The New Power Supply Battle for AI Data Centers

I. Background 

In recent years, generative AI technologies have rapidly emerged, including applications such as ChatGPT, AI image generation, large language models (LLMs), and AI agents. These applications heavily rely on the massive parallel computing capabilities of GPUs. As model parameter sizes continue to expand, the computational density required for AI training and inference has also increased dramatically, causing the power consumption of AI servers to rise rapidly.

Traditional data centers primarily adopted 48V/54V power architectures, with single-rack power consumption typically ranging from 10kW to 30kW. However, in the AI era, high-density GPU racks have already reached 120kW, 250kW, and are even progressing toward 1MW and beyond. Conventional power delivery and cooling architectures are gradually reaching their limits.

To address ultra-high power and high-density computing demands, next-generation AI data centers are increasingly adopting the following key technologies:

· High Voltage Direct Current (HVDC)

· Liquid cooling technologies

· Centralized Power Shelf architectures

· SiC/GaN power semiconductors

· Solid State Transformers (SSTs)

· Vertical Power delivery architectures

The core objectives are to:

· Improve power delivery efficiency

· Reduce energy loss

· Minimize heat generation

· Support ultra-high-density AI computing architectures

· Build next-generation AI energy infrastructure

II. The Importance of the Power Shelf 

In next-generation AI server architectures, the Power Shelf has become the core of rack-level power architecture.

A Power Shelf can be viewed as the “centralized power delivery and power management platform” of an AI rack. Its primary functions include:

· Converting external power into the voltages required by servers

· Uniformly distributing power to GPUs, CPUs, and networking equipment

· Providing high-efficiency power conversion

· Stabilizing high-density GPU workloads

· Reducing power loss and heat generation

Traditional Power Supply Units (PSUs) were mostly designed in a distributed architecture. However, as AI GPU rack power consumption has rapidly increased, conventional PSU architectures have begun facing several challenges:

· Insufficient space

· Excessive cabling

· Extremely high current levels

· Cooling difficulties

· Increased maintenance complexity

As a result, new-generation AI racks are gradually adopting centralized Power Shelf architectures, where high-power AC/DC or HVDC conversion is centralized at the rack level and then distributed uniformly to GPU and CPU modules.

The Power Shelf is no longer merely an extension of traditional PSUs; it has become a critical infrastructure component that directly impacts AI system performance, power stability, energy efficiency, and thermal design.

III. Limitations of Traditional 48V Architectures 

Traditional 48V/54V power architectures were originally suitable for standard cloud servers, but their limitations are becoming increasingly apparent in high-density AI computing environments.

According to the power equation:

P=V⋅I

When power remains constant, lower voltage requires higher current to deliver the same amount of power.

This leads to:

· Increased cable heating

· Higher copper loss

· Reduced power efficiency

· Significantly increased cooling pressure

· Larger and heavier wiring infrastructure

When the power consumption of a single AI rack rises to several hundred kilowatts, traditional 48V architectures can no longer efficiently support high-density GPU computing requirements.

Current traditional power architectures generally achieve overall efficiencies of only around 85%–89%, making energy loss and cooling costs major bottlenecks for AI data centers.

IV. The Rise of 800V HVDC 

To solve the power delivery challenges brought by high-power AI computing, data centers are gradually adopting 400V–800V HVDC (High Voltage Direct Current) architectures.

Its core principle is:

“Increase voltage, reduce current.”

By increasing the supply voltage, current and transmission losses can be effectively reduced, thereby improving overall power efficiency.

HVDC architectures offer the following advantages:

· Reduced power transmission losses

· Lower heat generation

· Higher power efficiency

· Reduced cable usage

· Improved space utilization

· Better suitability for high-density AI racks

Compared with traditional power architectures, HVDC systems can achieve efficiencies of approximately 92%–95%, offering significant advantages in large-scale AI data centers.

V. Rapid Growth of AI Rack Power Consumption 

The power consumption of AI GPU platforms is increasing at an astonishing pace.

Taking NVIDIA’s next-generation AI platforms as an example, the power consumption of a single AI rack has increased dramatically from tens of kilowatts to several hundred kilowatts.

Some next-generation designs are even advancing toward the 1MW–2MW range, signaling that AI data centers have officially entered the “ultra-high-power era.”

This means:

In the future, the power consumption of a single AI rack may equal that of an entire office floor or even a small factory.

As a result, the importance of power delivery, cooling, and energy management has risen significantly.

