But just yesterday, ByteDance seemed to have stealthily rolled out the function across its video editing tools (in non-US markets). And making it feel more and more like a reality where AI-generated video will no longer just be slop, but become a vertical of its own, potentially replacing some aspects of acting as we know it and moving toward industrial use. People raved about the capabilities here and here.

The thing is, I think Hollywood is still looking at the wrong battlefield, at least for now. The first meaningful commercial win for Chinese AI video is unlikely to be the two-hour blockbuster that demands sophisticated acting, screenwriting, costumes, lighting, motions, character arc, and cinematography. It’s the 90-second microdrama that is made for the phone, vertical, bingeable, optimized not for Emmys but for thumb-stopping cliffhangers.

While I'm not a fan of these microdramas myself, I did give them a try to better understand the product. I have to say, the business model to integrate AI is quite smart.

Think of it as Content Factory vs. Content Studio. A conventional studio will push forward and hire a massive crew to create a masterpiece that often incites intellectual debate, is conscious of social and political context, and evokes emotions and fosters universal connectivity; that’s what usually makes these big-screen movies a hit. An AI-film factory, on the other hand, tests, iterates, acquires traffic, and monetizes retention. Hollywood is worried about the studio model being disrupted, but the real disruption is the factory, which, in many ways, looks more like streaming services and influencer content, and now AI is what makes the factory run.

As I have written before, the interesting part of AI is increasingly not who has the flashiest frontier model, but who turns that capability into an indispensable application with a repeatable business model. In creative AI specifically, the question is not “how many people tried the app,” but whether the product solves a commercial workflow with enough ROI to justify the cost of inference and enough urgency for people to pay. Microdrama is a near-perfect case study of that logic.

And we’ve seen that over the last year, there’s been growing interest in microdramas from China. MIT Tech Review wrote about it last year, WSJ wrote about it recently, and more people are starting to see the connection to AI-generated video. So let me break it down.

Share

What Microdramas Actually Are

To be clear, Microdramas are not just shorter TV shows. They are smartphone-native, retention-optimized, and monetized more like gaming than like traditional streaming. Episodes run one to a few minutes. They often end each episode on betrayal, revenge, fake marriages, secret heirs, werewolves, billionaires, and miraculous recoveries.

Just as you’re getting super worked up, the app will pause and ask you to do what is more similar to what mobile gaming has trained people to do for years: pay to unlock the next episode, wait, or watch an ad. To end on that hook, to make people yearn to find out what’s next (even though usually it’s quite obvious) is exactly the point, and people are willing to pay to find out what happens next.

The product is not pretending to be prestige storytelling, nor is it even attempting to be jaw-dropping cinematography; it is shamelessly positioned as narrative dopamine.

Reuters reported that one of their reporters visited a set where an entire season was shot in six days. A shabby room or studio that can be staged for essentially all the scenes, and some (no offense) D-listers or amateur or student actors. That’s all it takes. For different markets, studios may hire different ethnic actors to play out the same screenplay. Budgets typically range from about US$28,000 to $280,000, which is like a rounding error next to a Netflix original.

The content is cheap enough to be disposable, but sticky enough that one hit can pay for many misses. At its core, it is not pretending to be high-brow art, but rather visual dopamine hits created with SEO in mind and performance marketing attached.

Subscribe now

The Scale Is No Longer Niche

Micro-drama as a sector is not so micro anymore, it is already a real consumer-spending category and proven to no longer be an experiment. And to my surprise, the North American market is actually the main profit pool for most of these companies.

Sensor Tower says micro-drama apps ReelShort and DramaBox generated about $130 million and $120 million of in-app revenue in Q1 2025 alone, bringing their cumulative global in-app revenue to roughly $490 million and $450 million, respectively, by March 2025. Of which, the U.S. market now contributes 60–70% of Chinese micro drama’s overseas short-drama revenue, and the average payment per download for ReelShort and DramaBox in the U.S. is about six times that of other markets.

Why AI Fits This Format First

Then there is the obvious shift in embracing AI tools for creating microdramas. (plus investors in this space have told me about this trend even a few months ago)

AI video won’t replace Christopher Nolan or James Cameron; it also won’t replace Jennifer Lawrence or Leonardo DiCaprio. In many ways, what it will do is actually move value and influence toward the cream of the crop even further (yes, than it already is).

The argument I’ve heard from multiple sources is that the impact has the potential to be both good and bad, depending on whose perspective you’re getting. For actors who have not made their name yet, it might mean they no longer even get booked for these lower-budget films, the ones that they used to build up their portfolio. For directors, they may think they no longer need actors and can actually produce a film with minimal investment. For screenwriters, they may be empowered to try to make their vision a reality with fewer hurdles.

For micro-drama studios, the desire to reduce the cost and time required to produce something that is already “good enough” for a one-minute vertical episode watched half-distracted in bed will drive their rapid adoption of AI tools. Even the spokespeople at Kuaishou told me that, for AI-native film studios, microdrama does not require cinematic perfection. This vertical requires speed, emotional pull, and just enough visual coherence to keep the thumb/eyeball moving. A continuity and consistency error that would ruin a TV (mid-screen?) drama may barely matter here.

That is why AI finds product-market fit in the content factory before it reaches the content studio.

Kuaishou says Kling now serves more than 60 million creators worldwide, has generated over 600 million videos, and has partnered with more than 30,000 enterprise clients. In January, Kuaishou also said Kling’s monthly revenue had exceeded $20 million in December 2025, implying a $240 million annualized run rate. Now, Kuaishou is a publicly traded company with every incentive to showcase AI monetization to investors, so treat these numbers as directional rather than audited — but the trajectory is unmistakable.

The company spokesperson told me that even as early as last year, Kling has already been used by professionals across marketing, e-commerce, film and television, short plays, animation, and gaming. And recent partnerships with Tokyo, Hong Kong, and more film festivals, and even collaborations with film schools to empower indie or independent filmmakers, are a niche they continue to try to cultivate.

The Value Chain: IP, Tools, and the Funnel

This is also why the companies behind AI microdramas matter more than the shows themselves. The value chain has three layers, and each tells you something about where the moat actually sits.

A. The IP layer is China’s least appreciated advantage. Reuters reported that China had 575 million online literature readers in 2024 and 352 million overseas readers across more than 200 countries. China’s robust digital literary space was especially vibrant during the 2000s, where online-native writers turned web posts into bestselling books and then into blockbuster movies. That era enabled talented young writers like Hanhan, who is the face of that era of openness.

But that online ecosystem is not as simple as a database of essays; it is also a training ground for writers who have spent years optimizing for cliffhangers, reversals, binge behavior, and emotional hooks. Which is exactly the skills microdrama demands. COL Group, the Chinese digital content company that has a large stake in the studio that owns the app ReelShort and its ecosystem. COL itself claims to have more than 4.5 million authors and 5.5 million digital works. The point is not just that China has “a lot of content.” That is a hugeeee upstream reservoir of serialized IP and retention-trained writers ready to write microdrama scripts essentially. Thus, there have been rumors that companies like COL continue to buy out vast amounts of these IPs, in a way, playing the VC game. If they ship out enough series, one will become big, and that one will be enough to justify the production costs of the others.

B. The tool layer is where ByteDance and Kuaishou are competing the most fiercely. ByteDance’s edge has long been recommendation infrastructure, product iteration speed, and the ability to funnel users across a wider ecosystem. In many ways, ByteDance’s release of Dreamina Seedance 2.0 is probably going to give Kling’s give a run for its money. Both have the advantage of being able to train on vast amounts of video data, but ByteDance’s may be higher in quality and diversity. The company’s editing tools have also taught its AI taste, preference for subtle and nuanced ones.

For ByteDance, microdrama is not merely a customer vertical; it is a proof point for why multimodal AI belongs inside an existing distribution and monetization machine. Kuaishou’s strategy, although similar, has slight differences. Kling is already monetized through a combination of consumer subscriptions, API services, and customized solutions. and focused on relationship building directly with indie studios and institutions.

C. The app layer is where the funnel lives — and tbh, where I think the real moat forms. Operators like ReelShort, DramaBox, GoodShort, and ShortMax are not “studios.” They are mobile consumer businesses. Their real assets are distribution, recommendation, paywall design, localization, and user-acquisition know-how. In this world, AI is not just a production tool; it is a margin lever and a velocity lever. If AI lowers the cost of making scenes, dubbing material in any language of demand, generating trailers, or testing thumbnail variants, then these apps can push more content into the same funnel and learn faster which stories convert.

The Funnel Wins

So here is the strategic core, and this is the part I want to highlight: AI lowers production costs, but it does not automatically make content businesses suddenly lucrative. In fact, once production becomes cheaper for everyone, production itself becomes less scarce, and in many ways, competition becomes fiercer. The moat shifts.

The moat is in distribution, and that is what ReelShorts and DramaBox have been investing heavily in. Even though the microdrama business is not lucrative or even loss-making right now. The thinking is to first get as many users to download their apps as possible, and if they own the distribution, then once the cost of production goes down drastically, the opportunity to monetize at scale is theirs to take.

So, looking at the forces, I’d rank distribution and IP ownership as the two that matter most. Distribution because it determines who captures the economics of attention — and in a world of infinite content, attention is the only scarce thing left. IP because it determines who has a defensible supply of stories that are pre-optimized for the format, and that supply is not easily replicated. Payment conversion, localization, and recommendation matter too, but they’re features of the funnel, not independent moats. And that also proves my point that it is going for the streaming platforms’ lunch first, not Hollywood, nor even YouTube/creator content first.

The companies that own both the stories and the funnel — like the COL Group / Crazy Maple / ReelShort nexus — are structurally advantaged. You see what I mean? It’s less about who has the best AI model and more about who controls the full loop.

This is why I’d frame AI microdrama less as a “film” story and more as an application story. It sits at the intersection of content, gaming, and performance marketing. The model provider supplies the tool. The IP company supplies the story. The app operator supplies the funnel. The app store takes its cut. The ad platforms sell the traffic.

The highbrow studios and the traditional blockbuster model will survive for prestige, for cultural events, for the kind of storytelling that requires genuine human nuance and craft. But the content factory is where volume, velocity, and AI economics converge. And that factory is already running.

Risks and Quiet Winners

Disney and Paramount have already challenged ByteDance over Seedance. Domestically, China’s regulator is also tightening control over the category. Reuters reported that the country now requires licenses to broadcast microdramas, part of a broader regulatory push after earlier campaigns removed tens of thousands of shows and roughly 1.4 million episodes for lowbrow, violent, or vulgar content, or material that promotes ‘unhealthy behavior’. Regulation will probably further consolidate the market, which favors incumbents with the resources to comply and disadvantages smaller operators, thus making the first movers like ReelShorts and DramaBox in a much better position.

And there is a structural irony worth flagging: if AI sharply lowers production costs for everyone, the market may be flooded with even more supply, which could compress producers' margins and shift even more of the economics toward traffic owners and app stores. Apple and Google, who extract 15–30% of in-app purchases, are quiet winners in the microdrama economy. They will be collecting rent on hundreds of millions of dollars in transactions without producing a single frame of content.

As of now, it looks like these Chinese micro-drama studios are building something closer to an AI-native content factory: web-novel IP upstream, retention-trained, serialized writers, short-video recommendation systems, mobile paywalls, performance marketing, and, increasingly, capable video models that power production.

It is actually a consumer-application story in which production costs collapse and the moat shifts from creation to distribution. It happened in music, where Spotify and playlists matter more than studios. It happened in publishing, where Amazon’s algorithm became the gatekeeper. And it is now happening in short films.

Relevant reads:

• 大公报：国产微短剧海外受捧 下载量超YouTube

• AI Proem: Kuaishou’s Strategy: Riding the Rise of Short Mini Dramas + Industry Expertise Over AGI Chase

• AI Proem: The Ghibli Hype: Next Frontier is in Video & Introducing China’s Leading AI Video App, Kling

• AI Proem: ByteDance’s Second Act in AI + OEM Partnership + DeepSeek V3.2

• 人民网：竖屏新势力，中国微短剧如何“踏浪出海”

Subscribe now

Child-first product design

While I was in San Francisco last week, I was introduced to a very interesting startup called Dex Camera. It looks like a little fan, but it is actually a handheld camera with a small screen that kids can point at objects and ask what they are. It is an AI-powered educational device. The tone sounds like Ms. Rachel. The design is simple, like the Toniebox. If you have a toddler, you know exactly what I’m talking about.

The AI-powered educational product seemed consciously designed around the most relentless toddler question in the world: What is that? The child points, and then the device responds. It asks the child to repeat after it before moving on. It turns curiosity into a loop.

As a mother of a toddler in a trilingual (quadrilingual) environment, this made sense to me instantly. We speak English at home, she speaks Mandarin at nursery, and she is surrounded by Cantonese at the playground. All day I get some version of, “mama what is that, how do you say it in English, how do you say it in Mandarin? No, but what about Cantonese?” So when I saw Dex, I was quite enamored by a tech-enabled language tool built around how a child’s brain naturally works.

But Dex itself is not really the story here today. Dex just made me realize how rarely we design technology from the child upward rather than the adult downward.

In all our mommy groups/ playdate chats or whatnot, we keep debating whether screen time is good or bad, when maybe we should first ask whether the product was ever designed for a child in the first place.

Most of the digital products children encounter today are adult products with parental controls bolted on later. A smartphone, a tablet, and a chatbot are, by nature, made for adults. All we have done is add time limits, filters, controls, and guardrails, and hope that somehow transforms them into something child-appropriate.

But every once in a while, you see a product, and it becomes obvious that the people behind it have actually spent time with small children and tried to understand them. Toniebox is one example. Dex is another. 小天才手表, the Little Genius watches in China, is another. They are all doing different things, but they share one common trait: they are not adult devices shrunk down for kids. They are built around what children can physically do, cognitively handle, and naturally want.

The universally recognized IPs, such as Ms. Rachel, and devices like Tonieboxes, are a testament to the fact that good product design knows no borders. Ms. Rachel herself, a trained early-stage educator who combined her expertise and her husband’s, who works on Broadway, created one of the biggest musical/educational IP empires in modern day. They are the “best” kind of tech learning and entertainment for children, and probably would get you the least amount of judgment from the people who love to give unsolicited parenting advice.

