On November 5, 2026, refurbished Dell PowerEdge XE9680 servers with eight H100 SXM GPUs and near-identical CPUs and memory were listed between $180,000 and $375,000. The spec line on those listings leaves out what you most need to know: warranty, GPU memory health, firmware, and what is missing from the box. This guide covers current asking prices, what to inspect before you pay, and how to put dealer quotes side by side.
Prices below were checked on November 5, 2026. They move, so treat them as a snapshot.
What they cost right now
Public asking prices for complete refurbished 8x H100 SXM servers, checked November 5, 2026. Server, condition, and warranty are as each seller lists them.
| Seller | Server | Condition and warranty | Price | Status |
|---|---|---|---|---|
| ServerPartDeals | Dell PowerEdge XE9680, 2x Xeon Platinum 8462Y+, 2 TB, 8x H100 80 GB SXM5, no drives | Seller refurbished, Dell ProSupport ends November 4, 2028 | $180,000 | Sold out |
| The Server Store | Dell PowerEdge XE9680, 2x Xeon Platinum 8462Y+, 2 TB, 8x H100 80 GB SXM5 | Tested, 90-day warranty | $199,999 | Out of stock |
| Network Outlet | Dell PowerEdge XE9680, 2x Xeon Platinum 8462Y+, 2 TB, 8x H100 80 GB SXM | Refurbished, 3-year warranty (site-wide banner) | $249,000 | Available |
| Cloud Ninjas | Dell PowerEdge XE9680, 2x Xeon Platinum 8462Y+, 2 TB, 8x H100 80 GB SXM | Refurbished, 30-day dealer warranty, Dell warranty through April 2028 | $375,000 | Available |
| Cloud Ninjas | Supermicro SYS-821GE-TNHR, 2x Xeon Platinum 8462Y+, 1 TB, 8x H100 80 GB SXM5 | Refurbished, 18-month warranty | $375,000 | Unavailable |
We found these listings ourselves and have not bought from any of these sellers.
Two other reference points:
- American Compute, a valuation firm, puts fair market value at "$195,000 per used 8-GPU node in August 2026," based on "observed listings for whole nodes, adjusted to a reference configuration, not realized sales." It lists an orderly liquidation value of $173,000 and a forced liquidation value of $126,000, and labels all three "Indicative, not an appraisal" (H100 server values, updated October 17, 2026).
- Compute Exchange, a marketplace, publishes an indicative range of $26,000 to $30,000 per refurbished H100 SXM5 module for Q3 2026, "based on aggregated supplier quotes and broker market data" (refurbished H100 SXM5). A loose SXM5 module only runs on an HGX baseboard in a server built for it, so this number does not translate directly into a system price.
On November 5, 2026, every complete refurbished system we found was listed at $180,000 or more. The two lowest listings were out of stock, and the lowest one available to buy was $249,000. Prices under $180,000 showed up only as liquidation estimates. Similar XE9680 configurations were listed $195,000 apart, so get several written quotes.
What to inspect before you pay
Ask the seller for the evidence below before you send money, and check it again when the server arrives.
GPU memory health: ECC counters and row remapping
NVIDIA's nvidia-smi documentation defines two ECC counters. Volatile counters "track the number of errors detected since the last driver load." Aggregate counters "persist indefinitely and thus act as a lifetime counter," and "Clearing aggregate counts is not supported on Ampere+." On an H100, the aggregate count is a history the seller cannot clear with nvidia-smi.
