Google Cloud recently achieved a major milestone by running a 130,000-node Kubernetes cluster using Google Kubernetes Engine (GKE). This experiment, which doubles the previous limit of 65,000 nodes, showcases the platform's ability to handle extreme scalability required for demanding AI workloads. During the test, the cluster maintained a Pod throughput of 1,000 Pods per second and stored over 1 million objects in optimized distributed storage, highlighting not just the number of nodes, but the overall system's performance under heavy loads. The architecture behind this achievement involves several key innovations. To manage read requests effectively, Google implemented features like the Consistent Reads from Cache and Snapshottable API Server Cache. These innovations allow the API server to serve data from an in-memory cache, significantly reducing the load on the central object datastore and improving response times. For storage, a proprietary key-value store based on Google’s Spanner was crucial, supporting 13,000 queries per second (QPS) for critical cluster operations without bottlenecks. Job scheduling also received a significant upgrade with the introduction of Kueue, which improves management of complex AI and ML workloads. Instead of focusing solely on individual Pods, Kueue allows for job-level scheduling based on priorities and resource allocations. Future developments aim to evolve Kubernetes toward workload-aware scheduling, providing a more holistic view of resource needs. This shift will facilitate better orchestration of tightly coupled applications in large-scale environments, making GKE an even more powerful tool for AI and high-performance computing tasks.