Medium-voltage UPS for AI data centers.

Interconnection, load pulses, generators and backup.

Interconnection. Grid operators now require large computational loads to stay connected through voltage and frequency disturbances. 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. In dual conversion the battery carries the load through the disturbance. In parallel the halls’ own UPS carries the computers through a deep sag. In the scenario model, the blocks return the site’s draw to its pre-fault level within about one cycle of the voltage returning, up to the blocks’ rated power, well inside the 0.5 s NOGRR 282 allows. The interconnection study confirms the recovery for each site. The Battalion medium-voltage UPS is fully compatible with grid-fault ride-through interconnection rules such as ERCOT NOGRR 282, and the site study confirms the settings for each interconnection.

AI load pulses. 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 UPS block meets the swing at the feeder, so the grid sees the slow average. The same control damps power oscillations between the load and the supply.

Ramp rate and generators. Reciprocating engines and gas turbines ramp over seconds to minutes, and fuel cells and small reactors ramp more slowly still. The load moves in under a second. The battery takes the fast part of the ramp. In parallel, the blocks protect prime movers that share their medium-voltage bus from the torsional stress of pulse loads: on-site prime generation, or medium-voltage standby generators. With dual conversion, generation upstream of the blocks sees the grid-side converters’ slow average.

Backup and cost. A dual-conversion block carries the load through a loss of supply with 0 ms transfer. In the distributed parallel concept, if a feeder breaker opens and stays open, grid-forming blocks form voltage and frequency for their feeder section within their rating and duration, the halls keep running, and the existing generators remain the next layer. Capacity above the reserve lowers energy cost and demand charges where the tariff rewards it. These are secondary duties and never displace the reserve.

Parallel and dual conversion on one platform.

Battalion builds both parallel and dual-conversion medium-voltage UPS on one platform. The distribution, load tolerance and battery duty determine which connection the site needs.

Blocks are offered from 2.5 MW to 5 MW and above, in steps that include 3 MW, 3.5 MW and 3.75 MW, with durations from one to four hours and custom durations on request. The design is battery-system agnostic. Reference designs exist with validated battery and converter vendors, and a site can specify its own.

Fault tolerance is layered: N+1 at the block level, converter strings that fail one string at a time, paired controllers, redundant auxiliary power and a block that keeps running if the site network is lost. A dual-conversion block protects the hall from the grid and the grid from the hall. A parallel block protects the grid from the hall, and the halls’ own UPS protects the computers from grid sags.

Battalion is manufacturing medium-voltage UPS equipment for AI data center projects. The UPS is one offering within the Battalion platform, alongside the energy management system (EMS), remote operations and project modeling.

See a block in 3D or request the spec sheet.

Dual conversion can replace the hall UPS. Parallel keeps it.

Dual conversion puts the block in series with the load. Each load section is always fed through its own load-side converter with 0 ms transfer, so the block can take the place of the halls’ 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 puts the blocks beside the load, on each medium-voltage feeder or at the substation bus. It is for a site that keeps its hall UPS and generators and must meet a ride-through rule such as ERCOT NOGRR 282 or a power-variation limit in its interconnection agreement, or that is planning for ERCOT’s proposed limit of 10 MW in any 5 seconds. The blocks smooth the load pulses at the grid connection and, sized for the hall load they back, return the site’s draw to its pre-fault level after a grid sag. They also protect generators that share their medium-voltage bus from the torsional stress of the pulses.

The deep-sag rows describe a sag to 20 % voltage, as in the scenario model. In shallower sags the site study confirms how much power the site keeps drawing and how the blocks are sized for it.

Basis: the power variation readings come from one training job across a 1 GW campus of 600 MW of IT and 400 MW of cooling and mechanical load, with 50 ms of timing spread, on the published training shape of a GPU cluster (Choukse et al., arXiv 2508.14318, 2025). Both cases run in 1 ms steps with the production control code. The filtered reading uses the 0.1 to 55 Hz band named in the proposal, with filter corners that are our assumption, and the raw reading is unfiltered. The limit is one ERCOT has proposed and not yet adopted. The sag figures are a scenario model, and the interconnection study confirms the ride-through settings for each site.

The two configurations can be combined on one site. The four connection concepts below show where each one connects, and the NOGRR 282 ride-through page walks through a grid fault in each.

Class values for the medium-voltage UPS.

Local control in every block.

Each concept uses a medium-voltage PCS station, battery containers, ring switchgear, sub-block controllers in A/B pairs, an auxiliary power skid with two sources and a central energy management system (EMS). A module fault in a string PCS removes one string. The rest of the station remains available.

Every sub-block and block runs autonomously on loss of the EMS, the core network or the wide-area network (WAN).

Distribution, ride-through and duty set the choice.

Loop-fed sections with generator-backed pod feed units point to parallel. UPS-class isolation of the racks points to dual conversion. Hours of energy shifting point to parallel. Minutes of ride-through at full load point to dual conversion.

Combinations are possible. A site can use distributed parallel on the loops and dual conversion for one critical hall.

Availability design basis.

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.

String-level fault containment, paired controllers, redundant auxiliary power and block autonomy contribute to the system design. The dual-conversion concept on a shared protected load bus is drawn with N+2, two spare blocks, because the loss of a block there becomes a load step for the blocks that remain. Redundancy at each layer and the repair path are considered together.

Ride-through, power smoothing and ramp rate control.

Reserve for continuity. Dispatch around it.

Battalion EMS coordinates grid-facing power and the energy the load needs. State of charge (SOC) is the battery’s available charge as a percentage of its usable range. The required backup reserve remains a dispatch constraint.

Energy shifting, demand response and eligible market services use capacity above the reserve. Energence models the competing duties, battery life and warranty limits. Request an engineering review to compare the connection options against your single-line diagram.