AI is creating an era of app glut

Will you pay for apps?

Now that the starting point of software development has become a sentence generation, we send prompt words such as “help me build an accounting app”, and Codex or other agents will automatically complete a series of processes such as coding, review, testing, online and release.

Ideas that were originally shelved because they couldn’t program and couldn’t find developers now have a chance to be made.

AI programming tools lower the threshold for making applications, but they also put an old problem in front of “developers” earlier:After it is made, who will use it?

The latest weekly data report from a16z poured cold water on the “vibe coding entrepreneurial wave”: application supply has quadrupled, but demand has remained unchanged; unicorns are indeed getting younger, but the gold rush dreams of independent developers have not been realized in the data.

There are indeed many apps in the app store.However, users’ downloading, usage and payment have not kept up with the same speed.。

The era of App-Slop

According to a research paper “Writing Code vs. Shipping Code” published by researchers from MIT and the University of Pennsylvania in May this year, from the beginning of 2025 to April 2026, the number of newly released apps on iOS approximately doubled each month.

Chrome extension growth accelerated, and new apps on Google Play reversed a years-long downward trend.

Researchers mark February 2025 as the starting point of the “Agentic coding era”. From this starting point, the number of monthly new applications on the three major platforms took off collectively: iOS rushed from about 40,000 to 120,000 per month, Android and Chrome also doubled to four times respectively. The curve is almost vertical after the inflection point in early 2025.

The researchers then looked at the performance of these new apps in their first three months: how many downloads they received, how many ratings they accumulated, and whether they found their first batch of users.

The results are somewhat bleak. There has been a significant increase in new applications, but the total engagement received by new monthly applications in the first three months has been flat or declining. Even applications that have not accumulated a small audience, and their proportion is still rising.

The demand side of the second row shows the total number of iOS ratings, Android and Chrome downloads during the same period, which is almost a flat line, and Chrome is even continuing to decline.

What is more telling is the structural data. When divided by usage, the proportion of “zero/low usage” new applications, that is, ratings less than 10 and downloads less than 100, has increased significantly, while the proportion of applications showing any sign of “escape velocity” that needs to cross the 10/100 threshold has dropped significantly.

The researcher specifically tested the possibility that the total amount would not change, but the shares would be redistributed internally, and a new winner would be born? The answer is no.It’s not that the cake has not changed and the winner takes all, but that there are no new winners at all.

Sternstein gave this phenomenon a name: App-Slop,Application swill in the AI ​​era. This is reminiscent of the “AI Slop” in the content farm era, except this time it’s the app store that’s being flooded.

Data from Sensor Tower confirms this judgment from the commercial side: using December 2024 as the benchmark, by August 2026, the total US application revenue index is only 102.0, and the total usage time index is 107.0.

In one and a half years, revenue +2%, time +7%; more software was released, but did not bring new usage demand of a corresponding scale.The pie of the application economy has not become bigger because of AI.

The only winner is AI itself

However, the sentence “AI has not driven the application economy” needs to be added with a qualifier:The AI ​​application itself is doing very well.

The scatter plot by category shows that Productivity (efficiency) is the only category in the US market where revenue and usage time are growing rapidly at the same time.

Productivity application revenue increased by approximately 100% and duration increased by approximately 55%, driven by ChatGPT, Claude, Gemini and Grok. Another small but eye-catching category is developer tools, and the largest game category is still experiencing negative growth in revenue.

In terms of usage, general AI assistants have similar advantages, that is, different tasks can be completed in the same entrance. Modifying emails, organizing information, and discussing travel arrangements can all become reasons for users to open it again.

In contrast, a new app that solves a single problem needs to prove its value more clearly. Especially when its main functions can also be completed in existing AI assistants, developers have to answer how much more convenient it would be to install more of this app.

Figure | After the release of ChatGPT, tools like Grammarly have also been significantly impacted. Specialized AI is no match for a general-purpose AI.

Users are willing to pay for AI capabilities and will not be interested in every app made with AI.

This is also where the “App Explosion” makes it easy for people to misjudge. Even though the efficiency category is popular now, we have used AI Vibe Coding to create an efficiency application. It is almost as difficult to make money from it.

There is still a long period of work between the excitement of this AI programming development tool and the popularity of the development results.

