Aluminum to Silicon

A lot of ink has been spilled in recent days on GPU financing. I wanted to go back to a familiar framework to examine just how this market works. For me, I don't think it's software, the better comp is commercial aviation (specifically aircraft engine ABS). This is going to be a multi part series of blog posts.

Earlier in my career I spent a bunch of years trading ABS, including resi mortgages, solar, credit card loans, student loans, and also aircraft and aircraft securitizations. During my time on Wall St. I bought and sold billions of dollars of structured paper, and then, as an operator (including as Head of Finance at Ramp) structured over $2 billion of asset-backed facilities and other forms of debt financing. Suffice it to say, I did a lot of deals, and many similarities are exceptionally notable.

Over the past two years, private credit has put billions into CoreWeave, Lambda, Crusoe, and a long tail of smaller GPU clouds. I went back to the aircraft framework to see how well it holds up.

A GPU neocloud has basically the same operating model as an airline. Both borrow against their equipment to buy hardware from a concentrated set of manufacturers, park it in scarce physical infrastructure, and rent it out by the hour before the amortization schedule catches up with them.

Here's six specific areas where I think the businesses line up:

1. Inventory that expires

An airline sells Available Seat Miles. Once the door closes on a flight, an empty seat is revenue that will never be booked. There is no way to store Tuesday's unsold seat and sell it on Friday.

GPU capacity works the same way. Once an H100 or B200 cluster is energized in a colo, an idle hour can't be sold later, and the debt service accrues whether the rack is busy or not.

2. Fixed costs dominate, marginal costs round to zero

People ask why airlines don't just raise fares enough to stop going bankrupt every decade, and the answer is mostly in the cost structure.

Almost everything about operating a flight is fixed: the lease or EETC payment, crew, gate fees, and the fuel needed to get the airframe to cruising altitude. Those costs are the same with 12 passengers on board or 200. One more passenger costs a few dollars of fuel and a can of soda. Once the flight is on the schedule, a $65 fare beats an empty seat priced at $1,000.

Neoclouds face a similar dynamic:

• Fixed costs are an 8x GPU server ($300,000+), a 10-year lease on 60kW+ of liquid-cooled rack space, and 10% to 14% debt service. Together these run well over 85% of the cost base, and an idle cluster still draws meaningful baseline power and cooling.
• The marginal cost of taking a GPU from 0% to 100% utilization is just the incremental electricity, roughly $0.05 to $0.07 per kWh, which works out to pennies per GPU-hour.

With a cost base like that, one metric dominates: load factor for an airline, cluster utilization for a neocloud. Airlines generally need load factors around 80% to break even. A debt-financed GPU cluster needs roughly 65% to 75% utilization to cover debt service, hardware amortization, colocation, and power. Below that line the operator is burning equity, and nearly all of the return on invested capital is earned in the band between breakeven and full utilization (basically between 75% and 100% utilization).

3. Squeezed between the OEM and the landlord

Airlines are stuck between the companies that build the planes and the airports that control the gates, and neoclouds have the same problem.

NVIDIA is Boeing and Airbus rolled into one. It captures most of the industry's gross margin at the point of sale and sets a release cadence (Hopper, Blackwell, Rubin) that keeps operators on a permanent CapEx cycle, much like the airframers claw back airline profits through aircraft pricing and engine upgrade cycles whenever carriers start earning excess returns.

On the real estate side, planes are useless without takeoff and landing slots at constrained hubs like Heathrow, JFK, or LaGuardia. In compute, the scarce input has shifted from silicon to powered shells and grid interconnects. Locking up 100+ MW of high-density, liquid-cooled capacity with cheap power and an energized interconnect is the neocloud version of owning slots at a fortress hub.

4. Yield management: three cabins

No operator can afford to fly empty, and not every customer needs the same reliability. Neoclouds have landed on the same cabin segmentation airlines built forty years ago.

• First class (multi-year take-or-pay): Frontier labs and hyperscalers sign 3 to 5 year contracts for large, contiguous InfiniBand clusters of 16,000+ GPUs. A $500 million training run can stall on a single bad node or latency spike, so these buyers pay a fixed rate regardless of usage to lock in uptime. This is the cabin the debt facility gets underwritten against.
• Economy (on-demand bare metal and VMs): Startups and enterprise teams renting 8 to 64 GPUs by the hour or month for fine-tuning and inference. Switching costs are low, churn is high, and buyers compare hourly rates across aggregators the way leisure travelers sort Google Flights by price.
• Standby (spot and batch): 500 idle GPUs at 3am on a Sunday get sold at a 60% to 80% discount, priced just above marginal power. Fault-tolerant work (batch processing, synthetic data generation, academic jobs) fills the gap and gets preempted on 30 seconds' notice when a full-fare customer shows up at the gate.

5. The EETC of silicon

With that much operating overlap, Wall Street brought over aviation's financing playbook too.

For decades, airlines with sub-IG balance sheets have funded their fleets through EETCs and aircraft ABS. The pitch to credit investors is that you are lending against standardized collateral and contracted lease payments held in a bankruptcy-remote SPV, rather than against the airline itself.

Blackstone, Magnetar (I was a founding employee at Crayhill Capital, a former Magnetar spinout), Carlyle and others brought the same structure to GPUs. The neocloud signs a multi-year take-or-pay contract with an IG or well-capitalized counterparty, drops the contract and the NVIDIA servers into a bankruptcy-remote SPV, and borrows at 70% to 80% LTV and 9% to 14% to buy the cluster. Contracted cash flows amortize the debt over 36 to 48 months, and equity is left holding unencumbered hardware at maturity.

6. GPUs are engines, data centers are airframes

I'd push the analogy one step further. A GPU cluster maps better to an aircraft engine than to a whole plane, and the data center (powered shell, cooling loop, grid interconnect) is the airframe.

An airframe is a long-duration shell: a tube with wings, avionics, and a cabin. The engines (CFM56, LEAP, GEnx) do the work, and two of them often account for 40% to 50%+ of a narrow-body's value. Late in an aircraft's life, the engines can be most of what the package is worth while the airframe heads toward part-out.

In a neocloud, the powered shell is the airframe. The building, the 100 MW utility interconnect, the substations, and the liquid-cooling manifolds have a 20 to 30 year economic life, which makes them the long-duration collateral. The GPU is the engine, turning megawatts into tokens, and it comes off the wing every few years while the shell keeps running for decades.

This matters because on the ABS desk, whole-aircraft deals and standalone engine ABS (the Willis Lease and Shannon Engine Support shelves, for example) trade on different credit curves. Three underwriting variables change immediately:

Most GPU debt today is sized like whole-aircraft ABS, assuming slow, linear depreciation on a cohesive asset. The collateral behaves like engine ABS with no maintenance reserve.

Anyway, this post is running long. In Part 2 I'll talk about why the Aircraft Engine ABS analogue is so important, and where this analogy breaks down, and also where there are interesting opportunities and where I think smart guys with ABS experience should be building. Stay tuned.