Arista Expands AI Networking Strategy to Rack-Scale Ethernet Architectures

Arista Networks Inc. is expanding its artificial intelligence networking portfolio beyond standalone switches, introducing an open, rack-scale Ethernet architecture designed to act as a unified backplane for AI data centers. Unveiled ahead of the Open Compute Project (OCP) Global Summit in San Jose, the strategy shifts the company’s focus toward pre-integrated compute and networking systems engineered for large-scale AI workloads, as reported by SiliconANGLE (https://siliconangle.com/2026/10/07/arista-takes-ethernet-rack-network-becomes-backplane/).

At the core of the announcement is the Etherlink SU-144, an Ethernet-based scale-up architecture built on the Ethernet for Scale-Up Networking (ESUN) initiative. The system connects up to 144 accelerator chips (XPUs) within a single-hop domain and scales up to 1,024 XPUs across multiple racks. Arista is developing the technology alongside an ecosystem of partners including Advanced Micro Devices Inc., Arm Ltd., Broadcom Inc., d-Matrix Corp., Meta Platforms Inc., Microsoft Corp., and Qualcomm Corp.

For scale-out fabrics, Arista is introducing liquid-cooled rack reference designs in eight-, 12-, and 16-switch configurations based on its 7060EX7 switches, which provide 102.4 terabits per second of throughput. The rack systems incorporate cooling manifolds, active leak detection drip trays, rack management controllers, power shelves with battery backup, and fiber patch panels. Integration partners will assemble, test, and deliver the configured racks directly to data center operators.

The rack architecture supports Arista EOS or open network operating systems on a shared Netdi foundation. It complies with Ultra Ethernet Consortium specifications and features Multipath Reliable Connection resiliency, multi-plane routing, and cluster load balancing. Arista also stated that next-generation optical technology, including XPO pluggables and open co-packaged and near-packaged optics, will boost rack capacity from 1.6 petabits per second to 6.5 petabits per second, cutting data center footprint requirements by nearly half.

Arista's scale-up expansion specifically targets non-Nvidia accelerator environments rather than seeking to replace Nvidia Corp.'s proprietary NVLink interconnect. During an analyst briefing, Vijay Vusirikala, Arista’s distinguished lead for AI systems and networks, noted that there is "no near-term opportunity" to dislodge NVLink in proprietary Nvidia clusters, positioning the new architecture around heterogeneous accelerators such as AMD hardware and Microsoft's Maia chips.

The transition also highlights a shift toward factory-integrated reference architectures. With a 16-switch configuration requiring up to 2,000 fiber connections, on-site assembly introduces risks of cable miswiring, bend-radius violations, and damaged optics. Arihant Jain, senior product line manager at Arista, explained during the briefing that "the whole old approach of spare parts or things assembled at site has to change," noting that co-designed liquid cooling and leak response systems require tightly coupled facility and network engineering.

Arista plans to demonstrate a fully assembled liquid-cooled rack at the OCP Global Summit running Oct. 12-15 in San Jose, where co-founder Andy Bechtolsheim will deliver a keynote on Oct. 13 titled "The Gigawatt AI Data Center Era." Initial adoption is expected among hyperscalers, neoclouds, and AI infrastructure providers deploying heterogeneous compute clusters.