Meitu's Xinhong Wu on AI apps: deepen vertical use cases and control model and monetization costs

AI Product Aying distills a podcast's lessons on product choices, small-team validation, subscriptions plus service fees and global markets, alongside reflections on the author's own startup.
I have to recommend this podcast episode with Meitu CEO Xinhong Wu. It is very down-to-earth.
Meitu is hardly a star company these days. After all, people usually talk about DeepSeek, Kimi or major internet companies such as ByteDance, Tencent and Alibaba.

But for a grassroots entrepreneur like me, this episode contains so much experience and judgment that I can put directly to use.
For my first ten years as an entrepreneur, I mostly built utility products. Over the past two years, though, I have felt somewhat lost. Traffic to our earlier utility apps kept declining, and I had little direction for new products.
As CEO, I lacked confidence myself.
I keep feeling that many AI products—I will not name examples and offend people—solve problems that ChatGPT can handle too.
We also researched several use cases and found that users' subscription fees could not cover the token costs. Unless we kept going down the fundraising route.
Hearing this episode just as I was planning for the new year at year-end made things much, much clearer.
#01
Where exactly are the opportunities for startups building AI applications?
Back in 2013, mobile-internet entrepreneurship was booming. When founders gathered, the conversation was basically about what kind of app each wanted to build.
But in this AI wave, I do not think building just any app offers an opportunity.
Look around: how many independent developers genuinely make money from AI applications themselves?
Very few. Many ultimately monetize through content creation, consulting or training instead.
As the barrier to developing applications has fallen sharply, the barrier to making the business model work end to end has risen considerably. I previously saw this comment under a friend's Weibo post, and it resonated deeply.

I used to pay for memberships in all sorts of utility apps. Now I no longer need many of them; I just have ChatGPT handle things.
Many AI applications are also easily swallowed up by general-purpose models or super AI apps such as ChatGPT.
For example, a content-creation product I used connected several leading industry models, wrapped a layer of specific use cases around them, and added features such as a Skill store and templates.
After using it a few times, I found it less smooth than using ChatGPT directly with my own customized Skill and a Feishu workflow.
More importantly, I already pay for ChatGPT. If my quota is sufficient, why spend another $20 every month on something ChatGPT can do itself?
That logic just does not make sense.
Meitu CEO Xinhong Wu's experience is that you must find use cases that general-purpose models cannot easily replace directly, yet users are willing to pay for, then commit to them and go deep enough.
Take Kaipai, which they developed.
The product began as a teleprompter feature in BeautyCam. But analyzing the data, they found that this small feature ranked third among all paid features.
The product manager therefore went to interview users specifically and discovered that many were real-estate agents, insurance agents or creators selling knowledge-based content. They needed talking-head videos for marketing and customer acquisition.
That was when Meitu spun the teleprompter out into Kaipai.
But that was not the end.
As the product developed, they kept digging into these industries, studying different users' workflows and specific needs. They even went to insurance-industry exhibitions to promote it.
Wu said they did not initially understand these industries either. Users gradually taught them much of that industry knowledge.
This really struck a chord with me.
Looking back at my own AI product work over the past year, I seem to have become increasingly unwilling to do the hard work.
When we started businesses before, entering an unfamiliar field meant willingly spending lots of time researching users, understanding what they did every day and their pain points, then working out how to make the product good.
Now, with AI, developing a product is too easy. Spot a use case, and a demo can be ready in a few days. If few people use it or model costs seem too high, we quickly switch to the next direction.
We are instead rarely willing to immerse ourselves in an industry as before and thoroughly understand users' needs.
After hearing Wu's Kaipai example, I wanted to slap myself. I really have not done well on this over the past year. Combining what the podcast shared, I summarized several ideas for small teams building AI applications:
1) Go deep into a field the team knows well or is interested in, and find pain points involving essential needs within it.
2) Stay firmly away from scenarios where ChatGPT + Skill can handle 70%.
3) Find scenarios where ChatGPT can solve only 30% but users have strong needs.
4) Go deep and build barriers through industry know-how.
#02
Have the courage to give up and cut losses in time
This is another point I particularly admire.
The startup stories we hear tend to be about someone betting on something, overcoming every obstacle and finally making it succeed.
But most startups' experiences may be exactly the opposite.
Take Meitu.
After ChatGPT launched, everyone saw the opportunity in generative AI. Meitu already made image and video applications, had plenty of users and years of industry know-how.
Why not take another step and build its own models?
In fact, Meitu had begun applying generative image models in products even before ChatGPT became popular.
It later began developing its own Qixiang large model, starting with images, then investing in large video models in 2023 and 2024.
After all, everyone knows video models may offer a bigger commercial opportunity than image models.
But as they worked on it, they found several large companies had overwhelming resource advantages in many respects. Their talent pools, compute and data were simply on a different scale.
