The Most Important Software Release in History
Nvidia CEO Jensen Huang said something at Morgan Stanley's Technology, Media & Telecom conference that set the tech world instantly ablaze: he called OpenClaw "probably the most important software release ever" — full stop, no qualifiers. This isn't some hyperbolic headline from an obscure blogger. It's a direct quote from the CEO of the world's most valuable company.
The Lobster Arrives: AI Goes From Chat Tool to Working Agent
I covered the lobster, OpenClaw, in an earlier column. Quick recap: it's an AI agent that "lives inside your computer," running continuously in the background — it can manage your inbox, schedule your calendar, run code, track project progress. It started life as Clawdbot, got renamed Moltbot, and is now OpenClaw. The name keeps changing.
Its most distinctive feature is a local-first design: the model runs directly on your own machine, so you don't have to send your data to the cloud, and in theory you don't have to keep paying API fees (though in practice, if you want the lobster robot to be genuinely smart, you'll still burn through money fast). If you've got a reasonably powerful RTX GPU, you already have the infrastructure to run OpenClaw. I've had it installed for a while now, and honestly, I'm hooked — I can't seem to walk away from the lobster robot.
Historically, "agentic" development was something only engineers could really feel in their bones. But the lobster robot has made the whole concept of an agent tangible. The security risk is still real, but the fact that it handles your inbox — reading, writing, replying to email — and chats with you over messaging apps to talk through problems, means ordinary people can now feel it too, very concretely. Under the hood, the lobster robot still runs on a large language model, so essentially every major LLM is benefiting right now: cost-conscious users run open source, users who want the best AI run one of the major labs' models, and either way, every user's token consumption is exploding. Its creator, Peter Steenberger, built OpenClaw on Claude, which makes Claude arguably the single biggest beneficiary so far — the user experience and the lobster robot turn out to be an exceptionally good fit.
Huang compared it to Linux, and honestly, the analogy is dead on. Linux is the infrastructure of our entire computing era, and it took 30 years to reach today's level of ubiquity. OpenClaw got there in three weeks. Three weeks! If you look at the GitHub star growth curve, the line basically goes straight up — it looks almost like it's merging with the y-axis, the kind of chart you'd assume was printed wrong.
But what I actually found most interesting was Huang's read on compute demand. He said a typical generative-AI query produces one response, and that's roughly how much token consumption it costs. Move up to the agent layer, and each task burns through roughly 1,000 times as many tokens. And an always-on background agent like OpenClaw? A million times as many.
Huang's point is unambiguous: everyone's been saying AI compute demand will keep growing, but "how fast" now looks like it was badly underestimated. "Our company's compute demand just went vertical," he said. For Nvidia, that's obviously the best possible news — no matter how aggressively you build out data centers, demand is always going to be the thing running ahead.
Some people might say: of course you should discount what Huang says — he sells GPUs, so the more people run AI agents, the happier he is. Calling OpenClaw the most important software release in history is basically free advertising for him. I think that's a fair objection, but it's also jumping to conclusions a bit fast.
Linux taking 30 years and OpenClaw doing the equivalent in three weeks isn't something Huang made up — it's GitHub's actual numbers. Agents consuming 1,000 times the tokens of ordinary AI isn't marketing either — it's a fact dictated by the technical architecture, and if you understand how agents actually work, that number is entirely reasonable. Huang is describing something genuinely happening. Over the past three weeks I've fully felt this with the agent robot myself — check the bill, and you really are burning through token counts in the thousands multiplied over, and getting all of humanity to tens of millions of AI queries suddenly isn't hard to imagine at all.
What I actually want to talk about is the paradigm shift OpenClaw represents. We used to describe AI as: you ask, it answers. A prompt was a question-and-answer exchange, conversational in nature. An agent is a different thing entirely — you give it a goal, and it goes and does it. "The old prompt was 'what is,' 'when did,' 'who is' — query-type questions. The new prompt is 'go create,' 'go finish,' 'go write this' — action verbs." That shift isn't just technical. It redefines the entire relationship between user and AI. You're no longer using it to look things up. You're using it to get things done.
Agents Rewrite the Compute Ceiling — Nvidia's Moat Gets Even Deeper
The investment implication is this: if agents really do penetrate as deeply as Huang describes, the ceiling on compute demand needs to be recalculated — and there may not really be a ceiling at all right now. Nvidia's moat is deeper than anyone previously assumed. At the application layer, meanwhile, the companies that can genuinely integrate agents into enterprise workflows haven't even begun to have their value recognized by the market. For now, I think this story is still very early, and it's worth watching closely.
That said, having something running in the background 24 hours a day doing work for you does feel a bit like hiring a full-time assistant who never sleeps and never eats — one who lives inside your computer. It still takes some getting used to and some exploring. But I think this is going to become normal faster than we expect — fast enough that we'll soon forget what life was like before AI agents, the same way it's already hard to imagine a time before smartphones. So I don't think the first act of the AI era has even wrapped up yet. We're maybe ten minutes into the show.