Financing
GPU financing: what aircraft finance teaches neoclouds
GPU financing copies aircraft finance's playbook. Here's what 50 years of aviation lending already learned—and how to apply it to your next GPU deal.
LegalBooks Compute Infrastructure Practice
Attorneys licensed in the US (California & Nevada) and Canada (Ontario)
Aug 25, 2026 · 6 min
A GPU loan and an aircraft loan are the same kind of deal. In both, a lender puts up most of the cash for an expensive machine and gets repaid from the money that machine earns, holding the machine itself as collateral. Aviation has run this model for 50 years. GPUs have run it for about three. So the older market is the best place to see, in advance, what tends to go wrong—and how to fix it in the contract.
General information for compute-infrastructure operators, not legal or financial advice.
The short version
- Same deal, faster clock. GPUs lose their value in 2–3 years; planes take 15–25. Almost everything hard about GPU financing comes from that one gap.
- Don't assume the resale value—lock it in. Aviation gets a third party to guarantee it. GPU deals should too.
- You can't repossess a GPU like a plane. It's bolted into a building, so your right to recover it has to be in the contract, not assumed.
- Make the customer pay even when the machine sits idle—and check they can. Who pays the bill matters as much as the hardware.
A GPU loan copies aviation's playbook almost exactly.
Both deals rest on the same idea: borrow against a costly machine, repay from what it earns, and use the machine as security. GPU financing has borrowed aviation's whole toolkit—the ownership structures that ring-fence the asset, the leases, the sale-and-leaseback deals, the long-term customer contracts, and the bundling of those contracts into bonds. A 2026 study of the AI buildout says today's data-center deals "resemble established forms of equipment finance, such as aircraft leasing" (Van Nieuwerburgh, 2026). And the money at stake is enormous: roughly $8.2 trillion USD of AI-infrastructure investment is projected through 2032. When that much capital copies an old playbook, it's worth knowing how the old playbook plays out.
The catch: a GPU loses its value years faster than a jet.
This is the whole game. Airlines write a plane down over 15–25 years to a leftover value of 0–20% (IATA). A top GPU can lose half its value in 2–3 years, because a faster chip arrives and buyers stop wanting the old one (American Compute, 2026).
Roughly how much of its original value the asset still holds:
| After… | Aircraft | Nvidia H100 | Nvidia B200 |
|---|---|---|---|
| 1 year | ~95% | ~60–75% | ~65–80% |
| 2 years | ~90% | ~45–60% | ~55–70% |
| 3 years | ~85% | ~35–50% | ~40–55% |
Value still held after 3 years
Aircraft ██████████████████████████████████ ~85%
B200 ████████████████░░ ~40–55%
H100 ██████████████░░ ~35–50%Aircraft figure is an illustrative straight line; GPU figures are mid-range estimates from a young market.
A GPU can keep running for up to about eight years, yet most GPU loans are written over roughly three (American Compute, 2026). Lenders match the loan to how long the chip stays valuable, not how long it works.
Three lessons aviation already paid for.
1. Don't assume the resale value—get someone to guarantee it. Aircraft lending lives or dies on what the plane will be worth years later. Aviation's fix is a residual value guarantee: a promise—from the customer or an outside party—to buy the machine, or top up its price, at a set amount on a future date. That way the lender still gets repaid even if the resale price has collapsed. GPU deals are starting to copy this; Meta's large data-center financing used exactly such a guarantee to make lenders comfortable (Van Nieuwerburgh, 2026). The catch for GPUs: that guarantee is only as good as the number behind it, and guessing what a chip will be worth in three years takes real technical judgment.
2. You can't repossess a GPU the way you repossess a plane. If an airline stops paying, the lender takes the plane back and leases it to someone else—it's built to move. A GPU cluster is bolted into one building's power and cooling. If the borrower fails, physically pulling the chips out and running them elsewhere is slow, expensive, and sometimes not worth it. Aviation could count on the collateral walking away; GPU lenders can't. So the right to get in, take control, and clearly own the chips has to be written into the contract from day one—you can't assume you'll grab the hardware later.
3. Make sure the customer keeps paying—even when the machine earns nothing. Aircraft leases usually require the airline to keep paying every month even if the plane is grounded and flying no one (Structured Finance Association). That's why lenders study the airline's finances as hard as the plane's. GPU deals need the same two things: a contract where the customer pays even if the chips sit idle, and a customer solid enough to actually pay. As one analyst put it, a data-center bond backed by Amazon is safe because Amazon earns money everywhere; the same bond backed by a smaller operator is only as safe as that operator's own customers (Junk Bond Investor, 2026).
Where GPUs are riskier than planes.
The structures copy cleanly. The assumptions underneath them don't—and that's what to stress-test before signing.
| Aircraft | GPUs | |
|---|---|---|
| Track record | 50+ years of default and resale data | ~3 years; never tested in a downturn |
| How fast it loses value | Slow, predictable | Fast; a new chip can gut the old one's price |
| Resale market | Deep, with trusted appraisers | Young; a used cluster is hard to price (CipherTalk) |
| Getting collateral back | The asset flies away | Bolted into a building; hard to recover |
| Hidden debt | Well understood | Large lease commitments often sit off balance sheets—Moody's counts 970BUSD * *, with * *660B USD off the books (Van Nieuwerburgh, 2026) |
Frequently asked questions
Why is GPU financing compared to aircraft financing?
Both are the same kind of deal: a lender puts up most of the cash for an expensive machine and gets repaid from the money that machine earns, holding the machine as collateral. GPU deals have copied aviation's structures almost exactly. The key difference is speed: a plane holds its value for 15 to 25 years, while a top GPU can lose half its value in two to three years.
How long do GPUs last, and what term should a GPU loan use?
A GPU can keep running for up to about eight years, but most GPU financing is written over roughly three years (with a 24 to 60 month range). The short term is not about the chip breaking—it is about the next chip making the old one uneconomic, so lenders match the loan to how long the GPU stays valuable, not how long it works.
Can you repossess GPUs the way you repossess an aircraft?
Not easily. An aircraft is built to move, so a lender can take it back and lease it to someone else. A GPU cluster is bolted into one building's power and cooling, so physically recovering and redeploying the chips is slow and costly. Because you cannot count on grabbing the collateral later, GPU contracts need to spell out access rights, control rights, and clear ownership of the chips up front.
Is there a resale market for used GPUs?
There is one, but it is young. Used GPUs and clusters do change hands, yet pricing is volatile and there is no trusted appraisal system like aviation's. That makes a used GPU cluster hard to value, which in turn makes it harder to lend against with confidence.
What this means for your next deal
The structures are borrowed from aviation and mostly solved. The hard part is the judgment underneath them: how fast the chips will age, whether the building can actually deliver the power and cooling, and what the collateral is really worth if a deal goes sideways. Those are engineering questions that decide legal terms.
That's how LegalBooks is built. We're lawyers licensed in the US and Canada who are also data-center engineers, so we read a GPU financing the way an engineer and a lawyer would read it together. Our platform has helped negotiate $266M+ USD in compute-infrastructure contracts across 160+ companies, with drafts and redlines back within six hours, guaranteed. Financing a buildout, or reviewing a lease or customer contract? Book a call with counsel and we'll pressure-test the resale assumptions, the payment terms, and your right to recover the hardware—before you sign.
General information, not legal or financial advice. Figures come from the linked sources and reflect a fast-moving market; confirm current terms before relying on them.