“A common misconception about Mixture-of-Experts models is that unused experts are not accessed; efficient serving requires large batch sizes where all experts are accessed, but each processes only a fraction of the batch.”