Aluminum to Silicon

As some of my followers know (hi, all 3 of you), earlier this year my wife bought a horse. At first I was like "wtf". And then the finance nerd in me decided "wow this thing is pretty cool", because it turns out horses are a fairly liquid and tradable asset class (somewhat). And funnily enough, they have their own residual value curve, which is not strictly speaking a "monotonically downward sloping" one. Instead, their intrinsic value goes up from the time they are born, and peaks around 10 years as they mature and gain experience, flatlines for a bit, and then eventually that value finally goes down until they pass away (thoroughbreds and racing horses, of course, exhibit different market dynamics, for obvious reasons).

It turns out, none of this discussion of residual values matters, because just the ANNUAL cost to maintain this damn horse is more than what we paid for it in the first place!

RIP me.

Which leads me back to aircraft engines, lol.

In Part I, I walked through six areas where GPU neoclouds run on the same operating and financing principles as planes, and specifically also drew the distinction between aircraft metal, and aircraft engines. (Btw thank you for all the feedback on Part I, including the folks who pushed back, and to Chamath for a nice nod).

Here's a continuation of that discussion, and where the analogy breaks down, and also some identifiable and specific ways where we can close the gap between this novel asset class and more conventional ones so as to create more liquidity and better financing in this market. This will also address some of the pushback on Part I.

1. Why engines are such good collateral

An airframe follows a fairly standard residual value curve. It loses value with age until it's eventually scrapped and parted out. Engines don't work that way.

People outside of aviation don't usually appreciate that a well-maintained engine barely depreciates. Like a Porsche 911 or a Ferrari 458, aircraft engines generally trade flat for many years.

Most of an engine's value sits in its hot section and life-limited parts (LLPs), which get overhauled or replaced on a set schedule, so every shop visit puts value back in. Appraisers price engines off remaining green time far more than calendar age. Same as the horse, the money's in the upkeep.

The line isn't flat. Between shop visits it saw-tooths: value runs down as green time burns, then jumps back after the overhaul, which is why the reserve mechanics exist. But over 15 or 20 years a popular engine like the CFM56 holds its value far better than the airframe it hangs on, and even a run-out engine has part-out value in its LLPs.

GPUs are the opposite. There's no part you can swap to restore value, and there's no real teardown market for a 4-year-old accelerator.

2. Green time with no shop visit

In engine ABS (the Willis Lease WEST or Shannon Engine Support shelves), you're mostly underwriting green time: the hours and cycles an engine has left before it comes off the wing for a heavy shop visit.

To protect bondholders as green time burns down, the airline pays monthly maintenance reserves (supplemental rent per flight hour or cycle) into a segregated account in the trust. When the engine runs out, the lessor uses that cash to send it to an MRO shop, replace the hot section and LLPs, and reset the collateral close to full-life value.

GPUs burn green time too, but there's no shop visit at the end. A cluster burns through its relevance against NVIDIA's 12 to 18 month release clock (Hopper to Blackwell to Rubin), with each generation cutting cost per token and power per FLOP. You can't send an H100 to a hangar, spend $50,000, and get back a Rubin.

That's a real mismatch between accounting and cash flow. Plenty of operators depreciate GPU servers straight-line over 5 to 6 years (many of the hyperscalers and neoclouds have been extending this, btw), which flatters operating margins, but the take-or-pay contracts behind their SPVs usually run 36 months. When the contract rolls off in month 37, the cluster reprices to whatever the market pays, and that range is wide, depending on market factors and supply/demand. Software matters, too: better inference engines and kernels squeeze more tokens out of the same H100 every year, which a jet engine never does once it leaves the factory.

So residual value doesn't always fall off a cliff, but the variance in outcomes is not great from a Quality Of Earnings perspective, and a lender only cares about the worst of them. Nobody sizes a loan to the bull case. And aircraft lenders have decades of lease rates and recoveries across multiple cycles, while GPU lenders have about three years of data, so the worst case gets a lot of weight.

Since nothing resets the collateral, amortization inside the initial contract has to do the job maintenance reserves do in an engine deal. If the debt doesn't mostly pay down inside that window, your tail is equity risk at a credit spread. So structure 90%+ of principal to amortize out of take-or-pay cash flows within the contract term, haircut any loan-to-cost figure that runs past it, and get step-in rights on the powered shell. By year 4, the interconnect and the rack lease are likely worth more than the chips in them. If there was a way to contain residual value risk, the market would be very happy. More on this later.

3. From engine pools to welded racks

There's one area where the engine analogy used to work in favor of GPUs, and that's modularity.

