As enterprises move artificial intelligence into production, AI factory networking is becoming a core part of the infrastructure equation, shaping performance, scalability and cost.

That shift is the focus of a new analysis by Bob Laliberte, principal analyst at theCUBE Research. The analysis draws on a recent discussion with Gilad Shainer, senior vice president of networking at Nvidia Corp., hosted by Laliberte and theCUBE Research chief analyst, Dave Vellante.

The central argument is that enterprise AI infrastructure can no longer be evaluated primarily by the number or speed of its graphics processing units. Networking, compute, storage and software must operate as one coordinated system.

Nearly 95% of organizations responding to theCUBE Research’s “Networking for AI” study said networking is more important to achieving business objectives than it was two years ago.

“That is a remarkable change in perception and reflects how AI is reshaping infrastructure priorities,” Laliberte wrote.

Why AI factory networking matters for production workloads

Traditional enterprise applications often run on individual servers that operate largely independently. AI training, distributed inference, retrieval-augmented generation and agentic applications require GPUs, central processing units, data processing units and storage resources to exchange information continuously.

That makes AI factory networking more than a system for moving data, Laliberte explained. It helps determine whether distributed infrastructure functions as a unified AI platform or simply a collection of connected servers.

“The success of an AI factory depends on how efficiently distributed resources collaborate,” he wrote in his analysis. “The network is no longer simply transporting information between servers; it is coordinating the operation of an entire distributed computing environment.”

Production inference adds further complexity because it can involve processors, databases, storage platforms, retrieval systems and incoming user requests simultaneously. Agentic AI increases that activity as autonomous systems retrieve information, call tools and exchange context across multistep workflows.

The network must allow those resources to operate together, according to Shainer.

“If you want to build an AI supercomputer, you need a network that makes those compute engines work as a single unit,” he said.

The analysis also examines Nvidia’s “extreme co-design” strategy, which treats networking, compute, storage and software as interdependent parts of one platform. This systems-level approach is becoming increasingly important as AI environments expand in scale and complexity, according to Laliberte.

Ethernet also features prominently in the discussion. Nvidia’s Spectrum-X platform is designed to reduce congestion, jitter and uneven performance across distributed AI systems. Shainer maintained that the platform continues to rely on standard Ethernet protocols.

“Ethernet is open by definition,” he said. “We’re not using any proprietary protocols.”

For enterprises, the larger issue is economics. Networking can affect GPU utilization, power consumption, resiliency and cost per token. Delayed communication can leave expensive accelerators idle, while predictable network performance can help organizations generate more output from the same computing investment.

Laliberte’s full analysis examines the architectural, operational and economic implications of networking’s expanding role inside enterprise AI factories.

Image: SiliconANGLE/ChatGPT