NVIDIA is reportedly evaluating additional memory configurations for its Rubin Ultra AI accelerator rather than committing exclusively to twelve-layer HBM4E. Possible alternatives include using HBM4 instead of HBM4E or reducing each stack from twelve DRAM dies to eight. Reducing stack height would allow a fixed number of known-good DRAM dies to produce more complete HBM stacks. TrendForce reportedly believes an eight-layer configuration is more likely than abandoning HBM4E completely because Rubin Ultra’s principal memory upgrade concerns interface speed. NVIDIA may offer several Rubin Ultra variants with different memory capacity, stack height or qualified suppliers rather than using one universal design.