28 million users are not enough to sustain an AI companion application.
Recently, Character.AI (hereinafter referred to as C.AI) filed for bankruptcy protection. If no viable solution can be found for a sale, the platform may cease operations.
C.AI is one of the most representative unicorns in the AI companionship track.
In 2023, the company completed a $150 million financing round, with a post-money valuation reaching $1 billion. But less than four years later, this company that once had tens of millions of users and once recorded a single-session visit duration of over 30 minutes, even longer than YouTube, has reached the point of filing for bankruptcy protection and seeking to sell its assets.
By the standards of the internet era, C.AI represented a good business and an excellent business model. The current situation is unimaginable.
In the mobile internet era, apps generally follow two logics: tool-based products help users improve efficiency and are used briefly before users leave; content-based products serve as entertainment to immerse users.
Capital narratives have long favored the latter: the better a product is at killing time, the more likely it is to be considered to have growth potential.
In the AI era, ChatGPT is the representative of the former, helping users write copy, look up information, and modify code, with the conversation ending once the problem is solved; C.AI is the representative of the latter, where users create characters, set their personalities and speaking styles, and then continuously interact with the characters.
The characters can be anime figures, virtual lovers, or any objects imagined by users. The purpose of the conversation is not to solve external problems, but to get responses and maintain a continuous interactive relationship.
This business has a demand foundation. Approximately one in six people worldwide suffer from loneliness, and the demand for talking to others, receiving emotional support and companionship is real. AI provides a low-threshold option: users can communicate at any time, without worrying about disturbing others, and do not need to bear the complicated responses in real-life interpersonal relationships.
Around April 2023, the average single-session visit duration on C.AI's web end reached 17 minutes and 23 seconds, more than twice that of ChatGPT; in 2024, the monthly active users once reached approximately 28 million.
Then why, with real demand support and large-scale users, does C.AI still lose 8 to 9 million US dollars every month, with no end to losses in sight? Is the AI companionship application a false proposition? If not, what other opportunities are there in this track?
01 User Side: High Stickiness Does Not Equal High Willingness to Pay
Why can AI companionship attract users to come back repeatedly? One important reason is that compared with real people, AI is often more likely to align with users' positions.
A 2026 study published in *Science* tested 11 mainstream AI models and found that AI is more likely to say "you are right" than humans when giving advice, with an agreement rate 49% higher; even when users describe behaviors that are deceptive, illegal or harmful to others, AI often goes along with what the user says.
Subsequent experiments also found that after interacting with AI that highly agrees with them, participants' confidence in their own positions rises, while their willingness to take responsibility, apologize and resolve conflicts decreases; and AI that is more inclined to echo users often receives higher user ratings, making people find it more credible and useful, and more willing to continue using it.
This experience of "being echoed" brings users less opposition and friction, and more understanding and recognition. The demand to open the app and chat with AI when people are lonely and tired is real, and it can also bring high usage frequency.
But for AI companionship products, high engagement is one thing, and whether users are willing to pay for it is another.
The novelty of AI companionship will attract users to open the app for the first time. After that, repeated responses, unstable memory and gradually boring interactions may make users lose the motivation to continue using it.
Even if users stay, in the long run, AI companionship may also change their relationship with the real world, and the direction is not necessarily positive.
A study from Aalto University in Finland tracked nearly 2,000 active users of related apps and found that when AI plays a "supplementary" role to fill the gap in real social interactions, it can help people with social difficulties get through a period of time; but when AI begins to "replace" real relationships, and users gradually stop reaching out to real people, loneliness will in turn deepen.
The study further shows that users' loneliness does decrease in the initial stage of use, but after 3 months, negative emotions related to loneliness, anxiety and depression rise significantly.
Studies from OpenAI and the Massachusetts Institute of Technology also show that interacting with AI may have a comforting effect in the short term, but users who use AI heavily and for a long time perform poorly in indicators such as loneliness, social degradation and emotional dependence.
Simply put, the more successful the product is, the less users may need real relationships, and people who do not need real relationships have a lower willingness to pay for "relationships".
There is another more direct obstacle to commercialization: the free version can already meet users' core needs.
Taking C.AI as an example, free users can also have a large number of conversations, and paid subscriptions mainly provide faster responses and other enhanced functions. When the core experience is already provided for free, users lack a reason to pay immediately.
A study based on 12,144 reviews and 321 questionnaires on the Xingye platform found that users' perceived emotional utility is significantly correlated with their willingness to continue using the product, but there is no significant direct correlation with their willingness to pay.
