Arista introduces Etherlink SU-144 for rack-scale Ethernet AI fabrics

Ethernet moves deeper into the rack

Arista announced an Ethernet portfolio for rack-scale AI infrastructure on October 7, 2026. The company describes Etherlink SU-144 as a scale-up architecture supporting up to 144 accelerators in a single-hop network and up to 1,024 through cross-rack topologies. These figures describe Arista’s architecture, rather than results from a LANGOM deployment.

Three physical approaches are outlined: an orthogonal chassis, a cabled backplane and a modular cross-rack system. Arista specifies 100–400 kW liquid-cooling support for the orthogonal chassis. Its scale-out rack designs combine liquid-cooled 7060XE7 switches, rated at 102.4 Tbps each, with cooling manifolds, active leak detection, rack management and battery-backed power shelves.

Arista says the scale-out architecture shares a Netdi foundation and can run its EOS software or an open network operating system.

Source: Arista Announces Rack-Scale Architectures for Scale-Up and Scale-Out AI Fabrics

LANGOM perspective

The practical change for an infrastructure specification is the boundary of the network package. A switch-only comparison is insufficient when the proposed design includes the rack, fluid connections, power and management controls. Ask bidders to identify which components are mandatory for the stated topology and which remain the operator’s responsibility.

Keep scale-up and scale-out requirements separate in the acceptance plan. Document the accelerator mix, connection layout, failure behavior and software versions for the exact system. Validate an intended cross-rack expansion using its actual cable lengths and service procedures, rather than assuming a single-rack result transfers unchanged.

Facilities review should establish the available electrical capacity, coolant conditions and maintenance access before committing to density targets. The stated cooling-support range is not a substitute for a configuration-specific load calculation. For operations, choose the network operating system alongside the telemetry and support model, and include fault recovery in the trial. These are LANGOM’s evaluation considerations; no delivery date or customer savings is inferred from the announcement.

← All industry news