How the numbers are built, what they can and cannot tell you, and the rules this publication holds itself to.
There is no single price for a used GPU. There are at least two, and they differ by more than most readers expect. A quote is the price at which supply is offered — a broker's indicative range or a dealer's posted ask. An executed price is what hardware actually sold for. Across the parts tracked here, asks currently run between 1.3× and 2.3× above executed prices.
The rule this publication holds: the two lanes are never averaged, never blended into a single range, and never shown without their sample size. Blending them produces a number that looks authoritative and means nothing — and the blending is invisible to the reader, which is what makes it dangerous. Anyone sizing an advance rate against GPU collateral needs the executed figure and the spread, not a midpoint.
Channel matters as much as lane. Our quote lane is institutional and volume-weighted; the executed lane currently available is retail secondary — largely single cards and small lots. These are not the same market, and a quote and an executed figure on the same row are not a like-for-like pair. Where a spread ratio is published it is computed within a single channel, and it is labelled as such. Institutional bulk transaction data is the gap this publication is working to close; no public source measures it today.
Where the two lanes contradict each other, the contradiction is published rather than resolved by preference. The H100 PCIe row currently shows an executed median above our quoted range on a sample of four sales — that is flagged as unresolved on the row itself, not quietly smoothed away.
Compute Exchange hardware market. Indicative price ranges from aggregated supplier quotes and broker market data, per SKU and condition grade.
Named dealer listings. Posted asks from named refurbishers and dealers (currently IT Creations), captured with listing date.
eBay sold listings. Being integrated — executed-transaction medians per SKU will carry their own sample counts when they print.
Executed transaction medians. Retail secondary sale medians per SKU, sourced from CCIR (ccir.io) and carrying their published sample counts and trailing window. Attributed on every row that uses them. Own-sourced executed data — including institutional bulk transactions — is in development and will be graded and labelled separately when it prints.
A — three or more independent sources with executed-transaction evidence. B — two independent sources. C — single source or carried forward, flagged. All current price figures are grade B or C; no grade-A figure exists yet and none will be claimed until the evidence supports it.
Corrections are published, never silently edited. Every material revision to a published figure is noted with the original value, the corrected value, and the reason.
GPUdebt.com holds no positions in anything it prices. No trading, no brokering, no inventory, no paid placement in the data. The publication's only products are the data and the analysis.