According to @_avichawla, MoE routing fragments batches, making grouped GEMM suboptimal at low concurrency and requiring weight-streamed GEMV kernels.