Independent applications are simultaneously facing three problems: how long it takes for the product to be absorbed by the model, whether growth can bring profits, and who holds the user entrance.
Written by 丨 Zhu Yingli
Edited by 丨 Zhao Lei
In July this year, an industry insider who has interacted with hundreds of AI start-ups talked to us about the AI application industry, and his first sentence was: "It's completely dead." He then retorted, "What applications do you think are worth doing?"
Applications are certainly valuable, but the traditional software entrepreneurial path has begun to fail: finding demand, acquiring users, and generating revenue are not sufficient for an independent company to exist long-term.
At the beginning of 2026, AI knowledge space product Kuse was gaining tens of thousands of new users every day. However, founder Wu Xian-kun was instead trying to suppress growth.
Losses forced him to do so. Kuse was giving each new user a trial value worth several dollars; the more users there were, the higher the inference fees paid to model and cloud vendors; even if users converted to paid, the cost structure was still poor. The team repeatedly adjusted pricing and payment models, only managing to barely reach a slight positive gross profit.
Some entrepreneurs can find capital market support to cover the losses from high growth; however, Kuse has never raised funds, relying on the investments of Wu Xian-kun and a few co-founders along with operational cash flow, and has no plans to raise funds. In Wu Xian-kun's view, financing does not solve the core issue: if growth merely hands over more revenue to upstream companies, what is the meaning of scale?
Similar problems occur in various AI application companies. Over the past three years, model capabilities have continuously expanded, and the functional innovations made by start-ups are increasingly being incorporated into the models; applications bear the inference costs, and companies selling models come into the field to directly grab users. Independent applications are facing three problems: how long it takes for the product to be absorbed by the model, whether growth can bring profits, and who holds the user entrance.
Some entrepreneurs have left pure software businesses, changing to sell delivery services, enter customers’ operations, or train their own models. They need to answer: when capabilities, costs, and entrances are all in the hands of upstream players, what can application companies retain?
The business of selling tokens: the faster the growth, the more losses incurred
In August 2025, Kuse released a press release stating that in 60 days it had achieved $9 million in ARR (Annual Recurring Revenue) without VC financing and paid advertising. Three months later, Wu Xian-kun declared in an English podcast that ARR had exceeded $10 million.
ARR was once the only "success" marker for AI application companies, and Kuse used to take pride in it. However, when asked again in August this year, Wu Xian-kun began to reflect on this metric: "You are actually selling goods; everyone treats GMV as ARR. If it were an e-commerce sale, $10 million isn’t a particularly large figure." He stated that all AI applications are essentially selling tokens, "You secure a few big clients; annualized, isn't that $10 million?"
Selling tokens and being a distributor for large models grows quickly. In 2024, payment platform Stripe analyzed the top 100 AI companies’ revenues globally: achieving an annualized revenue of $1 million took a median time of only 11.5 months.
Achieving $5 million took approximately 24 months. For comparison, leading SaaS companies in 2018 required 15 months and 37 months to complete the same two steps.
An investor has seen many absurd algorithms for increasing ARR: taking unsustainable one-time revenues or annual fees received in a particular month and multiplying by 12 to write it as annualized revenue; even more exaggerated, taking the highest income figure from one day or week in a year and multiplying by 365 or 52.
Even without inflating the figures, the "annualized revenue" of AI applications is far less reliable than traditional SaaS. The aforementioned investor compared: "In a typical SaaS, a customer paying this month will have a 95% chance of paying again the next month, so it is indeed recurring. Current AI applications are coding plans, and renewal rates are likely to be far lower than traditional SaaS."
The gross profit margin difference is also significant. A research study from the established American venture capital firm Bessemer examined 20 rapidly growing AI companies and found that many approached or achieved $100 million ARR in their first year of commercialization, but the average gross profit was only around 25%; traditional cloud software's mature period usually sees gross profits around 70%.
