China's 'Lobster' Agents: How OpenClaw Became Beijing's Default AI Infrastructure When Western Models Were Blocked

What Happened

Chinese developers and entrepreneurs adopted OpenClaw at unprecedented scale in March 2026, building custom AI agents they nicknamed “lobsters” on top of the open-source codebase. The BBC published a detailed feature on April 5 documenting the phenomenon: because Western frontier models — ChatGPT and Claude — are inaccessible in China, OpenClaw’s open-source architecture became the vehicle for Chinese users to build their own agentic systems on top of local AI models.

China has already surpassed the United States in OpenClaw adoption, according to American cybersecurity firm SecurityScorecard’s tracking dashboard, as CNBC reported in March. Nvidia CEO Jensen Huang told CNBC’s Jim Cramer that OpenClaw is “definitely the next ChatGPT.”

The term “raise a lobster” became shorthand across Chinese social media for training a personal AI agent. Celebrity comedian Li Dan told millions of Douyin followers he talked to his lobster in his dreams. Cheetah Mobile CEO Fu Sheng posted relentlessly about raising his lobster. An IT engineer interviewed by the BBC claimed his lobster could process 200 TikTok Shop product listings in two minutes — he previously managed about twelve per day.

The Government Push

The adoption wave did not happen organically. Chinese local governments actively subsidized OpenClaw projects, turning an open-source tool into national productivity infrastructure.

In Shenzhen, the AI and robotics agency of Longgang district proposed subsidies of up to 2 million yuan (US$290,000) for approved OpenClaw projects, according to the South China Morning Post. In Wuxi, eastern Jiangsu province, the local hi-tech district offered between 1 million and 5 million yuan for industrial applications like quality inspection and equipment maintenance. The BBC reported some government incentive packages tied to OpenClaw reached up to 10 million yuan, though that figure appears to include combined subsidies and financing rather than pure grants.

“The government is pushing, making a direction. And so that is why the big enterprises like Tencent, Alibaba have the motivation to build OpenClaw better for normal people,” Huang Dongxu, co-founder of software provider PingCAP, told CNBC.

Tech giants matched the government’s energy. Tencent organized public installation sessions at its Shenzhen headquarters where nearly a thousand people — retirees, students, entrepreneurs — lined up to get OpenClaw set up for free by company engineers, per SCMP. Baidu hosted similar events in Beijing, drawing hundreds. TechRadar reported that the queues became a cultural moment in themselves.

Many government subsidies specifically mentioned “one-person companies” — startups run by a single individual with AI handling the rest. “Human employees need rest, but OpenClaw can run 24/7,” one new user, Wang Xiaoyan, told CNBC. With youth unemployment above 16%, the one-person company model offered Beijing a policy answer that aligned AI promotion with job creation — or at least job substitution.

“Some say that in 2026, if you don’t ‘raise lobsters’, you’ve already lost at the starting line,” state newspaper People’s Daily published in a commentary, as quoted by the BBC.

The Pullback

The cycle followed a pattern familiar in China’s top-down system: local governments race to promote what Beijing signals it wants, then pull back when problems surface.

China’s National Vulnerability Database (NVDB), a cybersecurity center under the Ministry of Industry and Information Technology, warned that improper configuration of OpenClaw would expose users to “high-level security risks,” including cyberattacks and privacy leaks, according to SCMP. Government agencies began banning staff from installing the tool. The trend shifted from offering installation services to offering removal services.

“It’s disorder with control,” Rui Ma, founder of the Tech Buzz China newsletter, told the BBC. She added that Beijing’s intervention does not necessarily signal discouragement — local governments compete for central approval, rush in, then recalibrate as challenges emerge.

“Everyone in China knows that the government sets the pace, and the government tells you where the opportunities are,” Ma said. “It’s practical for most people. That’s probably a better plan, to just follow the government directive than to really try to figure it out on your own.”

What This Means for Agent Builders

Strip away the geography and this is an open-source infrastructure story.

When governments block frontier models, open-source agent frameworks fill the gap. ChatGPT is blocked in China. Claude is blocked in China. But OpenClaw’s codebase is open, and Chinese developers forked it, built on top of it, and connected it to local Chinese AI models like those from Baidu, Tencent, and the survivors of what Chinese media call the “Hundred Model War” — more than 100 AI models emerged since 2023, with only about 10 still viable, according to the BBC.

This creates a structural dynamic that matters for every builder betting on the agentic future:

The model layer is replaceable. The agent infrastructure layer is the moat. Chinese users proved that OpenClaw agents work just as well on local Chinese models as they do on Claude or GPT-4. The “lobster” that processes 200 TikTok Shop listings does not care which LLM sits underneath — it cares about the orchestration layer, the tool integrations, the memory system. OpenClaw provides those. The model is interchangeable.

Government adoption cycles are fast but volatile. Shenzhen was subsidizing OpenClaw installations on a Saturday. By the following week, the NVDB was warning users about security risks. Builders who depend on a single government’s regulatory posture — in any country — are building on sand. The open-source nature of OpenClaw means that even if Beijing restricts official use, the codebase is already distributed. As one MERICS think-tank researcher told the BBC, the enthusiasm that turned OpenClaw into something trendy was “uniquely Chinese,” but the underlying dynamic — open-source infrastructure routing around geopolitical restrictions — is universal.

The one-person company model is a leading indicator. China’s explicit promotion of “one-person companies” powered by AI agents is not a Chinese-only phenomenon. It is what happens when agentic infrastructure becomes cheap and accessible. The IT engineer doing 200 listings in two minutes was not running a tech company. He was running a TikTok gadgets shop. The implications for small-business automation globally are the same: agents that can handle marketing, finance, and administrative work collapse the cost structure of running a business to near zero.

Tom van Dillen, managing partner at consultancy group Greenkern, told CNBC: “China is turning an open-source tool into national productivity infrastructure at a speed no other country is matching.”

The Broader Geopolitical Arc

China’s OpenClaw frenzy arrives in the same week that two other geopolitical AI stories are playing out. The Trump administration is at the Ninth Circuit trying to enforce a Pentagon blacklisting of Anthropic for refusing military use cases. The UK government, meanwhile, is courting Anthropic with London expansion incentives and dual-listing proposals.

Three different national strategies. Three different answers to the same question: who controls the agent infrastructure that will run the next decade’s economy?

The United States is restricting: punishing companies that impose ethical limits on military use. The United Kingdom is courting: offering incentives to attract the company the US is pushing away. China is building its own: taking open-source infrastructure and turning it into a national project.

For builders, the takeaway is pragmatic. The agentic future does not require Claude or GPT-4 if the infrastructure layer is open. China just proved it with millions of “lobsters.”