Rack economics: October 2026
October’s secondary-market evidence shows why a single depreciation rate can give a misleading view of a financed server fleet.
Two component markets, different directions
Check the installed memory.
A basket move cannot be applied to every server. Record memory generation, capacity, condition and the composition of comparable systems.
Check the generation and sale basis.
A component index is an external signal. Establish the actual configuration, price basis, sample limitations and net disposal costs before changing a collateral mark.
Source: American Compute, Rack Report October 2026, data through August 2026. Reported monthly movements in different component baskets; not whole-server returns. InfraCompute has not reconstructed the underlying dataset.
Keep the observation date visible
American Compute’s October 2026 Rack Report uses data through August 2026. Its memory index rose 19.3% in the month while its GPU index fell 12.3%. Those are movements in American Compute’s component baskets, not an August return earned by every server owner. The publication date, observation period and basket definition belong beside any chart using those figures.
A rack contains components traded in several overlapping markets. Memory availability, accelerator generation, CPU age and networking compatibility can move differently. Applying one annual depreciation percentage to the full purchase invoice can hide those differences. Updating a collateral view therefore begins with the actual configuration and condition of the financed assets.
Compare prices on the same basis
A dealer asking price, a completed sale and an income-based valuation answer different questions. An ask expresses a seller’s position. A sale records a transaction under particular lot sizes, conditions and timing. A rental-income model estimates a future cash stream. None should silently replace the others in a collateral mark. The Rack Report distinguishes asking prices from realised transactions and flags limited samples.
Whole-system comparisons introduce another layer. Support, memory, storage, interconnect, freight and installation may differ between quotes. A per-GPU figure can obscure those inclusions. Reconcile a complete node to its component schedule, then identify which parts have transferable warranties and an accessible resale channel. Preserve a separate estimate of removal and disposal costs.
Use rental economics as a scenario
A high current rental rate can support near-term cash generation without establishing a durable resale floor. To value rental income, assumptions are needed for paid utilisation, realised price, operating cost, downtime, refresh spending and remaining service life. Test price compression and shorter economic life together. A result based on today’s rent continuing for years should be labelled as a scenario rather than a market sale price.
The difference between continuing-use value and disposal proceeds can be economically significant. A functioning cluster may need complementary assets and a customer relationship to retain its earning power. A forced sale may break those links. Present a bridge from the operating case to net recovery instead of choosing the higher number and calling it collateral value.
Set the next monitoring trigger
Track configuration changes, customer renewal dates and new comparable sales as well as component prices. Escalate when one of those changes undermines the assumptions supporting the remaining debt. American Compute’s benchmark is an attributed external market signal; it is not an InfraCompute appraisal or an independently reproduced transaction dataset. An asset-specific financing decision still needs current evidence on the actual fleet.
Questions for the credit paper
- Are the quoted values asks, completed sales, normalised system estimates or rental scenarios?
- Do the compared systems include the same memory, networking, support and installation?
- What remains after removing the equipment and selling it within the available time?
About this guide
Original InfraCompute analysis informed by the American Compute material below. The article structure, examples and interactive visuals are InfraCompute’s work. Source-specific findings remain attributed to American Compute; attribution does not imply affiliation or endorsement.
- American Compute, Rack Report, October 2026 ↗
October 2026 issue, with data through August 2026. Benchmark methods and limited-sample qualifications apply.
We reviewed the published edition and selected public references. We have not independently verified American Compute’s proprietary datasets, every market claim or confidential financing terms. Frameworks and hypothetical models describe questions to investigate; they do not establish protections in a particular financing. Legal and engineering conclusions depend on the relevant documents, system and jurisdiction.
Read our research methodology →