Chinese media recently put two sets of figures side by side. On one side, output: 221,900 new AI short drama titles appeared on Douyin in the first half of 2026, more than 1,500 additions a day. On the other, returns: using 50 million views as the break-even line, only 1.3% of titles reached it, roughly one in 77 made its money back, and around nine out of ten AI short drama companies are still losing money. Both numbers describe the same situation: the scarce thing is no longer supply, it is attention.
Output climbs fastest, revenue falls fastest
The structural shift is more interesting than the headline figures. Live-action drama accounts for under 5% of new titles but takes more than 90% of total viewing; during the Spring Festival period it released one fiftieth as many titles as AI drama and still drew 25 times the views. The economics moved the other way as well: per-episode compute cost for AI short drama rose from 17 yuan to 80–100 yuan, while revenue per 10,000 views fell from 30–100 yuan in late 2025 to 5–10 yuan, a drop of close to 90%, and the number of active production vendors in the first quarter fell from 1,216 to 698. Making a title got cheaper; getting it seen got more expensive. Once that arithmetic stops working, so does "produce more, earn more."
One more shift is happening underneath. The platform subsidies these teams relied on are receding, and a model that lived on minimum guarantees and revenue sharing cannot hold. Even a hit is no guarantee: one AI-generated live-action-style adaptation of Journey to the West drew 300 million views in 48 hours and more than 800 million in total, briefly outperforming live-action drama of the same period, and was taken down across platforms less than a month later over compliance issues. Making a title, getting it seen and keeping it online are three different things.
We took on a project the platforms could not deliver
Those figures matched something we ran into on a real project this year. A client brief arrived with a portion of content that could not be generated directly under the general rules of closed platforms: original material the client did not want uploaded to a third-party service, and several shots that had to be controlled frame by frame to match the storyboard. Nothing in the brief was problematic in itself; the obstacle was that "only using what the platform offers" led nowhere.
So we worked on two tracks. The shots carrying the weight — image quality, camera movement, performance — went to the newest closed models for their consistency and mature cinematic language. The restricted and precision parts were completed in ComfyUI with open models, where a node-based workflow lets you see and adjust every step and keeps the weights and the generation process on hardware you control. Joined together, the project went through. There is nothing secret about it; it just takes someone willing to learn both sets of tools properly.
The most common answer in the tender: we can't deliver
The tendering stage was the revealing part. Plenty of teams asked about the brief, and a good number came back saying they could not do it, almost always for the same reason: platform restrictions. That answer says more than a quotation does. When a team's entire capability sits inside the boundaries of one closed platform, whatever the platform will not generate is simply beyond them.
The dependence is learned, not inherent. Closed models have improved so quickly — longer single generations, native audio, stronger multi-reference consistency — that a few lines of prompt can produce usable footage. Brute-force generation is genuinely cost-effective, so there is little reason left to study how nodes connect, how licences work, or how a shot is controlled. The price of that comfort is a fading craft: the question used to be "how do I achieve this shot," and is quietly becoming "will the platform allow it."
A hundred salespeople, eight people making films
That mindset eventually shapes how a company is built. Some have turned AIGC into a sales-driven business: over a hundred salespeople bringing in orders, eight people actually running generations and editing. While subsidies were generous and briefs were loose, the structure moved fast and the output numbers looked good. When subsidies recede, compute costs rise and platform review tightens, it cannot deliver demanding work. Most of the teams we met in that tender carried some version of this model.
An official from China's National Radio and Television Administration put it plainly at a press conference: AI is a tool, and people are the subject of creation. Progress in tools is welcome, but a tool has no agency. It cannot judge which option is worth pursuing, and it cannot be held responsible for the result. A prompt lottery can produce attractive footage; it cannot produce the judgement behind it. Use AI as a tool, keep your own technical judgement and your own craft, and what you make stops being something anyone can copy. One line from that reporting lands well: once making a title stops being scarce, judging what is worth making becomes the valuable part, and the creator's centre of gravity shifts from production to selection.
If you are holding a brief like the one above, something the platforms cannot deliver, and you need a team willing and able to connect open and closed models, start by looking at the work and the teams listed under AI short drama production.
