Oct 2, 2026 · @Sai krishna
Video, made for 1
Video moves from something made once and sent to many, to something directed once and rendered for each person, often while they watch.
tap a cell
scroll. every idea draws itself ↓
the setup
The bakery and the kitchen
A bakery bakes one batch of bread and sells it to everyone who walks in. A kitchen cooks one meal for the person at the table.
Video has always been a bakery.
A crew shoots the film once. An editor cuts it once. A studio ships one master file, and a million people watch the same two hours. Every step in that chain costs real money, so the chain runs once and the output gets copied. Copying is free. Making is expensive. That one fact set the shape of the whole business: studios, networks, channels, ad slots, the thirty second spot, the three act structure, the Friday release date.
The internet changed half of that equation. It made copying and sending free, which moved the hard problem from distribution to discovery. The companies that solved discovery (YouTube, Instagram, Netflix) became the new power centers, and they gathered the audience that every bakery now has to sell through.
Generative video models change the other half. They make the making cheap. If that holds, the hard problem moves again, and so does the power.
three verbs
Record, digitize, generate
There are three verbs in this story.
The first was recording. Cameras and projectors arrived in the 1890s, and video became something you capture.
The second was digitizing. Through the 1990s, video turned into files, and files can be copied, edited and sent at no cost.
The third is generating. A model writes the pixels, and the cost of a new shot trends toward the cost of compute.
the whole argument
One picture
Video lived in the top left for a century.
Phones and platforms pulled a lot of it to the bottom left.
Generation opens the bottom right, and that is where this essay is going.
The thesis is simple to say. Over the next several years, video moves from something made once and sent to many, to something directed once and rendered for each person, often while they watch.
The rest of the essay builds that in order. First, how the models got here and what still breaks. Then India, where the screen is a phone and the habits are a few years ahead. Then micro dramas, the first format that already behaves like a kitchen. Then YouTube, which has begun to test the video itself. Then real time. Then the question that matters most: when every shot can be generated, what is left to direct? Last, the stack a company would need to serve all of this.
One rule for reading it. Facts carry a number or a name. Forecasts say Forecast in the sentence.
One bakery. One batch.
A million identical loaves.
Or a kitchen that cooks for each one.
how the models got here
From clips to scenes
Four technical changes took video models from a few silent seconds to a scene with a cast, a location and dialogue. Each one is simple to state.
Compression. A minute of raw video holds too many numbers for a model to read. So the model reads a compressed version, called latents, and a decoder turns latents back into pixels. MiniMax says the tokenizer in H3, which it calls H3-VAE, gives a fourfold gain in effective sequence length. Sequence length sets how much video a model can hold in mind at once. More compression buys longer scenes for the same compute.
One transformer for space and time. Current systems treat a video as one long sequence of patches and denoise the whole sequence together. A hand in second two has a better chance of staying the same hand in second twelve, because the model sees both at once.
Sound moved inside the model. Seedance 2.5 keeps the unified audio-video joint-generation architecture that arrived with Seedance 2.0. MiniMax H3 models voice, sound effects and music together in the same pass. Lips, footsteps and music land on the same beats. A character who speaks on screen is the line between a clip and a scene.
The prompt grew into a brief. ByteDance released Seedance 2.5 on July 31, 2026. It takes up to 30 images, 10 video clips and 10 audio clips as references. It makes up to 30 seconds per generation and extends across rounds. It edits by timestamp, swaps backgrounds and moves the camera after the fact. H3 models multiple shots natively and outputs 2K video with stereo sound. Both models accept a cast and a location as input. That is the shift that matters: describe a video → direct one.
- earlytext → [ model ]a few seconds, silent
- 2026text + 30 images + 10 clips + 10 audio → [ model ]a scene, with sound
- nextall of that + the viewer, live → [ model ]frames, as you watch
the price of a second
Weights get opened, then tuned
MiniMax prices H3 below a third of mainstream models per second at 2K, and it opened the weights. HeyGen took that open base and post-trained it for business video. HeyGen Video 1.0 launched on September 30, 2026.
At its October rate of one cent per second, 1,000 seconds at 768p costs $10, against $80 on base H3 according to HeyGen's own catalog. A thirty second ad costs thirty cents in compute. HeyGen's later rate is higher, and its own pages give two different figures.
