---
title: "A Business That Turns RMB 2M into 50M Is On Fire — But AI Games Are Still Doing Token Labor"
author: "Unique Research"
sourcePublication: "Unique Research Substack"
originalPublishedAt: "2026-10-02T14:09:45+00:00"
canonical: "https://ffcap.cn/en/research/a-business-that-turns-rmb-2m-into"
source: "https://uniqueresearch.substack.com/p/a-business-that-turns-rmb-2m-into"
language: "en"
---

# A Business That Turns RMB 2M into 50M Is On Fire — But AI Games Are Still Doing Token Labor

[![cover](https://substackcdn.com/image/fetch/$s_!JzTW!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd0309ef8-29ed-408c-8b70-ddf06cb3370d_2730x1536.jpeg)](https://substackcdn.com/image/fetch/$s_!JzTW!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd0309ef8-29ed-408c-8b70-ddf06cb3370d_2730x1536.jpeg)

**The inflection is real. But sitting right on top of it is a ledger that hasn’t been balanced yet.**

Let’s start with the math.

The breakout interactive live-action romance game **“Oh No! I’m Surrounded by Beautiful Women”** (_Wandàn! Wǒ bèi Měinǚ Bāowéi Le_) cost roughly **RMB 2 million to make** and pulled in about **RMB 50 million in sales**. The follow-up titles didn’t hit as hard, but they still moved around 200,000 copies and brought in nearly RMB 10 million.

For a few million RMB, the risk and barrier are far lower than shooting a film—and the payback is fast. No wonder traditional film and TV people are piling into this space.

These numbers come from a panel at the Chengdu AI Entertainment Conference. The person doing the math is Wang Jialun, CEO of Game Tea House (游戏茶馆).

Around the same table sit three others: Sun Weizhe, co-founder of ReelFork, an interactive content platform; Cao Na, brand and marketing director at Vertens3D—behind it sits 3D Home (三维家), which has accumulated billions of model assets; and Liu Liping, general manager of Chengdu Qi You You (柒玖游), a develop-and-operate game company behind titles like _Shendiao_ (射雕) and _New Chinese Paladin: Sword and Fairy — Questions of Love_.

Four very different backgrounds, gathered to talk about AI interactive content. And a subtle temperature difference emerged: when they talked about interactive movie-games with pre-scripted branches, everyone was doing the math and getting more excited by the second. When they turned to truly AI-generated content, the math stopped working.

That is probably the most honest water level of this industry right now: the inflection is real, but above it sits a ledger that hasn’t been balanced.

Interactive movie-games are not new. Sun Weizhe traced the timeline: from _The Invisible Guardian_ (隐形守护者), to _Surrounded by Beautiful Women_, to the “gold-digger games” and _盛世天下_ (Flourishing World). The common thread is clear—users walk along a route the creator designed in advance, and every branch is pre-built. “The user is more of an observer than a participant.”

But Wang Jialun offered an easy-to-miss angle: single-player games are not entirely competitive with each other. They bring new users to the category.

Someone who played _The Invisible Guardian_ will very likely go on to play _Surrounded_ and _Flourishing World_. He estimates _Surrounded_ brought about 2 million new users into the genre, and _Flourishing World_ may have brought 10 million. Every breakout hit enlarges the whole pond—rare in other content industries.

Even sexier is the derivative spending. _Flourishing World_ held offline fan meetups; fans played the interactive game while chasing the on-screen leads, snapping up signed photos and small cards without hesitation. Wang says it is approaching the logic of fandom. It’s even clearer in otome games: players treat characters as imaginary romantic partners. If a sequel swaps the lead or changes a character’s personality too much, the backlash is fierce.

Model capability is now adding leverage to this business. Sun’s judgment: as text and video models get stronger and production cost falls, ordinary people’s creative energy will be unleashed. On his own platform, there are already many non-professional creators—”maybe stay-at-home moms.”

His endgame: original authors build the rules and framework; the audience enters that content world, makes derivative creations within the framework, tunes plot branches, and forms their own version.

The consumer is simultaneously the creator. The more they participate, the deeper the emotional projection—and that participation itself is consumption value.

This balances the interactive movie-game ledger on another level: users are not just consuming content; they are producing it.

Traditional game characters are fixed content assets designed in advance by planners—finished the moment the game ships. Liu Liping’s definition of an AI NPC is clean: it is more like a running system that responds in real time, within the worldview’s rules, based on the player’s language, interaction history, and understanding of the character.

On AI memory, her view is counterintuitive: an AI NPC does not need to remember everything.

“People also remember important choices and forget irrelevant details.” The real point is whether the game world actually changes after the player talks to a character. “If it’s just a few extra lines of dialogue, it means little.” What the NPC should remember are the player’s key preferences—leaning aggressive or conservative—so the player feels, in later battles and interactions, that “it knows me.”

And she drew the boundary very clearly.

“We don’t pursue AI being 100% human-like. We want it to be a stable character.”

Players come back because of a stable character; they form long-term expectations toward it, and only a long-term relationship can support LTV.

This pulls the design goal of AI characters back from showing off to doing business.

Sun Weizhe added another angle. He told two on-set anecdotes: in _My Fair Princess_ (还珠格格), the character Consort Rong is ruthless on screen, but the actress was completely different on set; in the classic _Romance of the Three Kingdoms_, the actor who played Zhang Fei, Li Jingfei, loved reading in private—a huge contrast to the character—and Tang Guoqiang (Zhuge Liang) also had his card-playing, joking side.

His conclusion: characters need to be rounded, not just the one face from the main plot. The same goes for AI characters. Through side plots, different states, and behavioral motives, you let the user see more sides; only then does the character have logic and hold people.

