Wu Xinhong Interview: AI, Globalization and Competing with Giants
Summary
- Meitu’s AI application thesis rests on depth in vertical workflows. Wu Xinhong sees cropping, filters, basic retouching and one-line generation becoming table stakes; industry knowledge, batch processing, team collaboration and final delivery are what will keep applications relevant. Kaipai grew out of a teleprompter that ranked third among paid features in BeautyCam, showing how users can vote with their wallets before a product expands into industry services.
- Proprietary image models paired with a container for external models give Meitu a path to differentiation and cost control. Wu says in-house models now handle about 96.5% of demand and cost just 1/20 as much as commercial models for comparable results; R&D spending was 25% of revenue last year. Meitu exited the capital-intensive video-model race but kept investing in image models: if everyone calls the same 70-point model, it is hard to deliver an 80- or 90-point result on your own.
- AI’s organizational impact is already showing up in development workflows and team size. Product managers with no coding skills can build a demo in 3 days, while Pitch’s roughly 20-person team is delivering capabilities that might once have required 200 people—or more. Since early 2025, Meitu has borrowed from the game-studio model to launch more than a dozen AI innovation studios, with early validation teams kept to 10 people or fewer; Wu’s own meeting load is down to about 1/3 of what it was.
- Meitu wants to layer outcome-based services on top of subscriptions to unlock more pricing flexibility. Wu calls it “not so much a shift as an addition”: basic subscriptions remain, while services that consume substantial compute and directly help businesses operate will be priced separately. He sees “tens or hundreds of times” the pricing flexibility of subscriptions—not revenue multiples already realized. The current balance between enhancing mature products and incubating new ones is roughly 8:2.
- Globalization has been validated by users and paying customers, but Wu is cautious about setting an overseas revenue target. Overseas users account for more than 42%, and VIP members make up over 53% of AirBrush’s U.S. monthly active users. Wu sees overseas users and revenue each reaching 51% as a sign of successful globalization; he says overseas revenue could reach that level “in 2 years,” while stressing that growth should take its natural course and China remains a market Meitu must keep investing in.
- The yardstick for overseas growth has shifted from buying monthly active users to monetization, retention and local culture. From 2016 to 2018, Meitu spent heavily in markets with weak monetization; India briefly became its second-largest user market after China before policy changes damaged the investment. The current approach segments markets by cultural affinity, builds local operating hubs and redesigns products around different aesthetic and religious preferences, rather than simply relaunching under a new brand.
- The lessons of post-IPO expansion have made Wu more focused on matching ambition to capability and building a sustainable business. Meitu grew from around 1,000 employees at the end of 2016 to more than 3,000 in 2018, as both the business and organization began to slip out of control; Wu admits his mindset then was “I wanted it all.” Against giants today, he is willing to say “we can’t beat them,” opting for a specialized, “small and beautiful” business rather than betting everything on a low-probability win.
- Wu expects applications and models to coexist, but is revisiting his old belief in economies of scale. He admits that the possibility of models moving into Meitu’s territory remains unsettling, and says the only answer to skeptics is “results.” Even if existing products were eventually swallowed, he would keep building and adjust the boundaries as long as there were a few more years to work. The bigger unknown is whether super-individuals working with Agents will let small teams do the work of thousands; he welcomes that possibility and accepts that Meitu may have to disrupt itself first.
Deep dive
1. Meitu puts AI investment and profit delivery on the same scorecard
Li Xiang’s two impressions of Meitu are that it is “restless” and resilient: the company has tried hot businesses such as smartphones and short video, weathered downturns, and remained in the game since its founding in 2008. He opened by citing first-half results—revenue growth of more than 20% and net profit growth of nearly 40%—and asking how a tools company can benefit from AI.
Wu Xinhong’s answer starts with a distinction in positioning: “Models and applications occupy different niches.” Models and applications reinforce each other, just as Microsoft did not wipe out every software company in the PC era and mobile internet giants did not capture every opportunity. As chips, models and other parts of the stack mature, applications still have ample room to find use cases and business models.
Meitu’s AI investment predates ChatGPT. Wu says the company set up MT Lab in 2010 to research computer vision; R&D spending was 25% of revenue last year, correcting Li’s recollection of 20%. He wants to sustain a high level of R&D while continuing to deliver profits, pointing to revenue and profit growth over the past 6–7 years as evidence that Meitu is benefiting from AI.
2. Shallow generation features become table stakes; deep industry workflows matter
Wu groups cropping, basic filters, retouching and “fix a photo with one sentence, design something with one sentence” among shallow capabilities any model or assistant could offer. Image and video generation are also basic capabilities: they “can’t build any product differentiation” on their own. Adding generation does not mean a product has a moat.
