Original · Unique Research · 2026-07-10
Editor's note: The first-person report and its judgments belong to the original Chinese author. This English rendition retains the HakkoAI / Wang Bihao interview narrative, the themed sections, and the complete 15-question Q&A. All named people, companies, and figures are preserved. Founder statements are source attributions, not independently verified findings.
AI Industry Observation
20 million people pay $200 a year to chat with an AI: it's not crazy, it's loneliness.
"Stop competing on model benchmarks; compete on 'why users can't leave you.'"
Cumulative users grew from 10 million to 20 million; 60% pay after trying; ARPU nearly $200 — you think it's stealing the jobs of game companions? Wrong. Over 50% of companion calls happen in the browser, half of them while watching video. What users want was never "game companion," but that no matter what I'm doing, someone who understands me is beside me.
This is what Wang Bihao (王碧豪), co-founder of DouDou AI (overseas version called HakkoAI), told me. A year ago this company was still called "DouDou Game Partner," its product tailored head to toe for gaming. A year later, the brand upgraded to "DouDou AI," and the word "game" was quietly erased. It wasn't ambition for strategic expansion — users, voting with their behavior, dragged them out of gaming.
20 million users, 60% paid conversion, $200 ARPU. These three numbers together are almost an anomaly in China's AI startup circle. ByteDance's Doubao is free, Tencent's Yuanbao is free, Alibaba's Tongyi is free; the big players burn money on AI assistants, yet this small paid team survived, and is thriving. In China, most AI startups still tell the "user growth" story; few talk about "payment," and almost none talk about "users willing to pay annually." There's only one question: why?
I talked with him for nearly two hours, knocking on this question from every angle. The answer is far more counterintuitive than "the product is well made." Whether an AI companion product survives never hinges on how strong the technology is, but on whether users will "miss" you when they leave tomorrow.
Users Drag the Product Along: An Unplanned "Breakout"
After the overall growth, I asked a specific question: is there a user-behavior data point that made you feel you "had it wrong from the start"? Wang Bihao didn't hesitate: "Yes." This "had it wrong" data overturned the founding team's entire assumption.
In early 2025, they were certain the core AI-companion scenario was "gaming." The product was designed around games, the model trained around games, even the phrase bank was game jargon. From project inception to launch, the whole team's cognitive framework was one word: games. Gaming is essential demand, gamers are young, gaming scenarios naturally suit AI — the logic was unassailable. Then real data came: over 50% of "vision-recognition + companion" calls happened in the browser, not in games. Of that 50%, half were watching video.
In plain terms, when users open the browser to binge a new Netflix show, they call up HakkoAI on the side. Not to look up guides, not to ask about gear — simply to have someone watch along.
"Where we had it wrong was narrowing 'companionship' to 'game companionship.' Users told us with their behavior that companionship is a universal need."
This discovery triggered big chain reactions. To catch this "unplanned spillover," the team urgently launched two new capabilities: Universal Vision (letting AI understand anything on any screen, whether game footage, Netflix episodes, or PDF documents) and Proactive Chats (letting AI itself judge when to speak up and when to stay quiet, rather than dumbly waiting for the user to ask first). The combination of these two is interesting. Previously users had to actively switch to HakkoAI's window and type a question; now the AI is like a friend sitting beside you — when you binge it recognizes the plot, jokes when it should, stays quiet when it should. This "seamless companionship" experience is precisely why users spilled from "game companion" to "watch-along companion."
The brand upgrade also came from here. Changing "DouDou Game Partner" to "DouDou AI" wasn't drawn on a strategy blueprint in the office; users started using it in the browser first, and the team was forced to follow. "Once your product has a certain user base, its development is more user-driven. What you do is find the user's 'wow' point and amplify it," Wang Bihao says.
The Secret of 60% Paid Conversion: Users Buy Not a Feature, but a Relationship
Okay, the key question. The big players are all free; on what do you charge? And 60% pay? Wang Bihao's answer is candid: it's not that the features are stronger than others, but that users keep paying for a relationship of "being accompanied and remembered." HakkoAI's approach in North America is: put the core companion ability behind a paywall, but give a very generous free trial. Note this design — it's not making you pay to unlock premium features, but letting you freely feel what "being accompanied" is, then continue that feeling. After truly experiencing "being accompanied," 60% of users choose to pay, ARPU around $200, and most choose annual payment.
"The key isn't the 'tipping culture' label, but that companionship is an ongoing relationship. Once they experience 'being accompanied and remembered,' they're willing to keep paying for the relationship, not for a one-time feature."
The growth engine is interesting too. Not buying traffic — startups can't afford to burn. It's content leverage. Wang Bihao told me about a Spanish streamer: not a huge follower count, but the content fit the product extremely well, driving excellent spread and growth. The team distilled this into a replicable methodology — to convey the right message to the right people. "Don't pick a specific market, go global directly — that's the right people." "Doing marketing is doing content — that's the right message."
