A monthly salary of 20,000 yuan can't afford to feed lobsters? Mobile phone manufacturers are trying to break the cost deadlock by offering "free computing power with phone purchase".

Author: Deep Web Tencent News

Just as OpenClaw became a top trend in the AI ​​field due to the craze for "shrimp farming" and the controversy surrounding "shrimp killing," leading mobile phone manufacturers, who have been deeply involved in edge AI, couldn't hold back and began to deploy and "tame" their own Claw.

On March 6th, Xiaomi's mobile agent, Xiaomi miclaw, officially launched a small-scale closed beta test via invitation codes, becoming the first domestic mobile phone manufacturer to internally test "Claw." Following this, Huawei, Honor, OPPO, and others announced their own internal beta testing of Claw.

Among them, Huawei officially announced the addition of the OpenClaw mode to Xiaoyi and then launched the Xiaoyi Claw Beta version; Honor announced the launch of "Honor Lobster Universe", which supports one-click shrimp raising on PC and tablet, and will be compatible with the shrimp access of other ecosystem devices in the future; Chen Xi, the design director of ColorOS of OPPO, showed some of the functions of Xiaobu Claw on social media and stated that "Xiaobu Claw still has security issues to be resolved".

In other words, the current "shrimp farming" initiative by mobile phone manufacturers is mainly in the internal testing phase, and there is no clear timeline for its large-scale rollout.

For example, Xiaomi's miclaw app is currently only available for limited closed beta testing on the Xiaomi 17 series, Xiaomi 15S Pro, and Redmi K90 series. Users can access the Xiaomi miclaw app after receiving an invitation code and updating their system. "There are currently no plans to charge users during the closed beta period," said Lu Weibing, partner and president of Xiaomi Group.

Regarding mobile phone manufacturers deploying the mobile version of "Lobster," an industry insider revealed that "OpenClaw is essentially an open-source framework that includes a third-party skills and plugin ecosystem, and can also call various large models. For ordinary users, deploying OpenClaw has a high barrier to entry, but for mobile phone manufacturers, there is no technical difficulty. The difficulty lies in issues such as permission acquisition, user information security, and legal compliance."

"Mainstream mobile phone manufacturers are dealing with hundreds of millions of ordinary users. Any AI function must be fully verified before it can be pushed out, ensuring a mature, safe, and stable experience," a mobile phone manufacturer employee revealed.

Mobile phone manufacturers are flocking to shrimp farming.

Large model vendors are keen to deploy "Lobster", which can be simply understood as a business of "monetizing computing power", that is, making intelligent agents call the model more frequently and perform complex tasks, thereby consuming more tokens and directly boosting API revenue.

However, this logic doesn't hold true in the mobile phone industry. After spending thousands or even tens of thousands of yuan on a phone, users are rarely willing to pay extra for each specific task. Since they can't directly profit from "selling tasks," why are leading mobile phone manufacturers still willing to bear the costs of computing power and tokens to internally test a dedicated mobile version of "Claw"?

One reason is that, on the road from traditional mobile AI assistants to "personal intelligent agents," OpenClaw is getting closer and closer to the ideal form of a "super assistant."

Unlike previous voice assistants that could only respond passively, OpenClaw is more like a "digital employee" that is online 24/7, allowing ordinary users to truly experience the real possibility of AI replacing human labor for the first time.

From the underlying logic of OpenClaw, its core value lies in its strong "autonomy." It breaks through the boundaries of the chat box. As long as the corresponding Skills are configured and sufficient Tokens are authorized, OpenClaw can remember the user's habits and tasks, autonomously plan steps, call tools and operate software, until the final result is returned.

However, to truly tame this "autonomy" that floats in the cloud onto the small screen of a mobile phone, simply adding apps is clearly not enough. It requires mobile phone manufacturers to carry out a deep, bottom-up reconstruction of the operating system.

In terms of specific implementation paths, both Huawei's Claw and Xiaomi's miclaw have chosen to enter the market as "system-level applications." This approach essentially encapsulates previously discrete software functions, system permissions, and even cross-device capabilities into unified Skills that can be invoked by the Agent, and then organically connects them through a self-developed inference-execution engine.

