Smarter, Not Faster: The Storage Reality Hidden in 85,000 AI Checkpoints

The race to build frontier-scale AI factories has created immense pressure to pair massive GPU clusters with equally massive storage systems, chasing theoretical multi-terabyte-per-second performance. But what if this common approach is based on a myth? What if production workloads behave differently?

In this VAST Live session, Glenn K. Lockwood, Principal Technical Strategist at VAST Data, lead a data-driven exploration of the real storage demands behind training large language models. Drawing insights from a comprehensive analysis of over 85,000 model checkpoints, Glenn revealed that the actual I/O requirements for global storage are surprisingly modest. He unpacked how the universal practice of asynchronous checkpointing decouples training from storage performance, making "checkpoint overlap" – not raw bandwidth – the most critical metric for efficiency.

Stream now to gain a new framework for designing and provisioning AI infrastructure, including a practical sizing model to calculate real-world bandwidth needs. Learn how to stop overprovisioning, avoid unnecessary costs, and strategically invest your budget in what matters most: maximizing your GPU capacity.