News

AI data center load profiles hit generation with fast, repetitive load steps, and rotating machinery feels every one of them. A grid can average those steps out across thousands of other loads. A behind-the-meter plant cannot. The turbines carry the facility, so the facility behavior becomes the turbine’s duty cycle.

This article explains how AI load differs from conventional data center load, what steps do to a gas turbine fleet, and why the answer depends on the whole plant, not the machine on the data sheet.

How does AI load differ from traditional data center load?

A traditional data center draws a fairly steady load. Servers and cooling change slowly, and UPS systems smooth what the grid sees. Operators plan around a load curve that behaves like a ramp.

AI training does not behave that way. The North American Electric Reliability Corporation’s (NERC) July 2025 white paper on emerging large loads describes AI training executed in large clusters as showing rapid fluctuations during training and while checkpoint progress is saved. It shows a 50 MW block of a 200 MW AI training data center changing demand at a rate of 1.9 per unit per second for about 250 milliseconds. It also reports that the shift between training and checkpoint saving can take under a second. While training runs, the demand shows jittery spikes up and down (NERC, Characteristics and Risks of Emerging Large Loads).

Two qualifiers matter:

  1. Inference differs from training. The same white paper notes that the rapid ramps seen in training and checkpoint cycles are not observed in current inference methods.
  2. These are grid-side observations. The recordings come from facilities connected to a bulk power system. They show what the load does. They do not show what a given turbine fleet does in response.

The practical result is a load made of steps, not ramps. A step can arrive faster than a turbine is allowed to change output, and it repeats.

AI compute load steps compared to load-smoothed generation response in a behind-the-meter data center plant.

What can go wrong on the generation side?

The list below is SISO Engineering’s own reading of that load behavior, not measured field data from a specific plant. Each item needs a dynamic system study on the actual configuration, with the limits of the original equipment manufacturer (OEM) in hand.

  1. Frequency excursion. In an islanded plant, no large grid holds frequency. A step that outruns the fleet’s response pulls frequency down, and under-frequency protection can shed load or trip units.
  2. Turbine control limits. A gas turbine accepts load only as fast as combustion stability, exhaust temperature, and compressor margin allow. The control system enforces those limits. A step larger than the allowed rate has to be absorbed elsewhere.
  3. Load rejection. Load can also vanish. NERC’s incident review of a July 2024 event describes about 1,500 MW of data center load transferring to backup power during a six-fault sequence on a 230 kV line (NERC, Incident Review: Considering Simultaneous Voltage-Sensitive Load Reductions). On a grid, that raised frequency slightly. Behind the meter, the same behavior is a sudden load rejection on the turbines.
  4. Uneven load sharing. A fleet of mixed sizes, ages, or makes will not divide a step evenly. The unit with the fastest response takes the most, and it may reach a limit first.
  5. Repetitive thermal and mechanical duty. Repetition matters as much as magnitude. Cycling that a unit tolerates occasionally may affect maintenance intervals when it happens all day. The right question for the OEM is how much cycling is acceptable, at what rate, and over what period.
  6. Torsional interaction. Repetitive load at frequencies near a shaft’s torsional modes can excite oscillation. This is a study item, not a number to guess.
  7. Protection misoperation. Settings tuned for conventional loads may trip on behavior they were never meant to see.

None of these items is exotic. Each is a known phenomenon meeting a load that the original design did not anticipate.

How do designs absorb fast, repetitive steps?

Three responses show up in most designs, and they work together.

  1. Load smoothing. Reduce the step before it reaches the generation bus. This can happen inside the facility, through how the compute load is scheduled and managed, and at the plant, through storage.
  2. Battery pairing. Let the battery energy storage system (BESS) take the step and let the turbine take the trend. The battery responds in a fraction of the time and carries the fast component. The turbine picks up the slower change and returns the battery to its operating range.
  3. Controls coordination. Make the turbine controls, battery controls, and plant-level controls agree on setpoints, response times, and priorities. This is where most designs are weakest, because each supplier tunes its own equipment.

