The Secondary Market
What a deployed GPU is worth as an earning asset, and what machines list and sell for.
This page publishes one value per chip: what a machine is worth deployed and earning. It triangulates that value against the market. The value is model-implied, computed from CCIR's own benchmark. It is an income approach over the CRI rental rate and committed term curve. Every assumption is disclosed, and the discount rate is anchored to the same collateral's financing yields in the credit ledger. The market checks are observational: executed sales and posted asks, in separate lanes. They are inputs to recovery analysis, not a recovery estimate.
§ Income-derived value
Model-implied value of a deployed unit, derived from the CRI T2 (neocloud) US guaranteed on-demand rate and the neocloud committed term curve, recomputed on each daily data refresh. It is not a transacted price: the corroborating market records are beside each figure.
Model assumptions & sensitivity
Base case: year 1 at the 1Y committed rate; years 2–3 bootstrapped from the 2Y/3Y committed curve; 18%/yr decay beyond the last observed tenor of each chip's curve, a measured base calibrated to the realized decline of prior-generation rates (Method); 75% utilization; 65% net margin after power, colocation, and operations; 15% discount rate; 6-year service life from volume availability. Where a chip's committed curve ends short of year 3 (B200: 1Y), decay applies from the end of its observed curve. Missing tenors are never extended flat. A chip with no secondary-market record is published modeled only, so marked, and carries no ask or executed corroboration until one exists. The sensitivity band varies one assumption at a time around base: discount rate 10–25%, service life 5–7 years, decay 10–30%/yr (the 30% bound is the worst case on record: V100 from its scarcity peak). The band shows the min–max. Jointly optimistic or pessimistic corners are not stacked. Utilization and margin are held at base. A life bound at or below a chip's current age is omitted rather than shown as a zero. The discount-rate band is anchored to observed GPU-asset-backed financing yields in CCIR's credit ledger (see Method below). The no-contract floor is the same model with one input changed: the published Neocloud interruptible rate as the earning path from day one. That path is flat in year 1, decays beyond, and has no committed curve by construction. Utilization stays at the base 75% deliberately. A floating interruptible rate is the price at which spare capacity clears, so the rate already carries the demand risk. Haircutting utilization as well would count it twice. The executed lane is an observed disposition record, not a recovery estimate. Single-unit retail clearings can sit above or below a bulk disposition, and no current record measures market impact at volume. Loss-given-default is a transaction-level parameter (lien position, priority, workout cost, timing, pool composition) and belongs to the party doing the recovery analysis. This record is one of its inputs.
§ Contribute to the record
Do you buy, sell, broker, finance, or insure this hardware? If you hold executed price records (dealer sales, disposal and auction results, broker quotes, portfolio liquidations), CCIR accepts contributed data under the evidence-grading and provenance rules that govern every row on this page. Contributions are held source-confidential and enter published statistics only in aggregate, with their evidence grade disclosed.
Underwriting against GPU collateral? If you need a series observed (a model, a venue, a spread), get in touch.
§ The record — executed sales & posted asks
Three years of executed sales (July 2023 – August 2026; 3,009 sold listings, 10,911 units, $26.3M documented) beside the asks posted today across an official marketplace API and a named dealer cohort. The rental market prices an hour of compute. This page prices the machine itself.
Chart notes
Per-GPU prices, log scale, as of 2026-08-14. Executed dots require ≥3 sold listings in the window (models without one show ask only). The gap between dots is the posted premium. It narrows as a model commoditizes: ~2× on H100, ~1.2× on V100.
Chart notes
One panel per model with a locked launch-price basis. x = the executed record's calendar window, July 2023 – July 2026, shared across panels. y = percent of the evidence-graded launch basis, log scale fitted per panel to that chip's own range, so panels are not directly overlayable by eye. The table carries the comparable levels (bases: V100 $18,625 [A] · A100 $24,875 [A] · H100 $33,600 [B] · RTX A6000 $4,650 [B] — system-price, teardown, and announced-price bases; NVIDIA publishes no datacenter MSRP). The line is the executed sales record, a 3-month rolling median of sold listings. Every point is sourced sales. Thin months can move on few sales (the early-2025 H100 notch is a single month's 7-unit quantity liquidation, not a market level). The open circle at each panel's right edge marks the CURRENT posted-ask median (multi-source; the ladder above carries the levels and counts). Asks lead and overstate executed prices, roughly 1.2–2×, and the lanes are never blended. Archived ask history and dated contemporary reports are collected and retained in the dataset but not charted: sticky catalog prices and carried-forward quotes describe dealer posture, not the market's path. Executed levels before mid-2023 are not observed. Bars below: units sold per month across the executed record, calendar time. Models with thin or young records (H200, H100 NVL, A100 80GB) and L40S are tracked in the table but not paneled. A rolling median through one-listing months would whipsaw, and no L40S launch basis meets the evidence bar.
