AI Data Center Power Systems: How 800VDC, SiC and GaN Are Changing Power Architecture

Legacy AC power systems can become increasingly inefficient and difficult to scale as AI rack power rises because power is lost across multiple conversion stages, generating heat while low-voltage, high-current distribution requires substantial conductor cross-sections.

Traditional data center power delivery can rely on substation transformers to convert high-voltage transmission-line AC to a local three-phase AC distribution voltage such as 480VAC in North America or 400/415VAC in other regions. Battery backup units (BBUs) and energy storage systems (ESSs) can introduce additional conversion stages between the facility supply and the rack.

Each conversion stage adds losses and generates heat, while the low-voltage, high-current distribution required further downstream increases conductor requirements. Intermediate and point-of-load supplies then convert the available bus voltage to the lower voltages required by processors, GPUs and other electronic loads.

The industry is developing power architectures for individual AI racks approaching the megawatt range. As an example of the scaling challenge, NVIDIA estimates that delivering 1MW at 54VDC could require up to 200kg (440 pounds) of copper for rack busbars alone (source: https://developer.nvidia.com/blog/nvidia-800-v-hvdc-architecture-will-power-the-next-Generation-of-ai-factoriesNVIDIA 800 VDC Architecture Will Power the Next Generation of AI Factories | NVIDIA Technical Blog) At data center scale, the resulting conductor requirements become a significant design consideration alongside electrical losses, thermal management and available space.

The costs and physical requirements of copper distribution, together with the heat generated by low-voltage, high-current equipment, are directing development efforts toward two key innovation avenues: high-voltage DC distribution to reduce current and conductor requirements, and wide-bandgap semiconductors to improve power conversion performance.

One emerging architecture for high-density AI infrastructure uses an 800VDC distribution bus to reduce current at very high rack power levels. The higher distribution voltage can reduce conductor requirements and I²R losses compared with distributing the same power at lower DC voltages. Battery backup units and energy storage systems are inherently DC-based, creating additional opportunities to integrate energy storage into DC-oriented power architectures with fewer unnecessary conversion steps.

The scale of the difference follows directly from I = P/V: distributing 1MW at 54VDC corresponds to approximately 18.5kA, while the same power at 800VDC corresponds to 1.25kA, before redundancy and derating. This reduction in current is one of the fundamental reasons why higher-voltage distribution becomes attractive as rack power approaches the megawatt range.

The transition does not eliminate downstream voltage conversion. An 800VDC rack bus can feed an intermediate conversion stage to approximately 48VDC, 50VDC or 54VDC, followed by further conversion to 12VDC and ultimately the low voltages required by processors and GPUs. These conversion layers should be distinguished from conventional 1U CRPS/M-CRPS hot-swap server power supplies.

Electrons in materials conduct electricity only when they are in an atom's conduction band. A conductor, such as copper, has overlap between the valence and conduction bands, so electrons move freely. An insulator has a very large gap between the bands, so electrons cannot cross under normal conditions. Semiconductors sit in between: their bandgap is narrow enough for electrons to cross with the addition of energy to the system.

Silicon has a bandgap of about 1.1eV, while SiC (about 3.3eV) and GaN (about 3.4eV) are called wide-bandgap materials because their gaps are roughly three times larger. This wider gap is what gives them a higher breakdown field strength and allows them to operate at higher voltages, temperatures and switching speeds.

The wider bandgap and associated material properties of SiC and GaN enable operation under electrical conditions that can be advantageous in high-frequency and high-voltage power conversion.

In suitable converter topologies, these properties can reduce switching losses and allow operation at higher switching frequencies. Higher frequency can reduce the size of magnetic and passive components, while lower semiconductor losses can reduce cooling requirements. The actual efficiency and power-density improvement depends on the converter topology, voltage levels, switching frequency, thermal design and operating point.

AI data centers are no longer just about raw computing power. Resources such as copper, water for cooling and utility grid capacity are reaching environmental and physical limits. Reducing the thermal, electrical and materials footprint is therefore an important consideration for continued infrastructure scaling. Today's facilities must prepare for future power architectural changes, even if not all downstream systems are ready today.

For future high-density AI installations, the industry is evaluating architectures that integrate energy storage more closely with high-voltage DC distribution. One emerging approach places localized BBU or ESS capacity alongside an 800VDC rack bus, reducing current at high power levels and potentially avoiding some intermediate conversion stages. Such architectures are currently evolving alongside established AC and 48/54VDC power systems rather than replacing them universally.

Higher DC distribution voltage also introduces additional design and safety requirements. An 800VDC system must account for appropriate isolation coordination, creepage and clearance distances, DC-rated connectors and protection, grounding strategy and service procedures. These requirements depend on the actual system topology and applicable standards and cannot be derived from the bus voltage alone.

One transitional architecture uses an 800VDC rack-side distribution bus while retaining the facility's existing AC infrastructure, providing a migration path toward the higher rack power levels expected for future AI systems. Transitioning to this model can require a rack-side power system to convert the facility's existing AC distribution to the high-voltage DC bus. Depending on the system architecture, this equipment can also integrate localized battery backup or energy storage.