Original · Unique Research · 2026-08-20
Editor's note: The first-person report and its judgments belong to the original Chinese author. This English rendition retains the opening essay, five insight sections, closing checklist, and the full roundtable transcript. All named companies, products, and people are preserved. Game sales figures, funding details, and revenue claims are speaker claims, not independently verified findings.
AI Industry Observer
Making Users Earn on Your Platform: The Hardest Retention Play in AI
"Models will catch up generation by generation. Traffic dividends will recede wave after wave. Letting users earn money on your platform, grow assets, and make friends — that is the real ticket through the cycle."
What will be the most important factor in building AI products over the next 12 months?
When Unique Capital founder Wu Wei posed this question, he helpfully offered four options: retention, user acquisition spend, gross margin, unit economics.
The question got cut immediately. Liu Meiling of Haiyi Interactive (海艺互娱) said: 12 months is too far out, absolutely too far; seeing 3 months clearly is already good.
"This is probably the most real rhythm of the AI application industry right now: last year people were still talking three-year plans; this year even annual plans have collapsed into quarterly rolling forecasts."
Models iterate monthly, new products sprout one after another; users use your tool today and switch to someone else's tomorrow. iReader (掌阅科技) overseas business head Xiao Wei's observation is direct: C-end users come in for novelty, and once the novelty wears off, they leave.
So what actually keeps people?
At an AI industry summit roundtable hosted by Yangfan Going-Global, three guests from three different tracks. Xiao Wei's Paoman (泡漫) is iReader's AI comic-drama tool, backed by the web-novel IP library. Li Zhe co-founded Elser.AI (心影次元), moving from comic-drama tools to interactive movie-game tools — the latter being the category of 完蛋！我被美女包围了 (I Was Surrounded by Beautiful Women) and 盛世天下 (盛世天下), which sold over 3 million copies. Liu Meiling's Haiyi Interactive runs an AI interactive entertainment platform; its newly spun-out digital human companion product has grown rapidly in 4 months, though she declined to give specific numbers.
Different tracks, but the three answers quickly converged on the same thing: first have a money-making scenario that keeps people, then talk about everything else.
And listening closely, their understanding of "retention" is completely different from most people's intuition.
Don't Blame User Churn on the Model
Li Zhe first dismantled an industry collective hallucination.
Many teams think user churn happens because the model integration is slow: someone else shipped a new model first, and users rush away. This phenomenon exists, but Li Zhe says it isn't the root problem. Video models have developed to today where one dominant player has emerged in China; the gap between models will shrink until it flattens out. On that day, the model won't constitute a reason for users to choose.
What's the real problem? Li Zhe says it plainly, almost "badly":
"To keep users, we have to do a bad thing: make it hard for them to leave."
The method is letting users' assets accumulate on the platform. Prompts, materials, works, creative relationships — the longer they accumulate, the higher the migration cost. Of course, locking people in alone doesn't work; the prerequisite is that your product truly meets the vertical needs of the largest paying demographic. Good, cheap, smooth — users naturally won't bother leaving.
After locking people in, the next step up is community. Elser is building a creator community plus commercial order matching: the platform itself operates an overseas short-drama distribution APP, subcontracting orders to creators on the platform. Users on your platform aren't just consuming tools; they can also take orders and earn money.
In Li Zhe's words: you help them solve problems, they naturally help you solve problems.
A detail worth noting: their platform's creators are mainly PUGC — small teams of 2-3 people, between pure beginners and professional studios. This also explains why commercial order matching works: small teams with stable delivery capacity are the ones who can handle orders.
Let Users Earn, and They Won't Bother Leaving
If Li Zhe's approach is "deposit," Xiao Wei's approach is "carry them along."
Paoman v1 hit a classic pitfall: users come in, do one job, and leave. Xiao Wei's post-mortem conclusion: only doing one link in the ecosystem can't keep people; you must string the entire chain from IP to monetization.
This chain is long. Upstream: bring all of iReader's domestic web-novel IPs, scripts, and publications into Paoman, so users have material to create with. Midstream: users use AI to adapt scripts and generate simulated dramas. Downstream: bring in production partners, and finished dramas go through quality inspection and rating directly on Paoman. Xiao Wei calls this "stamp-down": even if the drama you make isn't good, the platform walks you forward.
But the most critical link in the entire chain is money.
iReader gives production partners floor guarantees; newcomers also have insurance money. How is cost calculated? Production partners first use test credits to evaluate a drama's cost — say it comes out to 400-500; iReader marks it up 20-30% to take it. That 20-30% is the production partner's gross margin, covering labor and compute costs. After the drama is taken back, there's also revenue share.
