Published figures for a used H100 server price disagree before anyone opens a chassis. A Hashrate Index article of 8 April 2026 states that a used eight-GPU H100 server "trades at roughly" $150,000 to $180,000, with no source, date or configuration [1]. American Compute values an HGX H100 8-GPU server at $195,000 per node in its October 2026 report, built from "observed listings for whole nodes, adjusted to a reference configuration, not realized sales" [2]. On 8 November 2026, two Dell PowerEdge XE9680 listings with eight H100 GPUs, neither stating the unit's condition, asked $199,999 and $275,000 [13, 14].
One is an estimate with no stated basis, one a modeled level built from asking prices, and two are public offers seen on one day. None records a sale. A particular server's price then moves with its contents, condition, warranty, diagnostic records, licenses and export screening.
Used H100 server prices in public sources
| Source and date | What it prices | Figure | Evidence class |
|---|---|---|---|
| Hashrate Index, 8 April 2026 | Used 8-GPU H100 server, no configuration given | $150,000 to $180,000 | Indicative price [1] |
| American Compute, updated 17 October 2026 (data through August 2026) | HGX H100 8-GPU node, standard configuration | $195,000 fair market value, $173,000 orderly and $126,000 forced liquidation | Indicative price built from listings [2] |
| The Server Store, viewed 8 November 2026 | Dell XE9680, 8x H100 80GB SXM5, condition not stated (seller of refurbished equipment), out of stock | $199,999 | Listing [13] |
| Vipera, viewed 8 November 2026 | Dell XE9680 bundle, 8x H100, condition not stated | $275,000 | Listing [14] |
| Interpro, viewed 8 November 2026 | Decommissioned Dell XE9680, 8x H100 | Contact sales | Listing, no price shown [16] |
A listing is a public offer seen at a stated time, and an indicative price is a reported price level without a verifiable counterparty. The evidence classes note sets out all six classes and why mixing them misleads.
Asks also run above sales. In American Compute's October 2026 Rack Report, a used H100 SXM card was listed at a median of $16,576 and sold for $14,250 typical, a ratio of 1.16 [3, 4].
Card, server or system
A per-GPU figure does not multiply into a server price. Publishers' 2026 indicative ranges for a used or secondary-market H100 SXM card run from Standard Mobile's $15,000 to $21,000 secondary-market range for the second quarter to BuySellRam's $21,000 to $28,000 for the first, with Mercatus and Compute Exchange inside that span [7, 9, 11, 12].
NVIDIA's H100 page compares two versions: the 80 GB H100 SXM for HGX partner systems and the DGX H100, and the 94 GB H100 NVL as a PCIe card [17].
The server around the GPUs varies too. NVIDIA's DGX H100 carries two 56-core CPUs, 2 TB of memory, eight ConnectX-7 InfiniBand cards and ten NVMe drives [18]. The Server Store's XE9680 listing describes two 32-core CPUs, 2,048 GB of memory, one ConnectX-7 card and no drives, and its option list prices memory separately [13]. Interpro's decommissioned XE9680 lists two 48-core CPUs, eight ConnectX-7 400G cards and ten drives [16]. Vipera's $275,000 page states no CPU, memory, storage or network build [14].
A title that names eight GPUs need not include them. SabrePC's "Supermicro SYS-821GE-TNHR Barebone - 8U - 8x NVIDIA HGX H100 SXM5", a listing seen on 8 November 2026 at $23,283.36, is a barebone system whose contents list no GPU baseboard [15]. The HGX server price explainer walks the same parts list for new B200 and B300 builds, and DGX, HGX and MGX explained covers what a DGX adds to an HGX board.
Condition labels and what each asserts
No condition scale is shared across sellers. On eBay, "Certified - Refurbished" means work by the manufacturer or a manufacturer-approved vendor, the Excellent, Very Good and Good refurbished grades are "backed by a one year warranty", and "Seller refurbished" means restored "by the eBay seller or a third party" [21].
Publishers disagree on which costs more. Compute Exchange for the third quarter of 2026 and BuySellRam for the first price refurbished H100 SXM cards above used ones, at $26,000 to $30,000 against $22,000 to $27,000, and $30,000 to $34,000 against $21,000 to $28,000 [7, 8, 11]. Mercatus, in a 27 March 2026 page updated 9 October, lists refurbished H100s at $18,000 to $22,000, below its $22,500 to $25,500 for used cards 18 to 30 months old [9].
A label says only as much as the records behind it. Compute Exchange describes refurbishment as field diagnostics, memory error and thermal checks, 24 to 72 hours of full-power stress testing and connector cleaning, shipped with a per-unit test report [8].