Li compared the risk to mobile internet applications that operating systems could easily absorb. Wu clarified that he means shallow use cases where a prompt produces an output. It is different when a product enters a complete industry workflow involving extensive batch processing and collaboration among multiple people.
Kaipai is the example he returns to. Behind talking-head videos are industry-specific needs, know-how, skills and relatively vertical model training; a general-purpose model may not be able to handle them directly. Li added that model companies might be able to build these capabilities, but going deep into each industry could be inefficient for them. Wu agreed.
3. Model boundaries keep expanding, while past hit-making can become a habit
Since November 2022, Wu says, his understanding of models has “kept getting challenged,” and at a high frequency. Competition and rapid iteration keep pushing capability boundaries outward; some application developers have even become a way for models to test markets. Meitu therefore sees models as intelligent brains while staying alert to their potential entry into its own use cases.
In imaging, the latest development to astonish him was this year’s Seedance 2.0: video models can now produce footage that looks live-action, making it hard to tell real from generated. He summed it up as “launch and it’s live,” and noted the rapid growth in AI short dramas, animated dramas and related video applications over the past few months.
Meitu has been good at helping users record and share their lives, and at creating popular formats, but those features can be used once and quickly replaced by the next trend. Asked whether that means it should be wary of viral hits, Wu put it more precisely: the risk is “getting path-dependent on how we made hits in the past,” and carrying a toy-like mindset straight into productivity products.
4. Kaipai became a standalone product after users voted with their wallets
Kaipai started as a teleprompter in BeautyCam. The team found it ranked third among all paid features, with many users willing to upgrade their memberships for it. Interviews revealed that what they really wanted was to make talking-head videos for marketing and livestream commerce. Li called it a standalone product that users had voted for “with their wallets,” and Wu agreed.
After the product stood on its own, demand continued to come from real industries: real-estate agents, insurance brokers, paid-content creators and local tourism operators all needed video marketing. Wu stresses that the team did not understand these industries from the outset; users’ suggestions drove iteration. “The users really taught us.”
5. Serving industry users means rebuilding offline growth capabilities
To reach these users, Kaipai began attending industry events such as insurance trade shows and trying direct, on-the-ground sales. Wu acknowledges that this is very different from Meitu’s past growth playbook for consumer lifestyle apps, which involved little offline promotion and limited engagement with specific industries. “This is quite difficult for Meitu.”
The team is working with agency partners and has also tried turning Kaipai into vocational training, with schools offering courses. Wu says feedback from schools and students has been very positive. Once the product enters an industry, growth also means building training and user proficiency.
Li compared weather apps with Uber and Didi: an operating system can provide weather by default, but it generally will not persuade drivers one by one to install and use a ride-hailing app—yet the latter can become a huge business. Wu agreed with the analogy. “It’s challenging, but we enjoy it,” he said of entering unfamiliar work. “It feels like opening up a new world.”
6. Two shocks in 2022 put generative AI at the center of strategy
In November 2022, Meitu Xiuxiu’s AI art feature went viral globally, reaching No. 1 overall in the app stores of nearly 20 countries. Wu says it was the company’s first use of a diffusion model in a product, arriving even 1–2 weeks before ChatGPT’s launch. User feedback gave the team an immediate view of how quickly the new capability could spread.
Li initially thought the product did not use generative large-model technology. Wu explained that it was built on the open-source image-generation model Stable Diffusion, or SD, with Meitu’s own training and productization layered on top. “It was generative AI too”; the difference is that one is primarily for images, the other for language.
After using ChatGPT, Wu felt “a new era had arrived,” with an impact “much like the iPhone launch at the end of 2007.” Conversations that had previously felt clumsy were now approaching human-to-human interaction. More important was the advanced intelligence behind them, which could serve as the brain for many products.
Meitu quickly revisited its strategy and organization and decided to invest heavily in generative AI. Asked whether it saw an opportunity or a threat, Wu said, “All we ever see are opportunities.” His focus was on the many vertical use cases that had not been claimed, where the new brain could help build products again from the ground up.
7. The application opportunity is here, but global competition is shortening the window
Wu places the previous mobile internet boom roughly between 2011 and 2015; Meitu launched its iPhone app in February 2011. Back then, simply claiming a use case early was enough. Users tolerated missing features and bugs that did not affect basic use, and competition was far less intense than it is today.
He believes the generative AI application boom has already begun: users and ARR at many AI-native products are growing explosively, and some global applications “are genuinely profitable.” But the application era has yet to fully unfold, while companies around the world are afraid of missing the entry point. That is compressing the window.
