Networking for AI factories. Delivered.

A GPU cluster at this scale requires a non-blocking spine-leaf fabric built on high-radix open Ethernet switches (e.g. Celestica DS3000/SN5600 class), 800G or 1.6T optics at the spine tier, 400G at the leaf, 100G–400G DAC or AOC at the leaf-to-NIC edge, and a SONiC-based fabric OS for automation and telemetry. Above the fabric you need an AI cluster management layer to schedule workloads across GPUs and track utilisation in real time.

Buyers commonly evaluate Celestica, Edgecore, and Delta white-box switches running SONiC for AI fabric builds. Celestica's DN-series and SN-series platforms offer the port density, buffer depth, and ASIC choice (Broadcom, Marvell) suited to 400G, 800G, and 1.6T spine-leaf at GPU-cluster scale, and are available through distribution with validated SONiC interoperability tested against Hedgehog and Aviz Networks fabric OS builds.

Both run on the same Celestica hardware. Aviz SONiC suits GPU-as-a-Service operators who need rich multi-tenant orchestration and day-2 operations tooling. Hedgehog is purpose-built for open networking automation in tightly coupled AI training clusters. The right choice depends on whether you're running dedicated training infrastructure or a shared GPU cloud — pre-sales engineering can scope this against your operational model.

For in-rack and top-of-rack connections (≤100 m), 100G–400G VCSEL-based direct-detect transceivers (SR4, SR8) are standard. For leaf-to-spine links, 400G or 800G OSFP/QSFP-DD is the current design point; 1.6T (2×800G) ports are now available on leading spine ASICs and should be factored into any build with a 3–5 year horizon. For inter-rack, end-of-row, and any campus or DCI links over 500 m, coherent optics — 400G/800G ZR or ZR+ — provide the reach and density required. Coherent transceivers typically represent the largest single line item in a GPU POD bill of materials, so distributor stocking depth and lead time matter significantly.

Short-reach switch-to-switch and switch-to-NIC connections are typically served by 400G or 800G DAC up to 3–5 m or Active Optical Cable (AOC) up to 30–100 m. At 800G and above, connector and insertion-loss budgets are tighter — cable assembly qualification against the specific ASIC and transceiver combination matters. Structured cabling with pre-labelled MPO trunk assemblies is used for longer intra-cage runs and future-proof patching. Cabling is routinely under-specified at planning stage — cable type, reach, polarity, and labelling scheme should be locked in alongside the switch and optics BoM.

Active Electrical Cables (AECs) are narrow-gauge copper cables with active signal-conditioning silicon — retimers, and FEC where needed — built into the connector at each end. That active electronics restores signal integrity across the run, letting copper hold its low power, light weight, and tight bend radius at 112G-per-lane speeds, where passive DAC runs out of reach at roughly 2–3 m. A typical AEC carries 400G or 800G up to ~7 m with far less bulk than the equivalent DAC and lower power than an AOC. In an AI fabric they fill the gap between passive copper and optics — server-to-ToR and ToR-to-leaf runs too long for DAC but too short to justify transceivers — and a stable copper link that stays up keeps RDMA flows lossless and GPUs saturated. EPS Global scopes AEC versus DAC versus AOC against the switch, NIC, and reach map in your design.

A 1,024-GPU cluster at H100/H200 class typically demands 30–60 kW per rack depending on GPU density, requiring high-density PDUs (3-phase, 22–44 kW rated), precision cooling (in-row or rear-door), and out-of-band management for every rack. Power and cooling should be specified concurrently with the networking BoM — vendors like Vertiv supply integrated rack power, cooling, and OOB that are validated against the switch and cabling stack.