Samsung is sending Nvidia’s custom high-bandwidth base die to TSMC. The short-term win and the longer risk are not the same. Nvidia wants one standardized logic die so Samsung, SK Hynix, and Micron all deliver the same performance without separate qualifications. TSMC’s 3nm process is mature, yields are higher, and capacity is already booked by the same AI customers Samsung wants. By accepting the rule, Samsung keeps a seat for high-volume HBM4 and HBM4E instead of watching orders go elsewhere. The upside is real and near term. It cuts customer friction, speeds validation, and lets Samsung’s memory business capture AI demand while its own foundry is still closing the gap on advanced nodes. An industry source notes Samsung is also running parallel qualification on its own lines, so this is not pure surrender. The longer risk is structural. Every custom base die that ships on TSMC silicon is foundry volume and process learning Samsung does not get. If the arrangement becomes the default, Samsung’s foundry stays in the role of memory supplier while TSMC deepens its hold on the high-margin logic under every AI accelerator. Yields and cost matter less if the strategic node never fully moves in-house. Standardization buys time, but it also normalizes dependence on the competitor that already leads advanced foundry. A process request is not a finished supply agreement. #Samsung $TSM