Source

Tags: ai-power, 800v-dc, gan, sic, vrm, ev-supply-chain, infineon, navitas, ti, vicor, vertiv, eaton, schneider, delta, mt-diablo, kyber, rack-power-density, wolfspeed, onsemi

• Source: The AI Power Crisis — Part 2
• Author: Nutty (@NuttyCLD)
• Date published: 2026-04-17
• Date captured: 2026-05-01

TL;DR

Google moved racks from 12V to 48V in 2010 and got ~30% better distribution efficiency. AI has now broken 48V — at 1 MW per rack, 48V implies 20,800A and ~200 kg of copper busbar PER rack (200 tons per GW). The industry is converging on 800V DC, but the supply chain chose 800V, not engineering — Porsche/BYD/Hyundai E-GMP/VW PPE built a mature 1200V SiC + 650V GaN ecosystem over 5-7 years that AI hyperscalers can step onto without inventing it. Two competing paths: NVIDIA's full 800V DC architecture (2027 Kyber timeline) and OCP Mt. Diablo sidecar (Microsoft + Meta + Google, 2025-2026, keeps 480 VAC backbone + AC/DC sidecar).

Key takeaways

• 48V → 800V is the same physics as 480V → HVDC. P=VI, loss = I²R. Raise V by 4x, cut current to 1/4, cut loss to 1/16.
• At 1 MW/rack, 48V = 20,800A = 200kg copper busbar/rack = 200 tons/GW. Meta's ORv3-HPR hits a practical air-cooled ceiling around 300 kW. Beyond that, thicker copper is no longer the answer.
• Why 800V won: the EV supply chain. Porsche Taycan (2019), Hyundai-Kia E-GMP, VW PPE, BYD Super e-Platform (1,000V), Tesla Cybert...