After node replacement, rate of (re) initializing shards is decreasing

I started off replacing a node, and having 1200 unassigned shards.
The cluster happily started reassigning 8 shards at a time.
Then it hit somewhere under 1000 shards unassigned.. and suddenly, started to only do 4 shards at a time.
Um, what?

Might anyone have insight into why this is happening, and more importantly, how to fix it?

There is a limit per node, so if fewer nodes are involved at this point it could lead to fewer concurrent recoveries.