This May and July, we published two articles observing Chinese AI micro-dramas, discussing who made money after AI production costs dropped to zero and whether newcomers can still enter the field, respectively. A few months have passed, and our team has kept a close eye on the sector. Looking back against our earlier assessments, several points need correction. The most striking contrast is in costs: while the API call price per second of footage has indeed plunged to a fraction of a yuan, the total expenditure to assemble a finished piece that can pass censorship and go live is actually rising.
The money saved on footage generation did not go to creators; instead, it flowed into the computing power spent on trial and error and the bills for buying traffic. The industry’s barrier to entry has shifted from whether one can generate footage to whether a title can pass review and be seen by audiences—two switches that remain firmly in the platforms’ hands. Cheapness is merely on the surface. Although platforms have loudly proclaimed support for live-action dramas, over 90% of newly released micro-dramas are still AI-generated. After a reshuffling across hundreds of thousands of series, the ones making real money remain the upstream tool makers and distribution giants.
Logically, when tools become cheaper, finished products should follow suit with lower prices. Yet the balance sheet of the micro-drama industry this year has done just the opposite. Look first at the API pricing of underlying models. Taking low-cost developer tiers as an example, data compiled in an industry report by 36Kr shows that the cost for mainstream models to generate one second of video has dropped to the order of a few tenths of a yuan.
Yet when it comes to producing a complete animated comic drama / manju, costs are substantially higher. According to surveys by DataEye and TMTPost, the production cost for finished manju footage has risen from around 100 yuan per minute to 200–300 yuan. Statistics from Yiou show that the pure compute cost for an episode of manju has jumped from 17 yuan to 80 to 100 yuan.
One reason for this inverted ledger is that viewers have grown much pickier. In the early days, people would watch even when facial features warped wildly; now, if the visual quality is rough, many swipe away immediately. To retain viewers, production teams have had to figure out how to keep characters’ faces and lighting consistent across shots. This leaves no choice but to multiply generation runs, turning what was originally a unit price cheaper by a few tenths of a yuan into a much higher total cost due to the sheer volume of generations.
Falling unit prices alongside rising total expenditures is not unique to micro-dramas. As we wrote earlier in To Cut AI Token Costs, Don’t Slash Model Unit Pricing First: token unit prices of models have fallen steadily, yet corporate AI bills keep climbing because people use much more when it gets cheaper. The micro-drama space sees the exact same dynamic: unit prices dropped by a few tenths of a yuan, but usage multiplied manifold.
Looking solely at model quotes makes it easy to misjudge costs; only by examining the production pipeline can one see how the money leaks away. When staff enter prompts to make characters perform actions, the footage frequently suffers from deformed limbs or jittery lighting. According to industry estimates, the usability rate of a single generation run is often under 20%. To piece together a set of acceptable shots, the same shot often has to be generated 5 to 7 times, relying on luck and repeated curation—a process commonly referred to in the industry as “gacha pulling” (chouka).
With back-and-forth attempts like this, computing expenses naturally multiply. As the same shot is generated over and over, total rendering expenditure surges. Although single generations are cheap, needing so many tries just to get one usable clip means that all the money saved on unit costs is spent right back on the extra attempts.
Even if a finished production is painfully pulled together, the biggest expenses still lie ahead. A micro-drama sitting on a hard drive is meaningless; someone has to watch it. Micro-dramas make money mainly through viewer payments for unlocks or ad revenue sharing. But relying solely on natural algorithmic recommendations rarely gets a new drama in front of massive audiences. Most project teams must pump money into ad systems to buy exposure and entice viewers to click and play—a practice commonly called traffic acquisition or user buying (touliu / mailiang).
According to Lin Qiwen, Vice President of DataEye’s Juchacha, and reporting by Jiemian News, traffic acquisition consumes roughly 70% of gross revenue in short-video platform monetization. Even for professional teams, the return on ad spend (ROAS) often hovers at mere break-even or razor-thin margins. If ad efficiency slips even slightly, returns drop below ad spend—the more you spend on acquisition, the more you lose.
