Are AI apps getting too crowded?

Are AI apps suddenly everywhere?

Yes. AI apps are genuinely flooding the market right now, and the increase is much bigger than the usual startup hype cycle.

RevenueCat’s State of Subscription Apps 2026 gives us the clearest measure of what changed. Its dataset covers more than 115,000 apps and $16 billion in subscription revenue. Around 2,000 new subscription apps were launching each month in early 2022. By early 2026, that had climbed above 14,700 a month. That is more than seven times as many launches in four years.

The acceleration became especially sharp on iOS as AI coding tools became mainstream. RevenueCat found that iOS now represents roughly 77% of new subscription-app launches, up from about 67% in 2023.

Sensor Tower sees the same explosion from the consumer side. More than 200,000 apps now mention AI-related terms in their store descriptions, and those apps were on track for nearly 10 billion downloads during the first half of 2026, 25% more than a year earlier.

There is another reason AI suddenly seems to be everywhere: the label now covers products that would previously have been considered ordinary software. CapCut uses AI for effects, background removal, captions and video generation. Canva has built AI throughout its design suite. Notion’s paid AI attach rate went from roughly 20% to more than 50% within a year, according to a16z, and AI features now represent roughly half of its ARR.

So yes, the flood is real. Easier development has massively increased the number of products competing for the same users.

Measure	Earlier level	Recent level	Change
New subscription apps launched monthly	~2,000	14,700+	More than 7×
iOS share of new subscription apps	~67% in 2023	~77%	+10 percentage points
Apps mentioning AI	—	200,000+	Record high
Downloads of apps mentioning AI	—	~10B in H1 2026	+25% YoY

Is AI app supply growing faster than people can actually use it?

Yes. AI app supply is growing much faster than human attention, even though demand for AI itself is still booming.

Demand has clearly not collapsed. Sensor Tower estimated that people spent 17.2 billion hours in generative-AI apps during the first half of 2025. A year later, that figure was on track to reach roughly 36 billion hours. Usage more than doubled.

Spending is moving just as quickly. Generative-AI mobile apps produced less than $60 million in quarterly in-app revenue in early 2023. By the first quarter of 2026, they were generating around $1.9 billion. Over the preceding twelve months, Sensor Tower measured $6.1 billion of generative-AI app revenue, up 232% year over year.

Broader AI apps were heading for more than $4 billion of in-app purchases in the first half of 2026 alone.

People therefore want more AI software and are paying more for it. The problem is that developers can create new supply even faster.

Human attention has a hard ceiling. There are only so many apps someone will discover, test, learn and keep paying for. Meanwhile, AI-assisted coding allows thousands of teams to attack the same obvious use case at almost the same time.

That is how the market can look incredibly healthy in aggregate while becoming much harder for an individual founder.

Can new AI apps still make serious money today?

Definitely. New AI apps are still reaching revenue levels that would have looked absurdly fast only a few years ago.

Lovable is one of the strongest examples. The company launched the current version of its AI software-building product in late 2024, reached $100 million in ARR within roughly eight months and then doubled to $200 million about four months later. By mid-2026, Lovable said its annualized revenue had passed $500 million.

Higgsfield has shown a similarly extreme trajectory in AI video. After moving from more novelty-driven video generation toward professional content and advertising workflows, the company reported annualized revenue climbing from roughly $20 million to around $700 million in about a year. Commercial advertising became the majority of its usage.

Manus also reported crossing $100 million ARR within eight months of launching its general-purpose agent. By then, the company said its system had processed tens of trillions of tokens and created tens of millions of cloud computers for user tasks.

These companies are obviously outliers. We should not use them to estimate what the average AI founder will earn.

But they do answer one important question. A market where companies can still enter and reach nine-figure revenue this quickly is not closed.

The real deterioration is in the odds of becoming one of those winners.

Are most new AI apps barely making money?

Yes. The gap between the breakout AI apps and the typical new app is enormous.

RevenueCat found that only 4.6% of newly launched subscription apps reach $10,000 in monthly revenue within their first two years. Only 17.3% even make it to $1,000 a month.

The gap becomes more striking when we look at where subscription revenue already sits. Apps launched before 2020 still generate 69% of all subscription revenue in RevenueCat’s dataset. Products launched in 2025 or later generate only about 3%.

Age explains part of that difference. Older apps have had years to build reviews, SEO rankings, audiences, email lists, paid acquisition systems and recurring subscribers.

Recent performance is also extremely uneven. Median monthly recurring revenue across subscription apps grew just 5.3% year over year. The top 10% grew more than 306%. An app shrinking by more than 33% was only bad enough to fall into the bottom quarter.

That is a rough market to be average in.

