Source

Tags: ai-power, vrm, mpwr, mps, aosl, infineon, vicor, delta, navitas, ti, openvreg, vertical-power-delivery, vpd, kyber, rubin, blackwell, gpu-current, ir-drop, factorized-power

• Source: The AI Power Crisis — Part 3
• Author: Nutty (@NuttyCLD)
• Date published: 2026-04-19
• Date captured: 2026-05-01

TL;DR

Between 800V at the rack and 0.8V at the GPU die, there are only a few centimeters — and that's where the hottest battle in AI power lives. Blackwell B200 needs ~1,500A on the core rail; Vera Rubin pushes ~2,875A. To hold 1,500A within 20mV requires a total path resistance of ~13 micro-ohms — physics, not taste. Four competing topologies for 48V → 0.8V; over 18 months the MPS / Infineon / AOSL / Vicor / Delta lineup has flipped twice. MPS is projected ~70% Vera Rubin VRM share with ASPs ~60% higher than Blackwell. Morgan Stanley estimates Kyber rack power-solution value rises >10x vs GB200. The 25-year-old VRM is now the protagonist.

Key takeaways

• The arithmetic is brutal: 1,500A through ~13 micro-ohms or board voltage starts moving. A few cm of board, a few vias, copper thickness, package bumps, decap placement — all matter.
• Pentium III ~ tens of amps. Blackwell B200 ~ 1,500A. Rubin ~ 2,875A. Same circuit class is now ~60-115x more current. Loss scales with I² — same path → 100x more heat. Two options: more copper, or shorter path.
• Four competing topologies, spec...