Warranty and support calendars
Warranty depends on who issued it, to whom and when. Supermicro's standard limited warranty gives GPU systems three years of labor and one year of parts from the date of Supermicro's invoice, and it "applies only to the original purchaser" [26]. On those table terms, a Supermicro GPU system invoiced before 8 November 2023 was past both terms on 8 November 2026 (arithmetic on [26]).
Under such terms, a second owner relies on the seller's warranty. Compute Exchange states that "Most refurbishers offer a 90 to 365 day limited warranty against defects, scoped to the refurbisher rather than NVIDIA" [8]. The Server Store's XE9680 page promises a "90 warranty" without naming the unit [13]. Some listings rule out returns: Vipera's says "All sales final" [14].
Support calendars bound the rest. Cisco set 19 July 2025 as the end of sale for its UCS C885A M8 servers with HGX H100 GPUs and 31 July 2030 as the last date of hardware support [20]. Microway relays that NVIDIA's DGX H100 end-of-sale program took final purchases by 25 January 2025 [19].
Test records that travel with the GPU
An H100 keeps its own memory history. NVIDIA's nvidia-smi documentation states that aggregate error counts "persist indefinitely and thus act as a lifetime counter", that clearing them "is not supported on Ampere+", and that each board's serial number is a "globally unique immutable alphanumeric value" [22]. Under NVIDIA's memory error return (RMA) policy, a GPU meets the RMA criteria for row-remapping failures when its row-remapping failure flag is set and validated by the NVIDIA Field Diagnostic [23].
A seller's "tested" needs a scope. NVIDIA's DCGM diagnostic suite takes under 35 minutes at level 3 and under 2.25 hours at level 4 on eight-GPU Hopper systems, and NVIDIA states that it is not designed to replace field diagnosis tools or to facilitate any RMA process [24]. A study of 1,056 A100 and H100 GPUs over 2.5 years found 3.2 times lower per-GPU mean time between memory errors on the H100, alongside improved resilience in critical hardware components [25]. Rillor's GPU systems page states that surplus, refurbished and used units are graded and tested before sale, and that every unit comes with its own report.
Licenses, drives and export screening
Software terms follow the card. NVIDIA's licensing guide states that each H100 PCIe or NVL, H200 NVL and A800 40GB Active GPU includes an NVIDIA AI Enterprise subscription [27]. The software can be installed on the system the GPUs sit in "but not on a different server or instance", and the start date is tied to the card [27]. These included subscriptions apply to those cards and not to the H100 SXM GPUs on an HGX board, for which NVIDIA lists NVIDIA AI Enterprise as an add-on [17]. DGX systems on the Hopper architecture include it in the DGX software bundle [27].
Drives come from a prior owner: Interpro lists its decommissioned unit with ten [16], and American Compute states liquidation values before de-installation, shipping and data sanitization [2].
Export screening reaches servers too. Guidance of 13 May 2025 from the U.S. Bureau of Industry and Security (BIS) counts "servers classified as ECCN 5A992.z" among the commodities that contain advanced computing integrated circuits [29], and NVIDIA's 10-Q says the U.S. restrictions announced in August 2022 "specifically impact" its A100 and H100 chips and "DGX or any other systems or boards which incorporate A100 or H100 integrated circuits" [30]. BIS's red flags include a buyer reluctant to state the end use and a freight forwarder listed as the final destination [28]. This is a summary of the regulations' text, not legal advice.
Reading a used price against comparable observations
One server has several values. In October 2026, American Compute put the same HGX H100 node at $216,000 installed and earning (limited sample), $195,000 removed and sold whole, about $188,000 parted out, $173,000 in an orderly sale and $126,000 in a forced sale, all indicative [3]. The installed value rests on what the server earns, and rent moved: AWS cut on-demand prices for its H100-based P5 instances by 44% from 1 June 2025 [31]. Book lives do not settle it: from 2025, Amazon cut the useful life of a subset of its servers and networking equipment from six years to five, while Meta raised certain servers and network assets to 5.5 years [32, 33]. How lenders weigh residual value is covered in GPU financing structures.
A used price is read against dated observations of the same configuration, condition and seller type. Rillor's GPU systems pricing method records each configuration's condition, such as new, refurbished or used, and its seller type, so two systems that differ only in condition or seller are different configurations, and each observation keeps its evidence class, so a listing is read as a listing and a quote as a quote.
Read the full report
The full report adds the card-level figures side by side, three server builds part by part and the diagnostic run levels.