And if you think about it, the underlying technology for Toniebox is not new. Audio stories are not new. But the product design is intuitive enough for a small child to use independently. Put the figure on top, the story starts. Remove it, the story stops. The ears act as volume buttons and are physically obvious, unlike an iPhone screen, where children do not understand the lack of texture. As a mother, I can’t overstate how much that matters. It really is product design with a children-first mentality.

So with that, it made me think about how to approach AI for kids, especially the children who will be the first AI-native generation, and why there is still no clear leader in this vertical.

The overlooked business vertical

I think part of the problem is that we may be misclassifying the category.

If you call these products toys, they look expensive or overly digital. If you call them gadgets, they look niche. But if you think of them as child-native learning and communication devices, the economics start to look different.

What makes this category commercially interesting is that parents are not really making an entertainment purchase. They are making a risk-management and convenience decision with a learning upside.

In practice, these products are competing less with toys than with a messy bundle of substitutes: the old iPhone handed down too early, the extra tutoring session, the endless parental explaining, the need to stay in touch, and the desire to give a child some autonomy without opening the entire internet. That is why the category can support higher pricing than a normal toy aisle product, because if done right, it solves several jobs at once.

That is why I think this may be a surprisingly overlooked vertical, even for me as a parent to young kids. It kind of sits awkwardly between consumer hardware, toys, education, and family safety, so nobody quite owns it. Consumer investors may dismiss it as a low-ticket novelty. Educators may think it is too gadget-driven or risky. Hardware investors may think it is too small a niche. But that blind spot may actually be an opportunity for innovation and something truly new.

Unlike AI schools or tutoring services like Alpha School, Khanmigo, or Ello, I want to look into the three business models of ‘made-for-kids’ tech products that I think could showcase the potential viable business models for future products to follow.

One model is Tonies. One is a Dex-like product. One is Little Genius.

A. Tonies = IP + content library

Toniebox is not really a hardware story. The box gets into the home, but the business is the library that forms around it. Tonies reported H1 2025 group revenue of €176.6 million, of which €136.7 million came from figurines and only €32.3 million came from Tonieboxes. The company also says it has sold more than 10 million Tonieboxes and about 134 million Tonies worldwide. What looks like a toy is really an IP and content-library business with household lock-in. It actually becomes a habitual lock-in, and then it becomes a hardware and software switching cost.

The Tonies model is clever because it combines licensed characters with original IP. The licensed characters do the obvious job: they attract attention, take up shelf space, and secure the first purchase. And they let Tonie own more of the habit, the release calendar, and the relationship with the child and the family. The trade-off, of course, is that licensed IP is helpful for customer acquisition but can also compress margins and leave the company more dependent on third-party franchises than investors may initially appreciate. Recent figurine growth, by the company’s own description, has been driven by a mix of licensed third-party Tonies and Tonies Originals, including Sleepy Friends.

And that is why I think Tonies is such an important template for what AI-native children’s products could become because the underlying design principles are so strong: durability, simplicity, low-friction engagement, and real autonomy. The child has agency, but inside a bounded system. They choose and react, and through that autonomy, they build taste, and they are not just swiping through an infinite sludge of screens.

B. Dex-like = hardware + recurring AI service

A Dex-like product is a very different business. Dex currently sells a $249 device (which I do think is a bit $) and offers a free plan alongside paid tiers at $9.99 and $19.99 a month. The plans add features like LTE connectivity, daily activities, and more customized learning, while the current version still relies on internet access, whether via Wi-Fi or an always-on eSIM.

That is a fundamentally different economic model. Tonies can sell the same figurine again and again. A Dex-like product, by contrast, likely needs a subscription (much like Spotify's business model) to justify the ongoing compute, moderation, safety layers, and fresh content. And that is where the challenge becomes more interesting. Parents do not want to feel like they are paying per token or being nickel-and-dimed every time a child points at a flower or a bus. But the company also cannot pretend that inference, safety, and personalization are free forever. So the product has to make pricing feel simple, while the economics underneath are much more complex.

Which is why I do not think the moat here can just be “we wrapped an LLM in a cute device.” At this rate, anyone can build a wrapper. The moat would have to come from a combination of hardware and software design, including child user behavior, trust, guardrails, saved progress, family language setup, curriculum, and habit. In other words, the defensibility is not solely in the model. It is in how well the entire experience is designed for a child and how trusted it is by a parent.

(btw I asked the founder where I can get one of these, he pointed to their Shopify site)

C. Little Genius = network + controlled communication

Then there is Little Genius 小天才手表, which points to a third model again. Not IP. Not token usage. More like a network and family-safe ecosystem as the selling point. Bright colors and big buttons. Easy calling and location tracking. Saved contacts and controlled communication. No truly open internet, but basically all the technology that modern parents want their kids to carry around.

According to SCMP, citing Counterpoint Research, Little Genius held more than 48% of the global children’s smartwatch market in the first half of 2025. That number matters not just because it is large, but because it tells you this category is not imaginary. I think there is real demand for devices that sit between no technology at all and full internet access.

Share

It also hints at where this business model can be a bit tricky. A communication device can become a network. A network can become a social graph. And once children want to be where their friends are, the moat is no longer just the hardware. It is the network density and the norms that form within it – it’s like a child-only, safe-space social media network. That may turn out to be the most defensible model of all, but it can also be a slippery slope, as it can become the one most vulnerable to regulation and interoperability, probably both. Wired recently reported on how social features inside Xiaotiancai have already produced like-chasing, status games, and cyberbullying concerns, which is exactly what happens when a child-safe device starts drifting toward social media logic.

So when people talk about “AI for children” as if it is one market, I think they are already thinking too loosely. Tonies, Dex-like devices, and Little Genius are three very different businesses. One is IP plus content-library lock-in. One is hardware plus a recurring AI service. One is network plus controlled communication.

Technology does not have to replace thinking

I feel that the broader debate around AI and children right now often gets too binary. It also falls under a bigger discussion that often gets collapsed into one: is technology improving our lives, or is it ruining them? My own view is that we cannot reverse technological progress, and eventually, they will all need to learn how to co-exist with technology. What we can do is design technology more mindfully for children.

Technology, broadly speaking, has improved the quality of life throughout human history. The question is what kind of technology we are placing in front of children, and what kind of habits it is training into them.

On the one hand, there is the view that AI and smartphones are melting children’s brains, killing attention spans, and replacing real thinking. On the other side is a kind of super-charged techno-optimism that puts children in front of technology designed for adults and expects them to just learn to use these tools early because the future is inevitable anyway.

I find both sides are lacking nuance, and the reality is that there is great tension between these evolving thoughts.

I do not think it is paranoid to worry about addiction, bad content, persuasive design, privacy, or the habit of outsourcing thought. A lot of those concerns are justified. And this is where I do agree with a lot of Jonathan Haidt’s points in The Anxious Generation, and it has basically become gospel among millennial parents. My husband certainly treated it that way for a while and preached it to everyone we met.

The reason the book hit so hard is that it captured something many parents already felt, especially for us who did not see the dark side of social media coming when we were introduced to it. The gist of his book is that there is an irony in how we parent today. While we have become overprotective of children physically, we have also turned a blind eye to how they’re exposed to the digital world. We monitor where they go, but hand them devices that expose them to distraction, negativity, age-inappropriate content, and potentially unsafe contacts.

But not all technology belongs in the same bucket. Bluey and YouTube autoplay are not the same thing. A reading coach and an AI companion are not the same thing. A handheld language camera and a general-purpose chatbot are not the same thing. We flatten these categories, then act surprised when the debate goes nowhere and yields no real constructive solutions.

That nuance matters because the key to learning is still thinking. Observation. Retrieval. Pattern recognition. Frustration tolerance. Curiosity. Technology does not need to replace any of that. In fact, the best technology for children should do the opposite. It should protect those muscles, because that is what teaches critical thinking and continuous learning.

A child using a camera to identify a leaf, repeat the word, hear it in another language, and then go look for another object is still observing the world. A child using a reading tool that listens while they decode words is still doing the reading. A child using an always-on AI companion as a best friend or all-purpose answer engine is in a very different kind of setting.

Common Sense Media warned in January that no child under 5 should be given an AI toy companion, urged extreme caution for ages 6 to 12, and found that 27% of AI toy outputs in its testing were inappropriate for children. That is not a small miss rate when the user is a child.

Subscribe now

A good child-first product should encourage the child to engage with the real world and do what they are intuitively curious to do: try new things and learn by making and asking. A bad one does the opposite. It fills the silence, removes friction, gives answers too quickly, and slowly trains the child to eliminate uncertainty rather than sit with it.

The reality is that children today will grow up in a world saturated with AI and devices. In fact, I think it would probably harm them if they were completely siloed away from technology and then suddenly became the only ones in the classroom who did not know how to use a device, or were seen as weird for being the only one not online. What I think we should really be pushing tech companies, and maybe educators too, to do is build child-safe and children-first products.

It is accepted across all parenting philosophies that children need some form of guidance on how they interact with others, what literature to read, and where to go running without falling into a pit, but shouldn’t that consensus now extend to how they interact with technology, too?

The future of children’s technology should not be just about making kids better at using machines. I reject the “vocational school” approach to it all too. It should be about designing tools that still leave room for noticing, wondering, remembering, struggling a little, and making sense of the world.

In many ways, AI can absolutely be used to kill thinking if what you expose children to is a general LLM. But it can also be designed to harness better thinking, and a more realistic embrace of the digital world we already live in.

I know some people may dismiss this topic as “soft,” but I do not think it is soft at all. It is about our children, about technology, and about a business vertical that still feels underbuilt and underthought. I hope more people take this topic seriously and think about the consequences of our children x technology and design and build more mindfully.

Disclosure: this is not a sponsored piece. I just came away genuinely intrigued by some of the devices now being designed for children, and by how much more intentional they feel than simply handing a child an adult device with a few filters slapped on top.

Hong Kong is going through a heartbreaking tragedy right now. My heart goes out to the families affected by the deadly fire.

It’s been deeply sickening to watch parts of X turn it into fuel for political narratives, when the only thing that should matter at this moment is getting survivors to safety and supporting the families who’ve lost so much.

My family and I are doing what we can, gathering winter clothing, basic necessities, and direct donations to local charities and NGOs. I’m acutely aware of how small that is compared to the scale of what’s needed, but if you’re in a position to help, I’ll share a list of organizations at the end of this piece.

And a day after Thanksgiving, I just want to say thank you for supporting AI Proem, and remember - there is always room for more kindness in this world.

Ok, now, let’s talk AI.

I met with a representative of Kling, Kuaishou’s AI team, this week. The meeting was meant to be a fairly “normal” catch-up and check-in on models, product roadmap, and use cases. But it ended up becoming something else: a window into how AI is diffusing into creative fields, and how differently investors fixated on value creation, builders focused on technological advancement, and creatives who hold the utmost respect for craftsmanship experience and evaluate AI.

Are we creating AI for the current generation or the next?

What stayed with me from that conversation wasn’t the tech demo. Kling’s models are objectively impressive: cinematic motion, controllable camera paths, and increasingly photoreal characters. Apparently, the “cat skating on ice” video was a huge deal in the industry because of how sophisticated the movement and physics were.

But the thing that really lingered was something the Kling representative said almost in passing: their team had been genuinely surprised by the kind of feedback they were getting from people in the creative fields.

When directors, actors, and editors start using Kling, their biggest concerns are not about resolution or latency. They ask: Am I accidentally infringing on someone’s IP? Does this cheapen what I’ve spent my life learning to do? Who actually owns the shot if AI generated half the frames? What about the martial art artists, the chefs, and the musicians we hire for those expert shots? Copyright, authorship, and respect for craft remain top of mind for them.

In tech, people are often so fixated on building the best system that they can become oblivious to the potential harm or disruption that system creates. It’s not that the developers are evil; it’s just that they’re optimising for a different objective. Investors focus on monetary value creation. Policy analysts focus on safety and guardrails. Developers focus on “进步” — advancement — and, as she put it, that can become the only thing they see. They don’t necessarily know what the real pain points of end users are. That’s why Kling has tried to flip the sequence: let users lead the product, and let the product guide the technology.

She shared another anecdote that I found fascinating. At a film school event, a well-known director and professor talked about how he always has to think about the physical space he can work with when staging a scene and planning actors’ movements.

A current student then challenged that assumption: why should we constrain ourselves to physical limits at all if we’re working with AI? Why not design the world exactly as we want? He went further and said that using AI means a director has to give up some control. The model will develop its own interpretation of the prompt, and it won’t always give you the precise image in your head. If you prompt “a sofa in a room,” you can’t expect it to perfectly match your imagination. In a way, making someone literally make that sofa in real life is easier, but that is also much more costly. Working with AI, in other words, means treating it as a creative partner rather than a tool completely in your control - like Photoshop?

That same student gave a simple example: when you want a cat to move in the frame, why must it always walk from left to right? What if the AI decides the cat should walk from right to left? He seemed far more relaxed about that kind of flexibility than many older generations in the industry. At the same time, anyone who has seriously used text-to-image or text-to-video tools knows that it’s incredibly hard to surgically tweak small details after the image or video is generated. So the way you prompt, including the language, the know-how of the industry, the context, and all, becomes highly technical. So it seems like prompting itself is turning into a craft.

All of this made me wonder if we are thinking about AI integration in the wrong way or in a narrow-minded way. We spend so much time discussing how AI will increase productivity within the workflows we already recognise. But what if AI-native generations reshape the workflows themselves? Thirty years ago, cartoonists drew with pencils on paper and spent hours sometimes to mix the perfect color. Today, they have digital tablets, animation suites, audio tools, and colour grading, so you can see that the entire work process is probably unrecognisable to their predecessors. The craft of creating cartoons has essentially changed completely.

So now the question is, do we evolve the craft or evolve the result? How do we not lose that craftsmanship, or at least continue to respect the aesthetics and skills that took decades to hone? The friction that is creating “slop” is actually because the creators of the tools do not understand the needs of the users.

Very focused, very clear target market

So that brings me to this point. Amid so many AI startups and big tech I’ve met, many have been pivoting their strategy from A to B within just months (or weeks trying to capture the gold rush). But what makes Kling interesting to me is that they’ve actually been quite committed to their vertical since the beginning. And they’re quietly inverting the usual AI hierarchy.

