Three and a half years after the explosion of generative AI, the market has reached a brand new level of divergence: optimism continues to be accelerating, whereas skepticism can be accumulating. Judging whether or not a “bubble” is coming just isn’t sufficient to clarify the present complexity. The “AI Perception and Bubble” sequence will search for key variables from completely different views of the market, expertise, business, and corporations.
Within the first half of 2026, a batch of AI purposes that have been as soon as extremely favored by capital are steadily withdrawing from the market. It isn’t simply money – strapped startup groups; giants like OpenAI and Google are additionally actively scaling again their beforehand expanded product strains.
In March 2026, OpenAI introduced its plan to discontinue the Sora video generator, which had been launched solely half a 12 months in the past. This utility, which featured a “social – like” expertise and as soon as topped the Apple App Retailer, was in the end axed attributable to a steady decline in downloads and the massive every day consumption of computing energy.
In the identical month, the AI mannequin analysis platform Yupp.ai introduced its shutdown. It was led by Chris Dixon of a16z crypto and raised $33 million in financing. It had amassed 1.3 million customers in lower than a 12 months however did not discover a robust sufficient product – market match. The founder admitted that because the mannequin capabilities improved quickly and customers’ work shifted to agentic methods that would name instruments and recollections, crowdsourced analysis primarily based on the chat layer was changing into more and more unimportant. The positioning will probably be retained till April fifteenth for customers to export their historic knowledge.
As well as, Google has additionally began to reduce its inside AI utility strains. In June, the core picture – technology perform of Pixel Studio was shut down within the v2.3 replace, and customers have been redirected to Gemini and Nano Banana. The browser Agent experimental challenge, Undertaking Mariner, was closed on Might 4th, and its capabilities have been built-in into bigger product methods akin to Gemini Agent and AI Mode.
The AI utility layer is shifting from the early stage of perform testing to a extra brutal industrial screening interval.
A lot of this market clearing is going on to utility – layer merchandise “constructed on single – level mannequin capabilities”: some are inside perform integrations inside giant firms, some are commercialization failures of startups, and a few are experimental initiatives being built-in into bigger platforms. Though they’ve completely different types, all of them expose the identical drawback: when the underlying fashions proceed to improve, has the applying layer fashioned a thick sufficient impartial worth?
Darren Mowry, the worldwide head of Google Cloud’s startup enterprise, mentioned in an interview with TechCrunch: If a startup primarily depends on the backend mannequin to finish its work, this type is nearly like white – labeling Gemini or GPT – 5, and the business has little persistence for it.
The so – referred to as “white – labeling” means repackaging others’ mannequin capabilities with one’s personal interface and model: customers see a brand new utility, however the core capabilities are nonetheless supported by main giant fashions akin to Gemini, GPT, or Claude.
These purposes which can be solely propped up by mannequin dividends are shedding the rationale to exist independently.
Now that the storm has landed and the underlying fashions are consistently changing into extra accessible, the place ought to the moat of the applying layer be?
The image is generated by an AI instrument.
After the big – mannequin capabilities develop into extra accessible, the applying layer is being re – priced
The downfall of many utility – layer firms does not imply that they had no worth from the beginning. The issue is that their worth was established at a stage when the fashions weren’t ok, customers weren’t mature sufficient, and eventualities wanted to be repackaged. As soon as the mannequin capabilities attain the person entry level, this a part of the worth will probably be shortly re – evaluated.
Jasper AI was one of many first firms to be hit by this logic. It was as soon as a star within the AI writing utility subject, counting on GPT – 3 to routinely generate inventive advertising and marketing content material and shortly changing into a unicorn with a valuation of as much as $1.5 billion. Nevertheless, with the popularization of ChatGPT, “producing advertising and marketing copy” shortly modified from the core promoting level of an impartial utility to a fundamental skill of huge fashions. Later, Jasper went via layoffs, a valuation adjustment, a management change, and shifted its focus to enterprise advertising and marketing workflows.
The same story additionally occurred to Chegg.
Chegg is a web based schooling firm that has been severely impacted by AI instruments akin to ChatGPT and Google AI Overviews. Within the first quarter of 2026, its income was $63.3 million, a 48% 12 months – on – 12 months decline. Chegg then laid off staff, its income dropped, and it shifted its focus to AI and vocational expertise enterprise.
Dan Rosensweig, the CEO of Chegg, publicly admitted that the rising curiosity of scholars in ChatGPT has affected the corporate’s new person development. Customers did not discover one other Chegg; as an alternative, they instantly shifted their must ChatGPT. For the applying layer, probably the most harmful substitute is commonly not a peer however the underlying mannequin abruptly changing into the person entry level.
Prior to now, there was a large hole between the unique capabilities of fashions and the true wants of finish – customers: fashions have been highly effective however troublesome to make use of, choose, and implement; customers had wants however did not perceive fashions, did not know the right way to modify parameters, and have been reluctant to bear the price of trial and error.
The worth of the applying layer lies in translating “what the mannequin can do” into “what you should use it for” and charging for this “translation”. The broader this hole, the higher its revenue margin.
Nevertheless, the fact is that this area is being infinitely compressed.
An entrepreneur within the giant – mannequin utility layer mentioned: “Now, upstream mannequin producers are additionally getting concerned within the utility layer. This hole is being crammed from each ends. As well as, downstream enterprise prospects are additionally maturing quickly. With the widespread use of huge AI fashions, a spherical of market schooling has been accomplished, and enterprises are clear about the principle core capabilities. Extra importantly, there are increasingly non-obligatory suppliers.”