Remarketing a repossessed wide-body 777 is a headache, since the cabin's built to one airline's spec. Engines don't have that friction. A CFM56 or LEAP doesn't care whether the people in the back are flying Emirates First Class or Ryanair, and engine lessors run global pools, swapping spares across dozens of airlines to keep planes from going AOG (aircraft on ground).

Air-cooled 8-GPU servers like the HGX H100 worked the same way, since they're standard boxes that fit almost any colo rack. Blackwell NVL72 and Rubin end that. Rack-scale systems need 130kW+ densities, facility-level liquid cooling, and proprietary copper backplanes, so the engine is now plumbed into the airframe. For lenders, that kills the engine-pool exit. (Check out this awesome Dwarkesh Jane Street video on Youtube. There is zero modularity here.)

It matters most in a default. In aviation, your backstop is mobility plus Section 1110 of the US Bankruptcy Code: if the airline doesn't cure within 60 days, you send a ferry crew, and the plane can be on lease in Southeast Asia within weeks, at least in a friendly jurisdiction.

Try that with 10,000 liquid-cooled GPUs. Draining, crating, shipping, and re-networking takes months, and in a soft market an idle cluster can lose value fast, so that friction eats a lot of your recovery. So some incremental modularity would be nice. Turns out that might be where we're headed.

4. Who's renting the metal aka Tenant Linked Risk

ABS is mostly about the collateral, but tenant risk still matters, and it cuts very differently in the two markets.

The fun thing about aircraft ABS is that most airlines aren't great credits, and lessors lease to them anyway. A weaker airline pays a higher lease rate and a bigger deposit, reserves cover the engines, and an ABS pool spreads the risk across 15 to 25 lessees. If one defaults, you fly the plane somewhere else.

GPUs work the other way around. Tenant risk (or lack thereof) matters a lot more. A cluster under a Microsoft or Meta take-or-pay is worth roughly the present value of that contract. Lease the same cluster to a VC-backed lab, or to an offtaker fronting for Chinese demand (where export controls or cash repatriation restrictions can kill the deal overnight), and it's worth whatever the spot market pays for aging silicon, and you can't fly it anywhere. Concentration makes it worse, since many GPU SPVs rest on one contract or a few labs whose fortunes move together. And a training cluster only earns first-class rates while all 16,000+ GPUs run as one system; if the fabric degrades, it breaks into standalone nodes that clear at spot.

The folks at Liquid Compute recently published the GPU Offtaker Credit Map, which is awesome (credit to Stanley Lee and the team), and I hope more folks refer to it. Their tiers, condensed:

In aviation, the asset bails out a bad credit. In GPUs, the credit is what makes the asset worth lending against, which is why advance rates and pricing drop off fast as you move down the offtaker stack.

The exception is scale. As GPU deals get bigger and lenders start pooling contracts across dozens of offtakers, single-name risk starts to wash out. It's the difference between a SASB CMBS or CRE CLO deal (single or maybe dozens of assets) vs a conventional RMBS or credit card pool (tens of thousands of borrowers). Most GPU financing today looks like SASB. A pooled, multi-tenant GPU deal with 20+ offtakers across tiers would underwrite a lot more like an aircraft pool, and some of the hedge would come from diversification alone, as long as the offtakers aren't all riding the same funding cycle.

5. Who actually services the metal (the GECAS and Willis playbook)

It's worth double-clicking on how aviation ABS handles servicing and workouts, because it exposes what I think is the biggest gap in GPU financing.

In consumer ABS, a backup servicer (Vervent, SST, Carmel Solutions) sits on hot or warm standby. If the originator blows up, it takes the loan tape and keeps billing borrowers.

For the GECAS (START) or Willis (WEST) deals you'll see servicing carved up across three roles:

In GPU debt, this middleware layer barely exists. Lenders got comfortable without it because the first wave was backed by IG contracts, and you don't need a technical backup servicer when you're really factoring investment-grade corporate credit.

That won't hold as GPU financing moves down-market into multi-tenant clusters and rated ABS. The rating agencies won't hand out investment-grade ratings on that collateral until someone solves servicer continuity risk. It's hard to overstate just how important this is to the maturation of the asset class: even as a Series A start up back in 2020 with a TINY amount of receivables outstanding, when I negotiated Ramp's first debt facility with Goldman Sachs, they required a back-up servicer for our card receivables as a closing condition. It was one of the least negotiated and most obvious provisions in the entire credit agreement.

Ok, I'll stop here for now. In Part III I'll get into the best business in commercial aviation (hint: it isn't flying planes) and what the neocloud version of it looks like. The key is that there's a bunch of things you can do to cap your downside and maintain some protections on your asset's residual value.