AI training and inference clusters are pushing network fabrics well beyond traditional data centre speeds. As GPU clusters scale out across multiple racks, 400G and 800G optics have become essential for keeping east-west traffic non-blocking and latency low. This guide covers what Australian data centres need to know when planning 400G/800G deployments for AI workloads.
Why AI clusters need 400G and 800G optics
Large-scale AI training relies on GPU-to-GPU communication across leaf-spine fabrics, where bandwidth and latency directly affect training time and cluster utilisation. 400G QSFP-DD and OSFP optics have become the baseline for new AI/ML clusters, with 800G modules increasingly deployed for spine and super-spine links as GPU counts scale into the thousands. Insufficient bandwidth between nodes creates bottlenecks that leave expensive GPU compute idle, so optics selection is a critical part of any AI infrastructure build.
Form factors: QSFP-DD, OSFP and QSFP112
QSFP-DD (Double Density) is the most widely deployed 400G form factor, doubling the electrical lanes of QSFP28 while remaining largely backward-compatible with existing QSFP28 cages. OSFP (Octal Small Form Factor Pluggable) is a newer, slightly larger form factor purpose-built for 400G and 800G, offering better thermal performance for higher-power coherent and DWDM optics. QSFP112 uses 100G-per-lane electrical signalling to deliver 400G in a smaller footprint. Your choice depends on switch port compatibility, power budget, and reach requirements.
Reach, cabling and DWDM considerations for AI fabrics
Most 400G/800G intra-rack and leaf-spine links use short-reach multimode or active optical cables (AOCs), while longer leaf-to-spine and inter-building runs typically require single-mode SR4/DR4/FR4 optics, most commonly terminated with MPO/MTP connectors as outlined in our fibre connector guide. For AI campuses spanning multiple buildings or sites, DWDM SFPs and coherent optics can extend high-bandwidth fabrics over existing dark fibre without a full re-cable, following the same wavelength-multiplexing principles covered in our DWDM SFP selection guide. Planning your fibre plant and channel allocation early avoids costly rework as GPU cluster sizes grow.
Sourcing vendor-compatible 400G/800G optics
As with earlier-generation SFP and QSFP optics, third-party 400G/800G modules that are MSA-compliant and properly vendor-coded can deliver significant savings over OEM-branded pricing without sacrificing interoperability. Always confirm firmware and platform support before ordering, and test a small batch before scaling to a full cluster rollout, following the same due diligence outlined in our SFP compatibility guide for Cisco, Arista, Juniper and Brocade. If you’re still weighing up form factors and speeds across your broader network, our SFP vs SFP+ vs QSFP28 vs QSFP-DD buyer’s guide is a useful starting point.
Tikka Consulting supplies vendor-compatible 400G and 800G optics for AI and HPC data centre fabrics across Australia, backed by local stock and technical support. Contact our team to discuss optics, DWDM channel planning, and cabling for your next AI infrastructure build.
