Can you make money with a constant livestream?

Why did constant AI livestreams suddenly become possible?

Constant AI livestreams are becoming practical now because video models can finally generate new footage faster than viewers consume it.

That threshold sounds small, but it changes the product completely. fal says H3 Max can generate a five-second video in under three seconds. Pieter Levels reported getting roughly 15 seconds of video in around nine seconds while building Infinite Slop. The system can therefore prepare the next scene while the current one is still playing.

A few generations ago, the timing was completely different. Levels said a 15-second generation could previously take two to five minutes. That was fine for making clips, but hopeless for a genuinely continuous stream because playback would eventually catch up with generation.

Infinite Slop and similar experiments now work around that bottleneck by generating short segments continuously and queuing them before viewers need them. Rehan Sheikh used the same basic idea for an endless Twitch channel, while fal has since demonstrated its own H3 Max Live experiments where audience input changes what appears next.

That is a real technical line. A livestream can now invent footage after the viewer arrives and still keep broadcasting without stopping.

What did Infinite Slop actually prove?

Infinite Slop proved that people will show up for interactive AI television, but we still have no proof that they will keep coming back once the novelty wears off.

The product generates 15-second scenes continuously while viewers suggest what should happen next. Because only four 15-second clips fit into a minute, everybody is effectively competing for a small amount of shared airtime.

That interaction is probably more important than the infinite video itself. During the launch period, Levels said roughly 37,000 people watched over its first full day, while concurrent viewership at one point was around 400. His posts about the experiment also reached a much larger audience through X.

Those numbers are meaningful for something built extremely quickly, especially because viewers were actively trying to influence the stream rather than just watching passively.

But launch traffic tells us very little about retention. Levels already had a large audience, the concept was genuinely new, and people had a strong reason to click simply to understand what “infinite AI television” looked like.

The more interesting test starts later. Do people return when they already understand the trick? Do they watch for 30 minutes instead of two? Do recurring characters, jokes or conflicts emerge? Do viewers care enough about the stream to spend money inside it?

Infinite Slop has already passed the curiosity test. The habit test is still wide open.

Is Infinite Slop actually making money right now?

Infinite Slop currently looks more like a sponsored technology experiment than a proven standalone media business.

Levels has openly said that running the stream would be very expensive and that fal was generously sponsoring it. There is no public evidence showing that ordinary viewers are already generating enough subscription, advertising or transaction revenue to pay the full cost themselves.

Commercially, though, the project is still relevant. fal gets something unusually valuable from sponsoring Infinite Slop because the stream constantly demonstrates the exact advantage fal wants developers to notice: fast video generation.

Every uninterrupted sequence shows that H3 Max can stay ahead of playback. Every audience request shows that the model is usable interactively. Every discussion about how Infinite Slop works ends up talking about the infrastructure underneath it.

So the sponsorship itself already points toward one viable business model for constant AI livestreams. A stream can make economic sense when another company values the demonstration enough to subsidize the underlying compute.

We just should not confuse that with proof that random AI television can currently finance itself through viewers alone.

How expensive is a 24/7 AI livestream today?

A fully generated 24/7 AI livestream is still brutally expensive at current retail AI-video prices.

fal currently lists H3 Max at $0.05 per generated second for 480p and $0.08 for 768p after its introductory discount. A full day contains 86,400 seconds.

If we generated every second of a 768p livestream from scratch, the video-generation bill alone would come to about $6,912 per day.

That works out to roughly $207,000 over 30 days.

Even 480p would cost around $4,320 per day at the standard rate.

Those figures are API list prices rather than fal's internal cost, and large customers could negotiate much better economics. Infinite Slop itself is sponsored, so Levels almost certainly is not paying those theoretical retail amounts.

Still, the numbers show where the industry currently stands. The huge breakthrough has happened in speed. The same thing has not yet happened in cost.

Streaming delivery, moderation, encoding, databases and engineering would all come on top.