VI. Comparison of High-Density AI Rack Power and Cooling Trends 

The trends above clearly show that:

· AI rack power consumption is rapidly surpassing traditional data center design limits

· Liquid cooling technologies will gradually replace air cooling

· High-voltage HVDC and centralized Power Shelves will become mainstream

· Future systems may fully evolve toward rack-level energy platform designs

VII. Two HVDC Architecture Models 

1. Transitional HVDC Architecture

Most current AI data centers adopt a “transitional HVDC” solution with the following power flow:

Utility Power (AC)
→ UPS
→ 800V DC Bus
→ Power Shelf
→ 48V/54V DC
→ DC-DC Modules
→ GPU/CPU

Architecture Features

Utility Power Input

The power grid supplies 10kV–33kV high-voltage AC power.

UPS System

Converts AC to DC while providing battery backup functionality.

HVDC Bus

Forms a 400V–800V high-voltage DC backbone power system.

Power Shelf

Converts 800V DC into 48V/54V DC and distributes power throughout the AI rack.

Server-side DC-DC Modules

Further convert the power into the low voltages required by GPUs and CPUs.

Advantages

· Existing data center infrastructure can still be utilized

· Faster deployment

· Lower retrofit costs

Disadvantages

· Multiple power conversion stages

· Energy loss remains relatively high

· System efficiency is approximately 89%

2. Solid State Transformer (SST) Architecture

Solid State Transformers (SSTs) are regarded as one of the most important power technologies for next-generation AI data centers.

The biggest feature of SSTs is:

They can directly convert 10kV–33kV high-voltage AC into 800V DC.

Compared with traditional architectures, SSTs significantly reduce the number of power conversion stages, providing:

· Higher power efficiency

· Lower energy loss

· Smaller equipment size

· Better power quality

· Higher integration levels

Currently, SST technology is still in the validation and pilot stage, but companies including:

· Hitachi

· ABB

· Schneider Electric

· Ampersand

· DG Matrix

have already begun investing in SST technologies for AI data centers.

In the future, large-scale AI data centers are expected to gradually adopt SSTs as core power architectures.

VIII. What Is a Solid State Transformer (SST)? 

A Solid State Transformer (SST) is a new type of power delivery system that replaces traditional iron-core transformers with power semiconductors and high-frequency power electronics technologies.

Its core technologies mainly come from:

· SiC (Silicon Carbide)

· GaN (Gallium Nitride)

and other high-efficiency power devices.

SSTs can perform highly efficient AC/DC conversion at high frequencies.

Key Features of SSTs

High-Efficiency Conversion

Efficiency can reach 92%–95%, with some designs approaching 98%.

Miniaturization

High-frequency transformers significantly reduce system size.

Bidirectional AC/DC Conversion

Supports future DC-native data center architectures.

Intelligent Control

Provides digital control, remote monitoring, and harmonic suppression capabilities.

Modular Design

Can be flexibly expanded and maintained according to load requirements.

IX. Comparison Between SST and HVDC 

Advantages of HVDC

· Mature architecture

· High cost-effectiveness

· Existing data centers can be upgraded gradually

Advantages of SST

· Higher efficiency

· More compact systems

· Better suited for future ultra-high-density AI racks

X. Future Development Directions 

The Era of 1MW Racks

High-voltage HVDC, high-density power delivery, and three-phase power systems will become standard.

Widespread Adoption of SiC/GaN

These technologies offer:

· High voltage tolerance

· High efficiency

· Low power loss

· Compact size

Universal Adoption of Liquid Cooling

Liquid cooling and two-phase immersion cooling will gradually replace traditional air cooling.

Vertical Power Delivery

DC-DC modules will move closer to GPU chips in order to shorten power paths and reduce losses.

Rack-Level Energy Platforms

Power Shelves will gradually evolve into integrated platforms combining:

· Power conversion

· Energy management

· Thermal management

· Real-time monitoring

forming the core energy platform of AI infrastructure.

XI. Challenges Facing AI Data Centers 

Increasing Pressure on Power Grids

AI workloads fluctuate dramatically. Future data centers will require:

· Large-scale energy storage systems

· Smart grids

· Dynamic energy management

to maintain power stability.

Insufficient Rack Space

GPUs, switches, and liquid cooling equipment consume large amounts of space, making Power Shelf integration increasingly difficult.

Rapidly Rising Energy Consumption

By 2030, global data center electricity consumption is projected to reach:

· 700TWh–1,000TWh

Therefore, renewable energy, energy-saving technologies, and highly efficient power architectures will become essential.

XII. Conclusion 

The development direction of AI servers and data centers is now very clear:

Higher voltage, greater power, higher efficiency, and stronger cooling capabilities.

Key future technologies will include:

· 800V HVDC

· 1MW high-density racks

· SiC/GaN power devices

· Liquid cooling technologies

· Solid State Transformers (SSTs)

· Centralized Power Shelf architectures

Among these, the Power Shelf will gradually evolve from a traditional power module into a core energy platform within AI infrastructure, becoming one of the key technologies influencing AI computing performance, power stability, and energy efficiency.