In early March, engineers from Tencent set up stalls in the north plaza of their Shenzhen headquarters, offering free installations of the "Lobster" OpenClaw to users. The scene resembled a bustling market, with lines stretching endlessly as people brought NAS devices, MacBooks, and mini PCs, evoking memories of tech enthusiasts flashing Android systems a decade ago. In reality, numerous major tech firms are aggressively advancing their own "Lobster" initiatives. Xiaomi, for instance, has begun internal testing of MiclawAgent, aiming to embed AI agents into its "Human-Vehicle-Home Full Ecosystem" to transform phones, cars, TVs, and appliances into AI execution nodes. As cloud providers start deploying these "stalls" and device manufacturers integrate agents into operating systems, the "Lobster" storm signals the beginning of the large language model era's next phase. This is not merely a competition over AI tools but a covert battle for the next-generation "super gateway."

Currently, all players face a dilemma: the pure "Chat" model fails to sustain a viable business. Over the past two years, domestic cloud providers and tech giants have engaged in an arms race, funneling thousands of high-end computing cards into data centers. By 2026, ByteDance, Alibaba, and Tencent are projected to collectively exceed $60 billion in capital expenditures. However, idle compute power incurs steep depreciation costs daily. Relying solely on consumer-facing chat interactions cannot absorb this massive compute reserve or generate revenue from users accustomed to free services. Occasional tasks like drafting emails or creating images consume minimal tokens, insufficient to cover the depreciation and operational expenses of underlying compute clusters. To monetize these expensive resources and generate steady cash flow, giants urgently need a "token black hole" that continuously and automatically consumes compute power.

The emergence of locally deployed agents like OpenClaw fulfills this role. When users issue complex commands, OpenClaw decomposes tasks, searches the web, calls local software, identifies errors, and self-corrects. Each step sends requests to cloud-based APIs. A single complex task can consume hundreds or thousands of times more tokens than a basic conversation. An AI analyst noted, "Chinese open-source models are adopted by OpenClaw primarily for their cost-effectiveness. Compared to overseas rivals, lower costs encourage more frequent API calls, directly converting into cash flow for cloud providers and preventing waste of massive compute investments." This explains why Tencent and others are willing to invest manpower in offline "stalls" to help users deploy open-source agents, and why Alibaba aggressively promotes OpenClaw's cloud integration. Each deployment plants a "compute pump" that runs 24/7 on users' local or cloud machines. Regardless of whether open-source models are used upfront, as long as inference and tool-calling APIs route to their cloud services, countless micro-requests accumulate into substantial B2C and B2B cash flow. Under intense scrutiny from capital markets demanding commercial returns from large models, this agent-driven API revenue stream is a critical lifeline sustaining compute expansion.

Beyond cash flow, the second objective behind pushing local agents addresses a ceiling in large model development: the exhaustion of high-quality training data. In recent years, compute power and training data have been the core resources in large model competition. As capabilities improve, another resource gains importance: task trajectory data. Consensus holds that high-quality public text data online (e.g., Wikipedia, news, books) has been largely exhausted by existing models. Continuously feeding static text will only produce more knowledgeable "bookworms" rather than advancing toward true AGI. What next-generation models need is understanding how humans "take action" in the digital world—this is the coveted "trajectory data." When users task AI with completing an action, the AI undergoes steps from understanding needs to searching, tool invocation, form filling, and payment processing, each leaving a record that forms a complete task chain. For agent models, this data is more valuable than plain text as it reflects real-world action logic—precisely the data giants previously struggled to obtain, hidden within fragmented software, closed apps, and corporate intranets inaccessible even to search engines with vast crawler networks.

OpenClaw deployed on user devices and system-level agents like Miclaw act as "data probes" behind enemy lines. Alan Feng, OpenClaw China community manager, stated, "After installing OpenClaw, users often expect magical automation, but the real value lies in defining clear tasks. Trajectory data feedback allows continuous model optimization, enabling vendors to enhance agent capabilities." When users run agents locally to perform operations, the agents record every operational intent and software interaction. The intensive promotion of agent applications by domestic giants is essentially a distributed, unprecedented data crowdsourcing effort. Users may believe they are getting free AI labor, but in guiding and correcting agents, they are providing high-quality reinforcement learning fine-tuning data at no cost. Once these trajectory data points flow back to the cloud, they become core barriers for training next-generation agents with strong logical reasoning and execution abilities—akin to Tesla collecting real-road data from millions of electric vehicles to refine its FSD algorithms. An insider from Alibaba's Qwen project mentioned, "The probability of China leading in new paradigms is below 20%, but through agent trajectory data, Alibaba can rapidly iterate models to narrow the gap." Now, giants are turning users' computers and phones into "data collection vehicles" for the AI era. Those who amass the most trajectory data will be first to train truly "embodied" super models.

From this perspective, promoting local agents is not about new tools but about competing for the operational gateway in the AI era. China's internet has undergone several gateway wars: early portals vied for homepage traffic; the search era made Baidu the information gateway; the mobile internet era shifted gateways to apps, with WeChat, Alipay, and Douyin becoming traffic hubs. AI is restructuring this landscape. Alibaba's Qwen continues investing in "AI task execution," enabling users to place orders with a single command; Xiaomi is embedding Miclaw deep into its phone OS. These moves signal that future user-digital world interactions will be reconfigured. When users habitually express needs verbally, operational paths change—they no longer open specific apps but delegate tasks to AI, which decides platforms, services, and payment routes. In such a system, apps persist but become service nodes, while the true gateway is the agent assisting users. In this new context, "competing for app gateways" is outdated; the real war is about becoming the underlying agent that takes user commands and controls the ecosystem. If a giant's agent dominates user devices, it wields ultimate commercial power—intent distribution—easily directing orders to affiliated businesses or travel needs to its payment ecosystem. In this new "walled garden" built by agents, once-dominant super apps could be reduced to "pipes" providing underlying services, losing direct user engagement, brand premium, and traffic value. This sensitivity explains why giants are aggressively pursuing agents—all aspire to be the controlling platform.

OpenClaw's popularity may be just a signal. The real shift is AI evolving from a "talking tool" to an "acting system." For two years, the industry focused on boosting intelligence; now, companies increasingly ponder how to equip AI with action capabilities. Once AI reliably completes tasks, internet structures will transform. Many applications may recede to the background, with users interacting primarily with an agent for most digital life operations. In this world, agents form a new operational layer connecting users to all services. Reflecting on tech history, platform-level shifts often begin inconspicuously: Android was initially a geek flashing system, WeChat Official Accounts started as simple content tools, and Mini Programs resembled lightweight web pages. Yet these products later became new platforms. If AI enters an agent era, OpenClaw may be among the first names remembered. China's internet may be on the cusp of this storm's eve.