Li Wenhao is a micro-drama actor in Chongqing. When he entered the industry in 2023, he could shoot for 50 consecutive days without stopping. In March of this year, he worked for 6 days.
This is not an isolated case. After the Spring Festival last year, the entire industry pivoted to AIGC. Writers who don’t use AI to write scripts no longer have anyone to collaborate with. Production costs dropped from 1.5 million yuan per show to under 300,000 yuan. Overseas production fees dropped from 3,000 yuan two months ago to 300 yuan. About 470 AI micro-dramas are released online every day, and the proportion of AI-generated content in the top 100 rankings rose from 7% a year ago to 38%.
This is the largest and most advanced real-world case of AI-native content production globally. It answers not “whether AI can make videos,” but “how the profits of the industry chain are distributed when AI compresses production costs to near zero.”
The mainstream narrative says AI has democratized content creation. In the first quarter of this year, about 128,000 micro-dramas were released online across the industry, of which AI productions accounted for 122,000, exceeding 95%, according to data from CCTV Finance and the China Netcasting Services Association. A senior dance major and a friend can set up a studio, and one person can produce 40 minutes of distributable content a day with a 45% profit margin. The threshold has disappeared, and production capacity has exploded. It sounds like the creators have won.
But the data tells another story.
To understand what AI has done to this industry, one must first understand what micro-dramas were without AI.
The business model of live-action micro-dramas has always been “three parts production, seven parts traffic acquisition.” A drama’s production cost might be 1 million to 1.5 million RMB, but 80% to 90% of the revenue must be spent on buying traffic (ad placement). The production side bears the production risk, while the platform collects advertising fees and revenue-sharing cuts. Creators do not possess pricing power. The algorithm decides who sees your content, and the algorithm also decides how much you pay for these exposures.
What has AI changed? It slashed production costs by 80%, compressed the schedule from three months to one month, and shrank a team of over a dozen people down to one person. Where did the saved 1.2 million go? It did not become the creators’ profit. In March 2026, the daily ad spend for AI micro-dramas on Douyin surpassed 70 million yuan, exceeding the ad spend for live-action micro-dramas for the first time. The saved production fees flowed directly into the platform’s advertising system. The more you save, the more aggressively you spend on ads.
There is a deeper signal here. According to first-hand information from practitioners, ByteDance’s Hongguo platform reduced the revenue-sharing ratio for AIGC micro-dramas to about one-twentieth of what it was. Not a 20% reduction, but reduced to 1/20. Why does the platform dare to do this? Because creators have no alternative channels. Your content is generated using ByteDance’s Seedance, advertised on ByteDance’s Douyin, and distributed on ByteDance’s Hongguo. The entire closed loop is in the hands of one company, and the revenue-sharing ratio is unilaterally determined by them.
At this point, most people would reach a conclusion: the platform is too greedy and ate all the profits. This judgment is correct, but insufficient. It does not answer a more fundamental question: why did the platform win?
On March 24 this year, OpenAI announced it would shut down Sora. This model, which shocked the entire industry with an AI-generated mammoth video in February 2024, survived in the consumer market for less than a year and a half. Sensor Tower data shows that Sora’s global in-app revenue since its launch was $1.4 million. During the same period, ChatGPT achieved $1.9 billion. A Forrester analyst called it a “resource black hole with limited monetization paths.” On the same day, OpenAI also canceled its $1 billion content partnership with Disney.
At the same time, ByteDance’s Seedance is being used as the default production tool by the micro-drama industry. Kuaishou’s Kling AI had a Q1 revenue of 650 million yuan, a year-on-year increase of over 300%, with an annualized revenue running up to $500 million. “The best video generation model” and “the most successful video generation product” happen to be two completely different things at this point in time.
The reason is not model capability. Sora’s technical path, using a diffusion Transformer to generate on spatiotemporal patches, is an academic breakthrough. The problem lies in the business structure. Video generation is an extremely cash-burning product: each generation consumes massive GPU computing power, users need to iterate repeatedly to get usable footage, and the monetization path relies on a professional creator ecosystem. OpenAI possesses none of these three conditions.
It has no creator ecosystem. Sora is an isolated web tool; after users generate a video, the next step is to figure out how to distribute it themselves. There is no algorithmic recommendation to let it be seen, no advertising system to monetize it, and no ready-made audience for it to find a market.
It has no user behavior data flywheel. Training video generation models requires massive amounts of video data. Not just public videos scraped randomly, but real consumption data with completion rates, likes, comments, and share tags. This data only exists in one type of product: short video platforms. ByteDance and Kuaishou have hundreds of millions of users swiping through videos every day; every swipe, every pause, every bounce is a signal for model training.
It has no sustainable cash flow. OpenAI’s main business is ChatGPT subscriptions and enterprise APIs; reasoning and programming tools are its most profitable product lines. Sora’s computing cost is far higher than text generation, but its paying user base is far smaller than ChatGPT’s. Cutting a cash-burning product aligns with any company’s financial logic.
Look at Google conversely. In September this year, YouTube announced at the Made on YouTube event that it would embed DeepMind’s Veo 3 video generation model directly into the creation workflow of YouTube Shorts. Users can click the create button for Shorts and use Veo for free to generate video clips with sound. Once generated, it can be published with one click. The audience can see it immediately. The ad revenue-sharing system can settle it immediately.