The claim that Meitu had entered a “life-or-death” moment once circulated publicly. Wu explained that the company was then focused on building generative AI; its chief product officer, Xiaobai, turned down a partnership visit by saying the company was “too busy keeping itself afloat.” He meant the team was racing to seize the opportunity, while revenue and profit were still growing rapidly.
Li asked where the boundary between applications and models lies. Wu gave a clear answer, with a condition: suppose most of Meitu’s current products will be swallowed up, but the company can still operate for a few more years—“why not keep doing it?” He says it is hard to predict the path, so the choice is to build first and “deal with whatever comes,” adjusting the boundary along the way.
8. Code that gets you to 50 or 60 points cannot replicate 90-point imaging
Meitu Xiuxiu has been around for 18 years, and competition is nothing new. Big tech companies and phone makers have built camera and editing products, and now AI is in the mix too. Wu says the company has to keep adapting; if it does nothing, users may drift away even if competitors never launch an active assault.
On the claim that “AI can copy any product,” Wu says it is “true in theory.” But copying the UI, interaction and features may get you only to 50 or 60 points. For users to feel the results are excellent and the experience is polished across the board, the product often needs to reach 90 or more. That requires substantial engineering and investment in image quality; writing the code does not replicate the accumulated work.
That effort also creates scale advantages for Meitu: the company has more than 2,000 employees, many of whom have spent years working on imaging. Outsiders may see a collection of filters and effects and miss the passion required to understand imaging and keep refining results. “Imaging really has to be driven by passion.”
9. First and second attributes give the classic product manager room to make a call
By a person’s “first attribute,” Wu means their core job and professional expertise—for example, an experienced cross-border e-commerce seller understanding sellers’ real needs. Their “second attribute” is a passion or skill outside work, such as photography or music. That experience can also become product insight. Wu himself brought a love of photography and familiarity with editing software to product building.
He is less convinced by a style of product work that emerged in the later mobile internet era: once the market structure was established, product managers focused more on strategy and A/B tests, building a huge funnel to sort through outcomes. Someone with a strong second attribute may know before testing whether a need is real, instead of handing every judgment over to the data.
Li recalled one explanation for Apple’s iPod: Steve Jobs loved music. Amazon’s Kindle, in turn, reflected Jeff Bezos’s love of reading. Wu agreed with the idea of bringing a second attribute into work. Meitu’s values include “a love of imaging”; he believes people are more likely to do something well over the long term when they genuinely love it.
When hiring designers and people responsible for image quality, Meitu looks at portfolios and photo collections, and uses conversations to spot genuine enthusiasm: “People who really love it have a light in their eyes.” As AI lowers the cost of execution, intuition, individual strengths and judgment matter again. Wu calls the shift “a renaissance for the classic product manager.”
10. AI-native products complement the existing Meitu Xiuxiu; disruption is best done in-house
Peachy can remember a user’s editing preferences, learn a particular KOL’s retouching style and then batch-process a set of photos automatically. Wu says the person in charge is a photography enthusiast who also understands firsthand how women want their photos edited—an example of a second attribute leading to sharper product insight.
On what an AI-era product like Meitu Xiuxiu might look like, Wu said the company had already wanted to build an AI-native product from scratch, with less baggage from the old framework. When Li asked what would happen to the existing product, Wu corrected his wording: a more accurate description is “a product like Meitu Xiuxiu, born in the AI era.”
The original Meitu Xiuxiu is also evolving. Wu says more than 97% of its features are AI-driven, even though they still look like familiar features to users. The envisioned new product starts from an Agent framework, while mature products keep increasing their AI content. At this stage, the two can complement each other.
Meitu therefore maintains internal competition between products and teams. If someone in the market is going to build an AI-era Meitu Xiuxiu, the company should take its own shot. Wu’s preference is for disruptive products to “emerge from inside the company rather than outside it.” Li compared this to the iPhone displacing the iPod; Wu agreed.
11. Meitu exited video models because the investment did not fit, but kept building image models
Meitu’s proprietary model work started with images. In 2023 and 2024 it also invested in video models, but quickly realized the bar was too high. The cost was not just algorithms and engineering teams; it also involved organizing, cleaning and labeling huge datasets, as well as scarce compute. Wu thinks video foundation models may ultimately be concentrated in the hands of a few giants.
With insufficient investment and scale advantages, Meitu pulled back to image models. Asked whether building its own models had been the right call, Wu first clarified that the company gave up on video models; it never abandoned image models, which it continues to iterate on. The pullback should not be read as a wholesale exit from proprietary models.
