The rapid expansion of data centers around the globe has taken center stage in the media, and for good reason as these new data centers are fundamentally reshaping technology and the way people interact with data. However, modern artificial intelligence (AI) factories, the primary focus of this attention, are fundamentally different from the enterprise and colocation data centers that preceded them.

Traditional enterprise data centers were largely static facilities. Their energy loads were relatively predictable, and their cooling requirements were much more stable. As a result, these facilities were often constructed and managed similarly to large commercial buildings.

AI factories represent an entirely different operating model. These sites have much higher rack densities, more dynamic workloads, greater power consumption and more complex cooling systems than conventional enterprise and colocation data centers. Where a conventional data hall was underwritten at roughly 5 to 15 kilowatts (kW) per rack, rack-scale AI systems now draw 120 kW or more — well past the practical ceiling of air cooling and into direct-to-chip liquid cooling served by coolant distribution units (CDUs) and facility water loops.

In addition, multi-year interconnection queues have increasingly pushed AI factories toward large-scale behind-the-meter (BTM) power generation, operating either grid-parallel or islanded. That adds load-shedding, black-start and load-sequencing responsibilities in which generation, cooling and compute load are interlocked variables that must be controlled together rather than monitored separately (Figure 1).

This paradigm shift, from data center facility to AI factory, requires an operational shift. To operate a complex facility requires transitioning from siloed control architecture to integrated process control and automation strategies. The organizations that grasp this concept the earliest and use it to improve both project execution and lifecycle operation will be the ones securing competitive advantage in the years to come.

AI changes the nature of control

In AI factories, traditional approaches to both project execution and lifecycle operation have begun to transition as teams explore new strategies. AI workloads are highly dynamic, so power consumption shifts rapidly, with cooling demands fluctuating as workloads are assigned and executed.

To meet these challenges, operations teams must be positioned to continuously and rapidly react, but traditional architectures that require mapping across databases—along with manual engineering to create graphics, trending and alarming—were not designed for advanced coordination and dynamic optimization.

These and other challenges of traditional data silos have outsized effects as data centers scale from facilities to factories. Operators face alarm flooding conditions, and manual maintenance and engineering efforts make it challenging to maintain their competitive edge.

To meet the challenges in the years to come, AI factories do not just require more control; they require a reimagined architecture purpose-built for AI factories.

The new reality of AI infrastructure projects

New AI factories face a wide array of project execution pressures. Today’s AI projects are executed at unprecedented speed. Investments are high, and the rewards for being first to market are even higher. As a result, owners often begin construction projects with only a fraction of the final design, allowing remaining requirements to evolve throughout project execution.

In many cases, this process of finalizing the design in parallel with executing the build out has become a necessity. Equipment availability changes constantly, and supply chain challenges can force design modifications mid-project. Successful project teams will be those that can pivot to accommodate changes at every stage of execution.

However, traditional automation systems often struggle with project changes—particularly late-stage changes. Project teams need a flexible automation architecture that can absorb project uncertainty rather than amplifying it.

Consider a modern AI factory project. Modular construction, multiple phases of execution as the project expands across a campus, “build while you go” execution models, and even the potential for building onsite energy generation as a core capability are all increasingly common. These strategies create a need for automation systems that accommodate change without requiring extensive redesign at every stage (Figure 2).

Enter the modern automation platform

As flexibility becomes the key to competitive advantage, modern automation platforms have become one of an AI factory’s primary project risk reduction tools. Today’s advanced automation platforms feature built-in flexibility, empowering project teams to adjust on the fly to meet changing goals or supply chain availability.

Today’s most successful organizations rely heavily on the flexible I/O architectures offered by advanced automation systems. Electronic marshalling capability in modern automation platforms provides flexibility to add I/O anywhere in the facility without impacting the control room cabinets. This allows teams to finalize and order cabinets before every detail is known, resulting in reduced rewiring, reduced redesign, faster commissioning and easier future expansion. In addition, electronic marshalling provides the ability to repeat core designs across multiple AI factory areas or even multiple sites, while still accommodating variation between projects.

The projects that finish soonest are often the ones that can absorb change with minimal disruption. Planning from the earliest stages to leverage the flexibility advantages of a modern automation platform can provide that capability.

Traditional architecture has hidden costs

One of the key reasons that the architectures designed for enterprise data centers are insufficient for AI factories is that what works for individual equipment and small facilities does not necessarily scale well to multi-building campuses and gigawatt-scale power demands. Even if the lack of flexibility in project execution is navigable, many of the limitations of traditional architectures will show up in long-term AI factory operations.

The traditional model consists of many programmable logic controllers (PLCs) custom integrated and monitored via a supervisory control and data acquisition (SCADA) system, sitting alongside a building management system (BMS) for mechanical equipment, an electrical power monitoring system (EPMS) for power and data center infrastructure management (DCIM) tools for asset and capacity data. This often results in sub-optimal configurations, with flat control networks, large quantities of manually integrated PLCs, and a significant engineering burden—issues that will only become more complex as systems evolve and new solutions are added to the network.

