A familiar cycle is repeating itself in China’s AI industry.
Last year it was DeepSeek-R1: chip makers scrambled to support it, local governments raced to follow suit, and all-in-one appliances became the hottest business model. This year, the frenzy has shifted to OpenClaw — while DeepSeek-4 remains in development. Cloud providers and foundation model companies are competing to deploy solutions on the ground, local governments are launching subsidy contests, and new AI hardware is on the horizon.
OpenClaw bills itself as “AI that actually gets things done” — an open-source, self-hosted agent framework that individuals and enterprises can use to build agents, interact with them through everyday applications like social media, and have them autonomously complete complex, multi-step tasks on their own computers or servers. Since the “lobster” was brought into the world by Peter Seinberger late last year, it has been hailed as the embryonic form of a truly usable agent. Tech enthusiasts see it as pointing toward AGI, and it has spread virally.
Last week, Tencent set up a street-side booth offering free on-the-spot cloud installation of OpenClaw for users. At the same time, Shenzhen’s Longgang District announced subsidies for OpenClaw deployment and applications. This string of moves rapidly ignited market sentiment. People have already taken to calling OpenClaw simply “the lobster.” When a technology earns a Chinese nickname, it has usually already completed its localization.
No major internet company wants to miss this wave. ByteDance livestreamed deployment tutorials and usage tips, Feishu launched an official plugin, and the free-tier API call quota was increased a hundredfold to one million calls. A fresh subsidy war has erupted. In fact, as early as the Spring Festival, Alibaba Cloud, Tencent Cloud, ByteDance’s Volcengine, China Telecom’s Tianyi Cloud, JD Cloud, and other domestic cloud providers had already begun offering tailored cloud services and deployment guides with customized features.
The major internet companies are also rapidly localizing OpenClaw and packaging it as a native platform capability. ByteDance launched ArkClaw, Tencent Cloud rolled out WorkBuddy, Alibaba countered with CoPaw, and Xiaomi began closed beta testing of MiClaw.
Local governments are equally willing to invest real money. Wuxi High-Tech Zone also plans to introduce policies that have been jokingly dubbed the “Lobster Twelve Articles” — one-upping Longgang’s “Lobster Ten Articles” in sheer number. The subsidies are pointedly targeted: if OpenClaw can empower manufacturing enterprises, a single project can receive up to 5 million RMB in rewards.
This frenzy is no accident. AI companies believe technology can reshape virtually every industry. Hardware makers, compute infrastructure providers, and software companies believe value will ultimately settle on their platforms. Local governments expect it to bring new jobs and tax revenue. This pattern of industrial diffusion has become the most characteristic growth pathway in China’s AI industry.
Commercialization efforts around OpenClaw have thus emerged rapidly. Dedicated local hardware, cloud servers (VPS), and model-provider-hosted products represent different deployment approaches — each a reflection of the broader ecosystem’s scramble to capture the OpenClaw dividend. This overflow of demand is already showing up in model providers’ revenue. Since this year’s Spring Festival, Kimi’s revenue has surpassed its entire previous year’s total, and MiniMax’s token consumption has topped the OpenRouter leaderboard. At least for now, the business model of selling tokens on the back of the trend is proven.
Last year, DeepSeek’s open-source models already previewed this dynamic. Regional compute centers raced to deploy, cloud providers rushed to list container images overnight, and domestic GPU manufacturers announced compatibility one after another. More than a year later, people are still waiting for DeepSeek’s next version. Now, the same anticipation has transferred to OpenClaw. People are staying up late waiting for new feature releases.
Everything is accelerating. Last year, open-source models like DeepSeek solved the problem of frontier model capabilities being prohibitively expensive. This year, open-source applications like OpenClaw are beginning to lower the barriers for AI to enter real industries. DeepSeek proved that frontier AI capabilities can spread rapidly through open source; OpenClaw is proving that model capabilities can be turned into application capabilities that scale quickly.
Agent capability is not the same as model capability. Reusable Skills represent a new dimension of scaling. The competitive focus in 2026 is not just about who has the bigger model, but about whose agent system is more engineered — packaging the model layer and the execution layer (the agent harness) into Skills that can be deployed, invoked, and even embedded into various business systems.
Venture capitalist Zhu Xiaohu of GSR Ventures has said that what struck him about OpenClaw was not the product itself, but the speed of the ecosystem’s growth: “Look at this past month — hundreds of thousands of new Skills have been created worldwide.” It reminded him of the early days of the personal internet era. He believes the opportunity for entrepreneurs today is to “first ride the lobster’s traffic wave, then see if you can build your own ecosystem.”
From Chat to Code to Claw, the center of gravity for AI applications is shifting from “conversation” to “execution.” The claw reaches further, and the value embedded within it grows. The market has long been waiting for a true killer application — or at least a scenario capable of incubating one. Coding is one such scenario; the personal agent is another direction that is widely anticipated.
Claude Code is pushing the boundaries of white-collar work automation, while OpenClaw’s rapid breakout stems partly from FOMO — the fear of being replaced by AI — and partly because people can see in its trajectory the early shape of a personal agent. Usable AI agents that can actually work are becoming real, becoming tangible. Since last year, OpenAI had been reverting to internet-era competitive logic, fighting for existing market share in advertising and e-commerce. It was not until OpenClaw appeared this year that things changed: Sam Altman swiftly brought Seinberger into the fold and had GPT-5.4 support computer use for the first time.
Yet coding tools are themselves a high-value market, while the future of personal agents remains unclear. An agent’s value ultimately depends on the value of the market it serves. Many scenarios that currently look like “a hammer looking for a nail” — especially consumer-facing applications — are not enough to sustain a stable business model. Users consume a limited number of tokens, tasks still require human follow-up after completion, and value density is not high.
With deployment and usage still requiring a certain learning curve, the personal agent is unlikely to become a true mass consumer market in the short term. But that does not mean it won’t produce important products. The early market of the PC era was assembling computers; the early market of mobile internet was downloading value-added services. Today’s OpenClaw ecosystem may be just a lively starting point — the real products are yet to emerge from the ferment.
On the other hand, the key to agents landing in high-value scenarios is not capability but permissions. Privacy, system access, and security liability are the fundamental challenges. OpenClaw’s explosive popularity has made this clearer than ever: once AI obtains system-level access, the boundary of a general-purpose agent is no longer cognitive ability but the boundary of responsibility. Meta’s super-intelligence team alignment lead exclaimed that OpenClaw nearly batch-deleted all of his emails.
What this tests is not the model, but whether institutions and organizations can adapt. Nearly every high-value service ultimately needs to be embedded within institutional and organizational structures, constantly answering questions about liability, compliance, and governance. This is also why a company like Palantir — willing to embed itself deep inside institutions and build complex systems for governments and large organizations — has been so highly valued by markets in the AI era.
Precisely for this reason, many Chinese entrepreneurs believe this is where their opportunity lies. Unlike Silicon Valley’s emphasis on universal products and standardized platforms, Chinese startups are long accustomed to building customized systems within complex organizational environments and taking on the “dirty work” of integration. Meanwhile, attempts to embed OpenClaw into consumer electronics and embodied intelligence devices are also emerging rapidly.
Over the past two years, open source has accelerated AI innovation and adoption. Some in Silicon Valley have praised China’s open-source models to the skies, while others view them with wariness and suspicion. Chinese entrepreneurs, for their part, have swiftly devoured America’s open-source OpenClaw whole. When it comes to open source, every layer of China’s AI technology stack seems a degree more fervent and earnest than its American counterpart — each one eager to “claw” out a little more.
