OpenClaw in China: why the buzz around the AI agent became a global market test

China is the first to turn OpenClaw from a developer toy into a mass test of whether AI agents are ready for real life. As users line up for installations, business scales cloud and token sales, while regulators try to prevent the experiment from becoming a problem.

Why China Specifically

OpenClaw is not a chatbot, but an agent that can be entrusted with actions: reading files, working with email and messengers, running tasks, and accessing external services. In China, this idea landed in an environment where new AI tools are quickly adopted not only by developers, but also by ordinary users. At offline installations organized by major tech companies, students, office workers, housewives, and retirees came. For many, OpenClaw became their first attempt to get not just a 'smart answer,' but a digital executor.

The Chinese platforms themselves gave additional momentum. Major cloud providers put OpenClaw in the spotlight, offered quick deployment, and linked it to local models that are significantly cheaper than Western counterparts. This dramatically lowered the barrier to entry: users no longer need deep infrastructure knowledge to launch an agent for news, social media, commerce, or everyday tasks. And when such services start to integrate into familiar products like corporate messengers, the barrier drops even further.

Who Makes Money on This

The buzz around OpenClaw is useful not only to users themselves. For Chinese AI companies, it is a way to shift interest in artificial intelligence from 'chatting with a bot' mode to a mode of continuous computation consumption. An agent runs longer, calls models more frequently, and requires more cloud resources than a regular chat. So each new installation is not a one-time demonstration, but potentially a long revenue stream for cloud, models, and related services.

• Cloud providers sell servers and ready-made installation templates
• Model developers get more tokens for long multi-step tasks
• Messengers and platforms embed agents into familiar scenarios
• Startups get demand for new skills, interfaces, and vertical products
• Local authorities view agent services as part of the AI economy

This is evident from the scale of consumption. According to Chinese authorities, the average daily number of token calls in the country by March 2026 had grown to 140 trillion compared to 100 billion in early 2024. Such growth cannot be explained by chatbots alone. Agent scenarios, where a model performs chains of actions in the background, are far more resource-hungry — and this is exactly what the entire ecosystem is trying to profit from now. So OpenClaw in China became not only a user trend but also a commercial stress-test for the entire AI supply chain.

Where Problems Begin

Mass adoption revealed the flip side. For OpenClaw to actually perform tasks, it is usually given extended rights: access to local files, APIs, external services, and plugins. This makes the product useful, but simultaneously increases the attack surface. China's cyber incident response center listed four key risks in March: prompt injection, erroneous agent actions, malicious skills, and vulnerabilities in the software itself. And this is no longer an abstract list for a report, but a set of quite practical scenarios.

In practice, this means very concrete problems. An agent can misunderstand an instruction and delete important emails or documents. A malicious actor can hide a malicious command in a web page that the agent reads as part of its work. Finally, a user can install an insecure plugin and give an attacker keys, files, or access to their machine. For a non-technical audience, there is another risk: OpenClaw promises magic but often requires fine-tuning. As a result, some people encounter not an 'autonomous assistant,' but endless configuration, strange glitches, and unexpected token bills.

Authorities responded quickly. Starting March 11, 2026, government agencies, state companies, and major banks in China began restricting OpenClaw installations on work computers. Later, regulators issued separate recommendations: isolate the agent in a container or virtual machine, do not run it with administrator rights, do not expose the service directly to the internet, and do not store sensitive data inside it. For authorities, this signals that agent systems have left the sandbox and reached infrastructure with real risks.

What This Means

The OpenClaw story in China is not just a local trend, but an early stress-test for the entire AI agent industry. If mass adoption shows that such systems can be made useful, cheap, and manageable, other markets will quickly replicate this scenario. If, however, users tire of errors, leaks, and bills, the backlash could slow the entire agent boom far beyond China.