# Arista Unveils Ethernet Architecture for AI Racks With 1,024 XPUs By Simon Yoon Canonical URL: https://www.tokenpost.com/news/technology/27617 Published: 2026-10-07T20:06:08.000Z Updated: 2026-10-07T20:06:08.000Z Section: Technology > The Etherlink SU-144 architecture connects up to 144 accelerators in a single-hop domain and up to 1,024 XPUs through cross-rack topologies. Arista introduced an open Ethernet architecture for rack-scale artificial intelligence systems on Oct. 7, extending networking into the tightly integrated infrastructure used to connect accelerators within and across AI racks. The Etherlink SU-144 architecture supports up to 144 accelerators, or XPUs, in a single-hop domain and up to 1,024 XPUs through cross-rack topologies. The portfolio includes three scale-up designs: an orthogonal chassis, a cabled backplane and a cross-rack system. The orthogonal-chassis design supports liquid-cooling loads ranging from 100 kilowatts to 400 kilowatts. Arista developed the rack-scale systems with AMD, Arm, Broadcom, d-Matrix, Meta, Microsoft and Qualcomm. “The network in the AI era is no longer just an independent pipe between servers. It’s now a tightly integrated rack-scale backplane that is essential to maximizing compute,” Tyson Lamoreaux, Arista’s senior vice president of cloud and AI networking, said. Arista also outlined scale-out racks built around liquid-cooled 7060EX7 switches rated at 102.4 terabits per second per switch. The racks include fiber patch panels, cooling manifolds, drip trays with active leak detection, rack-management controllers and backup-battery power shelves. The systems can run Arista EOS or an open network operating system through the common Netdi foundation. The scale-out fabric incorporates Ultra Ethernet Consortium standards, Multipath Reliable Connection resiliency, multi-plane routing, Dynamic Load Balancing and Cluster Load Balancing. Current OSFP-based racks deliver 1.6 petabits per second, while next-generation optical designs are projected to reach 6.5 petabits per second per rack. The change could reduce overall data-center footprint by nearly 50%. The portfolio is part of Arista’s Ethernet for Scale-Up Networks initiative, which covers connections within a rack, between AI racks and across larger AI environments. The approach uses common software, diagnostics and telemetry across those network types. Arista’s 7060XE7 portfolio supports 100 terabits per second of system bandwidth and 1.6 terabits per second of port throughput. The air-cooled 7060XE7-64PS/PRS systems are scheduled for the fourth quarter of 2026, while the liquid-cooled 7060XE7-64PRS-RV3-L and 7060XE7-128PE systems are scheduled for the first quarter of 2027. The announcement follows [AMD’s Helios platform for integrated AI racks](). “Arista is delivering an open, Ethernet-based framework for both scale-up and scale-out fabrics,” Lamoreaux said. ## Links in this article - [AMD’s Helios platform for integrated AI racks](https://www.tokenpost.com/news/technology/27262)