SK hynix Sees AI Infrastructure Growing 20-Fold by 2030 - Businesskorea

SK hynix's "20x" is less a demand forecast than a capacity-locking signal to capital markets and the supply chain. By 2030, HBM bit demand could leap from roughly 20 billion to 400 billion units annually — an exponential stress test on TSV yield and CoWoS-class packaging throughput. The real bottleneck isn't memory but power delivery and thermal management. Once per-GPU TDP crosses 1,500W, liquid cooling shifts from optional to mandatory, rewriting PCB, copper interconnect, and transformer roadmaps. SK hynix's HBM4 hybrid bet is essentially using the memory layer to absorb compute's thermal budget. US BIS HBM export controls are granulating toward customer- and use-case-level restrictions, inflating SK hynix's compliance audit costs by 15–20%. CXMT's HBM2E catch-up window compresses to under 18 months, forcing Samsung and SK hynix to accelerate HBM4 mass production to build a generational moat. NVIDIA's Rubin platform will split HBM4 orders across three suppliers; allocation hinges on 16-high stack yield. SK hynix clearing 85% by Q1 2026 locks ~40% share; otherwise Samsung's 4nm logic-die integration may flip the script. 12–24-month tail: CXL 3.0 memory pooling opens a non-HBM AI memory track, structurally benefiting LPDDR5X. But power-infrastructure lead times have stretched to 18 months — the "20x" storage narrative is ultimately capped by the grid's "1.5x" physical ceiling.