In 2026, why does capital favor "profit-making" AI applications more?

Ahead of this year's WAIC, Moonshot AI released Kimi K3, which instantly became a viral hit. However, the attention of the capital market has largely shifted to another development.

Over the past six months, valuation of this star large model company has surged 6 times, targeting $30 billion, while simultaneously advancing its Hong Kong IPO plan. In the all-staff letter delivered on New Year's Eve at the end of 2025, founder Yang Zhilin wrote a line: In 2026, we will focus on Agent, and will not take absolute user count as our goal.

While its valuation sextupled in half a year, the company has also voluntarily abandoned DAU, the metric that the capital market had been closely tracking over the past two years.

This is sufficient proof that while user scale remains important, it is no longer the sole hard currency for AI application valuation. The new core narrative has returned to the long-term sustainable commercialization of product forms.

The capital market's benchmark for evaluating AI has changed: Over the past two years, the prevailing approach was to trade computing power subsidies for scale, but book DAU can only prop up valuation, not gross profit. The capital market has realized that the land-grabbing strategy of the internet era does not work in AI. Computing power costs and retention curves have "shown their true colors" one after another, and the once highly anticipated hit products have begun to exit the market in batches.

The industry inflection point has thus arrived. Capital no longer pays for scale itself, but shifts to verifying the quality of commercialization.

Under this new logic, companies with robust business models have gained more opportunities to stay in the market as "good money" and secure better valuations. For example, Haiyi, which recently completed a Series B financing of over 100 million yuan, is a stable enterprise favored by capital: based in Chengdu, targeting the global market, taking "born for C-end users" as its business philosophy, it has continuously launched hit AI applications in the past few years, and its financial data is already quite impressive.

After the tide recedes, a more thought-provoking question arises: What kind of AI applications can "make money with dignity"?

Commercialization Quality Verification: The Inflection Point Has Arrived

The mainstream narrative in the AI application layer over the past two years was that capital and enterprises joined forces to burn money faster than Silicon Valley, prioritizing scale, providing computing power subsidies, clearing out competitors first, and leaving the gross profit problem "for later". This strategy worked in a period of loose capital, but it began to show signs of fatigue in the second half of 2025.

With persistently high computing power costs, rising customer acquisition costs, and generally poor retention data, VCs can no longer afford the money-burning strategy, so the valuation logic naturally switches to commercialization verification.

However, the commercialization efforts of mainstream large models and AI applications have never stopped in the past few years, and the solutions vary.

Stable subscription-based computing power sales is a very common path, which ChatGPT Plus, Claude Pro, and Kimi are all following, with API billing targeting individual users and developers. The advertising monetization path has been adopted by some high-frequency lightweight applications, while the credit point system of AI creation tools is the mainstream model for platforms like Midjourney. Every path is being tested by someone, and phased results have been achieved.

The commercialization boom has lasted for two years. There are always people making money, but very few people have been making money consistently.

A poignant example is Character.AI, the global benchmark in the AI character interaction track. According to data from Business of Apps, its peak monthly active users reached 28 million in August 2024, but the number did not increase in the following months; instead, it decreased, losing about 8 million monthly active users by January 2025. Also in 2024, after Google signed an agreement to acquire the two founders and obtain a non-exclusive license for its large model, the company stopped developing its own model.

Inflection AI followed a similar trajectory: it raised $1.3 billion, with a valuation of $4 billion at one point. Later, its product Pi faced traffic restrictions, and its founder left for Microsoft to shift to enterprise services.

The common point of these two star companies is that they once had enviable user scale and financing figures, but failed to pass the test of computing power costs and retention curves.

Therefore, everyone is working on commercialization, and the industry has basically reached a consensus in 2026: the development of AI cannot be judged solely by commercialization success, and monthly active users and payment figures are far from the end point. In the face of high computing power costs, high-quality commercialization is actually a very difficult task.

So for AI, what counts as high-quality commercialization?

No one has seriously answered this question in the past two years, because everyone was busy seizing market share, but now it has reached a point where the question must be answered. The answer is actually not complicated: whether users are willing to pay continuously for real value, and whether a combination of positive gross profit and high retention can be formed.

