Arista Networks has introduced the 7060XE7 Series, a new portfolio of 1.6T networking platforms that represent a strategic shift from standalone switches to rack-scale systems engineered specifically for AI infrastructure. The announcement, made on June 9, 2026, underscores Arista's commitment to addressing the extreme bandwidth, latency, and thermal challenges posed by large-scale AI workloads.
The key facts are clear: Arista's 7060XE7 family includes fixed switch platforms and configurable rack-scale systems built on Broadcom Tomahawk 6 silicon. The portfolio supports air, liquid, and hybrid cooling, and runs Arista's Extensible Operating System (EOS) with advanced packet buffering. Specific models include the 7060XE7-64PS and 7060XE7-64PRS (4U air-cooled), the 7060XE7-64PRS-RV3-L (2OU liquid-cooled), and the 7060XE7-128PE (128 800G ports in 4RU). All platforms support the Open Compute Project's Multipath Reliable Connection (MRC) protocol, an RDMA-based transport that enables efficient load balancing across multiple network paths.
Platform Overview
The 7060XE7-64PS and 7060XE7-64PRS, available in Q4 2026, are 4U air-cooled switches that support pluggable integrated heat sink (IHS) and riding heat sink (RHS) optics. IHS optics target current air-cooled data centers, while RHS optics are designed for future liquid-cooled AI fabrics requiring extreme port density. These models provide 64 ports of 1.6T each, delivering aggregate switching capacity of over 100 Tbps per chassis. The 7060XE7-64PRS-RV3-L is a specialized 2OU liquid-cooled platform that uses DC power from the ORv3 rack and contains no internal fans. It integrates directly with liquid-cooled XPU servers to maximize power efficiency and is slated for Q1 2027. The 7060XE7-128PE, also arriving in Q1 2027, offers 128 800G ports in an air-cooled 4RU design using 100G SerDes, providing backward compatibility and deployment flexibility for heterogeneous environments.
Each switch in the 7060XE7 family leverages Broadcom's Tomahawk 6 ASIC, which delivers 1.6T per port with 224G SerDes technology. This silicon architecture is critical for AI scale-up and scale-out fabrics, enabling low-latency forwarding and support for advanced congestion control mechanisms. Arista has also announced a partnership with AMD to jointly develop next-generation compute silicon and NICs that will further optimize AI fabric performance.
Core Technologies
At the software layer, the 7060XE7 Series runs Arista EOS, which includes low-latency and intelligent packet buffering to manage the intense microbursts common in AI communication patterns such as all-reduce and all-gather. The operating system also supports open-source alternatives like SONiC and OpenSwitch, giving operators flexibility in their network stack.
A standout feature is full support for the Open Compute Project's Multipath Reliable Connection (MRC) protocol. As described by Arista's CTO Kenneth Duda and distinguished engineer Alan Judge, MRC is an RDMA-based transport that allows a single reliable connection to simultaneously use many network paths over Ethernet. Endstation NICs stripe traffic across multiple links and paths, with out-of-order packets handled automatically at the receiver. MRC responds to congestion signals (ECN and packet trimming) by shifting load to the best-performing paths and avoiding those with link errors or failures. In production, Arista has demonstrated very high fabric utilization with good load balancing, all while interoperating seamlessly with scale-across and WAN networks using standard dynamic routing protocols.
Beyond MRC, the software stack includes advanced load balancing, congestion management, telemetry, and diagnostics—all essential for AI networking. Arista has also emphasized support for Open Compute Project's Ethernet Switch Abstraction Interface (SAI) and other community-driven standards to ensure multivendor interoperability.
Industry Context
The 7060XE7 Series enters a rapidly expanding ecosystem of 1.6T Ethernet solutions, with competitors including Cisco, Nvidia, and Celestica. Analyst Sameh Boujelbene of Dell'Oro Group noted that the platform signals a clear direction for large-scale AI fabrics: higher bandwidth, better power efficiency, and tighter integration between compute, optics, silicon, cooling, and network software. She pointed to strong customer and ecosystem validation from Microsoft Azure, Oracle Cloud Infrastructure, Meta, AMD, and Broadcom as a key differentiator.
Arista's move to rack-scale systems reflects the growing need for holistic AI infrastructure that goes beyond individual switches. AI training clusters now require thousands of accelerators communicating simultaneously, generating enormous east-west traffic with microsecond-level latency constraints. Traditional data center networking, designed for general-purpose workloads, cannot handle the bursty, all-to-all communication patterns of model training. Rack-scale systems like the 7060XE7 Series integrate power, cooling, optics, and software into a single optimized unit, reducing deployment complexity and improving total cost of ownership.
The portfolio's support for liquid cooling is particularly timely. As GPU and XPU power densities exceed 1000W per device, air cooling becomes insufficient for dense clusters. The 7060XE7-64PRS-RV3-L, with its fanless design and ORv3 DC power, allows operators to deploy switches directly in liquid-cooled racks alongside compute nodes, eliminating the need for separate cooling zones. This integration reduces energy consumption and improves overall system reliability.
Arista's engineering choices also reflect an understanding of the evolving Ethernet standards landscape. The inclusion of 224G SerDes and support for both IHS and RHS optics positions the platform for future 1.6T and 3.2T generations. The company has stated that the 7060XE7 architecture can be extended to support upcoming 800G and 1.6T optics without a complete component redesign, protecting customer investments.
From a software perspective, Arista's commitment to open standards like MRC and SONiC aligns with the broader industry push toward disaggregation and vendor neutrality. The MRC protocol, in particular, addresses a persistent challenge in RDMA over converged Ethernet (RoCE): load imbalance due to path polarization and congestion. By striping traffic across multiple paths at the NIC level, MRC achieves near-optimal utilization without requiring complex switch-based load balancing. This approach simplifies network design and enables higher effective throughput without sacrificing reliability.
The 7060XE7 family also includes extensive telemetry capabilities, including streaming data from the Broadcom ASIC's programmable pipeline. Operators can monitor per-flow latency, drop distribution, and buffer occupancy in real time, feeding data into AI-driven network optimization tools. Arista's CloudVision management platform integrates with these telemetry feeds to provide end-to-end visibility across the AI fabric.
Looking ahead, Arista's collaboration with AMD on next-generation compute and NICs suggests a deeper integration path. The combination of AMD's upcoming EPYC processors and Instinct accelerators with Arista's switches could deliver optimized fabric configurations tailored to specific model architectures. While specific product timelines remain unannounced, the partnership underscores the importance of co-optimization between compute and network for AI performance.
The 7060XE7 Series represents Arista's most ambitious platform to date—a comprehensive rack-scale solution that addresses the unique demands of AI infrastructure head-on. With its support for multiple cooling technologies, open protocols, and high-bandwidth silicon, the family provides a foundation for building efficient, scalable AI data centers capable of supporting the next generation of large language models and scientific computing workloads.
Source: Network World News