NVIDIA described Agent Substrate’s Kubernetes architecture and upstream development on 10 November 2026, saying the open-source runtime multiplexes agent sessions across a smaller pool of warm worker Pods. The company identifies it as an independent Cloud Native Computing Foundation (CNCF) Sandbox project integrating with NVIDIA DSX OS components, rather than a named DSX OS component.[1]
According to the post, a separate control plane maps actors, meaning agents and their environments, onto workers. Idle actors are snapshotted from gVisor or microVM sandboxes with memory and filesystem state, then restored onto available workers. This avoids creating a new Kubernetes Pod for each activation; ordinary Pod scheduling, quotas and placement policies still apply.[1]
NVIDIA describes request parking with bounded waits and queue limits, plus per-container CPU and memory limits for microVM sandboxes. GPU device support using Kubernetes Dynamic Resource Allocation is still being developed. The proposed GPU checkpoint design records device compatibility properties and restores onto matching hardware, falling back to a cold start when no match is available. Parts of the Dynamo integration are experimental, and APIs are changing.[1]
NVIDIA claims support for millions of agents and tools at 10 times the density of standard containers. The saved post supplies no benchmark methodology or results establishing that comparison.[1]
[1]: NVIDIA Technical Blog, “Agent Substrate: Advancing Kubernetes for Agents at Scale,” 10 November 2026. https://developer.nvidia.com/blog/agent-substrate-advancing-kubernetes-for-agents-at-scale
Sources
Check or correct this page
Verification record
The digest below shows whether the text you are reading matches the published source. It does not show that the reporting is correct. The build computes it from this page's Markdown source and lists it in the site manifest.
e922c5301a2161912eee1fc90cb646c95ea9fdb31761459eafa3d08fd1364e9a