An AI data center presents two power problems at once. Graphics processing unit (GPU) and tensor processing unit (TPU) clusters can swing tens of percent of facility load in under a second, every 20 seconds or so, thousands of times a day. The grid sees those changes at the connection. The compute equipment also depends on continuity of supply. A voltage sag that trips rack power supplies forces training to restart from its last checkpoint.
Utilities respond to changing demand with ramp and variability limits in interconnection agreements. Large loads also face ride-through rules. In ERCOT, Nodal Operating Guide Revision Request (NOGRR) 282, in force since August 1, 2026, sets voltage and frequency ride-through requirements for large computational loads: under Nodal Protocol Revision Request (NPRR) 1308, sites of 75 MW and above where at least half the demand is computational load. Those requirements belong in the power design from the start.
Seven reasons, in the order they come up
Interconnection risk
Grid operators require large computational loads, 75 MW and above in ERCOT, to demonstrate ride-through capability before they connect. A power smoothing battery answers that requirement in the interconnection study and can move the approval forward. For most data center projects, the interconnection date sets the schedule. The battery is one of the few items in the design that reduces the risk to that date. See data center interconnection.
AI load pulse management
AI workloads create power transients from idle to peak in under one second, every 20 seconds or so, thousands of times a day while a job trains. These load pulses cause voltage instability and flicker that utilities will not tolerate at the point of interconnection without mitigation. The battery takes the pulses and the grid sees a flat draw. See data center power smoothing.
Generator protection
AI pulse loads create torsional stress on the shafts of prime movers, both turbines and reciprocating engines. A site that runs AI load with behind-the-meter generation needs a battery for load smoothing, even with a partial grid connection, or the generators carry the pulses. The battery protects generation that shares its medium-voltage bus. In a parallel connection, low-voltage standby generators behind a hall’s own transfer switch are outside the battery’s path once the hall transfers.
Ramp rate support
Reciprocating engines cannot ramp fast enough for AI loads and gas turbines are slower. Fuel cells and small modular reactors ramp more slowly still. The battery handles the rapidly changing ramp rate that generators cannot, and holds the grid draw to the rate the utility allows.
Ride-through requirements
ERCOT NOGRR 282 with NPRR 1308 mandates voltage and frequency ride-through for large computational loads of 75 MW and above. Outside ERCOT, the North American Electric Reliability Corporation (NERC) issued a Level 3 alert on computational loads on May 4, 2026, and the Federal Energy Regulatory Commission (FERC), in Docket RD26-7-000, has directed NERC to file reliability standards for computational loads by December 31, 2026. In dual conversion the battery carries the load through the fault and the grid-side converter recovers after it. In a parallel connection the hall’s own UPS carries the computers through a deep sag. In the scenario model, blocks sized for the hall load return the site’s draw to its pre-fault level within about one cycle of the voltage returning, well inside the 0.5 s NOGRR 282 allows. The interconnection study confirms the recovery for each site. See NOGRR 282 ride-through.
Backup power and UPS replacement
In a dual-conversion block, grid-forming backup protects IT loads from voltage sags, frequency deviations and outages. Zero transfer time keeps servers inside the ITIC tolerance curve, the industry curve for how long equipment survives a voltage deviation. The block can take the place of the hall’s conventional UPS, and of its standby generators only where the battery duration, 1 to 4 hours or custom on request, covers the site’s continuity requirement, which is the owner’s decision. A parallel connection keeps the hall’s UPS and generators. See the medium-voltage UPS.
Energy cost
Four coincident peak (4CP) and five coincident peak (5CP) charges, energy price spikes, demand response and wholesale market services turn the battery from a cost into a return where the interconnection and the duty allow it. These come after the power quality duties above, never instead of them.
The battery works on the grid side and the load side
On the grid side, the battery supplies the difference when the load rises above the permitted grid draw and absorbs the excess when the load falls below it. It shapes the net power at the point of interconnection (POI), where the facility meets the grid. A training start can happen inside the facility while the grid-facing demand rises at the permitted pace. Recharge must fit those same limits.
On the load side, a dual-conversion block supplies energy through a disturbance or a separation from the grid. In the distributed parallel concept, blocks supply their feeder section after a separation, and the halls’ own UPS carries the computers through a deep sag. The connection determines how much isolation the compute load receives and which part of the distribution remains energized. The reserve held for that duty reduces the energy available for other uses. Our AI data center solution treats these two jobs as one design problem.
Four ways to connect the battery
Centralized parallel
One battery plant connects on dedicated feeders at the substation bus. It acts on the net power at the grid connection while the existing loops and load distribution stay in place. This can suit a project whose main battery duty is smoothing, ramp control or energy shifting at the meter. The study still needs to establish how the existing backup equipment protects the load during a disturbance.
Distributed parallel
A battery sub-block connects to each medium-voltage feeder through ring switchgear, also called a ring main unit (RMU). On feeder loss, the sub-block can establish voltage and frequency for its section in grid-forming mode. Existing generators and transfer switches stay in place. This arrangement follows the sections of a loop-fed site, bringing battery support closer to the loads that must remain supplied.
Dual conversion per load section
A block sits in series with each dedicated load section. A grid-facing power conversion system (PCS) supplies a direct-current (DC) bus, and a load-facing PCS feeds the racks. The battery connects to the DC bus. A static transfer switch (STS) provides a path to a backup bus. The load is supplied through its converter, with the additional conversion equipment and transfer path included in the integration scope.
Dual conversion on a shared protected bus
Grid-facing converters draw from the loop, while load-facing converters connect in parallel on a protected load bus. The battery supports the DC side of each block. There are no static transfer switches in this concept. If a block is lost, the remaining blocks take the load step. The design therefore needs sufficient remaining capacity and a defined response to that change.
Distribution, ride-through and duty set the choice
Start with the existing distribution and the decision to keep or replace the low-voltage uninterruptible power supply (UPS) and generators. Then establish the load’s disturbance tolerance and the grid’s ride-through requirements. Finally, define the battery duty, including repeated smoothing, reserve, recharge and longer energy shifts. Each consumes power capability, energy or battery life differently.
The choices can be combined. Distributed parallel support may serve the loops while dual conversion serves a critical hall. The medium-voltage UPS page explains the connection concepts and the system availability design basis.
Bring the diagram, load profile and agreement
A first conversation needs the single-line diagram, the load profile and the interconnection agreement. Include the time resolution of the load data, the backup equipment already planned and the disturbances the load must survive. These inputs establish what to simulate and which connection options deserve comparison. Request information to start that review.