The Architectural Imperative of 800 VDC in Next-Generation AI Factories

By: Maurizio Di Paolo Emilio, Contributing Editor at Data Centre Digest

The digital infrastructure landscape is changing radically. For decades, data centers were designed around “compute space” where power and cooling were secondary considerations in terms of footprint. But the arrival of large-scale AI models has changed this equation, creating “AI factories” where the power infrastructure can now compete with, or even exceed, the space dedicated to compute.

The impetus for this change is the exponential increase in the power density of GPUs. The traditional CPU and GPU improvements have seen a 20% increase in thermal design point (TDP) per generation, but the move to synchronous GPU networking with NVLink has disrupted this trend. By networking GPUs over copper to create a single massive compute domain, performance increases can be as high as 50x, but at the expense of near exponential increases in power requirements – with rack densities now approaching the megawatt scale.

The Physical Limits of Traditional AC Distribution

To support these unprecedented densities, the industry must move beyond traditional 415V or 480V 3-phase AC power systems. These systems are reaching their practical physical limits in several key areas:

• Whip Size and Cable Management: Standard AC whips are constrained to 60A or 100A by thermal limits and connector standards like IEC 60309. A megawatt- class rack would require an impractical number of cables, consuming valuable rack space and complicating thermal management.
• Copper Efficiency: Transmission at 800 VDC allows for 157% more power to be transmitted through the same cross-sectional area of copper compared to 415 VAC.
• Real Estate Value: In an NVLink-enabled rack, the area closest to the GPUs is the most valuable “real estate” for performance. Traditional in-rack AC/DC power shelves occupy this space, forcing a choice between power capacity and compute performance.

The 800 VDC Architecture

The 800 VDC architecture addresses these challenges by taking 800 VDC directly into the rack and distributing it to the compute nodes. This shift enables several critical technical advantages:

Simplified Power Interface

By migrating AC/DC conversion upstream (either to a side- car power rack or a facility-level rectifier), the physical interface to the compute rack is streamlined. This reduces the number of cables and components, improving both deployment speed and operational agility.

High-Efficiency Conversion

The 800 VDC architecture utilizes a 64:1 LLC converter with a matrix transformer to convert 800 VDC directly to 12V. This eliminates the intermediate 50V stage found in traditional DC/DC power trees, resulting in:

• A 1% improvement in overall efficiency
• A 26% reduction in component area near the GPU, where space is most critical

Safety and Standardisation

Safety is a core design feature of the 800V system. It leverages mature technologies from the electric vehicle (EV) industry, which has already transitioned from 400V to 800V platforms. Key safety mechanisms include:

• Touch-safe connectors: Used in all user-accessible locations
• Mechanical interlocks: Ensuring that no connection or disconnection occurs under load, thereby eliminating the risk of electric arcing

Implementation Strategies: A Phased Approach

The transition to 800 VDC is envisioned as a multi-stage evolution to accommodate
both legacy facilities and new builds:

• White Space Retrofit (near-term): For existing 415 VAC facilities, “side power racks” containing local rectifiers can be used to provide 800 VDC to high-density compute racks. While not the most efficient end-to-end due to extra conversion stages, it allows for immediate scaling of compute density.
• Facility-Level 800 VDC (mid-term): As 800V equipment becomes more available, rectifiers are moved further upstream, providing centralized DC distribution to the entire data hall.
• MV to 800 VDC (future): The ultimate goal is direct conversion from medium voltage (MV) to 800 VDC using Solid State Transformers or MV rectifiers, eliminating multiple low-voltage transformer and switchgear stages.

Advanced Isolation Stage: GaN-Based ISOP Converters

The isolation stage poses one of the biggest technical challenges in the 800 VDC ecosystem – how to convert high voltage DC directly to the 12V or 48V required by compute nodes. The architecture uses an input-series output-parallel (ISOP) topology to do so. The 800V stress is distributed across multiple modular LLC resonant stages connected in series in ISOP configuration as compared to the traditional single stage converters. This modularity allows for the use of low-voltage gallium nitride (GaN) devices instead of high-voltage Silicon MOSFETs.

Low voltage GaN switches provide much lower figure of merit (FoM) and switching losses to enable higher switching frequencies. This frequency increase drastically reduces the footprint of magnetic components and transformers, allowing for a power density that fits within the constrained “NVLink radius” of the GPU tray. The experimental data has confirmed that these ISOP based isolation stages can have a peak efficiency of more than 98% with a current capability of 500A. Moreover, the parallel output connection provides natural interleaving, which minimizes the output voltage ripple and improves the transient response, which is a critical requirement for the fast load swings typical of AI training workloads.

Conclusion

The 800 VDC architecture is no longer a theoretical preference but an architectural imperative for the AI era. By reducing current, copper usage, and conversion losses, it provides the only viable path to power the megawatt-scale racks required for the next generation of AI models. By leaning on the proven safety and component ecosystem of the EV and solar industries, 800 VDC provides a scalable, efficient, and safe foundation for the AI factories of tomorrow.

References

Huntington, J.; Tu, M., “800 VDC Architecture for Next-Generation AI Infrastructure,”
NVIDIA White Paper, 2025.