null
Server & Workstation RAM at Wholesale Volume Pre-tested Ships in 1–2 business days

GPU Servers

Refurbished Dell PowerEdge and HPE ProLiant servers built to host NVIDIA and AMD GPUs for AI training, machine learning, and high-throughput inference. Every unit is custom-configured to your GPU, CPU, memory, and storage spec — fully tested and warrantied. Configure a build below or request a quote.

Compare models

Most popular GPU servers

Choose a pre-configured build, or build your own on a chassis with the CPUs, memory and storage you need.

  • 1-year warranty, 90-day returnsParts, labor and shipping; up to 5 years
  • Tested in MichiganAssembled and tested before shipping
  • Free shipping in the lower 48Most builds ship in 1–2 business days
ModelSpecsBest forPre-configuredBuild your own
HPE ProLiantGen10 · 2UModel page › Specs2× Xeon Scalable 1st/2nd gen24 DIMM slots · DDR4Up to 24 × 2.5″ or 12 × 3.5″ Best forAll-round 2U with many storage options Pre-configuredSee builds ›  Build your ownSee chassis › 
Dell PowerEdge14th gen · 2UModel page › Specs2× Xeon Scalable 1st/2nd gen24 DIMM slots · DDR4Up to 16 × 2.5″ or 8 × 3.5″ Best forAll-round 2U; room for GPUs Pre-configuredSee builds ›  Build your ownSee chassis › 
Dell PowerEdge15th gen · 2UModel page › Specs2× 3rd Gen Xeon Scalable32 DIMM slots · DDR4Up to 24 × 2.5″ or 12 × 3.5″ Best forNewer 2U platform with PCIe 4.0 Pre-configured—  Build your ownSee chassis › 
HPE ProLiantGen9 · 2UModel page › Specs2× Xeon E5-2600 v3/v424 DIMM slots · DDR4Up to 24 × 2.5″ + 2 rear Best forBudget 2U with lots of drive bays Pre-configuredSee builds ›  Build your ownSee chassis › 
HPE ProLiantGen10 · Tower or 5U rackModel page › Specs2× Xeon Scalable 1st/2nd gen24 DIMM slots · DDR4Up to 24 × 2.5″ or 12 × 3.5″ Best forExpandable 2P tower that converts to rack Pre-configuredSee builds ›  Build your ownSee chassis › 
Dell PowerEdge13th gen · 2UModel page › Specs2× Xeon E5-2600 v3/v424 DIMM slots · DDR4Up to 16 × 2.5″ or 8 × 3.5″ Best forBudget 2U, homelab; GPU-capable Pre-configuredSee builds ›  Build your ownSee chassis › 
Dell PowerEdge14th gen · 2UModel page › Specs4× Xeon Scalable 1st/2nd gen48 DIMM slots · DDR4Up to 24 × 2.5″ Best for4-socket 2U for data analytics Pre-configured—  Build your ownSee chassis › 
Not sure which model fits? Send us your requirements.Request a quote
Price on request

More models, configured to your spec

By generation, newest first. Open a model to pick a chassis and configure it, or send us the spec and we reply with a quote.

Dell 14th gen

Refurbished GPU servers from PCSP put NVIDIA accelerator hardware to work for AI training, LLM inference, machine learning, rendering, and HPC — at a fraction of new-OEM pricing. Every system is a Dell PowerEdge or HPE ProLiant accelerator platform, custom-configured to your GPU count, CPU, memory, and storage, fully tested and backed by warranty. In stock builds ship fast; anything not in stock is available to configure by quote.

Shop GPU servers by model

Dell PowerEdge GPU servers

HPE ProLiant GPU servers

How to choose a GPU server

Inference vs training. For serving LLMs and running inference, a single-GPU or dual-GPU 2U server (R760xa, R750xa, DL380a) with L40S or A100 40/80GB is usually the cost-per-token winner. For fine-tuning and training, favor a multi-GPU chassis (C4140, R7525, or a 4×-GPU R760xa) with NVLink-class bandwidth and more total VRAM.

Match VRAM to model size. A single 48GB L40S or A100 comfortably serves quantized models up to ~30B; 70B-class models want 2 GPUs or 80GB cards; 100B+ MoE and multi-model serving want 4-GPU nodes. Add up the VRAM you need first, then pick the chassis that holds those cards and their power/cooling.

Generation & PCIe. Newer platforms (R760xa, DL380a Gen11/Gen12) bring PCIe Gen5 and DDR5 for the highest GPU bandwidth; previous-gen (R750xa, R7525, DL385 Gen10) deliver PCIe Gen4 at a much lower price and remain excellent for inference.

Which GPUs fit

These chassis accept data-center accelerators including the NVIDIA L40S, L40, A100 (40GB / 80GB), H100, and prior-gen A30 / A40 / V100, plus professional RTX cards depending on the platform and power envelope. Not sure which GPU your workload needs? Request a quote and our engineers will spec it — including GPUs that are not listed on the site.

Why buy a refurbished GPU server

Refurbished enterprise GPU platforms deliver the same silicon at 40–70% less than new, with the same testing and warranty. With DRAM and accelerator supply diverted to hyperscaler AI build-outs, refurbished hardware sidesteps the shortage and lead times entirely — you can build a fully configured GPU inference server today for what the memory alone would cost on a new-build quote. Every PCSP system is fully tested, backed by a 1-year warranty (extendable to 5), and ships fast from the US.

Training and inference nodes are read-heavy on the dataset volume and write-heavy on checkpoints, which pulls the array in two directions. What is a RAID array? covers where striping helps and where parity write penalties will cost you throughput; the server buying guide handles the CPU, memory and power side of a GPU build.

Frequently Asked Questions

What is a GPU server?

A GPU server is an enterprise rack or dense-node platform engineered to power, cool, and connect multiple data-center GPUs — with the PCIe lanes, power delivery, and airflow that standard servers lack. It is the hardware foundation for AI training, LLM inference, and GPU-accelerated compute.

Which GPU server is best for LLM inference?

For most inference workloads, a Dell R760xa or R750xa (or HPE DL380a) with one or two NVIDIA L40S or A100 GPUs offers the best cost-per-token. The L40S in particular wins on price/performance for serving. Move to a 4-GPU C4140 or R7525 when you need to serve very large or multiple models.

How many GPUs can these servers hold?

The R750xa and R760xa hold up to 4 double-width GPUs; the R7525 up to 3; the C4140 up to 4 in a dense 1U; the HPE DL380a is built for multiple accelerators. Exact limits depend on GPU model and power configuration — we confirm your build before it ships.

Can I configure a GPU server that is out of stock?

Yes. Any GPU server not currently in stock is available to build to your spec by quote — request a quote and we will configure and price it.

Is refurbished GPU hardware reliable for 24/7 AI workloads?

Yes. Enterprise servers are designed for continuous operation; every PCSP system is fully tested and warrantied. Refurbished platforms run production inference and training reliably around the clock.

Build your own configuration online, or request a quote for a custom GPU server spec.