Arista Brings Ethernet to 1.6 Tbps for Rack-Scale AI Networks

Arista Networks has announced the 7060XE7 series, a new family of 1.6 Tbps Ethernet platforms designed for high-density AI networks. The company aims to address a reality that already influences data center design: as training and inference loads grow from thousands to hundreds of thousands of accelerators, the network ceases to be an independent layer and instead functions as part of the computing system itself.

The announcement expands Arista’s Etherlink family and places the company in one of the most relevant technical debates today: how to build open, efficient, and predictable networks for massive AI clusters. The demand is no longer just for connecting servers. New infrastructures need to move massive volumes of data between XPUs, GPUs, NICs, memory, and storage systems with low latency, congestion control, and moderated energy consumption.

The network as the backbone of AI

According to Arista, the 7060XE7 marks a transition from high-performance switches to complete rack-scale systems. These platforms are positioned as a foundation for AI fabrics, both scale-out and scale-up, capable of adapting to air-cooled, liquid-cooled, or hybrid environments.

This technological shift is driven by a specific pressure: in training large models, the network directly influences job completion times. An unstable link, poorly managed congestion, or microcuts can impact thousands of accelerators and delay tasks that cost millions in compute capacity. In large-scale inference, latency and predictability are crucial, especially when models are run for millions of distributed requests.

Arista states that the 7060XE7 is designed to operate as a “AI supersystem” at rack scale, with low latency, intelligent buffering, and EOS capabilities tailored for communication-intensive patterns. This includes intra- and inter-rack traffic, as well as the semantics of computing and memory specific to current AI systems.

Meta, Microsoft, and Oracle Support the Ethernet Approach

Arista correlates this launch by highlighting some of its major cloud and AI clients. Meta, Microsoft, and Oracle are mentioned in the release as examples of operators that need denser, more efficient, and stable networks for their next-generation infrastructure.

Meta emphasizes the need to evolve physical infrastructure to support higher density and energy efficiency. Microsoft links the 1.6T Ethernet interfaces to the next generation of AI clusters for Azure Maia and their Fairwater data centers. Oracle underscores the necessity of networks that deliver performance, determinism, and stability for RDMA-based AI fabrics.

Mentioning these clients is deliberate. Ethernet aims to establish itself as an open, widely operable alternative to other specialized interconnection technologies. For major cloud providers, the promise of Ethernet is not just bandwidth but operational capabilities: familiar tools, open standards, ability to integrate multiple vendors, and continuity with existing data center networks.

This argument gains strength as AI clusters transition from experimental setups to industrial infrastructure. If a cloud provider wants to deploy tens or hundreds of thousands of accelerators, it needs a network that can scale, recover from failures, and be managed with the same discipline as the rest of its infrastructure.

Broadcom, AMD, and the Leap to 1.6T

Moving to 1.6 Tbps networks depends on several elements: switching silicon, NICs, optics, cabling, cooling, software, and rack design. Arista notes its collaboration with AMD on next-generation compute silicon and NICs for open standards-based scale-out AI fabrics.

Partnerships with Broadcom are central. The new series leverages Tomahawk 6, Broadcom’s Ethernet silicon with 102 Tbps switching capacity. Arista states it combines these capabilities with EOS or other open network operating systems to turn switches into a unified, rack-scale radix, suitable for air or liquid-cooled environments.

A highlighted feature is support for Linear Pluggable Optics (LPO), which Arista says can cut interconnect energy consumption by about 60%. In AI data centers, this reduction is significant. Energy isn’t consumed only in GPUs or accelerators but also in networking, optics, cooling, and power supplies. Any reduction in the interconnection layer can improve compute density per kilowatt.

EOS Adds Operational Intelligence for AI Fabrics

Software is as critical as hardware. Arista integrates several EOS functions into the 7060XE7 aimed at resilience, congestion management, and stability in AI environments. These include Dynamic Load Balancing, Cluster Load Balancing, support for Multi-Route Cache (MRC), Link Layer Retry, PFC-aware DLB, PFC-aware ECN, telemetry, Congestion Signaling, and Fast CNP.

In practical terms, these features aim to reduce one of the major issues in AI networks: how a minor failure can amplify into a complete job halt. In a traditional cluster, a degraded link might only affect limited traffic. But in distributed training involving thousands of accelerators, a single problematic point can slow or halt the entire workload.

MRC, or Multipath Reliable Connection, specifically targets avoiding this “failure amplifier.” Link Layer Retry helps maintain physical layer stability. Rapid congestion signaling mechanisms allow proactive reactions to blockages or saturation. Telemetry and consistent diagnostics facilitate problem detection in large-scale environments where manual intervention isn’t feasible.

Arista also mentions NetDI and PDI for diagnostics and portability. NetDI offers a uniform way to access telemetry and device diagnostics, while PDI manages low-level hardware details. For large operators with diverse environments, this uniformity can reduce operational complexity.

AI’s Network Race Is a Network Race

Over recent years, much of the focus on AI infrastructure has been on GPUs, accelerators, HBM memory, and power capacity. Arista’s announcement reminds us that the network is equally crucial. Without sufficient interconnect, accelerators wait for data. Poor congestion control prolongs jobs. Lack of energy efficiency constrains rack density.

The 7060XE7 series enters a market where AI growth is prompting a redesign of racks, power systems, cooling, and topologies. The appearance of liquid-cooled platforms, fanless systems, and support for ORv3 racks demonstrates that networking is becoming physically integrated into cluster design, not added later as a separate layer.

For Arista, the strategy is clear: Ethernet must support next-generation AI clusters without renouncing open standards or consistent operation. For customers, the key question is which combination of cost, latency, power, reliability, and ease of management enables AI scaling without the network becoming a bottleneck.

Artificial intelligence isn’t just increasing demand for chips; it’s also driving a new generation of data center networks. With the 7060XE7, Arista aims to position itself where network infrastructure ceases to be auxiliary and instead becomes a core contributor to AI system performance.

Frequently Asked Questions

What has Arista announced?

Arista has announced the 7060XE7 series, a family of 1.6 Tbps Ethernet platforms aimed at rack-scale AI fabrics.

What is the purpose of 1.6T networks in AI?

They are designed to move large volumes of data between accelerators, servers, memory, and storage systems with low latency and high efficiency—crucial in training and inference clusters.

Which major companies are linked to this announcement?

Arista references Meta, Microsoft, and Oracle as key cloud and AI clients working with their platforms to scale next-generation infrastructure.

When will the new equipment be available?

The 7060XE7-64PS and 7060XE7-64PRS models are expected in Q4 2026; the 7060XE7-64PRS-RV3-L and 7060XE7-128PE in Q1 2027.

via: arista