This is the second H3 post-train in five weeks. The first was fal's H3 Max in August.
Forecast: base models turn into a shared layer, like a database engine. The edge moves to whoever owns the tuning data and the job. HeyGen owns business video: talking presenters, product explainers, localized spokespeople. A commerce company, a studio and a school will each want their own post-train.
One caution on the numbers. HeyGen's quality claim comes from an internal test of 4,800 votes. 55.8% preferred HeyGen Video over H3 Max, with a range of 49% to 62%. The low end of that range is a coin flip. No third party has tested it yet.
the honest list
What still breaks
Length is the first limit. Generations run 15 to 30 seconds. A micro drama episode runs longer, so scenes get stitched.
Detail is the second. HeyGen lists its own weak spots: long on-screen text, organic motion like hair and petals, and hands doing detailed work with equipment.
Consistency is the third. Seedance says extensions keep characters and environments consistent. ByteDance published no benchmark for that claim.
Rights are the fourth. Seedance 2.0 arrived in February 2026 and drew cease-and-desist letters from the Motion Picture Association, Disney and Paramount Skydance. Two US senators asked ByteDance to shut it down. Forecast: every model that makes film-grade scenes meets the same wall.
Speed is the fifth. These models run as jobs. You send a brief and wait for the render. Real time needs a different design, and a later section covers it.
A model that takes a brief at a cent a second changes what a video costs. The next question is who needs that many videos. India is the place to look.
where the limits press hardest
Why India sees it first
The easiest place to watch a medium change is where its limits press hardest. India presses on three: the screen, the language and the money per viewer.
The screen is a phone. The IAMAI and Kantar report counted 958 million active internet users in India in 2025, up 8% on the year. 548 million of them, 57%, live in rural India.
588 million watch short video, which is 61% of the base.
Only 20% use more than one device, and 18% borrow someone else's phone. For most of the country the phone is the television, the cinema and the shop window. Ericsson's November 2025 report puts the average Indian smartphone at 36 GB of data a month, the highest in the world.
FICCI and EY add the time. Indians spent 1.23 trillion hours on their phones in 2025, and 59% of that went to media and entertainment. Instagram, Facebook and YouTube Shorts carried 46%, 29% and 26% of short video viewing, since many people use more than one.
The ad money already moved. Digital advertising reached INR 947 billion in 2025, up 26%, and now makes up 63% of all ad revenue. Linear TV advertising fell 10%. The fastest pool was e-commerce and point-of-sale advertising at INR 220 billion, up 50%. Paid video subscriptions stand at 216 million across 143 million households. Connected TV grew from 30 million to 40 million sets in a year, and those sets stream 85 hours a month of OTT.
The language is many. 56% of content produced for digital platforms in 2025 was in regional languages, up from 48% in 2023. A story that reaches Hindi viewers misses most of Tamil Nadu, Telangana, West Bengal and Maharashtra. Every piece of video in India carries a version problem.
× 5 languages that matter first (Hindi, Tamil, Telugu, Bengali, Marathi)
× 3 formats (vertical, square, wide)
× 4 audiences (age, city, taste, price band)
= 60 versions(illustration)
A bakery handles that with a dub in two or three languages and one cut. It covers a slice of the audience and calls it reach. A kitchen handles it by rendering the version each person needs. India has the most to gain from that shift and the least budget to pay for the old way.
The budget is the third limit. Micro drama apps earn about $15 per paying user, against $35 for OTT, per BusinessToday. A low price per viewer forces a low cost per minute. A low cost per minute pulls production toward generation. Markets with a high price per viewer can keep the bakery longer. India cannot.
The next section looks at the one format that already works like a kitchen.
958 million phones.
Each one is a TV, a cinema and a shop.
Each one speaks its own language.
a kitchen with a crew in it
Micro drama
The format looks small. A micro drama episode runs one to two minutes, shot vertical at 9:16, and a series holds 20 to 100 episodes. Every episode ends on a cliffhanger. The first few are free, then a paywall asks for coins or a subscription.
The format began in China around 2018 on Douyin and Kuaishou. By June 2024, 576 million Chinese users watched it, which was 52.4% of internet users.
