Nvidia stacks up agentic AI infrastructure

Nvidia has unveiled a broad software and infrastructure stack aimed at helping enterprises move AI agents from experimentation into production, introducing an open-source toolkit, a secure runtime environment, and a new processor architecture designed specifically for agentic AI workloads.

CEO Jensen Huang announced the new products on Monday during a keynote speech at GTC Taipei at Computex, targeting enterprises looking beyond generative AI assistants toward autonomous agents capable of executing tasks, accessing enterprise systems, and interacting with business workflows with limited human oversight.

The centerpiece of the stack is Nvidia Agent Toolkit, a collection of software components that combines Nemotron AI models, agent-development blueprints, CUDA-accelerated libraries, and a new secure runtime called OpenShell.

Nvidia said companies including Cadence, Siemens, Dassault Systèmes, CrowdStrike, Palantir, Microsoft, Red Hat, and Canonical are already integrating parts of the stack into products and enterprise platforms.

OpenShell may be the most significant element for enterprise technology leaders: It places governance and security controls beneath the agent layer rather than inside the model or orchestration framework itself. The runtime enforces access policies across filesystems, networks, and processes, while also providing sandboxed execution and privacy controls for AI workloads.

This architectural approach reflects a broader shift occurring across the enterprise AI market as organizations grapple with how to secure agents that can access applications, invoke tools, and perform actions autonomously.

Yugal Joshi, partner at Everest Group, said “Most of the runtime controls were at the agent process level. Nvidia is going a level below, making it more embedded and harder to escape.”

The industry has spent much of the past year attempting to scale AI agents through orchestration and governance layers, Joshi said, but Nvidia is now “pushing for building agent-native infrastructure layer and control planes rather than repurposing existing layers, which has been happening for quite a while.”

Agentic infrastructure

Alongside the toolkit, Nvidia introduced Vera CPU as a standalone product. The chip is already part of the Vera Rubin CPU-GPU double act, but Nvidia is now positioning Vera as a standalone CPU for agentic AI, reinforcement learning, and data-processing workloads. The company said Vera completes up to 1.8 times more tasks per second than x86 processors operating within the same power envelope and is being evaluated by organizations including Anthropic, OpenAI, SpaceXAI, ByteDance, CoreWeave, and Oracle Cloud Infrastructure.

Taken together, the launches position Nvidia as a supplier not only of AI models and accelerators, but also of the runtime, security, orchestration, and processor infrastructure it believes enterprises will need to support long-running autonomous AI systems.

Early deployments

“AI agents will use more tools than ever before,” Huang said during the keynote, arguing that agentic AI will drive a new generation of software and computing infrastructure.

Huang said Nvidia is working with companies including Cadence, Crowdstrike, Dassault, Palantir, SAP, and ServiceNow to build AI agents for semiconductor design, engineering simulation, and software and industrial workflows.

Nvidia said it is already using Cadence’s ChipStack autonomous verification agent internally, reducing chip verification cycles by more than 40 times compared with manual processes.

CrowdStrike is deploying Nemotron models in security operations, while Palantir is integrating them into its Forward Deployed Engineer platform to automate complex tasks inside air-gapped enterprise environments.

According to Joshi, the concentration of early adopters in engineering, manufacturing, and cybersecurity reflects where enterprises are currently most comfortable deploying autonomous systems.

The partner mix points toward industries “with structured workflows that have significant data availability and existing visible pain points,” he said, rather than heavily regulated sectors such as financial services or healthcare, where governance requirements remain more complex.

Building an agentic enterprise stack

Alongside the toolkit, Nvidia introduced Nemotron 3 Ultra, a 550-billion-parameter mixture-of-experts model designed for coding, research, and enterprise workloads. The company said the model has been optimized for agent frameworks including LangChain Deep Agents, OpenClaw, OpenHands, and OpenCode.

Microsoft, Red Hat, and Canonical are also integrating OpenShell into Windows, Red Hat AI, and Ubuntu environments, extending the runtime beyond Nvidia’s own infrastructure. SAP and ServiceNow had previously incorporated OpenShell into their enterprise AI initiatives.

For CIOs, the significance extends beyond another model launch. Nvidia is arguing that enterprise AI agents will require their own stack spanning models, runtime controls, governance, observability and compute infrastructure.

Whether enterprises embrace that approach remains to be seen. Additional security and control layers can introduce complexity, latency, and potential vendor dependencies. But Joshi said the market “appears to be converging toward a common architecture when it comes to scaling AI agents across control, security, runtime, and observability.”