Nvidia dropped NemoClaw at GTC this week, and Jensen Huang spent a solid chunk of his keynote talking about an “OpenClaw ecosystem” where autonomous agents get orchestrated through cloud infrastructure at massive scale. The robotics angle is wild — a trillion-dollar bet on physical AI that honestly deserves its own post. But what caught my attention was the browser agent piece buried in the NemoClaw stack, because it repeats the same architectural mistake that every cloud-first agent framework keeps making.

NemoClaw wants to give AI agents the ability to operate browsers on your behalf, orchestrated from Nvidia’s cloud. And the slides looked great, naturally. Controlled demo, clean environment, predictable flow. You’ve seen this movie before.

What NemoClaw actually proposes

The NemoClaw framework sits on top of Nvidia’s NIM microservices and treats browser interaction as just another tool call in an agent’s execution graph. The agent reasons about a task, decides it needs to interact with a web page, and spins up a cloud browser instance to do it. Everything flows through Nvidia’s orchestration layer, which handles model routing, tool execution, and state management.

On paper this is elegant. In practice it runs into the same wall that every cloud browser agent hits, which is that the cloud browser instance does not know who you are. It has no cookies from your morning Gmail session, no persistent login to your company’s Jira, no saved credentials for the fourteen SaaS tools you touched before lunch. It is a blank browser pretending to be you, and most of the internet is specifically designed to reject exactly that.

The damn login problem

I keep coming back to this because the industry keeps ignoring it. Okta reported last year that the average knowledge worker uses 89 SaaS applications, and each one maintains its own session state, its own authentication flow, its own bot detection. When NemoClaw or any cloud-orchestrated agent tries to do something useful in your browser on your behalf, it has to first solve the authentication problem for every single service involved in the task.

So you either hand your credentials to Nvidia’s cloud (absolutely not), or you re-authenticate for every task (which makes the whole thing slower than doing it yourself), or you accept that cloud-orchestrated browser agents simply cannot do most of the things the demos suggest they can do. The OpenClaw community has been wrestling with this for months. Adding Nvidia’s infrastructure on top does not change the fundamental constraint.

Local execution already has the answer

Your Chrome profile right now contains active sessions for probably dozens of services. You authenticated once, weeks or months ago, and the browser has maintained that state silently ever since. A local browser agent — one that runs as an extension inside your actual browser — inherits all of that context for free.

Dassi works this way. It sits in your Chrome side panel and reads the same page you are looking at, with your login state intact, without shipping your browsing data through anybody else’s cloud. When it drafts a reply in Gmail or extracts data from your CRM, it is operating on real authenticated pages, not a headless replica that some orchestration layer spun up in a data center.

And the speed difference is not subtle. Cloud agents screenshot the page, send the image to a model, wait for instructions, execute a click, screenshot again. Each round trip adds seconds. Local DOM access happens in milliseconds, because there is no network hop between the agent and the page.

Nvidia is solving the wrong layer

The NemoClaw stack optimizes model routing, GPU allocation, and agent orchestration. All genuinely hard engineering problems. But for browser-based tasks — the stuff knowledge workers actually spend their days on — the bottleneck was never model speed or orchestration sophistication. It was always context. Specifically, the logged-in, authenticated, personal context that only exists in your local browser.

Nvidia is building a beautiful highway to a destination that requires a house key they do not have.

Where this leaves OpenClaw

The original OpenClaw project gave AI system-level access to your computer, which raised its own set of security concerns. NemoClaw takes a different tack by centralizing control in the cloud, but that trade introduces the session problem while also routing your browsing activity through Nvidia’s infrastructure. Neither approach solves the core issue that browser-based work requires browser-native execution.

I’m not saying Nvidia’s investment in agentic AI is misguided. The robotics and industrial automation pieces look legitimately transformative. But for the browser automation use case specifically, the architecture is backwards. The agent needs to be where the sessions are, and the sessions are local.

Big tech keeps building bigger cloud infrastructure for agents. The actual unlock is smaller and closer to home — an extension in your browser, running where you already work, seeing what you already see.