In Wuhan, a content company with a little over 200 staff runs a batch tool that turns out microdramas (短剧) the way a print shop turns out flyers. At peak it has pushed more than 100 titles through in a day. The generation queue once ran so hot that the servers nearly fell over.
The cost line is the part worth sitting with. A 120-minute AI microdrama that used to cost RMB 300,000–500,000 (≈US$41,700–69,400 / HK$326,000–543,000) now costs about RMB 5,000 (≈US$694 / HK$5,435) on that pipeline, according to Xuanjia Technology (炫佳科技), the company that built it.
Now hold that next to the supply figure. At a State Council Information Office briefing on September 17, 2026, Wang Xiaoliang of the National Radio and Television Administration said 430,000 microdramas had gone online in the first eight months of the year — thirteen times the whole of 2025 — and that more than 90% of them were AI titles.
What happens when the marginal cost of a title approaches zero
A drama is now cheaper than the ad spend used to promote it, cheaper than a single day of a location shoot, and in many cases cheaper than the thumbnail.
That inverts the economics of the category. Microdramas were built on a simple machine: spend aggressively on paid traffic, hook a viewer in the first fifteen seconds, monetise before they leave. Content was a cost centre tuned to a marketing budget. When a title costs US$700, the constraint moves entirely to distribution and to the platform's willingness to slot you into a recommendation feed.
It also changes what a studio is. If the asset is generated rather than shot, the durable thing is not the cast or the set — it is the reusable library: a character's face, a costume, a street corner, a lighting setup. That is exactly what the tool vendors started selling this autumn. On October 26, ByteDance's Jimeng (即梦) shipped a 短剧工作室 workspace that keeps scripts, character and scene assets, storyboards, generation and post in one project file, on the theory that consistency across episodes is the product.
The demand side did not grow thirteenfold
Here is the uncomfortable half. DataEye's data for the first half of 2026 found that 84% of live-action microdramas scored below 1 million on its heat index, and only about 1% cleared 10 million. Those numbers describe live-action titles, not AI ones, so they are not a measurement of this year's flood. But they establish the shape of the market that the flood is entering: a long tail that is already extremely long.
Attention is the one input that did not get cheaper: there are still 24 hours in a day, still a finite feed, and still a viewer who abandons a drama in eight seconds if the first shot looks wrong. Thirteenfold supply against single-digit demand growth does not produce thirteenfold revenue. It produces a sorting problem.
The early evidence is visible in how platforms are behaving. Incentive programmes now pay for quality thresholds rather than volume. Jimeng's own bundling — assets, storyboards, post in one place — is an argument about consistency, not throughput. Nobody is advertising "make more"; they are advertising "make the ones that hold together".
What zero cost does to the signal
Cheap production is not the same as cheap judgement, and the gap shows up in a specific place: the first thirty seconds.
A human crew shooting a microdrama has to decide, on set, what the opening shot is. That decision is expensive and therefore considered. A generated pipeline has no such forcing function. It can produce twelve openings, and the selection becomes the whole job — which means the quality of the output now depends almost entirely on who is choosing and how much time they have.
That is a skills problem before it is a technology problem, and it explains why the same tools produce wildly different results across studios. The one thing a model cannot yet do is want the viewer to keep watching.
Labels, exports and the part that still costs money
Two structural forces are pushing back on the flood.
The first is disclosure. China's rules on labelling AI-produced content (人工智能生成合成内容标识) have been in force since September 2025, requiring explicit and implicit markers on synthetic media. For a producer shipping a hundred titles a day, labelling is not a creative decision but a compliance line in the pipeline, and one that platforms are now checking. It does not slow production much. It does make the "we just won't mention it" option unavailable.
The second is export. Overseas short-drama apps have been the growth story for two years, and AI production plus automated dubbing lowers the entry price for a small team to test a foreign market. But the overseas market has its own sorting problem, its own ad costs, and its own audiences who are, if anything, less forgiving of a bad first shot.
Honest limitations
The 430,000 figure and the 90% AI share come from a single official briefing as reported; the administration does not publish a public dataset I could audit, and "AI title" is not defined in the quote. The 84% and 1% figures are DataEye's, cover live-action titles only, and use a proprietary heat index. The cost comparison is Xuanjia's own claim about a client's pipeline and its own tooling — I have not seen an audited cost sheet, and a US$700 drama almost certainly excludes the writer, the editor and the person choosing between twelve openings. I have not measured audience retention for AI titles specifically, which is the number that would settle the question.
What readers can do now
- If you are commissioning microdramas, ask what the per-title cost excludes. The generation bill is now trivial; the judgement bill is not.
- If you are building a production pipeline, invest in the asset library and the selection step. Those are what survive when a title costs less than its marketing.
- If you are watching this market from outside China, treat the volume numbers as a supply signal, not a demand signal. The interesting metric is retention, and almost nobody publishes it.
