Powering the AI Factory: The Role of 800 VDC Distribution and ISOP Converters in Next-Generation Data Centers
By: Maurizio Di Paolo Emilio, Contributing Editor at Panel Builder US
Artificial intelligence workloads are rapidly reshaping data center power architectures. Conventional server infrastructure was not originally designed to sustain today’s extreme compute density requirements. As a result, modern facilities increasingly resemble “AI factories,” where maximizing computational throughput per rack is a primary objective.
The move to 800 VDC distribution is one of the main things that makes this change possible. This lets rack power grow from hundreds of kilowatts to megawatt-class levels, which are needed by next-generation AI platforms. In this architecture, the 800 V rail directly powers the compute trays. This creates new problems with insulation and device ratings, but it also gives us a chance to redesign the isolation stage. Input-Series Output-Parallel (ISOP) converter topologies are becoming strong candidates for this job because they are scalable and efficient.
Why 800 VDC Distribution Is Becoming the Backbone of AI Factory Power Architectures
The transformation of conventional data centers into high-density AI computing environments is fundamentally changing the requirements for electrical power delivery at rack and facility level. Unlike traditional server platforms, where power demand increased gradually from generation to generation, modern GPU clusters interconnected through technologies, such as NVLink, enable tightly coupled compute domains whose performance scales with physical proximity and electrical power availability. As a result, rack-level power density is no longer growing incrementally, but instead increasing in large steps, sometimes doubling or quadrupling between successive architecture generations. This rapid scaling is shifting the balance between compute hardware and supporting infrastructure, making power delivery systems a dominant factor in rack design and layout optimization.
800 VDC distribution is becoming a practical and scalable solution for next-generation AI factories to support these trends. An 800 VDC architecture cuts down on the amount of current, copper needed, connector complexity, and routing volume inside the compute domain compared to older 415/480 VAC facility distribution or 48–54 V rack-level buses. It also lets you move bulk conversion stages outside of the NVLink performance radius, which frees up valuable rack space for accelerators and makes the whole system more efficient. At the same time, synchronized GPU workloads cause fast load transients that move from rack to facility level and even to the utility grid. This is why distributed energy storage is such an important part of the architecture for smoothing out changing demand profiles. These things make 800 VDC distribution not just a small improvement over current systems, but also a key part of scalable megawatt-class compute racks and future AI-native data center architectures.
Figure 1: Rapid growth in GPU rack power density across architecture generations is accelerating the transition toward 800 VDC distribution in AI factory infrastructures (Source: NVIDIA white paper)
The role of ISOP converters in AI server power systems
ISOP architectures connect the inputs of several isolated converter modules in series and the outputs in parallel. With this setup, you can increase both the input voltage and output current capacity at the same time using the same modular building blocks.
If an ISOP system includes M modules operating from a total input voltage VIN and delivering output power POUT, each module only processes VIN/M and POUT/M. As a result, semiconductor voltage stress and thermal loading per module are significantly reduced compared with single-stage implementations.
Multilevel solutions based on LLC resonant converters are particularly attractive in this context. Proper selection of the magnetizing-to-resonant inductance ratio allows the converter to operate near resonance with minimal sensitivity to component tolerances. Under these conditions, modules naturally maintain voltage balance and current sharing without complex control loops. Operating at resonance also corresponds to the point of highest efficiency for LLC converters, although this benefit comes with limited output-voltage regulation capability. Consequently, these converters are typically implemented as fixed-ratio stages.
Compared with conventional single-module LLC converters, modular ISOP implementations provide several advantages:
- reduced voltage stress on primary-side switches
- simplified transformer insulation requirements
- improved thermal distribution across modules
- lower output-current ripple through interleaving
- increased achievable power density
These features make ISOP architectures well suited for isolation stages in high-power AI racks supplied by 800 V DC distribution.
Benefits of reduced primary-side voltage stress
Lower voltage stress per module allows for the use of semiconductor devices with lower voltage ratings and better switching performance. This is especially important when looking at high-frequency resonant converters that use silicon carbide (SiC) and gallium nitride (GaN) solutions.
For instance, think about a 64:1 isolated conversion stage that works from 800 V to about 12.5 V and has a total output power of 6 kW and a switching frequency of about 1 MHz. When used as a single LLC stage, primary switches need to handle the full input voltage. But spreading the conversion across several ISOP modules also spreads the voltage stress.
As the number of modules increases, the architecture becomes compatible with lower-voltage devices offering reduced conduction losses, lower gate charge, and smaller footprint. In particular, implementations using multiple modules with 150 V GaN transistors demonstrate significant improvements in efficiency and board-area utilization compared with solutions based on higher-voltage SiC devices.
Surface-mount GaN devices are especially attractive in this scenario because they support compact layouts and are compatible with advanced cooling strategies such as top-side liquid cooling.
Output ripple reduction through interleaving
Another important advantage of modular ISOP implementations is the reduction of output-current ripple through phase interleaving.
When converter modules operate with evenly distributed switching phases, the combined output ripple decreases while ripple frequency increases. This reduces the required output capacitance for a given voltage ripple specification and improves overall transient performance.
In practical terms, a conventional single-module LLC converter delivering hundreds of amperes at the secondary side may exhibit large peak-to-peak ripple current. By contrast, a multi-module ISOP implementation distributes this current across several interleaved channels, significantly lowering ripple amplitude while increasing ripple frequency. This translates into smaller passive components and higher achievable power density.
Example: 6 kW ISOP converter from 800 V to 12.5 V
A working example of an ISOP-based isolation stage shows how well this architecture works in AI applications with a lot of data. One design uses eight identical LLC modules connected in an ISOP configuration to make a fixed-ratio 6 kW converter that changes 800 V DC to 12.5 V.
Each module uses about 750 W and works as a separate 100 V-to-12.5 V resonant stage. The primary side uses small 150 V GaN half-bridge switches with on-resistance in the milliohm range. The secondary side uses low-voltage GaN devices set up for synchronous rectification.
Planar transformers built into multilayer PCBs keep things separate, which makes the structure very low-profile. The whole converter, which includes the control and reinforced isolation circuitry, takes up less than 5000 mm² of board space and is only a few millimeters thick.
Putting all the GaN devices on the top surface of the board makes it easier to manage heat. This design works with both heatsink-based and liquid-cooled cold-plate solutions.
Testing in the lab shows that this architecture works better than others. The converter works at its best when the input and output currents are around 500 A and the voltage is 800 V. It can reach peak efficiency of over 98% and full-load efficiency of over 96%. These results show that using low-voltage GaN devices with modular ISOP topologies for high-density isolation stages in AI infrastructure works well.
Conclusion
As AI computing platforms continue scaling toward megawatt-class rack power levels, high-voltage distribution architectures such as 800 VDC are becoming increasingly important. Within these systems, isolation-stage converters must simultaneously deliver high efficiency, compact size, and excellent thermal performance.
ISOP converters based on LLC resonant modules provide an effective solution by distributing voltage stress across multiple modules, enabling the use of low-voltage GaN switches, reducing output ripple through interleaving, and simplifying transformer implementation.
Experimental demonstrations of multi-module 6 kW converters confirm efficiencies above 98%, showing that ISOP architectures represent a strong candidate for next-generation AI server power systems.
References
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.
