How to Share GPUs Across AI Agent Workloads Using NVIDIA Time-Slicing in Kubernetes

Rising GPU costs and limited hardware availability are making efficient resource sharing a critical need for AI workloads running in cloud environments. A technical guide published on DEV Community explains how to implement CUDA-based GPU sharing inside Agent Substrate Actors using NVIDIA's time-slicing feature on Kubernetes. The approach allows multiple Actors — each hosting an HTTP app or AI agent — to take turns using a smaller pool of GPU-backed Worker nodes, reducing both hardware and cost overhead. While most agent tasks like tool calls and planning are CPU-bound, edge cases such as local LLM inference or on-device image generation can require GPU access. The setup requires a Kubernetes cluster with at least one GPU-enabled worker node, the Agent Substrate repository, and Docker for building container images.