Can infinite scrolling tiktok apps make money?

Why can Infinite Slop exist now when this was impossible before?

Infinite Slop is possible now because fal's H3 Max can generate a five-second 768p video in under three seconds, which finally lets AI video production run ahead of the person watching it.

That speed threshold changes what developers can build. Pieter Levels' Infinite Slop takes suggestions from chat, generates the next scene and tries to connect it with the previous one. There is no finished catalogue that needs to be produced before the stream starts. The content keeps being created while people are watching.

Fal's own launch benchmark says H3 Max produces a five-second clip in under three seconds and delivers roughly 35 times the throughput of the official MiniMax H3 endpoint. Levels described the speedup as around 50 times in his write-up. The exact multiplier matters less than the result: video generation has crossed real-time playback.

That sounds like a technical detail, but it removes a constraint that used to kill this type of product. If a five-second clip takes 30 seconds to render, an endless interactive stream constantly catches up with itself. At under three seconds, the application can keep several clips buffered ahead of the viewer.

So these days an indie developer can genuinely launch something that behaves like procedurally generated television. The harder question starts immediately after that: who pays for it?

Does Infinite Slop actually prove an AI TikTok can make money?

Infinite Slop proves that an endless AI video feed can exist and attract attention, but we still have no evidence that the consumer product can pay its own bills.

Levels says directly that Infinite Slop would be very expensive for him to run and that fal is sponsoring the project. Fal therefore absorbs the biggest variable expense in the experiment: video generation.

That arrangement makes sense for fal. Infinite Slop demonstrates H3 Max doing something memorable that competitors could not easily demonstrate a year ago. Developers can literally watch the model generate video faster than they can consume it. Fal is effectively buying a live technical showcase in front of the exact audience that might later use its API.

For the app itself, however, we do not have disclosed subscription revenue, advertising revenue, paid conversion, retention or profitable unit economics. The experiment has validated the product experience before validating the business.

There is one encouraging detail buried inside the architecture. Infinite Slop is a shared stream. When 1,000 people watch the same generated scene, the expensive generation happens once rather than 1,000 times. That distinction turns out to be central to whether this whole category can work.

Did OpenAI's Sora prove people actually want an AI-only TikTok feed?

OpenAI's Sora gave us a fairly brutal answer: millions of people will try an AI-only TikTok feed, but novelty alone did not create durable usage or enough revenue.

Sora was unusually close to the obvious version of this idea. OpenAI built a vertical social feed around AI-generated videos, with recommendations, sharing, remixes and characters people could place inside generated scenes.

The initial demand was real. Sensor Tower data reported by Bloomberg showed worldwide downloads approaching five million one month after launch. Yet monthly active users then fell every month at the beginning of 2026. By March, Sensor Tower estimated Sora at 4.7 million monthly active users, compared with 7.8 million for Kling AI.

The money was even less convincing. Sensor Tower estimated about $1.4 million in global net in-app revenue for Sora since launch, while ChatGPT generated roughly $1.9 billion through the app over the same period. Different analytics firms produced slightly different Sora revenue estimates, but they all point to the same order of magnitude: a few million dollars, despite enormous launch attention.

OpenAI has since discontinued the Sora web and app experiences, and its help center currently says the remaining API is also being wound down. That makes Sora an unusually valuable experiment because a company with enormous distribution, frontier video models and one of the strongest consumer AI brands still struggled to turn an AI social feed into something worth continuing.

We should not conclude that AI feeds are dead. Meta, for example, is still testing a standalone version of Vibes after saying its AI-video feed saw strong early traction inside Meta AI.

The narrower conclusion is more useful: getting people to open an AI TikTok is clearly possible. Getting them to keep opening it after they have seen the trick is much harder.

What would make people keep watching an AI feed after the novelty wears off?

An AI feed will keep people when viewers start caring about what happens next. Endless random novelty is weak at creating that kind of attachment.

The short-drama market gives us one of the clearest comparisons. Sensor Tower measured more than 850 million short-drama app downloads worldwide in the first quarter of 2026, up 140% year over year. Average daily time spent reached roughly 25 minutes by April, 85% higher than in January 2025.

Those apps look surprisingly similar to TikTok on the surface. The videos are vertical, episodes are short and the next piece of content is always one swipe away. The psychological structure underneath is very different because viewers follow recurring characters and unresolved stories. Each clip makes the following clip more valuable.

Large social-video platforms show another version of the same idea. Kuaishou currently has about 412 million daily users and 797 million monthly users. In its latest results, the company said the number of users with mutual follows who privately messaged each other increased more than 15% year over year. Kuaishou keeps investing in social relationships, creators, events and communities even though it already has effectively unlimited videos.