This was not a situation where persistence alone would bring victory. It is like warfare: we often hear stories of the few defeating the many, but in 99.9% of wars throughout history, the many have defeated the few.
Wu judged that video foundation models would very likely end up concentrated among a handful of giants with scale advantages.
If Meitu had only one-tenth of its competitors' budget, or even less, yet tried to do the same thing, its chances of winning would be far too low.
The prize is certainly lucrative, but if the team cannot win it, there is little benefit in continuing to struggle on.
After weighing this up, they ultimately decided to scale back research and development of video foundation models and refocus on the image models they were better at.
As an outsider, I think this decision was 100% correct. Looking now at the video-model battleground in China, it really is essentially ByteDance, Kuaishou and MiniMax.
Recognizing that you cannot win and decisively giving up after investing substantial resources seems harder to me than betting on the right direction. After all, people have egos.......
Wu said in the podcast that a company's goals should match its capabilities.
This sounds simple, but anyone who has actually started a business should know that sometimes the hardest thing is admitting you cannot make it work. Especially after spending so much money and having the team work for so long—who wants to give up easily?
#03
Develop models in-house for core use cases
After tightening its video-model efforts, Meitu concentrated more R&D resources on the image models it had been working on all along.
I have been thinking about this too. Look at Midjourney: it has built a good image model, and that market is not small. It may not offer as much room as video models, but there is still considerable opportunity.
In fact, investing effort and resources in image models has delivered substantial benefits.
About 96.5% of the capabilities in Meitu's relevant products are currently provided by its in-house models; only a small proportion actually calls external APIs.
According to their tests, achieving equivalent results with in-house models costs only one-twentieth as much as with external commercial models.
In-house models can also be optimized for specific use cases to differentiate the product from competitors.
For portrait retouching, for example, Meitu can conduct targeted training and optimization for specific needs such as portrait beautification, lighting adjustments and image-quality restoration.
After all, if everyone calls the same external model and it can achieve only 70 points, everyone may also be limited to 70.
Developing models in-house, however, offers a chance to bring particular effects to 80 or 90 points while also saving costs.
My takeaways here are:
1) For core use cases, in-house models should still be considered (post-training on top of open-source models).
2) In-house models can reduce costs.
3) Targeted post-training can also improve in-house models' performance on core tasks.
Of course, I do not think this is necessary for a startup just getting going. Large-model work is more something to consider after finding product–market fit (PMF).
#04
An AI-native organization
AI-native has been widely discussed this year too. There is already considerable agreement about what an AI-native organization is and about breaking down traditional job divisions so everyone becomes a super-individual, so I will not explain those again.
Starting in early 2025, Meitu drew on the experience of game companies such as Supercell to set up internal AI innovation studios.
There are now more than ten studios. Each team is kept below ten people as far as possible and handles the entire process from project approval to PMF validation. Some teams even have just two or three people.
Studio leaders resemble producers in the games industry. They have considerable decision-making authority and can independently decide how to build products and allocate resources.
They have also borrowed incentive mechanisms from the games industry.
For example, studios with approved projects receive startup funding and allocate their own staffing and promotional resources. If the product proves viable, core members can also receive a corresponding share of profits.
The overall idea is to let these small teams operate as independently as possible, developing and validating products like startups.
How does this work specifically?
First, they changed their previous product-approval process.
Previously, developing a product usually required extensive internal analysis, repeated meetings and discussions, and all kinds of documents and plans before action could begin, regardless of the company's size.
Now, however, AI has made building a demo extremely cheap.
Wu gave an example.
A Meitu product manager who previously could not code at all might now, after having a product idea, build the entire demo in three days with AI coding and run it personally first to see how it works.
They therefore now prefer building the product first and then discussing whether it merits formal approval.
This does not mean the company will fund and approve any demo someone casually makes, though. There are several basic requirements:
1. The product must fit the company's business direction; it cannot pursue every opportunity it sees.
2. Reuse the company's accumulated technical capabilities as much as possible. Meitu has spent more than a decade working on imaging, after all; there is no need to start from scratch with every new product.
3. The use case must be explored deeply enough; it cannot just be a simple AI feature.
The company uses these conditions and the actual demo to decide whether to approve the project formally.
After approval, it does not commit substantial resources all at once, but validates in stages. If the product performs well, investment increases. If the direction proves wrong, it stops promptly.
Of course, with the opportunity to build a startup comes the risk of entrepreneurship.
If a project performs poorly, it is stopped promptly, without dragging things out. The company cannot let a project without potential keep consuming resources indefinitely.
After a project stops, team members may return to their previous roles or join other projects and start again.
#05
Business models for AI applications
Let us also discuss how AI applications make money.
This is something giving me particular headaches right now. The business model for utility products used to be simple: users paid a few dozen yuan a month for a membership and could use all kinds of premium features.
That model does not necessarily work for AI applications now.
