In October 2026, China’s AI micro-drama circle was still lamenting that nearly 99% of projects were losing money. Many teams saved on location shoots and actor salaries only to pay steep compute bills for “gacha-style” video generation, rolling the dice repeatedly like opening blind boxes just to pick out usable shots. Once a series went live, crews typically had to funnel roughly 70% of their gross revenue back into paid traffic—buying ad placements to pull in viewers. Yet by the first weekend of November, the firsthand financial ledgers of two grassroots teams went viral across the industry.
One of those teams came from a university in Soochow. Last autumn, Soochow University mechatronic engineering junior Xiao Tianle put together a five-person group that produced three AI motion-comic micro-dramas in seven days, pulling in more than 200,000 yuan in revenue from one breakout hit before going on to produce several thousand episodes as a contractor for major platforms. Starting this April, however, falling barriers to entry in AI squeezed outsourcing margins, and in a retrospective interview with Gusu Evening News he admitted that the early fast-food contract work was a dead end. Leading an online team of more than ten people, he pivoted to original self-publishing, recently scoring 1.5 billion views and over 2 million yuan in revenue from a single title, bringing his cumulative revenue to more than 3 million yuan in less than a year; even so, he was blunt that strict and shifting platform review standards, the risk of carefully crafted shows sinking without paid traffic boosts, and relentless compute burn all keep cash flow tight.
When even a campus team grossing millions of yuan a year is walking away from fast-turnaround contract work, how can grassroots creators who refuse to buy traffic break through? Another crew in Shijiazhuang, Hebei, offered an answer. Last spring, 54-year-old former wholesaler Zhang Hansheng spent a little over 1,000 yuan on an online course to learn AI video, met classmates including executive producer Li Chaohua, and gradually assembled an online team of roughly 30 people. Within its six-person core crew are chief director He Xiushan, a former ride-hailing driver, alongside stay-at-home moms juggling school runs, delivery couriers, and office workers—none of whom ever studied film or television. In an exclusive interview with China Business Journal, they revealed that their photorealistic historical time-travel micro-drama Armed Into the Song Dynasty (持械入宋) racked up over 4 billion cumulative views with zero paid traffic from start to finish.
Let’s wring out the hype first: both sets of revenue and viewership figures come from the interviewees’ own accounts and have not been independently audited. Rumors had swirled that the Shijiazhuang team was led by an unemployed financial analyst and had pocketed tens of millions of yuan in returns. Zhang Hansheng specifically debunked this, noting that there is no financial analyst in the group and that the ten-million-yuan figure was purely an outside estimate; the team has not received nearly that much money, saying only that they are much better off financially than before and have yet to tally up the final books. Meanwhile, ByteDance’s free micro-drama app Hongguo Short Drama recently explained its revenue-sharing rules to the media, and putting those mechanics side by side with the Shijiazhuang crew’s practices makes this autumn’s industry shake-up crystal clear.
The old micro-drama playbook was to dump around 70% of gross revenue into paid user acquisition and bait clicks with exaggerated conflict in the first three seconds. Season 1 of Armed Into the Song Dynasty launched on August 21 to fairly lukewarm first-week numbers, collecting only 600,000 to 700,000 bookmarks. Its creators stated flatly: “We don’t buy traffic, and we don’t know how to buy traffic.” Riding word-of-mouth from returning viewers, Season 4 shot to No. 1 on Hongguo’s AI New Drama Chart the day after its release, Season 9 broke out completely, and Season 10 pushed its heat index to 82.06 million with roughly 11 million bookmarks on Hongguo.