Expanding on what I mentioned above, take a step back and think it through.

The standard playbook is: build a powerful model, ship a flashy demo, and only then scramble to figure out which industries want it. Most big tech companies have been building general-purpose tools and startups and finding ways to “wrap” around them for vertical use cases.

With Kling, they’ve only worked with videos essentially, given that its parent company, Kuaishou, is China’s second popular short-video platform.

They have also been very specific with their user demographic: film directors, marketers, post-production teams, and AI-native drama studios. And you can see that in the business mix: roughly 70% of revenue comes from prosumers, think professional creators and marketers who are close enough to the end audience to care deeply about craft, but pragmatic enough to pay for speed and scale. The remaining 30% comes from enterprise API customers across advertising, animation, gaming, and film.

So what they’re doing to ensure satisfaction is letting the users lead the product, which then reverts to the tech. The team first adjusts the product surface to make it easier to mark up shots, control motion, and line up start and end frames. Then the model itself is nudged in response: better multi-view consistency, more controllable motion, finer editing handles. You can see that evolution in the progression from the first release to things like the Master Edition 2.0 and now 2.5 Turbo. The model is no longer the hero on stage; it’s the part backstage that quietly updates to keep up with what the outside world is asking for.

They’re also hiring industry people, who act like translators between the film industry and the tech people, and partnering up with film festivals and associations across the world. What we’re seeing is industry → workflow → product → model, this is actually quite a radical rewiring.

Looking at the business numbers. In under a year from launch in June 2024, Kling had already crossed an annualised revenue run-rate of over US$100m by March 2025, according to the material Kling shared with me. Quarterly revenue stepped up from around RMB 150m in Q1 to over RMB 250m in Q2.

Its deliberate positioning as an AI creative platform for the professionals actually is a pretty overlooked vertical.

When Hollywood is protesting against the use of AI, there are people who will embrace it, notably in markets where there is less money for creatives. And it made me think of when TikTok took off, a whole sub-industry started from zero- MCN agencies that managed influencers. In some ways, I think we’ll see a subcategory of AI native content take off. As for deepfake concerns, she noted that in China, there is a strict requirement that all AI content be watermarked, and technically, if you remove that watermark, you’re breaking the law and can face legal repercussions. But the thing is, this requirement varies across markets. [I’m a bit scared by the lack of an international standard on this tbh]

And while I thought ByteDance would be its biggest competitor, she actually said it’s only ByteDance’s Seedance that is doing something similar. As we’ve been writing about at AI Proem, ByteDance seems to be doing everything and is probably the most aggressive of the Chinese big tech in talent poaching right now. It’s doing everything from LLMs, to consumer apps, to video editing, and well- this.

To note, one of the four tigers, Minimax, has also been shipping out multi-modal models and is said to be very competitive. But as it has been on an investor courtship, two investors have shared with me that they feel the company is a bit inconsistent in its strategy right now. It seems distracted with high-profile apps like Hailuo and Talkie, but it is also still trying to compete on frontier models. However, given its grassroots beginning, it is slowly running out of money. It doesn’t seem to have strategically chosen a lane to focus on yet.

And in this case, Kling, by comparison, looks like it’s the least ambitious among the three, yet the most focused.

New subcategory: AI-native entertainment

One of the most unexpected and interesting wedges for Kling has been vertical mini-dramas. One use case of AI-generated content that is taking off quietly is the potential of creating 1080p, smartphone–OK-quality videos. Think of the telenovela-like short mini-dramas that are taking off. Many of the producers and studios behind them are actually Chinese.

China’s 短剧 has actually taken a very unexpected life on its own. Zeyi Yang actually wrote about it for Wired in July. These short videos are pretty much like what he wrote: acting quite meh, plot somewhat the same, actors tbh usually some kind of a D-lister, not to be mean (it is what it is). You know what drives their popularity? – our universal appetite for romance, melodrama, storytelling.

I mean, who doesn’t love a good romance? I, for one, miss the 90s–00s romcoms and am too braindead at the end of the day to watch high-brow, deep-thinking, soul-searching sci-fis that seem to dominate Netflix / Apple these days (I do appreciate Severance, but it requires too many brain cells to follow tbh).

And everything seems to have a satirical take these days, but sometimes I just want The Summer I Turned Pretty, which I shamelessly binged religiously.

My husband made fun of me for coming home from work to find me glued to the screen, watching teenagers half my age fall in love and flirt with their eyes. (Team Conrad, thank you. Don’t judge me, talk to my hand.)

Anyway, you know how AI creation can be plugged into this?

Wellllll, it’s not like anyone’s zooming in on their phones to see the details of actors’ expressions or if the fingers match the melody played on the piano. So Kling is plugging into this space and co-creating AI-assisted short films with directors like Jia Zhangke and Timmy Yip, and these are explicitly framed as human–machine co-creation rather than full replacement.

So, to end my rant today. I think the conversation with Kling serves as a useful lens on where AI might actually be heading - beyond the industries as we know them. It’s not the first time traditional media/ content/ entertainment has been disrupted.

NEW ECONOMIES

The New Titans of Media

The future of media isn’t arriving — it’s already here…

Read more

5 months ago · 77 likes · Ollie Forsyth

[ actually put together a guide on “The New Titans of Media” here, and humble me is included too. SanQ.]

What we’re seeing is that on one hand you’ll always have the ‘AGI’ chasers, blinders on, optimising for benchmarks and rankings but in another case of “good enough” combined with deep tacit knowledge on the other hand, AI native videos will have a potential place in this world. I’m not sure how much I would like them, but I see it as a potential new subcategory and new business model.

For those who want are interested. Here’s a list of organizations you can reference to help the families affected by the tragic fire in Tai Po, Hong Kong. 🙏

In fact, official numbers say that cross-boundary passenger traffic between the cities has reached a record high of 1.34 million trips in August this year, according to the HK government.

Anyway, as I was saying, I made my journey up this past weekend, and to my surprise, it only took me 15 minutes to get to the train station, 15 minutes to cross the border, and 15 minutes on the bullet train before I was in the city center of Shenzhen -Futian District. That’s really just a normal intracity commute for most major Chinese or U.S cities. And honestly, I felt like an absolute bumpkin in awe of the speed and efficiency and gawking at the parks and buildings when I came out of the station.

I haven’t been to Shenzhen since COVID ended, and my first observation is that at least 30-40% of cars are EVs; it might be a bit anecdotal, but honestly, it felt like everything was battery-powered. [Plug: I have an excellent episode coming out tmr with , which will be all about the electrification of the economy, China’s energy mix, and how China is powering the AI boom and beyond]

Previously, when I thought of big cities in Mainland China, I always envisioned super loud and congested crosswalks and traffic. But the roads were remarkably quiet, with a mix of BYD taxis, small electric buses, electric motorbikes, and e-bikes.

The electrification of the economy couldn’t be more real. But then I questioned whether Shenzhen might be more progressive on this front since BYD is a Shenzhen homegrown firm. As you can see in the pics, pretty much every taxi has now been switched to BYD. In HK, even with so many government incentives, I’d say the majority of cars are still gas. And my family only just switched to an EV a few months ago, it felt like such a novelty to our friends, even.

An observation that Le Shrub and Paulo talked about in their recent trip to SZ as well. BYD is actually just affordable and cool, and people underestimate the power of cool.

XIN Summit

Now, enough with my rant. To the main point, the reason why I was in Shenzhen was to attend this tech conference.

The event, tbh, was underwhelming compared to the Beyond Conference in Macau this year, which was run by the same organizers.

Many of the exhibitors showed very little enthusiasm to showcase their work, while some were just scrolling on their phones and talking to friends. Maybe they could sense that I am not an investor, but they don’t know that I write to a couple thousand of you weekly…;p

The most fun bit was this small corner where startups could do their usual demo pitch, and then they could spin the wheel to match with investors who were there to assess their products. I listened in on a few pitches as well as the mainstage talks, which many focused on physical AI, spatial intelligence, and wearables as the next trend. Seems like the mainstream-nification of embodied AI has hit.

They’re still ‘going-global"‘

However, it was interesting to see that “going global” is even a bigger theme than 6 months ago. Echoing the themes at Beyond, it seemed like more and more companies are thinking of their global strategies.

In fact, a company like NOBODY, a startup focused on phones and wearables, adopted an interesting parallel strategy for the China market and the global market. Founded by Oppo’s co-founder Carl Pei, NOBODY is said to be HQ in London and selling to the European market sleek Gen-Z-sought-after smartphones.

Since they know that they don't have that pricing edge with smartphones in China, they focus on selling wearables like AI-empowered headphones, rings, and whatnot in China. It’s interesting to see how Chinese AI and tech companies are positioning themselves nowadays.

With founders and teams that are much more international than the previous generation, they are significantly less "obviously Chinese,” and many are leveraging their Chinese network of manufacturers and even R&D teams to develop two parallel business strategies/product categories, catering to both the China and non-China markets. This is another trend that will continue to grow I believe. The identity of “a Chinese company” will be more and more murky.

Global talent flow

Noted below is Wang Ning, Growth Lead at Nothing, to my point, a Tsinghua and Duke-educated businessman, with experience with McKinsey across “various locations” according to his LinkedIn. How do you define these tech firms purely by nationality/ geography now?

To my point in this article above about talent flow, AI and tech talent are incredibly global. It will be impossible to reverse the “globalisation of talent”, nor will it be beneficial for anyone.

The recent guest of my podcast, Yu Hang, Product Lead at Qoder, was another perfect example of this, where he lived/worked at Meta in Menlo Park for years before returning to China and is now based in Hangzhou working for Alibaba.

I acknowledge that this was just a short and less analytical update compared to our usual posts. But I wanted to take this opportunity to share with you all that I’ll be traveling quite a bit the next few weeks so that updates may come a little more sporadically.

Hit me up if you want to catch up in Singapore end of this week/ next week, and I’ll be going up to Shenzhen again next month to visit the Tencent HQ. Do you have any particular questions you’d like to ask them? Any particular interest in certain business verticals or products? DM me!

In the misty canals of Wuzhen, a county near Hangzhou, the six dragons all came out of their caves for this one big rah-rah. Probably due to a combination of encouragement from local authorities and pressure to share more ahead of IPO speculation and interest, the new generation of Chinese AI leaders laid out their vision for the next decade.

Dubbed the “Six Dragons,” these companies have been at the forefront of LLMs, robotics, and spatial AI. The founders and executives shared how they are navigating the transition from the internet era to the age of artificial intelligence, leveraging China’s vast data and manufacturing advantages while grappling with the profound risks and technical bottlenecks that lie ahead.

1. DeepSeek: The Open-Weight Accelerator Sounds an Alarm

“Since its establishment, DeepSeek has followed an open-source approach with the goal of achieving AGI,” said Chen Deli, senior researcher at DeepSeek. “Through exchanges with technical developers in the community, we have gained valuable feedback, which in turn has driven DeepSeek’s own development. Therefore, we believe that open-sourcing technology fosters better collaboration and sharing.”

DeepSeek, the LLM company that broke the internet, is staking its future on an open-source strategy in the pursuit of AGI. Senior researcher Chen Deli said this collaborative approach has been central to its progress, with community feedback sharpening the company’s technology.

He also struck a sober note on AI’s long-term trajectory, arguing that displacement could begin within five to ten years and that, over a 10- to 20-year horizon, AI could replace the vast majority of knowledge worker jobs.

Unlike past industrial revolutions, he suggested, AI may become the primary intelligence, potentially leaving humans without new roles to fill. While confident the technical problems will be solved, he said tech companies must also act as guardians of social order to manage a transition that could see humans “completely liberated from work.”

2. Unitree Robotics (宇树科技): Betting on an Imminent Breakthrough in Embodied AI

Unitree Robotics, a leader in legged robots (humanoids & quadrupeds), is betting that embodied intelligence is on the cusp of moving from sci-fi to daily reality. His thinking is: open algorithms + manufacturing depth = faster physical AI.

Founder and CEO Wang Xingxing voiced bullish optimism, arguing that robotics is closer to a breakthrough than more distant frontiers like nuclear fusion or Mars colonization, and that meaningful progress should arrive within two years. (Seems much more bullish than other industry players so far)

But he was candid about growing pains. Compared with large language models that benefit from abundant internet data, robotics lacks standardized data at scale. There is also no consensus on basic hardware principles, including where cameras should be placed on a robot.

Despite these early-stage hurdles, which he likened to the nascent days of AI models, Wang’s conviction is clear: cracking the general model for robotics is tantamount to achieving AGI. He credited China’s manufacturing ecosystem and the global open-source community for speeding the field’s pace.

3. Manycore Tech(群核科技): Open-source Spatial Data

Co-founder and chairman Huang Xiaohuang recalled 2021 as a pivotal year. With user growth and data compounding rapidly, Manycore open-sourced its spatial dataset, earning industry recognition and opening new opportunities in spatial intelligence.

Huang emphasized that more than a decade of China’s internet expansion created massive user bases and data, providing a solid foundation for training large AI models, and that spatial intelligence is becoming an essential direction for large-model development.

As the backbone technology for many robotics companies we’re seeing, Manycore may be pushed to the forefront now, especially as Feifei Li has just coined that spatial intelligence will be where we see the next breakthrough.

4. DEEP Robotics (云深处科技): Humanoid and Quadruped Robotics

Founder and CEO Zhu Qiuguo described the path from R&D to deployment. He shared a vivid glimpse into that evolution, noting that in 2015, their first robot dog had to be suspended by a steel cable because it couldn’t stand or walk on its own. The milestone of a freely standing dog in 2018 was followed by years of iterative development to address practical issues such as wear and tear and overheating for commercial applications. That persistence, he joked, has finally allowed them to evolve “from making dogs to making humans,” a reference to their recently launched humanoid robot designed to replace people in hazardous jobs.

This year, the company released a humanoid robot designed to replace or assist humans in hazardous environments. Zhu also pointed to industry-wide constraints: despite bold visions, like a robot soccer team beating humans by 2050, progress is checked by challenges in data collection and standards. (echoing what Unitree founder Wang Xingxing was saying) And added that, without external support, startups struggle with autonomous mobility, data capture, and the large compute required.