"'Right now it's all floor guarantees.' Xiao Wei says this lightly, but people in the know understand the weight: this is the platform using real money to underwrite creators' sunk costs."
The effect is direct: production partners and C-end users know they can earn on your platform, retention is very high. Xiao Wei's logic is endearingly simple: creation is interest; making money is productivity.
There's also a business calculation behind this playbook: live-action short dramas cost $100K+ to film overseas; with the same money, how many AI dramas can be made? Xiao Wei's judgment is that over the next 18 months, iReader overseas will dig in hard on AI short dramas.
Tools Are Only First; Ecosystems Have a Future
Liu Meiling's approach sounds the most "zen": tend your own plot of land well.
This has a context. Many AI teams are jealous of short-video dividends and want to chase everything. Liu Meiling's attitude is clear-headed: chasing won't win anyway; better to do your own thing deeply.
Haiyi's plot of land is called "Character World." Users create digital humans and generate characters on the platform; these characters can go to MoreShot to make short dramas, enter the digital human companion product, and join platform contests to win awards. Users are both consumers and creators; shallow play for shallow users, deep play for deep users.
She dropped a sharp line:
"'Carp leaps over the dragon gate — as a tool, you're only remembered as #1; everyone forgets the rest.'"
Translated: in the tool format, users only remember the first place; everyone else gets forgotten. So tools must grow into ecosystems; this is the baseline, there's no second option.
IP is the traditional content industry's highest barrier; Haiyi bypassed a head-on fight: no existing IP? Let users create IP themselves. To build the copyright foundation solidly, Haiyi's latest funding round went directly to Visual China (视觉中国). The logic is clear: characters grow on your platform, roots are on your platform.
Earn $100, Throw $90 to the Model? Not On
After retention, Wu Wei turned to the more piercing topic: gross margin.
AI applications grow on top of models; users double, inference cost doubles, plus user acquisition spend. Earn $100 and throw $90 to model vendors — the platform is basically an advanced API router; the more users grow, the more it looks like false prosperity.
Li Zhe's solution is engineering-oriented: make users spend fewer credits and make better content.
The specific approach is interesting. The system accumulates every user's prompt adjustment process: how many times they revised for one image or video, what prompt structures work best in what scenarios, then summarizes into reusable templates. Internally, Elser calls these skills, in two layers: general skills for real-person interaction and different genre scenarios; personal skills learning each user's own usage habits. Prompts users write are first optimized by the system before being fed to the model.
The result: users unconsciously get smoother, use fewer, and save more. This also explains why AI products give out free credits so stingily — credits are real money; giving too much is a real loss.
Liu Meiling's solution is more "bold": bargaining power.
"'As the company with the largest token consumption across the entire AI district, we naturally have our own methods and solutions.' She didn't elaborate on the specific method, but the logic is there: the larger your consumption, the stronger your negotiating leverage with model vendors, the more you can sit at the table."
Beyond that, Haiyi has an entire internal toolchain and skill system; new products must first pass internal use: if even internally it can't connect, don't take it outside. Products your own people won't buy are invalid products. This standard sounds crude but makes sense on reflection: if even the people who know you best can't be bothered to use it, why expect outside users to pay?
After Models Catch Up, What's Left on the Table?
The roundtable's closing question, set by Wu Wei, was "long-term operations": suppose model capability flattens out in the future, user acquisition gets more expensive, traffic dividends are exhausted — what assets truly remain on the platform?
Liu Meiling's answer was three words: creativity, aesthetics, digital assets.
"AI took stock analysts' jobs; it's competitive for many complex tasks, but aesthetics can't be taken away, entertainment can't be taken away." This is also where the name "Haiyi Interactive Entertainment" comes from. All products grow on the aircraft carrier called Character World: users create IP, use IP to make short dramas, use IP for companionship, loop back to creation.
Wu Wei added an observation live: the most popular small shops in malls now sell IP merchandise. A plastic piece sells for 100 RMB; people who don't get it think it's an IQ tax; people who do know it's the "谷子 economy" (acrylic charm/merch economy).
Liu Meiling pushed this observation one step further: each generation has its own IP. Her generation did games using Gu Long and Jin Yong IPs; just the approval process could take forever. Now teenagers grow up at home or in their residential compound; they're actually the group most prone to fantasy, each with their own inner world.