SISO Engineering’s hybrid power systems page covers turbine and battery storage integration in more detail.

A note on who does what belongs here. The BESS, the medium-voltage uninterruptible power supply (MV UPS), and the static synchronous compensator (STATCOM) come from their manufacturers. SISO Engineering helps integrate their controls and data so they work together, and analyzes the resulting system behavior. The battery supplier sizes and warrants the battery. The turbine OEM governs the turbine’s operating limits.

Why is this a system property, not a spec-sheet property?

A turbine data sheet lists capability under stated conditions. It does not say what the machine does when a battery, a protection scheme, and an AI facility all act on it at once.

Four facts explain why the plant, not the part, decides the outcome:

  1. Response depends on controls tuning. The same turbine behaves differently under different governor, limiter, and setpoint settings.
  2. Response depends on the other assets. Battery response time, inverter behavior, and protection settings change what the turbine sees.
  3. Response depends on the load. A training facility and an inference facility present different profiles to the same fleet.
  4. Nobody owns the interaction by default. The risk is rarely that one supplier is incompetent. The risk is that their assumptions do not align, and no one owns the seams.

This is why dynamic system studies and integrated test scenarios matter. They test the assumptions together, before energization, when a change costs a revision and not an outage.

Where does SISO Engineering fit, and where does it stop?

SISO Engineering is an independent controls and integration firm headquartered in Houston, working wherever the equipment runs across North America, South America, and Europe. SnapShot, its fleet intelligence layer, comes first. An AI facility is the mixed fleet it was built for: turbines, reciprocating engines, batteries, and balance-of-plant systems, often from different OEMs on different controls. Different assets. Different OEMs. Different controls. One operational intelligence layer.

SnapShot gives asset-level intelligence on each unit and fleet-level visibility across all of them. Its cross-OEM intelligence creates a common operational view across equipment from different manufacturers. The deviations, performance degradation, abnormal operating behavior, and recurring alarm conditions that a turbine control screen alone can hide become visible in one place. For a conventional power plant, lost generation is a commercial problem. For a data center, loss of generation can become a mission-critical infrastructure event.

Integration services and commissioning support are how SISO Engineering gets into a project. SISO Engineering’s background is gas turbine controls and automation. On a project like this, SISO Engineering coordinates controls engineering across platforms, supports integrated commissioning tests and baseline capture, and helps set the data and monitoring strategy. SISO Engineering’s TDCS (turbine data collection system) is built for data access from turbine control systems. That matters because NERC’s white paper notes that without high-speed recording data, diagnosing the root cause of a large-load event may be difficult or impossible.

Each capability has a place. SnapShot sits above the control and data layers. Plant control and asset dispatch stay with your distributed control system (DCS) and energy management system (EMS), and full plant optimization remains with your control and operations teams, supported by SnapShot’s insights. SISO Engineering works alongside the engineering, procurement, and construction (EPC) contractor, the owner’s engineer, the OEMs, the utility, and the lender’s independent engineer. The OEMs commission their own equipment.

For a related credential, see the Sentinel Energy Center case study. SISO Engineering’s controls work there covered the interaction between battery systems, turbine controls, and balance-of-plant equipment for a black-start application. It is a battery-plus-turbine integration example, not an AI data center project.

Next step: review the architecture against the load

If your plant will serve an AI facility, test the architecture against the load before the critical decisions are locked in. Integration services are how SISO Engineering gets into a project. It produces a risk register you can use with your EPC, OEMs, owner’s engineer, and investors, whether or not SISO Engineering continues into later phases. For the broader scope, see behind-the-meter data center power.

Talk to a SISO Engineer →

SISO Engineering’s goal is to strengthen the work of every partner on your project, making the entire power system more reliable, understandable, and operable.