That tells us where the real scarcity sits. TikTok, YouTube Shorts and Kuaishou already have more content than anybody could watch. Another trillion clips do not solve much by themselves.

AI feeds probably need memory. A character we recognize, a storyline we want resolved, a world we helped shape, a competition we can influence or a community we return to gives yesterday's consumption some value today.

That is a much stronger retention loop than asking a video model to surprise us again.

Can ads actually pay for a personalized AI TikTok feed?

No. Normal mobile ads currently come nowhere close to paying for a fresh H3 Max video stream generated separately for every viewer.

Fal is temporarily showing a 50% launch promotion for H3 Max, at $0.025 per generated second for 480p and $0.04 for 768p. The same product page says the normal prices return to $0.05 and $0.08 per second after the promotion. We use those stated normal prices for the business calculation because a temporary launch discount would give us a misleading long-term result.

At $0.08 per second, generating one continuous hour of 768p video costs $288. At 480p, the same hour costs $180.

Now compare that with mobile advertising. Appodeal's latest published eCPM report puts average U.S. interstitial eCPMs around $12.30 on Android and $13 on iOS. Using roughly $12.65 as a simple midpoint gives us a deliberately generous ad benchmark.

Imagine ten-second clips with one full-screen ad after every ten videos. That would be 36 ads during one hour of viewing, an extremely aggressive ad load for an entertainment product. At a $12.65 eCPM, those 36 impressions generate about $0.46.

We would be spending $288 on video generation to create roughly 46 cents of gross advertising revenue.

Even the temporary 50% H3 Max launch price only cuts the 768p compute bill to $144 per viewer-hour. The gap would still be more than 300 times.

These are already friendly assumptions. We have ignored bandwidth, storage, payment fees, engineering, content moderation, app-store fees, user acquisition, unfilled ad inventory and the fact that forcing an interstitial every 100 seconds could damage retention.

Personalized 768p feed	Ads per viewer-hour	Gross ad revenue at $12.65 eCPM	H3 Max generation cost	Compute / ad revenue
1 ad every 5 ten-second clips	72	$0.91	$288	~316x
1 ad every 10 clips	36	$0.46	$288	~632x
1 ad every 20 clips	18	$0.23	$288	~1,265x

What changes if one AI video is shown to thousands of viewers?

Shared generation changes the economics completely because the generation bill barely moves when another viewer watches the same clip.

Take the same 768p stream costing $288 to generate for one hour. With one viewer, the generation cost per viewer-hour is $288. With 1,000 simultaneous viewers, it falls to 28.8 cents. With 10,000 viewers, it falls to 2.88 cents.

That is why the architecture behind a product matters more than the fact that it looks like TikTok.

Using the same deliberately aggressive advertising assumptions from the previous section, one viewer generates roughly $0.46 per hour with 36 interstitial impressions. A shared 768p channel would therefore need around 633 simultaneous viewers to cover the raw $288 hourly generation bill. A 480p channel would need roughly 395.

At 10,000 simultaneous viewers, the generation cost becomes small compared with the hypothetical ad revenue. Bandwidth, moderation, distribution and acquisition still have to be paid, but we have at least moved into normal internet-business economics.

Caching creates another version of the same effect. A clip can be generated once, stored and then recommended to millions of people over several days. The more often each generated asset gets reused, the closer the economics move toward today's user-generated video platforms.

This gives us a useful rule for AI entertainment: generation cost should grow much more slowly than audience size.

Shared feed scenario	480p viewers needed for compute break-even	768p viewers needed for compute break-even
1 ad every 5 clips	~198	~316
1 ad every 10 clips	~395	~633
1 ad every 20 clips	~791	~1,265

Could TikTok Shop-style commerce make an AI feed profitable?

TikTok Shop suggests commerce can be far more valuable than advertising when the feed naturally creates buying intent.

Momentum Works and Tabcut estimate that TikTok Shop generated $11.8 billion of U.S. GMV during the first half of 2026, up 103% year over year. Globally, TikTok Shop reached about $50.3 billion during the same period, up 92%.

The interesting part for an infinite video product is where those transactions started. Video accounted for 40.4% of attributed U.S. GMV. Applying that share to the $11.8 billion total gives us roughly $4.8 billion of merchandise volume connected to video discovery in six months.

GMV is obviously much larger than TikTok's actual revenue, so we should not compare that $4.8 billion directly with advertising revenue. It still demonstrates how much more economic value a piece of content can create when it leads to a purchase.

That changes the design space for an AI feed. Imagine an endless interior-design feed where every room can be shopped, an AI fashion channel that continuously generates outfits from real inventory, or a collectibles feed where the generated entertainment revolves around items people can actually buy.

