400G Ethernet has fallen below the price premium that kept it hyperscaler-only. The question for enterprise architects in 2026 is no longer whether 400G works — it is which workloads justify pulling the trigger now.
The economics crossed over
Per-bit, a 400G QSFP-DD port now costs less than four 100G ports once you count switch slots, optics, fibre pairs and power. For greenfield spines the decision is effectively made: deploying 100G spines in 2026 locks in higher cost per bit for the life of the fabric.
Workloads that justify it today
- AI/ML clusters — even modest 8-node training pods saturate 100G uplinks during all-reduce phases.
- NVMe-oF storage — flash arrays speak at line rate; the network should not be the array’s bottleneck.
- East-west heavy virtualization — VM and container density has quietly tripled typical leaf utilization since the last refresh cycle.
- Backup and DR windows — 400G turns multi-hour replication windows into minutes, which changes RPO conversations.
Migration without forklift
The pragmatic path is spine-first: deploy 400G spines (64×400G in 2U is now standard), connect existing 100G leaves via breakout, then refresh leaves as server NICs migrate to 200/400G. Rail-optimized designs for GPU pods can coexist with a classic leaf-spine for general compute — they share the spine, not the philosophy.
Buy the spine for the fabric you will run in 2029, and the leaves for the servers you run today.
What to specify
- QSFP-DD over OSFP for enterprise reach unless you have a specific 800G roadmap commitment.
- Cut-through switching and telemetry (INT/sFlow) as defaults, not options.
- DR4/FR4 optics matched to your actual fibre plant — audit it before ordering.
PEXON supplies Cisco Nexus, Arista and NVIDIA Spectrum 400G platforms with compatible optics and cabling, staged and labelled per rack elevation.