5. BrainCo (浙江强脑科技): BCI Moves from Lab to Clinic

Founder and CEO Han Bicheng highlighted a staggering statistic—approximately 36% of medical spending is linked to brain-related conditions like Alzheimer’s and autism, which currently have no pharmacological cures. He said the company is working to introduce a non-invasive solution to clinics to help solve this issue. Their neural-controlled prosthetics already allow people with limb disabilities to use “thought” to accomplish fine-motor tasks such as writing and playing the piano, and to restore a natural gait.

As brain-computer interface (BCI) moves from lab to practical use, the industry is also exploring ways to help blind people “see.” Han framed BCI as part of human-centered digitalization—not only making machines smarter but improving care for people.

Challenges remain significant: complex neural-signal parsing and high stability requirements, especially under complex signals. Even so, BCI may offer therapeutic potential for Alzheimer’s and other brain disorders.

6. Game Direct (游戏科学): Creative Gains and Warnings of Concentration of Power

Game Science, the studio behind Black Myth: Wukong, sees its success as proof of China’s vast domestic market and rising appetite for top-tier local content.

Founder and CEO Feng Ji highlighted a dual reality. On one hand, AI empowers individuals; some creators already use AI to complete lyrics, music composition, and video production that once required a full team. On the other hand, he warned of the risk of excessive concentration of technological power.

With caution, he echoed the broader call to embrace AI rationally while guarding against risks.

Common Thread

Across the six, the message is consistent: open collaboration, internet-era data foundations, and manufacturing strength are speeding progress, while embodied-AI data and standards remain real bottlenecks.

Their consensus view is that once key thresholds are crossed, AI will advance in leaps, and building trustworthy AI will require technology and regulation working in tandem, so society can effectively embrace the present and AI.

What’s comforting to see is that all of them acknowledged the societal risks AI could bring, despite being the builders of the technology. And whether they’re in model or physical AI, the shared belief is that open source/ weight technology is actually the better model and pushes proliferation faster.

If you take a step back, the Chinese AI companies embracing open-source are essentially subsidizing the common goal of boosting the Chinese economy, as no one has really figured out a clear monetization strategy yet. The ultimate winners of this are Chinese consumers and the economy as a whole, through productivity gains and unlocked business opportunities.

Side note: I was listening to a podcast and editing this post while sipping on coffee at FinePrint today, to my left sat two millennial women dressed in all pink debating why women should even marry and isn’t traveling much more fun, and to my right were two older men talking about creating a demo and pitchbook for a robot protype that will clean and act as a security guard for office buildings at night. I turned off my podcast and eavesdropped for a solid ten minutes, and it was like listening to the two versions of myself talk past each other.

Btw, provides the full transcript of the event here:

Geopolitechs

Full Transcript: Six Little Dragons Wuzhen Dialogue

Wang Jian (Moderator): I’ve always been curious — what’s the connection between the “Six Little Dragons” and the World Internet Conference…

Read more

5 months ago · 8 likes · Geopolitechs

And some relevant reads FYI:

• China’s leading AI startups: ‘six dragons’, Hangzhou, adoption of AI in fashion designing, and DeepSeek mythbusters

• China’s AI Startups Pivot Amid DeepSeek’s Rise

• Rise of China’s Robotics Industry: from Manufacturing Arms to Embodied AI

• Introducing Unitree, China’s leading AI-embodied Robotics Company

• The Jevons Paradox in AI Infrastructure: DeepSeek Efficiency Breakthroughs to Drive Energy Demand

• Everybody Losing Sleep Over DeepSeek: Industry Implications to LLMs and AI Infrastructure

• DeepSeek V3 puts China AI on the global map: consumer use and capital expenditure implications

I recently joined a panel to discuss the intersection of AI and communications at CityU University in Hong Kong, alongside industry experts with diverse fields of expertise—a software developer, a lawyer, an educator, a policy analyst, and others.

One of the key points I brought up was the framework I walked everyone through, which outlined how to think about AI and its impact on the future of the professional field of communications. The focus is not on how AI is currently affecting our way of communicating and working, but rather on how agentic AI will disrupt our thinking about knowledge work as a whole.

Share

AI x Communications

Let me break down my framework.

Phase 1: Distribution as a Bottleneck

Twenty to thirty years ago, the bottleneck for companies to convey their message to the public or any stakeholder was primarily in distribution. If we use Hong Kong as an example, and I wanted to obtain information about my company’s performance for Wall Street, aside from direct channels of communication or through bankers. I would likely need to know the reporters at The Wall Street Journal or The New York Times who cover Asia.

Now, newspaper real estate was truly scarce as well, so as the company, you’d have to fight to sell the most interesting story, or what the reporter then deemed as the most important story for people sitting on the other side of the world to know. The distribution was the true bottleneck for information dissemination.

Phase 2: Rise of Digital Media

Then, let’s say about 10-15 years ago, we saw the peak of digital media platforms like HuffPost, BuzzFeed, Vox, Vice, and so on. They were able to cater to a broader audience, often younger than the traditional outlets. They carved out a niche for each outlet, and in some ways, you can say “lowered the standard” for stories to make it in a publication. Not to say these stories were less worth knowing it’s just that there was a lot more real estate for more stories to be told. However, what it led to was an overabundance of information and content, and the creation of what we now know as echo chambers of worldviews. Their business models couldn't be sustained, and now most of these platforms have shut down.

Phase 3: Rise of False Information and Need for Governance

Then we saw the rise of social media and the internet platforms themselves becoming news sources for the majority. Misinformation and disinformation became rampant, but on the other hand, there was a rise in citizen journalists. Information access was no longer an issue, and distribution no longer had a high barrier to entry. Expertise and the need to dissect information became scarce. And throughout these times, PR agencies revamped themselves to follow the trend, from helping companies access journalists and vice versa, connecting reporters to potential story leads, to then focusing on building out relationships and expertise with “trade media” or “regional media,” to then pivoting from pure corporate messaging to offering “digital solutions.”

The other big issue that arose from the democratization of information dissemination is the increasing number of false information and misinformation, and a recent study by the Reuters Institute at Oxford University showed that there was an enormous discontent from news consumers regarding the rampant spread of misinformation; however, most believe it is the responsibility of platforms to monitor, instead of the government’s. However, each company is led by a person and/or sponsored by a pool of capital backed by a party, which has led to the extreme political polarization of media and tech platforms.

Phase 4: AIGC Floods the Internet

This leads us to the current phase, where we’ve seen major tech companies implement content moderation processes, including both human content moderators and AI-powered moderation tools. However, on the other hand, content creators, whether professional or casual users, have also adopted AI for content creation. While there is no definitive percentage of the amount of content online that is generated by AI, some studies suggest that nearly 57% of all online text content is AI-generated or translated. An Ahrefs study found that as of early 2025, ~74% of new webpages included some AI content. And it’s not just the normie consumers, even the media experts sometimes cannot tell the difference between AIGC and real photography/ video/ content. (Much, much different from when I first covered deepfakes in 2019 - oh god, I am so embarrassed to link this, but here is a video of baby reporter me)

And in my own estimate, I think we can easily see AI automating at least 30% of the work that writers, researchers, and PR professionals took hours to do within minutes. I’m not saying they are at the top-notchh quality, but they can do the work. So, suddenly, media monitoring, interview briefings, research packages, press releases, callnotes and emails can be completed within minutes. We’re seeing that the ability towrite is no longer as vital, but to write well and write to influence is more critical than ever. However, to achieve that and avoid appearing like generic AI writing, industry expertise is necessary. We need more professionals who previously relied heavily on their writing capabilities to be now able to carve out their expertise. If you work with tech companies, you should have a basic understanding of technology. If you work with consumers, you need to know consumer companies inside and out. The execution bit is less difficult to accomplish, the strategic and critical thinking bit is more important, and that leaves many junior staff in an awkward position.

Phase 5: Coworking with Agentic AI

However, we then look forward to an Agentic AI phase, where, in 2-3 years or even less time, we may have agents embedded in customer service (if not already) and sophisticated work processes. Then, what will be vital is the ability to prompt and manage agents to complete tasks correctly. Additionally, it will be essential to train the agents to respond to different stakeholders and each individual interacting with them in a customized and tailored fashion. And more importantly, people who have industry expertise and can command AI tools.

In essence, we’ll need more “operators” or “project managers” and even “insight leaders” but fewer “executioners.” What we’ll likely see in this industry is the use of agentic AI to eliminate toil and keep humans in control for narrative, risk, and trust.

Subscribe now

The Future We Live In, Amongst Agentic AI

So that leads me to the bigger topic of today, the future of coworking with Agentic AI.

To start, this is how the smart smart researchers at Bernstein defined Agentic AI:

Agentic AI is the space that translates the LLM’s immense potential into autonomous action. The expectations from AI have already shifted to be outcome-driven - performing multistep, complex tasks like scheduling/prioritizing patients basis urgency after a conversation, booking tickets or assisting radiologists with diagnosis.

This ability to act, learn and adapt going beyond an LLM response is what makes Agentic AI the true frontier to watch out for, if one were to assess disruption due to AI.

Awaiting the Big Breakthrough

Agentic AI has been the talk of the town for the last 2 months. Every company is saying, “We have agentic AI". However, most layman consumers have not been “wowed” the same way when even the first ChatGPT consumer-facing app came out in late 2022.

The reason is well, for one, we humans have high expecations and become harder and harder to satisfy when fed with beter things - think about your taste buds, I took my husband back to my favorite thai food hole in the wall from my uni days and honestly it tasted so bad I couldnt even finish my meal, because since then I’ve been fed better food and my taste buds have become more and more snobby.

And two, we haven’t really figured out how to use agentic AI to really co-work/ collaborate/ unleash its potential yet.

How to understand the relationship between agentic AI and generative AI? Think of generative AI as the “brain” and the agentic AI as the “execution.”

An example, whether it’s your ChatGPT or your DeepSeek, it was able to plan you a perfect 5-day itinerary for Osaka that is toddler-friendly and combines scenery, history, and modern city culture. That is what the Generative AI could do for you. What the Agentic AI can now do is actually help you book the car, book the restaurant, and the tickets for the historical sites. (like a personal assistant)

For most, the agentic AI sphere is currently restricted to the human assistant stage in its evolution, according to Bernstein: “it automates most of the mundane back-end tasks, handles low-stakes customer conversations, and aids an expert for high-value tasks with human touchpoints. It is not replacing humans across the value chain, but rather unevenly - highest impact is in jobs like data entry, HR/IT support, administrative staff, junior sales representatives/software engineers, and customer support.”

Share

That said, while it has not become fully autonomous, it does threaten a major disruption in certain IT, Customer Service, and Sales roles.

In Bernstein’s recent report, AI vs. Human: Agentic AI: What will a Billion Agents change? 100 Startups show the way, the researchers tried out 100 Agentic AI apps, and they summarized that agentic AI’s current evolution is best understood as three practical “leaps” that map neatly onto knowledge work:

1. Truncating the Toil – automating routine, high‑volume tasks (data entry, simple inquiries).

2. Structuring the Unstructured – ingesting long documents/web pages/calls/news releases and turning them into queryable, structured data.

3. Assisted expertise (human‑in‑the‑loop) – tackling more complex tasks (drafting, diagnosis, outreach) with explicit human checkpoints.

The conclusion?

The review of ~100 agent startups shows most systems are still in “human assistant” mode, automating back‑office toil and aiding experts, rather than replacing entire workflows end‑to‑end.

And earlier, I pulled a number out of a hat, saying that AI could replace ~30% of knowledge workers' tasks right now, well, McKinsey (probably more scientific in their methods than mine) estimated that to be much higher. In fact, McKinsey estimates current gen‑AI can technically automate activities that absorb 60–70% of employees’ time, so agentic AI can be assumed to actually complete even more tasks and preserve employees’ time.

And plugging in this interview with Diana Wu David, who is the Director of Futures at Service Now, where we talk about the future of work. I think it’s important to remind us all that, instead of thinking about job replacement, let’s think about how to manage and control AI better.

So... what I’m trying to say is, (and responding to some people upset about how companies will always be future proof and we humans get laid off), we should not be limited to thinking the only way to be “future proof” is to stay at an organization or a role forever. Instead of staying future-proof, being adaptable to the future workforce. Organizations take months or years to implement new processes and strategies. But for each individual, it’s a personal choice and one that can be made much more swiftly.

If my father’s generation hadn’t learned to use Microsoft Office or how to better leverage the internet to communicate more efficiently, he wouldn’t have remained relevant in the workforce. Environments evolve, societies change, and the companies that stay relevant also change. It takes one’s effort to change with the times, too. In the mid-term at least, it’s not “AI replacing us,” it’s people who can leverage AI to better optimize their existing workflow/ demand that will “replace us” (of those who are complacent to learn and adapt).

​​The Agentic breakthrough has not come yet, but it is coming.

I know there has been a TON of AI (LLM) news from DeepSeek, Alibaba, and Tencent. I will provide a deeper analysis later in the month once the dust has settled. And coming to you on Thursday is an episode of Differentiated Understanding with an AI investor on Neoclouds, the VC bubble, and the capital race in AI. Stay tuned.

I’m no gamer myself, but my associate braved the crowds at China Joy in Shanghai two weeks ago. And surprise, surprise, AI was everywhere.

Subscribe now

For those unfamiliar, China Joy is famous not just for its gaming showcases, but also infamous for the “uncle problem” a few years back—middle-aged men crowding around cosplay booths and taking creepy photos, prompting a well-deserved public shaming. But beyond the cosplay drama, it remains China’s largest gaming conference, held just days after WAIC 2025, and still draws massive foot traffic.

That said, a casual stroll through the exhibition halls was overcrowded and visually stimulating. Still, according to my contact on the ground, it might have left an avid gamer somewhat underwhelmed.

Yes, there were AI-powered gameplay demos, but no “this changes everything” moment. No jaw-dropping mechanics. No booth where you thought: Ah, this is the future of gaming. Not yet.

However, if you look closer, we did see how AI is powering changes in the 2B development of gaming.

China’s Gaming Industry For Context

To start, despite headlines about the gaming industry crackdown, China’s gaming industry is said to have reached RMB 168 billion (USD 23.4 billion) in revenue within the first 6 months of 2025, according to a report by the Game Publishing Committee and the Expert Committee on Game Industry Research. That’s a year-on-year growth of 14%.

