Per-10,000-view revenue falls as low as 5 yuan: why the AI short drama business loses more the more it's made

By Tech Talk

In October in Hangzhou's Binjiang district, the damp chill of late autumn crept through the windows of a mixed commercial-residential building by the street. In a room of just over ten square meters, the computer tower was still humming, the screen covered with concept art and storyboards of cultivation-fantasy characters, while the settlement page in the short-drama backend made the revenue-share figure of under 800 yuan look especially jarring.

In front of the screen, stubble-faced Lao Lin sat slumped in his gaming chair without moving for a long while. On the whiteboard behind him were still the words he had written himself three months ago, when he and his friends first went into AI animated comics together: "Either make it, or die." The rent for the next quarter on this apartment had only been paid yesterday. Thinking back to the passion he had felt three months ago when he moved in with his startup capital, and looking at today's revenue share of less than 800 yuan, he was deeply confused: where had all the money gone?

Over the past month or so, this small three-person team had worked round-the-clock shifts to rush the job, burning through nearly 100,000 yuan in compute costs, repeatedly running tens of thousands of image-generation gacha-style trial-and-error adjustments, and hand-crafting three cultivation power-fantasy dramas of dozens of episodes each—yet the revenue share they actually saw was not even enough to cover utilities, broadband, and daily expenses.

But outside the screen, a different picture was on display.

AI short-drama concept stocks on the A-share market hit the daily limit one after another, and Jimeng AI also hung a Double 11 banner on its official website, advertising a promotional price of "Seedance2.5 as low as 0.177 yuan per second." Outsiders assumed the big platforms were once again throwing money around to subsidize a hot trend, but Lao Lin, who had been through multiple credit shrinkages during the year, knew very well that behind the promotional tricks lay steadily climbing compute costs and the invisible thresholds of ecosystem lock-in.

This is precisely the cruel reality of the entire track right now. Data disclosed by Chinese Online shows that in its own video-generation business, as much as 55% of Token consumption went to short dramas and video generation. Yet DataEye's statistics expose another truth: in the first half of this year, as many as 98.7% of AI micro-short-drama projects were unable to recoup their costs.

Peel back the three layers of accounts—secretly rising production costs, collapsing revenue shares, and monopolized traffic-buying channels—and you find that internet giants led by ByteDance have long since locked down the entire industry chain, achieving a closed loop that rakes in profits whether times are good or bad. But at the same time, the industry's own speculative bubble and the dilemma of content homogenization are also rapidly draining user trust and commercial value.

It is worth noting that while the vast majority of practitioners are either losing money to gain attention or digging into their own principal to help the big platforms consume their compute quotas, there remain a mere 1.3% of survivors—so what do they rely on to stay alive?

Closed source can't afford the compute tax, open source can't fill the engineering pits

What truly suffocates small and medium-sized teams is the "Jevons paradox" effect in economics. Although technology raises per-unit production efficiency, platform rule upgrades and industry-wide involution together spur the entire industry to consume multiples of resources, pushing total costs ever higher.

At the start of this year, the studio could still use cheap credits to make a few PPT-style animated shorts with static images panned and zoomed, and still get decent organic traffic. By the time the whole industry piled onto the table, platform algorithms immediately raised the entry threshold, directly throttling 720P footage and tilting traffic weight comprehensively toward higher-frame-rate, fully animated 1080P premium footage. To get a traffic admission ticket, producers were forced to upgrade to high-consumption models and burn nearly half their credits on discarded takes through repeated gacha-style generation.

Industry estimates show that the pure compute cost of a standard episode of AI animated comics has risen from 17 yuan in the early days to 80 to 100 yuan, while the production cost of premium AI short dramas using the live-action generation route has climbed to 800 to 1,500 yuan per minute; making a hundred-minute-scale short drama now requires nearly 100,000 yuan in investment.

English text is being prepared.

Since closed-source models keep raising prices, can free open-source models offer a way out?

In the second half of this year, the LTX series and MiniMax H3 (Hailuo 3.0) did deal a considerable blow to the closed-source camp, with the latter at one point taking away about 20% of price-sensitive customers. But when small teams actually moved their workflows over, they discovered that the membership fees they saved were almost entirely lost in the engineering quagmire.