"'A space recognized by yourself is called your little world; recognized by others it's an IP; recognized by more people it's a bigger world.'"
Li Zhe's view overlaps heavily. He listed three words for barriers: content copyright assets, user data, ecosystem community relationships. He also gave an unexpected example: tabletop RPG (TRPG). That old-school board game where a group sits together and co-creates stories is still wildly popular today. When Elser makes interactive movie games, they brought this mechanism into the tool: multiple people do TRPG simulations together, produce stories, polish them, then turn them into movie games.
Multi-person creation context all deposits in the system, migration is hard — retention and barriers solved together.
One more judgment worth noting: who are AI products built for in the future? Li Zhe says it's very likely the teenagers of today. They're still in middle school, but this generation has the wildest imagination and the most complete inner worlds.
Three Words, a Self-Check Checklist
Before dispersing, Wu Wei asked everyone to leave one metric: what proves your AI product has crossed the cycle?
Xiao Wei says positioning. Many people do everything — production and distribution both — they may make money, but never succeed. Find your ecological niche, go deep.
Li Zhe says retention. Especially at the moment the underlying model makes a step-change, the users who still stay are true retention.
Liu Meiling says monetization. Her original words are direct: if product retention is bad, positioning is bad, monetization is bad, don't do it; don't waste time, go quickly to a company with good monetization and study how they play.
Three words strung together form a complete self-check checklist: think clearly about your position in the ecosystem, use retention to verify if users truly need you, then use monetization to verify if it's worth continuing.
Models will catch up generation by generation. Traffic dividends will recede wave after wave. Letting users earn money on your platform, grow assets, and make friends — that is the real ticket through the cycle.
More Conversation Details
Wu Wei | Roundtable Host, Unique Capital Founder
Xiao Wei | iReader Overseas Business Head
Li Zhe | Elser.AI Co-founder
Liu Meiling | Haiyi Interactive PR Director
Most Important Factor in the Next 12 Months
Wu Wei: Please quickly introduce yourselves, then answer one question: today we're discussing how AI applications cross the cycle. In this process, what do you think is the most important factor in building AI products over the next 12 months? Retention? User acquisition? Gross margin? Unit economics? Self-introduce first, then answer.
Xiao Wei: I'm iReader's overseas business head, also head of Paoman. The host's question is professional. From my perspective, over the next 12 months, AI tool products will keep spawning new ones — including current ones like LibTV, Juli Lu, including our own Paoman; you'll find it keeps growing rapidly.
We see retention constantly leaking, from both ends: C-end users and B-end users — that's our experience summary. Because products keep getting more new ones, churn exists. Previously comparing C-end users, when they come to try this product or use it to make something, it's mainly novelty or value creation — this is very important. In the future, more and more new products will come; can the product create value? Repetition is relatively high; too many products, not enough demand. Maybe today use Paoman, tomorrow use someone else's product. So in this situation we need segmented analysis of value created for C-end. People who make games know there are onboarding features, so I think AI tools need even more guidance.
What's most important in the next 12 months? From iReader's perspective, we want to go all the way; everyone strives for #1, we strive too — this is our future development plan.
Li Zhe: Hello everyone, I'm Li Zhe, co-founder of Elser.AI. Our company has been founded two years. Last year we did the comic-drama track, making comic-drama tool products with overseas versions. Now we mainly focus on games; I'm responsible for interactive movie games. Previously I did more comic-drama work; comic-drama and interactive movie games are relatively connected. Now interactive movie games make more money, quite a lot.
Recent examples: 盛世天下 when it just launched — people who play interactive movie games know it's palace intrigue; it's sold over 3 million copies since launch, also on App Store. And back in 2023, 完蛋！我被美女包围了 went viral; even people who don't play interactive movie games know it. So actually AI short drama plus game interactive movie games — one word is Native, one is AI-able. AI-able I understand as improving efficiency and reducing cost of existing things; Native is producing new mechanisms that only exist because of AI.
What's most important for building an AI product in the next 12 months? My suggestion: the AI foundation may still change dramatically over the next 12 months; we can't lock onto one direction. Simply put, you still need to be in your own field. First, you must have a profitable scenario in hand, then continuously generate cash flow, and at different times keep thinking about what deep changes AI brings to product aspects, experience aspects, and game mechanics aspects, then build your own original, Native-leaning product. This way, whether for company lifecycle or your own future ceiling, it's a two-sided consideration.