And the number of gamers also reached 680 million people, reaching a new high (again, a recovery from the dip in 2022-2023). And the most prominent players are Tencent Games and NetEase in the country, so we’ll look at what they’re doing in the space.

Tencent is the world's largest gaming company by revenue and a significant player in the global gaming market, with a portfolio that includes popular titles like Honor of Kings, PUBG Mobile, and investments in companies like Riot Games and Supercell. While NetEase is more known for its mobile and PC games, including Identity V and Fantasy Westward Journey. And then there’s Baidu trying to stay relevant in the gaming space after its huge department cut a few years back. Competing on different fronts, they’re adopting AI in various ways.

2B: Where the Real AI Action Is

At the developers’ meetings, it was made clear that AI isn’t just tweaking game features; it’s starting to rewire the entire game development stack.

Baidu’s Junjie Du (VP, Baidu AI Cloud) put the size of China’s gaming industry at 300 billion RMB during his interview with Gaming Daily 游戏日报 and sees AI as a growth accelerant. In his view, Baidu’s role is not a game developer, rather it is as a full-stack technical partner, offering:

• Infrastructure Layer: GPU-accelerated compute for model training, with Baidu’s Baige AIHC platform improving utilization efficiency—key when GPUs remain expensive and scarce.

• Model Layer: The Baidu Qianfan platform, which blends ERNIE and leading open-source LLMs, plus fine-tuning services. Case in point: a partnership with 37 Interactive Entertainment on AI customer service.

• Application Layer: AI tools for text-to-image, text-to-video, and code generation, as well as advanced capabilities like text-to-3D modeling and text-to-skeletal animation—shortening production timelines and opening doors for AI-native game studios.

This “AI + Cloud” bundling isn’t just about selling compute—it’s about locking in developer mindshare before AI-first studios set the new baseline for speed and creativity.

2C: NPCs Get a Brain Upgrade

On the player side, AI’s influence is not like there’s been nothing; it’s subtler but growing. Although AI is not completely changing the 2C gaming experience yet, it continues to be deployed to enhance existing functions and features.

Case in point, just today, Tencent announced its 2025 Q2 earnings and its company’s executives highlighted that Tencent “deployed AI tools in games to accelerate content production, introduced AI-powered features enabling more realistic virtual teammates and non-player characters, used AI-powered marketing activities to increase user acquisition and engagement, contributing to the popularity and revenue growth of our Domestic and International Games.”

Which translates to faster content production, more human-like NPCs and virtual teammates, and easier-to-use AI-powered marketing to boost engagement and acquisition

AI-driven NPCs can now parse player intent, generate context-aware dialogue in real time, and adapt their behavior to in-game events—transforming NPCs from static quest-givers into dynamic participants.

NetEase’s Justice Online (逆水寒) takes this further, promising that no two player-NPC conversations are the same. Essentially, every interaction is personalized, enhancing immersion and replayability.

The High-Level Takeaways

• For 2B: The biggest commercial impact is in the production pipeline. AI is reducing time-to-market, cutting costs, and enabling game studios to scale output without scaling headcount. The real competition is in who owns the dev stack (cloud + models + tools).

• For 2C: While full AI-native gameplay hasn’t arrived yet, we can see that enhancements to NPCs and player interaction are already sticky engagement drivers. If this AI integration can improve, then it will likely translate into longer session times and higher in-game spend.

• Strategic Note: The companies that can marry 2B infrastructure dominance with 2C player-facing innovation will have the most defensible moat. Right now, Tencent has a distribution edge (WeChat ecosystem), Baidu has a full-stack cloud play, and NetEase is experimenting fastest with in-game AI behavior.

Why Open Weight Models Matter Here

The release of DeepSeek’s open-weight models has been a quiet but powerful accelerant for AI adoption in gaming. For indie studios, it means access to frontier-level models without the budget of a Tencent or NetEase, while making AI-driven NPCs, procedural content generation, and adaptive storytelling financially viable. For big publishers, it lowers experimentation costs and enables hybrid pipelines where proprietary IP is layered on top of open-source cores. In short, open weights reduce the barrier to entry for AI-native game design and will potentially compress the innovation cycle from years to months.

Thoughts on Chinese AI Talent

A piece published in the widely followed WeChat blog “The Intellectual” (知识分子) prompted me to reflect on the perception of Chinese AI talent. The key points the author made include:

• DeepSeek’s research team is relatively young and inexperienced in external views. Still, they have achieved significant academic milestones, with numerous published papers, high citation counts, and a solid h-index value, demonstrating consistent research output across the team. The core 31 authors demonstrate even higher academic metrics, not driven by a few stars, but by a broadly capable group with many having h-indexes above 10.

• Over half of DeepSeek’s authors were educated and employed entirely in China without overseas study, showing intense domestic talent cultivation mainly centered on institutions like the Chinese Academy of Sciences and top Chinese universities.

• As many have cited that over ~40% of AI researchers globally are Chinese, that definition is still vague. However, this author notes that many Chinese AI researchers now have US education or research experience, but often return quickly to China to work, indicating a reverse talent flow that differs from previous U.S.-dominant patterns.

• Resulting in his key point that there is a shift in global AI talent flow. Unlike earlier trends where Chinese (internet and) AI talent moved to and stayed in the U.S., many are now returning to contribute to China’s AI ecosystem, signaling China’s rise in AI innovation and talent retention. And this is a point I’ve made here to the leading Indian tech publication AIM as well.

Look, even I held ignorant views myself before moving to China and living in this part of the world for over a decade now. However, for the generations of the 80s, 90s, and even the 00s, cheaper labor is not a competitive advantage. (Heck, they’re not cheap.)

Anyway, it’s no news that ByteDance, Tencent, and Alibaba heavily recruited from Silicon Valley and then from each other and other Chinese research labs, shelling out big bucks - not the 100 million USD cheques Mark Zuckerberg is writing, but sizable ones like ~10-50 million USD for the top top top talent.

Unlike these big tech companies, DeepSeek and some other labs have committed to primarily hiring from leading domestic academic institutions directly, wanting them to be green and uncorrupted. Amid all this, the AWS AI lab in Shanghai was dismantled last week, and its leader, Wang Minjie, posted on WeChat that, due to geopolitical reasons, the research lab had been shut down. Noting that despite this setback, he’s optimistic because this is the golden era for AI research and talent. Companies can feel free to reach out to learn more about his team and recruit from the talented roster.

And in other news, Technology Source Share (技术源share) reports indicate that Manus’ official relocation to Singapore resulted in only 40 out of its 120 employees being relocated. And its attempt to “China shed” will mean it’ll have to hire locally in Singapore and even globally.

That’s why I’m covering an event that I think is quite interesting, showcasing the phenomenon in China: the unleashing of raw talent and increasing ambition. There still seems to be some misconception or prejudice from the West that thinks Chinese talent is imported or cheaper somehow in the AI and tech space. But what I want to highlight is that there is actually a lot of energy coming from the youth, and DeepSeek truly ignited that hunger again in the community.

I’m not saying the overall domestic consumption issue in China has been solved, nor is the property price drop no longer a drag on sentiment. BUT, I must say, it is exciting to see so many young people come together, energized and actively looking for ways to build.

With that...

AdventureX

What has been called the “biggest AI hackathon” in China took place this past week. The AdventureX Hackathon, held in Hangzhou (home to DeepSeek, Alibaba, NetEase, Geely, and many other notable companies), was organized by two high school graduates and a group of friends. What makes it special is that it targets empowering young talent, and the team of youngsters that founded the event used nine months to raise one million RMB in funding, as well as secure a free venue, accommodation, and meals for over 100 tech investors and leaders.

For its inaugural event in 2024, over 200 participants came together, with 90% of the hackathon participants being students, ranging from high school to graduate school, and an average age of ~22 years old. This year, the participants were capped at 26 years old, with the youngest participant being 11 years old. Nevertheless, nearly 900 young programmers, developers, and entrepreneurs came together.

The core founding team, including Ryan Zhu (also known as Zhu Yuhao), Max Liu, Marvin Cui, Christina Lee, and Kayna Huan, along with their group of friends (something like 20-30 teenagers), organized the entire event.

In an interview, it was said that on the day before the most important day of their lives thus far - the Gaokao, the national college entrance exam - the founders realized there were over 4,500 sign-ups from more than 182 countries and regions. And it was that moment they realized they had found their calling.

The group of young enthusiasts regards this sentence as a guiding principle and has made it the signature slogan of AdventureX, attracting like-minded individuals. It comes from an early interview clip of Steve Jobs, the tech icon who was also a geek in his youth. Ryan told reporters: “Everything around you that you call ‘life’ is made by people who are no smarter than you.”

“Once you realize this fact, life becomes much broader.” He encouraged young people to break the rules, not to accept fate passively, and to leave a mark during their short lives.

These are some candid video shots of the venue and exhibition halls that my contact, who has gone to Hangzhou to scout young talent, shot.

She said it was truly a bustling hub where the latest innovations in cloud services, AI, innovative hardware, and even some Web3 ideas converged. Overall, it was a mixed bag—while the buzz was definitely high, the number of truly standout projects was a bit smaller than expected. Still, a few cool innovations that caught her eye that show where things are heading.

Below are some key takeaways shared by my contact, an AI business leader from Singapore, as well as some consolidated points from various WeChat blogs, including Distill AI, Silicon Person Pro (硅星人Pro), and Nine Thousand Light Years (九千光年). Startups of all kinds showcased their demos, many obviously still a bit gimmicky, and do not exhibit a clear business strategy yet. However, my point for this article today is to highlight the enthusiasm in this space and some emerging trends.

First things first, there was a vast array of products, and the organizers themselves even created several applications to assist with the event's logistics and management. A chatbot helped answer inquiries regarding the event. A tool could help categorize emails automatically. A love match robot helped participants find romance based on their interests, age, and preferences, among other applications.

Brain-Computer Interfaces In Mental Healthcare

Noted by Distill AI, one exciting company that stood out is Emotiv. A company that is working on brain-computer interfaces (BCI). Consider gadgets that can read your brainwaves using EEG technology to track your level of focus and stress. What’s neat is that they’ve moved beyond just lab demos — their products are ready for everyday users. As mental wellness becomes more critical, these kinds of devices could be a game-changer for AI in healthcare.

This is definitely a growing trend in this space. I met with a company that was similar to this one at Beyond Macau as well, where the product uses a camera to read facial expressions and detect various levels of stress, happiness, anger, and other moods. It is reportedly being trialed in Shenzhen clinics for psychological well-being.

AR Glasses with a Practical Twist

So Meta glasses have been the dominant mainstream wearables in Silicon Valley, and now everyone’s waiting on Apple’s Vision Pro. My husband got a pair of the Ray-Ban Meta glasses to snap pics of my running toddler; I don't know how to feel about it, to be honest. The leading Chinese AR glass maker, Rokid, has made headlines across CNN and Bloomberg; others, such as Thunderbird and Xiaomi, are also pushing ahead with wearables.

recently wrote an interesting LinkedIn post, saying that AI glasses sit at the intersection of three drivers: 1) pressure to commercialize large models; 2) search for new endpoints in hardware; 3) quest for a post-smartphone interaction interface. Despite Snap’s failed attempt (perhaps they were too early), wearables like glasses (whether you like it or not) will likely play a much bigger role than they are currently perceived to.

EvenRealities, a Chinese startup presenting laser-projection AR glasses that are lightweight and priced more affordably, is not yet fully AI-defied, but it’s a display of what it can do. These glasses show text in real-time, help with translation, and provide navigation right in front of your eyes. The design is lightweight, straightforward, and priced realistically. This practical approach might help AR catch on before the big-name brands fully crack the consumer market.

AI, the Fortune Teller and the Companion

In many ways, we’re seeing companion bots emerge. And they’re now becoming more specific. FateTell showcased an AI-powered fortune-telling app. Utilizing large AI models, it provides personalized destiny readings and interactive divination games. Sure, it’s not about being scientifically accurate—it’s more about giving users an emotional experience and some fun interaction. This shows that AI isn’t just for knowledge workers' efficiency gains, but it could potentially be adopted to meet people’s emotional needs in creative and unexpected ways.

This piece is similar to the Hangzhou startup and fashion integration piece, where I explored a specific type of vertical use case of startups making a splash in the APAC/Greater China ecosystem.

[This was NOT a sponsored piece. However, I did gain access to the two founders featured below, allowing me to delve into more detail about their businesses. This is not a product recommendation, nor is it investment advice. I want to use these two startups to showcase a trend in the APAC AI space.]

As we’ve written quite extensively, there’s a massive cohort of Chinese AI companies where the founders are Chinese but they may have been educated in the West and/ or worked in the U.S. that are now setting up shop in Singapore due to mostly the reasons below:

1. Access to tech talent pool;

2. Desensitized geopolitical market;

3. English/ Mandarin language advantage;

4. Favorable government incentives and policies;

5. Relative proximity to China and more affordable talent in SEA;

6. Access to international capital.

As RockFlow and CapGo.AI are companies that have recently generated some buzz, let’s use them as examples to examine how they are building vertical AI apps and leveraging Singapore as their headquarters amid the ongoing tug-of-war between China and the U.S.

Subscribe now

Agents for Vertical Use Cases

There is an increasing number of agents being developed on top of LLMs, and companies are finding ways to build handy tools in specific verticals.

RockFlow is a Fintech, an AI-first trading App.

Vakee Lai, Founder and CEO of RockFlow, said to AI Proem: “We chose to build an Agent in this space because finance is a domain where precision, speed, and a minimal margin for error are essential. To truly scale AI in trading, an Agent must deliver not only insights but also real execution and reliable outcomes cost-effectively and efficiently.”

RockFlow’s mobile brokerage is built around what the company calls a “proprietary agent architecture,” a layer of LLM components trained on financial data that sits between the user and the market. The software combs more than 100 live data feeds, prices, news, social sentiment, and a customer’s trading history, to turn plain-English commands such as “Sell all my Apple” or “Find a seven-day baby-bear spread on Nvidia” into executable orders across equities, options, and crypto ETFs.