H3 Max output	Public price	Fresh video for 24 hours	30-day equivalent
480p at launch discount	$0.025/sec	$2,160	$64,800
768p at launch discount	$0.04/sec	$3,456	$103,680
480p standard price	$0.05/sec	$4,320	$129,600
768p standard price	$0.08/sec	$6,912	$207,360

Does broadcasting 24/7 actually make an AI livestream better?

A constant AI livestream only benefits from being live 24/7 when viewers have a reason to arrive at any hour.

Lofi music has that property. People study, code, work and sleep across every timezone, so immediate availability is part of the product.

Shopping can have it too. Someone can discover a product late at night and still ask questions or buy.

Persistent games and interactive worlds also benefit because viewers want to enter something that already feels alive rather than wait for a scheduled session.

Random AI comedy is harder. Generating another hour at 4 a.m. does not automatically create more value just because the server can do it.

This is where the economics get awkward. An uploaded YouTube video barely costs anything when nobody watches it. A frontier-video livestream can keep burning hundreds of dollars an hour in an empty room.

That makes “always available” a much stronger goal than “always generating.”

A smarter system could stay live continuously while generating heavily when viewers are present, slowing down during quiet periods, replaying older segments, or maintaining a buffer of content. The audience can still experience an endless channel without forcing the business to manufacture 86,400 completely new seconds every day.

Can ads pay for a constant AI livestream?

Ads can eventually support a constant AI livestream, but today the audience needs to be much larger than the early Infinite Slop numbers for the math to work.

Twitch currently lets qualifying creators receive up to 55% of net advertising revenue when they meet its ad-density requirements. We can use a simple model of six 30-second ads per viewer-hour to see the order of magnitude.

At $6,912 of daily 768p H3 Max generation, 400 average concurrent viewers create 9,600 viewer-hours per day. Six ads per hour gives 57,600 daily ad impressions.

The creator would need $120 of net advertising revenue per thousand impressions just to pay for the AI generation.

At 1,000 concurrent viewers, the requirement drops to $48.

At 5,000, it falls to $9.60.

At 10,000, it reaches $4.80.

That is where advertising starts looking much less ridiculous, assuming a platform such as Twitch or YouTube is absorbing most of the streaming infrastructure.

Self-hosting makes the threshold tougher. Cloudflare Stream currently charges $1 per 1,000 delivered video minutes. At 1,000 concurrent viewers, 24 hours of delivery would represent about 1.44 million viewer-minutes, or roughly $1,440 per day at that public rate. At 10,000 concurrent viewers, delivery alone would reach roughly $14,400 per day before discounts.

Those retail benchmarks will vary enormously at scale, but they reveal a useful pattern: once a self-hosted livestream gets big, distributing the video can become as important as generating it.

For a pure advertising model, platform distribution is therefore extremely valuable.

Average concurrent viewers	Viewer-hours per day	Net CPM needed to cover $6,912 generation
400	9,600	$120.00
1,000	24,000	$48.00
5,000	120,000	$9.60
10,000	240,000	$4.80
20,000	480,000	$2.40

Could paid prompts make Infinite Slop profitable?

Paid prompts are much more interesting than ads for Infinite Slop because viewers would be buying control over something genuinely scarce.

The stream may be infinite, but its airtime is not.

With 15-second generations, Infinite Slop has four new slots per minute. That creates 240 slots per hour and 5,760 over 24 hours.

At the current standard H3 Max price for 768p video, each 15-second generation costs about $1.20.

If every slot sold, an average price of $1.20 per prompt would theoretically cover the direct generation bill.

At 50% paid occupancy, the required average rises to $2.40.

At 25%, it becomes $4.80.

At 10%, it reaches $12.

Those amounts are far easier to imagine than asking a few hundred viewers to generate enough ad impressions to cover thousands of dollars of daily compute.

There is also a strong social effect. Paying $5 to make a strange scene appear in front of 10,000 other viewers feels more valuable than paying $5 to generate the same clip privately. The crowd makes the airtime itself worth more.

The risk is obvious: charge for every interaction and the chaotic crowd may disappear.

