Whole systems,sourced and brokered.
Used and refurbished DGX and HGX-class systems, OEM multi-GPU chassis, and complete nodes and racks, sourced from the same vetted supplier network as our GPUs and quoted per requirement.
Why buying a used GPU server is different
A complete system is more than the sum of its GPUs. Four things change when the unit of trade is the node.
The unit of trade is the node
Retired fleets come out of racks as whole servers, and most buyers want them that way: the GPUs, the baseboard, the host, the NICs and the storage as one working system. We source and quote the complete configuration, not eight loose cards plus a chassis.
Configuration is the specification
Two systems with the same GPU count are not the same system. Chassis, GPU form factor, interconnect, memory, NICs, storage and BMC state all move the price and the fit. We pin the configuration down before anything is quoted.
Provenance covers the whole chassis
A system's history is the history of every part in it. We ask the supplier for chassis and per-GPU serials, origin and service records, and we hold the condition representations on record for the system as a whole.
Acceptance is tested per system
The buyer acceptance-testing window verifies the system as delivered: every GPU enumerates at the stated count and memory, the BMC responds, power and network come up as configured, and a compute check completes across all GPUs.
GPU servers we broker
Described as goods, sourced as found. Clustermelt does not build, integrate or refurbish systems; it finds the one that already exists and brokers the sale.
- 01 /
HGX and DGX-class 8-GPU systems
NVIDIA DGX H100 and DGX A100 appliances, and HGX-baseboard systems from OEMs built around 8x H100 SXM5 or 8x A100 SXM4. The densest configuration in the secondary market and the one buyers ask for most.
- 02 /
OEM multi-GPU chassis
Dell XE-class, Supermicro, HPE and similar 4-GPU and 8-GPU PCIe and SXM chassis, including 8x A100 80GB PCIe and L40S-populated systems, sold intact with their host configuration.
- 03 /
Complete nodes and racks
Multi-node lots from cluster retirements, including switching and storage where a supplier is moving the rack as a unit. We describe the lot as it is and match it to a buyer who wants it whole.
- 04 /
Individual GPUs, when that is the fit
Loose H100, A100, L40S and other datacenter cards remain in scope, with published secondary-market ranges per model. If a system is not the right unit for your build, we quote the cards.
How server pricing works
Server pricing is quoted per requirement. A complete system is priced against its exact configuration, condition, documentation and the supply available for it, so there is no single published number that would be honest for every node.
The ranges we publish are per-GPU market references for individual cards, computed from observed secondary-market data with the methodology in the open. They are useful context for a system built around those GPUs; they are not a server valuation, and a card range multiplied by a GPU count is not a quote.
Informational only. Observed secondary-market estimates, not quotes, offers, or advice. Do not rely on these figures as the sole basis for a purchase.
Tell us the system you need.
Configuration first. The more of it you can give us, the closer the first number lands.
GPU server questions
Ready to get started?
Tell us what you need or what you have. We handle the rest, from first conversation to final payment.