Arista Networks is stepping up its effort to make Ethernet the common language of AI infrastructure, introducing new reference architectures and a roster of integration partnerships aimed at scale-up and scale-out networking. The move puts Arista in sharper competition with Nvidia, whose NVLink and InfiniBand technologies have become the default fabric for many high-performance AI clusters. Arista's message is that Ethernet, long the backbone of general-purpose data centers, can now meet the demanding bandwidth, latency and reliability requirements of accelerated computing without locking customers into a proprietary stack.

The company said the new designs bring together its 7060EX7 Series switches with physical infrastructure components such as fiber patch panels, integrated liquid-cooling manifolds, drip trays with active leak detection, rack management controllers and power shelves with backup battery units. Arista is pairing that hardware with software and technologies from partners including AMD, Arm, Broadcom, Meta, Microsoft and Qualcomm. The goal is to give large enterprises and hyperscale operators a practical path to build Ethernet-based AI clusters at rack scale.

Key facts at a glance

• Arista introduced Ethernet reference architectures and integration partnerships with AMD, Arm, Broadcom, Meta, Microsoft, Qualcomm and other industry players.
• The designs combine Arista's 7060EX7 Series switches with fiber patch panels, liquid-cooling manifolds, leak-detection drip trays, rack management controllers and battery-backed power shelves.
• Scale-up networking is based on the Open Compute Project's Ethernet for Scale-Up Networking initiative, or ESUN, while scale-out uses Ethernet features from the Ultra Ethernet Consortium, or UEC.
• Arista will not sell fully integrated racks. It plans to sell switches and software for partner-led rack integration and custom joint development.
• The Etherlink SU-144 architecture includes orthogonal chassis, cabled backplane and cross-rack designs, with the cross-rack option supporting up to 1,024 accelerators.

Ethernet as an open alternative

At the heart of Arista's strategy is the belief that the AI networking market is shifting from a single-vendor model to a more diverse ecosystem. For years, Nvidia's proprietary scale-up interconnect was the primary technology used to build tightly coupled AI systems. But as demand for AI compute expands, a growing number of non-Nvidia accelerators are entering the market. Those accelerators include Microsoft's Maia, Meta's MTIA, AMD's Helios and Google's Tensor Processing Units. Each represents an opportunity for Ethernet switch vendors, because customers want networking that can support multiple accelerator types rather than a fabric tied to one supplier.

Arista's cloud and AI group vice president and general manager said the company's differentiator is that its approach is strictly Ethernet. He noted that until recently Nvidia was the only major game in town and its proprietary scale-up technology dominated the market. As the total addressable market for scale-up networking grows and the non-Nvidia accelerator ecosystem expands, Arista sees a new opening. The company is positioning its switches, software and reference designs as a way to build AI networks that remain open at the accelerator, optics and rack level.

ESUN and UEC provide the standards foundation

The reference architectures rest on two industry efforts. ESUN, formed by the nonprofit Open Compute Project, is intended to advance Ethernet for scale-up connectivity across accelerated AI infrastructure. Its backers include AMD, Arista, Arm, Broadcom, Cisco, HPE Networking, Marvell, Meta, Microsoft, Nvidia, OpenAI and Oracle. ESUN is designed to give network operators flexibility across accelerators, switching silicon, optics, physical rack designs and network software while retaining the high bandwidth, low latency and reliable in-order delivery required for collective communication.

The Ultra Ethernet Consortium, meanwhile, focuses on a complete Ethernet-based communication stack architecture for high-performance networking. Its members include AMD, Arista, Broadcom, Cisco, Eviden, HPE, Intel, Meta and Microsoft. Arista is using UEC features for scale-out designs, including Multipath Reliable Connection fabric resiliency, multi-plane routing, intelligent dynamic load balancing and cluster load balancing. Those capabilities are aimed at improving AI token generation efficiency and handling the variable workloads that characterize inference and training environments.

Three Etherlink SU-144 designs

Arista's primary Ethernet ESUN-based scale-up architecture is called Etherlink SU-144. It comes in three reference designs, each targeting different density, serviceability and scaling requirements.

• Orthogonal chassis: This design uses direct orthogonal connectivity between accelerator and switch blades. It supports accelerator density of 144 XPUs within direct liquid-cooling envelopes ranging from 100 kilowatts to 400 kilowatts.
• Cabled backplane: This option combines a high-density cabled backplane with a modular rack structure to enable a highly serviceable design, making maintenance and component replacement easier in dense AI environments.
• Cross-rack: This modular architecture provides maximum flexibility to scale compute and networking across multiple racks. It extends the scale-up domain beyond a single physical rack and supports up to 1,024 accelerators while maintaining high-bandwidth, low-latency connectivity.

The cross-rack design is particularly significant because it challenges the traditional assumption that scale-up networking must remain confined to one rack. By extending the scale-up domain across racks, Arista is trying to give customers larger scaling domains that can span both scale-up and scale-out functions. That flexibility matters as AI models grow and as operators look for ways to balance accelerator density, power delivery and cooling capacity.

