Back to feed

Agent Substrate: Advancing Kubernetes for Agents at Scale

NVIDIA Technical Blog

Nov 10, 2026

11/10/2026

GPU Checkpointing Reduces GPU Interchangeability And Requires Compatibility-Aware Scheduling To Avoid Cold Starts

Agent Substrate: Advancing Kubernetes for Agents at Scale · NVIDIA Technical Blog

Science, Technology & Innovation · Nov 10, 2026

GPU suspend/resume requires preserving sandbox and GPU state, so restoration may be limited to compatible GPU models and driver versions; otherwise the system must cold-start, requiring compatibility-aware scheduling and capacity planning.


11/10/2026

Coordinated Agent-Environment Scheduling Driven by Inference Telemetry Optimizes Cache Reuse, Tool Execution, and GPU Utilization

Agent Substrate: Advancing Kubernetes for Agents at Scale · NVIDIA Technical Blog

Science, Technology & Innovation · Nov 10, 2026

NVIDIA proposes coordinating agent scheduling with real-time inference telemetry so systems can optimize KV-cache reuse, GPU utilization, agent suspension/resumption, and tool-driven workflows, reducing idle CPU capacity and repeated GPU context processing.


11/10/2026

Agent Substrate Replaces One Pod Per Session With Multiplexed Stateful Actors On A Smaller Warm Pool

Agent Substrate: Advancing Kubernetes for Agents at Scale · NVIDIA Technical Blog

Science, Technology & Innovation · Nov 10, 2026

Agent Substrate multiplexes many stateful agent sessions across a smaller pool of worker Pods by snapshotting idle gVisor or microVM sandboxes and restoring them on demand, aiming to reduce Kubernetes startup overhead and idle resource use; NVIDIA claims up to 10× container density, though the text provides no benchmark details.