Every time you send a row or record to your ML model, and it returns a result, it counts as a single prediction. For example - Sending a single record via API and returning a result counts as a single prediction. Uploading a CSV with 100 rows of data and producing prediction results for each row counts at 100 predictions.
This is different than uploading a CSV with 100 rows of data to build a machine learning model. That is simply considered training data and does not count toward a "prediction". It is only when you use that model on a new record to predict an outcome that is counts as a prediction.