Just looking inside software development, the acceleration brought by AI has not been fully transferred to the delivery results.

The study also combined GitHub activity and AI tool usage data from more than 100,000 developers. The researchers estimated that after cumulative adoption of AI programming tools across generations, code submission activity increased by approximately 180%, and the number of final software versions released increased by approximately 30%.

Suppose you use AI to make an accounting app. The income, expenses, categories, and monthly charts are all present, and the presentation is also complete. But then you have to decide who exactly it serves.

Figure | Most independent developers’ first projects are accounting, notes and todo lists

A college student who wants to control takeout expenses has very different requirements for accounting tools than a family who needs to manage joint expenses. The former may be most concerned about whether the input is fast enough, while the latter requires multiple people to use it together and distinguish the attribution of expenditures.

These choices will directly affect product design, and continuous addition of functions may not make it easier for a certain type of user to use it.

Commercialization brings another set of specific questions: How can potential users know about this application? If you need to buy ads, how much does it cost to acquire a paying user? If the product continues to invoke the model, will subscription revenue cover running and service costs?

For users to pay every month, they need a reason

If we say that people are not very willing to pay for AI-generated applications, but even if we look at those who pay for AI capabilities, the results are not very optimistic. These AI payments on the consumer side are far from reaching everyone.

According to Consumer Edge data cited by a16z, as of the first quarter of 2026, the proportion of individual US consumers paying for AI services will be approximately 3%.

The proportion of those aged 18–24 who pay for AI services is about 5.9%, those aged 25–34 about 5%, and only 1.4% of those aged 65 and over.

Young people are paying for AI at four times the rate of their elders, so while it’s a cliche that “we’re still in the very early stages of AI”, the data does back it up.

Pushing the discussion further to the grand question of “when will AI be reflected in GDP?”, his answer is:Wait until these apps get good enough.

Data from Consumer Edge also shows that the proportion of cardholders paying for at least one AI service has roughly doubled in the past year. The number of people paying directly is still limited, but at least it is slowly growing.

This report also analyzes the current status of startups, on one sideThe new unicorns are getting younger and younger, with the median age having dropped to 4 years, while the old guys are getting older, and the middle layer is disappearing.

In terms of revenue, SVB data quoted by a16z shows that since the beginning of 2022, the median revenue growth rate of multiple technology industries it tracks has dropped from about 40% to 70% to about 15% to 30%; at the same time, losses have narrowed significantly.

These companies have slowed expansion and focused more on improving profits and extending the life of their cash. Those companies that have recently completed financing still retain the characteristics of higher growth and larger losses.

Although the capital market is still willing to provide funds for growth, the financing situation of different companies varies greatly.

Based on progress in the first half of the year, SVB estimates that approximately 2,345 U.S. companies backed by venture capital may fail or close throughout 2026. Among the failed companies it counted in 2026, more than one-third were established during the zero interest rate period from 2019 to 2021.

This set of data seems far away from an independent developer, but it actually answers the simplest question at the beginning of the article.

When anyone can make an app in a few hours or even minutes,“Making things” itself has become increasingly difficult to become a business。

What AI programming really changes is the supply curve in the software world. In the past, for a small demand, no one might be willing to do it because the development cost was too high; now they can be quickly made into products. As a result, app stores will become more and more crowded, and software may be generated in large quantities and quickly, just like today’s web pages, pictures, and videos.

But we still only have 24 hours in a day. Users will not download four times more software just because the number of apps has quadrupled; nor will they be willing to pay four times more money just because developers have completed the work of the past month in one sentence.

Picture | 2026 Apple Design Award Winners

Therefore, the real value of Vibe Coding is in reducing the cost of “verifying an idea” rather than the cost of “making a business”. We can first make an MVP (minimum viable product, the simplest version of the product), test it with real users, use data to determine whether the demand exists, and then decide whether to continue investing.

This is actually very similar to what is happening in the large model industry. The so-called Tokenmaxxing, more tokens and more calls will not automatically translate into more revenue; similarly, more code and more apps will not automatically create more demand.

Even for independent developers, the situation may be harsher than before. More and more independent developers themselves have become users of AI applications, and former paying users have also become independent developers. However, there are no new paying users for applications developed by independent developers using AI.

AI has lowered the barriers to entry in this market, while also lowering the value of many simple softwares themselves.