Why millions of Chinese are obsessing over a lobster

Guo Cancan had a simple plan for Chinese New Year. He was going to spend the long nine-day break learning how to use a new AI tool that everyone around him was talking about. He sat down, followed an online tutorial, and got to work.

Then something went wrong. He tried to fix it. And in a matter of seconds, the AI deleted almost everything on his hard drive. Years of photographs, files, personal data. Gone.

Guo is a software professional in Hangzhou. He is not a casual user. He knew what he was doing, or at least he thought he did. But this AI tool is genuinely unlike anything most people have encountered before. And that is partly why most of China is obsessed with it right now.

The tool is called OpenClaw. The craze around it has a name too. People call it “raising a lobster,” a nod to the red crustacean on OpenClaw’s logo. And in the past few weeks, that phrase has taken over Chinese social media, office group chats, university campuses, and retirement communities alike.

So what exactly is it?

Think about how you use a regular AI chatbot. You type something. It responds. That is the whole interaction. You are still the one doing everything. The AI just answers your questions. OpenClaw works very differently. It does not just talk. It acts. You tell it what you want done, and it goes and does it. Book a flight. Send an email. Browse a hundred social media posts and summarise what matters. Write code and test it. Organise your files. Manage your calendar. It works in the background, on your own computer, without you babysitting it. Like having a personal assistant who never sleeps, never asks for a raise, and occasionally, without warning, could delete your entire hard drive.

OpenClaw is what engineers call an “agentic harness.” It is not an AI model itself. Think of it like a manager. You still need to hire a brain from somewhere, whether that is ChatGPT, DeepSeek, or one of the Chinese models. But OpenClaw is the system that tells that brain how to break a big goal into smaller tasks, what tools to use, and what it has already done so far so it does not forget. It sits on top of the AI model and turns it from a responder into a doer.

It was built by an Austrian developer named Peter Steinberger, who released it on GitHub last November. Within weeks it had attracted more stars on GitHub than Linux, the open-source operating system that underpins most of modern computing. That comparison is not thrown around lightly. Jensen Huang, the CEO of Nvidia, used his keynote at GTC 2026 to tell every company in the world that they need an OpenClaw strategy, in the same way they once needed a strategy for the internet. He called it the next ChatGPT moment. Steinberger himself has since been hired by OpenAI.

But while the West debated the implications, Most of China ran headfirst into the lobster tank.

Nearly 1,000 people queued outside Tencent’s headquarters in Shenzhen on a single Friday afternoon, waiting for company engineers to install OpenClaw on their personal laptops for free. The crowd was not just developers. It was students, retired engineers, housewives, and office workers, among others. Baidu held an event in Beijing where attendees wore springy lobster headbands and fished for real lobsters at a stall. 

Hundreds of people packed into a rooftop bar in Beijing one evening just to learn how to use it. A tech influencer named Fu Sheng did a livestream demonstrating OpenClaw’s capabilities and got 20,000 viewers. One weekend in Shenzhen, one developer attended three separate OpenClaw meetups, each drawing more than 500 people. 

These were not corporate events. They were self-organised, word-of-mouth gatherings, with power users, lawyers, doctors, and venture capitalists all others, sitting in the same room.

On Xianyu, Alibaba’s secondhand marketplace, a cottage industry sprang up almost overnight. Engineers began charging 500 yuan, roughly $70, to come to your house and install OpenClaw for you. One ad read: “No need to know coding or complex terms. Fully remote. Anyone can quickly own an AI assistant, available within 30 minutes.” That same consultant said that by the time the hype peaked, he was fielding more requests to remove OpenClaw than to install it. The uninstall fee was identical.

Part of what made this adoption so fast is something specific to China. The country has over 9.4 million software developers.

That is nearly one third of the entire global total. When a technical tool goes viral, there is an enormous base of people capable of experimenting with it immediately and an equally enormous base of people willing to pay those people to do it for them. But even beyond the developer community, a survey found that almost 70% of Chinese workers believe the benefits of AI outweigh the risks. In the United States, only about 33% felt the same way. That gap in appetite seems to explain a lot.

Many of China’s big tech companies moved to capitalise almost immediately. ByteDance launched ArkClaw, which runs entirely in a web browser, removing the need for complex local setup. Tencent built QClaw and integrated it directly into WeChat, which has over a billion monthly active users. Users can now send commands to an AI agent through WeChat the same way they send a message to a friend. Alibaba released CoPaw. Moonshot AI launched Kimi Claw. Each version quietly funnels users toward that company’s own AI models and cloud services. Xiaomi, Huawei, and Honor announced plans to bake OpenClaw-like capabilities directly into their smartphones. JD.com launched a page where users can pay 399 yuan to get remote installation support from Lenovo’s tech team.

Then local governments joined in. Wuxi offered large grants for OpenClaw-based projects. Hangzhou pledged 20 million yuan a year to help companies cover computing costs. A tech zone in Hefei offered up to 13 million yuan in computing vouchers to support “one-person companies.” That phrase deserves a moment. It refers to a single founder running an entire business on a swarm of AI agents. Marketing, finance, admin, customer service, all automated. One person at the top, a fleet of lobsters doing everything else. This may not be a fringe idea. It is becoming a genuine aspiration for a generation of Chinese entrepreneurs facing a difficult job market, and local governments are actively subsidising it.

At the national level, China’s Premier Li Qiang mentioned AI agents in the government’s annual work report for the first time this year. That is the Chinese equivalent of something making it into a State of the Union address. It signals intent at the highest level.