Wu says proprietary models currently handle about 96.5% of demand, with external APIs making up only a small share. Meitu can train its models for specific product needs and solve problems external models may not address. For comparable results, he says its costs are only 1/20 those of commercial external models.
Meitu never set out to build a general-purpose large language model; that was never in the company’s DNA. Wu says the company has a relatively objective view of its own capabilities and generally avoids setting goals far beyond them. Vertical image models are where it sees a competitive edge worth investing in.
12. A model container takes the best of each provider; proprietary models create separation
“Rather than build a large model, build a model container that best understands vertical use cases” is a phrase Li cited from Meitu. Wu says the idea took shape in 2024 and became clearer in 2025. Once the team began integrating different models, it found none was best at everything: for different outputs within the same product, Model A or Model C might be the strongest.
Meitu began calling external models in 2023 and 2024, putting them into the same competitive process as its own models and continually comparing quality and cost. But when Li asked why the company could not become a container for external capabilities alone, Wu replied that if everyone uses the same 70-point model, their results are likely to converge at the same level.
To bring certain use cases to 80 or 90 points, Wu believes proprietary models are still needed for differentiation, which is why Meitu is willing to keep spending heavily on R&D. He does not recommend the same path to every Agent company: “We generally don’t advise others to do this.” Every company is different, and understanding the business takes a lot of time. The priority is to get your own business right.
13. Coexisting with giants means surviving in the gaps; results will settle the debate
“Will models swallow applications?” is one of the questions Wu gets most often. He admits that explanations are unlikely to persuade people: “The only answer is results.” The market treats the concern as consensus, which may pressure listed application companies but could also create a contrarian opportunity. Over time, the fear may prove excessive.
Li asked whether Meitu’s ability to survive is exceptional or generalizable. Wu offered no blanket guarantee. Each model company has a main lane, such as coding or video. Applications can try to avoid those lanes and go deep into industries that large companies overlook. “Surviving in the cracks” may sound humbling, but it can work.
On an era in which researchers rule, Wu draws a clear line around Meitu’s remit. He estimates there are fewer than 10,000 top researchers globally, concentrated mainly at leading model companies. Meitu will not compete to build new learning frameworks; its job is to understand where users are, what they need, and whether the company can deliver outcomes that meet or exceed expectations.
“At our core, we deliver. We provide a service.” Wu compares applications to restaurants: spice makers and ingredient suppliers cannot do everything. Agents may replace expensive, long-running processes that once required coordination across multiple departments and companies, but delivering the final outcome remains a separate business that needs to be run well.
14. Building a demo in 3 days has changed the order of product development
In the past, launching a new product meant making the case, getting sign-off up the chain and then allocating development resources. Now, a product manager who could not previously code might use AI to build a complete demo within 3 days of coming up with an idea, then validate it independently. “Build first, get it running first,” and only then discuss whether to greenlight it—that is the workflow shift Wu sees.
Pitch is one example; its team currently has about 20 people. Before generative AI took off, building the full Agent capability on the backend might have required a team several times its current size. “Even 200 people might not have been able to build it.” That is a judgment about how difficult it would have been to build in the past, not a claim that 200 people have since been cut.
Internal product competitions have surprised Wu too: product managers, designers and even colleagues in functional roles can use AI coding tools to build credible products. People who once handled only one piece of a finely divided process are becoming generalists. Their competition increasingly includes AI-native teams with very few people.
15. Creator empathy and AI-native products offer two paths for product development
Meitu wants to build a creator culture because people serving creators need to understand firsthand how content is made. Kaipai serves talking-head video creators; MV Land serves music-video creators. The team has to try making talking-head videos, music videos and short dramas—and work to make them well—to understand users’ needs in the creative process.
Wu describes himself as introverted and not necessarily suited to recording talking-head videos every day, but he is passionate about music creation and video production. When he used to learn painting, he approached it mainly as an art student. Now he creates with a commercial mindset and a product portfolio in mind, accepting that he can “start out clumsy and get better over time.”
On what users feel, Wu says the biggest change is the impact of intelligence; changes in interaction come second. He describes the evolution of interfaces as moving from keyboard and mouse, to touchscreens and sensors, to text and voice, and perhaps eventually to immersive interaction. The more powerful shift is that a “super-smart brain” is now available to serve all kinds of ideas and needs.
Li summed up the main thread as adding intelligence to mobile internet products. Wu added that there are also AI-native “new species” that could not have been built before. Meitu is working on both: about 80% of its effort goes toward adding intelligence to mature products, with the rest used to incubate new products. The first is the core business; the second challenges the company.
