Industrial AI is becoming hybrid by design
Manufacturers spent much of the past decade centralizing data so it could be analyzed across plants, business functions and supply chains. That architecture remains valuable, but it becomes less convincing when an AI system has only seconds, or fractions of a second, to influence what is happening on a production line. Sending every decision through a distant cloud environment can introduce dependencies that factories cannot always tolerate, particularly when connectivity is interrupted or an operational process needs to continue regardless of what is happening elsewhere.
Saleh Al-Nemer, Regional Enterprise Architect and Distinguished Technologist at HPE, sees the debate shifting away from choosing between edge and cloud. The more useful question is how manufacturers divide intelligence between them so that each part of the architecture does the work to which it is best suited.
“Historically, the discussion was edge versus cloud, but now it is about how the edge and cloud work together,” he says. “The edge is important where the operation is time-sensitive, where you need local autonomy and where data locality matters. The cloud can provide the central controls and model training without affecting factory operations. It also allows manufacturers to bring information together across sites so engineering and product teams can use it over a much longer horizon.”
Machine vision shows why the location of intelligence can become critical. If a system identifies a defective component only after it has passed through several further production stages, the manufacturer may already have spent additional energy, machine time and material on something destined for scrap. Bringing the decision closer to the process allows the product to be diverted or corrected while there is still an opportunity to change the outcome.
Al-Nemer points to an application involving green anodes emerging from a furnace. AI-based inspection determines whether an anode is suitable to continue through the manufacturing process or should be redirected before further work is carried out. “If the product is not ready for the next stage, you can stop it there rather than putting it through another 25 steps in the factory,” he says. “That can save millions of dollars and a significant amount of energy. The important point is that the insight must arrive in time to influence the process. If it arrives afterwards, you may understand what went wrong, but you have already lost the opportunity to change the outcome.”
Put the decision where it can change the outcome
Where an AI workload sits is ultimately an operational risk decision as much as a performance one. Al-Nemer argues that manufacturers should identify which functions must remain available when connectivity degrades, rather than deciding architecture purely based on latency.
“Resilience is more important than latency in many critical operations,” he says. “You need to design the environment so the factory can continue safely when connectivity is limited. If you lose the network connection to the operating engine or dashboard, what happens to the factory floor and the process? You need to know that production can continue within defined limits rather than making connectivity a condition for operation.”
This makes local compute more than a way to accelerate AI. It provides a degree of operational independence, allowing essential decisions to continue at plant level while centralized systems handle activities that can tolerate delay or temporary disconnection.
Bosch Digital Twin Industries provides an example of that principle in asset management. Its digital twin-based industrial asset performance management solution uses edge AI to process equipment data locally and provide real-time insight into asset health. Reduced dependence on cloud connectivity allows the system to operate in remote or restricted environments while supporting predictive maintenance and earlier intervention when equipment begins to deteriorate.
The attraction of edge intelligence is not that every model should now be pushed onto every machine. AI workloads have different requirements, and manufacturers need to distinguish between training, inference and the operational decisions that genuinely need to happen close to the process.
Al-Nemer describes a developing architecture in which larger models can maintain a wider view of the business while specialized models are distributed closer to machines or manufacturing steps. Training can remain centralized where greater computing resources and accumulated data are available, while inference happens locally when production depends on an immediate response.
Edge intelligence still needs enterprise control
Moving AI closer to production creates another challenge if every factory builds its own isolated environment. Local autonomy cannot come at the expense of common governance, particularly when manufacturers need to update models or apply security policies across multiple plants.
Al-Nemer compares the approach with the way telecommunications networks manage large numbers of distributed cell sites. The local environment retains the ability to operate, while a separate management layer maintains oversight of the wider estate. “You effectively have two layers of operation,” he explains. “One allows the factory to operate with the compute and IT capability it needs locally. The second makes sure that environment is safe and working under the same governance as everywhere else. That covers areas such as security, authentication and authority management, so local operation does not mean losing enterprise control.”
Networks become increasingly important as more information is brought into that architecture. Individual OT messages can be small, but the frequency of communication grows quickly when machinery data is combined with machine vision and other sensor information.
In one factory modernization Al-Nemer describes, a network originally built around 100 Mbps connections was being upgraded to between 1 and 10 Gbps within parts of the plant, with 100 Gbps connectivity across the wider manufacturing environment. The requirement was driven less by any single industrial protocol than by the cumulative effect of more connected processes and greater use of data-intensive applications.
This is where the idea that manufacturers can simply add AI software to an existing plant starts to break down. Factory-floor intelligence depends on an infrastructure capable of supporting the model, moving the necessary data and continuing to operate in an industrial environment that may be far less forgiving than a conventional data center.
The factory infrastructure must catch up
Compute located close to production must contend with heat and other environmental conditions that enterprise infrastructure is not designed to experience. More demanding AI models may also require GPUs near the process, adding new requirements around power and cooling while potentially exposing hardware to vibration or other industrial conditions.
Not every application requires that level of infrastructure. Al-Nemer describes smaller edge devices that can collect machine data and run narrowly defined models, while more computationally demanding manufacturing processes may justify heavier AI infrastructure closer to the line.
Existing machinery presents a different problem because many plants contain assets designed long before anyone expected them to participate in an AI architecture. An older conveyor or production machine may have no native Ethernet connection or usable digital interface, leaving manufacturers to decide whether to retrofit enough sensing and control capability to make the equipment visible or replace it.
“You can add vibration or motion sensors to older equipment and change an analog control panel so you can begin collecting data and interacting digitally with the machine,” Al-Nemer says. “But it is not cheap, so the manufacturer has to decide whether modernizing that equipment is worthwhile or whether replacing it gives a better return. In some industries, such as mining, you may have equipment worth millions that is very difficult to replace, so adding those capabilities can make much more sense.”
Those decisions make infrastructure strategy part of the industrial AI business case rather than a technical issue to address after the use case has been selected. The value of a vision model or predictive application can be undermined if the factory cannot provide the compute environment or connectivity needed to operate it reliably.
Al-Nemer also sees a skills issue becoming increasingly important as manufacturers try to build for an AI landscape that will continue changing. Smaller models are already becoming capable of running on hardware that would have been inadequate only a few years ago, making it risky to design infrastructure around the requirements of a single generation of AI technology.
“The bigger investment factories need to make is in the people who will run and manage this,” he says. “There are not enough manufacturing subject-matter experts who also have a strong understanding of what AI can do for them. Most organizations are still exploring that. The technology will keep changing, so manufacturers need people who understand the process well enough to decide where AI genuinely creates value.”
Industrial AI is therefore moving closer to production without abandoning the cloud that manufacturers have spent years building around. The architecture is becoming more distributed because different decisions belong in different places. Training and enterprise-wide learning can remain centralized, while the intelligence that protects continuity or changes a production outcome must be available where the process is happening.
The manufacturers that manage that balance successfully will not treat the edge as another destination for workloads. They will design AI around the operational consequence of each decision, keeping intelligence close enough to act while retaining the wider control needed to improve it over time.
