Now

Gave a talk with Daniele Polencic at KubeCon EU 2026 in Amsterdam. We talked about what actually happens when you request nvidia.com/gpu: 1 - device plugins, kubelet, the whole thing.

The room was packed and the feedback after was really good. Honestly that is mostly Daniele. He is so good at taking a complex topic and making it clear, every time.

Also got to meet my team for the first time: Salman, Amin, Bart and Amit. I didn't know how awesome each one was.

Recording is coming to CNCF soon. Until then:
→ Rightsizing GPUs in Kubernetes
→ GPU-enabled Platforms on Kubernetes

We had been working on rightsizing GPUs in Kubernetes with Daniele Polencic.

A pod requesting 2 GPUs needs both on the same node. Free GPUs scattered across the cluster won't help you. The same 50% SM Active can be totally fine for inference and a real problem for training. And PyTorch holds onto GPU memory after it's done, so nvidia-smi shows high usage even when your tensors are small.

Daniele's diagrams in this one are really good. Some of the clearest GPU visuals I've seen.

Raw, unfiltered stories from engineers who've lived through cluster outages, bare-metal nightmares, and 3 a.m. incident calls—and somehow came out smarter on the other side.