The Indian version has its own shape. About 100 million people open a micro drama app each month and 17 million pay. The money works differently from OTT. Advertising is under 15% of revenue. UPI AutoPay carries 70% to 80% of subscriber authorizations, and up to 60% of subscribers would churn without it. Revenue per paying user is about $15.
What the format teaches
The paywall sits on a specific second of a specific video. A subscription pays for a library. A micro drama paywall pays for the next episode, so the business measures one cliffhanger against one decision to pay. Retention and revenue become the same number, read per episode.
The unit is small enough to measure. A 90 second episode has one hook, one turn and one cliffhanger. Platforms use AI to score scripts, locate hook points, pre-visualize scenes, dub, add music and replace props and vehicles in post, according to Hollywood Reporter India. Data picks the next scene before anyone shoots it.
The production line looks like a factory. In India a 75 to 80 minute show costs INR 20 to 25 lakh according to Hollywood Reporter India, about INR 20,000 to 30,000 a minute, and takes 15 to 20 days. FICCI and EY use a lower figure of INR 8,000 to 9,000 a minute. A Western series of 50 to 70 episodes costs $150,000 to $250,000 and shoots in 7 to 10 days. Platforms say AI-led production brings an Indian show down to INR 2 to 3 lakh, and BusinessToday reports timelines cut by more than 90%. FICCI and EY put the average AI cost saving at 20% to 25%. The gap between the two tells you adoption is early and uneven.
The genre is a template. CEO romance, revenge, a hidden heir. Templates are easy for a model to render and easy for a director to constrain. A format with known rules is where generation lands first.
Micro drama sits at bottom left. Data already decides what gets made and where the story stops. Crews still make the pixels. Each cost that generation removes moves the format one cell to the right.
The limits
Profit is thin. BusinessToday notes that few platforms disclose profitability, and Pocket TV has shut down. Completion rates run from 50% to 60% in one report and above 90% in another, so the definitions vary. Reelpulse's 2026 industry report describes the working model as hybrid: human writers, directors and actors keep creative control while AI handles iteration and localization.
That hybrid is the point. The format shows what happens when production cost falls and measurement gets precise. The audience becomes the editor. The next question is how far the audience can edit, and a platform that already holds the whole audience has begun to answer it.
the platform steps in
YouTube starts testing the video itself
On September 23, 2026, at Made on YouTube, Neal Mohan announced that creators will be able to upload up to three cuts of the same video and let YouTube show them to real viewers. The cuts can differ in hook, pacing, length and the way the story is told. YouTube compares retention curves, which show where viewers leave each cut. The creator reviews the results, picks a winner and locks it in. The feature reaches select creators in 2027 and covers both long videos and Shorts.
This is the end of a long climb. YouTube began testing thumbnails in 2024. It added three-way tests for titles and thumbnails in December 2025, and one report counts more than 40 million of those experiments. Each step moved the test closer to the content.
- thumbnailYouTube, automatically2024
- title + thumbnailYouTube, automaticallyDec 2025
- the cut (up to 3)the creator, from retention2027
- the video, per viewera model, in the loopforecast
The package came first because the package was cheap to vary. A thumbnail takes ten minutes. A cut takes a day of editing. The cost of making variant number two set the limit on what anyone tested. Generation lowers that cost, so the limit moves, and it moves toward the video.
Second-order effects
A test becomes a policy. A test with three arms and a human picking the winner gives one answer for everyone. A model that can render fifty hooks at a cent a second needs a different setup: show each viewer the variant that fits, and keep learning. That is a bandit, and it has no single winner. It has a rule for who sees what. Forecast: once variants are cheap, the unit YouTube manages is a set of cuts plus a rule for assigning them. The A/B test is the first form of personalization, dressed as an experiment.
The shared object breaks. Marques Brownlee put it plainly: "There's this sort of unspoken rule of community on YouTube that we're all watching the same video." Comments, quotes, reaction clips and timestamps all assume one artifact. If two viewers see different cuts, "3:41 is hilarious" points at different frames. He also asked whether viewers will know they are in a test. YouTube gave no answer to either question in the coverage I found. Forecast: comments anchor to scenes with stable IDs that survive across cuts, and not to seconds. A video becomes a scene graph with a master reference, and every variant maps back to it.