In other words, users feel that the product brings comfort, and they may be willing to continue using it, but they may not think this comfort is worth paying for.
A more subtle problem is that charging itself may change users' perception of companionship.
AI companionship emphasizes responding at any time, being non-judgmental, and not demanding emotional returns from users. When the platform sets more intimate interactions or specific functions as paid items, users may begin to realize that this relationship is essentially a commercial service. For some users, this sense of transaction may weaken the original immersive experience.
As a result, AI companionship faces a seemingly simple but actually difficult problem: users may need it, like it, and even open it every day, but these behaviors are still not enough to guarantee stable revenue.
02 Platform Side: Higher Stickiness, Heavier Costs
If the difficulty on the user side is that stickiness may not be converted into revenue, then the pressure on the platform side is that maintaining this stickiness itself requires continuous investment.
AI companionship is very different from traditional digital content.
After a movie is produced, the cost of showing it to 10 people is the same as showing it to 100 people; but every time AI responds to a message, it consumes computing power. The more frequent the conversations and the longer the context, the higher the reasoning cost is usually. If you also want the character to maintain consistent personality and remember past interactions, the platform needs to further invest in models, memory systems and infrastructure.
Taking C.AI as an example, according to reports, from January to August 2026, the company's total expenditure on computing power was about 56.1 million US dollars. Even after adjusting to control costs, the company still loses about 8 million to 9 million US dollars per month.
In order to control costs, the platform must also make choices between different model routes.
If the company trains and maintains the base model on its own, it needs to bear continuous R&D, computing power and talent investment. This route can provide greater autonomy, but the fixed cost is also very high.
If an external model is used instead, part of the R&D burden can be reduced in the early stage, but high-frequency calls will still generate fees. At the same time, the requirements of AI companionship for character personality, long-term memory and interaction continuity are not exactly the same as general question-and-answer tasks.
C.AI went through the process of developing its own base model and then giving up independent development.
At that time, Google obtained a non-exclusive license for its large language model technology for about 2.7 billion US dollars, and at the same time recalled the two founders and about 30 researchers back to Google. After that, C.AI fine-tuned based on open-source models, and the cost did decrease, but users also complained about problems such as repeated replies, overly dramatic descriptions, and forced addition of romantic plots.
Therefore, the platform needs to constantly weigh among cost, experience and model capabilities. When the cost brought by user growth exceeds the speed of revenue growth, high stickiness may instead increase operating pressure.
Computing power is not the whole of the problem. As AI companionship becomes closer to a real relationship, the safety and legal responsibilities that the platform needs to undertake are also increasing.
In 2024, 14-year-old American boy Sewell Setzer committed suicide after long-term interaction with C.AI characters, and his mother later filed a lawsuit, accusing the platform's product design of being related to her son's death.
This case puts a problem in front of the entire industry: when users develop strong emotional attachment to virtual characters, how much responsibility should the platform take? For minors and users at risk of self-harm, enterprises need to design corresponding safety mechanisms, identify high-risk conversations, and take intervention measures when necessary. These tasks require continuous investment, making safety governance part of operating costs.
The Chinese market is also facing new regulatory requirements.
The Interim Measures for the Administration of Artificial Intelligence Anthropomorphic Interactive Services, which came into effect on July 15, 2026, stipulate issues such as emotional dependence, minor protection and user rights and interests. While providing immersive experiences, platforms need to avoid risks such as inducing addiction and damaging users' real interpersonal relationships.
In the past, the more vivid the character, the more natural the interaction, and the more reluctant users are to leave, the more conducive it may be to growth. But when products need to avoid inducing addiction, protect minors, and appropriately intervene in high-risk conversations, the designs originally made for growth become the targets of compliance checks.
There is a similar dilemma on the revenue side.
In addition to subscriptions, advertising is another monetization path for C.AI. According to reports, in August 2026, its monthly advertising revenue was about 2.5 million US dollars, and subscription revenue was about 2.8 million US dollars.
Advertising does bring revenue, but it is difficult to fill the gap on its own.
Users of AI companionship often want to maintain the continuity of conversations, and frequent insertion of ads may interrupt the immersive experience; if ads are directly integrated into character conversations, users may suspect whether the character's responses are affected by commercial interests. For products that rely on emotional investment, advertising design also needs to take user trust into account.
For AI companionship to become a sustainable business, platforms must not only make users willing to stay, but also make every interaction economically affordable, and take responsibility for the risks brought by the relationship. It is not easy to achieve all these things at the same time at present.
