Arista Networks announced a portfolio of open Ethernet rack-scale systems for high-density artificial intelligence infrastructure. Developed with partners including AMD, Microsoft, and Meta, the systems offer a standards-based alternative to proprietary interconnects for scale-up and scale-out AI networks.
The architectures are designed to meet the high power and thermal demands of AI supercomputers, using liquid cooling to support dense accelerator configurations. Arista’s Etherlink SU-144 can connect up to 144 accelerators in one rack, with larger configurations possible across multiple racks.
The portfolio addresses growing demand for open, scalable networking as AI models expand.