Why SK hynix is weighing US memory production - The Korea Herald

SK hynix's deliberation over US front-end memory fabrication marks a potential inflection point in the global semiconductor supply chain, driven less by cost arbitrage and more by a convergence of geo... Read original →

Core judgment: SK hynix's US footprint is not cost arbitrage—it is a structural repricing of the AI memory supply chain. Technical cascade: HBM4 hybrid bonding demands storage and logic dies complete TSV interconnect within a single package. If SK hynix lands in Ohio, its 1b-nm DRAM node must undergo compatibility validation against Intel's 18A stack—DRAM and logic process curves carry fundamentally different yield profiles and overlay precision requirements. No precedent exists. More critically: domestically produced HBM compresses delivery to NVIDIA or Broadcom custom ASICs from six to eight weeks to under two, a magnitude shift for AI cluster iteration cycles. Compliance and risk: CHIPS subsidies carry five-year operational reporting and IP disclosure obligations, exposing HBM core parameters to Commerce Department scrutiny. Solidigm's Wuhan NAND faces tightening export controls; the US move is insurance against China-capacity unavailability. Yet 54.3 trillion won in domestic commitments confirms: the US is an option, not a relocation. Market dynamics: Micron faces its first direct domestic DRAM competitor. Samsung has Texas NAND but no US DRAM—SK hynix's move pressures Samsung into a 2026 parity decision. For Intel, SK hynix is the existential validation of Foundry 2.0; without an anchor tenant, 18A's $20B-plus capex is a bottomless pit. Trend: Within 12 to 24 months, memory localization shifts from procurement bonus to contract precondition. Hyperscaler LTAs will embed friend-shored memory clauses, replicating the TSMC-Arizona template. The memory-maker-plus-cloud-plus-foundry tripartite model may become the AI infrastructure standard, with Intel's Ohio fab as the first stress test of foundry-memory convergence.