Rack-scale integration and liquid cooling

Arista's announcement also reflects a broader shift in AI infrastructure. As clusters become denser and move toward liquid cooling, the rack is replacing the individual switch as the practical unit of deployment. Power distribution, thermal management, fiber density, leak detection and system validation must now be engineered as an integrated solution rather than as separate parts. Arista is addressing that shift by offering switches alongside rack-level components and by working with partners to deliver fully integrated systems.

The company is careful to note that it will not sell a rack SKU on its price list. It is not selling integrated racks directly. Instead, it is selling the switches that go into liquid-cooled racks, along with software and technology for integration into custom joint development designs. Arista is partnering with value-added resellers, system integrators and optical vendors, as well as manufacturing partners such as Foxconn, Quanta and Hive, so customers can work with the partner of their choice to deliver a fully integrated rack.

Bandwidth growth and footprint reduction

Arista said customer designs will vary by accelerator power, GPU count, bandwidth, thermal envelope and rack configuration. The company sees its capabilities in chassis engineering, thermal design and signal integrity as differentiators. One rack configuration, for example, features bandwidth that scales with next-generation interconnects. A 16-switch configuration approaches 1.6 petabytes per second per rack. Designs using XPO, CPO and NPO optical technologies could quadruple that bandwidth. A Tomahawk 7, 200G-based generation could support 6.5 petabytes per second in one rack while exceeding 100 kilowatts.

Compared with today's OSFP-based switches, which deliver 1.6 petabits per second per rack, next-generation optics increase that to 6.5 petabits per second per rack. Arista said this massive densification translates to nearly a 50 percent reduction in overall data center footprint, resulting in significant time and financial savings. The figures underscore why networking is becoming a central design constraint in AI data centers. As accelerators become more powerful, the fabric must keep pace without consuming disproportionate space, power or cooling capacity.

NetDI underpins validation and diagnostics

The foundation of all these designs is Arista's Network Diagnostics Infrastructure, or NetDI, a diagnostics and validation layer that runs between Arista's EOS operating system or third-party network operating systems such as SONiC and FBOSS. Arista's newer 1.6T AI switches integrate NetDI specifically to provide consistent diagnostics across different operating environments. NetDI can track switch diagnostics functions such as cable and optic health, device telemetry, troubleshooting and failure isolation. Those capabilities help cloud companies manage massive scale networks where small failures can have outsized operational impact.

Arista said NetDI offers deep hardware-level validation, signal integrity analysis, secure boot attestation and Single Event Upset resiliency. It also supports rich telemetry for the switch, physical optics, power shelves and liquid cooling infrastructure. That telemetry is increasingly important as AI racks become more complex and as operators need to correlate network performance with power, cooling and accelerator health. By building diagnostics into the reference architecture, Arista is trying to reduce integration risk for customers and partners that are assembling rack-scale AI systems from multiple vendors.

Competitive landscape and customer choice

The announcement lands in a rapidly evolving competitive landscape. Nvidia remains a dominant force in AI accelerators and networking, but customers are increasingly looking for alternatives that avoid vendor lock-in and support a broader mix of hardware. Broadcom, Cisco, HPE and other Ethernet players are also investing in AI networking. Standards groups such as the Open Compute Project and the Ultra Ethernet Consortium are trying to align the industry around common specifications so that accelerators, switches, optics and software can interoperate more easily.

Arista's bet is that open Ethernet will win share as AI deployments move from early hyperscale experiments to broader enterprise and cloud rollouts. The company is not arguing that Ethernet is a drop-in replacement for every proprietary interconnect in every workload. Instead, it is building a set of reference designs that let customers choose Ethernet for scale-up and scale-out while working with partners to tailor power, cooling and rack integration. The approach gives Arista a role in the rack-level conversation without requiring it to become a full-system rack vendor.

For customers, the appeal is flexibility. A reference architecture built around ESUN and UEC can accommodate different accelerators, different optical technologies and different rack designs. That matters as AI infrastructure evolves quickly and as operators try to avoid stranded investments. For Arista, the opportunity is to extend its data center switching franchise into the most demanding part of the network. The company is positioning its switches, software and partner ecosystem as the building blocks for Ethernet-based AI fabrics that can scale from a single rack to multi-rack deployments while maintaining the low latency and high bandwidth that AI workloads require.

The new designs also highlight how much of AI networking is now about physical engineering. Liquid-cooling manifolds, leak detection, power shelves with backup batteries and rack management controllers are not traditional switch features, but they are essential to deploying dense AI systems at scale. By addressing those elements in its reference architectures, Arista is acknowledging that the network is no longer just a collection of switches. It is an integrated system that must be validated, monitored and serviced as a unit. NetDI, partner integration and the company's focus on signal integrity and thermal design are all part of that broader message.

Source:Network World News