Why industrial connectivity belongs in your edge AI strategy

When enterprises plan an edge AI initiative, the first questions often concern models, compute, and data. Where should inference run? What hardware does it need? How will the application scale across sites?

An equally important question is how the devices, machines, and computing systems behind these applications will stay connected, and whether that connectivity can support the demands of AI-driven operations.

That question becomes more consequential as AI moves from generating insights to taking action. Identifying a defect requires access to inspection data. Coordinating equipment or directing a mobile robot also requires timely communication between intelligence and the physical world.

At ZEDEDA’s TRANSFORM 2026 conference, operators and technology providers described how that shift is taking shape. Their examples showed what edge intelligence enables in practice. Celona’s Puneet Shetty addressed a related architectural requirement in his presentation, "The ‘Last Mile’ of Edge AI”: connecting that intelligence to physical operations.

What Edge intelligence enables

Edge AI brings processing closer to where data is generated and decisions are needed. In industrial settings, it supports applications that inspect products, monitor equipment, guide work, and coordinate machines. Physical AI extends intelligence into systems that sense and act in the physical world.

The TRANSFORM discussions illustrated several aspects of this transition.

HP described a progression from reactive alerts toward agents capable of making local decisions, such as quarantining a part or stopping a production line. In one early customer example, HP reported more than 30 agents and four large models running locally, processing 11 million tokens in 90 minutes without slowdown. The example illustrates the sophistication of workloads moving closer to operations. 

Energy technology company SLB, formerly Schlumberger, described autonomous well sites where equipment coordinates in real time - for example, stopping fluid injection when an extraction pump fails. Here, intelligence supports a response involving several pieces of equipment.

Streamline Innovations described management by exception across approximately 70 remote hydrogen sulfide removal plants in seven states, reporting 99.95% uptime. Small teams oversee the facilities and intervene where attention is needed, illustrating the importance of managing distributed operations as a fleet.

Toyota Motor North America discussed edge AI for vision inspection and data filtering across 14 North American plants. Its emphasis on standardizing the platform while accommodating individual plants points to another requirement: making local applications repeatable across different environments.

These examples show intelligence becoming more closely integrated with physical work. They also show that successful edge deployments involve more than selecting a model. Local processing, equipment integration, fleet management, and communications all contribute to the operating system around the application.

Local intelligence still needs local connectivity

Scaling remains challenging. In its session, Chick-fil-A cited a figure of approximately 11% of edge projects reaching full production - a reminder that moving from a promising application to a repeatable deployment involves substantial work.

Connectivity is one part of that work, with different requirements inside and outside the site. 

The edge-to-cloud connection supports activities such as remote management, software updates, and data exchange. Decisions about which data moves can materially affect bandwidth demand. In the event’s edge-to-cloud data movement session, one example reported a 62% bandwidth reduction from sending metadata instead of raw video.

That is a workload-specific choice: teams need to establish which information must be retained and transmitted to support the application.

The local network serves another purpose. Puneet described an industrial AI stack comprising devices, edge infrastructure, and the connectivity joining them. Devices collect data, edge infrastructure processes it, and the network carries information and instructions between the devices and the edge infrastructure.

Moving inference on-site can reduce dependence on the cloud connection. It does not eliminate communication requirements between local compute and the cameras, sensors, or machines it serves.

A vision application still needs its images. A fleet-management system still needs information from moving robots. An application coordinating equipment needs a dependable path for the information and instructions on which that coordination relies.

When connectivity becomes an operational dependency

People routinely compensate for imperfect connectivity. We refresh a page, reconnect to a call, or move somewhere with a stronger signal. 

Automated systems respond differently. As Puneet explained, an automated guided vehicle experiencing a connectivity interruption may stop, disrupting the operation it supports.

The consequences depend on the application. An interruption might delay a data upload, interrupt a workflow, or cause equipment to enter a safe state. Understanding that behavior is part of designing the system.

Compute placement and connectivity therefore need to be considered together. Shared on-site compute depends on local connections to equipment. Processing aboard a robot can support local autonomy. It can also add hardware, power, and maintenance requirements across a fleet. Communication may still be needed for coordination, supervision, and updates.

Planning should bring together three decisions: where processing runs, how information moves, and how the application behaves when communication degrades. Establishing those relationships early gives equipment, AI, and network teams shared requirements.

Four requirements for industrial wireless

Puneet organized the wireless design challenge around four dimensions: coverage, mobility, reliability, and security.

1. Coverage must reach the places where work happens. Factory floors, warehouses, refineries, and outdoor facilities present different obstacles. Metal, concrete, moving equipment, and changing inventory all influence wireless design. The network design needs to be capable of supporting operations as conditions change. 
2. Mobility requires consistent performance along an asset’s route. Puneet described a Celona customer’s automated distribution center where mobile robots travel horizontally and vertically through a six-level metal structure, at speeds of 20–25 mph. Coverage at a stationary test point alone would not establish whether the network supports that operation.
3. Reliability includes predictable performance. Applications have different tolerances for delay, variation in latency, and interrupted communication. Those requirements should be defined for the workload and tested with the equipment involved. Puneet described a medical manufacturer that spent two years testing private 5G connectivity for programmable logic controllers, or PLCs. Its objective was greater manufacturing flexibility and less dependence on fixed cabling. The evaluation illustrates the rigor involved when wireless becomes part of an operationally critical application.
4. Security must preserve operational boundaries. Shared wireless infrastructure needs to support the segmentation and IT/OT separation required by the teams responsible for those environments.

Why converged wireless matters

Industrial environments combine existing equipment, new automation, mobile assets, and people. Their connectivity requirements span several technologies.

Private 5G supports industrial coverage, mobility, and traffic-management requirements. Wi-Fi provides high-capacity connectivity for a broad range of devices and applications. Public cellular supports connectivity beyond the site, while satellite can provide external connectivity for remote operations.

In his presentation, Puneet described mining environments using wireless connectivity for autonomous equipment, with Starlink providing wide-area backhaul. The local network and external connection serve distinct purposes within the same operation.

Using complementary technologies also creates an operational requirement: teams need to manage them coherently. Adding connectivity options can otherwise introduce more systems to configure, secure, and troubleshoot.

Convergence brings those technologies into a coordinated architecture, with management, policy, and security considered across the environment.

One platform for the last mile

Celona Orion is an agentic converged wireless platform that brings together private 5G, Wi-Fi, public cellular, and satellite connectivity into a unified platform for industrial operations.

Celona Orchestrator AI provides intelligence and orchestration for that environment, helping teams manage complementary wireless technologies as a coordinated system. Security is built into the architecture, including SIM-based identity, OT microsegmentation, and IT/OT separation.

For enterprises pursuing edge AI, this connects wireless design to the operational requirements Puneet outlined and provides a foundation that can evolve as automation expands.

Start with the application 

The industry perspectives at TRANSFORM demonstrate the opportunities emerging around edge intelligence and the operational discipline needed to deliver them.

Before selecting the architecture, ask what information must move, between which devices, and how quickly. Identify which assets move, what happens when communication is interrupted, and how the system will be secured and managed.

These questions help teams connect AI ambitions to operational requirements and evaluate the complete system before expanding deployment. As Puneet put it: The edge intelligence is only going to be as good as the connectivity behind it.

Watch Puneet Shetty’s TRANSFORM 2026 session, “The ‘Last Mile’ of Edge AI.”

ZEDEDA is a Celona alliance partner.