NVIDIA GTC teases developer AI tools

NVIDIA used its March 16 GTC keynote and pregame in San Jose to preview a developer push around AI agents, open models and toolchains rather than consumer features. (nvidia.com) The conference ran March 16-19, 2026, with workshops on March 15, and NVIDIA said the keynote would cover “the full AI stack” from open models to agentic systems and physical artificial intelligence. (nvidia.com 1) (nvidia.com 2) In NVIDIA’s own recap, Chief Executive Officer Jensen Huang said GTC would span “every single layer” of a five-layer artificial intelligence stack, and the company counted more than 450 sponsors, 1,000 sessions and 2,000 speakers. (blogs.nvidia.com) An AI agent is software that can take actions for a user, like a digital assistant with tools and memory instead of a one-shot chatbot. NVIDIA’s GTC materials centered that idea with references to “agentic systems,” “self-evolving agents” and open model runtimes. (nvidia.com) (nvidianews.nvidia.com) The clearest product signal was NVIDIA Agent Toolkit, announced March 16, which the company said includes OpenShell, an open-source runtime for building and running autonomous agents and “claws” with added safety and security controls. (nvidianews.nvidia.com) NVIDIA paired that enterprise pitch with NemoClaw, a stack for the OpenClaw agent platform that the company said installs Nemotron models and OpenShell “in a single command” and adds privacy and security controls. (nvidianews.nvidia.com) (nvidia.com) The company also tied those tools to local hardware. On March 17, NVIDIA said RTX personal computers and DGX Spark systems could run OpenClaw-style agents locally, and said DGX Spark has 128 gigabytes of unified memory and supports models with more than 120 billion parameters. (blogs.nvidia.com) That local-computing angle addresses two problems developers keep running into with AI agents: cloud cost and data exposure. NVIDIA said local deployment can cut application programming interface costs to zero on supported devices and keep data private on the machine. (blogs.nvidia.com 1) (blogs.nvidia.com 2) NVIDIA’s model announcements were aimed at the same audience. The company said Nemotron 3 Nano 4B targets compact local assistants, while Nemotron 3 Super 120B is designed for more complex agent systems on DGX Spark and RTX Pro workstations. (blogs.nvidia.com) The broader GTC press kit shows how NVIDIA framed the week: Vera Rubin chips and racks for “agentic AI,” Dynamo inference software for AI factories, open model families, and an open agent development platform all appeared alongside the keynote recap. Consumer graphics news such as Deep Learning Super Sampling 5 was present, but it sat beside a much larger set of developer and infrastructure releases. (nvidianews.nvidia.com) (blogs.nvidia.com) By the time the pregame rolled into the keynote, NVIDIA had already set the frame: this GTC was about giving developers more pieces to build, run and secure AI agents across cloud, workstation and edge hardware. (nvidia.com) (nvidianews.nvidia.com)