Cao Na, speaking from the technology supply side, poured cold water. Her words: the advantage of AI interaction is surprise—it gives users things beyond what was preset; the weaknesses are controllability, continuity, and real-time performance.

Dialogue generation is decent now, but if you need to generate new spaces and new characters in real time, whether the speed can keep up is an open question.

For a 3A game like _Black Myth: Wukong_, which guarantees experience through high-cost design, she is reserved about whether adding generative characters and interactions would hurt the experience.

Is AI companionship a real need or a fake one? The four agreed, unusually: the need is real.

Liu Liping gave numbers directly: AI companion products like Xingye (星野) have already hit 20 million MAU or higher, and game companion assistants are growing fast.

Sun’s angle is the most “philosophical”: whether it’s AI companionship or human companionship, it is essentially a relationship. Real celebrities can fall from grace (塌房); AI characters carry no such risk. You usually can’t reach a real idol, while AI characters can offer more everyday interaction.

Cao Na’s observation comes from around her: many young people on her team choose to stay single. Once food and shelter are solved, the need for companionship doesn’t disappear—it just shifts its object. A friend of hers, unmarried in her forties, pours a lot of emotion into celebrity fandom.

But Liu Liping immediately pulled the conversation back to the ground.

“User demand, revenue, and eventual profitability are three different things.”

Having run publishing and operations for years, she habitually asks first: does it make money? For a category to be viable long-term, it needs stable gameplay and stable monetization. Only when users keep coming back to the same character does a relationship form, and LTV hold up.

Wang Jialun added evidence. The otome market is genuinely big—Papergames’ _Love and Deepspace_ drew crowds at Gamescom, and he watched a free giveaway item resold for 100 euros. Many AI game teams that raised funding this year are building AI otome. But so far, “we haven’t seen a particularly successful case that keeps users playing and paying and runs for the long haul.”

The demand is there, the money is there, and the product that can sustainably catch it hasn’t appeared yet. That is the current state.

Wang Jialun told another case worth pulling out: _Rebirth Simulator_ (重生模拟器). He interviewed the team last month. The player plays Chongzhen (the last Ming emperor) and tries to change Ming dynasty history. Traditional history simulators pre-script most rules and outcomes; with AI, development becomes more random and more interesting. It performed decently in the market, but drew many negative reviews—aimed at the monetization model: after buying the game, players still had to pay for tokens.

Steam players are used to one purchase and no further spending. After the review flood, the game went free and charges for tokens. He understands the project is backed by a large platform: “a small team without funding probably couldn’t survive it.”

This is worth pausing on, because it exposes the fundamental difference in cost structure between AI games and traditional games.

Wang compared it with _Genshin Impact_: it may have cost a billion dollars, but later revenue can reach tens of billions—the more users, the lower the marginal cost. AI games are the opposite: every additional active user adds a model-call cost. Switch to a cheaper model? Quality drops, players refuse, and you can’t retain them.

Plainly: traditional games are built once and then collect money lying down; AI games burn compute live for every user served.

Until this contradiction is resolved, the business model stays twisted.

Liu Liping’s take on Vibe Coding confirms this from the side. She thinks hand-crafting games with Vibe Coding suits demos and hyper-casual games—arrow, puzzle, and number-driven mechanics that can break out on Douyin, WeChat mini-games, or overseas markets if you pick the right category.

“But that’s more like a shortcut to monetization; it’s not the same as a premium game.”

Big games still come back to gameplay: the fresh opening wrapped in AI may lower customer acquisition cost, but the novelty passes. “Without mature gameplay, it’s hard to support long-term retention and payment.”

What she is more looking forward to is AI reconstructing the interaction and content experience of traditional genres like SLG, card, and MMO. Only by building new experiences can a new track open.

The most certain opportunity in the whole panel was not on the content side—it was on the tooling side.

Cao Na’s world model has four pipelines: perceive the world, generate the world, edit the world, and manufacture the world. A concrete example: send a robot or robot dog through a venue and generate a 3D space with geometric logic; photograph one view of a space and the model extends and imagines around it, filling in the 360° scene and completing the modeling, directly saving labor and cost.

For the game industry, AI’s most certain value right now is on the production-cost-compression end, not the content-generation end.

Liu Liping’s practice points the same way: AI already contributes a large share of efficiency in concept art, art production, planning, and plot branching. What’s truly stuck is the deep numerical design of large games—AI’s understanding of overall architecture and numerical systems tends to “drift,” so teams rely on tools like numerical dashboards to have it assist verification, while the R&D team keeps control of the framework.

Cost reduction runs first; revenue growth is still being validated. The order is identical to the script AI has played in most industries over the past year.

Sun Weizhe says his founding habit is to “run a little ahead”—from big data and quant to SaaS, then video foundation models, now betting on interactive content, gambling on the window before applications explode.

That judgment is probably right. The interactive movie-game ledger is already clear, and the demand for AI characters and emotional companionship is plainly there. The only unbalanced ledger left is model cost and business model.

[![cover](https://substackcdn.com/image/fetch/$s_!JzTW!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd0309ef8-29ed-408c-8b70-ddf06cb3370d_2730x1536.jpeg)](https://substackcdn.com/image/fetch/$s_!JzTW!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd0309ef8-29ed-408c-8b70-ddf06cb3370d_2730x1536.jpeg)

---

Original publication: https://uniqueresearch.substack.com/p/a-business-that-turns-rmb-2m-into
On-site reading page: https://ffcap.cn/en/research/a-business-that-turns-rmb-2m-into