A clip has been going around claiming the future of television had quietly launched: an endless AI-generated channel that viewers steer by typing into a chat, running on a model that makes video faster than you can watch it. The figure attached to it — around $788,000 a year to run such a channel — got repeated everywhere, usually with the line “which is nothing, compared to broadcast.”
Both halves of that are roughly true. Neither is true in the way it was passed around. I checked the sources, because I have spent twenty years building real-time interactive visuals, and “real-time” is a word that gets used loosely by people selling things.
What actually shipped
fal.live launched on 31 August 2026 — a platform of always-on AI channels where the audience prompts what happens next. Within hours its streams were paused: the system was being refined, broadcasts would return soon. They returned on 1 September, and this time the platform describes them as running continuously, with prompts written by a language model that anyone watching can upvote to decide what gets made next.
As I write this the archive holds 405 recorded moments from seven streams, with new clips from 6 September. Several channels are live — an anime channel, a “chaos” channel running on H3 Max Director, a sitcom channel — and several others sit paused. An ongoing broadcast with part of the schedule dark, then, rather than a platform that is either fully on or fully off.
Infinite Slop, built by Pieter Levels with fal sponsoring the compute, went up two days earlier and reportedly drew 37,000 viewers on its first day. Same idea: an endless channel, the chat decides what airs.
Both grew out of something one fal engineer did for fun — hooking a continuously generating model to a Twitch stream and calling it “interdimensional cable”. He says he was banned from Twitch, then from Kick, and moved on. Neither platform has said anything publicly, so that part rests on his word.
What “faster than real time” means — and does not
This is the part worth being precise about, because it is where the excitement is.
fal’s H3 Max generates a five-second clip in under three seconds at 768p, by its own published timings. That is genuinely remarkable, and it is what makes an endless channel possible at all: if you can produce video faster than it plays, you can keep a buffer ahead of a viewer indefinitely.
But note what that sentence measures. It is assembly time against clip length, not a pipeline emitting frames on a deadline the way a game engine does. fal has not published the frame rate, resolution or latency of the live stream itself. For anyone who has shipped real-time 3D that distinction is the whole job: producing a second of footage in half a second is not the same problem as never missing a frame.
The more interesting claim is continuity. Chaining independent clips through their last frame is an old trick and it shows — characters drift, lighting jumps. H3 Max Director is fal’s answer: autoregressive, “native infinite continuity”, a context window of up to two minutes, holding the thread across scenes instead of resetting.
When I first looked at this, Director was an experimental checkpoint behind the streams with no public endpoint, and I wrote that announced is not shipped. That has changed. There is now a documented endpoint and a model id, and the “chaos” channel runs on it. It is still labelled experimental: it does not answer the ordinary request API at all, but publishes a session contract you open with a realtime client from an alpha package, and fal says it may change in a minor version. So it shipped — where “shipped” means you can build on it this week, not that you can plan an interface around it.
The number
Here is the part that moved, and how it moved matters more than where it landed.
The viral arithmetic is sound, and it quietly assumes four things: that the stream runs at 480p, that it generates exactly one second of video per second of airtime, that the price it used holds, and that the channel is built by chaining clips — which is not how the channel it described actually runs. Continuous generation is 86,400 seconds a day, so the rate does all the work.
| What you are paying | 480p, per second | A day, continuous | A year |
|---|---|---|---|
| Launch promotion, until 1 September | $0.025 | $2,160 | $788,000 |
| Deeper promotion, until 7 September | $0.0125 | $1,080 | $394,000 |
| List price, from 8 September | $0.05 | $4,320 | $1,580,000 |
The $788,000 everyone repeated is the first row. It was correct when it was worked out, and it is not a price you can pay now. The rate underneath it has already moved once since the model launched on 26 August: the page carried $0.025 on 27 August, half off and ending 1 September; that gave way to a deeper discount at $0.0125, and on 8 September it becomes $0.05.
That is the actual finding, and it is more useful than the number itself: at the list price that takes effect on 8 September, the same 480p channel costs about $1.58 million a year — twice the figure everyone quoted. Not because anything became more expensive in the ordinary sense, but because a launch discount ended, which is what launch discounts do.
Two things get blurred together here and are worth separating:
- Resolution is not a price change. 768p lists at $0.08 per second — about $6,900 a day and $2.5 million a year, 3.2× the viral figure. Getting there from the viral number means two changes at once, and only one of them is about price. The other is a decision about what you put on screen.
- Director is priced on its own clock, and it is the one to plan against. The continuous model that makes the whole idea work — the one the “chaos” channel runs on — bills a flat $0.02 per second of video generated, with no resolution tiers and a minimum of sixty seconds a session, until 14 September; $0.08 per second after that. Run continuously, that is about $1,700 a day and $630,000 a year today: cheaper than the viral figure, and cheaper than every row of the table above. From 15 September it is $6,900 a day and $2.5 million a year — the top of that table. The same fourfold step, on its own date, with no cheaper tier to retreat to.
And nobody is paying any of it yet on the channels you have seen — fal sponsors the compute for Infinite Slop, so the list price is not an invoice.
So is it cheap?
Against broadcast it depends on the resolution, which is a narrower answer than the one going around. A single regional channel’s playout, master control and staffing runs into millions a year before anyone makes a programme. At 480p list, $1.58 million a year for a channel that never repeats itself and answers its audience is still a different category of number. At 768p, $2.5 million a year is not undercutting broadcast — it is inside broadcast’s range, for a worse picture.
Against any other way of filling a screen, it is not cheap at all. A game engine rendering interactive 3D at 60fps costs approximately nothing per hour once it is built — which is exactly the business I have worked in for two decades. What generative video buys is not cheaper pixels. It is the removal of the production step: no modelling, no rigging, no animating, no set. You pay per second forever instead of paying once to build.
That trade decides where this belongs. For anything you will show a thousand times, build it. For something that must be different every time it is watched and is never watched twice — a channel that never loops, a backdrop that responds to who is in the room, an attraction that does not repeat for the person behind you in the queue — the economics now work where a year ago they did not.
Where this actually lands
Interactive entertainment, brand spaces, exhibitions, live events: the applications I would put this in first are the ones where content that never repeats is the point, and where 480p on a large surface at a distance is genuinely fine. That last clause carries more weight than it looks — the resolution that makes the arithmetic work is also the one you can least afford to have examined up close.
What I would do with Director now is prototype, not plan. On 3 September it was an announcement; today it is a documented endpoint with an experimental label and an alpha client, and the continuity problem it addresses is the one that separates a novelty from a medium. That is a real step. It is not a stable interface, and there is still no independent measurement of the two-minute context window.
Real-time generative video is real, it is here, and the honest version of the story is more interesting than the viral one. The number everyone repeated was a photograph of a discount. From 8 September the same channel costs twice that — and 3.2 times if you want it to look good.