Why Scaling AI Compute Performance Requires a New Power Architecture

Details

AI factories face a fundamental power delivery problem: each AC-to-DC conversion between the grid and GPU wastes energy, and as rack density climbs, small inefficiencies compound rapidly. NVIDIA's 800 VDC architecture reduces conversion stages by distributing power as high-voltage direct current from the facility level down to the compute rack. The initiative is a joint effort by NVIDIA, Google, and Microsoft through the Open Compute Project. A joint white paper was published in March 2026, followed by the LVDC Solid-State Transformer Specification v0.3 in July 2026. Over 80 equipment manufacturers and infrastructure companies are already building products to this specification. NVIDIA's MGX-compatible 800 VDC hybrid power rack arrives in H2 2026 and creates a practical bridge for existing AC facilities — it slots into current infrastructure and distributes 800 VDC within the rack row without changing the building's electrical system. More advanced configurations follow: a row power center supporting up to 2 MW per row (expected 2027) and full DC power blocks for new greenfield construction. Global data center electricity demand is projected to more than double by 2030, driven in part by AI workloads. Operators who built AC-era facilities need an upgrade path that doesn't strand existing land, power rights, or infrastructure investment.