Instant KAI Sandboxes with vCluster: Multi-Tenant, Multi-Scheduler GPU Sharing

Kubernetes offers many ways to share GPUs, but a single, cluster-wide scheduler often forces trade-offs between utilization, stability, and team autonomy. This talk shows how vCluster makes the NVIDIA Kubernetes AI Scheduler (KAI) run as an opt-in service for each tenant—so platform teams can raise GPU density while keeping operations predictable. What We’ll Cover Problem statement – why mixed workloads leave GPUs under-used and complicate on-call. vCluster fundamentals – lightweight control planes that isolate scheduling logic, not hardware. KAI at a glance – fractional GPU allocation, gang queues, topology awareness. Live demonstration – two vClusters on one host Key Takeaways A reproducible pattern for running different schedulers side-by-side. Practical steps to increase GPU utilisation without adding more clusters. An isolation model that lets teams experiment safely.