The Most Important Person in GPU Finance Might Be Your Most Boring Banker
What equipment finance knows that GPU lending may not
In credit, you hedge everything. Provisionally. Potentially. Could be. Might.
When someone in this business drops the qualifier, pay attention.
There is a version of this story that is mostly about technology — the hardware, the depreciation curves, the generation transitions, the question of whether an H100 is worth anything in three years. That version gets written constantly and well by people who understand the compute stack, candidly, better than I do.
This is a different version. This one is about what happens when the money shows up.
And the money has shown up. CoreWeave has raised north of $25 billion. Lambda Labs did a $500 million GPU-backed facility with Macquarie. The broader data center ABS market is pushing toward $25 billion in annual issuance. Blackstone, Morgan Stanley, J.P. Morgan, Blue Owl — the institutional credit complex is in this trade now, in size.
And it appears the underwriting infrastructure hasn’t kept up. Not because the lenders are unsophisticated. Because nobody has built the boring plumbing yet.
Here is what an equipment finance lender knows that most people in this conversation may not.
Collateral is a theory until you have to repo it.
In equipment finance, when a borrower defaults, the clock starts running on pure cost. Storage. Insurance. Deterioration — a machine that sat eighteen months at an auction yard is worth materially less than one that was running. Transportation, multiple times. Auction fees. Reconditioning. All of it while the asset generates exactly zero revenue.
That experience shapes how you underwrite from the beginning. You ask different questions. Not just “what is this worth today” but “what does it cost me to hold this if everything goes wrong, and how long can I afford to wait for an orderly sale.”
Those questions are not yet being asked systematically about GPU collateral. So the market improvises — bespoke deal structures, black box models, and a lot of quiet hope that the collateral holds. All while systemic risk is ratcheting up.
Here is something I'll say with only one qualifier. GPUs are potentially better collateral than traditional equipment in a workout. Not definitively — the data is early and the first real default cycle hasn’t arrived yet. But the structure is worth understanding now, before it matters.
A repossessed GPU cluster is already in a data center. The data center is the storage. No separate storage cost, no transportation, no deterioration from sitting in a field. And more importantly: it can keep earning while you figure out what to do with it. List it on a spot market, cover the colocation costs, run an orderly sale without a ticking clock of pure cash drain forcing your hand.
A combine cannot do that. No traditional equipment asset can do that.
That’s not a technology story. That’s a credit structure story. And it has implications for LTV, covenant design, and liquidation assumptions that nobody underwriting this asset class has formally written down yet.
Consider who is actually at the table for these deals. On one side: institutions whose credit histories are measured in centuries. Banks that financed railroads, farm equipment, industrial machinery — organizations that have outlasted every confident model ever built inside them. They run green screens because the green screens work, and because the thing that matters in credit isn’t the interface, it’s the judgment that accumulated behind it over decades of cycles. On the other side: companies measuring their roadmaps in quarters, their disruption in months, their compute capacity in hours. People building quantum infrastructure, neural networks, decentralized systems, things that didn’t have names five years ago. They move at a speed that makes an annual credit review feel like geology.
These two worlds are now in the same deal. The century-old institution is writing the check. The bleeding-edge operator is pledging the collateral. And the plumbing that connects them — the reference rates, the appraisal standards, the covenant frameworks that let both sides speak a common language — hasn’t been built yet.
Somewhere in the middle, someone has to write the methodology and build the infrastructure that both of them can cite.