Nadia Prasad

S3 durability protects against hardware failure, not credential compromise.

NVMe caching layers let GPUs wait less for S3 data.

Diagnosing and fixing five distinct startup bottlenecks that pile up before serving requests.

Shared NVMe caches between S3 and GPUs slash I/O latency by orders of magnitude.

Misunderstanding day-counting and filters turns cost-saving rules into budget disasters.

Cloud-native file services lock you into a single provider, making multi-cloud training costly.

Streaming data directly to GPUs eliminates costly delays before training actually begins.

Storage performance is now the top bottleneck keeping GPUs idle during AI training.