China is whipping up an OpenClaw storm.

March, beneath the Tencent headquarters in Shenzhen, Tencent engineers gathered on the North Plaza like a marketplace, setting up booths to install “Lobster” OpenClaw for users free of charge. The queue stretched endlessly—some carried NAS devices, others brought MacBooks, and some lugged mini PCs—evoking the geek meetups from a decade ago when Android systems were being flashed. In reality, many tech giants are intensively advancing their own “Lobster” projects. Xiaomi has begun internal testing of MiclawAgent, hoping to embed AI agents into its “all-ecosystem” system—enabling phones, cars, TVs, and appliances to become AI execution nodes. Cloud vendors are now “setting up booths” as well. As terminal manufacturers begin embedding Agents into their operating systems, the “Lobster” storm signals the opening of the second half of the large model era. **This isn’t a simple battle of AI tools, but a covert war over the next-generation “super entry point.”** --- ## Token-based Cash Flow Currently, all players face a dilemma: The pure “Chat” model cannot generate a healthy business model. Over the past two years, domestic cloud vendors and tech giants have been trapped in a long-term arms race, with tens of thousands of high-end computing cards being systematically pulled into data centers. By 2026, ByteDance, Alibaba, and Tencent’s combined capex will exceed $60 billion. But if users don’t engage, computing power is wasted, incurring hefty depreciation daily. The reality is, relying solely on C-end dialog modes neither consumes the massive compute reserves nor generates income from users accustomed to free services. A user might occasionally have AI write an email or draw a picture, but such single, low-token interactions cannot offset the depreciation and operating costs of their enormous compute clusters. To utilize expensive computing power and generate real cash flow, giants urgently need a “token black hole” that will continuously and automatically consume compute resources. **Locally-deployed Agents like OpenClaw fill this role.** When users issue complex instructions, OpenClaw breaks down tasks, searches online, calls local software, detects errors, and self-corrects and retries. Each step sends requests to cloud API interfaces. A single complex task can consume hundreds or thousands of times more tokens than a regular conversation. An AI analyst told Wallstreetcn: “Chinese open-source models are adopted by OpenClaw mainly due to cost-effectiveness. Compared to overseas competitors, lower costs allow more frequent API calls, which directly convert to cloud vendors’ cash flows, avoiding huge compute waste.” This is why cloud vendors like Tencent are willing to subsidize staff to help users deploy open-source Agents onsite, and Alibaba pushes for seamless OpenClaw cloud integration. Each deployment is effectively planting a 24/7 “compute pump” in users’ local or cloud PCs. **Regardless of whether the model running front-end is open-source, as long as the inference and tool-calling APIs point to their own cloud service, a flood of micro-requests will eventually converge into significant B2C and B2B cash flows. Under the capital market’s scrutiny for large model monetization, these API flows propelled by Agents are giants’ key lifelines driving compute expansion.** --- ## Mining Trajectory Data Beyond the first layer of cash flow, the second layer for giants pushing local Agents touches the ceiling of large model development: the depletion of high-quality training data. **For the past few years, the core resources in the large model race have been compute and training data. But as models grow more capable, another resource is becoming increasingly crucial: task trajectory data.** The consensus: Most high-quality, publicly available online text data (Wikipedia, news articles, books and papers) have already been “consumed” by large models. Feeding only static texts will make models mere bookworms—unable to reach true AGI. **What will next-gen large models need? They’ll need to know how humans “take action” in the digital world. That is the highly sought “trajectory data.”** When users ask AI to finish a task, the AI goes through a series of steps—from understanding requirements to searching, using tools, filling forms, making payments—each action leaves a trace. Together, they form a complete task chain. **For Agent models, this data is more valuable than plain text because it reflects real-world action logic.** This was previously the hardest data for giants to acquire, buried in countless fragmented software, closed apps, and enterprise internal networks—even search engine crawlers can’t reach it. But OpenClaw deployed on user terminals, and system-level miclaw, are “data probes” behind enemy lines. Alan Feng, OpenClaw China Community Manager, remarked: “Users install OpenClaw expecting magical automation, but the true value lies in defining clear tasks. Trajectory data feedback lets the model continuously optimize, letting vendors improve agent capability.” When users run Agents locally, letting them execute actions, the Agent records every intent and software interaction trajectory. The dense rollout of Agent apps by domestic giants is essentially a distributed, unprecedented-scale data crowdsourcing exercise. Users think they’ve gotten a free AI laborer; in reality, as they guide and correct Agents, they’re freely providing the giants with the highest-quality reinforcement learning fine-tuning data. **Once “trajectory data” flows back to the cloud, it forms the key moat for training next-gen agent models with strong logical reasoning and execution. It’s like Tesla once collected real-world driving data from millions of cars to power its FSD algorithms.** An insider from Alibaba’s Qwen project told Wallstreetcn: “China’s chance of leading a new paradigm is below 20%, but with agent trajectory data, Alibaba can iterate models fast and close the gap.” Now, giants are turning users’ PCs and phones into “data collection vehicles” for the AI era. Whoever controls the largest trajectory dataset can first train truly “hands-on” super models. From this perspective, local Agent promotion isn’t just about a new tool—it’s still about capturing the OS-level entry point of the AI era. --- **Entry Warfare Cycle Returns** China’s internet has seen several classic entry wars: early portals fought for homepage traffic; in the search era, Baidu became the info gateway; then, in mobile internet, Apps were the entry point, with WeChat, Alipay, and Douyin taking over traffic centers. But AI is changing this structure. **Alibaba Qwen continually invests in “AI Tasks,” letting users order with one phrase; Xiaomi internally tests miclaw, deeply integrating it in the phone’s OS. These moves signal that future user-digital world interfaces will be restructured.** When users grow accustomed to expressing needs in a phrase, operating paths change. Users won’t actively open an App but delegate tasks to AI. AI decides which platform, which service, and which payment chain to use. Thus, Apps’ status will change—they’ll still exist, but more as service nodes. The true entry point is the Agent that completes tasks for users. In this new context, “grabbing App entry” is outdated. The real war is becoming the “underlying agent” directly obeying the user, controlling the whole scene. If a giant can make their Agent dominate user terminals, they hold the highest power in commerce—the distribution of intent. They can easily direct food orders to their affiliates or funnel travel needs into their payment ecosystem. **In these new Agent-walled gardens, former super Apps will devolve into mere “pipes” providing service interfaces, losing direct customer dialogue, brand premium, and traffic premium.** This is why giants are so sensitive about Agents. Everyone wants to be the platform that controls Agents. --- ## **The Eve of the Storm** OpenClaw’s explosion may be just a signal. The real change is that AI is moving from “talking tool” to “acting system.” For the past two years, the large model industry’s core goal was boosting intelligence; now, more companies ponder—how does AI gain the ability to act? Once AI can reliably finish tasks, the internet’s structure will shift. Many apps may retreat to the background; users will only need one Agent to manage most of their digital life. In this world, the Agent is a new OS layer, connecting users and all services. Looking back, each major platform shift starts inconspicuously. Android began as a geek’s flashable OS, WeChat official accounts were just simple content tools, and mini-programs launched as almost lightweight webpages. But all these later became new platforms. If AI does enter the Agent era, then today’s OpenClaw might be one of the earliest names remembered. China’s internet may be witnessing the eve of this storm. --- Risk Warning and Disclaimer The market involves risks; investments require caution. This article does not constitute personal investment advice and does not take into account individual users’ special investment goals, financial situations, or needs. Users should consider whether any opinions, viewpoints, or conclusions in the article suit their particular circumstances. Investment decisions made accordingly are at one’s own risk.