Cisco's 2024–2025 industrial networking surveys put hard numbers on a shift test engineers have watched build for several years: 97% of industrial leaders expect AI workloads to change their network requirements, and 59% of manufacturing respondents are already deploying AI in some form. Samuel Pasquier, vice president of product management for industrial IoT networking at Cisco, presented the findings — drawn from more than 1,000 industrial customers, 350 of them manufacturing leaders — in an interview on the Visions podcast with Vision Systems Design senior editor Jim Tatum. The conversation centered on what changes when machine vision and robotics move from pilot projects to production lines, and the answer, repeatedly, was the network.

The scale of the change is measurable in camera counts alone. Pasquier said the number of machine vision cameras deployed in manufacturing has grown roughly 40x over the last three years. Each camera that once required a dedicated box running its software now feeds a switched infrastructure, frequently powered over Ethernet, with frames flowing either to a server for real-time action or to storage for later analytics. Either path puts traffic on the plant network that dwarfs what a PLC or a motion controller generates. An automotive example from Cisco's customer base makes the data volumes concrete: one manufacturer photographs every spot weld on every car body and retains the images for production records, so that any downstream field problem can be traced back to a specific weld on a specific shift.

Latency as a safety parameter

When AI crosses from observation into control — what Pasquier calls physical AI, the "camera sees something, we take action" loop — latency stops being a performance metric and becomes a safety parameter. A camera-guided pick-and-place robot making a decision too slowly is a robot that has missed its event window; a moving arm that responds late is a hazard to anyone near the cell. Pasquier's framing, borrowed from a customer, treats the network as the plant's nervous system: cameras are the eyes, robots are the muscles, and the speed of the connection between them sets how much of the physical world the system can actually govern. A larger network connects more assets and yields a more holistic view; smaller, isolated networks trade that visibility for simplicity.

Where deployments stall

Two failure modes surfaced in the survey work and customer engagements. The first is capacity. Pasquier described a recent customer that deployed cameras across a factory to observe assembly workers and optimize cycle times. Correlating video from many cameras simultaneously demanded more bandwidth than the existing network could carry without degrading PLC traffic — the one constraint no controls engineer will accept. The workaround, building a second network dedicated to the cameras, works technically but does not scale. Two parallel industrial networks double the operational burden.

The second failure mode is lifecycle management. A proof of concept involves one vision station. A production facility involves many, and someone must manage software updates and AI model upgrades across all of them. Cisco's response has been virtualization: running vision software in the data center, where virtualization tooling is mature, rather than replicating appliance boxes on the floor.

Edge, cloud, and data shelf life

Pasquier proposed a decision rule for where processing belongs: the shelf life of the data. Tracking a part's position matters only until the robot moves it; that data can live and die at the edge. Food and beverage traceability use cases, where cameras read QR codes on packaged goods and records must persist, point to cloud storage. The trade-off scales with data size: the bigger the retained dataset, the bigger the pipe to the cloud and the bigger the storage bill. Security rides along with any cloud path — Pasquier flagged intruder access to plant infrastructure as a design-level concern, not an afterthought.

Looking forward, he named mobility as the requirement most likely to be under-built today. Autonomous guided vehicles with robotic arms, and eventually humanoid platforms, will need wireless coverage engineered with the same rigor as the wired plant. His advice: manufacturers should decide how many separate wireless networks they are willing to operate, because proliferating single-purpose networks — one per use case — is the pattern that consistently breaks at scale.

Cisco has sold IT networking for more than forty years and industrial networking for more than twenty, and Pasquier was candid that his emphasis on architecture may reflect that vantage point. Still, the survey data is not vendor opinion: it is what 1,000 industrial practitioners reported. The question it leaves open for any manufacturer budgeting for physical AI is whether the plant network has been sized and characterized — bandwidth, latency, security posture — before the pilot's success is asked to survive a multi-line rollout.