Chart notes
The comparison view: executed 3-month rolling medians only (no ask lanes), every generation on one age axis and one shared log scale, so a newer chip's position can be read against an older chip's history at the same age. Bases as in the panels above. Each curve covers only the ages the executed record observes (from Jul 2023). The generations do not collapse onto one aging curve: that is the finding, not a defect.
| Model | Intro | Age | Executed median (90d) | n | Posted ask median | n · src | Ask / executed | % of launch | 3-yr units | 3-yr $ volume |
|---|---|---|---|---|---|---|---|---|---|---|
| H200 141GB | 2024 | 2.3y | $27,155 | 17 | $38,248 | 26 · 4 | 1.41× | 86.2% [C] | 39 | $0.95M |
| H100 NVL 94GB | 2023 | 2.9y | $25,750 | 4 | $34,050 | 11 · 3 | 1.32× | 76.6% [B] | 38 | $0.99M |
| L40S 48GB | 2023 | 3.0y | $6,695 | 20 | $9,000 | 37 · 4 | 1.34× | — | 106 | $0.75M |
| H100 80GB SXM5 | 2022 | 3.8y | $14,000 | 9 | $25,150 | 8 · 1 | 1.80× | 41.7% [B] | 101 | $1.18M |
| H100 80GB PCIe | 2022 | 3.8y | $25,617 | 4 | $36,222 | 17 · 3 | 1.41× | 76.2% [B] | 48 | $1.85M |
| A100 80GB PCIe | 2021 | 5.2y | $9,222 | 10 | $19,635 | 25 · 3 | 2.13× | 37.1% | 57 | $0.96M |
| RTX A6000 48GB | 2020 | 5.7y | $3,600 | 95 | $6,039 | 52 · 1 | 1.68× | 77.4% [B] | 1,934 | $7.36M |
| A100 80GB SXM4 | 2020 | 5.8y | $5,400 | 23 | $12,399 | 13 · 3 | 2.30× | 21.7% | 133 | $1.26M |
| A100 40GB | 2020 | 6.3y | $4,238 | 38 | $6,495 | 78 · 3 | 1.53× | 17.0% | 1,061 | $7.01M |
| V100 32GB | 2018 | 8.4y | $664 | 136 | $784 | 149 · 4 | 1.18× | 3.6% | 3,686 | $2.82M |
| V100 16GB | 2017 | 8.9y | $250 | 91 | $337 | 194 · 4 | 1.35× | 1.3% | 3,708 | $1.15M |
Table notes
Executed = median sale price across sold listings, trailing 90 days; cells with fewer than 3 sold listings show a dash. Posted ask = median current single-GPU ask across an official marketplace API and dealer catalogs in the US and UK (source count shown per row; named, with currency and VAT treatment, in Method). Hopper-class dealer pages are largely quote-gated, which is why source counts vary by model. Asks are opening prices, never pooled with executed. Ask/executed is the posted premium. It narrows as a model commoditizes. % of launch uses the evidence-graded basis in Method ([B]/[C] = non-vendor basis). 3-yr volume counts all hardware forms of the model (cards, lots, servers); accessories and mislabeled items excluded and tallied. Intro = general availability; age as of 2026-08-14.
