Infineon and NVIDIA collaborate on GaN-based power conversion - Compound Semiconductor

Infineon's lock-in with NVIDIA is not a co-development deal—it's a power architecture handover in data centers. With GB200 NVL72 racks exceeding 120kW, silicon MOSFETs hitting a 94% efficiency ceiling can no longer sustain AI cluster power economics. GaN's 10-50MHz switching frequency compresses magnetic component volume by 60%+—system-level restructuring, not parameter tuning. The technical ripple tears along two axes: upstream, GaN-on-Si epitaxy capacity becomes the scarcest resource by 2026, with I-HQ (Taiwan, China) and J-Device (Japan) facing order surges; downstream, UPS/PDU vendors like Vertiv and Eaton face forced BOM restructuring. The most fragile compliance node isn't Infineon's German fab—it's the deep dependency on MOCVD epitaxy equipment from Taiwan, China supply chains. If US export controls expand to cover GaN epi tools, lead times stretch from 12 to 26+ weeks. Competitively, Infineon's moat isn't GaN itself but its automotive-grade reliability infrastructure—data centers demand 5-year MTBF exceeding 1M hours, a bar Navitas and EPC haven't cleared. 12-24 month call: NVIDIA's Rubin platform (H1 2026) will likely make GaN power standard, not optional, expanding the global GaN power device market from ~$2B to $5-6B. The real long-tail effect: Infineon's reference-design entry will squeeze TI and onsemi's $8B server PSU controller share.