Maurizio Di Paolo Emilio, Marcom Director, EPC
Rapid growth of artificial intelligence is forcing a rethink of data-centre power architectures. Legacy server designs were never intended to support the extreme power densities of modern AI workloads. These facilities are increasingly becoming what we might call “AI factories” where the primary goal is to optimize compute capability in a constrained physical footprint.
A key enabler for this transition is the use of 800 VDC distribution. The high voltage approach enables power scaling at the rack level from conventional levels to the megawatt range, which is essential for next generation accelerator platforms. 800 V direct power delivery to compute subsystems introduces new challenges in insulation coordination and device stress, but it also presents an opportunity to rethink how isolation is implemented. Here, Input-Series Output-Parallel (ISOP) converter topologies are becoming popular because of their modularity, scalability and efficiency.
From Incremental Growth to Step Changes in Power Density
AI hardware evolution has shattered the traditional power scaling path of data centers. Earlier computing platforms saw incremental increases in power consumption, but today’s AI clusters - especially those built on tightly interconnected GPU fabrics - show much steeper growth curves. Availability of electrical power and physical proximity between compute elements are both critical system constraints that have a strong impact on performance.
Graphic Processing Units (GPUs) are made of billions of microscopic transistors that are compacted on a silicon substrate with features as small as 20 Å (That’s about the width of the DNA molecule’s double helix). The power needed to turn on these billions of processing agents is growing as fast as their computing power.
Thus, the rack-level power density is no longer increasing linearly. Instead, each new generation of AI hardware can cause dramatic jumps in the amount of power required, often multiplying previous levels. The resulting shift emphasizes power delivery infrastructure more than ever and is now a critical consideration in overall system performance and layout.
The application of 800-V high-voltage DC distribution is an attractive solution to these challenges. Higher distribution voltage directly translates into lower conduction losses, less copper, less complicated interconnect design and therefore lower power levels for a given power. This architecture dramatically enhances the power delivery efficiency and space utilization compared with the traditional AC-based systems or low-voltage DC buses.
The other advantage is the flexibility of architecture. Power conversion stages with high power density can be located outside of tightly packed compute zones, freeing up valuable space for processing hardware. But this transition also brings new considerations. The AI workloads cause rapid and large changes in power demands, which propagate across the entire system. To alleviate these impacts , distributed energy storage systems are increasingly used to smooth transient behavior and maintain stability .
In this sense, 800 VDC is not just an optimization but an enabling technology for the next generation of high density computing environments.
How to deliver the power
Traditional power architectures are facing challenges in providing more than 200 kW to a single rack. Today power is generally brought into the rack at 240 VAC , converted to about 400 VDC and then stepped back down to 48 VDC inside the rack in a power tray . Then each server board gets fed 48 V . This is further converted to sub-1 V levels required by the processors .
As the compute power increases, this approach reaches fundamental limits. Modern AI server boards can require several kilowatts, which means feeding tens or hundreds of amps at low voltage. At higher power levels , current quickly becomes unwieldy for standard connectors and cabling . This has both thermal and mechanical limits .
With that in mind, the industry is moving to deliver power to the server board at much higher voltages. The 800 VDC or ±400 VDC architectures result in lower current levels, making the power distribution more practical and efficient.
In these early systems, the high voltage power is usually generated outside of the system, often in a dedicated “sidecar” rack, and then distributed to the compute racks. Depending on the design, the voltage is either converted locally on the server board to intermediate levels like 6V or 12V, or converted to 48V in the rack before distribution.
Server boards are extremely valuable and space-constrained real estate, so any onboard power conversion must be highly compact, low-profile and exceptionally efficient, and support advanced thermal management strategies.
ISOP Architectures for Scalable Isolation
In high voltage DC systems, the isolation stage needs high efficiency, compact size and robust thermal performance. ISOP converter topologies (Inputs in Series and Outputs in Parallel) distribute electrical stress elegantly among multiple modules.
