Searching for shovel sellers under the desktop Agent dividend

```

The surprises in the AI circle keep coming.

Recently, a product named OpenClaw (formerly Clawdbot/Moltbot) has quickly gone viral across both domestic and international tech communities and social media.

As a deeply interactive agent that can run on your own computer, deeply access your computer system, files, applications, and chat records, users can instruct and collaborate with AI in the most natural chat interface.

In developer-shared use cases, this desktop agent can complete complex tasks such as comparing quotes from over a dozen car dealerships, automatically sending emails, tracking replies, and organizing price differences; it can also handle routine affairs such as bulk email unsubscription, insurance claim processing, flight booking and automatic check-in.

Importantly, it has long-term memory context, able to remember local projects, repetitive tasks, and personal preferences, and can even proactively send reports, reminders, or alerts without needing to be triggered, being dubbed by the industry as a "24/7 Jarvis on standby."

Startup founders, developers, and tech enthusiasts have all rushed in to try it, and overnight the "OpenClaw nanny-level deployment tutorial" became a traffic password on Xiaohongshu and Bilibili. Industry insiders bluntly said this is the ChatGPT moment for desktop agents.

With network effects and word-of-mouth spreading, as more people try to build their own "Jarvis," domestic model players and cloud vendors have quietly become the invisible winners behind desktop agents.

The "Shovel Sellers" of Jarvis

OpenClaw is not the first functional agent in the market, but it has once again ignited market enthusiasm after manus, Qianwen Assistant, and "Doubao Phone."

With the explosive popularity of OpenClaw, MacMini has also soared to become a "financial product"—many pioneers in the community claim to have bought large numbers of MacMinis to run OpenClaw. Google's AI product lead Logan Kilpatrick is among them.

The core design concept of this desktop agent is local operation, so deploying it on a standalone Macmini avoids mixing with the main work computer, maximizing privacy and system security.

However, as more people learned about the project, another voice quickly emerged: given its high-level permissions, OpenClaw is better suited to run in an environment isolated from your main computer, and easily deployable cloud servers soon became the "chosen" solution.

On January 28, Alibaba Cloud quickly launched a dedicated suite of cloud services, along with a detailed deployment tutorial.

Tencent Cloud’s lightweight application server Lighthouse also simultaneously released an OpenClaw application template, pre-configured with the required environment for running OpenClaw.

Later, JD Cloud, China Mobile Cloud, and UCloud successively joined the camp.

An AI application architect told Wallstreetcn that using a cloud server to try OpenClaw is a quicker and more cost-effective option, and since cloud servers naturally support 7*24 hour operation, it matches OpenClaw’s positioning well.

However, if you run "Jarvis" without a locally open-source model, you can’t escape the need to connect to a model API, and OpenClaw’s "money-burning" ability quickly reveals itself.

"OpenClaw’s appetite is too big. I intended to use hundreds of thousands of tokens for ten days or half a month, but in just half an hour, they were all used up." An independent developer from Shenzhen told Wallstreetcn. Recently, he used OpenClaw to clone a classic Snake game.

"At first, I thought it was amazing—OpenClaw wrote code, ran, and fixed bugs all on its own. I just watched it work like a boss. But when I saw the API bill, my smile disappeared," said the developer.

The traditional chatbot's "you ask, I answer" mode usually consumes only a few hundred tokens per interaction.

However, the agent model represented by OpenClaw is "autonomous looping." To fix a small rendering error, OpenClaw conducted over 40 self-dialogues and coding attempts in half an hour.

A model company executive pointed out that applications like OpenClaw rely heavily on two core capabilities: ultra-long context and cost-effective inference power. "Agents need memory—current mainstream practice is to store context in GPU memory, with each new question carrying along the previous questions and answers, so the agent’s input keeps growing."

Therefore, to run this super-strong "Jarvis," you need a high-performance, responsive, and affordable large model.

Under the recommendation of project author Peter Steinberger, domestic AI unicorn MiniMax’s M2.1 model, which excels at long texts and logical reasoning, gained significant traction.

Peter Steinberger said in an interview, "I’m currently able to run MiniMax M2.1 on it—I think it's the best model at the moment. Kimi has just launched, and I’ll probably try it later."

