Nvidia CEO Jensen Huang stood in front of roughly 25,000 developers at the SAP Center in San Jose on March 16 and posed a question that would have been incomprehensible six months ago: “What’s your OpenClaw strategy?”

The answer, according to Nvidia, is NemoClaw — an open-source enterprise stack that wraps OpenClaw’s autonomous agent capabilities in sandboxing, privacy routing, and policy-based security guardrails. The platform was announced during Huang’s nearly three-hour GTC keynote and represents Nvidia’s most significant bet on the software orchestration layer since CUDA transformed its business model two decades ago.

“Mac and Windows are the operating systems for the personal computer. OpenClaw is the operating system for personal AI,” Huang said in Nvidia’s official announcement. “This is the moment the industry has been waiting for — the beginning of a new renaissance in software.”

What NemoClaw actually does

NemoClaw is a reference stack — not a fork of OpenClaw, but a layer on top of it. It uses Nvidia’s Agent Toolkit to install two core components with a single command: Nvidia’s Nemotron open model family and a new open-source runtime called OpenShell.

OpenShell is the centerpiece. It sandboxes OpenClaw agents, limiting their access to sensitive data and enforcing an organization’s security policies. Kari Briski, Nvidia’s VP of generative AI software for enterprise, described it during a Sunday pre-briefing as “the missing infrastructure layer beneath claws to give them the access they need to be productive, while enforcing policy-based security, network, and privacy guardrails,” according to ZDNet.

Nvidia built OpenShell in collaboration with CrowdStrike, Cisco, and Microsoft Security to ensure compatibility with existing enterprise cybersecurity tooling, per ZDNet’s reporting.

A privacy router sits between local and cloud models, allowing agents to use frontier models like Claude or GPT-5 through the cloud while keeping sensitive operations on local hardware. The platform supports any coding agent and any open-source model — Nvidia’s Nemotron is the default, but the architecture is model-agnostic.

NemoClaw is also hardware-agnostic. It runs on GeForce RTX PCs, RTX PRO workstations, DGX Station, and DGX Spark supercomputers, per Nvidia’s press release. It can also run on non-Nvidia hardware — a striking departure from the company’s historically proprietary CUDA ecosystem.

Why Nvidia is doing this now

The timing traces back to a security problem that has dogged OpenClaw since January.

Meta banned employees from using OpenClaw on work computers. A Meta AI safety researcher publicly described an incident where an OpenClaw agent mass-deleted her emails without instruction. Multiple enterprises have balked at deploying an agent framework that grants autonomous access to local files, email, and network resources with minimal guardrails.

Nvidia sees this gap as its entry point. “Claws are exciting but they’re risky too, because they could access sensitive data, misuse connected tools, or escalate privileges autonomously,” Briski told The Register. The pitch is straightforward: NemoClaw adds the enterprise-grade security layer that OpenClaw lacks, making autonomous agents deployable in environments where compliance and data governance actually matter.

The strategic logic runs deeper than security. Nvidia’s dominance in AI has been built on hardware — the GPUs that train and run every major model. But as WIRED reported on March 9 when it broke the NemoClaw story, leading AI labs are now building their own custom chips, threatening Nvidia’s hardware moat. Controlling the software orchestration layer — the tools enterprises use to deploy and manage AI agents — creates a second moat that’s less dependent on chip sales.

Huang framed this explicitly during the keynote. “AI now has to think. In order to think, it has to inference. AI now has to do. In order to do, it has to inference,” he said, per CNET’s live coverage. The argument: as AI shifts from training to inference, and from chatbots to autonomous agents, the compute demand multiplies. Nvidia wants to own both the chips generating that compute and the software layer directing it.

The partnership web

Before the official announcement, Nvidia had been pitching NemoClaw to enterprise software companies for weeks. WIRED’s March 9 report identified conversations with Salesforce, Cisco, Google, Adobe, and CrowdStrike. None of those companies confirmed official partnerships at the time, though CrowdStrike and Cisco’s involvement in building OpenShell has since been confirmed.

OpenClaw creator Peter Steinberger — who joined OpenAI last month after the acqui-hire — worked directly with Nvidia on NemoClaw. “With NVIDIA and the broader ecosystem, we’re building the claws and guardrails that let anyone create powerful, secure AI assistants,” Steinberger said in the press release.

