The Secondary Market
What a GPU is worth as an earning asset, and what machines list and sell for.
One market value per chip. It is what a GPU earns at the rates the chip is being contracted at today, with every assumption disclosed. Fleet sales, retail sales and new cost sit beside it as checks.
§ Market value
Values and costs are per GPU, all-in: the GPU with its share of the server and network gear it runs in. That is the unit rent is paid on and the unit lenders take as collateral. The cost basis is the purchase price of that unit, the figure lenders test loans against. Neither includes the building, power or cooling. In the filed GPU facilities we track, the lien covers the GPU servers as equipment, perfected by a UCC filing, alongside the customer contract.
The value is the present value of the unit's cash flows: a contract signed today at the term and rate the market is signing, followed by re-leasing in the spot market at assumed future rates. It is a model output, recomputed with each daily data refresh. Beside each value: new cost, fleet sales between operators, and retail sales of the GPU alone, which set a floor. Values are for units in place. For CCIR's estimated stress-case values, based on the prior cases discussed there, see When GPU Contracts Break.
Model assumptions & sensitivity
| Chip | Contract $/hr | Band | Term | Re-lease $/hr | Net earnings / yr | From today's contract | Earning life left |
|---|---|---|---|---|---|---|---|
| B300 SXM | 3.85 | 3.77–3.92 | 3.8Y | 5.13 | $23.3k | 92% | 9.0y* |
| B200 SXM | 3.13 | 2.99–3.27 | 2.3Y | 3.74 | $18.3k | 81% | 8.2y* |
| H200 SXM | 1.64 | 1.63–1.65 | 2.0Y | 2.60 | $8.4k | 64% | 8.3y |
| H100 SXM | 1.40 | 1.17–1.63 | 2.0Y | 2.00 | $6.6k | 66% | 7.0y |
Net earnings / yr: what one GPU earns, all-in, in a contract year, after power, colocation, upkeep and overhead. Re-lease: the published neocloud interruptible rate, before utilization. From today's contract: the share of today's value earned under the contract the unit signs today; the rest depends on the market after it ends. * Vintage assumed.
Fleet sale: an executed sale between operators, disclosed in a filing, per GPU all-in. H200 SXM: WhiteFiber sold 126 H200 servers for about $26.1 million in the first quarter of 2026, about 90% of their cost. H200 SXM has no GPU-only ask or sales record yet.
How the value is built: the unit signs a contract today at the term and rate in publicly disclosed, executed deals for its chip. Deals from the last 12 months count, newest most; larger deals carry somewhat more. A bundled deal enters on a GPU-only adjusted basis. After the contract, the unit re-leases at the published interruptible rate at 70% utilization. That rate falls faster while a chip is young and more slowly after its third year. Costs come off each year. Contracted cash and re-lease cash are discounted at separate rates tied to GPU-backed financing in our credit ledger. The unit earns until its rent stops covering its costs. The range moves one input at a time. See Method.
§ Rack-scale systems: value per rack
GB200 and GB300 NVL72 sell as whole racks, so they are valued per rack, in place and earning, with the rack's share of networking and storage. No rack has a sales record yet, so the market line shows the cost basis only.
Rack model assumptions
| System | Rent $M / IT MW-yr | Band | Term | Basis | Value per IT MW | Life left |
|---|---|---|---|---|---|---|
| GB300 NVL72 | 14.7 | 14.1–15.2 | 5.0Y | Rent $14.7M per IT MW-year: 7 filed NVL72 contracts (4 GB300), 339 MW, hyperscaler anchor offtakes included; pooled with a GB300/GB200 factor of 1.07; 5.0-year term (GB300 deals, weighted like the rent) | $38.9M | 4.6y* |
| GB200 NVL72 | 13.7 | 13.2–14.2 | 2.1Y | Rent $13.7M per IT MW-year: 7 filed NVL72 contracts (3 GB200), 339 MW, hyperscaler anchor offtakes included; pooled with a GB300/GB200 factor of 1.07; 2.1-year term (GB200 deals, weighted like the rent) | $23.9M | 4.1y* |
Racks use the same method. The rent is the price per megawatt in executed NVL72 contracts, large hyperscaler deals included (see /term). After the contract, a rack re-leases at rack-scale market rates. The value includes the rack's share of networking and storage. The cost basis covers the rack and its network gear. GB200 has no filed purchase price. Its cost basis comes from 2025 analyst estimates plus network, checked against the GSA schedule price. * Vintage assumed.
