NVIDIA Advocates for 800 VDC Ecosystem in AI Factories

NVIDIA has proposed a significant shift in power infrastructure to support the growing demands of AI workloads. The company suggests transitioning to 800 VDC power distribution and integrating multi-timescale energy storage to address the challenges of high power density and workload volatility in AI factories.

The increasing power consumption in AI, driven by the pursuit of performance and enabled by high-bandwidth interconnects like NVIDIA NVLink, has made traditional low-voltage power distribution impractical. NVIDIA's solution aims to improve efficiency, reduce costs, and enhance the reliability of AI infrastructure.

Benefits of 800 VDC Power Distribution

The proposed 800 VDC architecture offers several advantages:

Native 800 VDC end-to-end integration: Eliminates redundant conversions, improving overall power efficiency and supporting high-density GPU clusters.
Reduced copper and cost: Allows the same wire gauge to carry 157% more power than 415 VAC, reducing copper usage, material, and installation costs.
Improved efficiency: A native DC architecture eliminates multiple AC-to-DC conversion steps, boosting efficiency and reducing waste heat.
Simplified and more reliable architecture: A DC distribution system inherently simplifies the infrastructure, with fewer components and potential failure points.

Energy Storage Strategy

To address workload volatility, NVIDIA proposes a multi-layered energy storage strategy. This includes short-duration storage using high-power capacitors and supercapacitors, and long-duration storage using facility-level battery energy storage systems (BESS).

Implementation and Adoption

NVIDIA's upcoming Kyber rack architecture will utilize the new 800 VDC architecture, distributing power at high voltage directly to each compute node. The company is collaborating with industry partners and publishing a technical whitepaper, "800 VDC Architecture for Next-Generation AI Infrastructure," to accelerate adoption. Details will be presented at the 2025 OCP Global Summit.