NVIDIA • Dell PowerSwitch • 100G–800G
AI Networking & High-Speed Data Center Switches
Enterasource sources, configures, and warranties the switching that AI clusters and modern data centers run on — NVIDIA Quantum InfiniBand, NVIDIA Spectrum Ethernet, and Dell PowerSwitch Z-series, from 100G up to 400G and 800G. Tell us the workload and the scale; we'll build the fabric to your exact specification, quote the full bill of materials, and back it with a 2-year warranty and free freight to the lower 48.
Refurbished or new. Built to your spec. Quoted fast.
The lineup: NVIDIA Quantum, NVIDIA Spectrum & Dell PowerSwitch
We supply the full AI networking stack — NVIDIA Quantum InfiniBand switches (formerly Mellanox) for the compute fabric, NVIDIA Spectrum and Dell PowerSwitch Ethernet for GPU-cluster and data-center networking, and the management leaf that ties a GPU pod together. Every InfiniBand and Ethernet switch below is configured to your BOM and available refurbished or new. Specs are per-switch; link through for the full datasheet on each.
800G switches — XDR InfiniBand & Spectrum-4 / 800GbE Ethernet
- NVIDIA Quantum-X800 Q3400-RA — 800G InfiniBand XDR. 4U, 144 × 800Gb/s ports across 72 OSFP cages, 115.2 Tbps aggregate, with fourth-generation NVIDIA SHARP in-network compute (FP8). The top of the InfiniBand fabric for the largest GPU training clusters.
- NVIDIA Spectrum SN5610 — 800G Ethernet (Spectrum-4). 2U, 64 × 800GbE OSFP plus SFP28 management, 51.2 Tbps. The Spectrum-X Ethernet spine for GPU clusters standardizing on Ethernet.
- NVIDIA Spectrum SN5600D — 800G Ethernet (Spectrum-4). 2U, 64 × 800GbE OSFP, 51.2 Tbps, DC-powered for high-density racks.
- Dell PowerSwitch Z9864F-ON — 800GbE spine. 2U, 64 × 800GbE OSFP112 + 2 × 10GbE SFP+, 51.2 Tbps non-blocking. Runs Dell SmartFabric OS10 or Enterprise SONiC.
400G switches — NDR InfiniBand & 400GbE Ethernet
- NVIDIA Quantum-2 QM9700 — 400G InfiniBand NDR. 1U, 64 × 400Gb/s ports across 32 OSFP cages, 51.2 Tbps, with an on-board subnet manager (internally managed) and SHARPv3 in-network compute.
- NVIDIA Quantum-2 QM9790 — 400G InfiniBand NDR. The externally managed QM9700: the same 1U, 64-port, 51.2 Tbps switch, driven by NVIDIA UFM elsewhere in the fabric — the choice for scaling many switches under one manager.
- NVIDIA Spectrum SN5400 — 400G Ethernet. 2U, 64 × 400GbE QSFP-DD + 25GbE SFP28, 25.6 Tbps. High-radix 400G Ethernet leaf/spine.
- NVIDIA Spectrum SN4700 — 400G Ethernet. 1U, 32 × 400GbE QSFP-DD, 12.8 Tbps. Compact 400G top-of-rack / leaf.
- Dell PowerSwitch Z9432F-ON — 400GbE spine. 1U, 32 × 400GbE QSFP56-DD + 2 × 10GbE SFP+, 25.6 Tbps non-blocking. Dell OS10 or SONiC.
- Dell PowerSwitch Z9664F-ON — 400GbE spine. 2U, 64 × 400GbE QSFP56-DD + 2 × 10GbE SFP+, 51.2 Tbps non-blocking. High-density Dell 400G aggregation.
Management & leaf
- NVIDIA Spectrum SN2201 — 1G management leaf. 1U, 48 × RJ45 1GBASE-T + 4 × QSFP28 100GbE, 448 Gbps. The out-of-band management switch every GPU pod needs.
See the full Dell & NVIDIA switch range →
Spec comparison
| Model | Fabric | Speed | Ports | RU | Switching capacity | Connector |
|---|---|---|---|---|---|---|
| Quantum-X800 Q3400-RA | InfiniBand XDR | 800G | 144 | 4U | 115.2 Tbps | OSFP |
| Spectrum SN5610 | Ethernet (Spectrum-4) | 800G | 64 | 2U | 51.2 Tbps | OSFP |
| Spectrum SN5600D | Ethernet (Spectrum-4) | 800G | 64 | 2U | 51.2 Tbps | OSFP (DC power) |
| Dell Z9864F-ON | Ethernet | 800GbE | 64 | 2U | 51.2 Tbps | OSFP112 |
| Quantum-2 QM9700 | InfiniBand NDR | 400G | 64 | 1U | 51.2 Tbps | OSFP (internally managed) |
| Quantum-2 QM9790 | InfiniBand NDR | 400G | 64 | 1U | 51.2 Tbps | OSFP (UFM-managed) |
| Spectrum SN5400 | Ethernet | 400G | 64 | 2U | 25.6 Tbps | QSFP-DD |
| Spectrum SN4700 | Ethernet | 400G | 32 | 1U | 12.8 Tbps | QSFP-DD |
| Dell Z9432F-ON | Ethernet | 400GbE | 32 | 1U | 25.6 Tbps | QSFP56-DD |
| Dell Z9664F-ON | Ethernet | 400GbE | 64 | 2U | 51.2 Tbps | QSFP56-DD |
| Spectrum SN2201 | Ethernet (mgmt) | 1G + 100G | 48+4 | 1U | 448 Gbps | RJ45 + QSFP28 |
Why buy AI networking from Enterasource
We build AI fabrics to spec and stand behind them. Every switch is sourced, configured to your bill of materials, tested in our lab, and shipped ready to rack — with a warranty and freight most resellers don't include.
