FSx for ONTAP now has AI-optimised storage options. A training read that crosses the S3 boundary costs more than the capacity it saves. Which constraint hits you first, throughput or capacity?