Moreover, for agents to think like humans, they can’t do without prompt orchestration tools like LangChain, which define the logic for AI tool usage; to remember thousands of file details and historical operations, vector databases like Pinecone or Weaviate become indispensable external "hippocampus."

Crucially, as AI gains permissions to delete files and modify the system, security becomes the #1 issue. So Docker containers and various security sandbox technologies become necessities, ensuring the AI won’t accidentally erase your system drive. These middlewares may not face users directly, but they’re the invisible backbone supporting stable agent operation.

Just like in the gold rush, regardless of the success or failure of desktop agent apps, the "shovel sellers" behind them pocket their profits first.

The Diffusion of Agent Dividends

OpenClaw's viral spread has made AI "working for humans" increasingly real. The emergence of such desktop agents has made the industry realize that future AI won’t be just another app, but a shadow butler reigning above apps.

Pushing this trend further, the software landscape will shift from the previous "battle of a thousand models" to a "battle of a thousand endpoints."

Currently, agent players are blossoming everywhere. In addition to Manus, which solves complex scenarios, and the viral open-source OpenClaw, products like Coze Workflow, Flowith, CherryStudio, MiniMax Agent, and Stepwise AI Desktop Companion are launching rapidly.

It's important to note that agents and models are a classic example of mutual achievement—Manus runs on a multi-model architecture including Claude and Qianwen, and building OpenClaw requires users to choose their own model. This demonstrates a principle: the basic capabilities of agents still depend on underlying large models.

As desktop agent interaction dividend explodes at the application layer, competition to some extent returns to the models themselves. Domestic and foreign foundation model startups—including OpenAI, China’s DeepSeek, and Kimi—have made Agent capability a major focus, by directly "internalizing" agent capability into the model.

This means that in the next six months to a year, more "Jarvis" will emerge both domestically and internationally.

On the other hand, as embedded system operators, giants like Apple, Android (Google), and Microsoft will not let system access be taken over by desktop robots.

Industry insiders expect Apple Intelligence and Microsoft Copilot will most likely evolve into comprehensive system-level agents.

After all, they have permissions unmatched by any third party: authorization for screen recording is hard to obtain, no need to simulate mouse clicks, and can directly call interfaces at the system kernel level.

Domestically, Huawei's HarmonyOS Next's "native intelligence," Doubao Assistant, and Alibaba Qianwen Assistant are doing the same.

This is an AI defensive war for system providers.

When built-in system assistants can not only chat but also help with takeout orders and sending red packets, OpenClaw-like agents will find penetration space in China’s phone and PC markets greatly reduced.

In short, whether model-focused players or hardware giants, all will engage in the battle for desktop agent dominance.

At the same time, as a terminal, the hardware market will usher in a dividend window period.

This time, it's Mac mini that's popularized by OpenClaw, but it’s not because of Apple hardware performance, but rather due to the geographical advantages of Mac’s system and memory architecture, MacOS’s convenience, and Apple's self-developed ARM SOC’s power efficiency.

But Mac mini is only the "best solution for now" and not the final answer.

Hardware vendors are sniffing out new opportunities. At present, swift-moving Huaqiangbei is entering the market with AI mini-hosts pre-installed with desktop agents.

These small boxes, similar to NUCs or Mac minis, run 7x24 hours and connect to your main machine via a local network.

Edge computing vendors are also preparing to "cut in."

Alibaba, Tencent, and China Mobile are launching "cloud computer boxes," essentially thin clients with computing power in the cloud. For users who only need lightweight agents, a cloud box for tens of yuan a month might provide a better experience than buying a Mac mini.

From this perspective, with Cowork and OpenClaw's viral growth, a big pie surrounding desktop agents is quickly taking shape.

The future competitive landscape is becoming clearer: at the software layer, third-party open-source agents will proliferate like bamboo shoots after rain; at the model layer, companies like minimax and Kimi will become the behind-the-scenes shovel sellers; finally, at the hardware layer, a batch of more cost-effective, large-memory mini hosts or cloud AI boxes designed specifically for AI agents are on the way.

A battle for desktop control, combining software and hardware, seems about to erupt.

Risk warning and disclaimerThe market has risks; investment needs caution. This article does not constitute personal investment advice nor does it take into account individual users' specific investment goals, financial situation, or needs. Users should consider whether any opinions, views, or conclusions in this article suit their particular circumstances. Investment based on this is at your own risk.

```