In the traditional SaaS era, once a product was validated by the market, it could usually scale confidently because the server costs of new users were very low. In the AI era, validation and becoming a sustainable business have become two separate matters: users might be willing to use and pay, but each use incurs inference costs, which do not decline with growth, and gross profits remain low.
Even more challenging, growth sometimes drags the product into losses. Kuse's suppression of growth at the beginning of the year was due to the realization that the costs of trial credits given to free users were too high, once causing the company to incur massive losses.
There are many precedents for this. According to The Information, Perplexity reported a gross profit margin of about 60% for 2024; however, that was because the company accounted for about $33 million spent mainly on free and trial users as research and development expenses; if included in operating costs, the gross profit margin would be negative.
Cursor is similar: its internal financial documents show that the quarterly gross profit margin was -23% as of January 2026; accounting for the inference costs of free users, it was about -31%.
Wu Xian-kun is skeptical that, in such an environment, it is still necessary for resource-limited start-ups to continue playing this game. He observed that productivity tools grew too sharply, with growth-stage gross profits plunging to extremes—down to -200%, some at -40%, and the better ones around 0% to 10%. Kuse barely achieved positive gross profit due to its focus on the business model.
He agrees with Fireworks CEO Lin Qiao's summary of the predicament facing start-ups—"scaling to bankruptcy": the more they grow, the closer they get to death.
When OiiOii was in beta testing at the end of last year, the queue of users reached 100,000; when it just went live, users kept increasing, exceeding 100,000 daily at most. However, the team soon realized that too many people were taking advantage of the system, so they deliberately suppressed growth.
OiiOii's founder Nao Nao stated that initially investors also looked at monthly and daily active users, but the team felt it was meaningless; by April, OiiOii had stopped tracking user activity, shifting focus to paid conversion rates, renewal rates, and average revenue per paid user.
However, the strength of the models and competition among peers has also made earning difficult. In video models, Seedance 2.0 became dominant and held pricing power after its release, offering prices of 1 yuan per second with little room for discounts even at high volumes; typically, application companies’ gross margins on paid users are only around 30%.
The two months following the release of Seedance 2.0 saw the industry's fiercest competition. In order to capture customers, some companies launched aggressive subsidies; when considering the trial credits given to free users and the marketing costs for creators, almost no company in the industry managed to achieve positive gross margin.
Each model upgrade renders a batch of applications redundant
In November 2023, Zhang Jiayuan launched the code search product Devv Search. At that time, programmers were already using ChatGPT to assist in coding, but the answers often had hallucinations and lacked sources. Devv indexes technical resources from GitHub, Stack Overflow, etc., providing citations and verifications for generated results, achieving hundreds of thousands of dollars in revenue within six months of launch.
He initially envisioned the company’s roadmap for the next three to five years: first, spend a year or two enhancing developer search, then develop code generation and debugging, and finally enter the automation stage that eliminates the need for humans to write code.
The pace of model updates quickly rendered this roadmap obsolete. In June 2024, Anthropic released Claude 3.5 Sonnet, claiming it solved 64% of the tasks in internal agent programming tests; with the addition of tools, it could write, modify, and execute code. The connection with Claude model was validated by Cursor.
In February 2025, Anthropic released the preview version of Claude Code, capable of searching, reading code, modifying files, running tests, and submitting code to GitHub. Zhang Jiayuan's initially expected search phase of one to two years lasted only six months; the market jumped straight to the automation development phase. "I didn't expect it to come so quickly."
Devv then pivoted to natural language generation applications. Zhang Jiayuan later determined that the real feasible window for this direction was only between September and November 2024: Cursor, Claude Code, and other programming agents entered from one end, while general agents expanded from the other. To continue acquiring users, a lot of tokens would need to be subsidized, leading to high costs. By mid-2025, Devv 2.0 stopped expanding and entered a maintenance phase.
Earlier AI writing applications enjoyed nearly a year-long window. A founder of a thesis writing product noted that in 2023 the cost of generating a paper was only a few dimes, while selling it was around 70 yuan; his company generated about 10 million yuan that year with peak daily revenues exceeding 100,000 yuan, and a net profit margin of about 50%. The product's market share was once very high; he recalled a time at a hotel when the front desk intern saw the invoice title and smiled knowingly: the thesis was written using this product.