Revenue milestone within two years	Share of new apps reaching it
$1,000 monthly revenue	17.3%
$10,000 monthly revenue	4.6%
Revenue generated by pre-2020 apps	69% of total
Revenue generated by 2025+ apps	3% of total

Did AI make building apps easier than finding users?

Yes. Building an AI app has become dramatically easier, while getting people to notice and keep using one remains hard.

The sevenfold increase in subscription-app launches captures this better than anecdotes about vibe coding. Tools such as Lovable, Replit and coding agents can generate interfaces, databases, authentication systems, integrations and production code that previously required much more time and engineering work.

The same thing is happening around the product. AI can create landing pages, ads, help-center articles, onboarding flows and customer-support responses.

The catch is obvious: every competitor gets the same tools.

A software idea that once required several engineers and six months of work might have attracted five serious teams. If people can prototype it over a weekend, dozens or hundreds can chase the same opportunity before anyone has built much of a lead.

We can see the consequence in RevenueCat’s data. Launches rose more than sevenfold, yet apps created before 2020 still collect almost seven dollars out of every ten spent on subscriptions.

Those older apps often have an advantage that has little to do with better code. They already own attention.

Are basic AI wrappers becoming a bad business?

Yes. Basic AI wrappers are getting much harder to defend because the large AI platforms keep absorbing the features that once justified separate apps.

Image generation shows what this looks like in practice. In a16z’s first ranking of major consumer generative-AI products, seven of the nine creative tools were mainly image generators. Three years later, only three dedicated image-generation products remained on the list.

During that period, ChatGPT and Gemini added increasingly capable image generation directly inside products already used by hundreds of millions of people. Midjourney, once a top-ten product in a16z’s ranking, had fallen to number 46 by its sixth edition.

Yet specialized creative apps have survived. Leonardo, Ideogram and CivitAI still appeal to particular creative communities. Suno built a strong music product. ElevenLabs has appeared in every edition of a16z’s ranking because professional voice cloning, dubbing and audio workflows require much more than a basic generation box.

Video has also remained much more open. Kling, Hailuo, PixVerse and Higgsfield have all built sizeable audiences while the frontier labs improve their own video models.

So “wrapper” is not a very useful diagnosis by itself. Almost every software company builds on somebody else’s infrastructure.

The useful question is what customers would still need from the product if everyone received access to the same underlying model tomorrow. A thin interface is vulnerable. A product with workflow, editing tools, integrations, proprietary context or a community has far more room to survive.

Is retention becoming the biggest problem for AI apps?

Yes. AI apps are unusually good at getting people to pay once, but they are currently much worse at keeping them.

RevenueCat compared AI-powered and non-AI subscription apps across its dataset. AI apps converted trials to paid subscriptions at a median rate of 8.5%, compared with 5.6% for non-AI apps. Download-to-paid conversion was also slightly higher at 2.4% versus 2%.

AI customers spent more too. Median first-month value per payer reached $18.92 for AI apps compared with $13.59 for non-AI products. After a year, realized value was $30.16 against $21.37.

Then the retention numbers flipped the advantage.

After twelve months, only 6.1% of monthly AI subscribers remained, versus 9.5% for non-AI apps. Annual-plan retention was 21.1% for AI products and 30.7% for non-AI products. Refund rates were also higher at 4.2% compared with 3.5%.

The pattern is awkward: AI wins the first payment and loses more of the renewals.

Crowding makes that retention problem worse because a user cancelling one AI app can immediately try another one that promises almost the same result.

Metric	AI apps	Non-AI apps
Trial-to-paid conversion	8.5%	5.6%
Download-to-paid conversion	2.4%	2.0%
First-month value per payer	$18.92	$13.59
Year-one value per payer	$30.16	$21.37
12-month retention, monthly plan	6.1%	9.5%
12-month retention, annual plan	21.1%	30.7%
Median refund rate	4.2%	3.5%

Will people really pay for several AI apps at once?

Yes, and people already do, but those apps need to solve meaningfully different problems.

a16z found that roughly 20% of weekly ChatGPT web users also used Gemini during the same week. People are clearly willing to use several AI products rather than choosing one permanent winner.

The behavior makes even more sense for specialized software. A developer can use Claude or ChatGPT while paying separately for an AI coding product. A marketer might use a general assistant alongside a video generator and an image editor. A company can simultaneously buy AI for coding, customer support, sales and internal search.

Menlo Ventures estimated that companies spent $19 billion on user-facing generative-AI applications in 2025. Coding alone represented around $4 billion, alongside billions more spent on general copilots, healthcare AI, marketing, customer service, design and other categories.

Redundancy is where things get harder.

