ONUG details a two-part framework for AI networking adoption

The vendor session outlines a two-part framework for adopting AI in enterprise networking, covering AI-ready fabric design and AI/ML use in network operations. For IT and security leaders, it focuses on evaluation criteria, pilot planning, and governance controls that link network capacity and observability to measurable operational outcomes.

Research Overview

The session presents a practical approach to assess AI in networking through two dimensions: “Networks made for AI” and “AI made for Networks.” It describes the work needed to build and operate network infrastructure that can support AI workloads while also using AI to assist network operations.

It also frames the adoption process as moving from narrow pilots toward broader production deployments, with attention to maintaining control of data, prompts, and models. The session materials emphasize measurable milestones and the role of governance checklists in scaling.

Key Findings

For “Networks made for AI,” the session calls out design separation between front-end and back-end paths, differentiated QoS classes, and support for lossless transport using PFC/ECN. It adds GPU/NIC awareness to topology planning and includes capacity and failure domain concepts such as POD sizing and traffic isolation.

For “AI made for Networks,” the session describes applying observability and AI/ML to operations by correlating fabric telemetry with GPU/NIC health and using flow-level analytics for job service-level objectives. It also includes intent-drift detection as the environment scales and proposes early use cases such as anomaly triage and guided troubleshooting.

Technical Breakdown

The session’s fabric design guidance includes interoperability as a requirement, focusing on choosing open, vendor-agnostic building blocks across GPUs, switches, and DPUs and front-end infrastructure. It contrasts adopting prescriptive reference architectures with using multi-vendor approaches based on SONiC for flexibility, with “Spectrum-X” cited as an example of a prescriptive RA.

On the operational side, the session describes observability requirements for telemetry correlation, job-level analytics, and detection of changes over time. For model and data strategy, it covers options including private or open-source LLMs, retrieval-augmented generation with network context, and governance controls such as guardrails, RBAC, and audit trails.

Operational Impact

The session proposes capacity planning and risk containment by defining oversubscription targets and using blast-radius containment for training versus inference workloads. It also describes moving from initial pilots to production by using governance checklists, success metrics, and a roadmap to expand AI capabilities without reducing reliability.

For MLOps within NetOps, it outlines practices such as versioning prompts and policies, evaluating model changes, and tracking outcomes including reductions in MTTD/MTTR and improvements in change success rate. It also includes an outline for hands-on walkthroughs covering orchestration of an AI-ready fabric and use of AI assistants from Day 0 bring-up through Day 2 troubleshooting.

This session presents a fact-based evaluation framework for assessing AI in networking across AI-ready fabric design and AI-assisted operations, with attention to observability, capacity planning, and governance. “Blog Signals brief” is a fact-based summary of the vendor blog.