Liu Meiling: Hello everyone, I'm Liu Meiling from Haiyi Interactive. I did about 13 years of game overseas publishing, global publishing; now doing AI for two years, and this product has been with us for over two years. In one sentence, Haiyi is an interactive entertainment AI platform. We've actually had three pivots: from platform to community, from community to ecosystem.
Answering this question, after hearing Mr. Xiao and Mr. Li, there's connection. I agree that 12 months is too far, absolutely; 3 months is OK. For our company, we build products around an IP matrix: Character World, Character World, Character World. Haiyi is an empowering product. Before we separately took out digital human for companionship, we already had digital human companionship in the product, in Haiyi, quietly polished for two years. The most important reason to spin it out as an independent product? Back to your question — which two points do you see? First, retention, absolutely retention. Because AI — whether matrix system or application — if you can't see that you can earn back the human cost, that's very realistic. But if you want a future, what's the expectation? Either LTV, or DAU, or retention, or monetization. Just these few points — if it hits one or two, do the product. Is it big? Otherwise, goodbye.
Retention: What's the Bottleneck to Keep Users Coming Back?
Wu Wei: Let's talk about retention next. But I want to understand what we can do to keep users coming back. Now AI tools don't just generate AIGC and call it done; we want users to keep coming back, even stay a long time, whether themselves or agents keep working. Not specific features — what do we do to make users come back the second week, the second month? What's the bottleneck affecting their return? Is the model not strong enough, prone to hallucination, or domestic vs. other?
Xiao Wei: Let's talk some real stuff, finally getting started.
When iReader was building Paoman, the first month of effort found one thing: all C-end and B-end users who came in would do one job, or one valuable job, then leave. Later we found you can't just do one link in the ecosystem; you need to do an ecosystem closed loop.
So we brought iReader's entire domestic IP line in, letting more domestic IP lines and script lines and publication lines all flow into Paoman. First, give him the ability to create content — material, series-like — after he has content creation ability, he tries adapting scripts. For a script to generate value, two things are needed: one is adapting into short drama, two is selling directly — this gives him monetization ability. Next, commercial production. Besides scripts ramping up, the downstream entire production partner side is also ramping up, including finished dramas. On finished dramas, we "stamp down" on the AI simulated dramas produced by Paoman.
Stamp-down means the drama you made might not be good, but I'll walk you forward. Stamp-down means we do quality inspection and rating. We think when he tries using this kind of AI tool, he has the intention to go overseas or domestic.
Wu Wei: You give them floor guarantees?
Xiao Wei: Right now it's all floor guarantees. Of course some newcomers come in for revenue share, and most will get some insurance money. Besides, we've produced a complete set of production partner quality inspection standards; all domestic and overseas partners are invited to test demos, everyone gets an entry opportunity, everyone gets a ticket.
The entire ecosystem closed loop is complete. From iReader IP source, to Paoman, finally to production partners, doing the whole ecosystem loop. Including overseas transcoding, micro-charts, social media, and TikTok, TikTok mini-programs and other overseas distribution channels are all in place. From root to middle transition, finally to script generation, then to traffic distribution — essentially walking him along one line forward. Ultimately, besides creation, besides interest-based creation, he can also earn on your platform.
Wu Wei: Mainly you feel it's their retention?
Xiao Wei: That's it. So far, based on Paoman v1 and its iterations, whether production partners or C-end users, knowing they can earn here, generate value, retention is very high. After all, this is the productivity-generated benefit.
Wu Wei: I also took a look at our platform. Indeed at first glance it's an image-anime generation tool. Users might just come to play at first, but now from being able to generate dramas, to being able to make interactive movie games — in this process, what do we do to keep users coming? This is also a retention topic; talk about it in connection with specific business directions.
Li Zhe: You have to think about one question: why did users leave? We were thinking about why users leave. There's an obvious phenomenon: integrate the model slowly, and users leave, waves at a time. But this actually isn't the root problem. As models develop further, the gap is really small. Now some domestic video models are already dominant; there aren't multiple model choices. At this point, what people consider isn't the model problem.
What do they consider? As Mr. Liu also mentioned, it's the asset-deposited community asset material problem.
To keep users, we have to do a "bad thing": make it hard for them to migrate. Some assets they deposit on the platform raise migration cost. But this is only that they won't leave. Meanwhile, doing this must make the product well, meeting the vertical needs of the largest paying demographic. After the product is good, they won't easily leave because they've used it a long time, assets already deposited. Then their decision is simple: pretty good, pretty cheap, assets have been there a long time, used smoothly. The material assets you produce on top deposit — for tool vendors, that's how it works.