The objective is thin. The test scores retention. Brownlee again: "It's helping you find which version has the best retention, which doesn't necessarily mean it's a better video." Micro drama already shows where a retention objective leads. A hook in the first seconds, a cliffhanger at the end, a cut tuned to the curve. Forecast: that grammar spreads from vertical drama to all video, because the test rewards it.
Meaning can change inside a variant. Critics already ask what happens when a creator cuts a positive and a negative version of the same review and keeps whichever draws more traffic. Forecast: platforms split variants into two classes. Presentation variants (hook, pacing, language, length) are allowed. Claim variants (what the video says about a product or a person) are restricted or disclosed. That line will carry more weight than any model benchmark.
The creator's job moves. Today the creator edits one cut and picks among three. With cheap variants the creator writes the space: which hooks are allowed, which lines must never change, which characters stay fixed, which languages come first. The direction becomes a set of rules and a taste for what to keep. The next section returns to this.
The platform gains the loop. Whoever holds the viewer, the traffic and the retention data runs the test. A third-party tool can make a hundred variants. It cannot serve them to an audience or read the curve. So variant generation spreads across many vendors while variant serving stays with the platform. YouTube already holds that position, and this feature uses it.
India sharpens each of these. A video there often needs a version per language before it needs a version per hook. A platform that can test a Tamil cut against a Hindi cut on the right viewers learns faster than any agency can brief.
YouTube's test changes the cut in advance. The viewer still watches a finished video. The next step changes the video while the viewer watches.
the fourth row
When the video is made while you watch
There are four moments when a video can be made.
- before releaseweeks · film, TV, a micro drama todayweeks
- before the testdays · YouTube's three cutsdays
- when you open ita localized, personalized adseconds to minutes
- while you watchMeta's Muse avatar, Flam's agentsunder one second
The first two rows are the bakery with better ovens. The third row is a kitchen that cooks when you order. The fourth row is a kitchen where the cook talks to you while the food is made. Each row down adds the viewer's own input to the video.
The fourth row now has a number. On September 23, 2026, Meta introduced Muse Realtime Avatar. It takes a photograph, an illustration, an animal or an object and animates it as a character that talks with you. The response arrives about 870 milliseconds after you finish speaking. It streams 448x768 portrait video at 25 frames per second.
How a model learns to keep up
A normal video model sees the whole clip at once and refines every frame together. That works offline. Live video asks for a frame every 40 milliseconds, and the model cannot see the future because the viewer has not said it yet. Meta's design shows the three changes this takes.
Causal chunks. The model is a diffusion transformer that generates short chunks, conditioned on a stream of speech tokens, reference images and a rolling window of its own recent latents. Each chunk covers 8 frames, which is 320 milliseconds of playback. The model needs about 20 milliseconds of compute to make it. That is 16 times faster than playback. The margin pays for concurrency: one NVIDIA GB200 serves 12 sessions at once.
Distillation. The teacher model needs 120 network evaluations per chunk. The student needs two. Meta reports a 60x cut. This is the same trade the whole field will make: spend heavy compute to train a slow model that is right, then teach a fast model to imitate it.
One stream for voice and picture. Muse Realtime Voice produces speech tokens, and the avatar model reads the same tokens. Lips, expression and voice stay locked together because they share a source. There is no separate lip-sync step to run late.
Research outside Meta points at the next target. A paper called Wonder, posted in July 2026, builds a camera-controllable world model on a 14 billion parameter base. It runs at 16 frames per second with four-step streaming generation, and it keeps a sparse-attention memory so that a place you leave looks the same when you return. Faces come first because a face fills a small frame and speech gives the model a strong signal. Scenes come second.
Cost changes shape
A broadcast video pays its compute once and spreads it over every viewer. A live video pays compute per viewer, per minute.
The price in that line is my placeholder. The structure holds at any price. Video that used to be a capital cost becomes a running cost that grows with attention. Forecast: the companies that win here sell outcomes like a sale closed or a lesson finished, and they charge per outcome, because per-minute pricing makes every viewer a cost line.
Flam shows where a buyer uses this today. It sells playback that starts in 50 milliseconds and characters that answer in under two seconds, to customers like Google, Grab and KFC. A shopper taps a product in a video and a character talks about it.