Gross profit margin answers the question "can this business cover computing power and customer acquisition costs", while retention and renewal rates answer "will users stay after experiencing the paid product". The former determines whether the business model is viable, and the latter determines whether the product's value is truly effective.

These two metrics better illustrate whether an AI application company has long-term self-growth potential and risk resistance capabilities than DAU and the number of paying users.

The public data released by Haiyi shows that its overall gross profit margin exceeds 40%, ARPPU (average revenue per paying user) is about $60, and the renewal rate of its core products has exceeded 60%. In the global AI ToC application market, most enterprises are still trapped in the cycle of losing money in exchange for scale, and very few companies in the application layer can reach this level in both gross profit and retention dimensions.

Haiyi is not a new company, nor is it operating in a new track. It started in Chengdu in 2023, with a globally native organization, and began as a multimodal AI creation community. Its product forms are familiar: generating images, training LoRA, making videos, and interacting with characters.

However, it is this seemingly "ordinary and familiar" business that has a particularly impressive financial structure amid this inflection point of commercialization quality.

Three Growth Curves, One Core Logic

High renewal rates and high ARPPU are prerequisites, and positive gross profit is the result. Behind the financial data, the real concern of capital is: how did Haiyi achieve this, and can it maintain this performance in the long run?

High-quality commercialization essentially comes down to the right product combined with the right business logic. Therefore, to analyze an application layer company, we need to go back and examine what products it has developed and what logical closed loops exist between these products.

Haiyi's product matrix currently includes three growth curves. The first is its original AI multimodal creation community SeaArt, which has accumulated 2 million AI-native creative assets over 36 months, covering models, LoRA, workflows, and templates.

The second is the AI short drama platform MoreShort, whose monthly revenue exceeded $1 million just 6 months after its launch.

The third is the AI character interaction product SeaSoul, which reached 500,000 daily active users in 4 months, with an average daily online duration of over 66 minutes per user.

It can be seen that the three curves have completely different forms: a creation community, short drama consumption, and character interaction. But their common point is also obvious: in fiercely competitive global sectors where giants and startups are frantically expanding their territory, they have quickly delivered results in terms of AI assets and revenue with extremely fast operational speed.

Standing in such a "battlefield-level" track, a company can continuously achieve rapid launches at the application layer, successfully run global operations with consecutive impressive data, and create hit products, which relies on two indispensable structural foundations.

The more visible layer is capability interlocking. Haiyi has broken down its product system into five capabilities: multimodal generation, agent interaction, content distribution and recommendation, paid user acquisition and growth, and commercial monetization.

Haiyi first validated the business logic from content generation, user retention to paid monetization in its creation community, then abstracted key links into reusable capability modules that can be applied to different product forms.

Conversely, the three product lines are also continuously deepening this capability system: the creation community scenario accumulates character assets and generation models, MoreShort trains distribution efficiency and monetization paths in the short drama consumption scenario, and SeaSoul polishes agent dialogue and retention in the character interaction scenario.

Under this positive cycle, the verification cycle for new products has been greatly shortened: SeaArt took 36 months, MoreShort reduced it to 6 months, and SeaSoul only took 4 months. The hit product effect brought by the same capability structure has been replicated across different product forms.

The reason why most AI applications cannot validate high-quality commercialization models is precisely because their capability structures are incomplete. For example, some companies excel at model generation but are not good at distribution and paid user acquisition, while others achieve rapid growth but cannot predict which content growth can directly correspond to commercial monetization. In the best case scenario, the business logic of a single-point application is validated, but it cannot be quickly replicated to the next product, missing the market window.

Looking deeper, the strategic logic is embodied in a complete user path: characters are the entry point, stories are for consumption, and interactions build relationships.

Users first enter the product through characters, generate content payments during story consumption, and settle into retention through interactive relationships. Therefore, the three product lines seem to be expansions of demands in different scenarios, but there is only one underlying user behavior model behind them.

The underlying logic of AI applications is very different from the land-grabbing strategy of the internet era: the internet logic is to acquire users first and then monetize, with scale itself being a barrier because the marginal cost tends to zero. However, the computing power cost of AI applications is ongoing, with each generation requiring payment for computing resources. The larger the scale, the higher the cost.

