A municipal rail-yard vibration set is tens of tebibytes of sequential files that training jobs will stream repeatedly. Object GET overhead dominates. Which initial storage decision fits that I/O pattern?

Think of tens of tebibytes of vibration files that trainers will stream over and over, with object GET overhead eating the clock. Put that read path on a high-performance FSx file system. S3 can still land the bytes, but GET-only is the wrong training path.