How vibe coding is changing startup economics, and why building software is no longer the hardest part.
Key insights
- ObservedMonthly launches of new subscription apps grew about 7× between January 2022 and January 2026. Apps launched in 2025 or later earned about 3% of subscription revenue in January 2026 (RevenueCat).
- ObservedAI-powered subscription apps earn more per paying customer in the first year and keep fewer of their subscribers after 12 months.
- HypothesisWhen building gets easier for many teams, the scarce resources shift toward knowing what to build, reaching the right buyers and keeping them.
- LimitsThe data covers subscription apps, not all of SaaS, and it does not show that vibe coding caused any of these patterns.
Everyone Can Build. Who Will Sell?
A founder with a free weekend and a coding agent can now put a working prototype in front of strangers by Monday. A few years ago, the same prototype would have needed a small team and a few months. Then Monday arrives, and the prototype meets a question that no coding tool answers: who will pay for this, and what will it cost to reach them?
This article is my attempt to take that question seriously. I work in growth and performance marketing, so I lean toward thinking acquisition matters. I tried to correct for that by using data I could check and giving opposing views their strongest form. The thesis is a hypothesis, not a law: AI is commoditizing parts of software creation, but not customer demand.
AI Is Changing the Economics of Building Software
McKinsey’s explainer on AI-powered software development, a discussion among three of its experts, describes a shift from coding assistants to agents that can specify, write, test and deploy software with limited human input. Its central point is organizational: the companies seeing real gains are not those that handed developers a new tool, but those that rethought how software gets made.
The examples are vivid, and they are examples. Matt Linderman describes a private-equity-owned software company whose CTO compared two full scrum teams with a single squad of one product manager and three developers. The squad finished a redesign in four days, against four weeks for a team four times its size. That is one client experiment reported by the consultant, not an industry benchmark.
The more useful idea is the gap between individual productivity and business value. Janaki Palaniappan says individual productivity gains may make a person more productive, but “it doesn’t lead to any real bottom-line or top-line impact” unless leaders redesign the whole workflow, from discovery to launch. Linderman adds a consequence: “As the code accelerates, the other steps become the bottleneck. First it was code review. Then it was product management.” Speeding up one stage moves the constraint to the next one.
A prototype is also not production software. Martin Harrysson cautions that full autonomy is not imminent, that it works best for greenfield builds and modernization, and that AI-generated code has so far “often been more verbose and less secure than human-written code.” In the 2025 Stack Overflow Developer Survey, 84% of more than 49,000 respondents said they use or plan to use AI tools, yet 46% did not trust the accuracy of the output and nearly 77% said vibe coding is not part of their professional work.
Measuring the effect is hard. In a July 2025 randomized trial, METR found that 16 experienced open-source developers took 19% longer with AI tools while believing they were faster. Its February 2026 update called the newer data an unreliable signal and said developers are likely more sped up now. I read that as a warning against both extremes: building has become much cheaper for some tasks and teams, and it has not become free or uniformly easy.
When Everyone Can Build, Competition Changes
If building gets cheaper for even a subset of founders, more products should reach the market. The best dataset I found is RevenueCat’s State of Subscription Apps 2026, built from more than 115,000 apps, $16 billion in revenue and over a billion transactions. It covers subscription apps on iOS and Android, so it is a view of one market, not of software as a whole, and its launch counts come from Appfigures.
Figure 1 · Observed data
New subscription apps launched per month
iOS and Android combined. Only the two reference points published in the report text are shown.
Bar chart. About 2,000 new subscription apps launched in January 2022. More than 14,700 launched in January 2026, roughly seven times as many. Source: RevenueCat State of Subscription Apps 2026, launch data provided by Appfigures.
About 7× growth. Bar lengths are drawn to scale. RevenueCat’s own chart shows monthly values in between; they are not reproduced here because the report does not publish them as figures. Source: RevenueCat, State of Subscription Apps 2026; launch data provided by Appfigures.