16. Layering outcome-based services on subscriptions creates more pricing flexibility
Asked whether the business is shifting from subscriptions to token consumption, Wu first corrected the premise: “It’s not so much a shift as an addition.” Subscriptions provide a baseline service, but some use cases consume substantial compute. If every user pays only the same basic subscription fee, the company may not be able to cover those costs.
He sees the opportunity in industry users’ willingness to pay for delivered outcomes: they need a service and will pay more for value. Meitu wants to build out this service line. Compared with subscriptions, future pricing could have “tens or hundreds of times the flexibility.” That is a capability the company hopes to establish, not a pricing result already achieved.
Li worried that tokens might be a fashionable label for packaging a business model. Wu brought the issue back to whether customers understand it: developers and coding users are familiar with tokens; owners of small eateries and general stores may not be. The product has to turn its capabilities into services that genuinely help a business grow before customers have a reason to pay.
17. More than a dozen AI studios use small teams to validate deep use cases
Starting in early 2025, Meitu gradually borrowed from the game-studio model to set up AI innovation studios and give producers significant autonomy. Wu pointed to Supercell: small studios, flat structures and versatile teams suit AI applications. The mature incentive, selection and exit mechanisms used in gaming are also worth learning from.
Meitu now has more than a dozen such studios. From project approval through PMF validation, teams are kept to 10 people or fewer where possible; some products can be built well by teams of just 2 or 3. Wu believes AI applications are becoming more commercially valuable and easier to monetize quickly, making this kind of incentive structure increasingly viable.
Projects are first assessed for strategic fit and whether they can reuse technical capabilities the company has already built; next comes the depth of the use case. For creators, Meitu often starts with a content format such as talking-head videos or music videos. Teams also have to produce a demo: “You can’t just show up with a business plan.”
Projects are assessed in stages, with decisions to expand validation or stop. After a project stops, some people return to their previous roles, some join other teams and some leave the company; their old positions may no longer exist. Wu acknowledges that participants take on real risk. This is a genuine form of internal entrepreneurship, not an experiment that guarantees them a place to return to.
18. Agents are taking over internal workflows, and Wu’s meeting load is down to one-third
The first step in organizational change was getting everyone to use AI deeply. Meitu gives employees generous token budgets and covers AI product subscriptions, because building AI applications requires firsthand experience. Casual experimentation is not enough to understand a product’s capabilities or rethink how to do the work.
The changes are already reaching translation, testing, operations and user acquisition. Multilingual releases that once required a large translation team can now be adapted quickly into dozens of language packs using translation Agents. The company is also testing AI that simulates real users to find product issues, while operations and paid user acquisition gradually become Agent-driven.
The change is not just about reducing headcount. Wu stresses that people in translation, testing and operations are directly involved in creating and running Agents, and are becoming AI product managers themselves. Problems in the business then feed back into organizational changes: fewer layers, emails and meetings, and more decisions returned to the people doing the work.
Wu’s own meeting load is now around 1/3 of what it used to be, while his time using products has increased. Building a creator culture means using the company’s own products as well as the best products from around the world. He says he still spends slightly more time on products than on organizational management.
19. Super-individuals are changing the talent model and forcing organizations to accept centrifugal forces
Wu roughly divides employees into 2 groups: people who quickly embrace AI and keep developing, and people who use it superficially and lack a sense of urgency or growth. A designer taking on engineering, translation, testing and operations would once have been hard to imagine; now it is becoming possible. He wants the organization to keep producing these “super-individuals.”
Li’s challenge was that the more capable these people become, the less dependent they are on the organization—and the greater the centrifugal force, with more potential to disrupt existing rules and management systems. Wu answered directly: “That’s inevitable.” The organization still has to embrace the possibility, or it will be pulled back into the old model. An environment that keeps producing high performers is itself a sign that the soil is fertile.
Wu offers a simple test for being AI-native: if the model interface goes down, the organization or application becomes paralyzed. He also acknowledges that AI-native organizations may look different across industries and use cases, making it hard to prescribe a single form. If the focus is on outcomes, there is no need to require teams to fit a standard profile in advance.
He points to Meitu’s early grassroots team as an example: no elite universities, overseas study, big-tech experience or much capital, yet the company got started. Li compared it to a player who shoots well enough not needing to worry about their form. Wu added that they did not need to worry about their height either.
20. Overseas users exceed 42%; subscriptions first proved their value in the West
Meitu launched BeautyPlus for overseas markets in 2013, and a hand-drawn selfie feature went viral in several countries in 2016, prompting continued investment in overseas products and operations. Overseas users now account for more than 42% and the business is still growing quickly. Wu says that after globalization took its first step, user feedback and new opportunities made it hard to “stop.”