However, in 2024, growth ceased. The team added features like PPT generation and AI content detection without any improvement in revenue. By 2025, Kimi and DeepSeek were able to write lengthy texts even cheaper or for free, causing this tool's revenue to drop from tens of thousands of yuan daily to a few hundred, eventually reaching zero.
Similar companies are also pivoting: Jasper transitioned to enterprise marketing, Copy.ai connected sales processes and customer management systems, "Wa Wa Writing" (Wave Intelligence) was acquired by a mobile phone manufacturer, and the aforementioned thesis product began to undertake magazine connections and paper guidance services from the end of 2024.
When it comes to workflow products, the first change was not even in revenue, but in capital market expectations. In December 2025, an entrepreneur who left a large corporation confirmed after trying browser plugins, AI writing, and open canvases that he would pursue AI workflows and secured over a million dollars in seed investment. These products allow users to describe needs in natural language, with AI automatically breaking them down into work nodes and calling tools to complete them; complex processes can also be made into templates for sale.
However, a month later, Skill began to flourish in agent products and gradually became a standard configuration. After launching this workflow, further financing became difficult as investors believed "Skill would replace Workflow." He could not secure the next round.
At that time, the product had already pivoted to enterprise services, integrating e-commerce, short dramas, and marketing. The founder mentioned that at the peak, there were tens of thousands of dollars in revenue per month, with a renewal rate of 70% to 80%. However, Agent and Skill have been devaluing independent workflow products, and the ceiling is quite visible.
This entrepreneur was pushed to seek a new direction: initially trying out general-purpose agent products, but discovering that the barriers to commercialization were very high, and the market changed too rapidly; he therefore decided to put it on hold. A month later, upon observing the upgrade in Claude Code's design capabilities, he turned to develop open-source design-type agent products. He stated that by the end of 2025, he found it hard to imagine being able to change direction twice in three months; now even if he changes once next week, he would not be surprised.
Investment firm a16z has been continually tracking consumer AI applications; among the lists from 2023 to 2025, only 14 companies made the top 50 each time. Each time a new capability is released by a model, a batch of products is generated, and during the next upgrade, a portion of them has their core value stripped away.
Over the past three years, from writing, translation, search, and code generation to workflows, the list of applications covered by model upgrades is growing longer, and the speed is increasing. According to estimates from the non-profit AI research organization Epoch AI, after April 2024, the pace of frontier model capability enhancement rose from approximately 8 index points per year to about 15 points, almost doubling.
For independent application companies, the absorption by models does not happen instantaneously. When three companies pivoted, their original products had not yet reached zero. Entrepreneurs are not losing current users but future potential; they must run faster to seize the next opportunity.
The people selling models will also come to take your users
In November 2025, Kuse upgraded from a canvas model to a folder structure, even before Codex; in January 2026, the product connected with Slack Bot, also preceding Claude Tag. However, in Wu Xian-kun's view, this short window of a few months cannot accumulate enough user habits, usage frequency, and revenue to allow them to escape the gravity of the model.
He believes the only critical question is: "Does 'model equals application' hold true?"
In March 2025, Manus launched and suddenly became popular; founder Xiao Hong's remark that "the shell has the value of the shell" once negated this proposition, giving confidence to many independent AI applications. However, since 2026, the rapid developments of Claude Code and Codex have swayed judgment to the other end.
Model giants have the capability and motivation to enter the application market themselves, seizing entry product opportunities. The Brookings Institution, a think tank in Washington, analyzed that training and operating models require a massive investment; as model capabilities converge, selling model APIs alone cannot support these investments, and developing applications has become a way for model companies to recapture profits.
There have been multiple incidents of upstream and downstream customers competing for business at Anthropic: after Cursor verified the demand for AI coding, Anthropic launched Claude Code, which gradually became more favored by developers; while deepening its cooperation with Figma, its executives resigned from Figma's board and launched Claude Design, directly entering the core market of the latter.