RevenueCat notes that the steepest iOS acceleration begins in early 2025, “coinciding with” the rise of AI-assisted development tools, and its founder’s letter calls the surge “likely driven by the AI app explosion.” Those are interpretations, not tested causal claims. Appfigures separately counted 557,000 new App Store apps in 2025, up 24%, naming AI and vibe coding as one of three drivers without quantifying it.
More supply does not mean more revenue for newcomers. In January 2026, apps launched before 2020 generated 69% of subscription revenue, and apps launched in 2025 or later generated 3%.
Figure 2 · Observed data
Share of monthly subscription revenue by app launch cohort
January 2026, all storefronts. The five cohorts sum to 100%.
Horizontal bar chart of monthly subscription revenue share by launch cohort in January 2026. Before 2020: 69 percent. 2020 to 2021: 9 percent. 2022 to 2023: 14 percent. 2024: 5 percent. 2025 to present: 3 percent. Source: RevenueCat State of Subscription Apps 2026.
Revenue share, not share of apps. Newer cohorts have had less time to build audiences, so this chart does not by itself show that new apps perform worse. Source: RevenueCat, State of Subscription Apps 2026 (data provided by Appfigures).
Newer cohorts have simply had less time, and RevenueCat itself credits older apps with years “to build audiences, optimize monetization, and compound growth.” What the data supports is narrower: far more launches have not changed who earns most of the revenue. The report also describes a “winner-take-more” pattern, with the top 25% of apps growing revenue 80% year over year while the bottom 25% shrank 33%. Its founder expects “more competition, higher CACs, and higher churn,” but that is an expectation. The report does not measure acquisition costs, and neither do I.
The New Bottleneck Is Distribution
When the number of products rises faster than buyers’ hours, attention becomes the constrained input. Martin Harrysson makes the point from the engineering side: if you can build almost anything quickly, then “everyone can do that,” including competitors and new upstarts, which puts a high emphasis on deciding what to build. Janaki Palaniappan frames it as a question: “Are those the right things to build? Will people actually want them?”
Distribution is the word people use for the answer, but it is easy to shrink it to “buying traffic.” Bryan Kim of Andreessen Horowitz argues in “In Consumer AI, Momentum Is the Moat” that default growth tactics are no longer enough and that paid acquisition and SEO “rarely create lasting retention anymore” in consumer AI. That is an investor’s opinion backed by company self-reports, not retention data, and he discloses an investment in ElevenLabs, a company the piece praises. The useful point survives: traffic is an input, and what matters is whether it turns into people who stay.
Figure 3 · Conceptual framework
How the bottleneck may shift when building gets easier
Conceptual framework · SciGrowth analysis
Flow diagram with four steps. One: AI-assisted development becomes more accessible. Two: more potential product launches. Three: more competition for attention. Four: distribution and retention become the harder problems. This is an analytical framework, not a measured causal relationship.
An analytical framework, not a measured causal relationship. Each link is a hypothesis that varies by market, and the arrows do not imply that every step happens in every category.
In practice, distribution is a bundle of separate jobs, and each can fail on its own:
- Customer research: knowing which urgent problem a specific person has, in their words.
- Audience access: owning or renting a path to those people, through a newsletter, community, creator or search position.
- Positioning and differentiation: a reason to pick this product over ten similar ones built last month.
- Organic and paid acquisition: channels with different cost curves, where paid is usually the fastest to test.
- Partnerships, brand and trust: borrowed credibility that makes a stranger willing to enter a card number.
The first session deserves its own mention. RevenueCat reports that 55% of all cancellations of 3-day trials happen on day zero, so distribution that sends a visitor to a product that fails to show value immediately is a cost, not an asset.
The Economic Paradox of AI-First Startups
Here is the hypothesis that interests me most as a performance marketer: if a team spends much less to build the first version, the money it saved may be redirected toward growth, and then competes with everyone else who made the same decision. The illustration below is a deliberately simple scenario, not a measurement.
