Hi @zackljones and @Mehrbod_Estaki,
The reason for this is that there is a parameter --p-n-reads-learn which is the number of reads to use to estimate the error model.
By default this is set to 1,000,000 and once it has acquired that many reads, it stops looking for more.
In practice this means you will see it run through a couple (or even most) of your samples before it has captured enough reads. Although I've seen this step completed in as few as 2 samples for very large datasets.