新闻

Neill Blomkamp released a thirteen-minute film in July. He made it with ByteDance's Seedance 2.0, working alongside concept artists and using the faces and voices of 32 real people, and he posted it himself as a test run before attempting a feature. Gizmodo called it "an absolute disgrace on every level", noted that the acting had not improved in a year, and settled on antiseptic. Kotaku's reviewer watched it so that nobody else would have to.

In the first three months of this year, Chinese studios released roughly 128,000 microdramas. About 122,000 of them, more than 95 per cent, were made with AI, according to figures the China Network Audiovisual Association published on 30 April in the first quarterly edition of its micro-drama creation guidelines. Almost nobody outside the trade press wrote a word about it. The audience paid.

Same technology, same year, opposite verdict. The comfortable explanation is that microdrama viewers will watch anything. The data does not support that, and the real answer is more useful to anyone deciding what to make next.

The audience is not being fooled

The most surprising number in this business arrived on 23 July, in the first industry report on the vertical series audience, published by Holywater Tech with the research firm Owl & Co and built on platform data plus a survey of 2,737 users in May.

Holywater runs two apps. My Drama is live action. My Muse is AI-generated. Conversion to paying subscriber on My Muse was 23.6 per cent. On My Drama it was 23.9 per cent. That is a gap of three tenths of a percentage point, which is to say no gap at all. On My Muse's flagship title, nine in ten paying subscribers were still watching after a hundred minutes of cumulative viewing, and 89 per cent of those surveyed said they wanted more AI stories.

Hold one thing in mind about all of that: Holywater publishes both apps, so this is a company reporting favourably on its own AI product. The figures are worth taking seriously anyway, because the incentive cuts against the finding. A vertical studio that wanted to justify its live-action slate had every reason to report a gap and did not.

Set the numbers against Pew's finding on 18 August that 52 per cent of American adults are now more concerned than excited about AI in daily life, rising to 55 per cent of the under-thirties, the first time a majority of that group has crossed the line. The same cohort reporting hostility to AI in a survey is converting to paid AI drama at the rate it converts to live action. Those facts are not in contradiction. They measure different things: what people think about a technology, and what they will pay for on a Tuesday night.

The money underneath is real and it is not speculative. Omdia puts the US microdrama market at $1.5 billion this year, heading for close to $2 billion next, as Reuters reported on 18 August. US monthly users went from 26 million in 2024 to 66 million in 2025. ReelShort's average daily mobile viewing time reached 38 minutes in the second quarter, on Omdia's analysis of Sensor Tower data, which puts it ahead of Netflix, Prime Video and Disney+ on that measure. On 6 August, Media Partners Asia projected ReelShort would finish the year on $1.05 billion in revenue and $40 million in net profit, reversing an estimated $12 million loss the year before. DramaBox sits second, on 21 per cent of the global market to ReelShort's 29.

One caution on that profit figure, because it is the sort of thing that gets misread. MPA does not credit it to content. It credits a fall in the cost of acquiring a customer, with marketing spend dropping from 55 per cent of revenue in 2025 towards a projected 44 per cent by 2028, helped by stronger franchises, telco deals and direct billing through ReelShort's own web store rather than Apple's and Google's. "What decides the next phase is distribution and unit economics, not necessarily content volume," MPA chief executive Vivek Couto said. Cheaper production is on that list. It is not at the top of it.

The whole thing comes down to shot length

Here is the arithmetic nobody puts in the trend pieces.

ByteDance released Seedance 2.5 on 31 July. It generates up to thirty seconds in a single pass, with audio, and can be extended twice from there. That is the top of the market. Google's Gemini Omni Flash makes ten-second videos.

A microdrama episode runs a minute or two. A series runs 50 to 75 of them on Reuters' reckoning, framed 9:16 for a phone, with a cliffhanger built into the end of nearly every episode because the next unlock is the business model. An episode is therefore two to six generations long at the current ceiling, and nothing in the format requires that shot four remember the geography of shot one, because by shot four the story has already cut to a slap, a reveal or a closing door.

A feature film averages 1,045 shots, on Stephen Follows' analysis of the Cinemetrics database, and every one of them has to agree with the others about a face, a coat, a light source and where the sofa is. Blomkamp attempted thirteen minutes of that, and his reviewers found the seams in the performances rather than in the rendering, which is worth noticing. The failure was not resolution.

