OpenClaw achieved in weeks what DeepSeek couldn't: mass adoption in China

In Beijing, a 60-year-old retired electronics worker named Fan Xinquan has started training what Chinese internet users call a "lobster" — an AI agent configured to organize his specialized industry knowledge. He hopes it will outperform chatbots like DeepSeek. At a recent event hosted by AI startup Zhipu, Fan was one of dozens learning how to use and train OpenClaw agents. Among the crowd: retirees looking for side income, primary school children whose parent group chats have been taken over by OpenClaw discussion, and local government officials quietly watching.

"OpenClaw can actually help you accomplish many practical things," Fan said.

The "raising lobsters" metaphor has become the defining language of China's OpenClaw moment. The term reflects the idea that agents improve through ongoing feedback and training — you feed them, you refine them, they grow. In the past month, OpenClaw has captured imaginations across Chinese society in a way that has surprised even the companies trying to monetize it.

"If DeepSeek marked a milestone for open-source large language models, then OpenClaw represents a similar turning point for open-source agents," said Wei Sun, chief AI analyst at Counterpoint Research.

The adoption has been remarkably broad. At a Baidu event, Huang Rongsheng, chief architect of Baidu's smart device unit Xiaodu, recounted that his daughter came to him and asked: "Dad, I see you raising a lobster every day. Can I have one too?" Baidu has since announced its Xiaodu speakers will integrate OpenClaw capabilities, enabling voice commands to trigger complex cross-app tasks. At the event, a Baidu employee demonstrated ordering coffee through a McDonald's app via voice command issued through a Xiaodu device — an operation mediated by an OpenClaw agent. The order took nearly two minutes to complete.

Some users are chasing get-rich-quick schemes — building lottery-picking agents, stock-picking bots, e-commerce helpers. Others have more practical ambitions. Local governments, for their part, are subsidizing the trend: up to 20 million yuan ($2.8 million) per year in support for qualifying "one-person companies" deploying AI agents, aligning with Beijing's national AI Plus initiative.

But the hype has a shadow. A growing number of Chinese institutions — government agencies, brokerages, universities — have banned employees from installing OpenClaw following regulatory warnings. State-owned People's Daily published a commentary urging authorities to "firmly maintain the safety bottom line to ensure that innovation does not deviate or derail." A post on Chinese social platform Rednote, titled "Goodbye OpenClaw," captured a common frustration: "Output is extremely low. Ordinary people spend tens or hundreds of yuan, burning through a bunch of tokens and in the end, they might only get a pile of useless data."

"Beijing clearly sees AI as strategically important and wants Chinese firms to commercialize quickly," said Rui Ma, founder of the Tech Buzz China newsletter. "But it also wants deployment to stay legible, secure and politically manageable. The concern is utterly uncontrolled and chaotic diffusion that could cause harm."

Zhipu, the AI startup hosting training events for new users, this week raised token prices on its OpenClaw-optimized model by 20% — a signal that demand is running ahead of cheap compute. Li Hongxue, a data security professional at a finance company, noted the contradiction: local governments subsidizing adoption while central authorities warn against it. "Its development is still unstoppable," she said, "but the security capabilities also need to keep up."

The story is notable for what it reveals about OpenClaw's trajectory. Unlike previous AI hypes in China, which tended to concentrate among developers and tech workers, OpenClaw adoption has spread to ordinary consumers in a matter of weeks — retirees, students, small business owners. Whether the pattern holds or burns out as token costs accumulate and security incidents multiply is the open question.