Compute Collateral
Independent research on compute-backed credit

Someone has to price the chip at the bottom of the structure.

Deal-level analysis of debt secured against GPUs, data-centre assets, and the operators between them — for the institutions holding the paper.

0510152025 0%25%50%75%100% Years from origination Debt outstanding Hardware value Cover lasts between one and five years depending on where secondary marks land Both series as a percentage of original hardware cost. Debt advanced at a 65% rate against a 22-year maturity. The band is the spread across secondary marks, not a confidence interval.

Illustrative. Residual paths, advance rate and amortisation profile are invented and must be replaced with sourced series before publication. The band is the range across observed secondary marks, not a confidence interval, and the central path is not a forecast.

How long the collateral covers the debt depends almost entirely on where secondary marks land, and the spread is wide. On favourable marks the facility is covered approaching year five; on stress marks cover is gone inside a year. The advance rate is set from a single residual curve, which prices none of that range — and past the crossing point, whenever it arrives, recovery depends on refinancing and re-equipping rather than on the assets named in the security package. Two things then decide the outcome, and neither is in the collateral analysis: the optionality the documents hand to related parties, and whether compute demand spreads through the wider economy or stays with a handful of buyers.

Coverage

Three places the risk actually sits

Equipment-level collateral

GPU-secured facilities, sale-leasebacks, vendor-financed structures. Perfection, custody, repossession logistics, and what a chip fetches in a forced sale.

Data-centre ABS and CMBS

Lease quality, offtake financeability, power contingency, and the take-out capacity the sector is quietly relying on.

Operator credit

Neoclouds, converted miners, integrated sponsors. Contract concentration, counterparty quality, and cash actually available for debt service.

About

Two halves that are rarely held by one person

Pricing this asset class takes someone who knows how a structure fails and knows what the hardware inside the SPV will be worth next year.

Martin Andrews covered structured credit through 2001–2011 and published on collateral quality before the market repriced it. Since then he has been active in machine learning research, having obtained his PhD in the 1990s, working on the systems that consume the very collateral being financed here. Reports assume you know credit and explain the compute, or the reverse, depending on which side of the desk you sit.

Writing

Recent writing

Recent thought pieces and research
Published Title Type
Advance rates in equipment-backed compute facilities: a survey of nine 2024–25 vintages Subscriber research
Where the take-out comes from if securitisation issuance stalls Subscriber research
Residual value methodology: pricing obsolescence on a two-year product cycle Silicon Duration
Reading a compute offtake contract as a credit analyst Silicon Duration
Access

Getting the work

Read one first

The residual value methodology piece is the clearest single sample of how the work is done. Full text, no registration wall.

Read the sample note

Institutional subscription

Research on individual debt instruments, released to subscribers on completion, plus the full archive and direct analyst access. Priced per seat and per desk.

Enquire about access