India fits the fourth row better than any market. 18% of internet users borrow a phone. Many type slowly in a second language and speak fluently in a first. A voice-in, video-out interface sits well with that user: a presenter who speaks Tamil, answers a question about a kurta and shows the fabric. Forecast: voice becomes the main input for shopping and learning on cheap phones, and the answer arrives as a face on screen.
The limits
The resolution is low, a little above a standard-definition portrait frame. The characters are faces and upper bodies. Meta offers no API, no price and no release date. Its comparison against Runway Characters (78% to 22%) and HeyGen LiveAvatar (88% to 12%) used Meta's own raters on conversations of two to three minutes, so long-session stability is untested. Meta's own note says that not every avatar in its examples ships in the Muse app yet. Its model carries the Video Seal watermark for provenance.
Watermarks matter here. A live synthetic face that sells you something needs a mark that says so. Forecast: provenance moves from a policy line to a technical requirement in real-time video, and platforms read the mark before they serve the stream.
One more shift follows. A live video cannot be edited after the fact. The only direction available is direction in advance: who the character is, what the character may say, how the voice sounds, what the character must never do. The director writes a spec and the model performs it. That is the subject of the next section.
what stays scarce
When every shot is free, direction is the product
When an input gets cheap, the scarce thing moves to the next step in the chain. Cameras on phones made photos free. Taste in photos got expensive. Printing made text free. Editors got expensive. Generative video makes the shot free. The next step in the chain is deciding which shots exist, in what order, and for whom.
Here is the map of what gets cheap and what stays scarce. The bottom row is the job. Six pieces of it follow.
Taste is selection. Films have always shot far more than they use. Generation takes that ratio to infinity. Someone has to say no again and again and yes once, and that person needs a reason. The reason is taste. HeyGen's own list of failures, long on-screen text, hair and petals, hands working with equipment, is a list of things a person must catch before a viewer does.
Character is an asset with a file. Seedance 2.5 accepts up to 30 images, 10 clips and 10 audio files as references. A character stops being a prompt and becomes a package: face, voice, wardrobe, mannerisms, the way the person walks into a room. The package belongs to someone. A brand mascot, a drama lead and a store presenter are each one of these. Forecast: the character package becomes the main intellectual property in generated video, the way a font or a logo is today.
Memory lives outside the model. A model holds a short memory in its context. Meta's avatar keeps a rolling window of recent latents. Wonder keeps a cache and retrieves from it so a revisited place looks the same. Both cover seconds or minutes. A micro drama runs 60 episodes and a creator runs for years. That memory has to sit in a database: who knows what, who wronged whom, which jacket the lead wore in episode nine. Forecast: the series bible turns from a document into a structured store that the model reads on every render.
Editing is rhythm. A model can cut between shots, and H3 and Seedance both model multiple shots natively. Where a scene breathes and where a cliffhanger lands is a judgment about a human attention span. Data now informs it. Micro drama platforms score hook points, and YouTube will score cuts by retention. Brownlee's warning applies: best retention is not best video. A good director knows which part of the curve to ignore.
Narrative is the set of things that cannot change. Personalization raises a design question that film never had to answer. Which parts of this story may vary by viewer, and which parts may never vary? The language, the city, the product on the table and the length can vary. The premise, what the character wants and what the ending means must hold. Forecast: directing becomes the act of writing those invariants, and the quality of a personalized video depends on how well they are written.
Context is a new input. The same video lands differently on a shared phone in a Tier III town, a connected TV in a living room and a commute. In India 18% of internet users borrow a device, and only 20% have more than one. Identity-based personalization breaks on shared phones. Forecast: personalization in India runs on the session (language, time of day, what was just watched, what the viewer said) and does not wait for a profile.
software has two forms
Source and build
Software has two forms. Source code is what the author writes. A binary is what ships. Video has only ever shipped binaries: a rendered file, identical for everyone.
The source is the direction: scripts, character packages, scene IDs, invariants, approved variants and the rubric that says what good looks like. Generation is the compiler. The compiler can run at the studio, at the platform or on the viewer's device.