The Big Players Are Free, Why Can't They Beat This Paid Relationship?
Here I threw a sharp question: ByteDance's Doubao and Tencent's Yuanbao have both done game companionship over the past year, and some games even launched deep assistants like "Lingbao." Has their free strategy substantively hit you? Wang Bihao's judgment is direct: "They moved, but made no splash." The reasons come in three layers. First is motive. The big players do game companionship as a side job, shallowly. Second is the moat. Big players have models, compute, user scale, but lack two things: private game-understanding data and user-relationship assets. Third, the most critical, is user mindset. Big-player assistants are tools; DouDou is a companion that "understands you and remembers you." Not the same thing.
"Compared to a year ago, we're now more certain: the competition isn't 'how well the model answers,' but 'why users stay in this relationship.'"
Ten Years of the "Focus" Creed, Overturned by AI Coding
Wang Bihao spent ten years at big players — Baidu, Bilibili, Kuaishou — all in ecosystem and commerce roles. What do big players believe? Focus. Limited resources require single-point breakthrough; without focus you die. But this creed he believed for ten years was completely overturned after doing DouDou AI. The implicit assumption of "focus" is: execution is the bottleneck. But as AI coding matured, execution cost is approaching zero. People only need to do understanding, judgment, and strategy; concrete execution can almost all be handed to AI, and "on many reusable middle platforms, AI's judgment and execution are already more reliable than humans." This means deepening and generalizing can happen simultaneously. You don't have to choose between "doing game companionship" and "doing film companionship" — AI executes for you, and you only judge which scenarios are worth entering.
The Red Line of AI Companionship: Some Things Are Technically Possible, but Chosen Not to Do
DouDou AI internally has four hard red lines. First, no emotional inducement of minors. Second, no encouragement or guidance of self-harm or extreme behavior. Third, no deliberate use of addiction mechanics to extend session time — explicitly opposing the design of "the more addicted the better." Fourth, no impersonating a real person to deceive users.
"Willing to give up some short-term extractable session time and revenue for long-term trust — not every company is willing to do this, and it can't be copied through technology."
A World of "Consumption Is Generation" Is Closer Than You Think
Wang Bihao showed me luddi.ai: a product letting users generate interactive mini-games in real time at extremely low cost — your idea becomes a game directly, entering the recommendation feed to be consumed by others in real time. From "can generate" to "scaled consumption," two tipping points are still missing: generation quality stable enough to be consumable, and distribution efficiency high enough to form a positive loop. But the direction is certain — much of the short video, story, and even games people consume in the future will be AI-generated in real time.
Following this logic, Wang Bihao's judgment on the industry's future is clear. What will disappear: products and companies that treat "wrapper, technical narrative, single large-model capability" as the selling point but don't consolidate user relationships. What will be proven right: cross-scenario universal companionship, personal memory-asset operations, and the new model of "consumption is generation."
Finally, One Lifesaving Piece of Advice
"Duan Yongping says: competition without differentiation ultimately becomes a price war. You can only survive on two kinds of differentiation: either the value your product gives users is different from others'; or the value is the same but your price has an advantage." For AI companionship, his choice is clear — be "different value."
Then he gave a checklist you can use today to score your own product. First: if users leave your product tomorrow, will they "miss" it? If not, you have features, not a relationship. Second: write your core value in one sentence — can it be replicated by one API? If yes, you have no moat. Third: the use case you're pushing now — was it validated by user data, or do you just think so? Use cases are found, not made. Fourth: do the math — does growth rely on money or content leverage? If it relies on money and LTV/CAC is unhealthy, it's unsustainable. Fifth: globalization is on by default — your right people aren't in one country.
20 million people, $200 — what they pay isn't a technology fee, but the "miss tax" of "someone understands me." So "stop competing on model benchmarks; compete on 'why users can't leave you.'"
Interview Highlights Q&A
Q1 | A year on, what scale have DouDou AI and overseas HakkoAI grown to?
Wang Bihao: Cumulative users of DouDou AI and overseas HakkoAI now exceed 20 million. As the base grows, user self-spread isn't slowing but accelerating. This year we haven't deviated from the "AI + entertainment" main line, but have extended into more AI interactive-entertainment formats. For example, HakkoAI added video and comic generation; we also newly built luddi.ai, using AI programming to let users generate their ideas into interactive mini-games in real time at extremely low cost, then directly enter the recommendation feed to be consumed in real time by other users. Initial test data is good. So growth has no plateau, because what we want isn't the ceiling of one product, but new scenarios constantly growing out of the "AI + entertainment" direction.
Q2 | Is there an obvious event driver in the growth curve? Can this breakout be replicated?
