Why edge AI cannot scale one factory at a time

The fastest way to undermine edge AI is to let every factory solve the problem for itself. A plant may prove that computer vision can identify defects, predictive models can anticipate equipment failure, or real-time analytics can improve throughput, but a successful local deployment is not the same as a scalable manufacturing capability. When each site selects its own hardware, applications, security controls and management processes, early progress can harden into another layer of industrial fragmentation.

That distinction matters because edge computing is no longer primarily an infrastructure decision. Pierluca Chiodelli, vice president of product management at Dell Technologies, argues that it has become an operational requirement as manufacturers attempt to turn growing volumes of plant data into decisions that improve quality, uptime, efficiency and safety. “Modern factories generate continuous, high-volume data from machines, sensors, vision systems, robotics and production environments, and manufacturers increasingly need to act on that data in real time,” he says. “The use cases that move the needle on the plant floor are AI-driven. Computer vision can catch defects before they ship, predictive maintenance can flag a failing motor days ahead of unplanned downtime, and process models can optimize yield and energy use shift by shift. These are low-latency applications where operators make decisions in seconds, so AI has to run where the data is created.”

Sending every video stream, machine signal and sensor reading to a centralized cloud can add latency, increase data movement costs and create dependency on external connectivity. What is changing is the scale of the ambition. Manufacturers are no longer considering one application on one production line, but how AI workloads can be deployed, secured, updated and governed across distributed locations.

That is where many edge strategies remain underdeveloped. Dell research cited by Chiodelli found that 94 percent of companies intended to deploy edge solutions, but only 10 percent had reached advanced production, while approximately 85 percent said they needed a simpler approach to edge setup. The gap is not evidence that manufacturers lack compelling use cases. It suggests that deployment complexity is preventing proven ideas from becoming repeatable operations.

From equipment project to operating model

The most persistent misconception, according to Chiodelli, is that edge computing is essentially a hardware placement exercise. Installing a server in a factory may provide local processing capacity, but it does not create the management structure required to operate distributed intelligence securely and consistently.

“Edge creates lasting value when it is treated as an operating model that includes secure onboarding, centralized lifecycle management, application orchestration, IT and OT convergence, data integration, and support for both legacy industrial systems and newer AI workloads,” he says. “It is tempting to spin up a custom, siloed solution every time a plant needs a new efficiency gain. But those snowflake one-offs add up fast, and what is left is a jumbled model with no real integration or cohesion. Approaching edge as a collection of isolated pilots recreates the same problems that slowed adoption in the first place, including sprawl, fragmented technologies, security exposure and difficulty scaling from one site to many.”

Factory-level autonomy is valuable when it allows teams to respond to local production needs, but it becomes counterproductive when every deployment requires a separate architecture and support model. A pilot dependent on specialist knowledge, manual configuration and bespoke integration may work in one facility while remaining impossible to reproduce elsewhere.

The alternative is to standardize the elements that do not need to be unique. Infrastructure can be provisioned through common processes. Applications can be packaged into reusable blueprints. Security policies, device onboarding and lifecycle management can be centrally governed, while workloads continue to run locally. Plants retain the low latency and resilience of edge processing without becoming disconnected technology islands.

Treating edge and cloud as competing destinations creates a false choice. Time-sensitive inferencing and operational decisions belong close to the process, while centralized environments remain useful for model training, fleet-wide analysis, governance and enterprise integration. The question is which work should happen where and how the environments should be managed as one architecture.

Prime Vision, a computer vision and robotics specialist serving postal, parcel and logistics operations, demonstrates that division of labor. Sorting centers generate large volumes of image and tracking data and must make routing decisions in milliseconds. Prime Vision and Dell developed a hybrid model in which real-time processing takes place locally, while centralized cloud resources are used when broader analysis adds value.

“Local device management handles the speed-critical work and centralized data analysis handles the heavier compute,” Chiodelli explains. “The model reduces cost and latency while increasing efficiency and security. It also improves control and compliance because data can remain inside the customer’s facilities, and it integrates with existing infrastructure rather than demanding a wholesale replacement. A sorting center and a plant floor face the same core problem: high-volume data from cameras, sensors and machines must drive decisions faster than a round trip to the cloud allows.”

AI use cases expose the architecture

Computer vision remains one of the clearest examples of why this architecture matters. Manufacturers have used vision systems for inspection, assembly verification and traceability for years, but AI allows those systems to make more complex judgments, identify subtle variation and adapt as products or processes change. The value depends on acting immediately. A defect identified after production has continued for another hour is information, but it is not effective process control.

Predictive maintenance follows the same pattern. Models need access to live equipment and sensor data, and the useful output is an intervention before failure rather than an explanation afterward. Real-time operational intelligence, condition monitoring and digital twins also become more valuable when they are connected to current production behavior instead of relying on delayed or incomplete data transfers.

Security is equally important because edge AI expands the number of systems operating outside a conventional data center. Manufacturers must protect OT environments, connected devices and the infrastructure hosting AI workloads. Physical safety use cases may also rely on AI to distinguish people from robotic movement or identify unsafe activity around automated equipment. The architecture therefore has to support operational resilience and cybersecurity at the same time.

“What ties these use cases together is the need to run AI across a large amount of distributed hardware,” Chiodelli says. “They suit edge environments because they require low-latency response, often generate too much data to move upstream cost-effectively and must keep running in distributed sites where security, resilience and local autonomy matter. Edge is where AI becomes operational. It is not insight after the fact, but inferencing, action and automation during production.”

Successful use cases can still conceal weak foundations. A vision application may achieve its purpose while introducing another management interface, data pipeline or security exception. Repeating that pattern across maintenance, energy, safety and quality creates technically successful projects that the enterprise cannot efficiently operate. As use cases multiply, common orchestration and governance become more important.

Scaling the deployment rather than the pilot

Eaton provides a clearer indication of what a repeatable model can deliver. The power management company operates more than 230 manufacturing facilities and needed to modernize a distributed footprint, connect IT and OT, reduce siloed systems and accelerate deployment while addressing cybersecurity, sustainability and skills constraints.

Eaton deployed Dell Distributed Private Cloud as an edge operations layer across its manufacturing environment. The approach introduced centralized management, zero-touch provisioning, application orchestration and reusable digital blueprints that could be deployed consistently across sites. It also connected with Eaton’s industrial software environment to support predictive maintenance, analytics, workload management, energy optimization and operational visibility.

The most significant result was not an isolated improvement in one production process. Software deployment time fell from between three and six months to a matter of days, representing a reduction of more than 90 percent. Eaton is using the model to support the modernization of more than 100 factories, with deployments planned across 90 priority locations.

“That example shows how distributed intelligence becomes valuable when it is deployed as a repeatable operating model rather than a one-off pilot,” Chiodelli notes. “The benefit was not simply putting hardware in plants. It was creating a secure and scalable way to push intelligence and automation closer to production across a global manufacturing footprint. That helps reduce disruption and downtime, improve energy use and support wider sustainability goals, but the foundation is the ability to repeat the deployment.”

For manufacturers, this reframes the measure of success. The question is not only whether an edge AI application improves inspection accuracy, detects failures or optimizes energy consumption. It is whether the company can deploy the next application faster, manage it with the same controls and reproduce the outcome in another plant without rebuilding the underlying system.

The manufacturers that gain most from edge AI will not necessarily be those with the greatest number of pilots. They will be those that convert local innovation into a governed, reusable capability. Processing intelligence close to production solves the latency problem, but only a common operating model solves the scale problem. Without it, every successful pilot risks becoming another exception. With it, each deployment becomes a foundation for the next.