Infineon bets on 800V DC for megawatt AI racks

Infineon's pivot to 800V DC is less a product update than a structural admission: AI's bottleneck has migrated from silicon to the power grid. At 100kW+ rack density, 48V distribution hits a hard ceiling in copper loss and conversion efficiency. Voltage escalation is physics, not preference. The cascade runs deeper than headlines. An 800V bus renders 650V SiC MOSFETs marginally inadequate, forcing 1200V+ devices and accelerating the 150mm-to-200mm substrate migration. GaN becomes irreplaceable above 200kHz in solid-state transformers. Vertical power delivery pushes modules from board-level to package-level, making chiplet-based VRM the dominant paradigm by 2026-27. Supply-chain exposure is underpriced. SiC substrate capacity concentrates in China (SICC, TankeBlue) and the US (Wolfspeed). Infineon, as a European, is squeezed from both ends. Decoupling could inflate 200mm substrate costs 15-25%, compressing margins. Onsemi has locked early positioning via NVIDIA's 800V reference design; STMicroelectronics leverages roughly 35% SiC device share to pressure pricing. Infineon differentiates through full-stack coverage from grid-side SST to chip-level VRM—betting on system integration over single-device performance. The real inflection is NVIDIA's Rubin power specification. Once finalized, the power semiconductor chain reshuffles within six months. Infineon's window is narrower than the market assumes.