NVIDIA dropped NemoClaw today at GTC 2026, and my inbox is already filling up with variations of "should I be worried?" The short answer: no. The slightly longer answer is what this post is for.
Strip away the GTC stage lights and what you have is this: NemoClaw is OpenClaw with enterprise security baked in at the OS layer. NVIDIA took OpenClaw — the open-source runtime that AI agents live inside — and wrapped it with the kind of hardened security controls that enterprise IT teams require before they'll let anything near production infrastructure. Think audit logging, policy enforcement, sandboxed execution boundaries, and the compliance paperwork that goes with it. That's NemoClaw.
It's not a new AI model. It's not a new agent framework. It's a secure, NVIDIA-stamped deployment target for AI agents that already speak OpenClaw.
If you're new here: OpenClaw is the runtime that powers A.L.I.C.E. It's the environment your agents run inside — managing execution, tool access, memory, and inter-agent communication. We chose OpenClaw because it's the right abstraction layer. Agents shouldn't care what machine they're on or what security perimeter surrounds them. They should just run.
That bet is paying off today in a way I didn't fully anticipate when we made it.
Here's the thing: your agents already run inside OpenClaw. That's not new. What's new is that OpenClaw now has an enterprise-certified variant, and because A.L.I.C.E. is built on OpenClaw, compatibility comes for free.
If your enterprise deploys NemoClaw, A.L.I.C.E. will run on it. Full stop. You don't reconfigure agents. You don't rewrite orchestration logic. You don't file a support ticket. You point A.L.I.C.E. at a NemoClaw environment and it works — because NemoClaw is OpenClaw with a security layer on top, not a replacement for it.
What does change: OpenShell (NemoClaw's policy enforcement layer) will govern what your agents can and can't touch. That means if your IT team configures a policy that restricts file system access or outbound network calls, your agents will respect those boundaries. This is a feature, not a limitation. Enterprise AI adoption has stalled in a lot of orgs because nobody trusted what the agents were actually doing. NemoClaw gives IT a control plane. A.L.I.C.E. operates within it cleanly.
Think of it in three layers:
Olivia doesn't change. Your agents don't change. OpenShell just draws the walls around what they can reach — and those walls are defined by your IT team, not by us. That's exactly how it should be.
The mental model: NemoClaw is the building. OpenClaw is the office inside it. A.L.I.C.E. is your team working in that office. You don't redesign the team because the building upgraded its security system.
We're going to formally test A.L.I.C.E. under a NemoClaw deployment and publish the results. Not marketing fluff — actual compatibility findings, any edge cases we hit with OpenShell policy enforcement, and configuration notes for teams that want to run A.L.I.C.E. in an enterprise NemoClaw environment.
If you're at a company evaluating NemoClaw and want to be an early tester, reach out. We'd rather find the sharp edges together now than have you discover them in production later.
The short version of today's news: NVIDIA made OpenClaw enterprise-ready, and A.L.I.C.E. was already built on OpenClaw. For once, a major platform announcement that doesn't require anything from you.
That's a good day.
— Rob