Upstream mannequin producers have core capabilities and may simply combine them as native capabilities. Downstream prospects have gotten extra educated, beginning to push for decrease costs, higher outcomes, and better ROI. There are additionally numerous substitutes, from ChatGPT, Gemini, Copilot to cloud suppliers and workplace software program. To make issues worse, new rivals can enter the market at any time.
Due to this fact, the applying layer within the center is altering from an “amplifier of technological dividends” to a “exhausting – hit space for worth proof”.
The surviving utility – layer merchandise should not simply “promoting AI alone”
On the opposite aspect of the sequence of shutdowns of Sora, Yupp.ai, and Pixel Studio is a nonetheless – booming market. Based on Sensor Tower knowledge, in 2025, the downloads of generative AI purposes doubled 12 months – on – 12 months to three.8 billion, and in – app buy income practically tripled, exceeding $5 billion. Sensor Tower additionally predicts that by 2026, the income of generative AI purposes is anticipated to exceed $10 billion. That’s to say, each cash and customers are there; it is not the business that’s collapsing, however primarily a batch of merchandise that “stood within the incorrect place”.
So, what precisely did the applying – layer merchandise which have really survived and even thrived do proper?
By wanting on the sixth – version listing of generative AI shopper purposes launched by a16z in March 2026, we will discover that the product types of the really profitable AI utility layer have modified. There are primarily three core varieties:
The primary kind of utility is the tremendous – utility that turns into the default entry level.
For instance, horizontal AI merchandise like ChatGPT, Gemini, and Claude are not conventional instruments; they’re all vying for the AI entry level. Customers regard these instruments as new workbenches: asking questions, looking for info, writing code, creating spreadsheets, connecting calendars, accessing emails, and invoking exterior purposes. a16z particularly talked about that each ChatGPT and Claude are constructing connector and app ecosystems. When a person connects their e-mail, calendar, CRM, paperwork, and work software program to an AI assistant, the switching value will rise quickly.
The second kind of utility is the one which initially occupies excessive – frequency or vertical eventualities.
Take CapCut for example. As a video – modifying instrument with over 800 million month-to-month lively customers, a few of its hottest capabilities, akin to background removing, AI particular results, computerized subtitles, and textual content – to – video conversion, are all AI – pushed. Nevertheless, customers do not come for “AI”; they worth the video – modifying perform itself, and AI simply makes operations that initially took ten minutes develop into one – click on operations.
There’s additionally Notion AI, which integrates AI into enterprise data bases, challenge administration, assembly data, and automation processes. That is why the paid penetration fee of Notion AI can improve quickly: customers should not simply shopping for a brand new instrument; they’re paying for a extra environment friendly approach of working in a system they cannot do with out.
The third kind of surviving product has developed from a instrument to an “Agent that does issues for customers”.
a16z particularly emphasised on this listing that Brokers have begun to emerge. For instance, Lovable, Cursor, Bolt, Replit, and Claude Code symbolize agentic behaviors within the growth state of affairs: they’ve began to assist customers construct merchandise, modify code, analyze initiatives, and advance duties. Horizontal Brokers like Manus and Genspark enable customers to assign extra open – ended duties, akin to analysis, spreadsheet evaluation, and slide technology, and the AI completes an finish – to – finish workflow.
Though all these merchandise have completely different types, the logic for his or her survival is identical: they do not simply depend on “having AI” to amass prospects; the core is to combine AI into the entry factors, eventualities, and duties that customers cannot do with out.
The AI utility layer just isn’t shrinking; it is simply that the brink has develop into greater
Due to this fact, once we focus on the closures and contractions of the AI utility layer right now, we won’t merely perceive it as “the AI utility layer is shrinking”.
What is actually exiting the market is a batch of sunshine purposes that bundle single – level capabilities into impartial merchandise. What continues to broaden are the applying – layer merchandise which can be embedded in excessive – frequency eventualities, occupy person entry factors, and enter actual – world workflows. They might not seem within the title of “AI purposes”, however the core is that AI has lengthy been built-in and has develop into a part of completely different product types akin to video – modifying software program, workplace suites, browsers, and design instruments.
The period when a single – level perform might be independently monetized is gone.
The story of the AI utility layer continues.
Within the upcoming Agent period, the brink of the applying layer will probably be additional raised: sooner or later, single – level capabilities will probably be much more inadequate. Whether or not a really precious product can enter the method, join methods, tackle obligations, and be protected and controllable, turning mannequin capabilities into an executable, traceable, and measurable enterprise closed – loop.
Within the developer state of affairs, we will see that this alteration has superior an awesome deal. Instruments like OpenAI Codex and Claude Code are pushing AI from “code completion” to “software program growth company”. These Brokers begin to perceive code repositories, modify recordsdata, troubleshoot errors, generate exams, and even constantly advance round a growth process.
Such capabilities are troublesome to be instantly changed by a basic chat field. Actual – world software program growth requires steady judgment, modification, verification, and supply in a fancy engineering system. As soon as a product is used for a protracted sufficient time, it’ll accumulate challenge context, crew habits, historic issues, and operation data, and develop into increasingly intently tied to customers’ every day work.
This screening will proceed. The merchandise which have survived right now nonetheless should consistently reply the identical questions. Who will nonetheless be on the sphere after the tide recedes? The reply will probably be given by time.
This text is from the WeChat official account “Tencent Technology”. Writer: Li Hailun, Editor: Xu Qingyang. Republished by 36Kr with permission.