The network becomes manufacturing’s AI infrastructure
The factory network used to be easy to overlook. It connected machines, moved data and sat beneath the more visible work of production. That view is becoming increasingly risky as manufacturers move AI, machine vision, robotics, mobile assets and edge applications out of pilots and into live operations.
Samuel Pasquier, vice president of product management, industrial IoT networking at Cisco, says the network has become critical because AI is no longer confined to isolated demonstrations. “It is very easy to do a proof of concept in a lab,” he says. “But when you start to deploy in production, the network becomes super critical because your environment is much wider, much bigger. You need to interconnect things so you can have more information.”
Cisco’s 2026 State of Industrial AI Report for Manufacturing shows how far that shift has moved. The report found that 59 per cent of manufacturers are actively deploying AI at scale, while 96 per cent expect AI workloads to affect industrial network requirements. Reliable connectivity, edge compute and bandwidth are now among the core requirements for scaling AI in production.
That matters because industrial AI is not an abstract software layer. In manufacturing, AI increasingly depends on cameras, robots, controllers, sensors, edge compute, wireless mobility and secure data movement between the plant floor and higher-level systems. If that infrastructure is not designed for production reality, a promising use case can become another local experiment that cannot be repeated safely across sites.
AI exposes the limits of old networks
Pasquier uses the example of an AI-enabled robotic arm to explain what has changed. Traditional automated assembly depended on fixed positioning, dedicated jigs and tightly controlled conditions. A part arrived in the expected place, and the robot performed a predefined action.
AI changes that by giving the robot vision and the ability to adjust to variation. “The robot now can adjust automatically to know where to put the screw, so they don’t have to have all those jigs,” Pasquier says. “That gives them more flexibility and allows them to move faster, but if you think about the technology stack behind that, you need to have a model, you need to have a camera that talks to a robotic arm, and you need to have a brain somewhere.”
That brain may be a GPU or CPU running close to the process, in a factory compute room, in a data center or in the cloud, depending on the use case. The model may need updating, the camera data must move reliably, and the response time must support the speed of the process.
That shift is already visible in the bandwidth expectations coming from manufacturers. Pasquier says plant-floor networks have moved from Fast Ethernet to Gigabit Ethernet, with some customers now looking at 10 Gigabit Ethernet as machine vision, robotics and edge analytics expand. “Over the last 10 years, there is 100 times increase in the amount of bandwidth and performance that customers are looking for on the plant floor,” he says.
Machine vision is one of the main reasons. A camera sees, the system processes and the operation reacts. The faster that loop closes, the greater the potential benefit for quality, safety and productivity. But vision systems require more than compute power. They also require bandwidth, low latency, reliable connectivity and power at the edge.
Wireless reliability becomes operational
The gap between ambition and factory reality is clearest in wireless. Cisco’s report found that 96 per cent of manufacturers say wireless networks are critical to enabling industrial AI, while 56 per cent say lack of reliability frequently affects AI-enabled mobility-related operations. For Pasquier, that tension reflects the rapid expansion of AGVs, AMRs and other mobile systems that were often first connected using standard Wi-Fi.
“Wi-Fi is an amazing technology,” he says. “But when you start to talk about low latency in an environment that has a lot of RF noise, think about a factory. You have metal everywhere, and this is not a very friendly environment for wireless communication. People have been able to make Wi-Fi work, but they reach the limitation of Wi-Fi.”
Planet Farms shows what this looks like in practice. The vertical farming company operates a 20,000-square-meter facility near Como, Italy, where robotic processes, real-time monitoring, data-collecting vehicles and advanced AI models depend on reliable wireless connectivity. Cisco Ultra-Reliable Wireless Backhaul resolved earlier latency and roaming issues, supporting real-time AGV and camera data collection in an environment with steel structures, humidity, low temperatures and terabytes of 3D camera data per day.
Power is becoming part of the same story. Cisco’s report found that manufacturers expect a 41-times increase in Power over Ethernet demand as AI vision systems are deployed. Pasquier says the number of cameras on plant floors has multiplied sharply, making the ability to deliver power and data through a single cable a practical scaling issue rather than a technical detail.
