Infinite Agents, Finite Kubernetes
summary
Argues that agent workloads are exploding as short-lived heavy-compute pods that spike GPU/CPU demand and stress the Kubernetes control plane, while space power and capital remain strictly capped. Proposes a concrete architecture: a trade-off curve plus hard enforced boundaries around workloads, standardized across a fleet to enable tight packing and cross-facility dynamic scaling. Claims this pattern preserves security and raises resource efficiency for nation-scale shared infrastructure; presents it as a proposed prototype rather than a deployed, measured system.
AI-written summary · about
related talks
06mFrom Containers To Agents: The Next Cloud Native
24mPropagating Programming Paradigms: Lifting High Performance Compute Into K8s
25mWhat We Learned Securing AI Agents at Scale on Multi-Tenant Kubernetes Clusters
19mNo Token Budget: Agentic Software Factories on Kubernetes
05mSponsored Keynote: The Cloud (And Your $$$) is Not Infinite: Dynamic Scheduling at Every Layer
29mKubernetes Workload Resiliency in Action: Beyond Basics
09mKeynote: Kubernetes for GenAI: The New Operating System of the AI-First World
30mSo You Want To Run AI Agents on Kubernetes: A 101 Guide