But moving one step further, you really need to build a community ecosystem. Let relationships between creators in the community, or incentives/business relationships between creators and platform... what specific actions? We're building a creator community on the platform while letting creators do commercial order matching. Including our own overseas short-drama distribution platform APP, where there are orders and external orders for creators, which can be subcontracted to them.
Equivalent to a subcontracting model, but subcontracting is still done on the platform. This is also bidirectional: help him solve problems, he naturally helps you solve problems.
Wu Wei: I'm curious, we now have many OGC; besides those doing applications, most people are doing material generation-related stuff. On our platform, are users mostly professional creators or OGC?
Li Zhe: We define the platform as PUGC. UGC is pure beginners or solo fighters; PUGC is small teams, more suited to OGC mode. It's actually not completely one person; it might be a three-person team, two-person team. In the creator community, it's more this group.
Wu Wei: I think this also solves the current problem of many OGC communities: they always earn, but where do orders come from? Weak monetization needs platforms like ours. Grind the product well, then find ways to get more orders to solve their problems.
Mr. Liu, do you have any retention secrets? Traffic keeps increasing; you just said emotional companion retention is also high.
Liu Meiling: Now it's not convenient to say how much retention is, because this product is being announced for the first time; it's only 4 months and truly very rapid.
Also Mr. Li has already answered many of my questions; let me add some. I strongly agree with what he said about ecosystem being very important. Carp leaps over the dragon gate — as a tool you're only #1, everyone forgets, so you must reach the ecosystem baseline. Including Mr. Xiao's IP-as-ecosword ecosystem matrix, giving them a production system; including Mr. Li's redundant orders and production system — all OK. We're doing all these things; we have UGC, PGC, OGC. We also built a certain degree of OGC platform with business interest.
Including the OGC Mr. Li mentioned — it's not necessarily one person; it might be a branch business of some company partnering with you, three or four person teams are all OK, companies OK, teams OK, individuals OK; the key is to grow your own ecosystem circle, play up and down well, do your own thing well. Don't look at others' things first — like short videos, no, impossible; chasing it now won't win, better to tend your own plot of land well.
What is our plot of land? They have IP; we can only say Character World. Character World specifically refers to characters users generate here. For example, IP is a threshold — I can't reach that. My Character World is: I don't have IP, right? But users pinch a lot of digital humans here, a lot of new images, their own characters. He can go to MoreShot to make short dramas, go to digital human companionship — users create IP. He joins our contest himself, wins an award, gets his own IP figured out. Users are both consumers and creators; shallow play for shallow, deep play for deep.
So this round of investment we took from Visual China, making IP or copyright this point solid. Get your own plot of land sorted out. If he creates on your plot of land, roots are on your land. IP and characters have already grown. He has a full toolchain; gameplay is also good, whatever he wants to play can be given. At OGC we can also invest in him; on the platform he can also earn — what's not to like?
Gross Margin: How to Keep User Growth From Becoming False Prosperity
Wu Wei: It shows opportunities are many. Next topic: unit economics, or gross margin. We just talked about keeping users; not talking about acquisition through various methods. Right now applications are relatively dependent on models, growing on top of models. If users double, inference cost increases, commercial level increases, acquisition level increases — it might be false prosperity. What engineering methods can we use? Some platforms mix real and fake... What methods can be used to raise or maintain product gross margin at a good level, not because of user growth and model cost increases, even not earning $100 and throwing $90 to the model, becoming the model's distribution channel or API router. What can be done to make unit economics healthier?
Xiao Wei: From my perspective it's hard to answer, but let me talk about what we're encountering now.
iReader Overseas has been issuing production partner orders, floor guarantee plus revenue share. But he needs to use Paoman and also recharge; Paoman also gives him orders, involving cost factors in between. Now the structure is whether domestic, overseas, or micro-drama models, letting them try one by one. Give test credits, let him evaluate how much this drama costs. The evaluation test cost is assessed by the production partner, we do quality inspection. When the evaluated cost range is say 400-500, we give him an overflow price of about 20-30%. Because he needs to recharge Paoman, then reverse-derive production partner earnings.
Currently the mainstream action is test-evaluate cost, deduct cost, then leave 20-30% profit. Our own overseas distribution channel is closed loop, so we can only look from this angle. This 20-30% is the production partner's gross margin, not mine. Essentially spending 20-30% to give him, letting him know this money can be earned, covering labor cost and compute cost all in. After the drama is taken back, there's also a certain revenue share, and the ecosystem spins.
