Analysis: AI supply chain expansion persists despite concerns over a downturn

Everyone watches the GPU; almost nobody tracks the real bottleneck migrating downstack into passive components and packaging substrates. Murata's MLCCs, Ibiden's ABF substrates, Sumitomo Electric's copper interconnects, and Kioxia's NAND form the capillary network beneath every AI accelerator. HBM stacking demands exponential growth in high-capacitance MLCCs; ABF layer counts are jumping from 12 to 20-plus. Capacity expansion here is a physics constraint, not a commercial bet. Geopolitically, Japanese firms occupy a neither-US-nor-China gray zone, making them irreplaceable neutral nodes in global AI infrastructure. Yet if Washington extends export controls to substrates and passives, their revenue structures face cliff-edge restructuring — the 2018 Toshiba memory spin-off could replay in a different costume. Kioxia's profit elasticity signals NAND's repricing from cyclical commodity to AI-structural demand. If Samsung and SK Hynix follow with capex, 2026 oversupply risk resurfaces, though AI inference workloads will absorb much of the incremental output. 12-24 month call: AI capex hits its first digestion phase in 2025Q4, but component-layer fab buildout runs 18-24 months, creating a demand-supply scissors gap by 2H26. The durable moat is process know-how, not capex. Firms with three-year offtake contracts in substrates and passives will outlast the cycle.