Taking Xiaomi's miclaw as an example, it integrates more than 50 system tools and ecosystem services to build a closed-loop engine of "perception-reasoning-execution". When faced with user commands, the engine will autonomously break down the steps, match tools, determine parameters, and continuously revise based on the execution results until the task is completely delivered.

Huawei's Xiaoyi Claw is built directly on the HarmonyOS platform. "Xiaoyi Claw has three major advantages: system-level permissions (direct access to underlying functions without third-party app redirection), full-scenario collaboration (seamless linkage between mobile phones, PCs, in-vehicle systems, and smart homes), and data security isolation (local processing of user privacy data)," a Huawei insider revealed.

However, deploying "Lobster" on mobile phones presents challenges beyond just technology and ecosystem. It also requires handling sensitive data properly while ensuring security and compliance, breaking down barriers across applications and platforms, and even reshaping the entire industry's profit distribution structure.

"The most important thing when deploying Lobster on users' frequently used mobile phones is to ensure information security," an employee of a mobile phone manufacturer emphasized.

These concerns about information security are not unfounded. Because OpenClaw's default security configuration is weak, attackers can easily gain complete control of the system, and security risks such as prompt word injection, accidental operations, and malicious plugin attacks have already emerged.

Faced with these hidden security "reefs," security governance has become an inviolable red line for mobile phone manufacturers when deploying "Lobster" on a large scale.

Taking Xiaomi's miclaw as an example, to prevent agents from arbitrarily executing high-risk operations such as payments in the cloud, miclaw directly "crippled" all tool registrations involving transfers and order placements at the code level. This means that without explicit user confirmation such as fingerprint verification or password input, any financial transaction will not be triggered, thus locking in the risk of automatic deductions at the source.

The battle for AI ecosystem entry points has begun.

OpenClaw's near-perfect approximation of the ideal "super assistant" is merely the surface-level incentive for mobile phone manufacturers to "raise lobsters" (i.e., develop a competitive strategy). The deeper game lies in the fact that as users gradually become accustomed to the interaction method of "getting things done simply by speaking," the old order of the traditional mobile internet, which is based on apps and where mobile phone manufacturers control the distribution rights of app stores, begins to loosen.

As Nvidia founder Jensen Huang said, "Mac and Windows are operating systems for personal computers, while OpenClaw is the operating system for personal AI."

In the PC era, whoever controlled the operating system controlled the gateway to the ecosystem. In the AI ​​era, this rule still applies, but the battle for access has shifted to intelligent agents.

Imagine if users got used to solving all their problems on third-party agents (such as web pages or standalone apps like OpenClaw), and smartphones might become nothing more than mere "hardware bases".

With major internet companies deploying mobile versions of "Lobster," the sense of crisis among mobile phone manufacturers is self-evident.

Just as mobile phone manufacturers announced the launch of mobile versions of "Lobster", internet giants such as Baidu and Alibaba also quickly took action and launched free internal testing of mobile versions of "Lobster".

On March 12, Baidu launched the "Redfinger Operator" app on Android, allowing users to directly experience the AI ​​assistant capabilities on their mobile devices and perform cross-application operations such as hailing a ride and ordering food delivery. Following closely behind, Alibaba Cloud launched the mobile version of OpenClaw, "Lobster"—JVS Claw—the next day, emphasizing "out-of-the-box usability." Users can operate applications, process files, and complete complex tasks in a secure, isolated cloud space using simple natural language commands.

Regarding the deployment of "lobster farming" on mobile phones by mobile phone manufacturers and major internet companies, Guo Tianxiang, research manager of IDC China, said, "At present, the practical application value of (raising lobsters) on mobile phones is limited. The key bottleneck is that if you try to call third-party apps, you will still face the problem of API authorization. If you force the call, you may encounter the situation of being disabled by these third-party apps, just like the previous Doubao phone."

Learning from the experience of the "Doubao Phone," mobile phone manufacturers such as Huawei and Xiaomi prioritized testing the mobile version of "Lobster" within their own closed ecosystems when deploying it.

For example, Xiaomi miclaw currently focuses on verifying the task execution capabilities of large models in the "human, vehicle, and home ecosystem"; while Xiaoyi Claw prioritizes achieving collaborative flow among Huawei's own devices such as mobile phones and tablets.