Relying on platform view subsidies is also shrinking as a viable path. The industry uses revenue sharing per 10,000 views to measure how much cash every 10,000 valid views can bring in. According to backend dashboard bills shared by creators and reported by Jiemian News, the payout per 10,000 views plummeted from 30–100 yuan in 2025 to 5–10 yuan in 2026. One title garnered 180 million views, yet ended up with only around 180,000 yuan in revenue share. The few cents saved on footage costs are virtually negligible in the face of colossal traffic bills.
Recently, platforms and regulatory authorities have issued consecutive documents encouraging live-action micro-dramas, creating an illusion that AI micro-dramas are on the way out. From a budget standpoint, live-action dramas have indeed received substantial, hard-cash backing. According to overviews by media outlets including 21st Century Business Herald, Douyin’s guaranteed budget for live-action micro-dramas this year exceeded 1.5 billion yuan, with the average guaranteed amount per drama rising by roughly 60% year-on-year; spearheaded by the National Radio and Television Administration (NRTA), six major platforms collectively pledged at least 6 billion yuan to support live-action dramas and guide platforms to set up dedicated live-action drama sections.
Judging from actual production volume, however, the situation is completely different. According to data disclosed by the NRTA at a State Council Information Office press conference in September and reported by Xinhua, 430,000 micro-dramas were launched across the entire web in the first 8 months of 2026, with AI-generated dramas accounting for more than 90%.
Regulation is tightening in tandem. NRTA Order No. 16 came into effect on September 1 this year, mandating that AI dramas display prominent labels on every episode. According to an NRTA notice on September 17, across the entire web 68,000 non-compliant micro-dramas have been taken down so far this year, over 90% of which were AI dramas. To go live, you must first pass review.
Behind these omnipresent feeds lies a massive user base. According to QuestMobile data, as of July 2026, daily active users of the standalone Hongguo app reached around 168 million, up 107% year-on-year, surpassing the combined daily active users of the four major long-form video platforms (around 155 million people). This metric excludes web and TV interfaces and counts activity driven by incentives such as check-ins and tasks. In addition, CBNData reported that AI dramas once occupied 77 of the top 99 spots on Hongguo’s trending chart.
This roughly reflects the platforms’ dual ledgers. Supporting live-action dramas serves more as a showcase to set benchmarks, meet compliance requirements, and win mainstream approval; the massively deployed AI dramas, by contrast, are fast-moving consumer goods that fill fragmented hours and absorb advertising budgets. Live-action drama production has not bounced back alongside the budgets; the vast majority of new releases remain AI micro-dramas.
After the industry produced hundreds of thousands of series with great fanfare and racked up hundreds of billions of views, who actually made money? The answer lies in the ledgers of upstream infrastructure. According to reports by TMTPost and others, primary tool Seedance generated monthly revenue exceeding 1 billion yuan, with a gross margin of around 70%; on the distribution side, Ocean Engine captured roughly 70% of the micro-drama traffic acquisition market.
Meanwhile, publishers and production teams caught in the middle face an extraordinarily difficult existence. Industry leader Dianzhong Tech generated around 14.5 billion yuan in revenue in 2025, but its founder previously stated publicly that the net profit margin of its micro-drama business was under 1%. Production teams are even more polarized: a handful of top-tier teams secure handsome profits through viral hits, but roughly 90% of companies across the industry operate at a loss. Statistics from DataEye are even more blunt: over 220,000 AI dramas were newly released on Douyin in the first half of the year; taking 50 million views as the break-even threshold, only 1.3% crossed that line (DataEye industry report).
Unable to sustain the pressure, production teams began exiting en masse. By Q1 2026, the number of active production houses dropped from 1,216 to 698, a decline of more than 40%.
This predicament is not unique to China. According to reports by 21st Century Business Herald and others, OpenAI’s video generation tool Sora generated approximately $2.1 million in in-app revenue in the six months after launch, against estimated daily costs of around $15 million and a 30-day retention rate of less than 1%, before eventually shutting down. While these figures remain unconfirmed by OpenAI, they illustrate a clear reality: having generation tools alone, without distribution and monetization, can hardly sustain the bills.
The unit price of generating footage has indeed fallen. But the cost of turning micro-dramas into something people actually watch and that can pass censorship has not dropped with it. What creators have at their disposal are mostly the generation tools; what truly determines whether a drama can go live and who takes the profit remains in the hands of the few players controlling regulatory compliance, algorithmic distribution, and monetization rules.