Figure 4 · Illustrative scenario
Where a startup’s first-year budget could go
Two stacked bars comparing hypothetical budget splits. Traditional development-heavy startup: development 70 percent, marketing 20 percent, other 10 percent. Illustrative AI-first startup: development 20 percent, marketing 60 percent, other 20 percent. These are invented numbers for illustration, not market benchmarks.
Illustrative scenario only, not observed market data. The percentages are assumptions chosen to show the logic of the hypothesis. Source: SciGrowth analysis.
Whether that redirection happens, and whether it raises acquisition costs, depends on the market. Customer acquisition cost (CAC), the total acquisition spend divided by the number of new customers, is not guaranteed to rise everywhere. It depends on how many products chase the same audience, on the channel and on how well the offer fits the buyer. Advertising spend is also not interchangeable with demand: ads can rent attention for a need that exists, but they cannot create the need, and scaling spend into a weak funnel mostly scales the leak.
That is why I prefer a chain of metrics over a single LTV:CAC ratio: new-customer CAC, first-order contribution margin (revenue minus the variable costs of serving the order), CAC payback period (months for a customer’s gross profit to cover their acquisition cost), repeat or renewal contribution, and then 6- and 12-month lifetime value. Keep three CAC numbers apart as well: platform-reported cost per acquisition, blended cost per new customer and fully loaded cost. They can differ by a large factor on the same account.
AI products add a wrinkle. RevenueCat observes that the cost to serve a marginal subscriber “used to be near-zero, but that’s not true for products using LLMs,” so money saved in development can return as inference costs that reduce contribution margin per customer.
AI Products Can Monetize Early, But Can They Retain Customers?
RevenueCat classifies an app as AI-powered if AI or machine learning models deliver its primary value, and about 27% of the subscription apps in its data fit that definition. Compared with other apps, AI apps show a clear pattern: they earn more from each paying customer and keep fewer of them.
Figure 5 · Observed data
AI-powered vs. other subscription apps: revenue per payer and retention
Two-panel comparison. Panel one: median realized lifetime value per paying customer after one year was $30.16 for AI-powered apps and $21.37 for other apps, 41 percent higher. Panel two: share of subscriptions still active after 12 months. Monthly plans: 6.1 percent for AI apps, 9.5 percent for others. Annual plans: 21.1 percent for AI apps, 30.7 percent for others. RevenueCat's headline summary says AI apps churn about 30 percent faster; the two panels use different metrics and denominators. Source: RevenueCat State of Subscription Apps 2026.
RevenueCat’s headline summary states that AI apps generate 41% more revenue per payer but “churn 30% faster.” The two panels use different metrics and denominators and should not be combined into one calculation. Bars in each panel are drawn to scale. Source: RevenueCat, State of Subscription Apps 2026, AI vs. non-AI segment.
The two panels measure different things. Realized lifetime value counts the value of an average payer over a year, while the retention panel counts the share of subscriptions still active by plan duration. RevenueCat’s summary also says AI apps “churn 30% faster,” but I could not find how that figure is derived in the report’s tables, so I quote it as the report’s own headline and do not combine it with the other numbers. The median refund rate is also higher for AI apps, 4.2% against 3.5%.
How can strong monetization and weak retention coexist? One plausible reading, my hypothesis and not something the report proves, is that curiosity can drive the first purchase while lasting utility drives the tenth renewal. Plan mix may matter too: AI apps sell 59.8% of subscriptions as monthly plans against 26.2% for other apps, which gives more billing cycles at which a person can leave. The data cannot separate these explanations. It does support the practical point that attracting a paying customer and keeping their value are different problems.
Distribution Is Not the Same as a Competitive Moat
Everything above could be read as “distribution wins,” and informed critics disagree. José Huergo argues in “Distribution Is the New Moat?” that code used to be the moat and distribution is now the scarce asset, because “marketing is a complex problem.” His piece is opinion and anecdote, and he runs a consultancy that sells go-to-market help.
