Three and a half years after the explosion of generative AI, the market has reached a new point of divergence: optimism is still accelerating, while skepticism is also accumulating. Judging whether a “bubble” is coming is not enough to explain the current complexity. The “AI Belief and Bubble” series will look for key variables from different perspectives of the market, technology, industry, and companies.
In the first half of 2026, a batch of AI applications that were once highly favored by capital are gradually withdrawing from the market. It’s not just cash – strapped startup teams; giants like OpenAI and Google are also actively scaling back their previously expanded product lines.
In March 2026, OpenAI announced its plan to discontinue the Sora video generator, which had been launched only half a year ago. This application, which featured a “social – like” experience and once topped the Apple App Store, was ultimately axed due to a continuous decline in downloads and the huge daily consumption of computing power.
In the same month, the AI model evaluation platform Yupp.ai announced its shutdown. It was led by Chris Dixon of a16z crypto and raised $33 million in financing. It had amassed 1.3 million users in less than a year but failed to find a strong enough product – market fit. The founder admitted that as the model capabilities improved rapidly and users’ work shifted to agentic systems that could call tools and memories, crowdsourced evaluation based on the chat layer was becoming increasingly unimportant. The site will be retained until April 15th for users to export their historical data.
In addition, Google has also started to scale back its internal AI application lines. In June, the core image – generation function of Pixel Studio was shut down in the v2.3 update, and users were redirected to Gemini and Nano Banana. The browser Agent experimental project, Project Mariner, was closed on May 4th, and its capabilities were integrated into larger product systems such as Gemini Agent and AI Mode.
The AI application layer is moving from the early stage of function testing to a more brutal commercial screening period.
Much of this market clearing is happening to application – layer products “built on single – point model capabilities”: some are internal function integrations within large companies, some are commercialization failures of startups, and some are experimental projects being integrated into larger platforms. Although they have different forms, they all expose the same problem: when the underlying models continue to upgrade, has the application layer formed a thick enough independent value?
Darren Mowry, the global head of Google Cloud’s startup business, said in an interview with TechCrunch: If a startup mainly relies on the backend model to complete its work, this form is almost like white – labeling Gemini or GPT – 5, and the industry has little patience for it.
The so – called “white – labeling” means repackaging others’ model capabilities with one’s own interface and brand: users see a new application, but the core capabilities are still supported by leading large models such as Gemini, GPT, or Claude.
Those applications that are only propped up by model dividends are losing the reason to exist independently.
Now that the storm has landed and the underlying models are constantly becoming more accessible, where should the moat of the application layer be?
The picture is generated by an AI tool.
The downfall of many application – layer companies doesn’t mean they had no value from the start. The problem is that their value was established at a stage when the models were not good enough, users were not mature enough, and scenarios needed to be repackaged. Once the model capabilities reach the user entry point, this part of the value will be quickly re – evaluated.
Jasper AI was one of the first companies to be hit by this logic. It was once a star in the AI writing application field, relying on GPT – 3 to automatically generate creative marketing content and quickly becoming a unicorn with a valuation of up to $1.5 billion. However, with the popularization of ChatGPT, “generating marketing copy” quickly changed from the core selling point of an independent application to a basic ability of large models. Later, Jasper went through layoffs, a valuation adjustment, a leadership change, and shifted its focus to enterprise marketing workflows.
A similar story also happened to Chegg.
Chegg is an online education company that has been severely impacted by AI tools such as ChatGPT and Google AI Overviews. In the first quarter of 2026, its revenue was $63.3 million, a 48% year – on – year decline. Chegg then laid off employees, its revenue dropped, and it shifted its focus to AI and vocational skills business.
Dan Rosensweig, the CEO of Chegg, publicly admitted that the increasing interest of students in ChatGPT has affected the company’s new user growth. Users didn’t find another Chegg; instead, they directly shifted their needs to ChatGPT. For the application layer, the most dangerous substitute is often not a peer but the underlying model suddenly becoming the user entry point.
In the past, there was a wide gap between the original capabilities of models and the real needs of end – users: models were powerful but difficult to use, select, and implement; users had needs but didn’t understand models, didn’t know how to adjust parameters, and were reluctant to bear the cost of trial and error.
The value of the application layer lies in translating “what the model can do” into “what you can use it for” and charging for this “translation”. The wider this gap, the greater its profit margin.
However, the reality is that this space is being infinitely compressed.
An entrepreneur in the large – model application layer said: “Now, upstream model manufacturers are also getting involved in the application layer. This gap is being filled from both ends. In addition, downstream enterprise customers are also maturing rapidly. With the widespread use of large AI models, a round of market education has been completed, and enterprises are clear about the main core functions. More importantly, there are more and more optional suppliers.”
