Doubao Summons Caocao for Ride-Hailing Services

On June 22, multiple media sources reported that the Doubao App had initiated grayscale testing of its "one-click ride-hailing" feature in Beijing and Hangzhou, with ride capacity sourced from Caocao Chuxing. However, Doubao's actual support for ride-hailing services had been operational since at least earlier that month.

In its official statements, Doubao refers to Caocao as its "inaugural partner in the mobility sector." This positioning implies that Doubao and Caocao are not exclusively tied, but rather that Doubao acts as an entry point, aggregating capacity from multiple providers.

To comprehend why Caocao was the first to partner with Doubao, we must first examine its current market standing.

On the surface, Caocao Chuxing is no minor player: backed by Geely, listed on the Hong Kong Stock Exchange, with a service network spanning nearly 200 cities and a market share in the low single digits, ranking second in the ride-hailing industry.

However, this "second place" is not as prestigious as it appears, given that Didi commands over 70% of the market. Ride-hailing is characterized by network effects, and within this industry, only Didi can ensure service availability independently. Both Caocao and T3 rely on external traffic sources.

The figures disclosed in the prospectus are revealing: in 2022, approximately half of Caocao's transaction volume originated from its own app; by 2024, 85.4% of its total gross transaction value (GTV) of RMB 17 billion came from aggregated platforms like Gaode, leaving only 14.6% from its own channels.

This indicates that Caocao is increasingly resembling an affiliated capacity provider.

Against this backdrop, Doubao's olive branch becomes readily understandable. With over 300 million monthly active users, Doubao is a premier AI application in China. For Caocao, partnering with Doubao means accessing another customer acquisition channel and hedging against reliance on aggregated platforms.

However, focusing solely on "additional entry points" underestimates Caocao's strategic calculations.

Caocao has recently unveiled its RoboX strategy, proposing a "Dual 100,000 Plan" to deploy 100,000 Robotaxis and 100,000 Robovans by 2030, and has established a dedicated AI division.

Taken together, these developments suggest that the significance of the Doubao partnership extends beyond merely directing traffic for ride-hailing. Today's ride-hailing orders through Doubao could evolve into orders for autonomous driving services tomorrow. This aligns with Geely's ambitions for autonomous mobility.

Transforming Chat into Transactions

The ride-hailing industry has become highly competitive in recent years, as evidenced by the decline in per-kilometer fares and multiple warnings from authorities about oversupply.

Generally, when an industry faces oversupply, the importance of demand or channels capable of absorbing demand increases. Thus, a notable aspect of Doubao's partnership with Caocao deserves attention.

According to media reports, when drivers receive orders dispatched by Doubao, the client app prompts them: "This is a Doubao service order. Providing excellent service for this trip will earn you a RMB 2 platform surprise service fee."

The accuracy of this information is uncertain. While an additional RMB 2 service fee could theoretically incentivize better driver performance, from a commercial standpoint, offering subsidies to passengers might currently yield better results than subsidizing drivers.

My review of testing by a Red Star Capital Bureau reporter suggests the opposite may be true. For the same origin and destination, hailing a ride through Doubao was more expensive than through the Caocao Chuxing app, possibly due to the lack of coupon interchangeability.

Compared to QianWen's integration with Alibaba's internal ecosystems like Taobao Flash Sales and Gaode Fliggy, Doubao must rely on partners to enhance its transactional capabilities, necessitating more refinement in the details and overcoming additional hurdles.

Doubao's integration of ride-hailing services clearly aims to complete its AI capabilities puzzle.

Alibaba's QianWen has been more proactive in this regard, spending billions during the Lunar New Year to subsidize food delivery—a move that likely left a lasting impression on many.

However, whether for QianWen or Doubao, meeting expectations for AI-driven transactions still seems distant. A more realistic goal for Doubao may be to convert non-revenue-generating chat traffic into revenue-generating transactional traffic, thereby mitigating its high operational costs as much as possible.

LatePost reported that as of the first half of 2026, the Doubao App, used by over 200 million people daily, generates less than RMB 1 million in daily revenue, primarily from e-commerce commissions. Based on this ratio, the daily e-commerce transaction volume facilitated by the app is only in the tens of millions.

While revenue is modest, costs are substantial. Reports estimate that by May of this year, Doubao's daily computational costs had reached tens of millions of yuan, surpassing Bilibili's operational costs—and this excludes investments in self-built computational centers for model training.