Security moves into the network
More connected assets also mean more cyber risk. Cisco’s report identifies cybersecurity as the top barrier to AI adoption, cited by 40 per cent of manufacturers, and security or segmentation as the biggest networking challenge for AI-enabled operations, cited by 46 per cent. Pasquier argues that every smarter device increases the attack surface because intelligence usually means software, communication and network access.
As AI connects more industrial assets, the security problem becomes harder to contain. “What does it really mean to be smarter in 2026?” Pasquier says. “It means that it can talk, it can think, it can exchange, so it’s connected to the network. And if it’s connected to the network, it’s increasing your attack surface.”
That matters because a large plant may have tens of thousands of connected devices, many of which cannot be patched as easily as corporate IT systems. “There is an economic reality,” he says. “Stopping a line to patch the PLC might not be something people want to do.” The network therefore becomes an essential control layer, defining which devices are allowed to communicate and limiting the spread of malicious activity if something goes wrong.
Segmentation has long been part of industrial security, but the challenge has become more complex as IT, OT, edge and cloud environments all need to exchange data. Cisco uses AI internally to cluster connected assets, identify devices that communicate with one another and generate segmentation policies. The purpose is not to add another isolated security tool, but to make visibility and policy enforcement part of the industrial network itself.
Audi provides the most advanced supporting example. At its Böllinger Höfe plant in Germany, Audi’s Edge Cloud 4 Production platform, developed with Cisco support, is part of a move toward software-defined manufacturing. Audi is virtualizing production-related applications and hosting virtual PLCs in an on-premise private cloud, while using a powerful network to connect applications in the data center with robots and automation cells on the plant floor.
The latency requirement is severe. Audi’s case study says some use cases must satisfy microsecond-range latency, and industrial communication with robots requires less than one millisecond of end-to-end latency. Cisco software-defined access supports micro-segmentation, while Cisco Identity Services Engine gives Audi greater visibility into which devices are communicating on the network. The case study frames the network as foundational to reliability, security and the ability to virtualize automation safely.
IT and OT must share the architecture
The stubborn problem is organizational as much as technical. Cisco’s report found that 43 per cent of manufacturing organizations still have little or no IT/OT collaboration, a figure Pasquier admits surprised him after years of discussion around convergence. The issue persists because it involves culture, ownership and long-standing habits.
The best model is not IT taking over OT, or OT resisting IT involvement. “You need people working and collaborating with each other,” Pasquier says. “IT brings the security expertise and the way to manage the network. OT brings the knowledge of what is happening on the plant floor and how things need to be organized.”
That collaboration becomes more important as manufacturers deploy software and applications at multiple levels of the industrial architecture. Some workloads need to run beside a robot. Others can run in a factory compute room, data center or cloud. Pasquier says the next requirement is common tooling to deploy, sustain and manage the lifecycle of applications wherever they sit.
That is also where modernization becomes less about ripping out old equipment and more about designing for the next layer of capability. Most manufacturers will continue to operate mixed environments, with legacy assets, newer connected systems and different levels of digital maturity across sites. The question is not whether every system can be replaced, but whether the network architecture can support what the business wants to add next.
Pasquier recommends beginning with security assessments, then using new lines and major upgrades as opportunities to design with the next decade of requirements in mind. Machine vision, device-level analytics and higher data capture should not be treated as afterthoughts once a line is already built. If the infrastructure is designed too narrowly, the next AI or automation use case becomes harder, more expensive and riskier to deploy.
The companies that scale industrial AI will be those that design the network, power, edge compute and security architecture for change. A more deliberate architecture gives manufacturers the flexibility to modernize cells, add technologies and collect more data without disrupting production.
Pasquier’s warning is blunt. “Don’t get bored with your network,” he concludes. “Your network will be critical no matter what you do. It will be connected, it will be on the network, and if you don’t invest properly and design it properly, it will bite you back.”
That may be the clearest sign of how manufacturing has changed. The network is no longer background infrastructure. It is the foundation on which AI, robotics, machine vision, cybersecurity and increasingly autonomous operations will either scale or stall.