Open-source models are often half-finished projects. Although some models open up basic 768P weights, they lock the core capabilities of complex instruction orchestration and 2K super-resolution refinement behind official paid APIs. More realistic obstacles come from technical and hardware costs: teams must hire technical artists at high salaries to wrestle with ComfyUI's complex node orchestration, LoRA training, and model tuning. Local deployment requires expensive multi-GPU clusters, while renting cloud GPUs by the hour often incurs per-second charges amid debugging errors. After all this trouble, many small workshops find that the overall output efficiency per person of open-source solutions is far less cost-effective than closed-source models. Whichever way they turn, small and medium-sized teams cannot escape the besieged fortress of compute inflation.

430,000 power-fantasy dramas flood in, and revenue per ten thousand views can't buy a cup of Mixue lemonade

In September, Wang Xiaoliang, director of the Network Audiovisual Program Management Department of the National Radio and Television Administration, disclosed that in the first eight months of this year, a total of 430,000 micro-short dramas went online nationwide, 13 times the figure for all of last year, with AI-generated micro-short dramas accounting for more than 90%. The frenzied deluge on the supply side quickly triggered a collapse of the pricing and settlement system.

It is understood that the contract production quote for standardized AI comic dramas was still around 5,000 yuan per minute at the beginning of this year, but has now fallen to a few hundred yuan, hugging the cost line. As a large number of teams rushed in and undercut prices, companies surviving purely on production orders have already exited, about 90% of them. The vast majority of these are workshop-style teams with no content R&D capability that rely purely on template-based volume output, and their exit is itself an inevitable result of industry consolidation.

English text is being prepared.

The value of a single view has shrunk so drastically because of the "lemons market effect" caused by bad money driving out good, as massive amounts of low-quality content have directly shattered the commercial credibility of the entire category.

What was truly hugely profitable in short dramas was the paid recharge model (IAP), but audiences lack any impulse to pay for AI fake faces that frequently look alike and have stiff expressions, forcing more than 90% of AI dramas onto the single-plank bridge of IAA, where viewers watch for free and revenue comes from ad sharing.

According to industry data from the China Netcasting Services Association's "Micro Short Drama Creation Guidelines," in the first quarter of this year, although live-action micro short dramas accounted for less than 5% of production capacity, during the Spring Festival period they contributed about 25 times the views of AI dramas and took more than 90% of all traffic across the internet.

In the free viewing model, the level of revenue sharing depends on the effective cost per mille (eCPM) paid by advertisers. When hundreds of thousands of formulaic AI cultivation, doomsday, and domineering-CEO power-fantasy dramas flood in, it has become the norm for viewers to swipe away in two or three seconds, and ad conversion rates are disastrous. The flood of bad money has not only scared off advertisers but also invited regulatory clean-up.

Since the beginning of this year, 68,000 violating micro short dramas have been taken down across the internet and more than 1,200 violating accounts have been dealt with, of which more than 90% were AI dramas. With the "Measures for the Administration of the Development of Micro Short Dramas" issued in September mandating that AI dramas add labels, authorities have further guided platforms to open "live-action drama" sections and support them with real money. Facing user loss, advertiser price pressure, and a pullback in platform resources, the revenue-sharing pool for AI short dramas has completely run dry.

ByteDance and others monopolize the entire industry chain, and more than 90% of players are working at a loss and paying out of pocket

DataEye's monitoring of Douyin data shows that among the 221,900 AI short dramas newly launched in the first half of the year, only 1,055 surpassed 100 million views, a rate of just 0.47%. On average, only 1 in every 77 AI comic dramas can be profitable. And the number of active production contractors in the first quarter sharply fell from 1,216 to 698, with 42% eliminated in three months.

Since the books are already this bad, why did 430,000 dramas still rush in one after another in the first eight months? Because the super platforms had long ago designed a precise casino mechanism. They use the myth of a 0.47% hit rate + the wealth illusion of consecutive limit-up boards on A-shares + Double 11 bonus points as bait, just like the sound of coins spilling from a slot machine, precisely capturing entrepreneurs' gambler psychology of being unwilling to leave the table.

