Meta Muse, which is being closely followed by Alibaba, Tencent and ByteDance, actually had Kimi take a seven-month head start ahead of it.

The 50-cent-per-time computing power black hole and the battle for ultimate payment authority: when Kimi starts to hit the brakes, what exactly are the big tech firms calculating for personal agents?

Personal Agent, this "third AI battlefield" for large tech companies, seems far more lively than the previous Chat and Work tracks.

Alibaba is conducting intensive tests on "Qianwen Go!"; Tencent is pushing forward LightVela at full speed and conducting internal beta tests for Handy Bot; ByteDance is ramping up automated control for Doubao. Following the pace of Meta Muse, domestic large tech firms are setting off a wave of rushing into the market to take over users' digital lives.

But a question suddenly occurred to me: Kimi launched a very complete product called Kimi Claw as early as February this year, the so-called Personal Agent, right?

If we define a personal agent as an AI system that can understand long-term goals, schedule tools across applications, and perform complex multi-step tasks, someone in China's internet industry has already waded through this river long ago. But it is intriguing that Kimi, the industry leader that first established the public perception of personal agents across the entire sector, now holds an unusually different posture.

7 months after its release, Kimi Claw has remained in the Beta phase, with the update frequency of core features dropping significantly. According to information I obtained, Kimi's internal team now puts very little effort into maintaining the Claw product.

For this "latest battlefield" that has attracted high expectations from the industry and swarms of large tech firms, why did the startup that was the first to test the waters take the initiative to hit the brakes?

Behind this is the core issue this article discusses: the financial model of personal agents.

The Computing Power Black Hole Behind the 199-yuan Subscription Plan

Let's start with the simplest calculation first.

Some Kimi users did the math: if the 199 yuan/month subscription plan is amortized over daily conversations, the average cost per interaction is about 50 cents. But this price point does not hold up when measuring complex Agent tasks, as the order of magnitude is completely different.

The computing power consumption of general Chat (such as asking "what's the weather today" or "help me write a leave application") is single-time, low-context, and ends in one round. But a real Personal Agent operates on a completely different logic.

An Agent needs to first break down the target, launch an isolated cloud sandbox, perform dozens of consecutive Tool Calls, web scraping, self-correction, and Chain of Thought (CoT) verification. Coupled with the tens of thousands or even hundreds of thousands of tokens in context retained to maintain active memory, a seemingly ordinary Agent task often triggers as many as a dozen to dozens of underlying model calls.

Adding the cost of cloud computing and storage, the actual token and sandbox computing power consumption of a heavy user running a complex Agent task far exceeds 50 cents.

The C-end SaaS subscription model (upfront charging) is inverted with the exponentially rising computing power consumption of Agents in terms of financial model.

Facts have proved that this mismatch between reasoning cost and revenue structure has long been reflected in Kimi's financial data.

According to public information, in the revenue structure of Moonshot AI (Kimi), the B-end API calls and developer ecosystem have always contributed the largest share of revenue, with its API business accounting for more than 70% of total revenue and continuing to rise.

For independent AI vendors, APIs are billed precisely by actual tokens, so more calls bring more revenue with clear gross profit margins. However, the C-end monthly fixed subscription Claw is subsidizing heavy power users with expensive computing power, easily turning into a gross profit black hole.

When more than 70% of a company's core cash flow comes from APIs, it becomes an extremely rational and inevitable choice to give the loss-making C-end Claw the cold shoulder and reduce its computing power occupation.

Another Accounting Logic of Large Tech Firms' "Back-end Discounting"

Since the leading player Kimi has proved that the personal agent model relying on C-end upfront subscription is not financially viable, why are Alibaba (Qianwen Go!), Tencent (LightVela/Handy Bot) and ByteDance (Doubao) still pouring huge amounts of capital into this battlefield?

Because large tech firms are calculating a completely different set of accounts. They are using Agents to reconstruct their most core back-end commercial assets.

Different large tech firms have completely distinct "back-end discounting" logic:

The first is Meta. Meta Muse helps users manage schedules and recommend products. On the surface it is a free assistant, but in essence it is legally obtaining users' extremely deep "strong intent context".

