The parallelism is nice, but it only helps cloud inferencers if there is an efficiency improvement. Autoregressive inference isn't parallel, but parallelism can be recovered with sufficient batching.