Three subscriptions that all behave like slightly different general chatbots compete for the same job in someone’s life. A coding agent, meeting recorder and healthcare workflow do very different things, so they can comfortably coexist.

That leaves plenty of room for multiple AI subscriptions, just not an unlimited number of interchangeable ones.

Are vertical AI apps less crowded than general AI assistants?

Yes. Vertical AI is currently one of the clearest places where startups can still build substantial businesses without fighting ChatGPT head-on.

General assistants compete on huge surfaces. They need strong models, brand recognition, compute, distribution and enough functionality to answer almost anything. Sensor Tower found that ChatGPT, Gemini and DeepSeek together already account for nearly 90% of time spent in mobile AI-assistant apps.

Vertical software plays a narrower game. An AI product for healthcare administration does not need to beat Gemini at general reasoning. It needs to understand one workflow deeply enough to save a hospital time or money.

Menlo Ventures estimated that vertical-AI spending reached $3.5 billion in 2025, almost three times the previous year. Healthcare alone represented around $1.5 billion, with AI medical scribes generating roughly $600 million.

Startups captured approximately 88% of vertical-AI spending in Menlo’s analysis. Across the entire AI application layer, startups held about 63% of spending, up sharply from 36% a year earlier.

That is especially interesting because incumbents should have an enormous advantage in enterprise software. They already own customer relationships, procurement approval, data and distribution. Yet new AI companies are still taking share.

The strongest opportunities today are often hidden inside very specific jobs: coding, medical documentation, legal work, customer support, sales workflows or industry-specific back-office tasks.

Those categories can support many winners because customers care more about whether the product solves their exact problem than whether it owns the world’s best general model.

Can ChatGPT, Gemini and existing software companies copy successful AI apps?

Yes, and this is one of the biggest risks facing simple AI products today.

Google, OpenAI, Anthropic, Microsoft, Adobe, Canva, Notion and other large platforms can add features to products that already have huge user bases. If a startup’s entire product is a feature that fits naturally inside one of those platforms, the threat is obvious.

Image generation has already gone through this compression. Standalone image generators lost relative ground as ChatGPT and Gemini became good enough for many mainstream users.

The same pressure is spreading into writing, document analysis, research, meeting summaries, presentations and basic productivity tasks.

Copying a feature and copying an entire workflow are very different jobs, though.

GitHub entered the AI coding market with almost every structural advantage imaginable. Even so, Cursor gained serious traction by redesigning the coding environment around AI, adding repository-level context, multi-file edits and agentic workflows faster than the incumbent.

Menlo Ventures estimated that startups captured around 71% of AI spending in product and engineering despite the presence of GitHub, Microsoft and other established vendors.

The dividing line is fairly clear. AI features that slot neatly into existing software are dangerous startup territory. AI products that force the workflow itself to change still give new companies room to move faster.

Will ChatGPT, Claude and Gemini become the new app stores?

They are already moving in that direction. AI assistants increasingly decide which outside service gets surfaced when a user wants something done.

OpenAI introduced apps inside ChatGPT and later opened an app directory where users can browse or search for them. Developers can build interactive products that appear inside conversations, and OpenAI has been experimenting with surfacing relevant apps based on context and previous usage.

Anthropic has followed a similar path. Claude’s connector directory had grown beyond 200 integrations by April 2026, covering productivity, design, finance, health, travel and other categories. Anthropic says users increasingly combine several connected products inside the same conversation.

Google pushed the model further recently. Gemini announced integrations with services including Granola, Otter.ai, Wix, Ticketmaster, Zocdoc, Angi and Thumbtack, allowing users to complete more tasks without leaving the assistant.

That could become a major opportunity for small software companies. A niche app no longer necessarily needs someone to remember its name, visit its website and form a daily habit. The assistant can surface it exactly when the user needs its function.

The trade-off is obvious: the AI assistant becomes another gatekeeper.

If ChatGPT or Gemini chooses which restaurant service, design tool or travel app gets called, being selected by the assistant could eventually matter as much as ranking highly in Google Search or an app store.

AI app distribution may therefore become less about convincing someone to install software and more about becoming the tool that an AI platform chooses to use.

Does cheaper and better AI make AI apps stronger or easier to copy?

Both. Better models make good AI apps more capable while making shallow AI apps easier to reproduce.

Frontier models have become much closer in quality across many mainstream tasks, while inference prices keep falling. Application developers can now switch between providers, route different requests to different models and benefit automatically when OpenAI, Anthropic, Google or other labs improve performance.

That is fantastic for products built around a real workflow. A medical documentation app, coding agent or video-production platform can become substantially better without training its own frontier model.

