AI Data Center Power Infrastructure: From Grid to GPU - Transformers, UPS, and Generators

GPU clusters have rewritten the power density assumptions that governed data center design for two decades. A single DGX B200 draws 14.3 kW at full load. One rack holding eight DGX systems plus top-of-rack networking requires 120-140 kW of usable power, compared to the 5-15 kW per rack that defined traditional enterprise data centers. At cluster scale, 1,024 H100 GPUs consume approximately 700-800 kW for compute alone, plus 200-250 kW for networking, storage overhead, and cooling, totaling roughly 1 MW per 1,024 GPUs at the facility level.

This 10-20x increase in rack power density forces infrastructure engineers to rethink every layer of the power train: utility service entrance capacity, substation and medium-voltage switchgear sizing, UPS and battery plant topology, generator configuration, and the final inches from busway to GPU voltage regulator module. Each layer has a capacity ceiling that becomes binding at AI deployment scale. A 50 MW GPU data center requires the same utility-grade infrastructure as a small aluminum smelter, but with the reliability requirements of a semiconductor fab.

AI data centers exceeding 10 MW take utility service at 12.47 kV or 13.8 kV medium voltage, stepping down to 480 V for facility use. The utility interface includes metering class revenue meters, protective relays, and main breakers rated for the full facility load. For a 50 MW cluster, the typical configuration is 3-4 incoming feeders from a 69-138 kV transmission substation, with site-owned metal-clad switchgear rated at 15 kV class, 40 kA interrupting capacity, with vacuum circuit breakers. The utility transformation step alone costs $2-4 million per MW and requires 18-36 months from application to commissioning.

Inside the facility, a ring-bus or double-ended substation configuration distributes 13.8 kV to electrical rooms located every 2.5-3 MW. Each room contains a 2.5 MVA liquid-filled transformer (13.8 kV delta to 480Y/277 V wye), main distribution switchboard with 4,000 A main bus, and distribution panelboards. At 50 MW, this means 18-20 electrical rooms, each occupying roughly 500 square feet. The secondary 480 V distribution feeds UPS modules, mechanical equipment, and lighting through separate switchboards for critical and non-critical loads per TIA-942 Tier III topology.

Online double-conversion UPS systems (VFI-SS-111 per IEC 62040-3) are the standard for AI data centers. The rectifier continuously charges the battery and powers the inverter, which regenerates clean AC output with zero transfer time on utility failure. For a 50 MW facility at N+1 UPS configuration, typical deployment is 8-10 UPS modules of 1.25-2.0 MW each in parallel. Each module's static bypass switch allows manual or automatic transfer during maintenance. The UPS must support 125 percent overload for 10 minutes and 150 percent for 30 seconds to handle inrush from GPU node PSUs during cluster power-up sequences.

Lithium iron phosphate (LFP) batteries have displaced VRLA in new GPU data center builds. LFP delivers 3,000-5,000 cycles at 80 percent depth of discharge versus 500-1,000 for VRLA, with 40-60 percent smaller footprint and 2-4 hour recharge versus 8-12 hours. The cost premium has compressed from 2x in 2020 to 1.1-1.3x in 2025. For a 50 MW facility with 30 seconds of ride-through, the battery plant needs approximately 420 kWh. At 5 minutes - common for cold-start generators - the battery plant scales to 4.2 MWh, occupying 2,500-4,000 square feet and costing $1.5-2.5 million at current LFP pricing of $350-450/kWh installed.

Tier III GPU data centers maintain N+1 backup generators sized for 100 percent of critical and mechanical load. For 50 MW, this requires 10-14 generators rated at 2.5-3.0 MW each, paralleled via synchronizing switchgear on a common 480 V or 13.8 kV bus. Each 3 MW diesel generator consumes approximately 250 gallons per hour at full load, requiring an 18,000-gallon tank for 72-hour autonomy. The tank farm totals 180,000-250,000 gallons across 6-8 double-walled steel tanks, occupying roughly 10,000 square feet in a dedicated yard with secondary containment per EPA SPCC requirements.

Battery Energy Storage Systems (BESS) are increasingly deployed alongside diesel generators. A 10 MW / 30 MWh BESS can ride through the critical 30-minute to 2-hour window covering 90 percent of grid outages while providing daily grid services (frequency regulation, peak shaving, demand response). BESS eliminates NOx emissions from monthly generator testing, reduces required generator count, and generates revenue of $50-150/MW-month in most ISO/RTO frequency regulation markets. Installed BESS cost at container scale is $250-350/kWh including power conversion systems, with 4-6 year payback in favorable markets.

At 80-120 kW per rack, traditional power distribution hits fundamental limits. A standard 60 A, 208 V PDU provides only 17 kVA. AI racks require overhead busway rated 400-630 A at 480 V three-phase (332-524 kVA), with tap-off boxes every 2-3 feet feeding 415 V or 240/120 V step-down transformers per rack. Each GPU rack receives 200-300 A at 240/415 V three-phase via two or four redundant feeds, each branch limited to 48 A continuous per NEC. The power path from busway to GPU silicon involves busway tap-off, rack PDU, branch breaker, power cord, node PSU, internal bus, and GPU VR.

Titanium-rated PSUs (96 percent efficient at 50 percent load) feeding VRs at 90-93 percent efficiency yield 12-17 percent total loss between busway and GPU. At 50 MW, that is 6-8.5 MW of waste heat requiring removal by the cooling system. Emerging approaches to reduce these losses include 400 VDC distribution with direct DC-DC conversion to GPU voltage (saving 3-5 points of conversion loss) and on-rack 48 V bus architectures with GaN-based intermediate bus converters achieving 98 percent efficiency. Both approaches are in early deployment at hyperscale AI facilities but have not yet been adopted in colocation.