# Arista Networks unleashes ESUN-based rack-scale architecture targeting AI infra > Source: > Published: 2026-10-08 19:36:42+00:00 Arista Networks unveiled an open Ethernet rack-scale architecture based on the Ethernet for Scale-up Networks (ESUN) initiative and with input from a handful of large industry names, with that architecture designed to power high-density scale-up and scale-out AI infrastructure. The Etherlink SU-144 architecture can support up to 144 XPU accelerators in a single-hop network and can scale to 1,024 XPUs through cross-rack topologies. That architecture is being offered through a trio of physical form factors to match user demand. Those form factors include a single-hop, single-rack orthogonal chassis architecture designed for “extreme bandwidth and density in intensive AI training and inference workloads.” This includes use of 100- to 400-kilowatt liquid cooling support. There is also a cabled backplane iteration that combines with a modular rack structure optimized for “signal integrity, performance, and long-term operational efficiency.” The final option is a cross-rack system that provides a modular standalone fabric to support greater deployment and upgrade flexibility. Arista cited development work on the architecture with AMD, Arm, Broadcom, d-Matrix, Meta, Microsoft, and Qualcomm. The scale-out architecture is built on Arista’s Network Diagnostic Infrastructure (Netdi) platform that can support the vendor’s EOS or an open network operating system (NOS). It uses Ultra Ethernet Consortium (UEC) standards, Multipath Reliable Connection (MRC) fabric resiliency, multiplane routing, and intelligence dynamic load and cluster load balancing. The rack design features multiple liquid-cooled Arista 7060EX7 switches with 102.4 Tb/s capacity per switch, fiber patch panels, integrated liquid-cooling manifolds, drip trays with active leak detection, rack management controllers, and power shelves with backup battery units. ## Big hitters are the target Arista noted the architecture is targeted at “cloud titans and enterprise innovators,” a target confirmed by Bob Laliberte, principal analyst and founder of Liberte Research Group. “From an enterprise perspective, near-term adoption will likely be concentrated among hyperscalers, AI infrastructure providers, and neoclouds,” Laliberte [noted in a LinkedIn post](https://www.linkedin.com/feed/update/urn:li:activity:7513588999396122625/). “However, the broader direction has wider relevance. Open, standards-based Ethernet is emerging as a credible option for both scale-up and scale-out AI fabrics, giving infrastructure operators more choice in how they connect diverse accelerators.” Dell’Oro Group [did note in a recent report](https://www.sdxcentral.com/news/ethernet-dominates-ai-networking-as-switch-sales-double-but-infiniband-rebounds/) that Ethernet switch sales for AI back-end networks doubled during the first quarter of this year and accounted for around two-thirds of data center switch sales in AI clusters during that quarter. That same report cited Arista as the No. 3 player in that space, behind market heavyweights Celestica and Nvidia. “The competitive discussion is moving beyond port speeds and switch capacity,” Laliberte added on the overall market. “It is now about who can deliver a validated, operationally manageable rack-scale system that puts expensive accelerators to work quickly and reliably.”