“At RockFlow, we see investing as a way of translating our observations of the world into strategy. That’s why we built Bobby as a unique multi-agent system that manages the entire trading lifecycle end-to-end, so that anyone’s observations can become actionable strategies, and investing becomes radically simpler,” Vakee added.

[See Demo video of Rockflow: “I think RWA is going to be huge, what stocks should I gain exposure to?”]

The firm’s first generation of that stack, launched in 2023 as TradeGPT, produced idea lists and simplified option plays (“Baby Bull” and “Baby Bear”). The hype this year surrounds its product, named Bobby, a 24/7 AI “investment partner” that promises to move beyond question-and-answer chatbots by rebalancing portfolios, building complex spreads, and monitoring risk in real-time, all within the same app interface.

RockFlow also lets users hit a “one-tap back-test” button to summon an AI Portfolio. The system combines the firm’s in-house quantitative factors with macro signals and a user’s risk profile, then outputs historical performance, Sharpe ratios, and a launch-ready basket. A forthcoming marketplace will enable creators to charge copy-trading fees, effectively turning their strategies into storefronts.

Another enterprise use case of AI is CapGo.AI. It is an AI marketing company focused on programmatic SEO and GEO. It uses AI to generate SEO-optimized webpages at scale in multiple languages, which then help businesses rapidly scale their online presence globally.

“We saw a huge opportunity where AI could radically scale webpage creation. At the same time, the rise of AI search (like Google's AI mode and ChatGPT) is reshaping how traffic is distributed online. Businesses need new ways to stay visible, and Programmatic webpage creation is a powerful strategy to adapt,” Yichen Guo, founder and CEO of CapGo.AI told AI Proem.

And there are many more startups like such. However, today’s story isn’t just about AI being used in the financial sector or for managing business efficiency. It’s to highlight a bigger trend that is reshaping APACs startup ecosystem.

Share

New Generation of Chinese AI Entrepreneurs in Singapore

The RockFlow story is a prime example of a rising trend in the AI space in the APAC region, where the HQ is set up in SG, the talents often come from China, and the revenue and customers are from around the world (esp. in the West).

Let’s go back to the beginning. RockFlow was incorporated in Singapore in July 2021 by Chinese founder Lai Yunqi, also known as Vakee in English, a former quantitative derivatives trader in London, product manager at Baidu’s search ad unit(Fengchao), and later an AI-focused venture capitalist. She holds an accolade of titles, including Forbes China’s 2025 “30 Under 30” list in the finance and venture capital category. This has helped raise the profile of a company that today counts approximately 120 engineers, quants, and designers spread across five offices.

But Vakee’s story is not an isolated incident. As introduced above as well, CapGo.AI represents a similar wave. We’re seeing the well-embraced trend of Western-educated Chinese AI founders and builders essentially using Singapore as a launchpad for their businesses, allowing them to have their cake and eat it too.

Yichen Guo, who was born in Beijing and educated at Berkley and Harvard Business School. Had a life as an investment banker in New York and worked in various tech roles in the U.S. before uprooting his life and moving to Singapore to found CapGo.AI.

He said Singapore serves as a strategic location: “Many of our clients are Chinese companies expanding abroad. They often have stronger products and cheaper pricing, but completely unfamiliar with global digital marketing. That's where we come in—bridging the gap with newest AI SEO solutions that get them found by international audiences."

Adding that, “Singapore gives us access to both worlds. On one hand, it’s a hub for Asian clients—especially Chinese companies with global ambitions. On the other, it’s deeply connected to the Western tech ecosystem.”

And about a year ago I spoke to him in detail about why Singapore was a clear winner when he was selecting a HQ location for his AI start-up.

“Singapore is an attractive place for many founders to set up entities targeting the global market because it shares the same internet ecosystem as the rest of the world, has good internet connectivity, and serves as an easy-to-access travel hub so many Singaporean AI companies can easily adapt a global roadmap.”

The truth is, they are one of many of these AI companies that want international capital as well as Chinese investment, but the increasing regulations haven’t been easy for them. Setting up shop in Singapore is not so much a “China-shedding” PR strategy, but rather a strategic operational decision.

“Chinese AI companies are desperate for foreign capital that they simply can't get at home. But with the current political shitstorm between the U.S. and China, setting up shop in Silicon Valley is basically asking for trouble,” said Keith Yap, the host of the Front Row Podcast.

Appreciating his candidness…

“Singapore feels like the perfect compromise. Close enough to China that founders can hop on a plane for a quick meeting in Shenzhen, far enough away that the U.S. VCs don't worry about geopolitical issues,” said Keith who regularly interviews guests in economics, business, and politics across Asia and brings together their perspectives on regional issues, spoke to AI Proem.

Anecdotally, he added, “I know of a few friends/ Chinese founders essentially commuting between their ops headquarters in China and their fundraising headquarters in Singapore. It's geography as a service.”

And since a significant number of these AI startups are founded by people who have worked at the Singapore offices of the likes of ByteDance, Alibaba, and Tencent. The teams, by nature, are a mix of Chinese employees, local Singaporeans, and talent from across the world, and the start-ups “feel Singaporean but also have Chinese DNA,” Keith added.

Global Appetite: Business and Product Strategy

Legal and Capital Structure Advantage

Kelly An, who has advised many AI companies’ global GTM strategy, told AI Proem:

“The Monetary Authority of Singapore enforces a 'same activity, same risk, same regime’ that closely tracks U.S. and EU standards, so licences and compliance expertise travel well.

The city-state’s bilingual talent pool works comfortably with RMB and USD capital flows, creating products that feel native on both sides of the Pacific. Its geographic position also offers a time-zone sweet spot that lets engineers ship overnight for U.S. product leads and demo to Asian clients the next morning, compressing the launch cycle. Together, these advantages produce fintechs, RockFlow included, built for global audiences from the very start.”

Challenge Remains in Willingness to Pay

Although Singapore has rolled out encouraging policies, offers lower taxes (compared to the U.S. and Mainland China), and a more business-friendly environment, companies setting up shop in Singapore are not problem-free.

“The biggest challenge for AI applications in Singapore is that the market size of the region (Southeast Asia) is still relatively small and fragmented. Software businesses, in general, thrive better in developed countries with a higher willingness to pay,” said CapGo.AI’s Yichen.

The reality is that the willingness to pay for enterprise and consumer software across Southeast Asia is still relatively low, even compared to Mainland China. For most founders, their eyes are on the prize, which remains in the developed software markets of Western Europe, North America, and parts of East Asia.

Case in point is RockFlow’s localization strategy. Its app is offered in nine languages and has users in more than 30 jurisdictions, with India, Japan, Vietnam, Mexico, and Taiwan next on the 2025 roadmap. And in addition, Customers can trade U.S. and Hong Kong equities and options, as well as 11 spot-Bitcoin ETFs and the Grayscale Ethereum Trust.

All things considered, Singapore remains a sweet spot for Asian enterprise apps trying to sell to the West, as its legal system follows the English common law tradition and is known for its transparency and judicial independence.

Vertical Competition Heating Up and ROI Still Unclear

Now, RockFlow is hardly alone in fusing AI and brokerage. U.S. rivals include Magnifi, Trade Ideas, and Robinhood’s experimental AI-chat features. Meanwhile, outside the United States, eToro relies on social copy-trading, while Wealthfront automates asset allocation. But most competitors stop at advice or alerts. By embedding an execution engine within its LLM layer, RockFlow aims to close the loop from idea to settlement - the AI element.

Meanwhile, Yichen said that the real challenge is in competition. “U.S. players are more product-driven; Asian firms lean toward traditional, service-heavy models.” Although, he prides the company for being able to bridge the two: combining AI automation with hands-on support, it also means that the team needs to be better and stronger on both fronts.

‘The Singapore Strategy’

Turning novelty into revenue is probably the next big test for many of these vertical apps. For these early-stage companies, it’s often challenging to determine whether there is an ROI yet. And obviously, an increasingly hostile relationship between China and the U.S. makes everyone nervous.

However, the optimism holds for many of these startup founders, who hope that by finding the right product-market fit, setting up shop in Singapore will enable them to propel their innovation to the global stage and prove to investors quickly.

As many have pointed out, “the Singapore strategy” seems to have worked out ‘O-K’ for ByteDance, Tencent Games, Lazada, as well as mid-sized Chinese heritage tech firms. Following the lead of the major players, they’ve established separate legal entities and management teams, localized operations for the APAC markets, and even global operations.

Keith, who has spoken to many top business and policy leaders in Singapore, said the business sentiment is welcoming in Singapore towards Chinese founders. And he believes “it's a smart business strategy for serving different user bases with different needs…You need different go-to-market strategies for Western versus Chinese markets, different compliance frameworks, and sometimes, different product features.”

“Singapore provides the infrastructure and regulatory clarity to set up these international operations properly, whilst maintaining the operational flexibility to serve global markets effectively.”

And he joked, despite what some outsiders might think, it’s actually not a tax break story, “if they wanted tax breaks, they can go to Dubai. Singapore solves real problems that Chinese AI companies can't solve anywhere else: access to networks (without political baggage). He said, ultimately, it's about Singapore being "China-adjacent but internationally credible enough for them to set up shop here.”

This has brought to my attention the rise of creative tools that have been receiving tremendous attention since the integration of AI. With Canvas being obviously at the forefront of adoption as an incumbent and new challengers like Midjourney and OpenAI empowered Dall-E in the West, and in the Chinese market, there is an increasing number of players in this space too, with ByteDance’s Dreamina (CapCut), Kuaishou’s Kling, and Meitu being strong contenders fighting for the preference of creatives.

And not to go unnoticed, OpenAI officially launched the image generation feature of GPT-4o in late March, which led to a flood of Ghibli-style images on social media. And Google launched Veo 2, which can generate high-quality video content based on text prompts, supporting 4K resolution and a variety of cinematic lens effects. In addition, Google released Gemini Flash 2.0 Image Generation, which also integrates an autoregressive model, significantly enhancing the usability of the model. The AI for Creative space is getting flooded with cool new tools and is setting itself up to be the next frontier technology.

[In this article, we will focus on the business of AI for Creatives, but I do want to note that the lack of regulations around image compliance and fake videos should be discussed, especially as we see text-to-video/image capabilities becoming more advanced by the day. If any safety/policy experts read this and are interested in contributing to AI Proem, please get in touch.]

China’s Creative AI Space

A closer look at China’s first choice for creative AI adoption reveals a space that has been extremely competitive. Just this past week, Kuaishou released Kling AI 2.0 globally, which includes a new video model and an image generation model. Kling’s multi-modal visual language editing capabilities enable users to input ideas through text, images, and other formats. Based on Jefferies’ report, image-generated video accounts for about 85% of Kling AI video creation, and over 15,000 developers and business clients worldwide are already using Kling’s API in different industry scenarios, generating 12 million images and over 40 million videos to date.

Following suit, Alibaba also released its latest open-source video generation model, Wan 2.1-FLF2V-14B, which is designed to simplify video content creation by accepting input of start and end frames. Creators can now develop their own AI models and applications at no cost and create short videos (5-seconds) at 720p resolution for free on the official website.

The big tech companies all have some form of a tool that they are trying to develop in the text-to-image/text-to-video space, from ByteDance’s Dreamina to Baidu’s Ernie, which is also attempting to do this in some capacity, as well as a handful of startups that are emerging. But the clear market leader right now is Kuaishou’s Kling.

We’ve previously compiled a list of Made in China AI apps, which you can reference.

Subscribe now

A focused monetization strategy: quality over quantity

However, whether it’s Kling or Midjourney, their ultimate user isn’t just the average knowledge worker, nor is it the average consumer. Unlike Yuanbao and Doubao, which adopt a consumer-facing approach, trying to catch them all. Creative AI tools are really focused on enhancing the workflow for professionals, and there are two main reasons for this. I spoke with the China market head of a stealth AI company in this field, and he explained the economics to me in the following manner.

Most users of these tools are advertising professionals, digital creatives, architects, designers, and so on. Although the base may not be huge in sheer numbers, they are willing to pay out of their own pocket if it means increasing their productivity.

Unit Economics

Whether it is his stealth AI or is it Kling (Kuaishou) or Dreamina (ByteDance) in this lane, a key difference in AI diffusion and monetization that sets it apart from the internet era is the unit economics. The logic is like this: the cost of maintaining a service for 1 or 100 users is the same for an internet product. But, the cost of providing services for 1 or 100 users in the AI world means it is quite literally 1x or 100x more. So companies need to figure out how to actually get their bang for their buck on each user, not just boost the sheer DAU number.

This also largely aligns with our analyses of the enterprise space, where vertical industry know-how is more valuable than overall market size.

Share

The video content industry is extremely commercialized and has a mature business model. According to industry estimates, ByteDance (the entire company, not just TikTok and Douyin) generated 400-500 billion RMB in ad revenue (~65 billion USD). This indicates there is a substantial demand for content advertising and an existing substantial market. Think about it, that’s ByteDance alone. Even if it’s a few percentage points of that market, it can still be a huge business.

For users, it's all about efficiency and achieving the same revenue with lower production costs. And AI companies like them will not help sort out ad placements, but the content creation cost segment is extremely price-sensitive.

Performance Marketing

For example, say if you can cut costs from $1.9 per 10-second video to $1.5 per 10-second video (even that small of a difference), that will make them want to make the switch. The user is a prosumer who will pay for the service and then produce something new with it. You can be either a company’s own marketing department or an agency that represents them.

He added that, in a market like China’s, most companies prioritize performance marketing, meaning they are concerned with the value or, in some cases, user conversion they can achieve when they pay for an ad, rather than focusing on brand value or creative aspects. So, that market is still huge in China. Ultimately, it’s not about the DAU of users, but the potential market size.

There seems to be a sense that the AI monetization strategy is becoming clearer, especially for specific industry cases. For creative-focused tools, it’s professional vertical adoption. The sheer number of DAU/MAU may not be high, but the price paid by a willing customer could still be quite large. And, look, the logic is that even a small percentage of the overall digital ad space market can be a good enough business.

According to Statista, ad spending in the Advertising market in China is forecasted to reach USD 251.91 billion in 2025, with Search Advertising accounting for the largest market volume of USD 69.38 billion in 2025. Specifically, in the Chinese advertising market, 89% of total ad spending is expected to come from digital sources by 2030. In comparison to the U.S., it is expected to lead in ad spending, reaching US$455.93 billion in 2025.