A generic comedy clip might earn a fraction of a cent through ads. A clip that helps sell a $120 product can support affiliate fees, marketplace take rates, merchant advertising or lead-generation revenue.

There is a catch. TikTok Shop has merchants, inventory, creators, logistics, payment infrastructure and years of recommendation data. An indie developer cannot reproduce that ecosystem.

The more realistic opportunity is a very narrow category where the feed, the audience and the transaction fit together naturally. If we can identify that category, infinite generation starts looking like a merchandising engine rather than an expensive entertainment toy.

Are ReelShort and DramaBox a better model than TikTok for AI video?

ReelShort and DramaBox may be better models for AI video because they make money from continuation rather than trying to monetize every minute with ads.

Sensor Tower estimates short-drama apps generated about $750 million of global in-app purchase revenue in the first quarter of 2026, up 20% year over year. DramaBox and ReelShort each came close to $140 million during that quarter alone.

This category is especially interesting because its interface already resembles the product people imagine when they talk about "AI TikTok." Viewers consume rapid vertical videos on a phone and can continue almost indefinitely.

The business model adds something that a generic endless feed lacks: scarcity. The viewer cares about one particular story, and the app can charge for the next part.

Generative video could make this model even stranger. A story can branch based on audience votes. Characters can continue indefinitely. Popular worlds can receive more episodes quickly. Different countries can get localized versions without reshooting everything. Paying users could even request a branch that the rest of the audience does not see.

The key is that infinite supply does not have to mean interchangeable supply. A thousand generated videos about characters we do not care about are worth very little. One generated scene resolving a cliffhanger we have followed for three days can be worth real money.

Video model	Recent commercial evidence	What creates the money
Short-drama apps	~$750M quarterly IAP; ReelShort and DramaBox near $140M each	Viewers pay to continue stories
TikTok Shop	$11.8B U.S. H1 GMV; 40.4% attributed to video	Content drives product discovery
Kling AI	More than RMB850M quarterly revenue	Customers pay for AI generation
Ad-funded short video	Mature global business model	Many cheap impressions served at scale

Should an AI feed charge people to generate videos instead of charging them to watch?

Charging for AI video generation is already a proven business, and Kling AI gives us the clearest recent evidence.

Kuaishou reported more than RMB850 million of Kling AI revenue in the second quarter of 2026, with revenue growing more than 200% year over year. That is roughly an order of magnitude larger than the entire lifetime mobile spending estimates we saw for Sora.

Kling's revenue includes a broader AI-video business rather than only consumer mobile subscriptions, so the comparison is imperfect. The underlying economics are still much cleaner. Expensive generation happens when a customer explicitly asks for it and is willing to buy credits, a subscription or professional access.

An infinite feed can use the same idea without putting the entire product behind a paywall.

We can let everyone watch a shared stream cheaply, then charge when somebody wants to take control. A viewer might pay to choose the next scene, create a private branch, put themselves inside the story, regenerate a clip, extend a character or start a personal channel.

Kuaishou's latest results contain another small clue. Users sent more than six million AI-generated gifts inside livestreams during the second quarter. People already pay to alter or enhance shared entertainment experiences when the action feels personal and visible.

That model lines up the compute bill with the person causing it. The passive viewer remains cheap. The expensive user becomes the paying user.

For a generative product, that is a much healthier relationship than giving everybody unlimited expensive computation and hoping advertising catches up later.

Could sponsorship be the easiest way for an indie AI feed to make money?

For a small AI feed, sponsorship may be the fastest route to meaningful revenue because one aligned sponsor can be worth far more than thousands of generic ad impressions.

As seen above, Infinite Slop itself already uses this structure. Fal covers the expensive generation while getting an unusually good demonstration of H3 Max in return.

The audience fit is unusually strong. Somebody watching an experimental interactive AI stream made by Pieter Levels is much more likely than the average internet user to care about AI models, developer infrastructure, GPUs, hosting or coding tools.

That means the sponsor does not need TikTok-scale reach. A few thousand highly relevant developers can be commercially interesting to an AI-infrastructure company even though the same audience would generate almost nothing through standard consumer ads.

This gives indie hackers a different path. A weird AI game for designers could attract a design-software sponsor. A continuous autonomous-agent competition could attract a developer-tool sponsor. A generative fashion channel could work with a retailer or marketplace.

Sponsorship has limits. Revenue usually requires direct relationships, and there is no automated marketplace guaranteeing that the next sponsor appears. The format works best when the audience has a clear commercial identity.

For experimental products, though, that can be enough. The first $5,000 or $20,000 of revenue may be much easier to find from one company that desperately wants those users than from millions of low-value impressions.

Is AI already making short-form video cheap enough to flood the market?

