SK Hynix says customers want more than twice the HBM it can supply

The most telling signal isn't the 2x demand gap—it's that customers are voluntarily accepting shorter stacks. In a balanced market, buyers optimize for peak bandwidth. Choosing lower performance to secure allocation is textbook rationing behavior, confirming HBM has shifted from a design choice to a hard constraint. Downstream, this gates the entire AI accelerator roadmap. NVIDIA's GB200, AMD's MI300X, and hyperscaler ASICs all depend on HBM availability before TSMC's CoWoS-L packaging even becomes the secondary bottleneck. The supply chain's metronome has moved from logic die to DRAM wafer starts. On competition: Samsung's HBM3E qualification with NVIDIA is the pivotal near-term variable. If volume orders materialize, SK Hynix's pricing leverage erodes. Micron's HBM3E is in production but at limited scale. US CHIPS Act subsidies are redirecting investment to Micron (Chandler, AZ) and Samsung (Taylor, TX), yet TSV and hybrid bonding—the core HBM processes—remain concentrated in Korea. No US-based capacity can close the deficit within 18 months. 12–24 month outlook: HBM will overtake logic as the semiconductor industry's margin engine. Customers will begin architecting HBM-agnostic designs to reduce dependency. HBM4 (12+ high) timelines compress. This isn't a cyclical squeeze—it's structural scarcity.