NVIDIA AI Updates: November 11, 2026
Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled
1. NVIDIA Donates OpenShell Agent Runtime to the CNCF Sandbox and Releases Open Source Kubernetes and Checkpoint Components
NVIDIA. NVIDIA announced at KubeCon + CloudNativeCon on November 10 that it is contributing OpenShell, an open-source runtime for securing and governing autonomous agents through policies on tool use, inference access and workload execution, to the Cloud Native Computing Foundation as a Sandbox project. NVIDIA says hundreds of organizations use OpenShell and that it has received tens of thousands of contributions in under a year. NVIDIA is also contributing to Agent Substrate, a Google-led framework for long-running agents on Kubernetes that was accepted into the CNCF Sandbox, and is working with Google on an OpenShell and Agent Substrate integration. The announcement adds three components of the NVIDIA DSX AI Factory Platform: NVIDIA Kubernetes Engine and NVIDIA Kubernetes Recipes, both in public preview, and NVIDIA Snapshot. The post does not name the licenses of the new projects. Source
2. NVIDIA Details Agent Substrate Contributions: Request Parking, microVM Limits and GPU Support Through Kubernetes DRA
NVIDIA / Agents. An NVIDIA technical post describes Agent Substrate as an open-source, zero-trust runtime that runs many sandboxed agent sessions on a small pool of warm Kubernetes worker Pods, checkpointing idle sessions and restoring them onto any free worker. The post says it targets millions of agents at about 10x the density of standard containers, with sandboxes running under gVisor or a microVM, and names coding assistants, RL post-training environments and GPU research sandboxes as target workloads. NVIDIA lists upstream work including request parking in the router, per-container CPU and memory limits for microVM sandboxes, and GPU support through Kubernetes Dynamic Resource Allocation, with NVIDIA GPUs first. Integrations with Dynamo (session-aware routing and signals on time to first token) and KAI Scheduler are described, and the post states the project is still evolving, with changing APIs and experimental parts of the Dynamo integration. Source
3. NVIDIA Snapshot Restores Checkpointed GPU Inference Replicas in 3.6 to 42.7 Seconds, 8x to 36x Faster Than Cold Starts
NVIDIA / Dynamo. NVIDIA released Snapshot, an open-source checkpoint and restore system for GPU workloads on Kubernetes that saves a fully initialized workload’s CPU and GPU memory and restores it onto any node with the same GPU model and driver. Examples exist for vLLM, SGLang and TensorRT LLM, and multi-GPU single-node workloads are supported with checkpoints up to 1 TB across eight GPUs. On B200 GPUs, restores took 3.6 to 42.7 seconds, for instance 1,110.65 s to 30.96 s for DeepSeek-V4.1-Flash on vLLM (35.87x) and 621.55 s to 42.69 s for DeepSeek-V4-Pro-NVFP4 on SGLang (14.56x). Design partner Crusoe cut GLM-5.3 pod readiness on HGX B300 from about eight minutes to 31 seconds. The code is at ai-dynamo/snapshot, and the Dynamo user guide for it is labeled experimental. Source
4. NVIDIA Kubernetes Engine Enters Public Preview as an Apache 2.0 Layer for Managed Kubernetes on GPU Infrastructure
NVIDIA / Kubernetes. NVIDIA introduced NVIDIA Kubernetes Engine (NVKE), an open-source software layer that lets AI factories offer managed Kubernetes clusters on their own GPU hardware. Each workload cluster gets its own control plane, etcd and node pools, bare-metal machines are provisioned through NVIDIA Infra Controller using a Cluster API provider, and nodes must present an attested SPIFFE identity before joining. NVKE works with NVIDIA Kubernetes Recipes, which builds the signed, version-pinned node images, though neither depends on the other at runtime. It is a Certified Kubernetes AI Platform tested on Kubernetes 1.36, licensed under Apache 2.0, and in public preview; NVIDIA does not yet recommend it for production. Firmus and SK Telecom are integrating it into their AI factories, and Mirantis, Rafay, Spectro Cloud, OpenNebula Systems and vCluster are incorporating components. Source
5. Alpamayo 2 Edge, a 2B-Parameter Driving VLA Model, Targets DRIVE AGX Thor at 160 to 190 ms Pixel-to-Trajectory Latency
NVIDIA / Autonomous Vehicles. NVIDIA introduced Alpamayo 2 Edge, a 2B-parameter vision-language-action driving model built on the Cosmos 3 Edge backbone with a diffusion action expert, designed for real-time operation on DRIVE AGX Thor. It emits Chain-of-Causation reasoning traces and trajectories, and can emit a <NOMINAL> token instead of a trace during routine driving to save compute. NVIDIA reports 15.15 km between at-fault incidents in closed-loop AlpaSim against 0.25 km for the 4B Qwen-Drive-1.0 baseline, and 160 to 190 ms pixel-to-trajectory latency on Thor depending on the amount of reasoning. The post also describes AlpaJudge, a human-calibrated judge for scoring reasoning quality that can serve as a reward in GRPO-style training, where NVIDIA reports 10.2% and 11.4% gains in judge-scored and human-evaluated quality. A recipe covers FP8 calibration, ONNX export and TensorRT engine builds. Source