NVIDIA today announced NemoClaw and Nemotron 3 Ultra, a complete open-source stack for building and running enterprise AI agents. The announcement came alongside a roster of major software partners already using the platform, making it the most comprehensive push NVIDIA has made into the enterprise agent layer above the hardware.
What Was Announced
NemoClaw is an open-source stack that gives enterprises a single, opinionated foundation for agent development. It bundles NemoClaw blueprints (pre-built agent patterns), Nemotron models, the OpenShell secure runtime, and CUDA-X libraries loaded with agent skills. The install is a single command, and the design separates the agent’s capabilities from the environment it runs in.
OpenShell is the governance piece. It defines exactly what an agent is allowed to do, what it cannot access, and which actions require a human to approve before proceeding. For businesses that have been reluctant to deploy agents because of compliance and control concerns, this is a meaningful development.
Nemotron 3 Ultra is the model powering these long-running agents. It is a 550-billion-parameter mixture-of-experts model built for coding, research, and enterprise workflows. NVIDIA reports up to 5x faster inference and up to 30% lower cost compared with open frontier models in the same class. It will be available June 4 on Hugging Face, ModelScope, OpenRouter, and build.nvidia.com as NIM microservices, with broad support from cloud partners.
Nemotron 3 Ultra has been post-trained to work with the leading agent orchestration frameworks: Hermes Agent, LangChain, OpenClaw, OpenHands, and OpenCode. Whatever orchestration layer your team is already using, the model is designed to plug in without re-engineering your stack.
Who Is Building With It
The partner list is notable. Cadence, Dassault Systèmes, Siemens, and Synopsys are using NemoClaw to build autonomous AI engineers that work as digital coworkers on simulation and verification workflows, compressing what previously took weeks into hours. These are not demos. These are production engineering workflows at companies that run complex technical infrastructure at global scale.
On the security and operations side, CrowdStrike and Palantir are deploying Nemotron-powered agents for long-running cybersecurity analysis and operational decision-making. Both companies deal in high-stakes, high-data-volume environments where agent reliability matters.
Microsoft is collaborating with NVIDIA to deliver native Windows AI agents built on OpenShell’s security primitives. Canonical and Red Hat are integrating OpenShell as a standard open-source runtime across PCs, data centers, and cloud environments. ServiceNow extended its NVIDIA partnership to deploy governed autonomous agents for enterprise desktops.
What This Means for Business
Three things stand out for business owners and technical leaders thinking about AI agents.
The cost equation shifted. If your team is running agents at any real volume, a 5x inference throughput improvement means significantly lower compute bills for the same work. Agents that previously took hours can finish in fractions of that time. That changes the math on which workflows are worth automating.
Governance is now built in, not bolted on. OpenShell’s approach to defining agent boundaries gives compliance-sensitive industries, including financial services, healthcare, and legal, a credible path to agent deployment. The “what can this agent touch” question has a structured answer, not just a policy document.
The stack is now complete. For two years the challenge with enterprise agents has been that the parts did not fit together cleanly. Great orchestration frameworks, but model performance was a bottleneck. Powerful models, but no standard deployment story. NemoClaw attempts to close that gap with a single installable stack that connects model, runtime, security, and tooling.
This does not mean agent deployment is easy. The hard work is still identifying the right workflows, mapping your data, and building the muscle to operate agents in production. Most businesses discover they have foundational gaps that need addressing before the first agent goes live. But the infrastructure excuse has significantly weakened.
If you are evaluating whether to run your first production AI agent, the platform landscape is more mature than it was six months ago. The question to ask now is not “is the technology ready” but “which specific process in our business should run first.” Understanding what a deployed agent actually does across a working day helps frame that decision clearly.
Source
GlobeNewsWire