TSMC says it will remain a logic foundry after a report that it is exploring 3D DRAM for Nvidia - MLQ.ai

TSMC's "logic foundry" declaration is not a clarification—it is a strategic anchor for the AI compute architecture power map. Technically, 3D DRAM's TSV stacking and logic chips' GAA transistors belong to entirely different process stacks. TSMC's real moat is CoWoS advanced packaging—dictating how GPU dies and HBM dies interconnect on silicon interposers. Diving into memory manufacturing would dilute capex efficiency and trigger a price war with SK Hynix and Samsung during a period of AI chip scarcity. That is strategic suicide. Competitively, SK Hynix has locked in Nvidia's HBM3E supply, while Samsung's HBM4 yield ramp is the critical variable for 2025-26. TSMC's "stay in logic" stance objectively signals green light for Korean memory makers to expand capacity. Nvidia's memory bottleneck will hinge on stacking-layer races, not foundry capacity. On compliance, US CHIPS Act subsidies for domestic memory keep intensifying, while Taiwan, China's memory production lacks equivalent policy hedges. TSMC concentrating capex on Arizona logic fabs is the risk-optimal play. Over 12-24 months, the AI bottleneck shifts from compute to bandwidth. CXL memory expansion and near-memory computing will reshape HBM demand curves. TSMC does not make DRAM—but it decides how DRAM talks to the compute core. That is the real power position.