A company has petabytes of audio recordings on-premises and needs to transcribe them to text and store the transcripts in S3 as quickly as possible. The on-premises network to AWS is limited to 100 Mbps. The transcription algorithm requires GPUs for inference. Which solution will deliver transcripts to the S3 bucket fastest?

The correct solution leverages Snowball Edge Compute Optimized with an NVIDIA Tesla module because it provides the necessary GPU resources for the transcription workload directly on-premises, overcoming the limited network bandwidth for the large audio files. Processing the data locally minimizes the amount of data transferred over the slow network. AWS DataSync then efficiently transfers only the smaller resulting transcripts to S3. Snowcone devices are too small for petabytes of data and do not offer GPU instances like Inf1. AWS Outposts would be a significant and time-consuming deployment for a one-time data migration and transcription task. Copying all petabytes of audio files to S3 first via the 100 Mbps link would be extremely slow, making the overall process much longer, even if Lambda functions could handle the transcription.