Input to converters are series connected in an ISOP arrangement to enable system to be used from high input voltage while outputs are parallel connected to give high current. The electrical and thermal load on individual components is reduced because each module handles only a part of the total voltage and power.
Resonant LLC converters are very well suited to this modular concept. These converters are designed to operate close to their resonant point, leading to high efficiency and natural voltage and current sharing between modules. This intrinsic balancing behavior reduces the need for complex control schemes. However, LLC converters usually have a fixed conversion ratio and therefore are best suited for stages where tight voltage regulation is not required.
The ISOP implementations have several advantages over monolithic designs, such as lower voltage stress on the switching devices, simplified insulation requirements, improved heat distribution and increased power density. These features are well suited for the rigorous requirements of AI server power solutions.
Figure 1: The basic ISOP concept is to use low voltage devices in a stacked series of LLC converters with their outputs connected in parallel.
Advantages of Lower Voltage Operation
A major benefit of ISOP configurations is the possibility to use lower voltage semiconductor devices. The total input voltage can be shared between multiple modules, so that each stage is exposed to a smaller range of voltage. This allows the utilization of faster and more efficient components.
This is especially useful when employing wide bandgap technologies such as gallium nitride (GaN). These lower voltage GaN devices have lower conduction losses, lower gate charge and better switching characteristics than higher voltage devices. These properties are needed for high-frequency operation and compact system design.
For a conventional single-stage converter stepping down from 800 V, the devices must be rated for the full input voltage. In contrast, the ISOP-based approach enables operation with devices rated for much lower voltages, e.g. 150 V GaN transistors, with the benefits of higher efficiency and smaller footprint. These devices are also well suited for advanced thermal management techniques such as direct liquid cooling.
Interleaving and Ripple Mitigation
The modularity of ISOP systems also allows good ripple reduction using phase interleaving. This is done by switching numerous converter modules with equally spaced switching phases so that the total output current is smoothed. This method reduces the current ripple amplitude while increasing its frequency, therefore reducing the output capacitance needs. The resulting design is more compact, with better transient response and higher overall efficiency. This advantage is especially crucial for high current applications characteristic of AI workloads. ISOP topologies divide the load on several channels instead of forcing huge ripple currents through one conversion stage, improving electrical performance and reliability.
Demonstration: A 6 kW Modular Converter
An example implementation shows the practical advantages of ISOP based designs. In this example, the isolation stage of 6 kW is achieved by eight identical LLC modules in an ISOP arrangement. Each module converts about 750 W, converting an intermediate voltage to a low-voltage output suitable for compute hardware.
The design features compact GaN devices on primary and secondary sides with synchronous rectification to maximize efficiency. Planar transformers embedded in multilayer PCBs provide isolation as well as a very low profile.
The surface-mount configuration of the complete system achieves a minimal footprint and enables advanced cooling strategies. The experimental validation demonstrates peak efficiency higher than 98% and over 97% at full load conditions, confirming the suitability of this architecture for high-density AI environments.
Figure 2: The EPC91123 is an 800 V – 12.5 V, 6 kW, eight-stage ISOP DCDC converter that is only 8 mm thick and is designed to be mounted directly on the server board. This converter has a peak efficiency of 98.3% and a full-load efficiency of 97%.
Reference
GaN Power Devices for Efficient Power Conversion, Fourth Edition – by Alex Lidow, Michael de Rooij, John Glaser, Alejandro Pozo Arribas, Shengke Zhang, Marco Palma, David Reusch, Johan Strydom.
Pozo, A.; de Rooij, M., “Using Low-Voltage GaN in ISOP Converters for AI Servers with 800 V Architecture,” Bodo Power Systems, March 2026.
Huntington, J.; Tu, M., “800 VDC Architecture for Next-Generation AI Infrastructure,” NVIDIA White Paper, 2025.
EPC91123 – 6 kW, 800 V to 12.5 V Isolated Converter Evaluation Board