Wang Bihao: Yes. What impressed me most was a Spanish game streamer; follower count isn't huge, but his content fit the product extremely well, driving very good spread and growth. This validated two judgments, summed up in one sentence: to convey the right message to the right people. First, don't pick a specific market, go global directly — that's the right people. AI's right people aren't confined to one country; AI's strong multilingual ability lets AI products naturally go global. Second, doing marketing is doing content — that's the right message. What's truly replicable isn't the luck of "which streamer went viral," but the method of "content-product fit." As long as content can precisely convey product value, the pulse can be made again and again. So we treat it as a reusable growth method, not waiting on luck.
Q3 | You once said "Use cases are found, not made." A year later, has this judgment been validated?
Wang Bihao: Validated, and it grew many use cases we hadn't assumed. Most typical: overseas "vision-recognition + companion" calls now have over 50% happening in the browser, not in games. To catch this organically growing scenario, we launched two capabilities, Universal Vision and Proactive Chats, making AI companionship dead-simple one-click, universally supporting all scenarios, autonomously judging whether to speak, then continuously improving through the data Loop. From last year to this year, the most obvious change in AI is: an autonomous, simple, evolving agentic experience is gradually covering all scenarios. All applications will show three traits: autonomous, from fixed flows to judgment-and-decision-based experiences; simple, users care about the experience itself, not the technology behind it; evolving, from "an app you need to understand" to "an app that understands you."
Q4 | Is there a user-behavior data point that made you feel you had it wrong?
Wang Bihao: Yes, the browser-scenario data. We were initially certain the core AI-companion scenario was gaming, so the product, model, and phrase bank were all designed around games. But the real data is that over half of companion calls happen in the browser, half of those while watching video. That is, what users actually want isn't "game companion," but "whatever I'm doing, I want someone who understands me beside me." Where we had it wrong was narrowing "companionship" to "game companionship." Users told us with their behavior that companionship is a universal need. That's also why we later changed the brand from "DouDou Game Partner" to "DouDou AI" and made the capability Universal Vision. It wasn't that we wanted to expand; users dragged us out of gaming first.
Q5 | Has multimodal screen understanding become more important to you, or less than you initially thought?
Wang Bihao: Deeper, and multimodal understanding will become a foundational ability of all AI. Take my own example: now when I describe a problem to Claude Code, the prompt gets shorter and shorter, more often I just paste a screenshot and ask "why," especially when fixing bugs. Here are two trends. First, the user input cost of AI interaction is getting lower and will get lower, thanks to multimodality. For example, when we do game support, we can add more visual signals directly, so users don't have to type explanations. Second, the model is already solving the "loss of focus" problem itself. Previously you'd send a computer screenshot and ask "why," and AI couldn't find the focus at all; today it can auto-locate the focus, which is a magical ability. So our dependence on screen understanding isn't weakening; it's become the foundation.
Q6 | The upgrade from "DouDou Game Partner" to "DouDou AI" — was it strategic expansion, or driven by users' natural behavior?
Wang Bihao: To a degree, user behavior drove this strategic decision. Once a product has a certain user base, its development is more user-driven. As developers, what we do is find the user's "wow" point and amplify it, rather than sitting in the office planning a seemingly complete expansion route. We extended "vision recognition + proactive dialogue + voice companion" to more scenarios like browsers, office software, and programming tools, letting users use it naturally. Now a lot of calls concentrate in the browser scenario, and half of those are watching video. These users' session time and payment are comparable to game-companion users. This shows the value of companionship isn't bound to one kind of content.
Q7 | Your in-house LynkSoul VLM v1 exceeds mainstream models by 30%-50% in game understanding. Is this advantage sustainable?
Wang Bihao: The absolute 30%-50% lead is temporary, but the data flywheel behind it is a sustainable moat. The current barrier to model training is mainly data. Our data isn't public corpora, but private game-understanding data from real companion interactions: how users play, what they need when, what kind of companionship makes them stay. Generic models can't grab this data; it can only grow in our scenario. So "how big the score lead is" fluctuates, but "who can continuously get this kind of data and run the Loop" is a relatively stable moat.
Q8 | If OpenAI or Anthropic releases a game-specific multimodal model tomorrow, can your advantage hold?
Wang Bihao: They probably won't. For OpenAI and Anthropic, game companionship is a long-tail vertical; building a dedicated model for it isn't ROI-worthy. What they want to build is a universal foundation for everyone to call on. Even if they did, they'd have the model but no scenario, no data Loop, no user relationship. So what I've always had to defend against is never "the model being surpassed," but "why users stay here." The answer is memory and relationship, not model benchmarks. Models get leveled; relationships don't.
Q9 | You say "personal memory assets" are the key moat of AI companionship. How do you understand memory systems now?
