PRODUCTION SCENARIO
An ecommerce company wants an Amazon Bedrock model to write a two-sentence highlight for each of 2 million archived customer reviews. The results are loaded into an analytics table the next morning. No user waits for an individual response.
Which invocation method will meet these requirements at the lowest cost?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: Batch inference with the prompts and results in Amazon S3
Batch inference reads prompts from Amazon S3, runs asynchronously and writes responses back to S3, and AWS prices it at half the on-demand rate, which fits a dataset where nobody waits on a single response. Provisioned Throughput bills hourly for dedicated capacity whether or not the job is running. A Step Functions map state works but rebuilds batch by hand at full on-demand prices. Cross-Region inference absorbs peaks and still pays the on-demand rate per token.