Why do other series lose money despite easily topping 100 million views, while a slow-burn long-form drama can snowball through organic recommendations? Hongguo Short Drama explained to China Business Journal that its payouts are not calculated by view counts, but are tied much more closely to viewers’ actual watch time. Like Douyin, Hongguo sits inside the ByteDance ecosystem, and according to the partnership guidelines of the Douyin Short Drama Creator Center, free micro-dramas settle payouts by multiplying viewers’ effective watch duration by a dynamic unit rate drawn from the platform’s net ad revenue and ad-free membership pool. Tapping into a video for three seconds and swiping away counts as a view, but earns almost nothing in watch-time payouts. Tencent’s WeChat Mini Program short dramas, whose weekly active users have neared 90 million, tie Tencent’s incentives for co-published new dramas to duration as well: as long as average watch time per viewer reaches at least 10 minutes, producers can receive up to a 100% revenue incentive.
In late October, platforms began a sweeping crackdown on watch-time farming and reskinned arbitrage. After reaching 430 million monthly active users and nearly 166 minutes of average daily time spent per user in September, Hongguo announced a phased shutdown of its “watch dramas to earn coins” feature, clearing out reward-farming bots and idle task grinders. Right on its heels, Hongguo removed 20,426 homogenized micro-dramas in a single sweep, targeting titles that stitched together forced conflicts, re-uploaded old shows under new names, or reused the same high-frequency AI face models across multiple series.
Douyin simultaneously rolled out new rules banning creators from using normal character dialogue in the first few episodes just to pass review before abruptly switching to voiceover narration over static visuals in the middle and later episodes. Once the platform detects low-quality motion comics with stiff, motionless characters cranked out in bulk by automated tools, it stops settling them at standard drama rates, downgrades them to the much lower-paying commentary tier, and throttles their recommendation exposure.
Once watch-time thresholds truly took hold, a much harder challenge emerged: how to use AI to shoot a long story that doesn’t break immersion. Photorealistic AI micro-dramas easily give themselves away through body movements, facial expressions, and lip-syncing, which is why photorealistic titles accounted for less than 8% of Hongguo’s top 500 AI micro-dramas in September 2026. Yet Armed Into the Song Dynasty deliberately chose a photorealistic historical war setting; even if a shot had no technical glitches, any take with flat emotion that scored a mere 60 out of 100 was scrapped without exception, and battle scenes featuring 20 to 30 characters often required 10 to 20 or more generation rolls.
Why didn’t such uncompromising standards trigger an explosion in compute costs? This 30-person online team had spent over a year finding its groove; at first, it split into five sub-teams working in parallel to churn out 20 micro-dramas a month. Li Chaohua, executive producer of Armed Into the Song Dynasty, explained that the team later deliberately cut its overall monthly slate down to single digits. Previously, rushing to meet deadlines meant skipping pre-production and jumping straight into video generation, which produced piles of unusable footage. During the half-month of pre-production for Season 1, the crew first read the original novel closely, locked in character designs, used ByteDance’s AI assistant Doubao to polish the script and dialogue, and then had chief director He Xiushan spend over a week crafting a 20-minute pilot for internal review. In an interview with Entertainment Value Officer, the creators noted: “Where we invest the most time and energy is actually adapting the script and generating the adapted dialogue, which our team sees as the single most effective way to save on video generation rolls.”
By locking down the emotional arc and action boundaries of every scene during the script and dialogue stage, the crew avoids constantly scrapping and regenerating video clips later. That mirrors the cost-saving logic of streaming giant Netflix using AI to run low-cost pre-visualization tests first: catching blocking and pacing flaws at the pre-production table is far cheaper than burning final-render compute on blind luck. As Zhang Hansheng and the creative team calculated in interviews with Shangyou News and other outlets, the entire 10-season, 1,000-episode series has been produced and released season by season since late August, costing roughly 100,000 yuan in compute per season and over 1 million yuan across all 10 seasons. Although total compute spend for this thousand-episode epic topped 1 million yuan, axing blind multi-team volume churning and putting the real work in before anyone hits the generate button slashed the per-episode discard rate—bringing the entire group’s monthly compute burn down instead.