§ Method
Income approach (model-implied). The headline value is a discounted-cash-flow over the income a deployed unit can contract for, with every leg either observable or disclosed. The observable legs: the CRI T2 US guaranteed on-demand headline as the spot rate, and the neocloud committed term curve (1Y/2Y/3Y per-operator median discounts) pricing years one through three. The curve is the market's own price for locked duration, not a forecast. The assumed legs are printed with the figure: decay beyond the curve, utilization, net margin, discount rate, service life. It is a model output and is labeled as one everywhere it appears. The sale and ask records below corroborate it. The model-to-ask ratio and executed price are shown beside every figure, so the reader holds all three numbers at once. Where no market record exists for a chip (B200), the value is published modeled-only and marked as such on its card. It is not triangulated against asks or executed sales. Its committed curve is observed to 1Y with decay applied beyond, its rate leg is a provisional headline, and its volume-availability vintage is assumed rather than sourced. It graduates to the corroborated treatment when a market record exists.
Executed record. Dated, human-captured snapshots of the eBay sold-listing research record (Seller Hub product research), July 2023 – July 2026: 3,564 sold listings, 15,155 units, $28.9M. Every row carries listing title, average sale price, sale count, dollar volume, format, and date last sold; every capture is archived. Rows are classified by title into single cards, multi-unit lots, servers, bare baseboards, converted cards (e.g. SXM-to-PCIe conversions; tracked, never banded), accessories, and mislabeled SKUs (V100S, L40, H100i and kin; excluded and tallied). Of the full record, 11,388 units and $26.0M are hardware of the models tracked here. Those are the figures this page reports. The remainder is the excluded tally. Multi-sale listings count once at their average price. Their sale history carries a date-last-sold only, so month assignment for those listings is approximate. Chart lines are 3-month rolling medians across sold listings, drawn solid throughout. Every point is sourced sales. Per-month listing counts ship in the series CSV, and thin months can move on few sales.
Posted asks. Current single-GPU asks from the eBay Browse API (official, keyed) and dealer catalogs collected under each source's terms: IT Creations and UnixSurplus (USD), Bargain Hardware and ETB Technologies (GBP, ex-VAT basis, converted at the ECB reference rate of the observation date: 1.3446 as of 2026-08-04). Additional dealers are tracked quote-gated (presence, no price). Collected sources are named. Data contributed through the channel below is held source-confidential with its evidence grade disclosed. Asks are a stock of open offers. They lead and overstate executed prices and are shown only beside them, never blended.
For credit and valuation work. Lenders writing advance rates against GPU collateral, rating agencies testing value-retention and useful-life assumptions, insurers underwriting residual value, and borrowers negotiating either side of those terms all face the same gap: no citable, neutral record of what datacenter hardware sells for. Today, credit documents overwhelmingly amortize GPU collateral to zero, not because the hardware is worthless, but because no reference existed. This page is the record: observed secondary-market prices with provenance. It carries executed levels, posted asks, volumes, and the spread between them. These are the raw inputs from which the parties who own those judgments can independently derive recovery evidence, depreciation assumptions, or residual analyses.
Decay beyond the curve. The decay base is measured, not assumed: archived pricing pages and the live record together show A100 and V100 rental rates fell 15–19% a year as those chips aged, and the 18%/yr base sits inside that measured range. The 30%/yr bound of the sensitivity band is the worst case on record: V100 from its scarcity peak.
Discount rate. The 15% base and the 10–25% band are anchored to observed pricing in CCIR's credit ledger: GPU-asset-backed facilities on record have signed at effective all-in yields from roughly 15% (2023) to 7% (2026, rated). A going-concern discount rate sits at or above the secured debt cost of the same collateral. The observed record skews to the sector's largest, highest-rated borrowers, a floor on the sector's cost of capital. The band's width above that floor carries the equity premium and the typical operator's higher funding cost. The 25% upper bound keeps the 2023 era in the band: debt on this collateral then cost roughly 15% all-in, with equity cost above that.
What is model and what is observed. Exactly one figure on this page is a model output: the income-derived value. It is labeled model-implied wherever it appears, with its assumption set printed beside it. Everything else is an observed price with provenance. Rate inputs are observed prices out to the last observed tenor. Beyond it, a decay rate calibrated from realized rate history applies and is disclosed with the figure. CCIR publishes no depreciation schedule as a product. The model prices today's curve plus disclosed, history-calibrated assumptions, not a predicted market. Statistics follow the methodology's stat policy: counts ship with every median in the series data, thin cells are gated, and windows are labeled on every display.