NGI Technologies Company Limited
The Power Revolution of Artificial Intelligence: In-depth Analysis of the 800V AI Server Power Architecture and Testing Challenges

Home>News>Applications

The Power Revolution of Artificial Intelligence: In-depth Analysis of the 800V AI Server Power Architecture and Testing Challenges

November 05,2025
share to:

The Troubles under the AI Boom—Power Consumption and Heat Dissipation

In 2025, generative AI applications represented by ChatGPT and Deepseek have deeply penetrated thousands of industries, driving global computing power to grow at an unprecedented rate. The power density of AI server cabinets has skyrocketed from the previous 10–20 kW to hundreds of kilowatts or even megawatts. The traditional UPS power supply architecture has low conversion efficiency, which restricts the development of AI computing power. Therefore, how to provide these “computing beasts” with more efficient, stable, and economical energy has become the core pain point of the entire industry.


lQDPKdMZ2h0JPGPNAifNA-iwfWrb2TsD7IcI5gpqHhoMAA_1000_551


The Breakthrough Path:

From AC UPS to High Voltage Direct Current (HVDC)

The typical power conversion chain of the traditional UPS solution is: Utility Power → UPS (AC/DC–DC/AC) → PDU (Power Distribution Unit) → Server Internal Power Supply (AC/DC). After multiple AC-DC conversions, energy loss is significant, and the overall power supply efficiency is usually less than 90%, making it difficult to meet the pursuit of ultimate energy efficiency in the AI era.

The ±400V, 800V, and 1500V (future direction) HVDC architectures can significantly improve end-to-end energy efficiency by reducing conversion stages and have become the mainstream trend of future AI data centers. Recently, NVIDIA released the white paper “800 VDC Architecture for Next-Generation AI Infrastructure”, clearly proposing the evolution direction toward the 800V DC architecture. 


c43f5a5e-405e-48b2-bc0c-a8ae67ae280b


800V DC Architecture: The “Optimal Solution” for Next-Generation AI Infrastructure

Core Value: 800V high-voltage DC, with efficiency improvement (3–5 percentage points higher than UPS), more than 45% reduction in copper loss, and 30% optimization in space, will become the mainstream power solution for future AI data centers. Among them, NVIDIA proposes the adoption of Solid-State Transformer (SST) technology to directly convert medium-voltage AC (such as 10 kVAC) into 800V DC in one high-efficiency step, feeding power directly into IT racks. This overturns the traditional multi-stage conversion model and becomes the cornerstone technology for realizing the “AI Compute Factory.”

The technical iteration path from the transformation of existing facilities to the ultimate SST solution is as follows:


lQDPJxJie6NjvGPNAZrNBAWw3HBboc9ojRQI5gpqICmkAA_1029_410


NGI

AI Full-Scenario Testing Solution

8116cc0a716bd75241e1b3f843509084


NGI has been deeply engaged in the data center energy industry for many years and has launched a full-scenario testing solution for AI server power supplies based on industry testing demands. It can meet testing applications in traditional UPS architectures, 400V/800V/1500V HVDC architectures, including PSU, Power Shelf, Rack Power, Board-Level Power, HVDC, BBU, and SST. The solution provides the following customer values:


High Voltage and High Power with Energy Regeneration

Taking the N35500-1500V power supply as an example, a 3U unit can integrate 42kW, providing a wide voltage range up to 1500V and megawatt-level power expansion capability. It seamlessly meets the testing challenges of 800V voltage and future higher power and voltage architectures. Adopting a source-load integrated design, the test energy can be fed back to the grid, with a feedback efficiency of up to 93%, effectively saving users’ testing electricity and cooling equipment costs.

In addition, NGI launched the world’s first 2400V high-voltage electronic load, which can match future 1500V high-speed testing requirements.


High-Speed Loading and Accurate Simulation of Server Dynamic Characteristics

The N69200 series supports a loading speed of 30A/μs, while the N67000 series supports 60A/μs, and multiple units can be paralleled to achieve up to hundreds of A/μs, meeting the high-speed dynamic loading test requirements of AI server power modules and Vcore power supplies.


lQDPJw3NPBuhHGPNASfNA-iwk3CUYiWR7-cI5gpqIU4VAA_1000_295


High-Precision Testing and MW-Level Parallel Operation

Taking the N69200-1200V model as an example, it supports three voltage ranges: 1200V/600V/120V, with measurement accuracy up to 0.015% + 0.015% F.S. It ensures that during 50V loading tests, the measurement error is only 0.025V, guaranteeing testing accuracy across a wide range from 48V to 800V.

At the same time, the N69200 series can be paralleled up to 1.44MW, supporting megawatt-level system verification of server power supplies. It can accurately simulate complex working conditions of AI cabinets under different computational loads, verifying the stability and dynamic response capability of front-end power systems.

Hot categories