Explore the challenges and solutions of running large-scale Kubernetes clusters for AI deployments in this 25-minute conference talk. Dive into CoreWeave's experience managing over 3,000 Kubernetes clusters on 5,000 bare metal nodes with massive GPU resources to power modern AI applications. Learn about the partnership between CoreWeave and Loft Labs, and discover how they overcame obstacles in security, GPU provisioning, and scalability. Gain insights into the pitfalls, design choices, and architectural challenges faced over three years while developing a serverless Kubernetes offering. Topics covered include secure tenant isolation on shared infrastructure, achieving 10-second autoscaling, on-demand cluster and compute provisioning, and day-2 operations for managing a large fleet of clusters at scale.