Generally you train each expert simultaneously. The benefit of MoEs is that you get cheap inference because you only use the active expert parameters, which constitute a small fraction of the total parameter count. For example Deepseek R1 (which is especially sparse) only uses 1/18th of the total parameters per-query.