AI is already making short-form video dramatically cheaper to produce, especially when creators generate the content once and distribute it repeatedly.

Kuaishou's latest results give us a useful aggregate rather than another isolated AI-video demo. The supply of short plays on its platform, including live-action and AI-generated productions, increased more than fivefold between January and June 2026.

Money followed that supply. Advertising spending associated with short plays increased more than 100% year over year during the second quarter.

Kuaishou explicitly attributes part of the expansion to AI reducing production costs and lowering the barrier to creating content. At the same time, the company has built a profitable AI-video product in Kling, so it sees both sides of the market: people producing more video with AI and people paying for the tools that produce it.

That is a much more convincing commercial pattern than a viral AI clip. We are seeing production economics change across an entire content category.

The important detail is how the output gets used. A short drama can be generated or heavily AI-assisted once, then viewed hundreds of thousands of times. An AI advertising creative can be produced once and shown to millions of prospects. A digital livestream host can work continuously while selling products to many viewers.

These models spread one production cost across a lot of consumption.

The same cost decline does much less for a system that throws each clip away after one person watches it. Even if generation becomes ten times cheaper, a personalized stream can still lose badly against content that gets reused ten thousand times.

AI is already flooding short video with cheaper supply. Reuse is what makes that cheaper supply commercially powerful.

What kind of infinite AI feed could actually become a real business?

The strongest infinite AI feeds currently look narrow, persistent and transactional: viewers return to the same world, and the product has something more valuable to sell than another ad impression.

A persistent AI microdrama is one obvious shape. Thousands of viewers could watch the same base story while voting on what happens next. The most popular branch becomes the shared stream, while paying users can create private branches or unlock alternative scenes.

An audience-controlled game show has similar economics. One shared generation serves everybody, while voting, special commands, character creation or competitions can create paid actions.

Commerce gives us another version. A narrow AI fashion, home-design or collectibles feed could generate endless entertaining demonstrations around products that actually exist. The commercial event then becomes the purchase rather than the impression.

The feed could also be built around a creator or existing community. This solves one of the hardest problems Sora exposed: generated video alone gives people very little reason to care who made the next clip. A recognizable host, character or group gives the content continuity.

The weakest concept is probably the easiest one to build: an app that continuously generates unrelated AI videos for each user and pays for everything with ads. The unit economics are terrible today, the videos have almost no accumulated value and the swipe interface itself is easy to copy.

There is also little moat in saying "our feed is AI-generated." Meta already has Vibes, major video platforms are integrating generative tools, and model APIs make the underlying capability available to almost any developer.

A useful moat has to grow as people use the product. That could come from persistent characters, a community, proprietary interaction data, creator distribution, merchant relationships, an owned fictional universe or years of user history that make the experience progressively better.

The winning indie product probably feels less like "a cheaper TikTok" and more like a new form of interactive media that happens to use the TikTok interface.

So can infinite-scrolling TikTok apps actually make money?

Yes, infinite-scrolling TikTok apps can make money today, but the viable versions will usually reuse AI content, sell something valuable or charge for generation rather than producing a private movie stream for every passive viewer.

We have enough evidence now to draw that line quite sharply.

Personalized real-time generation funded by normal advertising makes almost no economic sense at current video-model prices. Our aggressive U.S. ad scenario still left the raw generation bill hundreds of times larger than advertising revenue.

Shared feeds look completely different. Once one generated stream serves hundreds or thousands of simultaneous viewers, compute gets divided across the audience and starts behaving like a normal content-production cost.

Direct monetization improves the model again. Short-drama apps are already generating roughly three quarters of a billion dollars of quarterly in-app spending. TikTok Shop shows that video can drive billions of dollars of commerce. Kling shows that customers will pay directly for AI-video generation. Sponsorship can make a small but highly relevant audience valuable long before it reaches mass-market scale.

The strongest business combines two things: cheap passive consumption and expensive high-intent actions.

Let everybody watch the shared world. Charge the people who want to control it, branch it, personalize it, buy through it or reach its audience.

That also points to where the opportunity is for indie hackers. Building the feed itself has suddenly become surprisingly easy. Building TikTok's creator network, recommendation graph and advertising machine remains absurdly difficult.

There is little reason to copy all of TikTok.

A much better bet is to take the addictive interaction pattern and attach it to something TikTok could not easily have offered before real-time generative video existed: a story that never ends, a world the audience controls, an AI game that continuously improvises, or a commerce channel capable of creating fresh demonstrations forever.

AI has made the feed cheap to fill. Attention is still expensive to win, and personalized video is still expensive to serve.

That is why infinite AI TikTok apps can make money, but "infinite slop with ads" is probably one of the worst ways to do it.