400 Gigabit Ethernet

400 Gigabit Ethernet (400GbE) is defined by IEEE 802.3bs (2017). It uses PAM4 modulation across 16 or 8 electrical/optical lanes and is deployed as the spine-to-leaf and GPU server uplink speed in hyperscale AI factories, providing sufficient bandwidth for RoCE v2-based GPU-to-GPU traffic over standard Ethernet infrastructure.

800GbE, standardised in IEEE 802.3df (2023), doubles throughput to 800 Gb/s and is expected to replace 400GbE as the AI factory spine-leaf uplink from 2025–2026, matching the bandwidth demands of NDR InfiniBand-equivalent Ethernet fabrics at scale. NVIDIA's Spectrum-X platform and Broadcom's Tomahawk 5 ASIC target this transition. For operators planning AI factory infrastructure, the choice between 400GbE and 800GbE switching is a near-term capital procurement decision.