Upstream model manufacturers have core capabilities and can easily integrate them as native functions. Downstream customers are becoming more knowledgeable, starting to push for lower prices, better results, and higher ROI. There are also countless substitutes, from ChatGPT, Gemini, Copilot to cloud providers and office software. To make matters worse, new competitors can enter the market at any time.
Therefore, the application layer in the middle is changing from an “amplifier of technological dividends” to a “hard – hit area for value proof”.
On the other side of the series of shutdowns of Sora, Yupp.ai, and Pixel Studio is a still – booming market. According to Sensor Tower data, in 2025, the downloads of generative AI applications doubled year – on – year to 3.8 billion, and in – app purchase revenue nearly tripled, exceeding $5 billion. Sensor Tower also predicts that by 2026, the revenue of generative AI applications is expected to exceed $10 billion. That is to say, both money and users are there; it’s not the industry that is collapsing, but mainly a batch of products that “stood in the wrong position”.
So, what exactly did the application – layer products that have truly survived and even thrived do right?
By looking at the sixth – edition list of generative AI consumer applications released by a16z in March 2026, we can find that the product forms of the truly successful AI application layer have changed. There are mainly three core types:
The first type of application is the super – application that becomes the default entry point.
For example, horizontal AI products like ChatGPT, Gemini, and Claude are no longer traditional tools; they are all vying for the AI entry point. Users regard these tools as new workbenches: asking questions, searching for information, writing code, creating spreadsheets, connecting calendars, accessing emails, and invoking external applications. a16z specifically mentioned that both ChatGPT and Claude are building connector and app ecosystems. When a user connects their email, calendar, CRM, documents, and work software to an AI assistant, the switching cost will rise rapidly.
The second type of application is the one that originally occupies high – frequency or vertical scenarios.
Take CapCut as an example. As a video – editing tool with over 800 million monthly active users, some of its most popular functions, such as background removal, AI special effects, automatic subtitles, and text – to – video conversion, are all AI – driven. However, users don’t come for “AI”; they value the video – editing function itself, and AI just makes operations that originally took ten minutes become one – click operations.
There is also Notion AI, which integrates AI into enterprise knowledge bases, project management, meeting records, and automation processes. That’s why the paid penetration rate of Notion AI can increase rapidly: users are not just buying a new tool; they are paying for a more efficient way of working in a system they can’t do without.
The third type of surviving product has evolved from a tool to an “Agent that does things for users”.
a16z specifically emphasized in this list that Agents have begun to emerge. For example, Lovable, Cursor, Bolt, Replit, and Claude Code represent agentic behaviors in the development scenario: they have started to help users build products, modify code, analyze projects, and advance tasks. Horizontal Agents like Manus and Genspark allow users to assign more open – ended tasks, such as research, spreadsheet analysis, and slide generation, and the AI completes an end – to – end workflow.
Although these types of products have different forms, the logic for their survival is the same: they don’t just rely on “having AI” to acquire customers; the core is to integrate AI into the entry points, scenarios, and tasks that users can’t do without.
Therefore, when we discuss the closures and contractions of the AI application layer today, we can’t simply understand it as “the AI application layer is shrinking”.
What is really exiting the market is a batch of light applications that package single – point functions into independent products. What continues to expand are the application – layer products that are embedded in high – frequency scenarios, occupy user entry points, and enter real – world workflows. They may not appear in the name of “AI applications”, but the core is that AI has long been integrated and has become part of different product forms such as video – editing software, office suites, browsers, and design tools.
The era when a single – point function could be independently monetized is gone.
The story of the AI application layer continues.
In the upcoming Agent era, the threshold of the application layer will be further raised: in the future, single – point functions will be even more insufficient. Whether a truly valuable product can enter the process, connect systems, take on responsibilities, and be safe and controllable, turning model capabilities into an executable, traceable, and measurable business closed – loop.
In the developer scenario, we can see that this change has advanced a great deal. Tools like OpenAI Codex and Claude Code are pushing AI from “code completion” to “software development agency”. These Agents start to understand code repositories, modify files, troubleshoot errors, generate tests, and even continuously advance around a development task.
Such capabilities are difficult to be directly replaced by a general chat box. Real – world software development requires continuous judgment, modification, verification, and delivery in a complex engineering system. Once a product is used for a long enough time, it will accumulate project context, team habits, historical problems, and operation records, and become more and more closely tied to users’ daily work.
This screening will continue. The products that have survived today still have to constantly answer the same questions. Who will still be on the field after the tide recedes? The answer will be given by time.
This article is from the WeChat official account “Tencent Technology”. Author: Li Hailun, Editor: Xu Qingyang. Republished by 36Kr with permission.
该文观点仅代表作者本人，36氪平台仅提供信息存储空间服务。
36kr Europe (eu.36kr.com) delivers global business and markets news, data, analysis, and video to the world, dedicated to building value and providing business service for companies’ global expansion.
© 2024 36kr.com. All rights reserved.
Leave a Reply