Residual value methodology: pricing obsolescence on a two-year product cycle
Equipment facilities are sized on residual curves borrowed from asset classes that depreciate on a decade. Accelerators do not. This note sets out the method we use to mark hardware collateral, and what it does to advance rates when the curve is rebuilt from secondary prices rather than from vendor schedules.
The assumption under the advance rate
Every GPU-secured facility we have looked at rests on a residual value curve. The curve says what the collateral is worth in year two, year three, year four. The advance rate is then set so that the loan balance stays under the curve.
The curves in circulation are largely inherited. They come from asset classes — transport, industrial plant, IT hardware in general — where useful life is measured in years and obsolescence arrives slowly. Accelerators do not behave that way, and the reason is not that they wear out.
What actually depreciates
Three things move a chip’s secondary price, and only one of them is age:
- Successor performance per dollar. A new part at the same power envelope resets what the incumbent is worth for new workloads.
- Power and cooling fit. A part that needs a retrofit to install is worth less to the marginal buyer than the spot price suggests.
- Supply release. Collateral is correlated. The events that impair one borrower release inventory that marks the whole class.
The third is the one the borrowed curves cannot express, because in the asset classes they came from, defaults do not flood the resale market with the exact collateral being repossessed.
The method
We mark from observed secondary transactions where we have them, and from an imputed curve where we do not, built on successor performance per dollar rather than on elapsed time. Where a facility discloses its own residual assumption, we publish both curves side by side and let the gap speak.
The gap is usually large. That is the finding, and it is why this note is open rather than gated: the method should be arguable in public. The deal-level application of it is what subscribers pay for.
What this does not claim
This is a valuation method, not a forecast. It says what the collateral would fetch under stated conditions. Whether those conditions arrive is a separate question, and we do not answer it here.
Published in Silicon Duration, the open commentary of Compute Collateral. This is commentary, not investment research, not investment advice, and not an offer or solicitation. We do not trade the securities we cover.