When building AI, everyone talks about data, architectures, and compute. Almost nobody talks about the massive bottleneck hidden in plain sight: saving the model. As models grow, traditional synchronous saving stalls expensive hardware, costing time and money.
In this 15-minute lightning talk, we pull back the curtain on Orbax—the modern, open-source checkpointing powerhouse engine. We will explore how Orbax fundamentally changes state management through asynchronous writing, preventing your storage from bottlenecking your compute. Attendees will walk away understanding why modern ML infrastructure requires dedicated, smart persistence layers to scale seamlessly and efficiently.
