Arista Launched Ethernet Architectures for AI Data Centers

The new designs aim to shift AI scale-up and scale-out networking toward open Ethernet standards.

Updated on Oct. 8, 2026 in Data Centers

Isometric editorial illustration of stacked industrial server blades connected by copper liquid cooling pipes. — Arista has unveiled a new Ethernet architecture designed to support high-density AI data centers, shifting industry standards toward open, scalable networking. AI Illustration. Upload story photo >

Live Poll

Should the technology industry prioritize open-standard architectures over proprietary ones for future infrastructure?

Arista has introduced new Ethernet-based architectures for AI data centers, supporting up to 144 accelerators in a single rack. The designs mark a pivot into the scale-up market using open standards to support a broader range of AI chip suppliers.

Why it matters

As AI development demands more hardware, the industry is shifting away from proprietary standards toward open Ethernet to avoid vendor lock-in. This movement aligns with the growing diversification of AI silicon suppliers.

The Etherlink SU-144 architecture scales up to 144 accelerators per rack, with cross-rack designs supporting up to 1,024 units. These systems leverage Ultra Ethernet Consortium specifications for scale-out tasks and Ethernet-based designs for scale-up.

The players

Arista

A provider of high-speed networking switches and software for large-scale data centers.

Ultra Ethernet Consortium

An industry group developing open, Ethernet-based standards for high-performance AI networking.

Jayshree Ullal

The CEO of Arista who directed the company's expansion into the AI scale-up domain.

The details

The scale-up architecture utilizes orthogonal chassis, a cabled backplane, and a cross-rack design to manage high-density traffic. For scale-out operations, the architecture employs Multipath Reliable Connection — a protocol for maintaining high-speed fabric resiliency — alongside multi-plane routing and dynamic load balancing. These designs also incorporate liquid cooling to manage the thermal output of the high-density AI accelerators.

2024: Ethernet surpassed InfiniBand in the scale-across market.
June 2025: Ultra Ethernet Consortium released its first specification.
May 2026: Arista CEO announced plans to enter the scale-up market.
October 2026: Arista unveiled the new AI architectures.

The Tech Race

Arista is competing to capture a share of the $105 billion scale-up market by displacing proprietary interconnects with open Ethernet protocols. This follows a broader industry trend where Ethernet gained significant ground against InfiniBand throughout 2024 and 2025.

Data center operators can expect to begin evaluating these Ethernet-based designs as Arista matures its scale-up hardware portfolio. Arista plans to begin sampling open co-packaged optics for this domain in 2027, which will be the next major hardware integration phase for users.

The takeaway

The transition to Ethernet in AI compute marks a fundamental shift away from proprietary interconnect lock-in. Watch for Arista's sampling of co-packaged optics next year as a key indicator of Ethernet's viability in the high-density scale-up market.