§ 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 beside the asks posted today across an official marketplace API and a named dealer cohort. The record: July 2023 – September 2026, 3,356 sold listings, 13,555 units, $29.4M of the models tracked here, from a 4,205-listing capture. Sales record through 2026-10-05; values above use the 2026-10-09 data.
Chart notes
Per-GPU prices, log scale, as of 2026-10-05. 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 – September 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 NVL 141GB | 2024 | 1.9y | $32,295 | 9 | $42,995 | 13 · 4 | 1.33× | 102.5% [C] | 41 | $1.10M |
| H100 NVL 94GB | 2023 | 3.1y | $24,000 | 7 | $34,382 | 10 · 3 | 1.43× | 71.4% [B] | 44 | $1.14M |
| L40S 48GB | 2023 | 3.2y | $6,000 | 17 | $9,869 | 40 · 4 | 1.64× | — | 126 | $0.86M |
| H100 80GB SXM5 | 2022 | 4.0y | $11,700 | 7 | $26,888 | 7 · 1 | 2.30× | 34.8% [B] | 106 | $1.24M |
| H100 80GB PCIe | 2022 | 4.0y | $26,000 | 3 | $41,688 | 13 · 3 | 1.60× | 77.4% [B] | 49 | $1.88M |
| A100 80GB PCIe | 2021 | 5.3y | $12,500 | 7 | $22,549 | 24 · 5 | 1.80× | 50.3% | 62 | $1.02M |
| RTX A6000 48GB | 2020 | 5.8y | $3,733 | 115 | $5,941 | 54 · 2 | 1.59× | 80.3% [B] | 2,105 | $8.03M |
| A100 80GB SXM4 | 2020 | 5.9y | $5,200 | 10 | $18,644 | 12 · 2 | 3.59× | 20.9% | 139 | $1.39M |
| A100 40GB | 2020 | 6.4y | $3,695 | 39 | $6,810 | 82 · 5 | 1.84× | 14.9% | 1,172 | $7.50M |
| V100 32GB | 2018 | 8.6y | $649 | 154 | $890 | 162 · 5 | 1.37× | 3.5% | 5,374 | $3.98M |
| V100 16GB | 2017 | 9.1y | $225 | 118 | $338 | 194 · 4 | 1.50× | 1.2% | 4,337 | $1.29M |
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 (source count shown per row; 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 2026-10-05.
§ Method
Income approach (model-implied). The value is what a GPU earns, all-in, discounted to today. The unit signs a contract at the term and rate in publicly disclosed, executed deals. Deals from the last 12 months count, newest most; larger deals carry somewhat more. A bundled deal enters on a GPU-only adjusted basis. Where a chip has too few deals, posted on-demand prices less the measured gap to contracted prices fill in. After the contract, the unit re-leases at published interruptible rates at a stated utilization. Site costs, upkeep and overhead come off. The unit earns until its rent stops covering its costs. A chip with no sales record is marked modeled only.
Executed record. Sold prices of used GPUs, recorded by hand from eBay sold-listing observations, July 2023 – September 2026: 3,356 listings and 13,555 units of the models tracked here. Lots, servers, converted cards and mislabeled parts are left out. Every capture is archived. We publish rolling medians with their counts, never individual listings. Thin months can move on few sales.
Posted asks. Current asks for single GPUs from eBay's official API and from dealer catalogs, each collected under its source's terms. A dealer counts only while its latest capture is recent. Asks in other currencies convert to US dollars, ex-VAT, at the ECB reference rate. Asks run above what sells, so they sit beside sales and never blend with them.
Decline after the contract. Measured on the rental prices of earlier generations as they aged, from archived and live pricing.
Discount rate. Tied to the yields on GPU-backed financing in our credit ledger, plus a premium for equity and smaller operators. It moves as that market reprices.
What is model and what is observed. The value is the only model output on this page, and it is labeled wherever it appears. Everything else is an observed price with its source. Sales and asks check the value and never feed it. The value does not estimate recovery in a default.