To say special cost reduction, I think it's not cost reduction, it's taking production partners to play overseas at costs they can accept — that's my angle.
Wu Wei: Understood. This is ecosystem play, stringing all parties together, designing mechanisms so they're at least willing to earn.
Mr. Li, from doing AI comic-drama tools to interactive movie game tools — what engineering methods are in between to let us as application vendors and tool vendors also have good unit economics?
Li Zhe: We've been thinking about this. The biggest difference between AI-era products and internet-era products is marginal cost. The most direct and currently main solution is direct user charging. Advertising companies actually... if you look at many products, why do they give so few free credits? There's really no way — lots of users, credits are real money, hard to just come up with lots of free trial. Let me state the big premise first.
Second, think about how to make users use fewer credits, or feel credits are effective and satisfy them. The deposit idea: while users continuously adjust prompts, deposit the final prompt. See how many times he adjusted for one image or video prompt along the way, summarize which prompt structure and content in which scenario, and use prompt templates for reuse in the next scenario.
Similar idea, learning user behavior. From the user's angle, let him spend less cost to make better quality content.
Wu Wei: Would such an excellent skill be shared with other users?
Li Zhe: It's internal system capability, our own capability. Summarize his past many attempts into general experience, helping beginners get started quickly.
Wu Wei: On some models, do you do engineering optimization?
Li Zhe: Yes. Users write prompts on the system, but we optimize them. In other words, the prompts users actually receive aren't what they said directly; we optimize. The deposit and knowledge mentioned earlier can be understood as a knowledge base; skills have two layers: one is general skills, certain scenarios like real-person interaction or certain content genres have different genre skills; each user also has personal habit skills. Let him unconsciously get smoother, use less, save more.
Wu Wei: This is like interactive movie games; many dramas are simultaneously on one line, volume is huge.
Li Zhe: Yes, interactive movie game volume is huge.
Wu Wei: Mr. Liu, you have the whole ecosystem; same unit economics question, how to build good balance, at least the platform also has good gross margin?
Liu Meiling: Mr. Xiao and Mr. Li did what we also did. Doing ecosystem can't escape two links: first, the model side must find ways to reduce cost and increase efficiency; how to reduce cost and increase efficiency? Your bargaining must be strong.
Wu Wei: Strong, in what aspects?
Liu Meiling: What's different from me integrating it myself? Various situations, but hard to explain. As the company with the largest token consumption across the entire AI district, we have our own methods and solutions. Consumption of various models is relatively large, so of course there are advantages.
Besides, another strong capability is internal tool development. We have internal toolchains for everyone to use, and all kinds of skills. Four months ago, the whole team started using skill records, letting work capability train itself.
Internally developed things — why can products be pushed out in a short time for everyone to use? First, the product must be bought internally. Make a product, internal usage rate very low, nonstop complaining, called to change or not, the product is invalid. If internally it can't connect, don't fight outside. Our own people won't buy.
Internally there are all kinds of skills, including tools, and using them is still very efficiency-enhancing.
Besides, including the ecosystem chain Mr. Xiao mentioned, giving evaluation value — this includes partner interaction, also doing ecosystem links. After all, having a model platform, we're not doing it super ahead, but the ecosystem closed loop can bind deeply cooperating clients and friends. The value system can deepen larger token consumption, more say. The loop process: the more weight, the more influence, the more at the table. What everyone talks about is talking — see who consumes more; consumption means strong value.
Long-Term Operations: After Models Flatten, What Truly Remains?
Wu Wei: No worries, let me ask you next, to avoid feeling they've already said everything.
Actually this is the last question, second to last. Today we talk about crossing cycles; first we talked retention — though products get richer, before AI users looked for products, now products aren't enough for user penetration, retention gets harder, but we have to do something. Second, we talked about how besides model consumption, what do AI products do to keep unit economics and gross margin at a good level.
Finally, let's talk about long-term operations. Though the first question asked about the next 12 months, now it's become the next 3 months; originally it said the next 3 years. Now looking at a certain model being strong, but like endorsement models, sometimes open-source is strong, sometimes closed-source is strong, closed-source sometimes A strong B strong, unclear. The world is so big, different countries and regions have opportunities. Suppose model capability flattens in the future, acquisition cost keeps increasing, there's phased traffic dividends — like AI short dramas now going hard, with phased dividends. Over time, what assets truly remain? Is it IP? User digital assets? Role-based IP formed with us? Or created economic models and systems that can keep earning here? Over time, what remains with us? I'd like to ask everyone. Want to start? They said it pretty well actually.
