Why 800VDC Data Center Power Distribution Is Necessary
Rack-level power delivery is running out of room before AI compute is. Vertiv and other operators describe AI racks already at 140 kilowatts (kW), with 240 kW close behind, and 800VDC data center power distribution is becoming the response once conversion hardware can no longer fit inside a standard enclosure.
The pressure is structural, not speculative. Vertiv’s account of the transition describes racks scaling from 72 to 576 GPUs, with 600 kW racks and megawatt-class deployments already part of near-term planning among operators.
Between 350 and 400 kW per rack, the physical case for higher-voltage DC becomes hard to avoid: connector sizes, busbars, and the copper volume needed to carry that current at lower voltages stop fitting in the space available.
ST frames the same problem in blunt electrical terms. Delivering 600 kW to 1 MW racks at the 48 V distribution voltage that has served 15 kW racks for decades would require roughly 12,500 A of current at the top end, a figure that makes cabling and heatsink sizing impractical.
The Conversion Hardware Behind 800VDC
Raising the bus to 800 V does not eliminate conversion, it relocates it. Power still has to step down to the voltages GPUs, CPUs, and memory modules consume, and multiple vendors have shipped reference boards to do that work closer to the silicon.
ST’s NVIDIA-validated boards illustrate the range: a 6 kW 850 kHz LLC converter using a 700 V GaN transistor and a 40 V low-voltage MOSFET converts 800 V to 12 V at a peak efficiency of 97.5%, while a 20 kW 650 kHz eight-level stacked LLC converter takes 800 V down to 6 V at 96.5% for lower-voltage, higher-current rails.
A separate 12 kW board that NVIDIA moved into production testing sustains continuous delivery at over 98 % efficiency with a power density exceeding 2,600 W/in3 at 50 V output, figures ST presents as validation that dense point-of-load conversion is achievable at scale.
On the battery side, Infineon’s 24kW SiC-based backup unit reference design operates directly from a battery stack onto an 800V DC bus using 650V and 1200V silicon carbide devices, reaching a power density of 450W/in3 with efficiency exceeding 99% in the same footprint as existing low-voltage backup units.
That module packs a 24kW main power stage with a 2.4kW auxiliary supply into a 112mm x 88mm x 118mm enclosure, and its core switch, the CoolSiC MOSFET IMT65R033M2H, carries a 175°C junction temperature rating meant to survive thermal cycling in dense rack environments.
| Reference design | Voltage conversion | Peak efficiency | Power density |
|---|---|---|---|
| ST 6 kW LLC board | 800 V to 12 V | 97.5% | 2,500 W/in3 |
| ST 12 kW board | 800 V to 50 V | over 98 % | exceeding 2,600 W/in3 |
| Infineon 24kW BBU | 800V bus, battery-side | exceeding 99% | 450W/in3 |
Grid-Side Load Swings and Fault Ride-Through
The case for 800VDC is not only about density inside the rack. AI training loads can swing from almost nothing to hundreds of megawatts and back several times per second, a variation that stresses the equipment between the data center and the grid.
ERCOT’s Large Load Working Group is evaluating how to accommodate several 800 VDC facility designs on the Texas grid, and infrastructure provider Dimaag presented in April an approach meant to isolate that load volatility from upstream generation and voltage support.
Vertiv frames the same problem as a design discipline rather than a single fix. Racks are growing from 10 to hundreds of kilowatts and data centers into the hundreds of megawatts, and the resulting bursts can strain generators, the grid, and utility power quality upstream.
The proposed answer is fault ride-through: keeping the load flat and continuing operation through a grid disturbance instead of disconnecting, using batteries and local energy sources to decide when to draw from which source. Whether this holds under real utility fault conditions at scale is not established in the evidence.
What Standards and Grounding Still Need to Settle
Component-level efficiency numbers do not by themselves make an 800VDC deployment safe or serviceable. ABB’s OCP-aligned guidance focuses specifically on sidecar grounding and protection considerations for disaggregated 800V power racks, an area distinct from conversion efficiency.
Vertiv’s rollout plan treats the sidecar as an intentionally contained first step: an 800 VDC power center placed alongside the IT rack rather than built into it, which lets an operator keep most of its upstream AC infrastructure while introducing 800 VDC close to the compute.
Vertiv’s Chief Product and Technology Officer Scott Armul described the trajectory at the company’s May 2026 Investor Conference this way: “This is a story of DC power in the rack moving to DC power in the pod.” He framed the sidecar phase as transitional, with pod- and hall-level deployments beginning in 2027.
The Eaton-led session covering the same shift lists medium-voltage solid-state transformers, DC UPS, and sidecar designs as parallel building blocks still under evaluation, which signals that no single reference architecture has been settled across vendors.
Metrics to Verify Before Committing
None of the efficiency and density figures above substitute for validating grounding, protection, and fault-ride-through behavior at the facility level, since separate sources treat those as unresolved layers of the same transition rather than solved problems.
Before committing to an 800VDC data center power distribution retrofit, confirm the rack density threshold where your own busbar and cooling design breaks down, the grounding and protection scheme your equipment vendor validates for a sidecar topology, and whether your utility interconnect can tolerate the load-swing profile the AI cluster will actually produce.