There is a painfully self-mocking saying in the industry: AI has pushed the production cost of a single episode down to 1,000 yuan, but to make viewers see it, you have to spend 5,000 yuan to buy traffic. According to Huaxi Securities' research report "AI Short Drama Industry Perspective" and DataEye monitoring data, this year the traffic placement cost for AI short dramas has soared by more than 100% year on year, and buying traffic and boosting distribution directly consumed about 70% of total full-chain costs, with producers ultimately often receiving only 4% to 7% of the revenue share.

At present, the advertising ROI (return on investment) of mature AI short drama teams has been compressed to the extreme range of 1.03 to 1.07, meaning that for every 1 yuan invested they earn at most 3 to 7 fen; ordinary projects have an ROI of less than 0.8, losing 20% as soon as they buy traffic. Even Chinese Online, Zhangyue Technology, and Kunlun Tech, which hold massive IP, after greatly ramping up AI short drama and comic drama businesses this year, have seen their financial reports show revenue surging while profits were devoured by high sales and traffic-buying expenses, falling into losses.

Ride-hailing platforms at least leave drivers some soup: they cannot make big money, but at least a month's work can solve basic food and clothing. But at the AI short drama gambling table, giants represented by ByteDance simply took the whole pot away. Although traditional live-action short dramas lost money, at least the money flowed to extras, lighting technicians, and boxed-meal vendors in Hengdian; in ByteDance's ecosystem, however, compute power and traffic buying account for the vast majority of the principal invested by entrepreneurs, and most of it ultimately flows into the big tech company's servers and advertising systems.

Upstream is Tomato Novel, the country's largest free web fiction platform. To get the hit templates among its 500,000 works that have been validated by readers, creators first have to pay ByteDance an IP licensing fee. In the midstream production segment, according to statistics from AI Pulis and Tech Planet, ByteDance's Seedance accounts for more than 80% of the domestic AI short drama tool market; LatePost disclosed that its annualized revenue has reached $2 billion (about 14.3 billion yuan), with a gross margin as high as 70%.

When Volcano Engine opened the full Seedance 2.0 API in April this year, it also required major clients to sign annual framework agreements with a minimum consumption guarantee of at least 10 million yuan. Many companies were forced to sign in order to lock in high-concurrency compute power and review convenience, taking on tens of millions of yuan in debt before a drama was even shot.

After the finished product is made, promotion and distribution cannot avoid Giant Engine, which takes 73% of the domestic short drama ad-buying market. At the end point, Hongguo Short Drama, often treated as a free drama-watching app, is a giant toll booth with 168 million daily active users and 125 minutes of daily use per person, turning all user time into ByteDance's ad inventory. Advertisers' money all flows into the giant's pocket, and only a few drops leak through its fingers to creators as a coolie fee of 5 yuan per 10,000 views. The giant monopolizes all five roles—IP merchant, compute arms dealer, ad-buying vendor, advertiser, and settlement party—so small and medium teams have effectively become algorithmic consumables that bring their own provisions, pay out of pocket, digest compute quotas for the big companies, and test content for free.

Giving up the fantasy of platform revenue sharing is the calculation of those 1.3% survivors

When a super giant welds every link from IP and tools to traffic into an iron barrel, application-layer entrepreneurship at the C-end revenue-sharing table has already become a false proposition. Apart from heavy-asset giants such as Yuewen and Chinese Online, which hold massive proprietary IP and can reuse digital assets across dozens of dramas through industrialized pipelines, ordinary small workshops that stay in the game and fight head-on are doomed to lose.

So how do the 1.3% of independent teams that actually make real money survive?

The answer is simple: they long ago overturned the C-end revenue-sharing gambling table and went to earn money outside the giants' closed loop. Looking back at the compute distribution data, right after the 55% for short drama tokens comes e-commerce and advertising marketing at 24%. Operators who are doing well have thoroughly connected these two compute blocks, turning AI short dramas into a low-cost traffic funnel at the front end.

English text is being prepared.

English text is being prepared.

Right now the Double 11 promotional banners are still flashing, and many newcomers are still calculating when the 0.177 yuan per second points package will pay for itself. Lao Lin will not be the last person to shut down the compute backend, and the next Lao Lin will not be the last person to move into that room. In this compute gambling game that no one wants to quit, recognizing as early as possible who the banker is and who the consumable is matters far more than chasing the next model version. Because here the banker cannot be changed and the rules cannot be altered; the only certainty is that the rice bowl next to the iron pot is only enough to feed that 1.3%.