Samsung Electronics is betting on vertically stacked zHBM to move memory beyond the physical constraints of conventional 2.5D architectures, setting a system-level target of raising AI response throughput from about 100 to 1,000 tokens per second per user as workloads become more demanding.