This creates an unusual dynamic: Steinberger now works for OpenAI, which runs OpenClaw through a foundation. Nvidia is building enterprise tooling on top of that foundation. The relationship between OpenAI’s stewardship of OpenClaw and Nvidia’s commercial ambitions for NemoClaw will be worth watching as both products mature.

The Nemotron Coalition

Alongside NemoClaw, Nvidia announced the Nemotron Coalition, a multi-lab collaboration to advance open-source frontier AI models.

The founding members include Mira Murati’s Thinking Machines Lab, Perplexity, Cursor, Mistral AI, and Sarvam, among others. The first project: Mistral and Nvidia will co-develop an open model trained on Nvidia DGX Cloud, which will be open-sourced and serve as the foundation for Nvidia’s upcoming Nemotron 4 model family. Other coalition members will contribute data and testing, according to ZDNet.

The coalition positions Nvidia at the center of the open-source AI model ecosystem — a role Meta’s LLaMA initiative once monopolized. With Mistral, Perplexity, and Cursor as partners, Nvidia is assembling a coalition of companies that depend on open models and can collectively pool resources to compete with the closed-source labs.

From SaaS to agents-as-a-service

Huang’s most pointed claim during the keynote was that OpenClaw represents the same kind of paradigm shift as Linux, HTTP, and Kubernetes.

“Just as Linux gave the industry exactly what it needed at exactly the time, just as Kubernetes showed up at exactly the right time, just as HTML showed up — it made it possible for the entire industry to grab on to this open source stack and go do something with it,” Huang said, per TechCrunch.

Briski was more direct about what this means for the software industry: “Claws are the new application layer for AI, and they’re driving orders of magnitude more demand for compute,” she told The Register.

The implication: the SaaS model — where users interact with cloud-hosted applications through a browser — gets replaced by agents that operate locally, execute multi-step tasks autonomously, and interact with cloud services on the user’s behalf. If that transition happens at enterprise scale, the company controlling the agent orchestration layer captures enormous value.

Nvidia is positioning NemoClaw as that layer. OpenAI’s Frontier platform, launched in February, is the most direct competitor. Google, Microsoft, and Anthropic all have their own agent platforms at various stages of development. The race to become the default enterprise agent stack is now fully underway.

What’s missing

NemoClaw is currently in alpha. Nvidia’s developer documentation is candid about this: “Expect rough edges. We are building toward production-ready sandbox orchestration, but the starting point is getting your own environment up and running,” per TechCrunch.

Several questions remain unanswered. How does OpenShell’s sandboxing interact with OpenClaw’s core ability to access local files and system resources — the feature that makes it useful in the first place? What happens when an enterprise policy conflicts with an agent’s task execution? How granular are the privacy guardrails, and who defines them?

The hardware-agnostic claim also deserves scrutiny. While NemoClaw technically runs on non-Nvidia hardware, Nvidia’s Nemotron models and the full Agent Toolkit are optimized for Nvidia GPUs. Running NemoClaw on competitor silicon will likely mean degraded performance — a soft lock-in that mirrors the CUDA playbook even if the code is technically open-source.

And then there’s the competitive question. OpenAI, which now controls the OpenClaw foundation, is simultaneously building Frontier as its own enterprise agent platform. If OpenAI decides to prioritize Frontier over the open-source OpenClaw stack, NemoClaw’s foundation could shift beneath it. Nvidia is betting that OpenClaw’s open-source community is strong enough to sustain development independently — the same bet IBM made with Linux in 2001.

The bottom line

Nvidia’s $4.5 trillion market cap was built on selling the hardware underneath AI. NemoClaw is a bet that the next wave of value sits in the software layer on top — specifically, the security and orchestration infrastructure that makes AI agents safe enough for enterprises to deploy at scale.

GTC attendees can try NemoClaw at Nvidia’s build-a-claw event through March 19. Developers can access the Agent Toolkit and OpenShell at build.nvidia.com or download from GitHub to run locally.

Whether NemoClaw becomes the Linux of the agent era or another enterprise middleware layer remains to be seen. But with seven-figure partnership conversations, a Huang-level keynote slot, and a growing open-source coalition behind it, Nvidia is making the largest single investment any company has made in turning autonomous AI agents into enterprise infrastructure.