The vertical frame does the rest of the work. A 9:16 crop is a face-delivery mechanism. It cannot hold a wide shot with much information in it, so vertical drama stopped trying years ago and settled into close-ups, two-handers, reaction, and rooms you never need to see the whole of. Those are the shots generative video handles best. The things it still handles worst, sustained wide geography, hands doing fine work, an unbroken take with a moving camera, are shots the format had already priced out for entirely different reasons.

This is the part I find genuinely interesting. Microdrama did not adapt itself to AI. Duanju was built cheap and in a hurry on a phone-shaped canvas, by producers who needed sixty episodes for less than the cost of one sitcom episode. Reuters puts that comparison at $100,000 to $300,000 for a 50 to 75 episode series, against $2 million to $4 million for a single half-hour of traditional sitcom. Cheap, it turned out, looks a great deal like AI-friendly. The format was pre-adapted, and nobody designed it that way.

Watch where the viewer is sitting

Seventy per cent of vertical series viewers watch in bed before falling asleep. Fifty-nine per cent watch on the couch. Heavy viewers are putting in 13.1 hours a week, up from 9.5 five months earlier.

That is not the posture of an audience auditing a render. It forgives a great deal, and what it forgives is specific rather than general. A viewer lying in the dark with a phone held at arm's length, half asleep, is not going to catch a wardrobe shift between episodes seven and nine. She will catch a scene that bores her.

The producers know what they are competing for. Susan Rovner, a veteran of NBCUniversal and Warner Bros. Television who co-founded the vertical studio aTwist, put it to Reuters plainly. "We're not looking to win Emmys, we're looking to win the People's Choice Awards."

Where it breaks

Then there is the number that complicates all of this, and it deserves more attention than it has had.

ShareChat's Mohalla Tech committed ₹100 crore, a little over $10 million, to AI-led microdrama in early August. Its own QuickTV data shows the latest AI-generated series completing at close to 50 per cent by episode 20, against around 60 per cent for live action. Ten percentage points.

Read that alongside Holywater's conversion parity and a fairly precise picture emerges. Nothing about an AI microdrama stops a viewer buying it. Something about it makes them stop watching around twenty episodes in, which on a sixty-episode series is a third of the way through the story and roughly the point at which a viewer has to start caring what happens to someone.

Whatever is failing there is not visual. Resolution does not degrade at episode 20. Character does. That is the harder problem, and it is not one more model release solves, because it was never a rendering problem in the first place.

What this means if you are making one

Storyverse produces microdrama alongside film and television, branded content and animation, so we have a commercial stake in this format continuing to do well. That is the disclosure. The rest is what the production side actually looks like.

The temptation, once you have seen the conversion data, is to treat generative video as a rendering department: write the thing, then have the machine make it. Most of the disappointing output in this format comes from exactly there, and it is a sequencing error rather than a tooling one. The fit between AI and vertical drama is decided in the script, not at the render. If a scene calls for a continuous forty-second take, a wide establishing shot carrying plot information, or two characters handling the same object across a cut, you have written shots the tools will fight you on, and no amount of regeneration fixes a structural decision made three weeks earlier.

Writing to the grammar is not a compromise. It is what every format-native writer has always done. Soap writers wrote for standing sets and multi-camera blocking. Sitcom writers wrote for a proscenium and an audience. Vertical drama has a grammar too, it is unusually well documented by now, and it overlaps almost exactly with what generative video can currently do. Studios treating that as a scriptwriting constraint are producing work that holds. The ones treating it as a post-production problem are producing the ten-point completion gap.

If you are weighing a vertical series and want to know what the format will and will not carry before you commit a script to it, send us the brief.

The number to watch

None of this means AI video has arrived. It means one format got there first, for reasons that have as much to do with how vertical drama was built in 2019 as with anything a model learned in 2026. The next format to fall will be the one whose native shot is already short, close and cheap, and there are not many of those left.

Blomkamp's audience noticed in thirteen minutes. The microdrama audience notices around episode 20. That is a considerable head start. It is also a good deal shorter than the boom currently sounds.