[OCP in Focus] SK hynix Drives up to 5x Faster Time-to-First-Token with MARS; HBF Expands Capacity by 14x - TrendForce

Core thesis: SK hynix isn't "speeding up" inference—it's rewriting the storage economics of AI. Technical cascade: HBF pulls NAND into the high-bandwidth tier, directly eroding HBM's monopoly in inference workloads. Over 70% of LLM weights sit in "warm data" territory, previously forced to compete for HBM bandwidth. Now they have a dedicated lane. The compute-to-storage ratio shifts from 1:1 toward 1:4, fundamentally altering GPU/ASIC architecture paradigms. Compliance & risk: HBF fabrication depends on 200+ layer 3D NAND stacking—equipment still runs through ASML and AMAT. More critically, once OCP standardization solidifies, BIS will almost certainly follow the 2023 HBM-specific export control precedent and classify HBF under "advanced storage," materially raising compliance costs for US-bound shipments. Competitive dynamics: Samsung is the existential threat—holding both HBM and NAND lines makes HBF a "left-hand-integrates-right-hand" structural advantage. Micron, without a credible advanced NAND roadmap within 18 months, risks total marginalization in AI inference storage. 12-24 month outlook: HBF lands in mainstream inference server BOMs. The real long-tail effect: inference token costs drop an order of magnitude, and the storage chip industry migrates from "cyclical commodity" to "compute infrastructure" logic—restructuring the entire AI deployment economic model.