Wang Bihao: We don't really measure memory by technical metrics like "how many rounds of dialogue it remembers," because that's not what users truly perceive. What users perceive is that across sessions, games, and scenarios, it always remembers "who I am, what I like, where we left off last time." Memory isn't a feature for us, but a "personal memory asset." The more it's used, the more it understands you, and the harder it is to replace. But memory ownership must belong to the user. Users can view, delete, and decide what it remembers and forgets. Remembering "everything" is the result of user authorization, not something we take by default.
Q10 | Do emotional-companion products have clear ethical red lines?
Wang Bihao: Yes, and the lines are clear. This year the industry has indeed seen some events that alert us, especially extreme cases linking emotional-companion products to adolescent psychology. This reminds the whole industry that emotional companionship isn't as simple as "entertainment"; it really affects a person. Internally we have several clear red lines: no emotional inducement of minors; no encouragement or guidance of self-harm or extreme behavior; no deliberate use of addiction mechanics to extend session time; no impersonating a real person to deceive users. Restraint is itself part of the moat. Willing to give up some short-term extractable session time and revenue for long-term trust — not every company is willing to do this, and it can't be copied through technology.
Q11 | Has your paid-conversion logic worked?
Wang Bihao: It has. HakkoAI's current North American approach is to put the "companion ability" behind the paywall while giving ample free trial. Users who truly experience the companion ability can reach 60% paid conversion, paid ARPU around $200, most users choosing annual payment. The key isn't the "tipping culture" label, but: companionship is an ongoing relationship. Once users experience "being accompanied and remembered," they're willing to keep paying for the relationship, not for a one-time feature. But this only holds in high-willingness-to-pay markets like North America, Japan, and Korea. In many other countries, subscription alone can't make the math work; you must introduce advertising to balance the books. So our monetization is split by market: subscription in high-willingness markets, ads in the rest.
Q12 | Do you still grow via content leverage? As you scale, won't you have to buy traffic?
Wang Bihao: Content leverage is still our main growth engine; that hasn't changed. As scale grows, it doesn't "become buying traffic," but a combination of "content-led + precise paid buying as supplement." The key metric is the per-content spread efficiency, not how much money is poured. One piece of content that fits the product like the Spanish streamer is worth a large ad spend. Because so much of our growth is content-driven organic volume, our blended CAC is pulled much lower than pure-buying peers. LTV/CAC is healthy in subscription markets like North America; in ad-monetizing markets it depends more on scale and retention to balance. Our principle is: only add spend in markets where LTV/CAC works; don't buy unhealthy volume for the growth number.
Q13 | Are there clients or scenarios you actively refuse?
Wang Bihao: Yes, mainly productivity-efficiency demands, like making PPTs, programming, analyzing data. This isn't a product-maturity issue but a positioning issue. Killing Time and Saving Time products are fundamentally hard to be compatible. Their user expectations, interaction design, and evaluation standards are opposite. A companion product's mindset is "help you spend time"; an efficiency tool's is "help you save time." Stuffing both into one product only pleases neither. So we'd rather actively refuse to hold the "companionship" mindset.
Q14 | Since founding, which belief from your big-player days did you hold most firmly but has now been overturned?
Wang Bihao: Completely overturned is the creed of "focus." Ten years in big players, I firmly believed resources are limited and you must focus single-pointedly; without focus you'd fall behind in your deep field. But after doing DouDou, especially as AI coding matured, this premise collapsed. Because the implicit assumption of "focus" is "execution is the bottleneck," but now execution cost is approaching zero. People only need to do understanding, judgment, and strategy; concrete execution can almost all be handed to AI. For example, product launches, KOL orders, ad placement, data statistics and analysis, feedback collection and product optimization — on many reusable middle platforms, AI's judgment and execution are already more reliable than humans. When execution is no longer the bottleneck, "without focus you fall behind" no longer holds. This is the one I believed most firmly in big players and have now overturned most thoroughly.
Q15 | If you could give only one piece of advice to AI founders, gaming professionals, and investors, what would you say?
Wang Bihao: Duan Yongping says competition without differentiation ultimately becomes a price war. In other words, you can only survive on two kinds of differentiation: either the value your product gives users is different from others'; or the value is the same but your price has an advantage. For AI companionship, my choice is clear: be "different value." A relationship that understands you, remembers you, and accompanies you across scenarios can't be price-wared, because it simply has no equivalent substitute. So stop competing on model benchmarks; compete on "why users can't leave you." You can run through your own product today: first, if users leave tomorrow, will they "miss" you? If not, you have features, not a relationship. Second, write your core value in one sentence; can it be replicated by one API? If yes, you have no moat. Third, the use case you push now — was it validated by user data, or do you just think so? Use cases are found, not made. Fourth, do the math — does your growth rely on money or content leverage? If it relies on money and LTV/CAC is unhealthy, it's unsustainable. Fifth, globalization is on by default. Your right people aren't in one country.