This changes who has leverage. YouTube today holds the binary and can choose among cuts that creators made. To render a new cut it needs the source. Forecast: the contest of the next few years is over who holds the source. Creators who keep their character packages and invariants keep the ability to build for any platform. Platforms that hold the traffic want the source so they can build for each viewer.
what a company would need
What's under the kitchen sink
A consumer company that wants to serve video this way needs more than a model API. A model API is one layer in six. Tap a layer to see what exists in October 2026 and what someone still has to build.
The number that runs the stack
Orchestration hides one metric that decides the economics. Yield is the share of generations a director accepts.
The yield and the exchange rate in that picture are my assumptions. The one-cent price is HeyGen's October rate. Today's Indian micro drama runs INR 8,000 to 9,000 a minute (FICCI and EY, the low figure), which at about INR 90 to the dollar is roughly $140 for 90 seconds. Even with a pessimistic yield, compute costs a small fraction of a crew. Once compute is that cheap, the director's review time becomes the largest cost. Forecast: the best tool in this stack is the one that cuts review time per accepted second, with automatic checks that reject a bad take before a person sees it.
Who owns what
Forecast: generation and inference commoditize. Open weights, a second H3 post-train in five weeks and a price per second near one cent say the same thing. Whoever sells raw generation sells at falling prices.
The layers with lasting value sit at the top and bottom of the stack. The top is the source: characters, invariants, scene IDs and the rubric for good. The bottom is the loop: the audience, the personalization policy, the retention and conversion data and the payment rails. A consumer company holds the loop. A studio holds the source. The company that holds both can build a video for each viewer and learn from each one.
The source layer suits studios that serve more than one job at once. StarCent, which ran a front-page ad in The Economic Times on October 1, 2026 beside partners such as Flam, ShopOS and Grapes, lists its work as branding, marketing and entertainment. Those three jobs share one source. The brand's invariants are the branding. The variants are the marketing. The series is the entertainment. A studio that writes the source once and serves all three has a position that survives any single model.
the conclusion
Directed once, rendered for each
Video is turning from something you finish into something you author. The finished file stays, and it becomes one build of a source.
The evidence stacks up in order. Models take briefs now: references, sound and multiple shots, from Seedance 2.5 and MiniMax H3, tuned for business by HeyGen at a promotional cent a second. India shows the pressure that makes this necessary: a phone for a screen, many languages, $15 per paying user. Micro drama shows the loop: data picks the next scene and the paywall sits on the cliffhanger. YouTube shows the platform stepping in, with three cuts of a video tested on real viewers from 2027. Meta's Muse avatar shows the far end: a video made while you watch, 870 milliseconds after you speak.
Direction is what remains scarce through all of it: the source, the invariants, the memory and the taste to keep the right take.
Three stages
Stage one: variants. 2026 to 2027. Cuts, hooks and languages, tested before and during release. Stage one is under way. YouTube's cuts, micro drama hook scoring and HeyGen's localized presenters are all variants.
Stage two: builds. 2027 to 2029. A cut per context, picked by a policy that checks the invariants. Stage two needs three things that do not exist yet: a source format that carries invariants and scene IDs, serving that reads context without a profile, and quality control that rejects bad takes before a person looks.
Stage three: performances. 2029 onward. Characters and then scenes rendered live on a phone. Stage three needs real-time models that move from faces to scenes at phone resolution, and inference cheap enough to pay for from an ad or a sale. Distillation points the way, since Meta cut 120 network evaluations to 2 for one chunk.
The dates are my estimates. The order is the argument.
What would prove this wrong
Any one of these can hold in one market and fail in another. India has the least to lose from the old arrangement, so it is where the thesis is easiest to test.
Where this lands
For a hundred years the unit of video was the file. A crew made it, a distributor copied it and a viewer received it. The next unit is the source, plus a policy for building it for whoever is watching and wherever they are.
Creators who hold a clear source gain, because any platform can build from it. Platforms that hold the viewer gain, because they run the loop. Consumer companies with their own audience data gain, because they know what each person needs to see. The ones who lose sell finished files by the hour of crew time.
Bakeries stay open for the shared moments: a final, a premiere, the film everyone talks about the next day.
Everything else gets made, rendered, generated to order.
Directed once.
Rendered for each.
Bakeries stay open for the shared moments.
rendered for you. pick a language.
Directed once, rendered for each.