The startup time of a HDInsight cluster is quite long; it's not really suited for ad hoc clusters (which are really easy to spin up quickly in Databricks).
They spin up the VMs with either open source Spark or the Databricks runtime - you get to choose the distribution and version before spinning up the cluster.
If you have enough workloads you can run your own pool of VMs to provide a 'serverless' experience to your Spark users: https://databricks.com/blog/2017/06/07/databricks-serverless...