The Networking for AI Summit kicked off with a great market insights discussion with Zeus Kerravala of ZK Research. We discussed how AI is impacting networking environments, including the back-end, front-end, and WAN. Their central thesis: the network is not ancillary to AI, it is the foundation that determines whether training, inference, and emerging agentic workloads can scale reliably and securely across data centers, clouds, and the edge. We also covered the importance of leveraging AI to manage these network environments.
Below are some of the highlights of our discussion, however, take a listen to the full video below
Why networking is now pivotal.
- AI shifts traffic patterns from north–south to massive east–west flows; low latency and high bandwidth (400/800G today; Tbps soon) are non-negotiable.
- Market signals (vendor earnings across switching, optics, and security) indicate the “blade” of the hockey stick is ending; broad upgrades are beginning.
- Simplicity and speed-to-adoption matter: enterprises are leaning on vendor “AI factory” blueprints and validated designs as skills ramp.
Data center fabric: Ethernet vs. InfiniBand.
- InfiniBand led early for ultra-low latency lossless fabrics; Ethernet has rapidly closed the gap with RoCEv2, improved silicon, and deterministic features.
- Recent testing shows negligible training performance differences in many scenarios; industry momentum (including GPU vendors) favors Ethernet for scale-out and operational consistency.
- Next shift: scale across data centers (DCI) as power/placement constraints rise—expect more optical/photonics investment and fabrics that treat multiple sites as one compute unit.
Edge, access, and device explosion.
- AI value creation hinges on edge data; connectivity becomes strategic.
- Wi-Fi 7 and private cellular (4G/5G) will co-exist: Wi-Fi for high throughput/density, private cellular for deterministic URLLC-like needs (ports, mines, logistics).
- Gap to close: converged management across Wi-Fi and private cellular.
- Machine-to-machine/IoT growth and real-time requirements push more inference to the edge.
WAN & interconnect.
- More data must move between sites/clouds for fine-tuning and model ops; telcos are introducing AI-oriented transport offers.
- Expect a resurgence in interest in optimizing WAN traffic and careful capacity planning to prevent starving mission-critical applications.
AI for networking (ops transformation).
- Transition from reactive, human-intensive break/fix to predictive, generative, and agentic operations: forecasting saturation/failures, natural-language queries, digital twins, and closed-loop remediation.
- Human-in-the-loop remains essential for context, governance, and trust (“time to comfort”). Roles evolve rather than disappear; engineers shift from log-sifting to design, policy, and oversight.
Implications & guidance for enterprises.
We concluded by discussing practical advice for enterprises as they embark on their AI initiatives.
- Plan holistically: treat networking for AI (fabrics, DCI, edge/WAN) and AI for networking (agentic AIOps) as a single program.
- Prioritize determinism and security: deterministic latency, loss handling (RoCEv2), and end-to-end, distributed policy enforcement—no performance/security trade-offs.
- Prepare for always-on inference: capacity planning must assume sustained utilization plateaus, not spiky batches.
- Adopt proven patterns: use validated designs/AI-factory blueprints to accelerate deployment and de-risk operations.
- Invest in people + trust: start now to build operational confidence with human-in-the-loop automation; increase transparency and auditability of automated actions.
Our ANGLE
From GPU interconnects to wireless at the edge and the WAN in between, the network is a critical factor in scaling AI. The next wave combines high-performance Ethernet fabrics, optical DCI, and converged edge access with agentic, human-guided operations—turning networking into a proactive, autonomous foundation for AI business outcomes.