H100 also replaces failing memory rows with spare ones. NVIDIA's memory error management guide lists row remapping for GH100, the H100 chip, with support for up to 512 remappings in the frame buffer. Ask for this output from all eight GPUs:
nvidia-smi -q -d ECC,ROW_REMAPPER
Read these fields:
| Field | What it means |
|---|---|
| Aggregate uncorrectable ECC errors | Lifetime count of errors the GPU could not correct |
| Correctable Error (row remapper) | "The number of active row remappings due to correctable ECC errors" |
| Uncorrectable Error (row remapper) | "The number of active row remappings due to uncorrectable ECC errors" |
| Pending | A remap is waiting. "The GPU must be reset for the remapping to go into effect." |
| Remapping Failure Occurred | Indicates "whether or not a row remapping has failed in the past" |
NVIDIA's RMA policy sets the remapping failure flag when the GPU needs to remap a row for an uncorrectable error and cannot: the bank already has eight rows remapped for uncorrectable errors, the row was already remapped, or the GPU has reached 512 remappings for uncorrectable errors. Its RMA criteria are met "when the row-remapping failure flag is set and validated by the field diagnostic." Do not accept a GPU with that flag set, and ask the dealer what their warranty does with a GPU that sets it later.
Ask for the full nvidia-smi -q output too. It carries each GPU's serial number, which you can match against the units that arrive.
Run hours
We could not find an NVIDIA-documented power-on-hours counter for the H100. "GPU Uptime" in nvidia-smi -q counts only from when the driver initialized the GPU. Ask the seller where the server ran, for how long, and at what utilization, and treat the answer as a claim until the ECC and row-remapping history supports it.
A stress test under load
NVIDIA's DCGM diagnostics run in levels, and each level includes everything below it. On an 8-GPU Hopper system, NVIDIA lists level 3 at under 35 minutes and level 4 at under 2.25 hours. Level 3 adds the diagnostic, targeted stress, targeted power, nvbandwidth, and NCCL tests. Level 4 adds memtest and the pulse test.
dcgmi diag -r 3
Ask for a level 3 or level 4 report run shortly before shipment, then run it yourself on arrival. NVIDIA states that DCGM diagnostics are not designed to replace its field diagnostic tools or to facilitate an RMA, so use the report as an acceptance test and keep the warranty terms in writing.
While the test runs, watch the kernel log for XID errors. NVIDIA's XID catalog lists these memory-related codes for H100:
| XID | Meaning |
|---|---|
| 48 | Double-bit ECC error |
| 63 | GPU memory remapping event |
| 64 | GPU memory remapping failure |
| 92 | High single-bit ECC error rate |
| 94 | Contained memory error |
| 95 | Uncontained memory error |
Baseboard and firmware
The eight GPUs, four NVSwitches, and a management controller live on the HGX baseboard, and each one carries its own firmware. NVIDIA's DGX H100/H200 firmware guide shows the GPU tray package reporting separate versions for the GPU VBIOS, the NVSwitch, the HGX management controller, and the FPGA, and listing each of the eight GPUs on its own line.
Ask the seller for:
- the HGX baseboard part number and serial from its label or the BMC inventory;
- the installed firmware versions for the BMC, BIOS, GPU VBIOS, and NVSwitch;
- whether the server maker still publishes firmware updates for that model.
Power supplies and fans
Each H100 SXM GPU is rated up to 700 W, configurable (NVIDIA H100). Eight of them are 5.6 kW before CPUs, memory, NICs, and fans. Supermicro's SYS-821GE-TNHR ships with six 3000 W Titanium power supplies in a 4+2 redundant arrangement, rated 2880 W at 200 to 207 VAC and 3000 W at 207 to 240 VAC. Three of the XE9680 listings above show six 2800 W supplies.
Ask for:
- a BMC export (iDRAC or Supermicro IPMI) showing every power supply present and healthy, with redundancy intact;
- the system event log, so you can see past power, thermal, and fan events;
- the fan module status, and which power supplies and fans were replaced during refurbishment.
Confirm your rack can deliver the input voltage the power supplies need for full rating before the server ships. Our power planning guide covers circuits and PDUs.
Warranty terms
A refurbished server can carry two layers of support: the dealer's warranty and whatever remains of the OEM's support contract. The listings above range from a 30-day dealer warranty backed by a Dell warranty through April 2028, to a 3-year dealer warranty.
