Arista Networks pushes Ethernet into AI racks with Oct. 7 design

Arista has spent years selling high-speed Ethernet switches that direct traffic inside data centers, plus its Extensible Operating System (EOS) software for managing those connections. The company's pitch is that as AI models grow larger, GPUs stall unless data moves fast enough between chips, servers and storage — a bottleneck that wastes expensive computing power.

The Oct. 7 architecture marks a shift in scope. Rather than only linking separate servers and racks, Arista is designing networks that let large groups of AI processors communicate closely, positioning Ethernet as an alternative to proprietary networking systems. The company says the Etherlink SU-144 design handles single-hop connections among up to 144 AI accelerators and scales across racks to support up to 1,024 accelerators.

That ambition sits alongside a product cycle: June's 7060XE7 Series supports 1.6-terabit-per-second connections and, Arista says, linear pluggable optics that can cut interconnect power consumption by roughly 60% versus traditional pluggable optics — relevant for operators already facing heavy electricity demands. Arista still faces proprietary networking technologies and the risk that major cloud customers build their own equipment.