Next-Gen Scale-Up Networking for AI Fabrics
Arista Networks, Wednesday, October 7th, 2026
Arista explains why AI scale-up domains are moving from PCIe and proprietary links to open Ethernet across racks.
Arista argues that as models grow, local PCIe, CXL and proprietary interconnects can no longer serve as the primary scale-up mechanism and instead create constrained compute islands.
Operators are moving to rack-scale scale-up domains that link tens to hundreds of XPUs into a single tightly coupled system, making the network part of the compute system itself.
Arista positions Ethernet, already dominant for AI scale-out, as the open foundation for scale-up through the ESUN effort it helped launch.