Get these in writing:
- warranty length and start date (ship date or delivery date);
- repair, replace, or credit, and who decides;
- whether replacement parts ship in advance or after the failed part is returned;
- whether any OEM support contract transfers to you, with its end date and service tag;
- the dead-on-arrival window and who pays return freight (Dell lists the XE9680 at up to 114 kg).
Sellers to start with
These three are the refurbished sellers we would contact first. Their terms, from their own sites on November 5, 2026:
- IT Creations lists the Supermicro SYS-821GE-TNHR with the HGX H100 8-GPU board, priced by quote. Its warranty page offers 30 days, 60 days, 1 year, 3 years, or 5 years, chosen at checkout.
- ALTA Technologies is R2v3 certified and holds an A+ BBB rating. Its refurbished servers page says "Alta Technologies offers a 1-year warranty on full configurations," while its homepage advertises a 3-year warranty for end users, so ask which applies to a GPU server. It quotes H100 systems on request. The same refurbished servers page says "To date, we only sell AI GPUs within the USA," which matters if your server ships abroad.
- NetEquity offers end users a lifetime replacement warranty with the remedy at NetEquity's "sole discretion," and excludes firmware and BIOS issues. Its H100 pages currently read as buy-side, so expect to ask what it can source.
If you are working from an older list of builders: Lambda ended its on-premise hardware business as of August 29, 2025 (Lambda), and Silicon Mechanics' site now redirects to Thinkmate. Compute Exchange is a marketplace. It states that it does not certify, inspect, refurbish, or take ownership of hardware, and that condition and warranty terms are negotiated between buyer and seller (Compute Exchange).
How to compare quotes
Put every quote in one table before you compare prices.
| Line item | What to write down |
|---|---|
| GPUs | 8x H100 SXM 80 GB on an HGX baseboard, baseboard part number |
| Server | OEM and model, CPUs, memory size |
| Storage | Boot and data drives, or "none included" |
| Network | NIC model and count, optics and cables |
| Power | Power supply count and rating, redundancy |
| Evidence | nvidia-smi -q output, DCGM report level and date, serial numbers |
| Warranty | Dealer term, OEM contract end date, advance replacement or not |
| Delivery | Freight method, insurance, DOA window, ship date |
| Payment | Deposit, balance, method, quote validity date |
The listings above show why this matters. The $180,000 unit was listed without drives. The $375,000 XE9680 carries a 30-day dealer warranty and depends on the remaining Dell warranty after that. A quote that looks cheap can be missing storage, network cards, or warranty you would pay for separately.
Verify the seller and their payment instructions independently before you wire money, and read the return terms before you pay.
If the server is going outside the US, read our export controls guide first. The buyer's ownership can require a license even when the destination does not.
Frequently Asked Questions
How much does a refurbished 8x H100 server cost in November 2026?
Public asking prices for complete refurbished 8x H100 SXM servers ran from $180,000 to $375,000 on November 5, 2026. The two lowest were out of stock, and the lowest one available to buy was $249,000. American Compute put fair market value at $195,000 per used 8-GPU node in August 2026, based on observed listings.
Can a seller reset the ECC error history on an H100?
NVIDIA's nvidia-smi documentation states that clearing aggregate ECC counts is not supported on Ampere and later GPUs, which includes H100. The aggregate count is a lifetime record.
Can I see how many hours a used H100 has run?
We could not find an NVIDIA-documented power-on-hours counter for H100. GPU uptime in nvidia-smi counts from driver initialization. Ask for deployment history and judge it against the ECC and row-remapping data.
What does "Remapping Failure Occurred" mean?
The GPU needed to remap a memory row after an uncorrectable error and could not. NVIDIA's RMA criteria are met when that flag is set and its field diagnostic confirms it. Do not buy a GPU with it set.
Get a sourced quote
gpu.fm does not hold inventory. We source refurbished and new H100 servers through dealers and quote on request. Tell us the configuration, destination, and target date, and which of the evidence above you want with the quote.
Request a quote | H100 SXM product page | GPU server pricing guide
