Every AI startup founder has heard the same terrifying prediction: Large language models will eat the application layer. There’s no room for you. It’s a grim consensus that’s killed countless pitches and sent teams back to the drawing board. But if that prediction were absolute, one company should have been dead long ago: Meitu.
Meitu’s CFO, Gary Ngan, recently sat down with the AI Proem podcast for an hour-long conversation that should worry every believer in the ‘LLM-eats-everything’ story. Not because Meitu has a secret supermodel—but because they’ve built a business on something that can’t be trained into a general model: subjective human taste.
Here’s the raw number: Meitu is running at over 70% gross margins. In a world where every AI company is bleeding cash on compute, they’re printing money. How? By refusing to play the ‘scale at all costs’ game. Instead, they’ve carved out a niche in visual aesthetics—a domain where there is no objective answer, where ‘handsome’ in America is not ‘handsome’ in Asia, and where a crop tool is not a crop tool but a cultural statement.
Let’s unpack what this means for every AI founder, product manager, and strategist who’s been told that vertical depth is a death sentence.
Why ‘Beautiful’ Is a War You Can’t Win by Throwing GPUs At It
Ngan dropped a line that should be framed in every AI startup office: ‘Aesthetic judgment is too subjective for a single general model to conquer.’ He’s not being philosophical—he’s being tactical. Language models thrive on tasks with objective answers: a correct line of code, an accurate summary, a factual question. Visual aesthetics? There is no ground truth. The same prompt ‘make this look luxurious’ generates wildly different results for a Japanese e-commerce seller versus a British fashion brand. A general model, trained on the average of all global data, will naturally flatten these differences. That’s not a bug—it’s a feature for the generalist. But for the specialist, that flattening is a gap the size of a market.
Meitu’s entire product philosophy is built on exploiting this gap. They don’t just build a tool and hope it works. They let products grow out of user behavior. Meitu Design Suite started as a poster-making feature inside Meitu’s flagship app. Kaipai (a video creation tool) grew out of a teleprompter feature in Meitu’s camera app. Each new product is a container for an existing demand, not a speculative bet. This ‘soil-first’ product strategy means validation costs are near zero, and user needs are already proven.
The Golden Quote That Screenshots Itself
‘If you don’t train your own model, your aesthetic standards will be defined by a third party—and they don’t care about your niche.’ That’s Ngan’s blunt answer to why Meitu builds its own models, despite the conventional wisdom that app layer companies should just rent APIs. Over 90% of Meitu’s AI output comes from proprietary models. The rationale: if you rely on a general API, you’re outsourcing taste. And taste is the only thing that matters in a visual product.
But here’s the twist: Meitu doesn’t religiously avoid third-party models. They use them for long-tail scenarios where proprietary models are overkill. The priority is user satisfaction, not model pride. That’s a rare maturity in an AI world obsessed with ‘owning the stack.’
Localization Isn’t a Translation Budget—It’s an Embassy
The most compelling part of the podcast is Ngan’s story about Lunar New Year. Meitu’s Chinese team created a festive theme featuring red decorations—a hit in China. They launched it in Korea without changes. It flopped. Why? Because Koreans wear white during Lunar New Year, not red. A trivial detail? Only if you think culture is a static variable. True localization doesn’t mean translating your app; it means physically embedding your team into local lives.
Ngan describes sending product managers to live in target markets, to interview users face-to-face, to understand why a British influencer loved the jawline tool but hated the skin-smoothing feature. These are insights that no data report can capture. They are the kind of messy, human nuances that general models, trained on averaged global data, will systematically erase. The companies that survive the AI onslaught will be the ones that build organizational muscle for this kind of thick, grounded localization.
The Subscription Trap and the Coming Hybrid Model
Meitu’s revenue model is a lesson in patience. They shifted from advertising to subscriptions in 2022, and now subscriptions are their largest income stream. But Ngan admits they’re only at 5-6% penetration in China, while the theoretical ceiling is 10% (based on global benchmarks). That’s 80-100% growth just from the domestic market. In the US, their AirBrush app already has over 50% paid conversion. The math is compelling.
But the real innovation is in the new product, Picchi. It’s an agent that learns your personal editing style from a few uploads, then applies it automatically. The user pays not just for a subscription, but for the ‘right to train’ a personalized model. This is a hybrid model: subscription + usage-based compute. If you can crack the C-end hybrid monetization model, you unlock a revenue multiplier that pure subscription or pure ad models can’t touch. This is where the puck is going.
The Real Competition Isn’t Canva or Adobe—It’s Indifference
When asked about competition from ByteDance, Canva, or Adobe, Ngan doesn’t flinch. He divides the field into two buckets: consumer (where no major Western player is doing what Meitu does) and productivity (where Canva is a generalist, Meitu is a vertical specialist). His example: Canva won’t tell you that your e-commerce product image can’t feature a minor as a model. Meitu Design Suite will, because it’s built for that specific industry. General platforms don’t want to do the dirty work of vertical depth. That’s your open door.
So, will the LLM devour Meitu? Ngan’s answer: no, because taste is too subjective. I’d add a longer-term caveat: the technology gap will narrow, but the organizational capability to productize subjective aesthetics—that takes years of accumulated user behavior data, a deep bench of designers who train the models, and a culture that eats local nuance for breakfast. That is not a moat you can buy. It’s a moat you build, one user interview at a time.
Meitu’s story is a blueprint for every AI startup that’s been told to ‘go horizontal or die.’ The truth is simpler: go vertical, go local, and bet on the things that make humans irreplaceable. The LLMs will take the rest. But the rest was never the whole cake.
FAQ
Q: Isn't Meitu just a Chinese beauty app? How does that apply to global AI startups?
A: The core insight is universal: if your product deals with subjective human judgment (design, taste, culture), general models will always struggle to capture niche nuances. Meitu's approach—building vertical tools, training on local behavior, and keeping a human-in-the-loop for taste—is a playbook for any AI company in a 'soft' domain.
Q: Won't better models eventually make Meitu's proprietary training obsolete?
A: Model quality will improve, but aesthetic judgment isn't just a technical problem—it's an organizational one. The data advantage (millions of user preference signals) and the cultural embedding (designers who understand local tastes) compound over time. A general model can't replicate that unless it's trained on the same private data, which it legally can't access.
Q: Meitu's 70% margins sound unsustainable—aren't they just lucky with a niche?
A: The margins are real because most visual editing tasks are lightweight (crop, retouch, remove background) and don't require heavy compute. The 'heavy' AI is used sparingly, often for premium features. This is a structural advantage: visual editing has a fundamentally different cost structure than text generation or video synthesis. Other verticals can replicate this by choosing problem spaces with similar asymmetry.