The new factory architecture starts with the workload
Factory architecture once sat behind the production process. Servers, networks and storage were essential but largely treated as technical infrastructure rather than a source of competitive advantage. That distinction is disappearing as operational data moves between machines, local compute, enterprise platforms and cloud environments, carrying the models and automated decisions that increasingly shape factory performance.
The change is not simply that information technology and operational technology are converging. Manufacturers are asking the combined environment to support machine vision, predictive maintenance, real-time optimization and AI inferencing. Greg Hookings, Senior Director, EMEA Sales and Service at Penguin Solutions, argues that architecture has become inseparable from manufacturing strategy because where data is processed, protected and retained now influences how quickly production can improve.
“The output manufacturers want has not changed that much,” Hookings says. “They still want products faster, more flexible manufacturing, better quality and lower cost. What has changed is the amount of intelligence that can be put into those outcomes. Vision technology is not new, for example, but the way data can now be aggregated and used to identify imperfections is becoming much more detailed.”
Convergence has changed the workload
Early discussions of IT-OT convergence focused on connecting isolated assets and extracting information from the plant floor. That work remains unfinished, but the architectural challenge has moved on. Data once collected for reporting is now training models, supporting automated decisions and feeding applications that must operate close to production.
This creates a more demanding infrastructure requirement. Conventional industrial computing is being supplemented by GPU capacity, faster networking and larger storage environments, while trained models are moving back toward the edge for inferencing. The resulting architecture must support both the scale of AI development and the operational discipline of manufacturing.
“It is a moving target,” Hookings explains. “We have gone from making manufacturing data useful to aggregating much larger volumes because AI needs more powerful computing, more storage and faster networks. On the other side, trained models are being deployed onto the shop floor for vision inspection, quality inspection and other use cases. The traditional IT components are still there, but they are being augmented by a different level of infrastructure.”
The practical consequence is that architecture can no longer be designed around one destination. Some workloads need cloud elasticity during experimentation or model training. Others belong close to the equipment because they depend on continuity, local control or immediate response. Enterprise systems still need selected operational information, but moving every data point upward creates cost and unnecessary dependence on connectivity.
For manufacturers, the more useful question is not whether cloud or edge is superior. It is what each workload requires and what happens to the operation when part of the architecture becomes unavailable. That shifts the discussion from technology preference to production consequence.
Data placement is a strategic decision
Ownership is one of the factors bringing architecture into the boardroom. Manufacturing data can expose process knowledge, product designs and intellectual property that distinguish one company from another. Public cloud services may provide a fast route into AI, but the long-term economics and control model become more difficult as workloads scale.
“Who owns the data is very important,” Hookings continues. “A company may be comfortable putting ordinary email traffic into a public cloud, but it will think differently about product IP or the information behind a medical development. Architecture matters because manufacturers have to decide where that infrastructure sits, how it is sized and how much capacity they will need three to five years from now.”
Hybrid architecture is often presented as the natural answer, although it describes a series of choices rather than a settled design. A manufacturer may begin with a small private GPU cluster for experimentation or inferencing while using external capacity for larger tasks. As use cases mature, the balance may change because of cost, performance or the value of keeping critical knowledge under direct control.
Hookings sees hybrid infrastructure as an entry point rather than a permanent formula. “There is a large upfront investment in AI infrastructure, and manufacturers are competing for expensive components with some of the biggest companies in the world,” he says. “A hybrid approach may mean starting with a small cluster and relying on cloud capacity for larger processing. As the strategy develops and the data becomes more critical, companies will want greater control over both the information and the cost.”
That control depends on avoiding indiscriminate data movement. The plant floor generates far more information than the enterprise needs to store centrally, and much of it loses value quickly. Processing data where it is created allows manufacturers to filter, contextualize and act before escalating selected information to higher layers.
“The basic principle is to process data where it is relevant to the system,” Hookings says. “The edge should be the first point for processing key information, rather than creating huge volumes of data that are constantly moving around the organization. The information that matters can then move upward into other systems. You do not need to ship everything to the cloud if decisions can be made locally.”
Resilience must be designed in
Industrial architecture differs from conventional enterprise IT because application availability is tied directly to physical output. A failed server may stop a batching process, interrupt automated picking or leave operators without the information required to run a line. Redundancy, failover and maintainability therefore have to be treated as production requirements rather than infrastructure enhancements.
A specialty chemicals manufacturer working with system integrator EOSYS illustrates the problem. Its slurry batching process depended on nine physical hosts running separate Rockwell Automation workloads, including batch, HMI, historian and SQL applications. The aging environment was operating on Windows Server 2008 R2, the backup domain controller was not synchronized and the company had unknowingly been running without effective redundancy. When a final server failure took days to recover, the weakness of the architecture became an operational crisis.
EOSYS consolidated the workloads onto a fault-tolerant edge platform during a ten-week modernization program, with the final cutover completed in two days. The new environment virtualized the critical applications while reducing the burden on a two-person OT team that had previously been called out to reboot and repair systems. According to the case study, the plant has experienced no unplanned application downtime since the deployment in 2022.
The lesson is not that every manufacturer should adopt the same hardware. It is that consolidation only creates value when it removes single points of failure rather than concentrating them. Virtualization can simplify an industrial estate, but the underlying compute still must survive component failure, support recovery and remain manageable by the people available at the site.
Mitsubishi Heavy Industries (MHI) faced a related challenge when deploying an intelligent automated picking system for a large beverages company. Its warehouse execution, optimization and control applications needed continuous access to real-time data to coordinate automated guided forklifts and vehicles. Traditional PC clusters lacked seamless failover and required more specialist intervention than the warehouse could support.
MHI deployed the three core applications as isolated virtual machines on a dual-node edge platform. During tests simulating power loss and hardware faults, the system maintained uninterrupted operations, allowing the automated picking process to continue while failed components could be addressed without shutting down the application environment. The architecture combined local processing, fault tolerance and simplified administration because all three were necessary for the automation to be operationally credible.
Architecture becomes a manufacturing capability
The next phase of industrial architecture will be shaped by inferencing and increasingly autonomous operations. Manufacturers are likely to begin with contained use cases where value is visible, such as inspection, predictive maintenance or logistics optimization. Successful deployments will create demand for greater private computing capacity and more repeatable infrastructure across plants.
“I expect manufacturers to start with smaller clusters and experiment with their own AI infrastructure,” Hookings adds. “It will begin with a use case, perhaps quality control or identifying anomalies in a logistics center. Once they see competitive value from inferencing inside the business, that will structure further investment and lead to larger commitments.”
Those commitments cannot be separated from architecture governance. Each new model introduces questions about data lineage, cybersecurity, update frequency and what happens when a service or component fails. More software-driven manufacturing also shortens product and improvement cycles, which places additional pressure on infrastructure to support change without destabilizing production.
The new factory architecture is therefore unlikely to be defined by a universal edge-to-cloud template. It will be defined by deliberate workload placement, clear ownership of critical data and resilience matched to operational consequence. Manufacturers that treat those choices as an IT procurement exercise may still deploy impressive technology, but they will struggle to scale it reliably.
Those that treat architecture as a manufacturing capability will be better positioned to adopt AI without surrendering control of the data, costs or production environment that makes it valuable. The factory of the future will not send everything to one place. It will know what must remain close to the process, what should move elsewhere and how the operation continues when any part of that system fails.