How is the scaleTarget calculated when auto batcher is used? Is it based on the input 'instances' shape? or the batched request count

Let's say I'm using batcher with (max) batch size 32 and Auto scaler with scaleMetric:rps (or concurrency)
then if my scaleTarget is 64 am I getting (assuming perfect batching conditions):

64 calls to preprocess with "instances" shape (32,...) effectively making it 32*64 instances to process total (on one pod)

OR

2 calls to preprocess with "instances" shape (32,...) effectively making it 32*2=64 instances to process total (on one pod)

If it is the second case then what happens if scaleMetric < maxBatchSize ?