a rolling monthly arrangement with a start-up.
Concentration: most large contracts come from a small number of AI companies and hyperscalers. Analyses stress the loss of the largest one or two counterparties and the time and cost to re-lease their capacity.
Residual value: stresses assume the hardware is worth a fraction of appraised value at the end, with haircuts scaled to the product cycle. Realized secondary prices for the prior generation after a launch are the calibration point.
Operating dependencies: power contracts, colocation leases and networking are outside the SPV but essential to the cash flow; the analysis checks that the SPV or its servicer controls them for the life of the notes.
Servicer and backup: whether another operator could step in, and what it would cost, determines how much of the contracted cash flow survives a servicer failure.
Worked example
An operator places 4,000 current-generation GPUs, appraised at $30,000 each ($120 million), into an SPV with three-year capacity contracts paying $2.20 per GPU-hour at 85 percent contracted utilization: about $65 million of contracted revenue a year, roughly $45 million after power, colocation and servicing.
Senior notes of $75 million at 9 percent amortize over three years: debt service about $29 million a year, covered 1.55 times by net contracted cash flow. Loan-to-value at issuance 62.5 percent. A $7 million cash reserve covers about three months of debt service.
Year 2, a new generation launches and the GPUs' resale value falls to $15,000. The notes have amortized to about $50 million against $60 million of hardware; loan-to-value is 83 percent, but the contracts are unchanged and the notes continue to amortize from rent. The structure holds because it never depended on resale.
Alternative structure with a $40 million balloon at year three: the same price drop leaves the balloon dependent on selling hardware worth perhaps $10,000 each by then, $40 million against a $40 million balloon with no cushion. Same assets, same contracts, a different note profile, and one of the two is a residual-value bet.
How it compares with a GPU-backed loan and where the exposure reaches investors
GPU-backed loan: on the operator's balance sheet, secured by GPUs and contracts, with covenants on the operator (coverage, utilization, leverage), a borrowing base that moves with appraisals, and lender remedies against the operator. Fast to arrange, flexible to amend, and the model of the CoreWeave facilities of $2.3 billion in 2023 and $7.5 billion in 2024.
Securitization: assets isolated in an SPV, repaid from the assets alone, notes placed with a wider set of investors, more rigid once closed. Suits operators with a book of long, strong contracts who want to term out financing at a lower cost than a corporate loan.
Where individuals meet it: private credit funds and BDCs that buy the notes or the loans; marketplace platforms that sell participations in GPU-backed loans; tokenized compute products that pass through rental cash flows. In every case the diligence questions are the ones above: who pays the rent, for how long, and does repayment depend on selling the hardware.
Frequently asked questions
- What is GPU securitization?
- A structure in which GPUs and the contracts that rent them are sold to a special-purpose vehicle, which issues notes to investors. The notes are repaid from the rental payments and from the hardware's value when the contracts end. It differs from a GPU-backed loan, where a lender advances against the GPUs on the operator's own balance sheet.
- What are GPU-backed securities?
- The notes issued by a GPU securitization vehicle. Senior notes are sized so that contracted rental cash flows cover interest and principal with a cushion; junior notes and the operator's retained equity absorb losses first. They are a form of equipment ABS, and where they are rated, the ratings turn on the quality of the rental counterparties and on assumed residual values.
- How is GPU securitization different from a GPU-backed loan?
- A GPU-backed loan is a bilateral or club facility to the operator, secured by the GPUs, with financial covenants on the operator. A securitization isolates the GPUs and contracts in an SPV, removes them from the operator's credit, and repays notes from the assets alone; the operator typically stays on as servicer. The loan is faster and more flexible; the securitization can reach a wider investor base and, for strong contracts, a lower cost.
- What advance rate do GPU securitizations use?
- Note issuance is usually sized to contracted cash flows rather than to a percentage of hardware value, but the resulting loan-to-value has been reported in the 50 to 70 percent range for current-generation hardware backed by multi-year contracts with creditworthy counterparties, with lower advance rates for uncontracted capacity. Overcollateralization and cash reserves add to the cushion.
- What is the biggest risk in a GPU securitization?
- Residual value. GPUs depreciate on a technology cycle, not a wear cycle: a new generation can cut the resale value of the prior one by a third or more within a year. A structure that relies on selling the hardware at the end of the contracts is exposed to that; one that amortizes the notes fully within the contract term is not. Counterparty concentration is second: a handful of AI labs and hyperscalers sign most large contracts.
- Who has used GPU-backed financing at scale?
- CoreWeave's 2023 facility of $2.3 billion and its 2024 facility of $7.5 billion, arranged by Blackstone and Magnetar among others, were secured by GPUs and customer contracts and were widely reported as the first large institutional GPU-backed financings; they were structured as delayed-draw loan facilities rather than public securitizations. Market reporting since then describes securitization-style structures being arranged for other operators; terms vary and are mostly private.
- Can an individual investor buy GPU-backed securities?
- Not directly in most cases: the notes are sold to institutional investors in private placements. Individuals meet the exposure through funds that hold them, through platforms that sell participations in GPU-backed loans, and through tokenized compute products, each of which adds its own layer of terms. The reference pages on GPU rental economics and GPU collateralization cover those routes.