Traditional social likes and browsing are weak intent, while Agents record users' plans such as where to go on vacation next month, what car they want to change to, and what training courses to sign up for their children. Exchanging computing power cost for improved targeted advertising efficiency means leveraging computing power to realize the discount of the entire advertising ecosystem.

Alibaba is calculating GMV and performance take rate.

In Alibaba's latest test scenarios for "Qianwen Go!", typical tasks include one-click price protection and refund on Taobao, flight delay claim settlement, and car maintenance price comparison and reservation. Users use Agents to trigger real service fulfillment. The underlying computing power consumption of Agents helping users check discounts and complete orders can be easily covered by transaction commissions from Taobao/Tmall, commissions from Amap and Fliggy, as well as promotion fees paid by merchants.

Tencent is calculating private domain retention and financial/payment circulation.

Tencent's Agents control WeChat's private domain context and social relationship chains. As long as Agents can keep users' personal affairs within the WeChat closed loop, Tencent can obtain extremely high user lifetime value (LTV) through deposited funds, payment processing fees, financial monetization and video account e-commerce.

ByteDance (Doubao) is calculating content attention and the protection of traffic entrances, making Agents a pre-router for personal information filtering and consumption decision-making, preventing its short video, e-commerce and advertising traffic from being intercepted by other Agents.

This explains why independent AI companies can only helplessly scale down while giants have to fight this battle: Independent AI companies earn money by selling tools, while giants are seizing the agency right of the ecosystem entrance.

Payment: The Ultimate Destination of the Agent Entrance

Following this business logic to further deduce, the new question is: as the "third-generation entrance", where does the personal agent end up?

My point of view is: it cannot bypass payment.

An Agent without payment authorization can never be more than a "armchair strategist" suggestion box. Even if it can write a perfect travel guide, users still have to manually open the App to select a room and enter the password to pay. At the moment of redirection, the commercial value and ecosystem closed loop are cut off.

Only when users allow Agents to have the authority of "controlled withholding/authorized payment", can Agents truly take over the agency right of the real world.

This also explains why ByteDance (Doubao) quietly launched the utility bill payment function for water, electricity and gas recently.

On the surface, utility bill payment is an extremely tiny and inconspicuous function. But from a tactical perspective, it is a very precise move by ByteDance to make up for its shortcomings.

ByteDance has a huge DAU base, but users do not have fund accounts or payment habits in Doubao. Utility bill payment is a typical scenario of "low decision-making, high frequency, no price comparison", which does not require users to select brands or collect coupons, they only need to complete the payment.

ByteDance's entry into this scenario essentially aims to cultivate users' awareness of "binding cards, authorizing, and deducting payments directly in the AI assistant". Only by opening up the payment pipeline can all kinds of service fulfillment of Doubao be realized in the future.

But it is not easy to cross the ultimate high wall of "payment".

Alipay (credit system, risk control model and withholding agreement) and WeChat Pay (high-frequency social relationship chain and offline private domain) are two natural barriers built by Tencent and Alibaba over more than ten years.

Some time ago, when Doubao's mobile assistant tried to identify the screen through GUI and simulate users' manual clicks to complete cross-platform order payment, mainstream Apps including WeChat, Taobao, Alipay and Meituan quickly reached a tacit agreement to restrict such cross-platform automatic control operations.

The logic is very harsh: "We can allow you to search our data, but your Agent must never cross the boundary to control our payment accounts and order buttons. Because once the payment entrance is intercepted by you, we will degenerate into a pure underlying tool."

Looking back, this wave of boom brought by Meta Muse has not changed the essence of the business world.

For independent AI startups, the C-end personal agent is a very tempting but extremely dangerous computing power trap. For giants such as Alibaba, Tencent and ByteDance, personal agents are the third-generation "super gatekeepers" for internet traffic and transactions after "search" and "recommendation".

This covert battle has only just begun. Beyond the technology competition, it is also an ecosystem positioning battle about whether computing power costs can be discounted by the back-end ecosystem, and who can finally take over the end point of personal payment and service fulfillment.

This article is from the WeChat official account "Yangtze River Delta Momo", the author is Sanhuo, and 36Kr publishes it with authorization.