Is Ethernet Ready for AI Networking?

I attended AI Infrastructure Field Day 4 last week in Santa Clara, California, and these are my takeaways from the Cisco Data Center Networking session.

During the session, Cisco focused on recent innovations across the Nexus networking portfolio and HyperFabric, with a particular emphasis on how Ethernet is evolving to support modern AI workloads.

Cisco’s position is that modern Ethernet, delivered through the Nexus portfolio, can now support both AI training and AI inference workloads. The company argues this enables organizations to simplify operations by using a single Ethernet-based network instead of maintaining separate fabrics.

A recurring theme was the importance of clearly distinguishing AI training from AI inference. Training is the phase where models are built, typically processing large datasets in batches and prioritizing maximum throughput to reduce time-to-model. Inference, by contrast, is the operational phase where trained models are deployed and consumed, and where low latency and predictable response times are critical.

From a networking standpoint, AI training environments typically rely on large GPU clusters connected in a non-blocking topology. In this design, any GPU can communicate with any other GPU in a point-to-point fashion without impacting parallel traffic. Historically, these requirements have led many organizations to adopt InfiniBand for GPU-to-GPU interconnects.

Cisco’s message is that modern Ethernet, implemented through the Nexus portfolio, now delivers capabilities that directly challenge InfiniBand in compute and GPU networking scenarios. Because Ethernet is a mature and widely understood technology, many AI infrastructure teams find it easier to configure and operate. Consequently, this familiarity reduces friction, especially compared to InfiniBand, which often requires teams to learn an entirely new operational model and management toolchain.

Additionally, Cisco’s presenters noted that they increasingly see customers deploying both training and inference workloads on Ethernet-based network topologies. By converging these workloads onto a single fabric, organizations simplify operations, reduce infrastructure complexity, and make ongoing management and support significantly easier.

At HighFens, when we speak with customers about AI networking, we consistently hear the same three questions:

• InfiniBand or Ethernet?
  Cisco’s Nexus answer: Ethernet across the board.
• Should training and inference use different network architectures?
  Cisco’s Nexus answer: A single Ethernet topology for both.
• Does network selection differ for brownfield versus greenfield deployments?
  Cisco’s answer: No. Ethernet enables integration of new and existing gear at the network layer, making mixed environments manageable.

The Cisco team made it clear that Ethernet is positioned as the right networking choice for AI, regardless of whether the workload is training or inference. Ethernet benefits from a rich ecosystem of mature tools and operational practices, making AI networks easier to manage while avoiding the need to retrain teams already fluent in Ethernet networking.

HighFens stands ready to help organizations navigate this pivotal year and ensure AI delivers real impact.