SK hynix Unveils 3D FeNAND for AI Computing - THE ELEC, Korea Electronics Industry Media

FeNAND is not 'faster NAND'—it is the missing tier in AI's memory hierarchy. When HBM3E pushes bandwidth past 3TB/s yet inference latency remains the binding constraint, SK hynix's ferroelectric polarization mechanism sidesteps the charge-trap physics ceiling entirely, compressing read/write from microseconds into nanoseconds. That is the real weapon. Upstream, HfO₂-based ferroelectric thin-film yield will gate 2026 mass production; material suppliers (Tokyo Ohka, Shin-Etsu) face immediate allocation pressure. On compliance, if the 3D stacking process dips below 14nm, the US BIS 'advanced memory' definition could sweep it into export controls, forcing OSAT players in Taiwan, China (ASE, PTI) to re-qualify customer lines. Competitively, Samsung's FeRAM roadmap and Micron's HBM4 bet form a pincer—but FeNAND positioned as an 'AI-inference-dedicated cache' complements rather than replaces HBM. The genuine threat: it rewrites the BOM cost structure of NVIDIA's next-gen ASICs. 12–24 month call: sampling by Q4 2025, small-volume in H1 2026. Memory pricing power shifts from 'cost-per-bit' to 'effective bandwidth-per-joule,' and the entire AI infrastructure procurement logic gets restructured.