Subscribe now

Zooming in on Kuaishou’s Kling

ByteDance’s China short-video app, Douyin, primarily captured the domestic first-tier and second-tier cities. Its biggest competitor is Kuaishou domestically, which is traditionally the most popular short-video app in third-tier and fourth-tier cities.

Kling, super charging Kuaishou’s AI

But in just this past year, Kuaishou’s AI-driven video creation app, Kling, has really come out on top. It distinguishes itself in the competitive generative AI landscape through a multi-modal strategy that integrates advanced algorithms, extensive data resources, and targeted investments.

Central to its technological edge is the refinement and upgrade of large-scale models, particularly focusing on a Mixture of Experts (MoE) architecture that reduces parameter size and operational costs while maintaining performance. This approach enables Kling to deliver high-quality video generation with greater efficiency. (However, I’ve also heard rumors that their model training requirements are much less stringent compared to other companies, so it has had the advantage of just dumping all sorts of video content into training.)

On the commercial front, Kling leads the domestic market in AI revenue monetization among native apps by lowering barriers for content creators and seamlessly integrating AI capabilities with advertisers and merchants. By focusing on not just producing more AIGC but also enhancing both the quality and quantity of AI-generated content (AIGC), it has been able to contribute to the whole ecosystem. But that said, a side comment - the future of regulations around AIGC is still very unclear and could lead to potential safety concerns for us all.

Similar to other consumer-facing apps that we have explored, Kling’s business model also incorporates subscription and value-added service charges. Notably, it is diversified across three primary revenue streams: consumer subscriptions (ToC), API services, and customized solutions tailored for both domestic and international clients. This multi-pronged approach is to support scalable growth and broad market penetration.

Much like Meta and ByteDance, the platform has also used AI to enhance its recommendation capabilities and thus has driven user engagement by delivering more targeted content recommendations, as announced in its recent earnings results. In turn, it increases the daily time users spend on the app.

Financials

Based on Jefferies’ March report on Kuaishou’s earnings, from a financial perspective, capital expenditures related to Kling are expected to remain manageable in 2025, as significant investments were made in 2023 and 2024. The focus for 2025 will shift toward inference costs, which the company plans to optimize through operational innovations and better resource utilization. R&D spending is projected to rise, following peers, driven by increased talent recruitment to sustain continuous improvements in model quality and product features.

However, Kling could now be a turning point for Kuaishou. It represents a cornerstone of Kuaishou’s AI strategy, combining technological innovation with a multifaceted business model to drive user engagement and monetization, and leveraging its video content database. This puts Kuaishou back at the forefront of China’s tech companies and potentially positions it as a leading player in the evolving AI video generation sector.

Share

Appendix: Who is Kuaishou?

First, this is not a random company. It is a company with a market cap of ~224 billion HKD (~29 billion USD). Kuaishou Technology is actually China’s first short-video app. It was founded in 2011, before Douyin, which was founded in 2016 (TikTok).

It was co-founded by Su Hua and Cheng Yixiao, and it is listed on the HKEX (HKG: 1024). It is largely known for developing a mobile app for sharing user-generated short videos, a social network, and a video special effects editor. The app is known as Kuaishou in China but Kwai in many countries outside of China. It also has a presence in India, Pakistan, and Indonesia, under the name Snack Video. Although it is publicly listed, it has a majority government-backed shareholder, China Internet Investment Fund, a state-owned enterprise.

Before the anti-monopoly probes in China’s internet sector, the internet was notoriously split up into two camps: one within the Alibaba system, which often benefited from the behemoth’s incubation and financial support, and one within the Tencent camp, which prioritized WeChat and its related services within its offerings. And Kuaishou was a loud and proud Tencent camp player, having received over USD 2 billion in investments from Tencent to compete with ByteDance. While ByteDance remained independent of these two throughout.

Global Ambitions

In 2020, Kuaishou tried to enter the U.S. market with a product called Zynn, but that was just when Trump was forcing TikTok to spin off, and there was that huge PR crisis with the high-profile hire. TikTok brought in Disney executive Kevin Mayer to expand the U.S. market. Still, Mayer quickly exited the company as it was drawn into a series of probes by the Trump administration.

Only a year later, Zynn quietly exited the U.S. market and quickly pivoted its engines towards LATAM and other emerging markets, making it one of the most popular apps in Brazil right now, under a new name. However, whether domestically in China or globally, it never reached the level of success that TikTok did, always hovering in the #2 spot in popularity. In fact, the last few years have been rough, as you can see below. In the company’s Full Year 2024 results, it disclosed 735.6 million MAUs, compared to ByteDance’s ~755 million MAUs in China alone.

Today’s post is a sequel to China’s startups pivot to vertical AI piece and follows our thought process on why China and the U.S. chose divergent AI monetization strategies - hint* China in consumer application end, and the U.S. focused on enterprise software.

Subscribe now

Microsoft CEO Satya Nadella famously claimed in a BG2 podcast that traditional business apps (SaaS) will soon be replaced by AI agents. This statement sent the internet into a frenzy, with headlines like “Satya said SaaS is dead.” But as the argument has played out over the last few months, we’re realizing that that’s not exactly what he meant.

Basically, he claims that traditional SaaS apps, which are essentially CRUD (Create, Read, Update, Delete) databases with business logic, are becoming obsolete because AI agents will take over the business logic layer.

He envisions a future where AI agents manage workflows and decision-making across multiple applications and databases, effectively replacing the traditional SaaS model with what he calls "Agents-as-a-Service" (AaaS). This shift means SaaS apps will revert to simple data stores, while AI layers provide the intelligence and orchestration.

In a subsequent interview with Indian entrepreneur Varun Mayya, Nadella said that in his world, Excel and Word are going to be agents that go with the co-pilot, a completely new way of thinking about the workflow. And he’s not the only one thinking this way. Greg Isenberg, CEO of LateCheckout, has also been very vocal on this front, and we will explore his points below.

And despite macro challenges and still a very unclear ROIC to AI applications right now, VCs and startups are still charging the way and they appear undeterred. In fact, they seem more convinced than ever about AI’s long-term role in software. Leading VC firm Andreessen Horowitz just announced plans to raise a $20 billion megafund focused on AI, doubling down on AI investments, especially in application use cases, a shift away from just infrastructure investments.

As I wrote in my piece on China’s AI startups, a similar trend is taking off in the U.S. We’re seeing a decisive shift: startups are moving away from competing on frontier models and doubling down on vertical use cases. And I think there actually is more open disucssion about this in the U.S. right now also because of the different AI monetzation strategies the two countries have chosen - China in consumer application and U.S. betting on enterprise adoption.

Historically, SaaS vendors have had a tough time selling to SMBs. These businesses are far more cost-conscious, every tool needs to deliver either concrete cost savings or clear revenue upside. As Primary VC put it, “SMBs are tough to sell to and even harder to retain.” However, enterprise software is probably even harder to replace.

So where is the narrative of “AI will replace SaaS” getting its confidence?

Subscribe now

AI Transforming SaaS

The Shopify CEO message leak went viral last week and reignited a discussion that’s been simmering for months: Are AI tools replacing new hires? A while back, the prevailing narrative was that SMEs were freezing headcounts and relying more on GPT-like tools, while large enterprises were still hiring and integrating AI gradually. However, now even Shopify appears to be signaling a shift.

This has sent some frights and shivers across the industry, but another argument that has come up is that if we’re not bringing in new talent, are AI tools actually good enough to replace their capabilities? Will AI agents be able to replace human talent, or to extent, can they replace certain workflows that previously required humans to complete? What is the future of AI in SaaS usage and adoption? Is AI killing SaaS?

Satya Nadella said that in the future, when hiring talents, “on a prosaic level,” he means that people will have “a basket” of agents that they’ve built, and they’ll be hired not just for their competence but also for their agents. He actually acknowledges that SaaS won’t be replaced completely but will be complementary to how we do things now.

On the other side of the argument. Greg Isenberg, CEO of LateCheckout, took to LinkedIn last month and has been a vocal leader in the argument that SaaS is dead. And he shared his view that AI agents will dismantle SaaS in three phases.

1. What we’re mostly seeing now can be considered as AI as Co-Pilot, mostly in a supporting role to our existing SaaS operations.

2. In the next year or so, we will see AI as Autonomous Operator which it will be able to bypasses human interfaces, unbundles workflows.

3. And then in 2-3 years time, we’ll be in an era of “software invisibility” where AI agents will talk directly to APIs and bypass the human interfaces altogether.

He predicts that “this democratization of software creation means every company becomes a potential software producer rather than just a consumer.” Companies will eventually abandon subscription-based tools in favor of custom-built, AI-generated software. Ultimately, leading to SaaS (as we know it), will then just all die.

However, this vision oversimplifies how technology is actually adopted in the enterprise. It overlooks practical realities like slow adoption - incrementalism, the importance of data analysis and systems of record, and the sometimes inefficiencies and unsophistication of DIY software development.

Instead, what I think will happen is that SaaS will be forced to evolve into a hybrid model where vendor platforms and custom-built AI agents will coexist, much like what Satya Nadella has said.

Potential Trends in Enterprise Adoption

1. Data Analysis is Not Easy

Many existing platforms, such as CRM service providers, accounting platforms, and HR software, are actually systems of record where they store, track, structure, and help analyze enterprise data, whether that is customer information or sales figures.

Even if agents can help receive instructions and execute tasks, these mega-service platforms will remain critical in an AI-native world because enterprise data cannot simply exist in thin air or be disorganized in random folders on our desktops.

And beyond the storage, analyzing the data is a capability that a stand-alone AI agent will not be able to do. This is where Satya’s point about the agent will coexist with the existing softwarethat we know of, like, Excel or Word, makes more sense. Proprietary data and the systems of record will continue to serve as foundational layers for the company’s processes and operations, just with AI intelligence added on top.

2. Vibecoding + DIY

The phenomonon of vibecoding has allowed many non-technical players to build their own apps and build on top of existing available products. But frankly, these DIY projects probably aren't good enough for mass-scale adoption if it’s not done properly by a professional.

The thought of being able to build your own apps is great, and many have shown that you can do it. But at an enterprise level, it also ignores the operational burden of building, maintaining software and scaling. Enterprise adoption of tools will need to be adopted across firms to ensure processes alignment, and it requires upkeep and trouble-shooting. To be honest, it may work well for SMEs but not for large corporations when you need to control quality across markets and time zones.

Added on, DIY may make collaborations or cross-market projects more challenging, and they are definitely not things companies want to mess with. A AI-native application is different in this case.

(On top of this, you’re assuming people are still willing to pay for the base product first - which we’ve written it’ll be a massive hurdle in China given the open-source options there are now. Might work better in the U.S.)

3. Hybrid Baby: SaaS + AI

The most likely outcome is that existing SaaS and AI will give birth to some kind of hybrid. SaaS companies have already started embedding GenAI into their products (think Canva, Figma), opening up APIs for agents and serving as secure foundations, and this will only continue. Customers will then build custom logic on top for edge cases or proprietary processes.

In Bernstein’s recent report, Software: GenAI application winners/losers framework, backs this up as well. Highly customizable software that mirrors enterprise workflows becomes staff augmentation, which creates deep stickiness. In contrast, general-use AI tools often require enterprises to retrain their processes to fit the software, which creates friction. Switching costs in SaaS aren’t just about money; they’re often about process change. And that favors existing, trusted vendors with strong integration capabilities.

And to Nadella’s point, Microsoft will thrive in this new world, because businesses will likely still buy standard tools and build on top of them. Standardizing that within their operations to make it most business-friendly for their case and, in many ways, this is not that different from many of the SaaS offerings out there right now.

This hybrid model offers the best of both worlds: vendor reliability and customization flexibility.

4. AI Adoption Will Not Be Overnight

Lastly, any enterprise adoption will not be overnight, unlike the consumer phenomenon we’ve seen. Over the last two years since OpenAI’s ChatGPT release, consumer adoption and for the sake of this argument - vertical adoption has been rapid. Still, enterprise AI adoption isn’t going to be gradual and layered due to bureaucratic reasons and system processes, data security, talent and uptraining time, and switching costs.

LLMs excel at tasks focused on ideation and summarization but still often falter on deterministic tasks like accounting or forecasting. And let’s not forget the hallucination risks. Although we’ve written about how AI value creation will shift to application-end, that was more about consumer-facing applications.

Most enterprise adoption will more likely happen in three phases:

1. AI assists users within existing tools (e.g., summarizing CRM activity, media monitoring, streamline information). This basically has already been implemented across consulting firms;

2. AI takes on more tasks in an agentic way, but still under close supervision to ensure logical reasoning and accuracy, we’re seeing this with agents like OpenAI’s Operator or ManusAI;

3. And then it’s the supposed “AGI”, AI autonomy, but that likely will still be within structured, known workflows that have been proven by human labor - at least for the near-term.

[And I wrote about what tasks I think AI agents can replace and why certain tedious tasks should be done by AI in PR]

This approach reflects organizational realities because enterprises often have the mentality if - “dont fix what is not broken.” So, in many ways, that sets up enterprise incumbents to be well-positioned to embed AI incrementally and preserve (at least try to retain) customers and help them with AI upgrades in existing workflows. It will be hard for new players to challenge their existing reach and deep industry know-how, and even user behavior.

Subscribe now

The Other Race: Vertical SaaS

Then, there is the discussion on vertical AI. Vertical AI startups are gaining traction for good reason. And we’ve talked about how in the China AI startup space, most of not all have pivoted from LLMs to a vertical use case amid DeepSeek’s breakthrough.

In the same interview with Varun Mayya, Satya Nadella said the competition for companies will extend beyond LLMs and be in their ability to make breakthroughs in building AI agents for various industries.

Competing on foundational models is expensive and distribution-intensive. Instead, founders are chasing faster ROI in high-value, industry-specific use cases. Instead of a mass-market SaaS integrating AI, this is building a tool for a niche market or at least a very specific industry use-case, such as the financial services sector, law, healthcare, or even specific tasks within an industry, like the social intelligence tool Meltwater for PR and communications professionals.