Liu Meiling: OK. We've been doing this point. From the approach including this year's announced strategy, it's very clear. Previously Haiyi rarely came out to share, basically not externally, because we felt we should adjust the strategic posture to the best before telling everyone. Now we can share: we're doing the Character World approach.
All Character World uses Haiyi as the aircraft carrier base, depositing large amounts of user creation, videos, images, community, serving future digital humans, C-end companionship, serving current and future MoreShot and a series of products growing on this aircraft carrier, forming a Character World IP loop. Let users enjoy the fun of creating IP here, create IP to make short dramas, create IP for companionship, return to C-end and enrich IP again — that's what we do.
Future tool building will also revolve around Character World. Future deposits — when creation is flattened, acquisition is flattened, flatten anything — what remains is creativity, aesthetics, digital assets. These AI can't replace. Other things, since AI appeared and took stock analysts' jobs, it's competitive for complex work, but aesthetics can't be taken, entertainment can't be taken. That's why it's called Haiyi Interactive Entertainment.
Wu Wei: The Character World you mentioned, what users create here, is multi-character. Actually for a long time I've communicated with many young people, teenagers, slightly younger friends. Our generation, the next generation, the generation after — what they experience is different. You can see on the market, including China, overseas, Korea, Japan — what are the popular small shops selling? Many are IP merchandise, many things you don't understand — this peripheral, plastic piece, how is it worth 100 RMB? 谷子 you don't understand, but it is worth it.
Curious — above is user-generated Character World with IP attributes. Many brands have traditional, ancient, mature IPs, but user-created IP, industrialized production, or traditional IP — like designs, various images — what's the future relationship? In the future, will it be mainly user-created IP, everyone personalizing and pursuing niche IP, or will mass IP still exist? Haiyi's Character World can also make movies in the future, also do a lot.
Liu Meiling: Mr. Liu has price too, this question is sharp. Sharp means don't ask?
Coming back. In future character IP construction, I think each generation has its own IP, each generation has its own gameplay. I also had Gu Long and this series of novel IPs; previously doing games, many IP image characters must be licensed. Throw a game touching IP and it's out, must chase. At this point you must look at domestic and overseas all.
To our generation now, like doing games there's a problem: approval — really takes forever, really hard to handle. The new generation, like going to do IP — I've talked with many young people, haven't even mentioned 谷子. Now young people aren't like when I was smallest running in fields and on mountains, very free, constructing fantasy worlds less. Now young people are often at home, in the residential compound — actually they're a group that loves fantasy, with their own spiritual world. You can see many things young people play, with their own strong little character loop world, independent space. A space recognized by yourself is called your own little world; recognized by others is an IP; recognized by more people is a bigger world. So the future is such a loop. You can ask middle and high school students.
Another point, including Japan's lost 30 years, reference many later IP events — referenceable, predictable, observable.
Wu Wei: Understood. IP and IP, IP and events. OK, I'll summarize better than you.
Actually what finally remains — you just mentioned digital assets. We say 3 months, want as something capable of lasting longer, where's the barrier?
Li Zhe: I very much agree with Mr. Liu's view; observing high school students, younger age groups — it's very consistent with what we think. We think AI is now rapidly developing, from the bottom rapidly. Moving further, who are the AI products built for? Maybe for future teenagers. Thinking long-term, wanting to build a barrier, long-term company product, it's very likely built for them.
Wu Wei: Thought it was for agents, or for humans. Brilliant view.
Li Zhe: Talking about content side. Regarding this point, a few barriers: content copyright assets, user data, ecosystem community relationships.
Content-wise, there's now a lot of content that can be secondarily mined, used, reinvested, but people with content expression needs are still very many. Especially many high school students, with their own spiritual world, very rich, minds conceive many imaginations, rich imagination, conceive many worlds, even build their own worldview.
Has everyone heard of the board game called TRPG? It has history, but still very popular, within small circles very obsessed, like board game script killing. TRPG products are everyone sitting together completing stories and scripts; everyone has great interest in story creation. When we make interactive movie games, one tool leans toward story generation, where the mode is through everyone together or a few people doing TRPG simulation, ultimately producing a story, everyone evaluates and polishes, feels good, then pushes it into a movie game. Doing many such processes originally, multiple people create one world.
This is equivalent to what Mr. Liu mentioned — context migration is very difficult; multi-person context all concentrated in the tool and system, and what deposits afterward, forming content assets. Migration hard, retention solved, barrier solved. Think in this direction.