- Built to your specification. NVIDIA Quantum InfiniBand, NVIDIA Spectrum Ethernet, and Dell PowerSwitch Z-series — refurbished or new, configured to the exact port count, optics, cabling, and firmware your fabric needs.
- Tested and configured in our lab. Each switch is inspected, tested, and configured to your BOM before it ships — cages matched to the right optics and DACs, not a box of mismatched parts.
- 2-year Enterasource warranty. Standard on every switch — longer coverage than the typical 1-year secondary-market warranty.
- Free freight to the lower 48. On the switches, included.
- One BOM for the whole fabric. We quote the spine, the leaf, the optics and cables, and the out-of-band management switch together — so a spine-leaf or GPU-fabric build arrives as one coordinated order.
AI training clusters, GPU fabric, HPC & data-center spine-leaf
These switches are built for the networks between GPU nodes — the high-bandwidth, low-latency fabric that keeps accelerators fed during training and inference.
- AI/ML training clusters — NDR (400G) and XDR (800G) InfiniBand, or Spectrum-X 400/800G Ethernet, rail-optimized to the GPUs so collective operations don't stall.
- GPU fabric — the east-west compute fabric that lashes GPU servers into one training system; in-network compute (NVIDIA SHARP) offloads reductions from the GPUs.
- HPC — InfiniBand fabrics for MPI workloads where latency and non-blocking bandwidth decide job time.
- Data-center spine-leaf — Dell PowerSwitch Z-series and NVIDIA Spectrum for non-blocking 100–800GbE spine-leaf, running Dell SmartFabric OS10, Enterprise SONiC, or NVIDIA Cumulus Linux.
Building the compute nodes too? See our guide to GPU servers for AI — the fabric on this page connects the nodes that guide helps you spec.
Build your cluster fabric — we quote the full BOM
Send us the shape of the cluster — GPUs per node, node count, oversubscription target, and your longest cable run — and a networking specialist will scope the spine, leaf, optics, cabling, and management switch into a single quote. Not sure whether InfiniBand or Ethernet fits? We'll walk it through with you.
We'll only use your details to respond to your enquiry.
Frequently asked questions
Do you sell NVIDIA Quantum and Spectrum switches?
Yes. We supply NVIDIA Quantum-2 (QM9700 / QM9790) and Quantum-X800 (Q3400-RA) InfiniBand switches — the line formerly branded Mellanox — and NVIDIA Spectrum Ethernet switches (SN5610, SN5600D, SN5400, SN4700, SN2201), alongside Dell PowerSwitch Z-series. Each is configured to your specification and backed by a 2-year Enterasource warranty.
Can you supply a specific 800G or 400G switch and configuration?
Yes — tell us the model and the fabric you're building and we'll configure and quote it. We source and configure NVIDIA Quantum/Spectrum and Dell PowerSwitch Z-series to your exact port count, optics, cabling, and firmware, refurbished or new. Send the requirement and a networking specialist will scope it.
What is an InfiniBand switch?
An InfiniBand switch is a low-latency, high-bandwidth fabric switch built for the network between compute nodes — most often GPU servers in an AI or HPC cluster. It uses RDMA and, on NVIDIA Quantum switches, in-network compute (SHARP) to offload collective operations from the GPUs. Current NVIDIA Quantum-2 runs at 400G NDR and Quantum-X800 at 800G XDR.
InfiniBand or Ethernet for AI?
Both work for AI at scale; the right choice depends on your stack. InfiniBand (NVIDIA Quantum) leads on latency and in-network compute for large training fabrics; Spectrum-X and Dell PowerSwitch Ethernet give you a standards-based fabric your team already knows, now at 400G and 800G. We'll help you decide during the quote — and our guide, InfiniBand vs Ethernet: how to choose an AI fabric switch, covers the full comparison.
What is NVIDIA Spectrum-X?
Spectrum-X is NVIDIA's Ethernet networking platform tuned for AI workloads, pairing Spectrum-4 switches (such as the SN5600-series at 800G) with adaptive routing and congestion control to deliver InfiniBand-class behavior over Ethernet. It's the path for AI clusters that want to stay on Ethernet end to end.
Do these switches come with a warranty?
Yes — every switch ships with a 2-year Enterasource warranty as standard, longer than the typical 1-year secondary-market coverage. Extended terms are available; ask your specialist when we quote.
Can you quote a full cluster fabric, not just one switch?
Yes. Send us GPUs per node, node count, oversubscription target, and your longest cable run, and we'll quote the spine, leaf, optics, DAC/AOC cabling, and the out-of-band management switch as a single bill of materials — with cages matched to the right optics so it racks cleanly.
