That number isn't hypothetical. Cast AI's 2026 State of Kubernetes Optimization Report — based on real utilization data across tens of thousands of clusters — found average CPU utilization at just 8%,… See more

That number isn't hypothetical. Cast AI's 2026 State of Kubernetes Optimization Report — based on real utilization data across tens of thousands of clusters — found average CPU utilization at just 8%, and GPU utilization at 5%. You're paying for the other 92%. The problem isn't that teams don't care. It's that Kubernetes doesn't optimize by default, and most infrastructure decisions were made before AI workloads entered the picture. CPU overprovisioning has soared to 69%, memory to 79% — and both are getting worse year over year. This talk walks through what's actually driving the waste: how overprovisioning becomes a habit, why GPU idleness is a different class of problem, and what cost-aware Kubernetes infrastructure looks like in production. We'll also touch on what we learned building Kimchi an AI cost intelligence layer and the real friction points teams hit when they try to fix this. You'll leave with a concrete framework for diagnosing waste in your own clusters before your next cloud bill arrives. Sobre Esmira Bayramova: I’m Esmira Bayramova, Senior Manager of DevRel & DevOps at Cast AI. With 8+ years in DevOps and platform engineering, I’ve worked across space tech and cloud. I’m a multiple-time Dan Kohn Scholarship awardee, recognized for my CNCF community contributions. I’m passionate about mentorship, supporting women in tech, and building inclusive, hands-on spaces where engineers can grow, connect, and share real-world experiences. Show less