The commercial turning point came in 2018, when AirBrush, aimed at Western markets, first tested subscriptions. The results in 2018 and 2019 were successful and generated real profits. The team saw that global users recognized the value of the application and were willing to pay, strengthening the case for further investment.
AirBrush’s subscription penetration in the U.S. is now above 53%. After Li asked for clarification, Wu specified that this is the share of U.S. monthly active users who are VIP members. He called it the highest level of subscription penetration; many other countries are still a long way from it.
Wu sees overseas users and revenue each reaching 51% as a benchmark for successful globalization. He says overseas revenue could reach that level “in 2 years,” but Meitu is not insisting that it must exceed a particular percentage. China is a large market in its own right and should remain an important part of the business; the mix can take its natural course.
21. Rebranding alone does not work; products have to understand local culture
AirBrush can be thought of as Meitu Xiuxiu for Western markets, but its UI, interactions and feature preferences are distinctly different. The two products may not look like they come from the same company, and they have separate product teams. Wu admits that Meitu initially tried keeping the core product unchanged and simply giving it an overseas brand, but found that “reskinning” did not work well.
Lifestyle use cases are especially dependent on local aesthetics and social trends. Preferences differ widely between East and Southeast Asia and the West. For example, Asian users may seek whiter skin, while Western users may not see that as the right outcome. Productivity applications are easier to build around shared needs; some requirements of global e-commerce sellers are very similar.
Meitu therefore looks at markets through cultural groupings: Latin America, where Portuguese and Spanish are widely spoken; culturally similar English-speaking countries; and Southeast Asian markets that influence one another. Wu stresses that imaging products are deeply shaped by culture. Markets cannot be divided only by geography, with the assumption that the same content will work everywhere.
AI try-on is one example. Thai users may be more adventurous in their clothing choices, while neighboring Indonesia and Malaysia have large Muslim populations whose religious preferences need to be respected, including more conservative clothing and hijab options. The same feature requires different content in different markets.
22. China is the foundation, and globalization starts in product design from day one
Asked whether positive momentum overseas meant negative momentum at home, Wu said no. He sees China as the world’s largest single-language market, with improving willingness to pay. China remains the foundation even as Meitu expands overseas: “We have to do well here, no matter what.”
Overseas markets also help test business models. Subscriptions first succeeded overseas, then reached China in 2019; Chinese users’ willingness to pay has increased year by year since. Other new approaches are often tested first in smaller overseas markets, rather than rolled out across a domestic product that already has high market share.
The two types of market also differ in practice. Wu observes that Chinese users are more accustomed to downloading mobile apps, while PC usage remains high overseas, especially for productivity products that often operate as websites. Globalization requires choosing the right product format, not just translating the interface.
Wu now uses globalization to mean including China in the global product from the outset: build for the world from day one rather than spending a long time validating in a single market before expanding abroad. To identify common needs, product and R&D teams have to get deep into different markets and build real judgment instead of making assumptions about “overseas users.”
23. More than 100 days a year overseas makes aligning perspectives part of the job
Wu spends more than 100 days a year working overseas, mostly across 2024–2026. He tends to work from a few cities where Meitu already has teams, gradually establishing a regular rotation. It reflects how much more important globalization has become in how he allocates his own time.
Along with user research, those trips include roundtables, company visits and conversations with entrepreneurs and researchers. He repeatedly emphasizes “aligning perspectives”: AI is moving so fast that face-to-face exchange is valuable to the business. And it cannot be just him; key team members need similar firsthand exposure.
Meitu’s main competitors globally are from China and the U.S., with more companies from Singapore appearing as well, many founded by people of Chinese descent. Wu says competition is fiercest in China and the U.S.; there are many AI application companies in the U.S. too. Going global does not mean leaving intense competition behind.
Meitu studies the global product and operating strategies of ByteDance, Canva and video-creation company Higgsfield. It competes with ByteDance while also studying it, and some former ByteDance employees now work at Meitu. Wu is happy to “learn from successful companies,” but says he cannot go into detail about what the company has adopted.
24. Beyond Asia, organic growth in Brazil is bringing local investment
Meitu’s strongest markets remain in East and Southeast Asia. Wu cites Thailand, where Meitu Xiuxiu has more than 40% share and is the clear leader in its category. It has become a familiar mainstream product, close to a must-have app. Cultural similarities help Meitu understand these markets.
Brazil is a standout market despite its greater cultural distance. Wu points to its population of more than 200 million and active social sharing. Even without much investment in some products, Brazil became the second- or third-largest source of users; RoboNeo already has more users in Brazil than in any other market.