Model companies can anticipate the capabilities of the next generation of models and can use mass API calls to determine which scenarios are more active. Once a需求 is verified by a start-up, model companies can replicate it as their own product. OpenAI is similar; Deep Research competes with AI search and research tools; ChatGPT Work and Codex enter documents, spreadsheets, presentations, and internal tools; ChatGPT Sites expands into application generation and deployment.
This dual practice of providing infrastructure while entering downstream products is not exclusive to the AI era. Microsoft once bundled Internet Explorer with Windows, later integrating Teams into Office, triggering both US antitrust lawsuits and EU investigations.
Once model giants enter the market with similar products, start-ups have to consider not just functionality and cost structures but whether their ongoing projects are still necessary.
This year, Tencent, Alibaba, and ByteDance collectively entered the office agent field, launching Workbuddy, Qianwen Office, and Doubao Office, respectively. According to data from Analysys on 17 native PC office agents, the traffic in this field doubled from March to June, but two-thirds flowed to these three large companies' products.
An investor who looked into this direction earlier this year found that out of hundreds of projects, only a few persisted until now, with some entrepreneurs even transitioning to become agents for Workbuddy.
This competition is almost a gamble with little chance of winning. Start-ups face not only product differences, but also the free quotas, existing users, and enterprise sales resources of large companies. Taking Workbuddy as an example, the free version’s credits are worth about 25 yuan per month; if daily active users amount to 1 million, the monthly expenditure on free users alone would reach 25 million yuan. In the user acquisition phase, invited new users receive an additional credit worth 100 yuan; for every 1 million new users, the maximum expenditure could soar to 100 million yuan.
In a competitive field among major firms, this amount is trivial, but start-ups cannot afford it. One entrepreneur initially intended to pursue a similar direction but ultimately realized that it was "too big and too mainstream," only capable of fulfilling a large company's ecological niche. Start-ups have to shrink down, focusing on resolving a more vertical issue and becoming an add-on for a large company’s products.
He likened the future of general office agents to browsers, while the plugins they create are akin to web pages, merely waiting to be called upon by entry products.
Escaping competition also means losing users
Faced with the momentum of models and large companies consuming everything, AI application start-up teams easily view “what they won't do and can’t do” as the starting point for direction choosing.
When Wu Xian-kun launched a new project in 2024, he selected the canvas format, believing that the main entrances of model companies are all dialogue-based, while a canvas offers a completely different interaction. By the end of 2024, the team first developed a design product similar to Lovart, subsequently bringing the canvas into office scenarios.
Similar thinking can be found in many contemporaneous products: China's Lovart and Flowith, as well as overseas Miro and Subset, all sought to accomplish tasks that dialogue couldn't achieve through a canvas. Even Baidu launched "Free Canvas," integrating text, PPT, images, audio, video, and AI generation capabilities into the same whiteboard.
However, the starting point of "avoiding large models" brought about new problems. To prove its difference from models, the product tends to complicate interactions, pursuing niche and seemingly unique demands. Every added layer of difference reduces the number of people who can understand and use the product. The team presumed they found a position not entered by model companies, but it turned out to be a location with scarce users.
Kuse spent a long time polishing the PDF reading experience on the canvas: how to arrange files, how to allow users to read while invoking AI. After a period of time, Wu Xian-kun was forced to admit a simpler problem: there may not be that many people willing to read PDFs on a canvas.
A year later, Wu Xian-kun attributed this misstep to the initial intent when choosing the direction. When an entrepreneur focuses first on "avoiding" rather than releasing the capabilities of large models and then searching for demands in empty spaces, the path only grows narrower. Focusing on what model companies can't do does not result in creating a good product.
Cursor and Claude Code offered him another reference: allowing the model to comprehend code repositories and files to complete tasks directly instead of restricting models with new interactions. Kuse subsequently shifted from canvas to folder-based interactions.