Traditional software has very low marginal costs. A few extra uses by a user barely add to costs. AI is different: every generation consumes compute, and tasks involving images or video can be especially expensive.
If users pay only a few dozen yuan a month but consume a hundred yuan or more in tokens, would the company not lose more as it does more business?
Wu's approach is to retain a basic subscription while adding per-service fees.
Basic features remain available through membership subscriptions. Needs that consume considerable compute or directly complete a job for the user can be charged for separately.
Put plainly, package AI capabilities as services users understand and directly deliver results for them.
This also made me realize how narrow my thinking about AI product business models had been. I kept considering how much users would pay monthly, whether token costs could be covered and what gross margin we could achieve.
From another angle, however, if a job previously cost users thousands of yuan and several days, and AI can now do it directly at a lower cost, the price users are willing to pay may be different.
This returns to the first point, though: the core is still to go deep into vertical use cases, because a shallow application layer makes it difficult to deliver results directly.
#06
Practical experience in going global
Everything Wu shared in this part came from hands-on experience. I will summarize it directly as a list.
1) AI productivity applications should target global markets from day one. Many internet companies used to build products in China first and expand abroad only after validating success.
Wu believes global users can now be considered from the first day of project approval. This is especially true for productivity tools, since work needs across countries have much in common.
For example, cross-border e-commerce sellers have many similar workflows and needs whether in China or the US.
2) Lifestyle and productivity products face different globalization challenges. Lifestyle products are strongly influenced by culture, aesthetics and social trends. East Asian users, for example, may like skin lightening, while European and American users may not share that aesthetic.
Demand for productivity tools, however, is usually more similar. Such products more easily find common needs among global users and are suited to global development from the outset.
3) Going global cannot simply mean renaming a domestic product and translating its interface. Meitu tried this early on and found it did not work well.
It therefore developed AirBrush specifically for European and American users, with an independent team redesigning everything from the interface and interactions to individual features around local preferences. It has now become one of Meitu's main overseas products.
4) Overseas markets should be divided by cultural groupings, not geography alone. Thailand, Indonesia and Malaysia, for instance, are neighbors, but users' aesthetic habits differ greatly.
When BeautyCam introduced AI outfit changing, Thai users tended to prefer bold clothing styles, while some users in Indonesia and Malaysia preferred more conservative dress. Meitu therefore added clothing and headscarves suited to local cultural practices in those markets.
Cultural differences like these need to appear in specific product features.
5) Do not blindly pursue overseas user growth; ability to pay and long-term retention matter more. Meitu fell into this trap between 2016 and 2018.
At the time, pursuing monthly active user growth, it spent heavily on paid acquisition in markets with weaker ability to pay. Retention and monetization both disappointed.
Its user base in India was once second only to China's. Later, local policy restrictions caused heavy losses on the earlier investment.
The internet industry used to favor building scale first and thinking about revenue later, but we should now be more rational about that logic.
6) Use different markets to validate business models, then extend successful experience to others. Meitu first validated subscriptions overseas, then brought the model back to China and found domestic users' willingness to pay was rising too.
Meitu still prioritizes relatively small markets to test new monetization strategies. China's user base is too large: rolling an immature strategy out at scale could trigger substantial negative feedback.
7) Do not overlook overseas markets growing strongly on their own despite receiving little targeted investment.
Meitu initially put few resources into Brazil, for example. But because local users love sharing content on social platforms, Brazil has become the second-largest, third-largest or even largest source of users for some products.
Meitu established an operations center in São Paulo specifically to further develop Brazil and Latin America. It now places great importance on finding new market opportunities in organic-growth data.
8) Overseas markets need local teams, but R&D need not move abroad. Meitu has offices in places such as Silicon Valley, London, Sydney and Singapore, and local operations centers in markets including Brazil. Operations staff are mainly local.
R&D teams remain primarily in China, however. Chinese engineers' professional capabilities and efficiency are very competitive, while labor costs in Silicon Valley are much higher. There is no need to move the entire R&D team abroad simply for globalization.
9) To globalize, founders and core teams need to visit overseas markets themselves to understand them. Over the past two or three years, Wu has worked abroad for more than a hundred days each year.
Besides talking with local teams, he visits startups, attends industry salons and speaks with local researchers.
His reasoning is that technology and markets change too quickly in the AI era. Core teams need firsthand knowledge of what is happening globally, rather than relying only on secondhand information obtained in China.
That is enough of a summary for now.
This podcast episode really contains so many valuable insights. It is rare to hear such practical experience in building AI application startups, with far more substance than much of what today's internet-celebrity entrepreneurs share.
Next month, before formally finalizing our company's 2027 plan, I will definitely listen to the entire podcast again.
This article comes from the "AI Product Aying" WeChat account. Author: "AI Product Aying."
Related topics
After reading
Join the AINRK community
Read how to join the community and connect with people working in AI.