Meitu has therefore set up an operating center in São Paulo, staffed mainly by local Brazilians, to drive further localization of products and marketing. Wu calls it “going beyond Asia to discover a new continent”: first study markets that are unfamiliar but already performing well organically, then build the capabilities to serve them specifically.
Europe is still below his expectations, mainly because the team does not know the market well enough, languages are diverse and service capabilities are not yet adequate—not because of competition. Africa has also grown organically, but will need a dedicated team. Wu wants to move gradually from organic growth after launch to active management by regional operating centers.
25. Global resource allocation still centers on Chinese R&D and internal incubation
Meitu has offices in Silicon Valley, Sydney, London and Singapore, but its R&D hub remains in China. Wu values the expertise and efficiency of Chinese engineers. There are also R&D colleagues in Silicon Valley, but not many: R&D requires a certain scale, and a few people working alone may not be efficient. Labor costs are also much higher locally.
When it comes to R&D spending, the biggest costs remain people and compute. Wu has not spread R&D resources evenly around the world just because Meitu is a global company. He considers team collaboration, investment scale and whether those capabilities can keep supporting products in the same imaging category.
He remains cautious about investing in global imaging teams. A team taking investment may expect product and technical support that Meitu may not be able to deliver after the deal. Without control, the investment risks becoming purely financial. If the goal is purely financial, there may be no need to invest in imaging, so Meitu is currently focused on internal incubation.
With smartphones and Meipai no longer being pushed forward, the business model has become simpler and resources have shifted from being spread thin to being concentrated. At the same time, more of the team is going overseas to interview users, connect with KOLs and meet companies, replacing guesses about users with judgments grounded in experience.
26. User acquisition from 2016 to 2018 showed that MAU cannot substitute for monetization
Wu acknowledges that Meitu took a wrong turn in its globalization from 2016 to 2018: it focused too narrowly on MAU and spent heavily in countries with limited ability to pay or poor monetization. “Burning real money” did not produce good monetization. Paid users could churn easily, and long-term retention was not necessarily good.
India was once Meitu’s second-largest user market after China. Wu recalls that in 2020, the Indian government removed all Chinese apps and games from local app stores; they have not returned. Meitu’s substantial investment in the market was therefore damaged. The experience involved both user-quality issues and uncertainty in the market itself.
At the time, many mobile internet companies believed that “users would always convert into revenue,” and investors valued companies on user numbers. Loss-making companies could still raise money at high valuations and go public. Wu says it is a good thing that the market has become more rational about applications in recent years. Li said speculative pricing had shifted to models and embodied AI; Wu’s only firm conclusion was that “every era has its own hot trends.”
27. “Small and beautiful” is a choice; matching ambition to capability matters more than taking on giants
Wu says that early after Meitu went public, he had unrealistic ambitions to push beyond the company’s natural boundaries. Now he “likes small and beautiful.” By revenue or headcount, Meitu might qualify as “middle class,” but rankings mean little to him. He admires Germany’s hidden champions for their focus and specialization, and believes a small-and-beautiful company can still generate substantial revenue.
Asked whether Meitu can compete with ByteDance, he is willing to say directly: “We can’t beat them.” The Don Quixote-like passion of his earlier entrepreneurial years has given way, after ups and downs, to more objective goal-setting: match ambition to capability, or even set goals slightly below it, so execution can outperform and improve the odds of doing the job well.
Both smartphones and generative AI have shown him how important scale effects are. If Meitu has only 1/10—or even less—of a competitor’s budget and tries to do the same thing, it is unlikely to do it well. “We don’t gamble.” The company does not bet on a low-probability big win. It has been public for 10 years and has responsibilities to shareholders as well as employees; Wu is not willing to go all-in on one major strategic bet.
The “tragedy of tools” has also eased after Meitu succeeded with subscriptions. Tools were seen as having low barriers to entry, weak retention, easy switching and poor monetization. Wu has gradually come to recognize that they have their own value system. With hundreds of billions, even trillions, of dollars in market value among tool companies globally, Meitu no longer has to become a platform to prove its worth.
28. Post-IPO expansion took the company from 1,000 to 3,000 employees and gradually out of control
Going public had been Wu’s deeply held measure of success; the sense of achievement lasted about a year. Media attention, funding and a larger user base made him feel that Meitu could do more. The team grew from about 1,000 people at the end of 2016 to more than 3,000 in 2018, and things began to slip out of control in little more than a year.
Li asked whether, armed with what he knows now, Wu would have done better back then—or not gone through with it at all. Wu did not want to pretend he could travel back in time, but said that if he were advising a company at a similar stage, he would first review the business and make choices. The problem then was “I wanted it all.” Today he would look at the team’s DNA, capabilities, market size and returns.