Levels has already mentioned that people suggested bidding for airtime, while saying he liked the free, messy nature of the current version. That instinct makes sense. A large free layer keeps the stream active, funny and unpredictable. Money could instead buy priority, extra voting power, persistent characters, special scenes or access to side worlds.

The best monetization probably keeps free users inside the game while letting paying users influence it more.

Share of daily slots sold	Average paid prompt needed to cover generation
100%	$1.20
75%	$1.60
50%	$2.40
25%	$4.80
10%	$12.00

Can subscriptions and donations support a constant AI livestream?

Subscriptions and donations can help a constant AI livestream once a real community forms, but they are a weak foundation for an expensive stream with only a few hundred regular viewers.

YouTube currently pays creators 70% of net revenue from memberships, Super Chat, Super Stickers and Super Thanks.

To cover $6,912 of daily video generation through that system alone, viewers would collectively need to spend roughly $9,900 per day before YouTube's creator share.

With only 400 average concurrent viewers, that would imply more than $1 of fan spending for every viewer-hour watched.

At 10,000 concurrent viewers, the burden becomes much more reasonable.

Community funding also depends on what people become attached to. Traditional livestreams often revolve around a recognizable creator. A fully autonomous AI channel has to build that attachment in some other way.

Persistent characters could do it. Shared history could do it. Rival groups, ongoing jokes, recurring storylines and collective goals could do it.

Random scenes with no memory make that much harder because viewers have very little reason to become members rather than just stop by whenever a funny clip goes viral.

Is sponsorship actually the best business model for AI livestreams right now?

Sponsorship is currently one of the strongest business models for expensive AI livestreams when the stream naturally demonstrates what the sponsor sells.

Infinite Slop is a particularly clean example.

fal sells fast generative-media infrastructure. Infinite Slop shows fast generative-media infrastructure working continuously in public.

The sponsorship therefore creates more than logo exposure. Developers can see latency, output quality and real-time interaction directly.

fal has even pushed the idea further with its own H3 Max Live experiments, which confirms that it sees always-on interactive generation as a useful showcase for the model.

That logic will not transfer equally well to every sponsor. A random consumer brand would need to justify spending serious money purely for audience exposure. An AI infrastructure provider can also count developer acquisition, technical credibility and product demonstration.

The strongest sponsorships should have that same close connection.

A game engine could sponsor a persistent generated game world. A robotics company could back an endless simulated robot competition. An AI voice platform could fund a nonstop interactive radio station.

The tighter the relationship between the experience and the underlying product, the easier the economics become.

Do people actually keep watching endless AI livestreams?

Endless AI content currently has a serious retention problem, and Nothing, Forever is the clearest warning.

The AI-generated Seinfeld-style Twitch stream exploded in 2023. TechCrunch observed more than 15,000 simultaneous viewers during the viral period, and Twitch tracking recorded an all-time peak above 16,000.

The stream could theoretically continue forever, which was a big part of the fascination.

It did continue.

The audience largely disappeared.

Recent TwitchTracker data has shown only a handful of average viewers despite the channel still having more than 130,000 followers. Compared with its peak, active viewership collapsed by well over 99%.

Lofi Girl shows the opposite pattern. Its flagship 24/7 music stream can still attract roughly five figures of average concurrent viewers, while the broader channel has accumulated billions of views and more than 15 million subscribers.

The difference is repeat usefulness. People keep needing background music tomorrow, next week and next year. They do not keep needing proof that AI can generate an endless sitcom once they have already seen it.

That comparison is a warning for Infinite Slop. Viral curiosity can create a huge launch while telling us almost nothing about long-term demand.

A constant livestream needs some reason to return after the viewer understands the technology.

Stream	Core reason to watch	Peak or current scale	What happened over time
Nothing, Forever	Novel endless AI sitcom	Peak above 16,000 concurrent viewers	Fell to only a handful of average viewers recently
Lofi Girl	Background music for studying and working	Roughly five-figure average concurrency on the main stream	Remains one of YouTube's best-known 24/7 channels

Does letting viewers control an AI livestream make people stay longer?