A thousand-episode series also has to wrestle with the chronic AI headache of characters changing faces every time the camera cuts. When Season 1 began, the crew used ByteDance’s video generation tool Jimeng 2.0, later upgrading to Jimeng 2.5, which can reference more than a dozen images simultaneously when generating video. Before production started, the team prepared dozens of standardized reference images for characters, sets, and props and loaded them into a shared asset library for the whole crew, while “secret-recipe” prompts that industry peers usually guard jealously were made completely open-source inside the group. Even though every member phrased prompts differently, relying on a unified reference library and aligned character profiles kept the resulting visual style remarkably consistent.
The workflow these non-film-school creators pieced together is AI-native to the core—meaning the entire collaboration process is designed around how AI actually behaves. On a traditional film set, a director stands on location supervising how actors and cinematographers play off one another; here, stay-at-home moms and drivers scattered across the country never meet in person and are directing generative models that have zero human intuition. Their solution was to stop hoarding working prompts altogether, laying out character personalities, Song Dynasty costume details, and dozens of standardized reference sheets completely in the open for everyone to share—a practice that closely echoes the aggressive transparency software teams pursue when building AI-native organizations. When both human teammates and AI models work from the exact same unambiguous lore and visual assets, people can work independently without the show drifting off-model.
Avoiding visual glitches alone is not enough; to earn high watch-time payouts with zero paid traffic, a story has to keep viewers hooked for dozens of hours without wearing thin. Each episode of Armed Into the Song Dynasty runs under two minutes, making a single season of roughly 2.5 hours comparable to a feature film and bringing all 10 seasons to about 25 hours in total. Chief planner Zhang Hansheng dug through the free IP library in the Short Drama Creator Center and picked an obscure Fanqie Novel web serial written by “Chagen’er,” a working English teacher in Harbin, which had fewer than 100,000 clicks at the time. Most time-travel power-fantasy dramas hand their protagonist a built-in cheat code—like a portable upgrade mall that conjures up whatever supplies they lack out of thin air; this series stripped out every cheat code, so every bullet the protagonists bring to the Northern Song Dynasty is one less bullet left, every drop of gasoline burned is gone for good, and they still have to navigate Northern Song household registries, travel permits, and official dress etiquette the hard way. That grounded detail sparked bullet-chat debates over Song history and inspired viewers to build a fact-checking site that maps official ranks and battlefields across three layers—historical fact, fiction, and speculation—while state media outlet Banyuetan published a dedicated piece praising the micro-drama for cutting costs without cutting standards.
Affording that kind of slow, meticulous craft also depends on a collaboration model free of fixed payroll overhead. The 30-person Shijiazhuang collective runs on a one-person-company (OPC) model, where each member operates like an independent contractor teaming up on projects, with some members also moving into a local short-drama industrial center to claim policy subsidies. Team members sign no hierarchical employment contracts and draw no fixed base salaries; 100% of direct production costs go toward compute, and everyone splits revenue after the fact based on their role and production hours—the very structure that let them weather half a month of upfront script polishing before shooting and the slow build of the opening seasons.
Looking back at this autumn’s industry shake-up, models capable of referencing more than a dozen images at once and delivering sharper visuals merely saved grassroots creators the cost of soundstages and background extras. What enabled the six-person core crew behind Armed Into the Song Dynasty—none of them trained in film—to break out with a thousand-episode historical war drama was stumbling into a workflow built around the grain of AI itself.
By phasing out coin subsidies and pulling more than 20,000 homogenized titles offline, platforms have given real teeth to payouts based on effective watch time, bringing an end to the quick-buck era of brute-force generation rolls and reskinned paid-traffic arbitrage. Under the new rules, what truly saves grassroots teams from steep ad-buying bills and wasted compute is, on one side, the grounded research of verifying Northern Song household registries and travel permits and polishing dialogue line by line before production begins—and, on the other, the aggressive transparency of laying out reference libraries and prompts in the open for the entire crew to share.