Plain and simple collection creation (or even shard splits etc.) get progressively slower as more and more collections are in the system. The culprit is the autoscaling framework (possibly some unnecessary policy computation), even when no autoscaling is being used whatsoever.
Here is how bad the situation is:
Btw, even when using createnodeset parameter, there is still a gradual slowdown (but less steep than the graph here).
This is a matter of grave concern for anyone running a Solr cluster with more than a few hundred collections.