A product whose main advantage is one clever prompt faces the opposite effect. Each model upgrade reduces the amount of specialized engineering required to reproduce what it does.

So defensibility increasingly comes from what surrounds the model.

Cursor builds around repositories and development workflows. ElevenLabs combines voice models with cloning, dubbing, editing and production tools. Vertical enterprise products connect with customer systems and accumulate context that a generic assistant does not automatically have.

Integrations can create another layer of stickiness. Once a company connects an AI product to its systems, retrains employees and builds processes around it, switching becomes more annoying and risky.

Distribution, proprietary data, community and trust matter too. They are all slower to copy than the underlying AI capability.

Better models will keep wiping out weak differentiation. Strong applications should actually benefit from those same upgrades.

Is distribution now more important than having better AI?

For most AI apps, yes. Distribution is currently harder to reproduce than access to a good model.

The supply numbers explain it. Nearly 15,000 new subscription apps are launching every month, compared with around 2,000 four years earlier. At the same time, only 4.6% of new apps reach $10,000 in monthly revenue within two years.

There is no shortage of software anymore.

Established apps begin with advantages that compound: app-store reviews, SEO rankings, email lists, installed users, communities, partnerships, historical data and recognizable brands. That helps explain why apps launched before 2020 still earn 69% of all subscription revenue.

Some of the fastest-growing AI startups have found ways around that disadvantage.

Lovable benefits when users publicly share what they build. Higgsfield attached itself to growing demand for social and advertising video. Enterprise AI companies can spread employee by employee rather than depending entirely on slow top-down procurement.

Menlo Ventures estimated that product-led adoption represented roughly 27% of enterprise AI application spending, almost four times the comparable rate it measured for traditional SaaS.

Being able to create an excellent AI product is increasingly the entry ticket. The harder advantage is having a repeatable way to put that product in front of the right people.

Where is the AI app market already too crowded?

Generic AI assistants, basic image generators and one-function utilities already look overcrowded.

General assistants are the clearest case. ChatGPT, Gemini and DeepSeek account for nearly 90% of time spent in mobile AI-assistant apps according to Sensor Tower. ChatGPT alone became the fastest mobile app to reach one billion monthly active users.

Competition among the leaders is still real. ChatGPT’s share of unique cross-platform AI-assistant users dropped below 50% for the first time during 2026 as Gemini and Claude grew. a16z also found Gemini’s paid subscriber base growing quickly and Claude’s paid subscribers growing by more than 200% year over year.

But those numbers are hardly an invitation to launch another generic chatbot. A new entrant would be competing against companies spending billions on models, infrastructure, distribution and advertising.

Basic image generation has moved in the same direction as the major assistants bundle good-enough generation. Simple PDF summarizers, email rewriters, document chat tools and caption generators face a similar problem because ChatGPT, Claude and Gemini increasingly handle those tasks themselves.

Specialized video creation, professional agents, regulated industries, local services and deep enterprise workflows remain much more open.

The AI app market has become extremely uneven. Some categories still have plenty of room. Others have already reached the point where another nearly identical product adds supply without creating much new value.

Are AI apps getting too crowded?

Yes, especially at the shallow end of the market. AI has made app creation so easy that generic products are arriving much faster than users can give them attention.

The evidence is strong enough to be decisive. Monthly subscription-app launches have increased more than sevenfold. More than 200,000 apps already advertise AI capabilities. Only 4.6% of new subscription apps reach $10,000 in monthly revenue within two years. AI apps convert better and earn more from each payer, yet they retain subscribers substantially worse than non-AI apps.

That is what crowding looks like in practice: lots of launches, easy initial curiosity and brutal competition for lasting usage.

Calling the entire AI app market saturated would still go too far.

Demand is growing extremely fast. Sensor Tower measured generative-AI app revenue rising 232% year over year, while usage more than doubled. Businesses spent roughly $19 billion on AI applications in Menlo Ventures’ latest full-year analysis, and startups captured most of that spending. New companies such as Lovable and Higgsfield have also shown that a product launched into today’s competitive environment can still reach hundreds of millions of dollars in annualized revenue remarkably quickly.

The market is separating into very different layers.

At one extreme, a handful of giant assistants are fighting to become the default interface for AI. At the other, thousands of easily built utilities are competing for tiny slices of attention.

The best territory sits between those extremes: AI products that own a specific workflow, audience, industry, dataset, community or distribution channel.

A few years ago, being able to build the AI product was itself an advantage. Today, thousands of other teams can probably build something similar.

The question that matters now is whether users have a specific reason to choose yours, keep using it and keep paying for it.

A growing number of AI apps cannot answer that question. Those parts of the market are already too crowded.

The apps that can answer it still have a very large market in front of them.