Collateral Without a Curve: A Residual-Value Methodology for Compute-Backed Credit
TL;DR
Whether a defensible residual-value curve for data-centre accelerators can be built from public information, to build one, and to test how sensitive disclosed 2025 structures are to getting it wrong are determined.
Abstract
Background: Lending secured on graphics processing units reached record volume in 2025, yet compute assets have no published residual-value methodology of the kind aircraft, rail and shipping finance have relied on for decades. Lenders price an asset whose value curve nobody has estimated. Objectives: To determine whether a defensible residual-value curve for data-centre accelerators can be built from public information, to build one, and to test how sensitive disclosed 2025 structures are to getting it wrong. Methods: Two constructions are attempted on evidence to 31 December 2025. The market approach standard in transport finance proves unusable: the published secondary-price evidence is thin and mutually contradictory. We therefore build an income-anchored curve, Compute Base Value, as the present value of the remaining net rental stream, terminating at an endogenous economic life: the age at which net rental revenue falls to cash operating cost. It is calibrated on a 15 October 2025 rental cross-section, thermal design power and 2025 colocation and energy costs, then adjusted for illiquidity and distress and applied to seven structures. Results: Economic life falls from 7.2 years at 300 W to 4.3 years at 1,800 W. A 700 W accelerator retains 73%, 46%, 25% and 10% of delivery value over four years, close to the personal-computer profile and a third of the aircraft convention. Every disclosed structure running past four years loses collateral coverage within eighteen months. Conclusion: Compute collateral behaves like computer capital, not transport capital. Amortisation within four years is the binding constraint, and residual-dependent structures have no feasible advance rate.