The distance between the tool and distribution is zero. This is the real reason why Seedance and Veo could succeed, and Sora could not. It’s not that ByteDance’s AI happens to be powerful so creators are trapped; it’s that ByteDance had the channel first, which gave it the conditions to make its AI the most powerful. The sequence is channel, category, tool, not the other way around.
Applying this deduction to the global market, the pattern becomes even clearer.
Kuaishou had a short video platform first, with over 400 million MAUs in China, and then came Kling AI. Of Kling AI’s 650 million revenue in Q1, the fastest-growing segments were “enterprise API services” and “professional user subscriptions,” with the top three application scenarios being advertising marketing, film and television production, and gaming. Without Kuaishou’s advertising clients and creator ecosystem, Kling AI’s product iteration and commercial conversion could not have been this fast.
Google had YouTube first, the world’s largest video platform with over 2.5 billion monthly active users, and then came Veo 3. Veo’s 4K output and native audio generation are indeed technologically leading, but what truly makes it commercially viable is that embed button inside YouTube Studio.
Companies with only models and no channels have two paths for their ending. One is to shut down. Sora chose this path. The other is to pivot into B2B tools, working for companies with channels. America’s Runway and Pika, and China’s Minimax and Shengshu Vidu, all took this path. They don’t build consumer platforms; they act as API suppliers for professional film companies and advertisers. They didn’t die, but they don’t participate in making the rules.
There is an even more subtle variable: open source. If video generation models become fully open-source like Llama in the future, and the technical threshold drops to zero, what will happen? When everyone can use the same models to generate videos of the same quality, the competition will completely shift from “who generates better” to “who can get people to see it.” The value of channels will not only not decrease, but will rise even further.
What about going overseas? If the domestic market is locked down by ByteDance’s closed loop, can going overseas break this structure?
The overseas micro-drama market is indeed exploding. In the first eight months of 2025, overseas revenue reached $1.525 billion, a year-on-year increase of 195%. ReelShort’s 2024 revenue was about $400 million. AI has also driven the threshold for going overseas to the floor: traditional manual translation of 100 minutes of video takes 7 to 10 days, while AI completes 1,000 minutes in 12 hours. Manual dubbing costs 6,500 yuan for 100 episodes; AI drops it to a few hundred yuan. Overseas production fees dropped from 3,000 yuan to 300 yuan.
But going overseas did not change the landlord of the channels. What is ReelShort’s main customer acquisition channel? TikTok ads. What about DramaBox? Meta’s Instagram and Facebook ads. Your micro-dramas target middle-aged white female users in the US market. For this user persona, who holds the precise ad targeting data? Meta and Google.
The business model of micro-dramas is an ROI arbitrage model from start to finish. How much is the production cost, how much is the user acquisition cost, how much is the audience payment conversion, and profit equals payment minus cost. The user acquisition cost on top platforms usually accounts for 50% to 80% of total revenue. AI compresses production costs to near zero, so where does the saved money flow? It all turns into platform advertising fees.
In essence, going overseas is just changing locations to pay rent to the same group of channel giants. TikTok belongs to ByteDance, YouTube belongs to Google, and Instagram and Facebook belong to Meta. The settlement currency changed, but the platform’s pricing power over your content did not.
Returning to Li Wenhao’s story at the beginning of the article. He only worked for 6 days, not because he didn’t act well enough, but because in a platform-driven AI production closed loop, humans do not possess bargaining power.
The daily pay for top micro-drama actors in China once reached 50,000 to 60,000 yuan. After the Spring Festival this year, the number of scripts they could receive per month dropped from over 30 to seven or eight. AI doesn’t need schedule management, doesn’t need revenue-sharing negotiations, and doesn’t need portrait rights consultations. Some people directly scrape photos from social media to feed into the model as reference images, and the generated characters look almost identical to real people, except it doesn’t legally constitute “infringement.”
ByteDance not only controls distribution and tools but also controls the upstream IP sources. Tomato Novel is the largest free web novel platform in China, providing the core raw materials for micro-dramas: narrative templates whose click-through rates and willingness to pay have already been validated. Your script is adapted from a Tomato Novel IP, your video is generated using Seedance, your audience is bought on Douyin, and your revenue is shared by Hongguo. There is not a single process in the entire chain that does not belong to the same company.
Actor Hao Lei said something on a variety show that shook the industry: AI will replace 90% of actors. She wasn’t pitching a product; she was describing her own observation.
Returning to the initial question: after production costs approach zero, where did the profits go?
The first layer of the answer is: they went into the hands of those who control distribution power. This layer of the answer is correct, but it stops at describing the phenomenon.
The second layer of the answer is: it’s not that the platform won because its AI is strong, but that its AI is strong because the platform has channels. Companies with channels can make the best AI video products not because of their algorithmic genius, but because they have data, computing budgets, creator ecosystems, and zero-distance monetization paths. Models made by companies without channels can be stunning, but they cannot survive the test of business cycles.
The implications of this logic extend beyond micro-dramas. In any industry where AI causes production costs to plummet, if its profit structure was already tilted toward channels, cost reduction will only make the channels siphon money faster. Costs dropping to zero is not liberation; it is acceleration. It transfers the value of more segments to the nodes that control distribution power.
In the field of video generation, channels determine everything. This is both a summary of global AI video commercialization over the past year and a half, and a prediction of the competitive trends in this field for the next few years.