MOMO in the Yangtze River Delta

The Battle for Computational Resources at 50 Cents per Use and the Quest for Ultimate Payment Authority: As Kimi Taps the Brakes, What Strategies Are Tech Titans Pursuing with Personal AI Agents?

The arena of personal AI agents, deemed the 'third AI battlefield' for tech giants, appears even more vibrant compared to the preceding eras dominated by Chat and Work applications.

Alibaba is rigorously testing 'Qianwen Go!'; Tencent is making significant strides with LightVela and conducting internal trials of Handy Bot; ByteDance is intensifying its automation efforts for Doubao. Following in Meta Muse's footsteps, domestic tech behemoths are igniting a wave of enthusiasm to join the fray and dominate users' digital lives.

Yet, a thought suddenly struck me: Wasn't Kimi the pioneer in launching a comprehensive Personal Agent product—Kimi Claw—back in February of this year?

If we define a personal AI agent as an AI system capable of comprehending long-term objectives, coordinating tools across various applications, and executing multi-step complex tasks, then China's internet sector has long harbored pioneers in this domain. Interestingly, Kimi, as the early trailblazer in establishing the concept of personal AI agents, now appears to be adopting an unconventional stance.

Seven months post-launch, Kimi Claw remains in its Beta phase, with a discernible decrease in the frequency of core feature updates. According to my sources, Kimi is now allocating relatively minimal internal resources to maintain the Claw product.

In a 'new battlefield' eagerly anticipated by the industry and enthusiastically entered by tech giants, why has the startup that first plunged in decided to tap the brakes?

This brings us to the crux of the matter discussed in this article: the financial model underpinning personal AI agents.

01 The Computational Quagmire Behind the 199 Yuan Package

Let's commence with a straightforward calculation.

Some Kimi users have estimated: if the 199 yuan/month package is distributed across daily conversations, the average cost per conversation hovers around 50 cents. However, using this price to gauge complex Agent tasks is misguided.

General Chat (e.g., inquiring, 'What's the weather like today?' or 'Help me draft a leave request') consumes computational resources in a singular, low-context, and single-round fashion. In contrast, a true Personal Agent operates on a vastly different logic.

The Agent must first dissect the objective, initiate a cloud-based isolated sandbox, and continuously execute dozens of tool calls (Tool Call), web scraping, self-correction, and Chain-of-Thought (CoT) verification. Coupled with the tens or even hundreds of thousands of Token contexts maintained for active memory, a seemingly mundane Agent task can actually trigger numerous model calls at the underlying level.

Considering cloud computing and storage costs, the actual Token and sandbox computational consumption for heavy users executing complex Agent tasks far surpasses 50 cents.

The C-end SaaS subscription model (charging upfront) is financially misaligned with the exponential increase in computational consumption of Agents.

Facts have substantiated that this mismatch between reasoning costs and revenue structure is already evident in Kimi's financial data.

According to public information, in Yuezhi'anmian's (Kimi) revenue structure, API calls and the developer ecosystem from the B-end have consistently been the primary revenue drivers, with API business revenue accounting for over 70% and steadily rising.

For independent AI vendors, API services are billed based on actual Token usage, yielding clear gross profit margins; the more calls, the greater the revenue. In contrast, the C-end Claw, with its fixed package subscription, is subsidizing heavy users with costly computational resources, easily becoming a gross profit black hole.

When over 70% of a company's core cash flow stems from APIs, adopting a passive approach and reducing computational resources for the loss-making C-end Claw becomes an eminently rational and inevitable choice.

02 Big Tech's 'Backward Discounting' Strategies

Since even the leader, Kimi, has demonstrated that relying on C-end forward subscriptions for personal AI agents is not financially sustainable, why are Alibaba (Qianwen Go!), Tencent (LightVela/Handy Bot), and ByteDance (Doubao) still investing heavily in this battlefield?

Because tech giants are calculating differently. They are leveraging Agents to restructure their core backend commercial assets.

Distinct big tech firms possess unique 'backward discounting' logics:

First up is Meta. Meta Muse aids users in managing schedules and recommending products. On the surface, it's a complimentary assistant, but fundamentally, it's legally acquiring users' highly detailed 'strong intent contexts (Intent Context)'.