John Hwang’s reply, “Distribution is not a ‘moat’ in AI,” says that distribution does not equate to “defensibility,” and that in a low-barrier market vibe coding “just floods the zone with more competition.” His full text is paywalled, so I rely only on the free preview and do not represent his complete argument.
Kim’s essay lands between the two: it says there is no traditional moat in consumer AI, proposes momentum as a substitute, and adds that “traction from distribution only sticks if your product keeps up.” My synthesis is that distribution is an advantage that decays unless converted into something more durable, such as customer relationships, proprietary data, workflow embedding, switching costs, brand trust or network effects. None of these replaces a product people value. A strong channel with a weak product is a faster way to learn that it is weak.
Why Growth Strategy Becomes More Valuable
If building becomes cheaper for many teams, the expensive questions are the ones before and after the code. Linderman says that when engineering moves faster, “figuring out what to build becomes the constraint,” and Palaniappan says a human problem sits at the start and the end of the process.
This is where the role of a growth strategist may evolve, and I want to be precise: it does not replace developers. It treats customer discovery, positioning, acquisition, activation, monetization, retention and unit economics as one connected system, because a weak link anywhere changes the economics of the others. Founders who can reason across that whole chain, with or without a growth partner, probably have an edge that coding speed alone does not give them.
A Practical Go-to-Market Framework for AI-First Founders
This is a working checklist, not a proven formula.
- Define the customer and the urgent problem. Write one sentence naming who has the problem, what it costs them today and what they do about it.
- Validate demand and willingness to pay. Look for pre-orders, paid pilots or waitlists with a price attached.
- Identify plausible acquisition channels. Pick two or three where your buyers already spend attention and estimate a rough cost per customer for each.
- Build the smallest useful product. Use AI to compress this step, but keep security in scope wherever real customer data is involved.
- Measure activation and early retention. Track how many new users reach the first moment of value and how many return, before buying traffic.
- Validate contribution economics and CAC payback. Include model costs if the product runs on LLMs. If you sell through conversations, send real sales outcomes back to your ad platforms so the numbers reflect revenue and not only leads.
- Scale acquisition only when the fundamentals support it. Raise spend after payback and retention hold at small volume, and keep customers with lifecycle email and SMS.
Conclusion
AI can make software easier to build. It cannot manufacture customer demand. The evidence I could verify shows a large rise in the supply of new subscription apps, a revenue pool still held mostly by older apps, and AI-powered products that monetize well but retain worse. That is not a general law, and the sources have limits: the data covers subscription apps, McKinsey’s examples are interview anecdotes, and Hwang’s argument is only partly visible without a paid subscription. I treat the thesis as a working hypothesis to test in each market.
So here is the question I keep returning to: will the next generation of successful AI founders stand out through coding speed, through understanding customers, or through the rare ability to do both?
Building the product is only the beginning. Sustainable growth requires a strategy for acquisition, conversion, and retention.
If you want to talk through yours, send us a request, or see how we approach paid ads and the rest of our services.
Sources & Further Reading
- RevenueCat, State of Subscription Apps 2026. Primary quantitative source: launch counts (data provided by Appfigures), revenue by launch cohort (January 2026), AI vs. non-AI revenue and retention comparisons.
- McKinsey Explainers, AI-powered software development: How technology is rewriting the rules. Interview with Janaki Palaniappan, Martin Harrysson and Matt Linderman, edited for length. Quotes taken from the interview transcript.
- Bryan Kim, Andreessen Horowitz, In Consumer AI, Momentum Is the Moat (June 11, 2025).
- José Huergo, Distribution Is the New Moat? (May 11, 2026).
- John Hwang, Distribution is not a “moat” in AI (July 3, 2025). Paywalled; only the free preview was read.
- Appfigures, The App Store Just Logged Its Biggest Release Year in Nearly a Decade (December 5, 2025).
- Stack Overflow, 2025 Developer Survey press release (July 29, 2025).
- METR, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity (July 10, 2025), and its February 24, 2026 update.