Wu Wei: Content assets have context, social relationships. Building the entire barrier, user cost.
Now I'll say it. iReader's overseas future 3 months development, including some policies — this is quite dry, teachers might be more interested.
First, the next 3 months will definitely revolve around Paoman; the entire overseas business line will all move to Paoman. The overseas business line splits into script purchasing, finished drama purchasing, production partner distribution, and Paoman compute subsidy policy. Starting from the script step, we've already established cooperation with many owned IP parties and script studio partners. iReader-approved scripts are prioritized for purchase; unselected ones help find downstream buyers. Scripts have started advancing; after studios settle into Paoman, the editorial team requires all moves to Paoman custom commissions — this part's profit is可观, the amount is可观.
Besides, finished dramas will also move to Paoman later. Requiring the entire AI short drama, 4-5 quality inspection departments all move to Paoman quality inspection. Confirmed quality-inspected purchases are bought immediately, even exclusive; amount is可观. Jointly playing with other distribution channels. Many production partners make dramas and find nobody wants them, ending up with only 20+ revenue share, money runs over but no money earned. We want to help him solve monetization, so finished drama side is adding large volume; we've already started introducing.
Further down the most core, each has its own development: some want to do copyright, some scripts, some production partners. For production partners, we've already opened all Paoman testing; capable people come in, capable to take high floor guarantees. All partners settled with or cooperating with iReader are qualified and capable to get corresponding floor guarantee amounts; the amount won't be particularly low. Overseas first topic said certain revenue share ratio will be given; already in progress.
Finally main push core Paoman. Now compute cost reduction and compute subsidy main thinking point is cost. Doing ecosystem closed loop, not letting partners lose money — they can win small, but absolutely won't lose. Bundled with iReader, we take him continuously overseas. This is the currently drafted 3-month purchasing policy, cooperation policy; offline can communicate later.
Because now AI short drama going overseas is in an explosion period. In the explosion period, how much longer will it explode, how big an opportunity?
Xiao Wei: You can see AI short dramas aren't that high on the charts, but listed companies don't need ranking. Purchasing intensity basically won't be low. Over the next 1-2 years, iReader Overseas will still be a focus. Live-action dramas are too expensive; with $100K+, how many AI dramas can be made? Over the next year to next June or July, basically continuously deep overseas. Maybe there will be changes, but the overseas direction won't have problems for the next 18 months.
Wu Wei: Indeed the hottest. Seeing many teams expanding, doing production, doing sharing.
Besides, throwing out a cooperation metric point. Mr. Liu said will also cooperate with many small and medium production partners; iReader will also cooperate with small and medium production partners. After cooperation, take investment, contracted compute, contracted capacity. Maybe various companies will also do it, choosing cooperation based on cost. iReader has already invested in many small and medium production partners.
How much have you invested?
Xiao Wei: Next time offline, I also care about the data.
Wu Wei: OK, finally everyone has a closing statement. Back to AI long-term growth crossing cycles; if using one metric to prove an AI product has crossed the cycle, what might it be? Let me summarize, one line each.
Xiao Wei: I think positioning is very important. What is the positioning in going-global or domestic industry? What do you hope to become in the future? Especially critical. I find many people both do production and after doing it do everything; finally maybe earn money, but never succeed, not earning from excelling in one place. Positioning is very important. Is it possible you can't choose? No problem; now I very much hope to establish connections with more partners; finished dramas are already cooperating and advancing. Each company's positioning is especially important; find your ecological niche, go deep and thorough.
Wu Wei: Yes.
Li Zhe: A relatively conventional metric, but I think it's very important: retention. Especially when the underlying model makes another step-change, this moment's retention is most representative. If it can still retain.
Wu Wei: Specifically next-day, 7-day, or what retention?
Li Zhe: Next-day, 7-day aren't particularly important. Retention as a metric, relative to other growth, is the most important in my view.
Wu Wei: That's how it is, retention positioning.
Liu Meiling: Their two answers are already very key; I'll add monetization. Positioning, retention, monetization. No need to say wrong things for gimmicks; need to tell everyone successful experience, things done thoroughly from post-mortems. Mutual recognition is the cornerstone and gene for the three of us coming out to share. Whether positioning or retention, there's another point: revenue is also important. If product retention is bad, positioning is bad, monetization is bad, don't do it. Completely not standing from a market perspective, don't waste time; quickly go to a company with good monetization and retention and study how they play.
Welcome to join Haiyi; excellent students are welcome to join.