AI bottleneck shifts from compute to power; European supply chain sees power semiconductors as next shortage

The value anchor in AI infrastructure is migrating from compute-per-watt to power-conversion efficiency. Once clusters cross the gigawatt threshold and 800V DC bus architectures replace legacy 54V rails, power-semi content per megawatt jumps 3-5x. This is an architectural discontinuity, not a linear ramp. The binding constraint has shifted from GPU yield to 200mm SiC wafer capacity and automotive-grade qualification cycles. Infineon's Villach and ST's Catania lines cannot close the gap within 18 months. EU's Net Zero Industry Act, layered with tightening PUE mandates, effectively mandates localized power-semi sourcing. Firms single-sourced from China or Japan face a step-change in compliance friction. Infineon is staking its roadmap on an AI Power platform; ST leverages vertical SiC integration. Meanwhile, Chinese IGBT suppliers are compressing the cost curve by 30-40%, forcing European OEMs to re-architect vendor matrices. 12-24 month call: expect a power-semi supercycle mirroring the 2021 memory squeeze. The bottleneck is not demand — it is 200mm SiC substrate output and qualified device libraries. Whoever locks wafer capacity first captures pricing power over the entire AI buildout.