The First GPU Bond
In early 2024, Lambda Labs did something no one had done before: it securitized a rack of chips. Here's what that means, why it matters, and what the structure quietly assumes.
In April 2024, Lambda Labs closed a $500 million debt facility led by Macquarie Group. Lambda buys Nvidia GPUs and rents them to AI companies. It needed more GPUs. To get them, it posted its existing GPUs — and the rental contracts they generate — as collateral against the loan.
Lambda’s CEO called it the “first-of-its-kind” GPU-backed asset-backed securitization. That description is essentially correct. And the fact that it happened at all is genuinely significant: it means the financial system has started treating compute infrastructure as an asset class, not just an operating expense. The plumbing is being built.
But first-of-its-kind structures carry a specific risk that mature markets have learned, sometimes painfully, to address. Two years on the Lambda deal is worth examining closely — not to fault it, but because the assumptions it rests on are the same assumptions the entire GPU finance market is now building on top of.
How securitization is supposed to work
Asset-backed securities are one of the more elegant inventions in modern finance. The basic idea is that you can isolate a pool of income-generating assets — car loans, mortgages, equipment leases — in a special purpose vehicle, then sell investors the rights to those cash flows in tranches. Senior investors get paid first; junior investors absorb losses first; the structure is self-contained enough that even if the originator fails, the assets survive.
What makes the structure work is the independence of the collateral valuation. A pool of auto loans has NADA Blue Book. A pool of mortgages has licensed appraisers and a thick dataset of comparable sales. Before the deal closes, before the rating agencies assign their letters, someone independent has told you what the underlying assets are worth if the borrower defaults and you have to sell them.
That independence is not a formality. It’s the mechanism. It’s what allows investors to price the senior tranche inside 110 basis points over Treasuries rather than treating the whole thing as venture debt. Without it, you don’t have an ABS. You have a secured loan with extra steps.
What the Lambda deal actually relies on
Lambda’s SPV holds two things: physical H100 GPUs and the rental contracts they generate. If Lambda defaults, the theory is that bondholders can step in, continue operating the GPUs or liquidate them, and recover their principal.
For that to work, you need to know two things at the time of the deal:
What are the GPUs worth as hardware? H100 SXM chips were trading between $25,000 and $40,000 per unit in secondary markets in early 2024. But those prices were moving fast — in both directions — and “market price” for a GPU is not like market price for a bushel of wheat. There is no exchange. There is no published settlement price. There is a set of private transactions and a set of listed rental rates on platforms like Vast.ai, and from those you triangulate something that approximates a value.
What are the rental contracts worth as cash flows? This depends on whether GPU rental rates hold. In early 2024, H100 spot rates were elevated. The contracts written against those rates looked valuable. But GPU rental rates are not fixed by any independent body. They are set by the platforms — including Lambda itself — based on supply, demand, and competitive dynamics that shift quickly. When Nvidia accelerates its release cadence, when new supply comes online from competitors, when AI model efficiency improves and demand per training run falls, those rates adjust. The cash flows securing the bonds adjust with them.
None of this is hidden. The rating agencies flagged it. Early GPU ABS deals in 2024 were capped at A+ by S&P specifically because of the short performance history and the collateral uncertainty. The market priced the novelty: investors demanded around 150 basis points over Treasuries for senior exposure — roughly double the spread on comparable data center ABS — as compensation for not knowing what they didn’t know.
The deal closed anyway, and it should have. The capital was needed, the structure was sound, and pricing the uncertainty premium into the spread is exactly how markets are supposed to handle novel collateral. Lambda got its GPUs. The market got a data point.
The irony in the cap table
Fast forward eighteen months. Lambda has grown. It has more customers, more GPUs, and more financing. In February 2025, it closed a $480 million Series D.
One of the investors in that round was Nvidia.
Nvidia makes the chips. Nvidia sells those chips to Lambda. Lambda pledges those chips as collateral against bonds sold to third-party investors. Those investors are now lending money against assets whose market value is determined, in meaningful part, by Nvidia’s pricing decisions, product roadmap, and release cadence.
Nvidia is now also an equity investor in Lambda.
Nobody is doing anything wrong here. Nvidia investing in its largest customers is rational — it’s the same logic behind their investment in CoreWeave, in Perplexity, in a dozen other AI infrastructure companies. Strategic alignment makes obvious business sense. And the Lambda bondholders are sophisticated institutions who can read a cap table.
But the structure creates something worth naming. The party whose decisions most directly affect GPU collateral values — Nvidia, through its pricing and release schedule — has a financial interest in Lambda’s equity value. Lambda’s equity value is, in part, a function of the GPU rental rates that also secure its debt. The information flow between the party that sets hardware prices, the party that sets rental rates, and the party issuing bonds against those rental rates is entirely internal to a network of aligned interests.
This is not a conspiracy. It is a structure. And it is exactly the structure that mature credit markets have spent decades building governance mechanisms to address.
What a mature market builds next
The GPU ABS market is growing fast. One forecast puts issuance at $25 billion by 2028, up from around $8 billion in 2025. As more issuers come to market — not just Lambda and CoreWeave, but potentially hyperscaler financing arms and new entrants — the question of how to value the collateral becomes more consequential, not less.
In mature asset classes, the answer to that question is an independent reference price. Not a price set by the issuer. Not a price set by the hardware manufacturer. A price produced by an observable, reproducible methodology, sourced from market data, published on a schedule, and governed by parties with no stake in any particular deal closing at any particular level.
This infrastructure never gets built by the market on its own — at least not quickly. In aircraft finance, it took years of EETC spreads pricing in an aircraft-specific uncertainty premium before standardized appraisal frameworks compressed that premium and opened the market to a broader investor base. In energy, LMP electricity pricing emerged through regulatory mandate, not voluntary adoption. In rates, the lesson was starker: the absence of independent governance doesn’t just create pricing uncertainty, it creates a vector for capture.
The Lambda deal was historic. It proved that GPU compute can be securitized, that institutional investors will buy the paper, and that the structure can be made to work. That’s not a small thing — it opened a door.
What comes through that door next will depend on whether the market builds the valuation infrastructure proactively, or waits until the first stress event forces the question.
The Penstock publishes independent research on financial infrastructure for emerging technology asset classes.