Buying Decommissioned A100 and H100 Clusters — Any Size

Required Configuration

GPU platforms: HGX A100 (40GB or 80GB) and HGX H100 8-GPU systems

Lot size: From 8 servers to complete multi-rack clusters

Fabric: InfiniBand or Ethernet fabrics purchased with the compute or separately

Condition: Working, tested used; units with minor faults considered at adjusted pricing

Location: Specify current location; in-place removal and logistics can be arranged

Availability: Provide the planned decommissioning date

Looking for NVIDIA Spectrum-X 800GbE Switches and BlueField-3 SuperNICs

Required Configuration

Switches: NVIDIA Spectrum-4 SN5600, 64-port 800GbE — 8 to 16 units

SuperNICs: NVIDIA BlueField-3 B3140H or B3220, 400Gb/s — up to 128 units

Optics: 800G and 400G OSFP transceivers matched to ports

Software: Cumulus Linux or SONiC — state which is installed

Condition: New preferred

Availability: Ready to ship, or provide a confirmed lead time

Sourcing Multi-GPU AI Workstations for an R&D Team

Required Configuration

GPU configuration: 2 to 4× NVIDIA RTX PRO 6000 Blackwell Workstation Edition or RTX 6000 Ada 48GB per system

CPU platform: AMD Threadripper PRO or Intel Xeon W

Memory: 256GB to 512GB ECC

Storage: 4TB+ NVMe boot and scratch

Form factor: Tower or desk-side, standard office power

Quantity: 6 to 12 systems

Condition: New

Availability: Ready to ship, or provide a confirmed lead time

Looking for NVIDIA GH200 Grace Hopper Superchip Servers

Required Configuration

Platform: NVIDIA GH200 Grace Hopper Superchip (96GB HBM3 or 144GB HBM3e)

Form factor: 1U or 2U MGX servers, 1 to 4 superchips per server

Networking: ConnectX-7 or BlueField-3, 400Gb/s

Quantity: 8 to 32 superchips across the order

Condition: New or tested used

Availability: Ready to ship, or provide a confirmed lead time

Sourcing Liquid Cooling CDUs, High-Density Racks and PDUs

Required Configuration

Coolant distribution units: In-row or in-rack CDUs, 1MW total capacity across the order

Racks: 48U or 52U, 800mm wide, rated for 130kW+ liquid-cooled loads — 16 to 32 units

PDUs: 415V three-phase, 60A or higher, metered and switched

Rear-door heat exchangers: Considered as an alternative for air-cooled racks

Condition: New

Availability: Provide a confirmed lead time

Looking for DDR5 RDIMMs and Enterprise NVMe Drives

Required Configuration

Memory: DDR5-4800 or DDR5-5600 ECC RDIMM, 64GB or 96GB modules — 1,000 to 3,000 modules

Drives: Samsung PM9A3, Micron 7450 PRO or 9400 PRO NVMe, 3.84TB to 15.36TB — 200 to 600 drives

Brands: Samsung, Micron or SK hynix only

Condition: New or new pulls; no used memory

Availability: Ready to ship

Sourcing All-NVMe Storage Servers for AI Training Data

Required Configuration

Platform: 1U or 2U all-NVMe storage servers, 12 to 24 drive bays

Drives: Enterprise PCIe Gen4 or Gen5 NVMe, 15.36TB or 30.72TB

Capacity: 1PB to 2PB raw across the full order

Networking: 2× 200Gb/s or 400Gb/s (InfiniBand or Ethernet) per node

Software: Hardware-only acceptable; WEKA-, VAST- or DDN-certified configurations preferred

Condition: New or tested used

Availability: Ready to ship, or provide a confirmed lead time

Looking for NVIDIA L40S and RTX PRO 6000 Blackwell Server Edition GPUs

Required Configuration

Option A: NVIDIA L40S 48GB GDDR6, passive — 40 to 160 cards

Option B: NVIDIA RTX PRO 6000 Blackwell Server Edition 96GB GDDR7, passive — 16 to 64 cards

Interface: PCIe Gen4 x16 (L40S) / PCIe Gen5 x16 (RTX PRO 6000)

Condition: New preferred; new open-box considered

Availability: Ready to ship, or provide a confirmed lead time

Sourcing a 1,024-GPU NVIDIA H100 Training Cluster

Required Configuration

Compute: 128× 8-GPU HGX H100 80GB SXM5 servers (1,024 GPUs total)

Fabric: Non-blocking NDR 400Gb/s InfiniBand compute fabric, rail-optimized

Storage: High-performance parallel storage, 2PB usable or greater

Racks: Fully racked, cabled and labeled preferred

Condition: New or recently decommissioned with complete service history

Location: Specify current inventory location; in-place deployments considered

Availability: Ready to ship, or provide a confirmed lead time

Looking for NVIDIA Quantum-2 NDR InfiniBand Switches, Adapters and Optics

Required Configuration

Switches: NVIDIA Quantum-2 QM9700 or QM9790, 64-port NDR 400Gb/s — 12 to 20 units

Adapters: NVIDIA ConnectX-7 NDR 400Gb/s, single-port OSFP — up to 256 units

Transceivers: NDR OSFP twin-port and single-port optics, matched to switch and adapter ports

Cabling: Multimode or single-mode fiber, lengths to be confirmed

Condition: New preferred; tested used considered for switches

Availability: Ready to ship, or provide a confirmed lead time

Cookies on GPU Exchange

We use cookies to see how the site gets used and to improve it. Analytics only run if you accept.