AI data centers. Critical power at medium voltage.

A battery on the medium-voltage bus.

Battalion places a battery energy storage system (BESS) on the medium-voltage bus, with the energy management system (EMS) coordinating the load and the grid. The site’s distribution and ride-through duty select the connection.

Battalion builds both parallel and dual-conversion medium-voltage uninterruptible power supply (UPS) systems on one platform.

One training job moves the whole campus.

A training job runs in steps. The GPUs compute, pause together while they exchange results, then compute again. Every GPU in the job moves at the same moment, so when one job spans the campus, the whole campus moves with it. At 1 GW the load falls from about 970 MW to about 425 MW and comes back every 20 seconds or so, with rises of up to 147 MW in 100 ms and 525 MW in one second.

Timing differences between racks do not average this away, because the job re-synchronizes at every exchange. With 200 ms of spread the campus still swings from about 440 to 930 MW. Independent jobs and inference smooth the load: eight independent jobs move about 260 MW in a second, and three training halls among seven inference halls about 150 MW.

With dual-conversion UPS blocks the grid sees about 1 percent of the swing. The load-side converters serve every pulse from the battery and the grid-side converters follow a slow average, so the grid draw stays between 805 and 815 MW while the campus swings between 424 and 971 MW. On ERCOT’s proposed large-load limit of 10 MW peak to peak in any 5 seconds, the grid draw reads 5.7 MW filtered and 6.6 MW raw over two hours. Parallel blocks give the same reading in the same 1 GW case: beside the load, with a 3 ms total response from measurement through command to the converter, they inject and absorb the opposite of each pulse and read 5.7 MW filtered and 6.6 MW raw.

The backup equipment and grid requirements set the starting point.

First, establish whether the project will keep or replace its low-voltage UPS and generators. Dual conversion feeds each load section through its own load-side converter with 0 ms transfer, so it can take the place of the low-voltage UPS. It can take the place of the 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. Parallel blocks sit beside the load, on each medium-voltage feeder or at the substation bus, and the halls keep their UPS and generators. The existing distribution, load tolerance and required continuity determine which equipment remains and where the battery connects.

Second, identify the applicable grid requirements. 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. ERCOT has also proposed a large-load power-variation limit of 10 MW in any 5 seconds, which it has not yet adopted. The Battalion medium-voltage UPS is fully compatible with grid-fault ride-through interconnection rules such as NOGRR 282. For another independent system operator (ISO), bring its ride-through requirements and the interconnection agreement into the study.

A site that keeps its UPS and generators but must meet a ride-through rule, or is planning for the power-variation limit, can use parallel blocks. They smooth the load pulses at the grid connection, with the same reading as dual conversion in the 1 GW case above. In a deep grid sag the halls’ UPS carries the computers. 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. They also protect generators that share their medium-voltage bus from the pulses. The medium-voltage UPS page compares the two configurations line by line.

Four ways to connect the battery.

Ride-through, power smoothing and interconnection, each with the traces behind it.

Three pages take one duty each, with the simulations and scenario models behind the numbers on this page.

• NOGRR 282 ride-through: a grid fault with dual-conversion blocks and with parallel blocks. Fully compatible with ERCOT’s large-load ride-through rule.
• Data center power smoothing: the AI load pulse over an hour and over twenty seconds, played live, and ERCOT’s proposed 10 MW in 5 seconds limit simulated at 120 MW and 1 GW.
• Data center interconnection: ramp rate control at a training restart, and one EMS for a limited grid connection with on-site generation.

The reasoning behind all three is in why your AI data center needs a battery.

Training continuity and a controlled grid interface.

The battery absorbs load swings, shapes training starts and stops, and supports the section through disturbances. Operators gain a defined reserve and a grid-facing load profile that follows the interconnection limits. Without these functions, load volatility can restrict training schedules and grid events can force lost work to be repeated.

The availability model targets five nines (99.999 %) at the system level with N+1 block failover, one spare block beyond the load requirement. It is evaluated for each site.

Sizing accounts for power conversion, battery duration, redundant blocks and the energy held for ride-through. Market dispatch uses capacity above that reserve. Recharge follows the same ramp limits as the load.

The data center viewer connects the layout to the study.

Energence is Battalion’s energy intelligence platform. The data center viewer brings the site layout, energy data and analysis together so the distribution and connection options can be considered in context. The siting study screens the connection capacity available to the load.

Read why an AI data center needs a battery or request information for your project.