- Arista is moving deeper into AI data center networking with Ethernet-based scale-up and scale-out architectures
- Its new liquid-cooled Etherlink SU-144 design targets high-density AI racks with up to 144 accelerators
- The push reflects a broader shift away from proprietary AI interconnects like Nvidia NVLink and InfiniBand
Arista is known for switches, but the company is now tackling rack-level solutions as AI rack densities climb and data center operators demand more from networking partners. Ethernet-based scale-up has emerged as a key focal point in its evolution.
This week Arista unveiled new scale-up and scale-out architectures engineered in partnership with AMD, Arm, Microsoft and Qualcomm. Both architectures leverage open Ethernet specifications.
“Where Arista is coming from is a strong embrace of standards-based solutions. It is all ethernet: ESUN for the scale-up side and UEC-based Ethernet for the scale-out side,” Arista’s Distinguished Lead for AI Systems and Networks Vijay Vusirikala told Fierce.
ESUN, or Ethernet for Scale-Up Networking, is an Open Compute Project Foundation networking project launched last year. UEC refers to the Ultra Ethernet Consortium, which released its first specification in June 2025.
Its Etherlink SU-144 architecture for scale-up is capable of supporting up to 144 accelerators in a rack and up to 1,024 in a cross-rack setup. It comes in three configurations: an orthogonal chassis design with the switches mounted vertically behind the servers, a cabled backplane design with the switches sandwiched between servers, and a cross-rack design with switches positioned between server racks.
“Scale up is now a first-class networking citizen with everything you had traditionally expected in a full-featured network stack,” Vusirikala said.
Meanwhile, its UEC-based scale-out architecture features Multipath Reliable Connection (MRC) fabric resiliency, multi-plane routing, and intelligent Dynamic Load Balancing (DLB) and Cluster Load Balancing (CLB). It is rolling out reference designs for 8-, 12-, and 16-switch designs.
Both the scale-up and scale-out architectures are based around liquid cooling. That’s right, liquid cooling isn’t just for GPUs anymore.
“These switches are 100% liquid cooled,” Vusirikala said. “To get to this type of density, we cannot do it with air-cooled switches.”
Getting into the scale-up domain
The move comes as Arista pushes beyond its historical plays in scale-out (rack to rack) and scale-across (data center to data center) to serve scale-up (within the rack) needs as well. Arista CEO Jayshree Ullal first outlined the company’s plans to get into the scale-up game on the company’s Q1 2026 earnings call in May.
“We haven’t played in the scale-up market because it was dominated by Nvidia’s NVLink,” Vusirikala explained. But as the AI market has diversified with more chip entrants – everyone from Meta, Microsoft and Google to startups like Rebellions – demand for non-proprietary Ethernet based solutions has grown.
As Nvidia’s share of the chips market decreases, NVLink will see a corresponding drop in dominance. That’s the opportunity Arista is looking to capitalize on.
650 Group Co-Founder and Principal Analyst Alan Weckel told Fierce the scale-up market is expected to be worth around $20 billion in 2026 and grow to $105 billion in 2030. Ethernet is expected to account for less than $1 billion this year and $30 billion by 2030.
"As we get more accelerators(GPU/XPU/TPU), the market will looks towards Ethernet to have a consistent and high performance fabric across their accelerator fleet," Weckel said.
This trend has already been seen in the scale-across market, which was dominated by Nvidia’s InfiniBand technology for years before Ethernet flipped the script in in 2024 and 2025. While InfiniBand saw something of a resurgence in Q1 2026, Ethernet has continued to expand its lead in the scale-out domain, according to Dell’Oro Group.
Arista executives said on its Q2 2026 earnings call that the company is eyeing open co-packaged optics for the scale-up domain, with these expected to come into sampling and trials next year. Vusirikala added that validation and a manufacturing ramp is expected to follow in 2028 and beyond.
In terms of who it's up against in the scale-up market, Weckel said Nvidia is the dominant player today, with Astera Labs and Celestrica having footholds with Amazon and AMD, respectively.
Read more about data center networking on Fierce Network
Arista taps ex-AWS exec to scale data center biz
Cisco has its own take on how to ‘scale across’
Nvidia beefs up its Ethernet platform to let data centers join forces
Nvidia eyes data center Ethernet as its next multi-billion-dollar biz
CScale bets AI factories can ride out laser failures at gigawatt scale