China is Devouring OpenClaw Raw
From Chat to Code, and now to Claw, a familiar cycle is repeating itself in China’s AI industry
The focus of AI applications is shifting from dialogue to execution, extending its reach and increasing the value attached to it. The market has been waiting for a true killer application, or at least a scenario capable of nurturing one. Coding was one, and personal intelligent agents are turning into the next. This is not to say that they are not a widely anticipated direction.
Last year, we had DeepSeek-R1.
Chipmakers rushed to adapt to it, local governments followed, and AI all-in-one hardware became the hottest deployment format. This year, the same momentum has shifted to OpenClaw. DeepSeek-4 remains under development, while cloud vendors and model companies pivot toward deployment, subsidies, and distribution. A new wave of AI hardware is already on the horizon.
Now we’ve got OpenClaw.
OpenClaw positions itself as AI that actually does things: an open-source, self-hosted agent framework capable of executing complex, multi-step tasks across local machines and servers. On its own. Complex multi-step tasks on any computer or server are no sweat. Just costly. Hatched by Peter Steinberger last year, it has been hailed as a prototype of a usable intelligent agent. With Tencent’s viral installation booth and Shenzhen Longgang District subsidies for deployment and application, the market has started sniffing around. And when a new technology has a Chinese nickname, it has pretty much completed its localization.
No major internet company wanted to miss this opportunity. ByteDance conducted a live broadcast to introduce its tip for using OpenClaw. Lark launched an official plugin, and the free version’s API call limit was increased a hundredfold to 1 million times. A new subsidy war has begun again.
This demand spillover is already showing up in model revenues. Since the Spring Festival, Kimi’s income has reportedly surpassed its entire revenue from last year, while MiniMax’s token consumption has topped OpenRouter.
At least for now, the model of bundling and reselling tokens alongside deployment has proven viable.
Chinese cloud vendors all offered deployment guides during the Lunar New Year holiday, all with targeted functions. Major internet companies are also rapidly localizing OpenClaw and packaging it into their own platforms. ByteDance launched ArkClaw, Tencent Cloud launched WorkBuddy, Alibaba used CoPaw as a counterpart, and Xiaomi began beta testing MiClaw.
Local governments are also willing to invest real money. Wuxi High-tech Zone also plans to introduce policies, jokingly referred to as the “Lobster Twelve Measures,” even surpassing Longgang’s “Lobster Ten Measures” in terms of quantity. The subsidy policies are highly targeted; if OpenClaw can empower manufacturing enterprises, a single project could receive a maximum reward of 5 million RMB.
Everything is accelerating.
The key to competition in 2026 will not be about who has the largest model, but rather whose intelligent agent system is more engineered, encapsulating the model layer and the execution layer into skills that can be deployed, invoked, and even embedded into various business systems.
Rapid ecosystem growth.
Zhu Xiaohu of GSR Ventures said that what impressed him about OpenClaw was not the product’s inherent strength, but the rapid growth of its ecosystem. “Look at this past month, hundreds of thousands of new skills have been added worldwide.” This reminded him of the rise of the personal internet. He believes that the opportunity for entrepreneurs now is to “first ride the wave of popularity, and then see if they can build their own ecosystem.”
Claude Code is pushing the boundaries of automation in white-collar jobs, while capable AI agents are beginning to be truly usable and tangible. However, coding tools themselves represent a high-value market, while the future of personal intelligent agents remains uncertain.
The value of an agent ultimately depends on the value of the market it serves. Many scenarios that seem like “finding a nail with a hammer” today, especially applications for ordinary users, are insufficient to support a stable business model. Users consume a limited number of tokens, and manual completion is still required after tasks are finished, resulting in low value density.
Real products will emerge.
Given the learning curve for deployment and use, personal intelligent agents are unlikely to become a truly large-scale consumer market in the short term. This doesn’t mean it won’t give rise to significant products. The early market in the PC era was assembling computers; the early market in the mobile internet era was downloading mobile value-added services. Today’s OpenClaw ecosystem may just be a lively starting point; real products will emerge from this excitement.
The key to deploying intelligent agents is permission.
Privacy, system access permissions, and security responsibility are its fundamental challenges. The explosive popularity of OpenClaw makes this clearer: once AI gains system-level permissions, the boundary of a general intelligent agent is no longer cognitive ability, but the boundary of responsibility.
It is the system and organization that is being tested, not the model itself. Almost all high-value services ultimately need to be embedded within systems and organizational structures, constantly responding to issues of responsibility, compliance, and governance. This is why software companies like Palantir, willing to delve into the inner workings of systems and build complex systems for governments and large organizations, are highly sought after in the AI era.
This is where many Chinese entrepreneurs see their advantage.
While Silicon Valley prioritizes general-purpose products and standardized platforms, Chinese startups are accustomed to embedding themselves inside complex organizational systems: handling the messy, unscalable work of integration. In the age of open source, this makes all the difference.
Faced with OpenClaw, a US-born innovation, Chinese players have moved quickly to absorb, adapt, and commercialize it. Across every layer of the AI stack, the approach is more aggressive, more pragmatic, and ultimately more extractive. Not just using open source, but squeezing value out of it.