NVIDIA pivots to agentic AI at GTC
NVIDIA used GTC 2026 to shift from 'model as product' to 'AI as agent', releasing an open Agent Toolkit and OpenShell runtime so enterprises can build and run self‑evolving AI workflows across edge and cloud devices. That move accelerates production deployments that can directly automate manufacturing, inspection, and knowledge work at scale. (nvidianews.nvidia.com)
NVIDIA announced the Agent Toolkit and OpenShell runtime at GTC on March 16, 2026, during Jensen Huang’s keynote. (investor.nvidia.com) The Agent Toolkit bundles open models (NVIDIA Nemotron), an open agent blueprint called AI‑Q, optimization skills like cuOpt, and the OpenShell runtime as core components for enterprise agents. (investor.nvidia.com) NVIDIA says the AI‑Q hybrid architecture uses frontier models for orchestration and Nemotron for research, which the company claims cuts query costs by more than 50% while topping the DeepResearch Bench accuracy leaderboards. (investor.nvidia.com) Seventeen major enterprise partners — including Adobe, Salesforce and SAP, plus Atlassian, ServiceNow and Siemens — signed on to adopt the Agent Toolkit at launch. (venturebeat.com) OpenShell is published as an open‑source runtime that enforces policy‑based security, network and privacy guardrails and creates sandboxed execution environments that agents cannot override. (investor.nvidia.com) NemoClaw, introduced alongside the toolkit, packages Nemotron models and the OpenShell runtime for one‑command installation onto OpenClaw agents and adds enterprise security/sandboxing for deployments on RTX PCs and similar endpoints. (venturebeat.com) LangChain is partnering to integrate AI‑Q, OpenShell and Nemotron into its agent libraries, and several security vendors including TrendAI/TrendMicro announced support for OpenShell to secure enterprise agent rollouts. (eweek.com)
