“Running Mixture of Experts (MoE) models on traditional DRAM-based hardware requires a batch size of hundreds to be economically viable due to memory access overhead.”