LinkT&B Editorial Department
AI server power design combines conversion efficiency with rack distribution, load transients, redundancy and cooling. Evaluate architectures such as 800 VDC against the actual upgrade schedule, distinguishing supplier roadmaps from deployed infrastructure.
Contents
- Comparable conditions for power architectures
- Background: With the expansion of AI/GPU, the way we look at electricity has changed.
- Main technological battleground: From optimizing PSU units to optimizing power distribution architectures
- The 48V/54V class is not over, but its limits are beginning to be seen and it will gradually evolve to 800VDC.
- GaN and SiC are an era where it's not a matter of "who wins" but "where to use them"
- The axis of evaluation in the future will not only be 50% load efficiency.
- Comparison of PSU/power distribution methods
- Summary: In the AI era, power supply design is not a “converter” but a “power system” design
- Please contact us
- Related Articles
- Discuss your project
With the spread of generative AI, server power supply design can no longer just be about "high efficiency." The power consumption per GPU has reached the 1kW class, with racks exceeding 100kW, and MW class power consumption in the future. What is being questioned now is not only the conversion efficiency of a single PSU, but also the limits of 48V/54V class power distribution, low load efficiency during redundancy, and the cooperative design of power supply and cooling assuming liquid cooling. In this article, we will summarize what is changing in server power supplies in the AI era, based on publicly available primary information.
Comparable conditions for power architectures
| Item | What to check | How to assess it |
|---|---|---|
| Efficiency | Boundary from input to load | PFC-stage versus complete-system efficiency |
| Transients | Load steps, dips and redundancy transfer | Voltage excursion and recovery time |
| Implementation | Wiring, insulation, connections and cooling | Maintenance and fault isolation |
Background: With the expansion of AI/GPU, the way we look at electricity has changed.
Annual energy consumption and instantaneous power or facility capacity are different measures. Use kWh for energy and kW for power, and distinguish installed capacity from actual utilisation.
Furthermore, the IEA does not say that “AI operations account for more than 40% of electricity,” but rather that accelerated servers that support AI implementation will account for “almost half” of the increase in global data center electricity demand from 2024 to 2030 (IEA, April 2025). Importantly, AI is not only increasing demand, it is also changing the very configuration and power density of servers.
The scale of the facility has also changed. The IEA describes the average data center as 5 to 10 MW, and large hyperscale locations as over 100 MW (IEA, October 2024). On the node side, the NVIDIA DGX B200 is configured with eight B200 GPUs and six 3.3kW PSUs in 5+1 redundancy, delivering a maximum system power of 14.3kW and a maximum of 1,000W per GPU in the B200 class (NVIDIA, December 2025).
Main technological battleground: From optimizing PSU units to optimizing power distribution architectures
The 48V/54V class is not over, but its limits are beginning to be seen and it will gradually evolve to 800VDC.
NVIDIA has presented an 800 VDC architecture for future high-density AI infrastructure, with broader deployment aligned to its 2027 roadmap. This does not mean existing data centers have already completed the transition. As a calculation example, supplying 200 kW of DC power requires about 3.7 kA at 54 V or 250 A at 800 V. This comparison ignores losses; conversion stages, insulation and protection must be evaluated separately. NVIDIA: 800 VDC
As an example of this trend, Schneider Electric and NVIDIA published a reference design in December 2024 for liquid-cooled, high-density AI clusters at up to 132 kW per rack (Schneider Electric, December 2024). In October 2025, Schneider Electric announced an 800 VDC sidecar aimed at next-generation GPUs and racks of up to 1.2 MW (Schneider Electric, October 2025).
GaN and SiC are an era where it's not a matter of "who wins" but "where to use them"
Even within PSU, the main players are changing. GaN can easily take advantage of its high frequency and low loss, and as of March 2024, Navitas announced that its GaN-based 3.2kW data center power platform has achieved over 100W/in³ and an efficiency of over 96.5% (Navitas, March 2024). On the other hand, SiC is advantageous in high voltage and high power stages, and Infineon classifies SiC as 400V to 3.3kV class and GaN as 40V to 700V class, indicating the premise that they can be used depending on the application (Infineon, public information as of 2026).
In actual AI server PSUs, a hybrid of GaN and SiC is most likely. Navitas announced that it will be possible to reach 97% efficiency at 137W/in³ with 4.5kW CRPS in 2024, and 98% efficiency with 8.5kW AI data center power supply with GaN+SiC configuration in November 2024. In addition, the company's IntelliWeave control claims a PFC stage peak of 99.3%, but this should be read as the peak value of the PFC stage rather than the overall PSU efficiency (Navitas, November 2024/March 2025).
The axis of evaluation in the future will not only be 50% load efficiency.
AI servers assume a redundant configuration, and each PSU does not always operate under high load. DGX B200 is also 5+1 redundant (NVIDIA, December 2025). For this reason, a PSU that does not lose efficiency even under low loads is important. 80 PLUS Ruby, announced by CLEAResult in March 2025, requires 96.5% at 50% load and 90% at 5% load for redundant data center PSUs. This is a standard for reducing unnecessary losses even with redundant operation and oversized designs (CLEAResult, March 2025).
Comparison of PSU/power distribution methods
| Technology | Efficiency based on published examples | Power density based on published examples | Introduction cost | Main application areas |
|---|---|---|---|---|
| Conventional Si-based Titanium class PSU | 96% at 50% load (80 PLUS Titanium, for redundant DCs) | Disadvantageous than WBG type | low to medium | Update of existing CPU server and existing data center |
| GaN-based CRPS | Over 96.5% | Over 100W/in³ | Medium | High frequency/small PSU with 48V/54V output |
| GaN+SiC hybrid CRPS | 97〜98% | 137W/in³ class (4.5kW example) | Medium to high | High-density server for AI/HPC, 100kW class rack |
| 800VDC power distribution + in-rack DC/DC | Expected to improve up to 5% up to the terminal compared to the current 54V system | More effective for "system optimization" than PSU alone | High (initial introduction) | Future 100kW+~MW class AI rack |
The implications at the design site are clear. In the short to medium term, we will promote higher efficiency and higher density PSUs using GaN/SiC for 54V systems. In the medium to long term, there should be a scenario for transitioning to 800VDC power distribution, including at the rack, row, and room levels. These two things need to be considered in parallel.
Summary: In the AI era, power supply design is not a “converter” but a “power system” design
Power supplies for AI servers are no longer just a matter of selecting a PSU. As indicated by the IEA, data center power demand will expand to approximately 945 TWh in 2030, and accelerated servers will be responsible for a large portion of this increase (IEA, April 2025). In the field, it is necessary to ascertain the limits of the 54V system, support the immediate high density with GaN/SiC, and design with an eye toward liquid cooling, redundant operation, and future 800VDC. Competitiveness is determined not only by component efficiency but also by the ability to jointly optimize power distribution, protection, cooling, and operational efficiency.
Please contact us
LinkT&B supports the design and development of power conversion circuits that utilize high-efficiency power supplies, DC-DC converters, and next-generation power semiconductors. For more information on improving power efficiency, increasing power density, and considering future architectures for AI servers, please contact us through service page or contact us.
Related Articles
SiC MOSFETs in 2026: Technology Trends and Selection Criteria / Motor Technology Trends: Choosing the Right Motor for Your Application
Further reading: NVIDIA: 800 VDC architecture
Discuss your project
We support circuit design, prototyping and hardware testing. Tell us about the operating conditions and the problem you need to investigate.