Viewer control gives constant AI livestreams a much stronger reason to exist because the audience can create consequences together.

Twitch Plays Pokémon remains the best historical example of how powerful that can become.

During the original Pokémon Red run, more than 1.16 million people entered commands through Twitch chat. Twitch later said viewers sent more than 122 million commands, watched more than one billion minutes and pushed the audience above 120,000 concurrent viewers at its peak.

The game itself was old. The crowd made it unpredictable.

People returned because yesterday's actions changed today's situation. A bad command could release an important Pokémon. Thousands of viewers could spend hours trying to walk through one doorway. Groups formed around different strategies and characters.

Infinite Slop already has part of that dynamic because viewers influence what comes next. The weaker part, for now, is persistence.

A strange character can appear for 15 seconds and then vanish. The world does not yet reliably remember enough for viewers to develop the same kind of long-term investment.

Longer continuity could change that dramatically. fal has already experimented with models designed to preserve audiovisual continuity across successive generations.

If AI livestreams start remembering characters, locations, possessions, rivalries and previous viewer decisions, we get something much closer to a multiplayer game than an endless video generator.

That feels like a much stronger product.

Is livestream commerce the clearest proof that constant streams can make money?

Livestream commerce gives us the strongest current proof because viewers can buy something directly instead of producing tiny amounts of advertising revenue.

The scale can be surprisingly large.

The Wall Street Journal recently documented coin seller Bjorn Bergstrom running a 177-hour Whatnot livestream with a rotating team. The broadcast generated roughly $2.05 million in coin sales and around $155,000 in profit after costs and giveaways. His strongest months have gone well beyond that.

China has taken the model further, especially with digital hosts that can remain online when human sellers would normally stop.

The economics are attractive because one buyer might generate $20, $100 or $1,000 of gross merchandise value. A media stream might need thousands of ad impressions to create the same amount of revenue.

There is still a geographic difference. In the US, TikTok Shop is growing extremely quickly, but livestreaming remains a minority of purchases. Momentum Works estimated $11.8 billion of US TikTok Shop GMV during the first half of 2026, with live accounting for only 8.2% of attributed GMV. The Shop tab and normal shoppable videos accounted for much more.

So livestream shopping is hardly the universal future of commerce.

What it proves is narrower and more useful: constant broadcasting works much better when each minute can lead to a high-value transaction.

Can AI livestream hosts actually sell products?

AI livestream hosts are already generating serious sales in China, so this part of the model has moved well beyond demos.

WIRED reported on Chinese merchants using AI sales hosts built with systems from companies such as Baidu and DeepSeek. One virtual presenter selling Brother printers reportedly increased livestream sales by around 30% compared with human representatives and generated roughly $2,500 in two hours.

JD.com produced a much larger experiment by creating an AI version of founder Richard Liu. The broadcast attracted more than 20 million views during its first hour and generated around RMB 50 million in sales over the full livestream.

An AI recreation of entrepreneur Luo Yonghao later attracted more than 13 million views during a six-hour broadcast and generated more than RMB 55 million, roughly $7.7 million at the time. In several categories, order volume reportedly beat Luo's previous human-hosted debut.

These examples have obvious advantages over Infinite Slop. Digital sales avatars are cheaper to run than constantly generating cinematic video, and every successful session sells actual inventory.

The commercial lesson is strong. One of the first major categories for constant generative livestreaming may simply be salespeople who never need to sleep.

Will YouTube and TikTok monetize endless AI livestreams?

YouTube and TikTok currently leave room for AI livestreams, but automated low-value content can easily fall outside their monetization and recommendation rules.

YouTube now says monetized content should not be mass-produced, generic or excessively repetitive. Its guidance specifically raises problems with generic AI-made material and disconnected clips that offer little original value.

The platform still allows AI-assisted content when creators add real creative input, narrative, distinctive characters or other original value.

That gives interactive AI streams a better position than simple automated clip factories. A shared story controlled by viewers has a clearer creative structure than a bot endlessly pushing interchangeable generations.