This is why the computing power cost and customer acquisition cost of the AI application layer are easily stuck at the stage when a company just becomes a star in the capital market. A common phenomenon is that the more paying users there are, the higher the computing power cost. After scaling up, the company cannot afford the load, user experience continues to decline, and eventually users are lost.

The ideal of trading computing power subsidies for scale seems beautiful, but it cannot withstand the long-term test of commercialization quality. The companies that can truly succeed are those that can structure their capabilities and clearly calculate the unit economic model.

How the Ecological Flywheel Starts to Rotate

Where money flows in from and where it goes out is the most direct way to judge a company's development direction.

Haiyi, which has just completed its Series B financing, was jointly led by Visual China Group, Huagai Chuangying, and Vertex Ventures this round, with participation from GF Xinde, Tian Investment Capital, Sichuan Venture Capital, Guangzhou Hewei Yongsheng, and other institutions. Its list of historical shareholders also includes well-known industrial funds and institutions such as Alibaba, Tencent, CMBC International, Zhongding Capital, Actoz, Shanghai Artificial Intelligence Industry Series Funds, and Heying Capital.

Breaking down the list, the demands of the three types of investors are different, but they correspond to several criteria that AI applications need to meet simultaneously.

Investment institutions focus on growth curves, ecological differentiation, and the ability to continuously validate commercialization, investing for financial returns that can survive market cycles.

Industrial parties pay more attention to upstream and downstream ecological collaboration. For example, Visual China Group, a global A-share listed visual content copyright service platform, aims to secure content supply and B-end commercial scenarios in the AI era. Local government funds focus more on industrial implementation and building global AI application benchmarks.

The combination of "capital + industry + government" appearing in the same round of financing list proves that Haiyi can deliver all three things: hard data on commercialization quality, extensibility of industrial ecological collaboration, and the ability to drive regional industrial development.

Haiyi's global user scale and content assets, as its "foundation", can precisely meet these requirements: cumulative registered users reach 65 million, monthly visits exceed 30 million, overseas users account for over 90%, covering core markets such as Japan, the United States, Brazil, and Russia; users generate over 10 million images and 500,000 videos per day, which forms the foundation for all commercialization stories.

Haiyi's content assets cover models, LoRA, workflows, templates, and Agents, making it one of the particularly large-scale AI creation asset libraries in the world. Creators' content continuously enriches the platform ecosystem, and the improvement of platform growth and commercialization capabilities can in turn bring more distribution and monetization opportunities to creators.

In addition to its core ToC ecosystem, Haiyi is also opening up model services, AI asset management, Agent capabilities, and global advertising infrastructure to B-end, P-end, and OPC (one-person company) teams, connecting more external co-creators beyond its own ecosystem.

Creators can complete the full workflow from character setting, visual generation, story production to commercial distribution on a unified infrastructure. Independent creators and small teams can also do things that previously could only be accomplished by strong IPs and large companies.

With the content industry support from Visual China Group, Haiyi will accelerate compliance and channel expansion for digital content in extended scenarios such as advertising, marketing, publishing, and film & television, which are hard thresholds for the globalization of AI products.

The industrial chain rooted in Chengdu is another ecological foothold for this global company. Haiyi has partnered with Sichuan Tianfu New Area to launch the "Tianfu T·OPC Xinghai Entrepreneurship Camp", providing computing power, tools, and policy support for AIGC creators, nurturing the one-person company ecosystem, and helping lightweight teams complete the production and commercialization of AI-native content.

No one can currently give a clear answer to what the endgame of AI content consumption will look like. It may be an undefined product form, or it may reorganize the relationships among creation, consumption, and social interaction in a way we have never imagined, integrated with mainstream terminals.

Haiyi, which is riding the wave, is answering a more macroscopic and long-term question: In 2026, can an AI application company operate without burning money, achieve positive gross profit, and make users pay continuously for real value?

The answer has already been validated.

The remaining task is to continue expanding the boundaries of this hit product creation system and ecological collaboration, generating profits independently while enabling the capital market, industrial parties, government