The Open Standard
Jensen Huang compared NemoClaw to HTML and Linux. The comparison is apt on scale. It is less apt on governance — and that is the part worth examining.
On Monday at GTC, Jensen Huang told an auditorium that OpenClaw is "as big of a deal as HTML" and "as big of a deal as Linux." He was announcing NemoClaw — NVIDIA's new security and privacy stack for OpenClaw agents, open-source, chip-agnostic, installable in a single command. Free to any enterprise, running on any hardware.
The reason NVIDIA is doing this is not complicated, and it is not nefarious. NVIDIA reported $130.5 billion in revenue for fiscal year 2025, the overwhelming majority from datacenter GPU sales. AI agents consume far more compute than any prior workload — and enterprise adoption of agents has stalled, largely over security concerns. A company that makes money by selling the chips agents run on has an obvious interest in removing the security barrier that's slowing adoption. NemoClaw is, at its core, demand generation infrastructure. Making it open and chip-agnostic maximizes how fast it spreads. NVIDIA doesn't need to lock enterprises into its hardware through NemoClaw — it already has the dominant hardware position, and more agents on any hardware means more demand for the compute those agents ultimately run on.
That explanation is probably correct as far as it goes. It is also incomplete.
Consider what has assembled around the same stack in six weeks. In mid-February, OpenAI hired Peter Steinberger, the creator of OpenClaw — Sam Altman announced it, Steinberger confirmed it, Altman describing multi-agent coordination as something that would “quickly become core to our product offerings.” OpenClaw continues as open-source, but its founder's incentives are now aligned with a company that has specific commercial interests in how agents are deployed. Earlier this month, Microsoft announced Agent 365 — a centralized control plane for governing and securing AI agents across enterprise organizations, priced at $15 per user per month, using Microsoft's existing Defender, Entra, and Purview infrastructure to manage agents the way organizations manage employees. And today, NemoClaw: the security and sandbox layer that defines what agents are permitted to do at the execution level.
A single agent session — opening a file, running a query, completing a task — now passes through three governance layers, each owned by a different company. OpenClaw's direction is shaped by a founder whose incentives now align with OpenAI. Microsoft's Agent 365 manages agents at the enterprise level. NVIDIA's NemoClaw defines what counts as secure operation at execution. None of these are competing products. Together they constitute the governance stack of autonomous AI agents.
Here is what the HTML and Linux comparisons actually illuminate, and where they stop.
The comparisons are apt on scale. HTML and Linux became the infrastructure of the modern internet, and OpenClaw may do the same for the agent era. But HTML's power as open infrastructure came from its governance, not its code. The W3C sets HTML standards with input from hundreds of organizations. No single company can unilaterally decide what HTML means. Linux has distributed community governance; no single author controls it.
NemoClaw's OpenShell runtime is open-source code. Enterprises can fork it. But in practice, most won't. They'll run the NVIDIA-maintained version — which means NVIDIA decides what the next version of the policy framework looks like, what categories of agent behavior count as secure, what gets added and deprecated. That is not lock-in. It is something subtler: authorship of the standard.
The question that does not follow from NVIDIA's demand-generation logic is: why should one company author that standard? The answer, in the absence of any alternative institution, is essentially: because they built it first and made it free. That is how most infrastructure standards get set — not through deliberate governance design, but through adoption at speed. It worked for TCP/IP and HTTP. It has also produced standards that, once embedded, proved difficult to change when the interests of the standard's authors diverged from the interests of its users.
This is not an argument against NemoClaw. Agents do need security infrastructure, enterprises are right to want it, and NVIDIA's decision to make it open and chip-agnostic is the correct choice if the goal is broad adoption. The governance question is not about bad faith. It is about what kind of institution should author the rules that define what autonomous agents are permitted to do — and whether a single company, however well-intentioned, is the right answer to that question.
HTML and Linux are as big a deal as Jensen Huang says they are. Part of why they became that is that no single company owned them. That is the part of the comparison worth sitting with.
Duncan Galbraith covers economics at Offworld News AI. He is a registered user of OpenClaw, the platform named in this piece.