TikTok takes a similar approach with LIVE. Its guidelines can limit recommendation for streams dominated by repeated actions without a clear objective or meaningful interaction.

For an Infinite Slop-style product, interaction therefore helps twice. It makes the experience more interesting and also gives the platform a stronger reason to see the stream as an actual live product rather than automated spam.

Anyone building this today should assume that platform tolerance will depend heavily on how repetitive and autonomous the final experience feels.

Is moderation the hidden problem with constant AI livestreams?

Moderation is one of the biggest practical risks because an autonomous stream can generate a policy violation while nobody is watching the output internally.

Nothing, Forever learned this quickly.

Twitch suspended the stream after its AI comedian produced a transphobic monologue. The creators later explained that technical problems had pushed them onto a different model and moderation setup.

That failure happened in a relatively simple text-driven AI show.

Modern video models add far more surface area. Users can request celebrities, copyrighted characters, violence, sexual content, brands, political figures and realistic fake events. A single generation can create problems with platform safety rules, copyright rules or synthetic-media disclosure requirements.

The scale also compounds the risk.

A stream generating 5,760 clips per day would produce more than two million clips over a year.

Even a tiny failure rate becomes uncomfortable when the system is generating millions of outputs in public.

So moderation has to sit inside the generation pipeline. Prompt filtering, output checks, model restrictions, user penalties and emergency shutdown systems all become part of the basic operating stack.

Anyone treating moderation as something a human can casually review afterward is going to have a bad time.

What kind of constant livestream can actually make money?

The constant livestreams with the best economics today combine cheap output with a strong reason to return and something valuable viewers can do inside the stream.

The same pattern shows up across very different examples.

Lofi Girl works because people repeatedly want the product and keeping the channel alive is cheap.

Whatnot sellers and Chinese digital hosts work because viewers can buy products directly.

Twitch Plays Pokémon showed how a shared persistent state can turn spectators into participants.

Infinite Slop points toward another model where scarce influence over the broadcast itself could become the thing people pay for.

Sponsored experiments also work when the broadcast demonstrates an underlying technology, as fal is currently doing around Infinite Slop and H3 Max.

The difficult version is pure entertainment where every new second costs money, nothing persists, viewers have no useful action to take, and advertising is expected to pay the entire bill.

That setup requires either a huge audience or dramatically cheaper generation.

The most promising AI livestreams will probably look more like games, stores, communities and interactive worlds than traditional television channels.

Can you actually make money with a constant livestream today?

Yes, constant livestreams can make money today, but pure 24/7 AI-generated entertainment still has weak economics unless the stream adds commerce, paid interaction, sponsorship or a very large recurring audience.

The strongest proof already exists outside the “AI slop” category.

Long livestream commerce sessions can generate millions of dollars in sales. Digital AI hosts in China can sell products continuously. Lofi Girl shows that a cheap 24/7 utility stream can maintain a huge recurring audience. Twitch Plays Pokémon showed how powerful a shared interactive world can become.

Infinite Slop adds something genuinely new to that picture. Real-time video generation is now fast enough for viewers to influence an endlessly generated broadcast while it is happening.

The remaining problem is economic.

As seen above, continuously generating 768p H3 Max footage at current public pricing would cost roughly $6,900 per day before delivery, moderation and engineering. Advertising alone becomes comfortable only when the stream reaches several thousand persistent concurrent viewers or generation costs fall sharply.

Direct interaction has much better math. With only 5,760 fifteen-second slots available each day, a relatively small number of paying viewers could potentially finance generation by purchasing influence over what appears next.

Retention will decide whether any of this becomes a lasting category.

Nothing, Forever showed how quickly fascination with “AI that never stops” can disappear. A viewer only needs to understand that trick once. Persistent characters, shared history, utility, commerce and meaningful control give people reasons to come back after the novelty is gone.

So yes, we can make money from a constant livestream now.

The strongest opportunity these days is to use the livestream as the interface to something people value enough to revisit, influence or buy.