From 54V to 800V: Powering the Next Era of AI Data Centers

1. AI Compute Growth Is Reshaping Data Center Power Architecture

Traditional data centers were designed primarily as digital storage and application processing facilities. For years, rack power densities of 5–10kW were sufficient, and 48V/54V low-voltage DC power architectures provided a reliable and efficient solution.

However, the rapid growth of artificial intelligence is fundamentally changing data center requirements.

Modern AI data centers are evolving from traditional “data warehouses” into high-density compute factories. Large-scale AI training and inference workloads require unprecedented power delivery capabilities, pushing rack power consumption from tens of kilowatts toward hundreds of kilowatts and eventually megawatt-class systems.

NVIDIA’s accelerated computing roadmap provides a clear indication of this trend. Blackwell-based systems are already approaching 120kW per rack, while future platforms such as Rubin Ultra are expected to reach significantly higher power levels, with next-generation AI architectures targeting megawatt-scale rack power.

GPU power consumption is also increasing rapidly, moving from approximately 1.4kW per accelerator toward 3.6kW and beyond.

A single AI rack today can consume as much power as an entire residential community.

When power demand reaches hundreds of kilowatts or even megawatts, traditional low-voltage power distribution architectures begin to encounter fundamental physical limitations.

The challenge comes down to basic electrical principles:

P = V × I

For the same power output, lower voltage requires significantly higher current.

At the same time, transmission losses follow:

P_loss = I²R

meaning power loss increases exponentially as current rises.

For example, delivering 1MW of power through a 54V DC architecture would require nearly 18,500A of current. This level of current is beyond the practical limits of conventional cables, connectors, and power distribution systems.

Simply increasing conductor size is no longer a viable solution.

2. Three Major Engineering Challenges Are Emerging

2.1 Cable and Conductor Size Become Impractical

Supporting thousands of amps requires extremely large copper conductors.

At megawatt rack levels, traditional low-voltage systems may require copper conductors with cross-sectional areas exceeding several thousand square millimeters. Cable size, weight, cost, and installation complexity quickly become unacceptable.

For large-scale AI data centers, excessive copper usage also creates challenges in infrastructure cost, rack weight, and building load capacity.

2.2 Thermal Losses Become Increasingly Difficult to Control

Because current increases dramatically at lower voltages, resistive losses rise sharply.

Compared with an 800V architecture, a 54V system requires much higher current, resulting in significantly greater I²R losses.

The consequences include:

• higher conductor temperatures;

• increased cooling requirements;

• greater connector stress;

• reduced overall data center energy efficiency.

As AI workloads continue to scale, power losses that were once acceptable become major operational costs.

2.3 Valuable Rack Space Is Consumed by Power Equipment

Low-voltage high-current systems require larger power conversion stages and more parallel power modules.

In extreme cases, power infrastructure can occupy a significant portion of rack space that could otherwise be used for GPUs and computing hardware.

For AI infrastructure operators, every rack unit matters.

The objective is no longer only delivering power — it is delivering maximum compute density within limited physical space.

3. 800V HVDC: A Fundamental Solution for Megawatt-Scale AI Infrastructure

The industry is increasingly recognizing 800V high-voltage DC (HVDC) as a key architecture for future AI data centers.

By increasing voltage and reducing current, 800V HVDC addresses the fundamental limitations of traditional low-voltage systems:

• significantly lower transmission losses;

• smaller and lighter conductors;

• higher power density;

• improved rack scalability;

• better overall energy efficiency.

Unlike incremental upgrades to existing 48V/54V systems, the transition to 800V represents a fundamental redesign of the data center power delivery ecosystem.

As AI infrastructure moves toward megawatt-scale racks, high-voltage DC power distribution is becoming an essential foundation for next-generation compute facilities.