However, while operating "Lobster" in a relatively closed ecosystem can avoid some risks, it also restricts "Lobster's" actions to some extent, since users' high-frequency needs are often scattered across third-party national applications such as WeChat and Douyin.

In order to strike a balance between security compliance and full functionality, vendors have not chosen to completely abandon cross-application collaboration, but instead have explored a more cautious and controlled technical path.

Regarding collaboration services with third-party applications, a technical staff member close to Xiaomi revealed that Xiaomi miclaw currently collaborates with third-party applications mainly through two industry-standard methods: one is to launch the application or trigger specific actions through Intent drivers (SendIntentTool); the other is to promote the application to adapt to its AppTool SDK (based on the AIDL protocol), and to perform deeper function calls and task collaboration through preset data formats. Third-party apps can also proactively push notifications to Xiaomi miclaw to trigger tasks.

"Distant water" cannot quench "immediate thirst".

Currently, deploying proprietary "lobsters" from the underlying system is a crucial step in the evolution of smartphones into "AI phones." However, for manufacturers eager to seek incremental growth in the AI ​​wave, the primary challenge in building a super intelligent agent is cost pressure.

Deploying a localized "Lobster" is not a simple software upgrade; it also requires upgrades to hardware such as the core processor and storage. The high-frequency inference and real-time response of large models place higher demands on the NPU computing power of the core processor (SoC), and also significantly raise the requirements for the specifications of RAM and storage chips.

"Running large models on mobile devices is affected by a series of technical factors such as storage space and power consumption. The larger the number of parameters, the more difficult it is to run on a mobile phone. A model with 1 billion parameters will occupy 1GB of mobile phone memory, 7 billion will occupy 4GB, and 13 billion will occupy 7GB," revealed the director of the AI ​​solutions center of a leading mobile phone manufacturer.

Currently, storage chip prices are on an upward cycle, and every GB of memory upgrade directly squeezes the profits of the entire hardware system.

More troublesome than the one-time hardware investment is the ongoing usage cost after the mobile version of "Lobster" is activated. On the PC side, each task execution corresponds to real token consumption and computing power costs. The previous news that "a monthly salary of 20,000 yuan cannot afford to support Lobster" directly exposed this "cost anxiety" to users.

“Before using ‘Lobster,’ you need to figure out what you’re going to do with it,” explained Feng Nian, founder of Dianjinshou (an MCN agency). “In the video production process, the token consumption for editing and generating videos is actually very different, but many beginners don’t understand what Lobster can actually do.”

Fengnian used the team's actual operations as an example to calculate the costs: "Our editing primarily uses OpenClaw deployed on a Mac mini 4 to assist with the work. Specifically, 'Lobster' is responsible for generating scripts for restaurant review videos based on local hotspots. Some of these scripts are shot with real people, while others are generated using AI (such as Seedance 2.0 or Sora 2). Lobster can control the Mac mini to edit the videos while simultaneously calling Sora 2's API to generate the videos. It's cheaper to have people do some of these tasks, and more cost-effective to delegate them to AI. In a day, we can produce about 12 original + montage videos, with a corresponding token consumption cost of approximately 15 yuan."

"The core difficulty in decision-making lies in balancing the cost of token computing power with the salaries of junior editors," Feng Nian added. "The key to making reasonable use of 'lobsters' is to clarify which tasks should be assigned to 'lobsters' and which should be left to humans. Unfortunately, many companies are currently 'raising lobsters' purely to follow the trend and show off their skills, without generating actual productivity."

While a daily token cost of 15 yuan may seem low, it's a significant burden given the massive user base of mobile phone manufacturers. With hundreds of millions of users accustomed to the "buy the hardware, get the service for free" model, whether mobile phone manufacturers can sustain the resulting massive computing power and token costs in the long run remains to be seen.

“Mobile phone manufacturers may adopt a ‘buy a phone, get free computing power’ model in the future,” an industry insider predicted. “For example, a certain amount of free tokens may be included with the purchase of a phone to handle light daily tasks such as writing daily reports and booking tickets. As for complex and high-consumption operations such as video generation, they may charge separately based on the complexity of the task, or the user may have to bear the cost of tokens exceeding the limit.”