AMD has published its 2023–2027 data center networking roadmap, dividing networking into Scale-Up and Scale-Out domains with the goal of moving from a single-node 8-GPU architecture to 72-GPU and ultimately 256-GPU supernode configurations. In scale-up, the 2023–2025 period uses Universal Baseboard to interconnect 8 GPUs on a PCB, with bandwidth rising from 7x128 GB/s to 7x153.6 GB/s and channel speeds from 32 Gb/s to 38 Gb/s. In 2026, AMD will introduce the MI455X, increasing the logical GPU count to 72 with single-layer all-to-all switching, copper DAC cables with active retimers, per-GPU bandwidth of 7x800 GB/s, channel speeds of 200 Gb/s, and a dedicated switch chip delivering 102.4 Tb/s aggregate bandwidth. By the MI500X era in 2027, AMD plans 256 GPUs using a 3-rack Pod architecture, with co-packaged copper or co-packaged optics and per-GPU bandwidth of 2,400 GB/s. In scale-out, AMD continues the Pollara 400G NIC from 2023 through 2026, upgrading to the Pensado 800G in 2027; optical modules remain at 400G DR4/DR8 through 2026 and upgrade to 800G DR4 in