KubeCon + CloudNativeCon India 2026

Two years ago, every platform team was building developer platforms. Today the same teams, and the AI Cloud providers selling to them, face a harder question. How do you safely share GPU infrastructure across multiple AI teams without stepping on each other?

Kubernetes is already the production platform for AI inference. Containers, autoscaling, multi-tenancy, RBAC are all solved. The remaining problem is GPUs themselves. Today Kubernetes asks for them in whole numbers. One pod, one GPU, even when the workload uses 10% of it.

This keynote walks through how the CNCF ecosystem, and Kubernetes itself, has answered. HAMi virtualizes one physical GPU into multiple slices today, each with its own memory budget. DRA evolves the platform's resource model so Kubernetes finally understands GPUs as rich devices instead of opaque numbers.

Then, we move to a live demo. There will be a MacBook on stage, connected directly to an NVIDIA DGX Spark. A single Blackwell GPU will run two open source LLMs for two teams, generating answers simultaneously. There are no slides and no recordings.

That's the AI factory. And yes, Kubernetes has solved it, with a little help from its friends.