At GTC 2026, Jensen Huang did something interesting. He didn’t just announce new hardware or a model release — he held up a hobbyist open-source project built by an Austrian developer and compared it to Linux and Kubernetes. That project is OpenClaw. And NVIDIA’s enterprise answer to it is NemoClaw, announced at the keynote and built to take what Peter Steinberger created in his spare time and make it safe enough for a Fortune 500 to actually deploy. If you’re trying to understand where enterprise agentic AI is heading in 2026, this is the thread to pull.
What OpenClaw Actually Is (And Why NVIDIA Cares)
If you missed the OpenClaw story, here’s the short version. Peter Steinberger — the same person behind the popular iOS PDF framework PSPDFKit — published a project in November 2025 called Clawdbot. It was a local AI agent framework that let you run agents on your own machine and control them through messaging apps like WhatsApp, Telegram, Signal, Discord, and iMessage. MIT license, bring your own API key, works with Claude, GPT, DeepSeek, Gemini, or local models through Ollama.
It went viral. Sixty thousand GitHub stars in the first 72 hours. By March 2026 it had passed 250,000 stars — the most-starred software project in GitHub’s history. Along the way it got caught in a trademark dispute with Anthropic (the original name, Clawdbot, lasted until January 27, 2026 when it briefly became Moltbot before Steinberger landed on OpenClaw three days later). The red lobster mascot stuck. So did the tagline: “The lobster way.”
The core mechanic that made it click is the Skills system. Skills are just directories with a SKILL.md file that tells the agent what to do and how to do it. There are over 100 built-in skills and a community registry called ClawHub that grew fast. Matt Schlicht launched a companion app called Moltbook — essentially a social network for AI agents built on top of the platform. Steinberger announced on February 14, 2026 that he’d be joining OpenAI and moving the project to an open-source foundation.
Here’s the thing though: OpenClaw has real security problems. CVE-2026-25253 carries a CVSS score of 8.8. Researchers found over 30,000 exposed instances. The skills registry was compromised. Steinberger himself said it plainly: “It’s a free, open source hobby project that requires careful configuration to be secure.” That’s not a knock on the project — it’s honest. But it’s exactly why enterprises can’t just deploy it and call it a day. Which is exactly the gap NemoClaw is designed to fill.
NemoClaw: What NVIDIA Actually Built
NemoClaw is NVIDIA’s enterprise-grade AI agent platform, and it is built directly on top of OpenClaw. Not a fork, not an alternative — literally on top of it. Jensen Huang’s framing at GTC was deliberately echoing the trajectory of foundational infrastructure: “OpenClaw gave us exactly what it needed at exactly the right time… like Linux, like Kubernetes.” The implication is clear. OpenClaw is the raw infrastructure layer. NemoClaw is what you build on top of that for production.
What NemoClaw adds to the base OpenClaw stack:
- Enterprise security — addressing the exact vulnerability profile that made raw OpenClaw unsuitable for regulated industries
- Privacy guardrails — policy enforcement to control what data agents can access, process, or transmit
- Policy enforcement — organizational governance controls that a solo developer’s hobby project simply doesn’t have
- Integration with NVIDIA NeMo — the broader AI agent software suite that NVIDIA has been building
One detail worth flagging: NemoClaw is hardware agnostic. It doesn’t require NVIDIA GPUs. That’s a deliberate choice, and a smart one — it means NemoClaw competes on the software and governance layer, not by locking you into a hardware stack. NVIDIA obviously wants to sell you H300s, but they’re not making NemoClaw the mechanism for that lock-in.
This sits alongside two other open-source releases from GTC worth knowing about. NVIDIA OpenShell is an open-source runtime specifically for “self-evolving agents and claws” — with safety and security baked into the runtime itself. NVIDIA Agent Toolkit is a set of open-source models and software for building enterprise agents. Together, NemoClaw plus OpenShell plus Agent Toolkit form a layered stack that takes the OpenClaw ecosystem from hobbyist to enterprise-grade.
Nemotron 3 Super: The Model Underneath It All
No enterprise agent platform announcement lands without a capable model to run on it. NVIDIA announced Nemotron 3 Super on March 11, 2026, and the numbers are worth looking at directly rather than summarizing vaguely.
Nemotron 3 Super is a 120B total parameter model with 12B active parameters. It’s a hybrid Mixture of Experts architecture — meaning you get large-model capability without paying the compute cost of activating all 120B parameters on every inference pass. NVIDIA used two architectural choices that are genuinely novel here: LatentMoE routing and native NVFP4 pretraining.
The benchmark comparisons against GPT-OSS (OpenAI’s open-weight model at the same parameter class) are specific enough to be useful:
| Benchmark | Nemotron 3 Super | GPT-OSS-120B |
|---|---|---|
| SWE-Bench Verified | 60.47% | 41.90% |
| RULER at 1M tokens | 91.75% | 22.30% |
| Inference throughput | 2.2x higher | Baseline |
The RULER score at 1M context is the one that should stop you. 91.75% versus 22.30% on a long-context benchmark is not a marginal win — that’s a qualitatively different capability for enterprise agentic tasks that require working across massive codebases, lengthy documents, or complex multi-step workflows. The SWE-Bench number matters too. Coding capability is increasingly the proxy for general agent usefulness, and 60.47% is a meaningful lead at this parameter class.
It’s open weights with an open training recipe, which means enterprises can fine-tune it, audit it, and deploy it without being dependent on an API. For regulated industries, that’s not a nice-to-have — it’s a requirement.
The Broader GTC Stack: From Agents to Physical AI
NemoClaw didn’t land