In addition, the fragile, manual integration of PLCs often leads to fragmented, siloed data and multiple databases across the network. As teams add more solutions—such as layered remote access—into the mix, operational complexity continues to increase.

Modern automation architecture is designed differently

The importance of modern automation architecture in AI factories does not suggest that PLCs and SCADA no longer have a role to play. PLCs will still be critical elements of AI factory architecture for original equipment manufacturer packages, local control and high-speed applications. Similarly, SCADA will still be critical for monitoring, visualization and reporting. The question is not PLCs versus a distributed control system (DCS). The true question is, “What is the unifying operational layer seamlessly integrating all the technology necessary to run the facility?” (Figure 3)

A DCS-oriented architecture based on a modern, seamlessly integrated automation platform provides integrated data and native communications out of the box, encompassing PLC, DCS, SCADA and other automation technologies. This provides more coordinated control across the organization, built-in network segmentation and easier lifecycle management.

Ultimately, the most successful architecture is a blend of technologies united by a cohesive automation platform.

Reliability is a core differentiator

One of the key advantages of tying automation technologies together via a seamlessly integrated automation platform is the ability to drive significant improvements in uptime and return on investment (ROI). AI factories are built around expensive and sensitive technology, such as central processing units (CPUs) and graphics processing units (GPUs). Cooling capability and power usage are directly linked to output. CPUs and GPUs in AI factories quickly ramp up and down—often unpredictably—creating massive power and thermal swings. Cooling failures risk hardware damage and the associated downtime can be extremely costly, while thermal throttling degrades performance. Damaged hardware is expensive and difficult to replace, and any resulting issues related to service-level agreements can create significant financial consequences.

Data center designers have engineered redundancy into power and cooling equipment for decades, like 2N uninterruptible power system (UPS) topologies, N+1 chiller plants and concurrently maintainable distribution paths. The gap is at the control and data layer, where single points of failure, non-redundant controllers, and operational blind spots remain common. In contrast, the suppliers behind modern industrial platforms have spent decades engineering for uptime by offering redundant controllers with bumpless switchover, redundant networks and online configuration changes and upgrades that can be performed without interrupting operations.

Industrial control techniques improve dynamic load management

The extreme dynamism inherent in AI factories is exactly the type of challenge industrial process control systems were designed to solve. AI loads do not behave as steady-state workloads. Training jobs synchronize thousands of accelerators, producing megawatt-scale step loads and ramp rates measured in seconds. Demand shifts throughout the facility, cooling requirements move dynamically, and power consumption fluctuates—and running conservatively to avoid these challenges sacrifices performance and output.

An industrial automation platform is designed to support the advanced control strategies —such as model predictive control, feedforward and cascade loops, coordinated thermal and electrical load management, and others that empower facilities to operate confidently at higher utilization. These are the same technologies that have been supporting refineries, specialty chemical plants, life sciences facilities power plants and other critical infrastructure for decades—and make no mistake, AI factories are indeed critical infrastructure.

Critical infrastructure requires critical infrastructure security

It is also important to recognize that, as critical infrastructure, AI factories require a strong cybersecurity posture. Strong cybersecurity is a matter of operational discipline in addition to information technology (IT) protection. However, many traditional data center architectures place controllers directly on flat networks, exposing them to significant threats. In addition, as teams need to add remote access to systems, they further increase their exposure.

The growing convergence between operational technology (OT) and IT requires disciplined network segmentation. Modern automation platforms provide built-in segmentation, secure architectures aligned to ISA/IEC 62443 zones and conduits, secure-by-design software, unified cybersecurity management and built-in lifecycle support. Choosing systems built with security engineered into the architecture from the beginning makes it far easier to operate with minimal risk of operational disruption and/or safety incidents due to cybersecurity breaches (Figure 4).

Unshackling data

Another lesson many organizations operating AI factories are learning is that the best automation helps support broad business goals. AI factories rely on data for critical decision making. Sustainability and environmental reporting, power usage effectiveness (PUE), water usage effectiveness (WUE), emissions monitoring capacity planning, operational optimization, enterprise-wide visibility to ensure standardization and uptime, and AI-driven analytics are all highly dependent on quality, contextualized data. The siloed architectures common in traditional data centers often create siloed islands of data, duplicated effort and inconsistent reporting.

Modern integrated automation platforms provide contextualized data and easier enterprise integration via a unified data fabric to deliver enhanced decision support across the business (Figure 5).

More industrial facility than office building

Enterprise data centers shaped today’s architecture assumptions for AI factories. However, modern operation of AI factories is quickly exposing the limits of those assumptions. Today’s facilities require higher availability, greater flexibility, stronger cybersecurity, better data management and more sophisticated control strategies. Only a modern automation platform can deliver those capabilities with a seamlessly integrated solution that also maintains the flexibility necessary for success in an increasingly competitive marketplace.

The organizations that adopt an industrial automation mindset early will be better positioned to speed deployment, reduce risk, improve uptime and maximize ROI. AI factories are no longer simply places where computing happens but are instead becoming some of the most complex and mission-critical facilities ever built. Those organizations embracing the right infrastructure will be well positioned to meet those needs in the decades to come.