Key takeaways:

• NVIDIA launched its Agent Toolkit at COMPUTEX 2026 — NemoClaw blueprints, Nemotron open models, and OpenShell secure runtime.
• CrowdStrike, Palantir, Siemens, Cadence, Dassault, and Synopsys have announced deployments; broader ERP/CRM/HR adoption has not been confirmed.
• OpenShell is the policy and privacy control layer for enterprise agents — governing data access, output routing, and behavior rules.
• DGX Station for Windows (Q4 2026) is an announced local AI workstation; no pricing has been disclosed. Not a near-term purchase decision.

Your enterprise software vendors are making an AI infrastructure decision without consulting you. At COMPUTEX 2026 in Taipei today, NVIDIA announced its Agent Toolkit — an open-source set of models, blueprints, and a secure runtime that enterprise software companies are using to build autonomous AI workflows. CrowdStrike, Palantir, Siemens, Cadence, Dassault Systèmes, and Synopsys announced they are building on it today. Today’s named partners are in cybersecurity, defense intelligence, and industrial engineering — not yet ERP or CRM — but NVIDIA released the toolkit as open-source specifically to drive broader adoption.

What Did NVIDIA Announce?

The toolkit has three layers. NemoClaw is NVIDIA’s blueprint framework connecting AI models to orchestration and tool-use harnesses; Siemens, Cadence, and Dassault announced they are using it for autonomous engineering workflows. OpenShell — co-developed with Microsoft, Canonical, and Red Hat — is the secure agent runtime that sets policy and privacy rules for what agents can access and where data goes; it is being integrated across Windows, Linux, and cloud environments. Nemotron are NVIDIA’s published open enterprise AI models for long-running agent tasks; according to NVIDIA’s developer blog, they are optimized for higher throughput in agentic applications — independent benchmarks have not yet confirmed these figures.

Also announced: DGX Station for Windows, a desk-side AI workstation co-developed with Microsoft, supporting models up to 1 trillion parameters locally. It is expected Q4 2026 with no pricing disclosed. This is an announcement, not a buying opportunity — treat it as a signal to watch if on-premises AI processing is a near-term requirement.

Why Should Operators Care?

Enterprise AI conversations have focused on which model your vendor chose: OpenAI, Anthropic, or Google. NVIDIA is doing something different — inserting itself as the runtime layer between the model and the software application. That layer controls what agents can do, what data they access, and how governance rules are enforced.

The toolkit is open-source and designed for broad enterprise software adoption. Microsoft Build 2026 starts tomorrow; NVIDIA’s OpenShell collaboration with Microsoft makes enterprise agent governance a likely theme. If SAP, Salesforce, Workday, or ServiceNow build on this toolkit — which is NVIDIA’s stated goal — operators will need visibility into that governance layer before agents receive production access.

The near-term move is to start asking. When any software vendor announces AI agents or autonomous workflows, add these to the conversation:

• What agent runtime governs this feature? What are the default policy controls?
• What data can the agent access? Can you restrict it to specific systems or datasets?
• Where does execution run — locally, in your vendor cloud, or through a third-party AI cloud?
• Who audits agent behavior? How are errors or unauthorized actions handled?

These questions apply regardless of whether the vendor uses NVIDIA or another stack.

---

FAQ

What is OpenShell and why does it matter for enterprise security?

OpenShell is NVIDIA’s policy and privacy control layer for enterprise AI agents — it governs what data an agent can read, where outputs go, and what containment rules apply. When a vendor says their AI agents are “secure,” ask whether they use OpenShell or an equivalent, what the default policy settings allow in your deployment, and whether you can configure restrictions. Governance defaults are the accountability question vendors will most want to skip over.

My software vendors haven’t announced AI agents yet. Does this affect me now?

Yes — indirectly. NVIDIA released this toolkit as open-source specifically to drive adoption across enterprise software categories, and your vendors are evaluating infrastructure choices now. Start asking about agent runtime, data-routing, and policy controls in renewal and expansion conversations before any agent feature is on the table. The time to negotiate governance terms is before deployment, not after.