China’s ‘Raise a Lobster’ Movement Is Rewriting the Global AI Agent Race — And Zhipu Just Weaponized It
Long lines of ordinary citizens queuing at tech company offices to have AI agents installed on their laptops. Government subsidies flowing to individual developers building on an open-source framework. A publicly listed company releasing a frontier model trained entirely on domestic chips — then giving it away for free.
This is OpenClaw in China, April 2026. And according to a Forbes analysis by Tirias Research, it may be producing the most powerful competitive moat in the global AI race.
The Lobster Economy
“Raising a lobster” (养龙虾) — the Chinese internet’s term for installing and training an OpenClaw agent — has become a genuine cultural phenomenon. The term captures something the West’s developer-centric adoption missed: this is a consumer movement.
Tencent and other tech giants are hosting mass installation events drawing thousands. Shenzhen’s Longgang district is pouring millions of yuan into individuals and startups building on the framework. The ecosystem is self-reinforcing at a speed that makes Silicon Valley’s AI adoption look measured by comparison.
Zhipu’s Masterstroke: GLM-5 and Turbo
At the center of this is Zhipu AI (Z.ai), a Tsinghua University spinoff and China’s first publicly listed 100% AI company. The sequence of moves:
- Feb 11: Released GLM-5, a 744-billion parameter mixture-of-experts model under an MIT license — completely free to use, modify, and deploy
- Mar 16: Released GLM-5-Turbo, a proprietary API-only variant fine-tuned specifically for “OpenClaw scenarios”
- Price: ~$1.20 per million input tokens — a fraction of Western competitors
The benchmark numbers are striking. GLM-5 hits 77.8% on SWE-bench, 86.0% on GPQA-Diamond, and 75.9% on BrowseComp — competitive with Claude on agentic benchmarks. And it was trained entirely on Huawei Ascend chips, not Nvidia hardware.
The Data Flywheel No One Can Replicate
Here’s the strategic insight Forbes highlights: Zhipu isn’t just selling a model — it’s capturing the world’s largest real-world agent interaction dataset.
Millions of daily OpenClaw interactions in China generate high-quality data on:
- Real-world tool calls and their outcomes
- Task decomposition patterns
- Failure modes and recovery strategies
- Multi-step workflow execution
Zhipu’s Turbo is deeply integrated with OpenClaw’s “Packages” system, positioning it to capture the lion’s share of this data. The company promises to feed learnings back into future open-source GLM releases — creating a flywheel: better agents → wider adoption → richer data → better models.
As Forbes puts it: “Western players like Anthropic and OpenAI field powerful models, but they have nothing comparable in scale to the real-world framework-specific data and iterative feedback loop to which Zhipu has access.”
The Paradox Continues
This is the same China that banned OpenClaw in government agencies and SOEs in March. The contradiction is deliberate — Beijing is simultaneously:
- Restricting uncontrolled agent deployment in sensitive sectors
- Subsidizing grassroots development at the individual and startup level
- Weaponizing the adoption data for domestic AI model improvement
It’s not a contradiction — it’s a strategy. Control the ceiling, accelerate the floor.
What This Means for OpenClaw
Several implications:
For the project
OpenClaw now has 65,000+ forks on GitHub and has matured rapidly. The transition to foundation governance and additions like Nvidia’s NemoClaw security stack have hardened it significantly since early February. Chinese adoption is the strongest validator of the framework’s scalability.
For Western competitors
The “raise a lobster” data flywheel is a competitive threat that can’t be replicated by building better models alone. If Zhipu’s data advantage compounds over 6-12 months, the performance gap on agentic tasks could widen in China’s favor — despite inferior base hardware.
For security
Mass consumer adoption means millions of agents running with varying security configurations. CertiK issued an emergency warning this week that OpenClaw’s growth is outpacing its safety features, identifying pathways for malicious extensions to hijack local execution environments.
The Bigger Picture
The age of AI agents is arriving, and the largest competitive advantage may not be the most powerful model — it may be the largest and most active community putting agents to work.
Right now, China is doing that at a scale the rest of the world may struggle to match. The question is whether the security infrastructure can keep pace with the enthusiasm.
As Forbes concludes: “The Chinese are raising lobsters at a scale the rest of the world may struggle to match.”
Sources: Forbes/Tirias Research, CertiK Warning, OpenSourceForU