Expensive accelerators require high-throughput, low-latency data planes to maintain saturation. In this session, explore Google Kubernetes (GKE) primitives designed for stateful scaling to achieve optimal performance for AI and enterprise databases, such as ClickHouse. Learn to architect storage solutions with Lustre and Cloud Storage FUSE to drive peak throughput for training and inference. We’ll demonstrate how storage profiles and attribute classes simplify performance tuning, and how Run:ai Model Streamer and key-value caching accelerate inference tokens-per-second. Leave with a battle-tested blueprint for zero-downtime scaling and a reference architecture for high-performance data on GKE.