Meipai is the project he still regrets. It had achieved a strong first phase, but the company did not keep building it well because of competing priorities, gaps in understanding and other factors. If he could go back, he would “do Meipai properly.” When Li asked whether it could have become a third pole alongside Douyin and Kuaishou, Wu agreed that it was possible.
On Li’s reference to “one and a half successes and 3 failures,” Wu said failures had brought more lessons and growth in understanding. He is now more willing to embrace failure, identify the cause quickly and look for solutions. Products themselves are refined through many small failures along the way.
29. As chairman as well as CEO, Wu focuses more on the organization, self-sufficiency and self-audit
Since taking on both chairman and CEO roles in 2023, Wu has focused more on the organization, strategy and monetization, and taken on more investor-relations conversations. He sees the imaging platform as a core capability: identify common needs across products, keep refining the underlying capabilities, then reuse them across the business. Operationally, he stresses the need to generate enough cash internally at all times.
He has learned about management from colleagues, formal study and company visits. He studied for a total of 4 years in a program run by Tencent’s Qing Teng and Peking University’s Guanghua School of Management, and is now pursuing a DBA. Meitu does not try to copy another company wholesale; it looks for proven approaches to specific problems. In his view, a business combining art and technology is hard to match to any single template.
Asked about Tim Cook, Wu gave an unequivocal endorsement: Apple’s revenue and profits have grown during his tenure, and it has continued to deliver returns to shareholders, so “he is certainly a good CEO.” The regret is that Apple has produced fewer of the astonishing shocks associated with the Steve Jobs era, but the company and technology are at different stages. Wu is more curious about what Apple will bring to the AI era.
Wu admits that his own unilateral decision-making contributed to the consequences of past overexpansion. He and the team pushed most of the business forward directly, and by the time they realized something was wrong, it was too late. His expanded chairman responsibilities now add another review of compliance and whether investments are necessary. “Before, it was a bit like having no audit at all.” Li said he was mainly auditing himself; Wu replied, “Auditing myself too.”
30. From perfectionism to collective execution, young people need to be given real responsibility
The most difficult period was the shock of expansion in 2017 and 2018, followed by layoffs and project cuts in 2019. Wu describes himself as introverted and an INTJ; he internalized much of the pressure and did not ask others for much help. What kept him going was also “not wanting to admit defeat” and “pride.” Today, that pride is more about making the company successful and building world-class products.
In his view, early perfectionism also reflected individualism. His first 2 startups were small, so he did a lot himself and treated products like works of art to be polished. Now he is moving toward collective execution and thinking in terms of priorities and trade-offs: world-class products require strong organizations, and one person’s attention to detail cannot substitute for a team’s capabilities.
He worries about becoming too conservative and losing touch with younger people. The oldest members of Gen Alpha are now 16, while AI has lowered the barriers to starting a company even further. Young people need to join and take on real responsibility. One recent success has been using school courses, student discounts and campus ambassadors to help students become proficient with the products earlier and turn schools into global growth nodes.
Wu describes his current state as “a mix of excitement and anxiety,” with relatively little pain. The excitement comes from new product categories and seemingly unlimited possibilities; the anxiety comes from the pace of iteration and the question of whether models will enter Meitu’s territory—and how to defend it. He wants to turn anxiety into motivation and believes a moderate amount of external pressure is useful.
31. A moat compounds over time through good choices, but scale effects may still be challenged
Wu does not attribute a long-term moat to any one static feature. He cares more about making the right choices at each stage, allocating limited resources, using validation to select what works and then scaling those choices over time. “If it’s just a choice without time working in its favor, it won’t work.” Durable advantage comes from continuing to do the right things.
More of those choices are now made collaboratively with the team, using complementary strengths and reducing the unilateral, instinctive decision-making of the past. For Wu, a moat is also connected to how decisions are made: first decide what is worth doing, then give sustained investment time to compound, rather than assuming one successful call guarantees a durable advantage.
The final idea he is revisiting is economies of scale. “I’m not rejecting it now,” he stresses; it still applies to large models. But if applications can use Agents to coordinate automatically, with digital employees and super-individuals, it remains an open question whether small teams can do the work of thousands. He hopes the old rule can be broken.
If that possibility becomes real, Wu believes it would mark a genuine explosion in human productivity. Li immediately asked whether Meitu itself is not an organization of thousands. Wu returned to the choice that has run through the conversation: think ahead about these possibilities and look for ways to disrupt yourself. “It’s always better for disruption to happen internally than to be hit from the outside.”