Arista Networks Challenges Nvidia with AI Ethernet Fabric

Arista Networks is positioning Ethernet as the premier networking fabric for artificial intelligence scaling to provide an open alternative to proprietary systems. The company recently introduced new reference architectures and a wide range of integration partnerships with major industry leaders to help hyperscale customers implement these designs.

Establishing an Open Fabric for AI Infrastructure

Arista is moving to provide a significant alternative to the proprietary interconnect technologies that currently dominate the market. By focusing on standard Ethernet, the company aims to offer network operators greater flexibility across various accelerators, switching silicon, and physical rack designs. This initiative targets the growing need for high bandwidth and low latency in modern AI clusters without locking users into a single vendor ecosystem.

The company is rolling out its 7060EX7 Series switches along with an extensive list of supporting hardware. This includes fiber patch panels, liquid-cooling manifolds, and advanced leak detection systems. These components are designed to work with software from a broad ecosystem of partners including AMD, Arm, Broadcom, Meta, Microsoft, and Qualcomm.

The push for Ethernet-based systems comes at a time when the market for AI accelerators is diversifying. Many organizations are developing their own specialized chips, such as Microsoft Maia, Meta’s training accelerators, and Google’s Tensor Processing Units. Each of these represents a new opportunity for Ethernet-based networking to serve as the unifying fabric.

Arista is heavily involved in the Ethernet for Scale-Up Networking initiative, a group formed by the Open Compute Project. This group includes a wide variety of industry heavyweights who are committed to advancing Ethernet technology. Their goal is to ensure that standard networking can handle the intense demands of accelerated AI infrastructure while maintaining reliability.

By adhering to these standards, Arista ensures that its products remain compatible with a wide range of hardware. This approach contrasts sharply with closed systems that require specific, often more expensive, cabling and switching components. The company believes that as the non-Nvidia accelerator market grows, the demand for standard Ethernet will follow suit.

Developing Advanced Reference Architectures

To help customers implement these high-density systems, Arista has defined three primary reference designs. These designs, part of the Etherlink SU-144 family, provide a path for integrating compute, networking, and liquid cooling into a single functional unit. Each architecture addresses different physical constraints and performance requirements found in modern data centers.

The first design uses an orthogonal chassis which allows for direct connectivity between accelerator and switch blades. This configuration is highly efficient for liquid cooling and can support a density of up to 144 processing units within a single envelope. It is specifically built for environments where power consumption ranges from 100 kilowatts to 400 kilowatts.

A second option utilizes a cabled backplane combined with a modular rack structure. This design emphasizes serviceability, allowing technicians to replace or upgrade components with minimal disruption to the rest of the system. It balances the need for high-density cabling with the practical realities of maintaining a large-scale data center.

The third design, known as cross-rack architecture, is perhaps the most ambitious. It allows the scale-up domain to extend beyond the physical boundaries of a single rack. By doing so, it can support up to 1,024 accelerators while still maintaining the low latency required for complex AI workloads. This breaks the traditional limits of rack-based scaling.

Industry analysts note that this shift marks an evolution in how data center hardware is purchased and deployed. Instead of buying individual switches, the entire rack is becoming the basic unit of deployment. This requires a much higher level of engineering integration between power distribution, thermal management, and network validation.

Arista is also looking toward the future of optical technology to further increase density. Current configurations provide about 1.6 petabytes of bandwidth per rack. However, the company anticipates that next-generation optics will push this to 6.5 petabytes. This massive increase in capacity could lead to a 50 percent reduction in the physical footprint of a data center.

Integration and Diagnostic Management

While Arista provides the blueprints and the switches, it does not sell fully integrated racks as a single product. Instead, the company works with value-added resellers and system integrators like Foxconn, Quanta, and Hive. These partners take Arista’s technology and build the finished rack systems according to the specific needs of the end customer.

The foundation for these complex deployments is the Network Diagnostics Infrastructure, or NetDI. This software layer runs between the operating system and the hardware to provide consistent monitoring across different environments. It is a critical tool for managing the massive scale networks that power modern AI training and inference.

NetDI provides deep hardware-level validation and monitors the health of cables and optics. It can perform signal integrity analysis and secure boot attestation to ensure the system is both functional and secure. This level of telemetry is essential for troubleshooting failures in a system where thousands of components must work in perfect unison.

The software also manages power shelves and liquid cooling infrastructure. Because AI workloads generate an immense amount of heat, tracking the performance of the cooling system is just as important as tracking network traffic. If a leak is detected or a pump fails, the system must respond instantly to protect the expensive hardware.

In terms of scale-out networking, Arista is incorporating features from the Ultra Ethernet Consortium. This includes multi-path fabric resiliency and intelligent load balancing. These features are designed to handle the specific traffic patterns of AI workloads, which often involve large bursts of data that can easily congest traditional networks.

By focusing on these diagnostic and management tools, Arista aims to make Ethernet as reliable and easy to manage as proprietary alternatives. The company is betting that the combination of open standards, high performance, and deep visibility will win over enterprises looking for long-term flexibility in their AI infrastructure investments.