This is partner content, brought to you by FlexGen. Pasi Taimela is the company’s chief innovation officer.
For data center developers, recent power challenges have largely been framed as a race for megawatts. But perhaps most important is the ability to coordinate the needs of the grid, the data center energy requirements, and a variety of generation assets to operate seamlessly as one system.
Today’s AI data center power systems are being designed against a moving target. Developers are making decisions about infrastructure that will operate for decades, often without knowing exactly how much grid capacity they can rely on, what hardware will be available at what cost and on what timeline, or how compute demand will evolve over time. By 2027, Nvidia plans to support AI racks requiring one megawatt or more of power.
The rules governing these projects are moving, too. In June, FERC directed all six regional grid operators under its jurisdiction to justify or reform their rules for connecting large loads like data centers to the transmission system.
Layered on top of that uncertainty, there’s complexity. Data centers used to simply connect to the grid for pristine power backed by a system designed to withstand disturbances. That grid access is no longer a given. Now, gigawatt-scale AI facilities must design, deploy, and operate their own power systems. Campuses are increasingly combining battery storage, on-site gas and renewable generation, UPS systems, and limited grid power to get the reliable energy they need.
Coordinating all those resources to operate as a single power system while maintaining flexibility over time requires a well-planned orchestration layer — a sophisticated microgrid controller fit for the scale of an AI data center. Because of the scale of an AI campus, strong orchestration can mean the difference between cascading blackouts and a data center that runs smoothly while strengthening the surrounding grid.
Hybrid power architectures require sophisticated coordination
As much as a third of new data center power demand through 2030 could be met by behind-the-meter resources, according to McKinsey. Each resource added to the mix brings its own set of operating characteristics. Gas generators have limits on how quickly they can respond to changes in load. Batteries have state-of-charge and thermal constraints. Utility interconnections come with operating and protection requirements.
The data center itself needs tight control of voltage, frequency, and power quality to protect equipment and uptime. All of this needs to be coordinated at an unprecedented gigawatt scale, with millisecond response times, and happen simultaneously.
The orchestration layer turns those individual resources into a site-wide power system, deciding which resources should respond, when, and by how much. It consists of battery storage, power conversion systems (PCS), and energy management system (EMS) software working together.
Batteries provide a storage tank of electrons that can start filling up or emptying out in milliseconds, as needed. The PCS determines how quickly and precisely power can be converted and delivered to or from the battery. And the EMS serves as the entire site’s microgrid controller, sometimes called the master plant controller, by coordinating all campus energy resources and governing how the system operates in real time according to the power system’s control strategy.
All three components need to be planned together, and that work needs to start early. An EMS cannot call on energy that the battery wasn’t sized to provide or ask a PCS to respond in a way that it wasn’t designed to support. If hardware procurement and control architectures are treated as separate workstreams, reconciling them later in the process can add costs, integration work, and time to commissioning.
The orchestration layer also gives a data center site flexibility, so it’s easier and more efficient to make changes, such as adding new generation, securing grid interconnection, or updating commands. With sophisticated controls, a data center can evolve its power system over time without impacting campus operations or the local grid.
The moments when orchestration matters most
Beyond coordinating multiple energy assets, there are four friction points that define the design of an AI power orchestration layer: low-voltage ride-through events, AI load smoothing, peak-demand limits, and curtailment requirements.
Low-voltage ride-through helps protect both the data center and the grid during voltage dips. The battery absorbs any non-pristine power coming from the grid, then delivers pristine power to the servers. A well-functioning orchestration layer — the batteries, PCS, and EMS — can keep operations running through disturbances while preventing large data center loads from tripping offline suddenly and simultaneously.
AI load smoothing addresses rapid fluctuations in the data center’s electricity demand that can be difficult for generators and the grid to follow. The PCS and battery can immediately smooth those swings, while the orchestration layer can keep generators within their operating bands and batteries within their target state of charge levels, protecting on-site generation and the grid.
Peak-demand limits may require a campus to keep its draw on the grid below a defined threshold, even when compute demand exceeds that threshold. The orchestration layer can use on-site battery storage and generation to manage the difference.
Curtailment rules may require data centers to reduce grid demand during periods of grid stress. The orchestration layer must coordinate storage, on-site generation, and load to meet that requirement while maintaining as much critical operation as possible.
These challenges have nuanced solutions, but they all demand a multi-asset power system that can respond as one with a clearly defined software control layer. And not only that, but do it all at the same time.
Well-orchestrated data centers work with the grid, not against it
As projects move from securing power to designing the systems that will serve AI workloads for decades, developers should consider orchestration infrastructure alongside capacity, architecture, and procurement decisions, not after. Establishing the software control layer from the start gives data centers the flexibility to manage their own complex and changing energy needs while responding intelligently to conditions on the larger power system. For the next generation of AI campuses, being a good grid citizen starts with how the power system is designed.