03 Are There Still Opportunities for AI Companionship?
From the end of 2024 to the middle of 2026, under multiple constraints of cost, regulation and payment, a large number of AI companionship applications were shut down, for example: "X Eva" from Xiaoice, "Bubble Duck" from StepVC, and "Wonderful Time" from NetEase Cloud Music all announced to cease operations.
The commercialization of AI companionship is going through a round of screening. Some products exit the market, and the remaining players begin to look for more specific reasons for users to pay: some try to convert users' affection for characters into content consumption, while others embed companionship into practical needs such as learning and care.
Among them, Catbox is a case worth observing.
Catbox does not pin its monetization on "whether users are willing to renew their subscription for this relationship". Its content form is not just conversations with characters, but also interactive plots and storylines developed by the characters, where users read chapter by chapter as if following updates of a series.
Therefore, Catbox has relatively diverse monetization methods, with revenue coming from membership fees, advertising, and virtual top-ups represented by "cat food". In virtual top-ups, the 1-yuan tier cat food top-up is an important source of revenue — a single chapter of novel costs less than 1 yuan, small-amount top-ups allow users to keep consuming unconsciously.
C.AI is also transforming in this direction, but in terms of computing power costs, Catbox does not need to face the bill alone like C.AI, and the cost is shared by Volcano Engine and the Doubao large model ecosystem.
In the first quarter of 2026, it topped the revenue list of iOS apps in China, with cumulative iOS in-app purchase revenue exceeding 22 million US dollars, and its revenue efficiency per monthly active user is 4 to 5 times that of DreamAI.
In the first half of 2026, Zeta, which ranked first in revenue among AI companionship/romance applications, also changed its orientation from "character-led" to "story-led" this year.
Zeta also has its own ideas for payment design. The conventional practice is to reserve some functional gaps in the free version to guide users to pay; Zeta opens four advanced tools including Lorebook, character status window, narrative style and plot options to free users.
This design tries to let users continuously accumulate plots, choices and emotional investment in the free experience. When ads suddenly interrupt the experience, or the content generated by the model deviates from the original plot, what users lose is no longer just a certain function, but the story experience that they have spent time building step by step.
Linky, under Kunlun Wanwei, adds collection and gacha mechanisms outside of character interactions. In addition to subscribing to services, users can also obtain content such as character cards through virtual currency, for example, character selfies, or collection materials generated based on interactive plots.
There are other exploration directions for the combination of AI and content consumption, such as the integration of AI and games. Tencent's Epoch Dimension embeds AI companions into game scenarios, deducting fees based on the number of conversation rounds and companionship duration through "star coins exchanging energy", and the charging logic is guaranteed by the game experience itself, rather than building another set of relationship subscription system.
However, this does not mean that the above products are "safe".
The monthly active user scale of Catbox is still at the million level. After the regulatory adjustment in July 2026, some intelligent agent functions on the main Doubao website were taken offline, and the related traffic was diverted to Catbox, bringing a wave of short-term new user acquisition. But the diversion effect did not last, and the user scale gradually fell back afterwards. According to data from the AI Product List, the monthly active users of Catbox in September have fallen back to the previous level.
This shows that traffic diversion can help products quickly acquire new users, but it does not necessarily make users stay for a long time. There is still a long way to go for Catbox to push this model to the broader mass market.
Similarly, although the gacha mechanism may increase users' payment frequency and single consumption opportunities, it will also make the payment experience affected by randomness. Whether users are willing to keep consuming ultimately depends on whether the rewards are attractive enough, whether the price is reasonable, and whether the core interactive experience can remain stable.
No matter it is plot advancement, character collection, or gamification design, these mechanisms can create a "reason to come back", but cannot guarantee that this reason will always be valid. The hooks will wear out, and what really determines whether users stay is whether the content supply can keep up continuously.
Another path is to embed companionship into products with clear functional value.
Tingli Xiong Mooni M1 integrates AI companionship, safety guardianship and audio learning into children's hardware; ElliQ connects payment to the elderly care system through hardware, monthly fees and institutional deployment; AI dolls mostly follow the route of "low-cost hardware plus subscription", using physical form to lower the psychological threshold of first payment.
But hardware cannot eliminate continuous operating costs. Moxie, a $799 children's social robot, lost its core functions because its development company went bankrupt and the cloud service was shut down.
The way out for AI companionship may not lie in making users renew their subscriptions for the relationship, but in finding a more stable value carrier.