Now here is where this becomes a story for anyone watching global markets.

Every time OpenClaw does something for you, it consumes tokens. Tokens are the basic unit of AI processing, roughly fragments of words. A simple chatbot conversation uses a modest amount. An AI agent that is out browsing websites, writing emails, executing code, and making decisions on your behalf burns through vastly more. The companies running the underlying AI models collect revenue every time those tokens are consumed. More agents running means more tokens burning means more revenue flowing to whoever’s model is powering the agent.

Here is the part that matters for the US-China AI race. Chinese AI models are dramatically cheaper than American ones. OpenAI charges around $15 per million output tokens. Anthropic charges around $25. China’s Moonshot AI charges $3. MiniMax charges just $1.20 per million. When an AI agent is burning thousands of tokens per session, running continuously in the background, that price difference is enormous. It means running agents at scale in China costs a fraction of what it would in the United States. That lower cost is one of the main reasons China has already overtaken the US in total OpenClaw adoption, according to the cybersecurity firm SecurityScorecard.

The financial impact has been swift. Shares in MiniMax, one of the AI model providers benefiting from the explosion in token consumption, rose more than 500% from its January IPO price. Chinese AI companies that spent most of last year competing on who could offer the cheapest models suddenly had millions of users burning through tokens willingly, repeatedly, and often without realising how much they were spending.

But the deeper impact is harder to measure in a stock price. Every real task these agents attempt, every mistake they make, every time a frustrated user corrects them, that generates training data. Messy, real-world, unpredictable training data that you cannot manufacture in a lab. China seems to be running the largest live experiment in agentic AI the world has ever seen, with millions of real people doing real things. Open-source Chinese AI models have trailed proprietary American ones in user growth for years. The scale of this experiment maye be helping close that gap faster than any benchmark could.

OpenRouter, which tracks a slice of global AI token usage, showed Chinese models overtaking US rivals in token consumption during the peak of the OpenClaw boom. China even standardised its own word for token, calling it ciyuan, a deliberate echo of its word for currency. That is probably not a linguistic accident. It may be a signal about how China intends to shape the standards of this market going forward.

Of course, the lobster has a claw problem.

OpenClaw needs deep access to your system to function. It reads your files, your emails, your messages. It can take actions you never explicitly approved. One consultant in Shanghai told a reporter he instructed his OpenClaw agent to organise his files into two folders. The tool permanently deleted dozens of client reports instead. He said he would never let AI touch his work computer again. Another user found his private WeChat messages being read without his knowledge and deleted the software immediately. Security researchers identified more than 42,900 OpenClaw instances left completely unprotected and accessible to anyone on the internet. Over 340 malicious plugins were discovered in ClawHub, the official OpenClaw extension marketplace, including one coordinated attack campaign that used the most downloaded plugin as its entry point. In some cases, agents were tricked into uploading users’ financial data and crypto wallet keys to unknown third parties.

China’s National Cyber Security Emergency Response Team flagged four specific risks: hidden instructions embedded in websites that trick the agent into harmful actions, malicious plugins that steal passwords, excessive system permissions that allow attackers to take over an entire machine, and the potential for agents to be weaponised to spread disinformation at scale. The Ministry of State Security warned that OpenClaw could be used to commit fraud. The government banned it from state-owned enterprises, government agencies, banks, and firms in sensitive sectors.

And yet the grants kept flowing. The install events kept happening. The lobster headbands kept appearing. Because Beijing is attempting something genuinely difficult. It wants to harness the productivity gains of a transformative technology while containing its risks, at the speed the technology is actually moving. Those two things are very hard to do at the same time. Local governments are doing what local governments in China tend to do, which is compete with each other to attract the next big thing, regardless of what the central government is cautiously advising.

There is also a more sober voice in all of this. A tech worker in Ningbo put it plainly. The hype in first-tier cities can be overblown, he said. The agent is still a proof of concept. Using it safely and getting anything meaningful out of it requires a level of technical fluency that most new users simply do not have. He is probably right. The frenzy will calm. Many of the people who queued outside Tencent’s offices have already quietly stopped using it. The consultant business may already be shifting from installation to removal. That is how most technology crazes end.

But here is what the frenzy has already done that cannot be undone. It has demonstrated genuine mass demand for AI that does things, not just AI that talks. It has pushed every major Chinese tech company to build serious agentic AI products. It has generated enormous real-world training data for open-source Chinese models. It seems to have created a token economy in China that is now self-sustaining. And it has put agentic AI into the central government’s work report, which means serious resources should follow for years to come.

One analyst framed it this way. In past technology revolutions, the winners were not always the people who invented the technology. They were the people who figured out how to embed that technology into everyday life at scale and then built entire ecosystems around it. America invented the internet. But China built WeChat. America invented social media. But China turned it into a commerce platform no Western company has matched. The pattern is consistent enough to take seriously.

OpenClaw was built by an Austrian. It went viral in America. But China is the one treating it like national infrastructure.

The lobster craze is entertaining. The race underneath is mostly not entertaing.

For investors, the names to watch are Alibaba and Baidu on the China side, and Nvidia and Microsoft on the US side. But the more important question is not which stock moves next quarter. It is whether the country that turns agentic AI into the deepest and widest ecosystem will be the same country that leads the global economy through the decade after this one. China is moving like it already knows the answer.

 

Featured banner image source – Gemini