Traditional social media likes and views signify weak intent, whereas Agents record where users intend to vacation next month, what car they aspire to purchase, or what classes they plan to enroll their children in. Exchanging computational costs for enhanced targeted advertising efficiency is a means to harness computational power to bolster the advertising ecosystem.

Alibaba calculates based on GMV and fulfillment commissions (Take Rate).

In Alibaba's latest 'Qianwen Go!' testing scenarios, typical tasks encompass 'Taobao one-click price protection and refund, flight delay claims, and car maintenance price comparison and booking'. Users employ Agents to trigger actual fulfillment. The computational consumption behind Agents assisting users in verifying discounts and completing orders is readily offset by transaction commissions from Taobao/Tmall, commissions from Gaode and Fliggy, and merchants' promotion fees.

Tencent calculates based on private domain retention and financial/payment flow.

Tencent's Agent governs WeChat's private domain contexts and social relationship chains. As long as the Agent can retain users' personal affairs within the WeChat closed loop, Tencent can derive high user lifetime value (LTV) through trapped funds, payment fees, financial monetization, and video account e-commerce.

ByteDance (Doubao) calculates based on content attention and traffic gateway defense, positioning the Agent as a front-end router for personal information filtering and consumption decisions to prevent its short video, e-commerce, and advertising traffic from being intercepted by other Agents.

This elucidates why independent AI companies can only reluctantly scale back, while giants must engage in combat: independent AI companies generate revenue from tools, whereas giants vie for agency rights over ecological gateways.

03 Payment: The Ultimate Destination for AI Agent Gateways

Continuing along this business trajectory, a new question emerges: where does the 'third-generation gateway' of personal AI agents ultimately lead?

My perspective is: payment is inevitable.

An Agent devoid of payment authorization will perpetually remain a 'theoretical' suggestion box. Even if it can craft a flawless travel itinerary, users must still manually open an app to select a room and enter a password to pay. At that moment of redirection, commercial value and the ecological closed loop are severed.

Only when users grant Agents 'controlled direct debit/authorized payment' permissions can Agents truly assume agency rights in the real world.

This also clarifies why ByteDance (Doubao) recently discreetly introduced utilities payment functions for water, electricity, and gas bills.

On the surface, utilities payment appears to be a minor and unremarkable feature. However, tactically, it's a highly precise maneuver by ByteDance to address a weakness.

ByteDance boasts a massive DAU, but users lack funds accounts and payment habits within Doubao. Utilities payment is a quintessential 'low-decision, high-frequency, zero-price-comparison' scenario that doesn't necessitate brand selection or coupon collection—users simply need to pay.

ByteDance's foray into this scenario is essentially cultivating users' mindset of directly 'binding cards, authorizing, and deducting payments' within the AI assistant. Only by establishing payment channels can Doubao's future service fulfillments proceed seamlessly.

However, surmounting the 'payment' ultimate barrier is no easy feat.

Alipay (with its credit system, risk control models, and direct debit agreements) and WeChat Pay (with its high-frequency social relationship chains and offline private domains) are two formidable fortresses constructed by Tencent and Alibaba over a decade.

Recently, when the Doubao mobile assistant attempted to employ GUI screen recognition and simulate user manual clicks for cross-platform ordering and payment, mainstream apps like WeChat, Taobao, Alipay, and Meituan swiftly formed a tacit agreement to restrict such cross-platform automated takeover operations.

The rationale is ruthless: 'I'll permit you to search my data, but your Agent absolutely cannot come over to control my payment accounts and order buttons. Because once you intercept the payment gateway, I'm reduced to a mere underlying tool.'

Reflecting on this, the recent surge propelled by Meta Muse hasn't altered the fundamental nature of the business world.

For independent AI startups, C-end personal AI agents are an enticing yet perilous computational trap. For giants like Alibaba, Tencent, and ByteDance, personal AI agents represent the third-generation 'super gatekeepers' of internet traffic and transactions, following 'search' and 'recommendation'.

This covert war has just commenced. Beyond technological competition, it's also a battle over whether computational costs can be offset through backend ecosystems and who can ultimately seize personal payment and fulfillment endpoints—an ecological positioning war.