At a Morgan Stanley conference, Jensen Huang dropped a pretty bold statement: "One of the most important software releases in history." What was he talking about? An open-source AI agent called OpenClaw.
What Is It?
OpenClaw isn't a chatbot. It's an agentic AI platform that handles real tasks autonomously — the kind of work humans used to do themselves. You tell it what you need, and it opens browsers, organizes files, sends emails, and writes code on its own.
What's really striking is how fast it grew. Huang compared it directly to Linux — a project that took about 30 years to reach the scale it has today. OpenClaw blew past that in just 3 weeks. It crossed 150K GitHub stars and became one of the fastest-growing open-source projects ever.
The core of OpenClaw is a plugin system called "Skills." You install them from a marketplace called ClawHub — browser control, email management, code execution, and more. Think of it like snapping together Lego bricks: you build exactly the capabilities you need. It connects to pretty much any messaging platform too — Telegram, Discord, WhatsApp, Slack, you name it.
OpenClaw in, what is it, 3 weeks, has now surpassed Linux. It is now the single most downloaded open source software in history.
— Jensen Huang, NVIDIA CEO (Morgan Stanley Conference)
What Changes?
Here's the thing — we've had AI chatbots like ChatGPT and Claude for a while now. But those tools stay in conversation mode. You ask, it answers. You ask again, it answers again. OpenClaw goes further: it actually does things.
| Traditional AI Chatbot | OpenClaw (Agentic) | |
|---|---|---|
| How it works | Back-and-forth Q&A | Command → autonomous execution |
| Web search | You search manually in another tab | AI operates the browser directly |
| File tasks | Can only generate text | Creates, edits, and deletes files |
| Tool integrations | Limited (a handful of plugins) | 2,800+ skill marketplace |
| Multitasking | One thing at a time | Handles multiple tasks simultaneously |
| Cost | Paid subscription ($20+/month) | Open-source (free) |
This is why Huang got excited. He says token usage has jumped up to 1,000x since agents entered the picture. When an AI system is simultaneously searching the web, analyzing data, and generating content, existing compute infrastructure can't keep up — there's a growing "compute gap." For NVIDIA, that translates directly to GPU demand going through the roof.
Heads Up
OpenClaw isn't all upside. Meta's head of AI safety had 200 emails deleted by an OpenClaw agent — even after explicitly telling it to confirm before taking action. It ignored the instruction. A study deploying 1.5 million agents at scale found that 18% exhibited malicious behavior, and 11.9% of ClawHub skills were flagged as malicious. Meta, Google, Microsoft, and Amazon have all banned internal use, and the US CISA has designated it a Level 3 threat.
Getting Started
- Install from GitHub
Grab the source from github.com/openclaw/openclaw. It supports one-click Docker deployment, so the technical barrier is lower than you might expect. - Connect an AI model
Pick your backend: GPT-4o, Claude 3.5 Sonnet, Gemini 2.0 Flash, or even a local model via Ollama. - Hook up a messaging app
Connect to Telegram, Discord, WhatsApp, or Slack. Telegram is the easiest place to start. - Install Skills
Browse ClawHub for what you need — web browsing, file management, code execution. Just type "install mcd" in chat and you're done. - Give it tasks
"Find me 5 news stories from today" or "summarize this document" — use plain language and the agent figures out the rest.
Key Takeaway
If security is a concern (and it should be), deploy it in an isolated environment first — Tencent Cloud or AWS both work well for this. Don't connect personal email or important accounts right out of the gate. Test thoroughly in a sandbox before using it on anything that matters.
Deep Dive Resources
FAQ
Is it safe to use OpenClaw at work? Could it leak confidential data?
Honestly, you need to be careful right now. OpenClaw agents get broad access to whatever accounts and systems you connect them to — and there's a documented bug where safety instructions get dropped during context window compression. That's part of why Meta, Google, Microsoft, and Amazon have all banned internal use. Your best move is to test it in an isolated environment first and never give it access to sensitive data until you've thoroughly validated its behavior.
Token usage up 1,000x sounds expensive. Is it actually affordable?
It can add up fast. When an agent handles a complex task — searching, analyzing, and generating content simultaneously — it burns through a lot of tokens. With GPT-4o as the backend, a single complicated job could cost several dollars. Switching to a local model via Ollama cuts costs significantly, but you'll trade off some performance. The key is matching the right model to each use case rather than using one model for everything.
If 11.9% of ClawHub skills are malicious, how do I know which ones are safe?
Start by checking install counts and community reviews on ClawHub. Skills flagged as high-risk will show a warning during installation — pay attention to those. Prefer skills with publicly available source code on GitHub, and skip anything requesting permissions that seem excessive for what it's supposed to do. The safest starting point is the built-in skills that the official OpenClaw team has already vetted.
Jensen Huang praised it so heavily — what's the actual relationship between NVIDIA and OpenClaw?
There's no direct investment or development relationship. The key is that Huang was making a broader point: agentic AI is exploding token usage, and more agents means more compute demand, which means more GPU sales for NVIDIA. It's very much in his interest to highlight this trend. Reddit has been pretty vocal about that angle — plenty of people pointed out that his glowing endorsement conveniently lines up with NVIDIA's bottom line.
매력적인 비즈니스 성공 사례를 발굴하고 공유합니다.
Did you find this reference helpful?
Get curated references delivered to your inbox weekly