Zhang Jiayuan faced similar issues in the several pivots of Devv. There was no clear user for Copilot Hub, and Devv Search was overtaken by model advancements. By the time of Devv 2.0, the team had narrowed its positioning to "help users create AI applications" to differentiate from other natural language generation products, which took about half a year to develop. When the product launched, the opportunity had already passed, leading to low user acquisition and retention.
The failure of Devv 2.0 made him realize that focusing on competitor-designed features might not necessarily align with user needs. When a product moves from 0 to 1, if there are still no stable users and retention is low, the primary questions to answer must be who is using it and why, rather than how it differs from similar products.
Confronting competition could also present a way out. In the latest open-source design agent project, the founder of the aforementioned workflow products saw a possibility: when demand is validated, and large companies invest heavily, perhaps staying in this market is unnecessary—market penetration is still in its early stages, and a different position might be found.
His choice is to develop open-source products in large companies' main battlefield. For example, his current project was initiated after noticing that Claude Code was developing design products, leveraging this enthusiasm to create an open-source version. Within a week, the new project garnered nearly 50,000 stars, with revenues even surpassing previous workflow products. Now over 90% of their resources and manpower are allocated to this matter.
Moving downstream: selling results instead of products
Exploratory entrepreneurs generally follow two paths: one moves deeper into client businesses, transitioning from selling tools to being accountable for results, transactions, and profits; the other moves upstream technologically, creating their own models and training confirmed needs and data into those models.
OiiOii chose the first path. This AI video application company no longer engages in a subsidy war with peers, beginning to enter specific sub-industries, training smaller models post-production to solve issues in specific scenarios, delivering video content and workflows to clients.
Nao Nao mentioned that they have validated multiple vertical industries. For example, in the industry of converting comics into animations, Seedance 2.0's 2D effects are inadequate, as it overfits real or 3D, making 2D appear overly smooth, which seems unnatural in Japanese animations. "We utilized a method of training that made the 2D presentation resemble Japanese animations closely, and clients were extremely satisfied and willing to pay."
By solving actual problems, the value of their delivery has significantly increased. Nao Nao compared that if the deliverables for clients were created using tools, producing a two-minute video could cost merely 200 yuan, but providing a compliant video could be priced at 10,000 yuan for two minutes. "Just think, how many times more is that?"
The past enterprise service business often failed because deliveries were too hefty and human costs too high. AI has resolved this issue; as long as the production tools are good enough, the number of users needed is minimal. Nao Nao stated that products originally meant to outwardly showcase usage value have evolved into an efficient internal production system.
The validated solutions in vertical industries will subsequently accumulate into skills that can be provided to more users—this is also OiiOii's upcoming product direction.
The previously mentioned thesis writing tool team has not created new consumer tools, but returned to the accumulated 5 million student users: undertaking customized requests for thesis publication and conducting AI training, teaching students how to write copy, edit, and operate content, becoming a "one-person company."
This service online accompanies students for several thousand yuan in fees, while offline courses cost 10,000 to 20,000 yuan. The founder refers to it as "heavy delivery": clients are paying for the results from learning tools to generating content and earning money. He believes that students nowadays have difficulty finding jobs, and compared to the nebulous demand for "releasing creativity," most people's urgent need is to earn money using new productivity tools, while their accumulated experience in social media and productivity tools perfectly aligns with this requirement.
Wu Xian-kun delves deeper. Kuse previously created "digital employee" Junior for enterprises, but each company's data, permissions, and processes differ, necessitating diagnosis, consulting, and deployment first; once the system is ready, clients may not necessarily be willing to share sensitive data with external vendors. Projects do have demands, but replicating them like standard software is challenging.
He later chose to explore another method: collaborating with investment institutions to first acquire service companies and then use AI to reform processes, sharing profits through equity stakes. This approach can penetrate deeper into business than FDE and aligns the interests of both parties. Wu Xian-kun stated that the team has acquired a company.