From this perspective, although QianWen and Doubao are converging in their activities, their strategies may differ.

Doubao has already become the chat assistant with the highest monthly active users, and internally, some now question the product's value. However, QianWen does not yet face such pressure, with market share acquisition remaining its primary objective. Thus, Doubao has an incentive to generate cash flow wherever possible, exemplified by its introduction of paid plans and the traditional internet approach of monetizing traffic.

Three Paths to AI Monetization: If One Fails, Will Another Succeed?

Expanding our view to the entire industry, AI companies today face roughly three paths to monetization. These are not parallel options but rather three tiers arranged by "ideal, reality, and fallback."

The first path is charging users subscription fees. This is the most straightforward and intuitive approach: if your product is good enough, users will willingly pay for it.

The direction of Doubao's planned paid version focuses on productivity scenarios like PPT generation, data analysis, and film and television production. The rationale is sound: complex tasks do consume more computational power, and someone should pay for that. However, getting Chinese users to pay for AI subscriptions on a large scale is extremely challenging.

The domestic market has been accustomed to free offerings, with ByteDance being a major contributor. Through free services like Tomato Novels, Hongguo Short Dramas, and Qishui Music, ByteDance disrupted Tencent's long-established paid models for online literature, short dramas, and music.

Now, as ByteDance seeks to charge users, it somewhat contradicts this trend. Overcoming this is as challenging for ByteDance as it would be for Tencent. With competitors like QianWen and DeepSeek offering free services that are adequate in many scenarios, Doubao lacks a strong rationale for charging when its model intelligence does not clearly outperform theirs.

Moreover, efficiency tools have traditionally been free in China: while Zoom earned over USD 1 billion annually from paid video conferencing, DingTalk, Feishu, and Tencent Meetings have long offered core features for free. In such an environment, pursuing subscriptions feels like swimming against the tide.

The second path is selling capabilities to businesses.

This is currently the only repeatedly validated viable path.

Across the ocean, Anthropic demonstrated this with Claude Code. However, a prerequisite exists: you must possess "capabilities" to sell, not just "traffic."

ByteDance already has successful products on this path. Seedance generates USD 2 billion in annualized revenue with a 70% gross margin, far outperforming the chatbot business.

Most of this revenue comes from enterprise clients, with short drama companies paying for Seedance and then spending their marketing budgets on platforms like Hongguo, completing a closed loop within the ByteDance ecosystem.

Another representative case is Zhipu. Recently, Zhipu released and open-sourced its new model, GLM-5.2, achieving first place among globally available models on the frontend development blind testing platform Code Arena.

Shortly thereafter, on June 22, Zhipu's Hong Kong stock market value briefly surpassed HKD 1 trillion, doubling within days. It follows the same programming and productivity path as Anthropic. Although Zhipu's market performance resulted from multiple factors, the rationality of betting on programming and productivity scenarios has been proven.

The third path is monetizing traffic, the most familiar approach from the mobile internet era. First, scale up through free offerings, then convert traffic into money through commissions, advertising, and e-commerce. This does not require selling sophisticated capabilities—only that you have users.

Doubao's 300 million monthly active users cannot all be potential clients for productivity scenarios. Only a small proportion has the demand and willingness to pay for productivity tools, and they may not choose Doubao.

For the vast majority of ordinary users, no paid plan can extract real money from their pockets. Thus, exchanging scale for transactions and monetizing traffic through commissions and advertising may be the only viable path in sight.

However, the issue is that in the mobile internet era, the flywheel of "scale up for free, then monetize gradually" worked because the marginal cost of adding one more user was nearly zero—the larger the scale, the closer to a money-printing machine.

Yet every AI invocation burns computational power, making it more akin to manufacturing with raw material costs. Under this new cost structure, more traffic does not necessarily bring profitability closer but may instead bring losses nearer.

Thus, the real question may be: Can the tried-and-true old path from the mobile internet era still work in AI's new, resource-intensive landscape?

Just as the success of Doubao's ride-hailing feature does not depend on how many cars it summons but on something more fundamental—whether the small commission it earns from each ride can cover the computational cost bill it incurs.

Doubao does not yet have an answer to this question. Neither, perhaps, does the entire industry.