None of these paths have been fully navigated yet. Nao Nao's method of selling deliverables theoretically addresses immediate profit margins and cash flow, but long-term, they still have to tackle the most challenging aspect of B2B: how to continuously expand their customer base. AI training aimed at students has not yet demonstrated through data that it can help participants earn money on a scalable level. Wu Xian-kun has immersed himself deeper into operations, but efficiency improvements and integration could ultimately fail.
However, under the prevailing narratives of "model equals product" and "models consume everything," these methods at least carry a higher possibility of survival for now.
Moving upstream: creating your own models
Another segment of people chose the opposite direction, moving upstream to create their own models. Cursor, Krea, Captions, Perplexity, HeyGen, and LiblibAI have all transitioned from applications into the model layer, primarily along verified routes in programming, search, design, imagery, video, and virtual human scenarios, training more vertical models.
Yang Bolin's team initially developed Cuflow, transforming user-uploaded course materials into interactive videos. He mentioned that during public testing, tens of thousands registered, achieving over 20 million exposures overseas; however, more than half of the uploaded materials were not academic but consisted of corporate training materials, comics, novels, and product detail pages. In April 2026, the team stopped updating Cuflow and redirected their focus to training a real-time interactive video model centered on code.
This decision stemmed from not only user observations but also a sense of insecurity accumulated during their application phase. Cuflow had to integrate language, speech, and video models from different companies; with every upstream update, engineers needed to retest capabilities and costs before deciding which functionalities required remodification. Yang Bolin expressed that application companies, to some extent, have become "model evaluation bloggers," testing new models immediately after their release, only to find themselves adapting again shortly after receiving an update, without prior notice.
What was even harder to accept was that some updates could completely erase several months of engineering effort. In the early days, they had to process lengthy documents by first performing text recognition and segmenting the text, then sending images and text to the models for search; after the context lengths of the models grew, directly inputting the entire material sometimes yielded better results. The old processes became not only redundant but slowed down speed.
From this, he felt that AI applications, seemingly under their control, ultimately possess their most crucial abilities outside of their grasp: "We are always just tenants."
This situation is not unmanageable. Application companies can bind client workflows, and strengthen brands and channels. However, Yang Bolin believes that these efforts do not leverage the team's strengths: they have algorithm engineers and model training talent, and if most of the time is spent merely testing and switching others' models, this advantage is wasted.
Returning to the model layer represents a choice with "higher odds of winning" for him. There have been many teams originally engaged in AI education; after pivoting to real-time interactive videos based on diffusion language models, direct competition faced fewer teams. More importantly, he believes he holds information that others have not yet fully verified: the cost of generating video from code is low, speeds are fast, and it can accurately edit and audit, naturally supporting interactivity and multi-person collaboration, making it suitable for corporate training and product demonstrations—primarily where information needs to be conveyed in video.
As the team trains models, they also discuss needs with clients from industries such as education, marketing, and corporate training. Two to three months after pivoting, they already secured interest in cooperation from 21 enterprise clients and plan to understand requirements from the client's perspective for targeted training.
Yang Bolin believes their true advantage lies not in developing models themselves but in their ability to continually reinforce one another between scenarios and models: first identifying truly valuable scenes that can continuously produce data, then understanding those scenes and embedding that understanding into the models through follow-up training and feedback.
The cost of solely developing models is also very clear. With a team of 25, building processes from supervised fine-tuning, preference optimization to reinforcement learning for follow-up training pipelines requires managing reward functions, containerization, and GPU scheduling— a single misconfiguration can incur significant losses.
According to Yang Bolin's estimates, subsequent training requires several hundred GPUs. The company is currently undergoing another round of financing, with a substantial portion allocated for training resources. Creating models has not eliminated commercialization issues. While 21 clients signed intentions, there are still few actual paying customers.
Regardless of whether focusing on delivery or follow-up training models, entrepreneurs must continue moving forward, for they have already secured funds, built teams, and gained exposure; they also believe that as long as AI continues to develop, model giants cannot handle everything.
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。