Jason Shuman, General Partner at Venture Capital firm Primary, who specializes in AI vertical investment, wrote on LinkedIn that founders are working on the same vertical SaaS use cases for AI at “breakneck pace”.

The three most important things right now are 1) Speed, 2) distribution advantages, and 3) strategic product decisions. (which we’ve explored in details in consumer-facing apps) And in many ways, the same logic applies for SaaS.

Especially where speed, distribution leverage, and smart product sequencing matter most. It’s about building moats fast, testing quickly, and thinking two steps ahead.

Trace Cohen at a family office echoes this: “The future is vertical AI—deeply specialized startups solving high-value problems in industries like healthcare, finance, and defense.” These players combine domain expertise, proprietary data, and utility-driven design. Even if the market size is smaller, the adoption curve is steeper. This is echoed in the Chinese AI market.

Conclusion: AI won't Kill SaaS

This brings me back to Greg Isenberg’s post from above, which suggests that AI will replace all our existing SaaS tools. I think he has some fair points, but that vision underestimates enterprise adoption challenges and is definitely over-optimistic about corporate efficiency and openness to change. I am more aligned with what Satya Nadella is saying.

AI probably won't replace software in the near term, but it will rewire the anatomy of these platforms and workers' behaviors in using these tools. As RMIT Professor Jonathon Boymal recently pointed out as well, research shows that Microsoft 365 Copilot is changing work habits of knowledge workders but it’s in areas like “email management” that is “adjusted independnetly” and not in “behaviors requiring coordination, like meetings,” again to my point - scaled adjustments in the processes across enterprises will take time but independent efficiency gain or productivity enhancing tasks will adopt AI at a much faster scale.

On AI Proem, we’ve explored many consumer AI applications and their speedy adoption rate and why, but for enterprise SaaS, the future isn’t “AI vs. SaaS.” I think it’ll be SaaS that behaves like AI and AI that becomes infrastructure. And the future belongs to those who (human or company) adapt and hence also my repeated emphasis on the importance of how we educate our next generation about AI.

SaaS vendors that open their platforms to AI agents, retain control of data pipelines, ensure customer security, and empower customer innovation will thrive. What we’ll see is more startups focusing on vertical use cases or creating agents on top of existing SaaS to complement the existing workflow. What that means is we might have more players than we originally imagined and incumbents like the big techs would likely thrive in consumer end or use cases of large distrubtion and reach. Meanwhile, newcomers will have their chance to lean on their industry expertise.

Thank you for your attention and interest, please hit subscribe or share with a friend if you found this useful/ insightful/ entertaining. Also feel free to leave a message or DM me if you think I’m spewing absolute gibberish. Cheers!

China’s artificial intelligence (AI) ecosystem is undergoing a seismic transformation catalyzed by the rapid adoption of DeepSeek’s groundbreaking R1 model. The startup’s meteoric rise is not only reshaping the AI application market but also redefining the global narrative on AI innovation and accessibility. With a focus on cost efficiency, open-source frameworks, and technological breakthroughs, DeepSeek has emerged as a key driver of growth, innovation, and competition in China’s burgeoning AI sector.

DeepSeek’s open-source approach has been pivotal in democratizing access to advanced AI technologies. By partially releasing its model architecture and parameters, DeepSeek has fostered collaboration among developers and enterprises, enabling them to customize and enhance its capabilities, as CNBC reported with a long, extensive list of open-source adoptions across the sector.

The open-source movement initiated by DeepSeek has prompted other major Chinese firms, including Baidu, to follow suit. Baidu recently launched its ERNIE X1 and ERNIE 4.5 models with open-source frameworks, signaling a broader shift in China’s AI industry toward accessibility and innovation.

For many, the choice of open-source is not just a business decision but a bigger movement in China.

Vital Timing of DeepSeek

DeepSeek’s emergence comes at a critical juncture for China’s economy. After facing a domestic slowdown in areas such as real estate downturns and trade tensions with the U.S., Beijing has embraced AI as a strategic growth driver. Goldman Sachs projects that AI adoption led by DeepSeek could boost China’s GDP by 20 to 30 basis points by 2030 through enhanced productivity and automation. The government has also ramped up investments in AI infrastructure, with companies like Alibaba forecasting significant capital expenditures to advance frontier models.

DeepSeek’s breakthrough has injected confidence back into the economy and society at large. Young entrepreneurs are bright-eyed, and investors are eager to find the next big deal or trend. As Bloomberg puts it, it’s “fired up a long-dormant Chinese tech industry.” (By the way, this is a really good long read; I highly recommend it.)

Artificial Intelligence Generated Content

As DeepSeek's swift adoption has spurred unprecedented growth and shifted user demographics, AI applications are taking off at rocket speed across the country, according to a recent analysis by Questmobile (China’s leading mobile app service provider).

The rise of DeepSeek, along with other key players like ByteDance’s Doubao, Tencent’s Yuanbao, Moonshot AI’s Kimi, and Baidu’s Ernie, has propelled the aggregate monthly active users (MAU) for AIGC (Artificial Intelligence Generated Content) apps to a staggering 248 million in February 2025. This growth trajectory is remarkable, with AIGC apps achieving a 20% internet user penetration rate in just 23 months, outpacing even the rapid growth of short video apps, which took 30 months, according to Citi Research.

Zooming in on DeepSeek’s achievements over the last few months. It achieved the fastest daily active user (DAU) growth in the history of the Chinese internet, reaching 50 million DAU in a mere 42 days. This is a stark contrast to established giants like Douyin (371 days) and PDD (741 days) in reaching the same milestone. While DeepSeek's DAU has since stabilized at around 45 million after peaking at 53 million in February 2025, its initial surge has left an indelible mark, according to industry blog Democratizing Information 信息平权-Jason . This stabilization occurred as more Large Language Models (LLMs) integrated DeepSeek and launched their own reasoning models.

The dramatic increase in overall user numbers clearly illustrates the rapid expansion of the AIGC app market. In February 2025, the total MAU for AIGC apps reached 245 million, a significant jump from 148 million in January 2025, 37 million in February 2024, and just 8 million in May 2023 in China, based on Questmobile’s figures. This explosive growth reflects a burgeoning appetite for AI-powered tools among Chinese internet users. Correspondingly, aggregate time spent on AIGC apps reached 464 million hours in February 2025, an eightfold increase year-over-year and a 66% month-over-month rise. On average, users spend approximately ~9 minutes per day engaging with major chatbots. Specifically, DeepSeek users spend around ~8.5 minutes daily, compared to ~10.4 minutes for Doubao, ~9.8 minutes for Yuanbao, and ~6.6 minutes for Ernie.

AI for All: Adoption Through the Roof

Beyond sheer numbers, DeepSeek's emergence has also triggered a notable shift in the demographic profile of AIGC app users.

Previously, AI apps attracted a larger proportion of male users. However, post-DeepSeek, there has been an increase in the activity of female, younger, and older users. China seems to be taking on a “全民AI” (AI for All) approach. From grandma/grandpas in their 80s to primary school pupils, everyone’s eagerly downloading AI apps onto their smartphones.

In February 2025, out of the 245 million AIGC app MAU, 61% (149 million) were male, and 39% (96 million) were female, representing substantial growth of 89% and 128%, respectively, since December 2024.

Furthermore, when total MAU grew by 102% from December to February, the age group below 30 grew by 108%, while the age group above 40 saw an even more significant increase of 238%, compared to a more modest 54% growth in the 30-40 age bracket.

By February 2025, 41% (100 million) of AIGC users were aged 30 or below, 35% (85 million) were between 30 and 40, and 24% (59 million) were above 40. This broadening user base suggests a wider adoption and integration of AI tools into the daily lives of diverse segments of the Chinese population.

While DeepSeek's initial momentum was strong, the Citi report notes that DeepSeek’s download volume has started to decline as more LLMs incorporate its technology and release their own models. Despite this, DeepSeek's DAU has remained resilient. The report also highlights the significant growth of Tencent's Yuanbao since early February 2025, suggesting a dynamic and competitive landscape where different players are vying for user engagement. It remains to be seen whether Yuanbao can sustain this momentum through increased promotional efforts.

Empowering the Whole Value Chain

Many have expected the parade of open-source models ignited by DeepSeek’s initiative to spill over into adjacent fields, from smart EVs to humanoid robotics and beyond. Which we’ve started to see already…(Bear with me. This is the topic for a piece that’s been long in the works - China’s strength in Physical AI, powered by the prioritization of the “real economy.”)

According to Bloomberg’s interview with Balaji Srinivasan, tech investor and former general partner with Andreessen Horowitz, he said that with China’s strength in hardware, the cheaper and more accessible the AI models, the more the demand for devices powered by AI.

In fact, its impact has truly been seismic. There has been a rally across all the industries AI touches in China- across the supply chain.

And a 1 trillion dollar wipeout in the U.S. markets. (see the Fortune article here, a very nuanced piece about DeepSeek’s global industry implication, and they cited me ;p)

A Catalyst for Transformation

By prioritizing cost efficiency, open collaboration, and broad accessibility, DeepSeek has set new benchmarks for innovation while challenging American dominance in the field.

DeepSeek's rapid ascent has acted as a catalyst for the explosive growth and evolving user demographics within China's AI application market. While the competitive landscape is heating up, DeepSeek's initial impact demonstrates the significant potential and transformative power of AI in engaging a broad spectrum of Chinese internet users (as I said, consumer apps can scale so fast in China).

As China embraces this transformative wave of AI applications, the implications extend far beyond its borders—reshaping global perceptions of the future of AI.

The dust has settled, and the hype surrounding DeepSeek has tempered. Since the release of the R1 model, both hyperscalers and LLM startups have revisited their business strategies. And China’s AI startups are undergoing a strategic transformation as competition intensifies. Once focused on foundational large language models (LLMs), these companies are now shifting toward consumer applications and vertical markets, aiming to carve out sustainable niches in a crowded field.

While mainstream media has highlighted China’s EV and home appliance sectors adopting DeepSeek’s technology, these integrations remain largely consumer-facing enhancements. The more significant impact has been on China’s LLM startups - more specifically, the “four tigers,” which are now rushing to make a second pivot.

Big Tech Remains in the Lead

Despite the shifts among startups, the biggest winners remain China’s tech giants since DeepSeek’s R1 release. As seen in the screenshot of the WeChat account Weixi, the "Terrific Ten" has seen a 93% return year-to-date, far surpassing the 48% growth of the U.S.'s "Magnificent Seven." This underscores the dominance of incumbents in capital-intensive areas like the models and infrastructure layers.

The Strategic Shift Among Startups

The original “AI four tigers”—Zhipu, MiniMax, Moonshot AI, and Baichuan—were once leaders in China’s LLM space during the peak of AI hype. Over the past two years, as big tech began catching up with proprietary LLMs, these startups pivoted toward consumer-facing applications with tools like Kimi, Talkie, and Hailuo. However, following DeepSeek's R1 release, the landscape has become crowded, and expectations for free and accessible frontier LLMs have risen.

In response, these startups are revamping their strategies yet again. According to an analytical piece by The Financial Times, they are now shifting focus to enterprise solutions to remain competitive amidst DeepSeek’s widespread adoption across China. Notably:

• Baichuan: Narrowed its focus to healthcare applications and established itself as a leader in this vertical.

• Moonshot AI: Reduced marketing efforts for its Kimi chatbot to prioritize model training and performance improvements.

• Zhipu AI: Strengthened enterprise sales while preparing for an IPO and continued developing open-source AI models.

• MiniMax: It seems to be staying put as of now. Staying focused on its consumer apps.

Even 01.ai, the AI startup founded by Lee Kaifu has stopped pre-training large language models to focus on selling tailored AI business solutions using DeepSeek’s models. This shift reflects a pivotal moment where hyperscalers dominate infrastructure and model layers due to their ability to invest heavily in R&D. Startups are increasingly focusing on application use cases, and more specifically, with vertical expertise— similar to patterns seen during the internet era.

Subscribe now

Timeline of Strategic Adaptations

Let’s just take a look at the timeline, which was put together with the help of Perplexity Pro.

Pre-2023: Initial Business Strategies

• Baichuan: Founded in April 2023 by Wang Xiaochuan, Baichuan initially focused on developing open-source LLMs. They released models like Baichuan-7B and Baichuan-13B in 2023, aiming to contribute to the AI research community and establish a foothold in the foundational model space.

• MiniMax: Established in December 2021 by former SenseTime experts, MiniMax concentrated on creating LLMs and multimodal AI systems. Their early efforts were directed toward foundational AI technologies, including the development of the MiniMax-01 model.

• Moonshot AI: Founded in March 2023 by Yang Zhilin, Moonshot AI aimed to develop LLMs capable of handling extensive text inputs, emphasizing long-context processing capabilities. This focus was intended to address complex language understanding and generation tasks.

• Zhipu AI: Established in 2019, Zhipu AI initially concentrated on building LLMs and providing AI solutions tailored for enterprise applications, targeting sectors such as local governments and various industries.

2023–2024: Pivot to Consumer Applications

As the AI market evolved, these companies recognized the potential of consumer-facing applications and adjusted their strategies accordingly:

• Baichuan: Transitioned towards developing AI applications for the healthcare sector, including an AI doctor assistant designed to aid in medical diagnoses.

• MiniMax: Launched consumer apps like "Talkie" and "Xing Ye (星野)." "Talkie," introduced in June 2023 for international markets, allows users to create and interact with virtual characters. "Xing Ye," released in September 2023 for the Chinese market, offers similar functionalities.

• Moonshot AI: Developed "Kimi," an AI chatbot that gained popularity for its conversational capabilities. However, the company faced challenges related to service outages and increased competition, prompting a reevaluation of their focus.

• Zhipu AI: Expanded into consumer applications while continuing to build its enterprise business, offering personalized AI solutions to local governments and companies.

2025: Strategic Shifts Post DeepSeek

The introduction of DeepSeek's R1 model in late January 2025, noted for its rapid adoption and advanced capabilities, prompted these companies to reassess their strategies:

• Baichuan: further narrowed its focus to the healthcare industry, emphasizing the development of AI tools for medical applications. It has successfully established itself as a vertical leader.

• MiniMax: Continued to innovate in consumer applications, enhancing products like "Talkie" and "Xing Ye," and developing new AI-driven features to maintain a competitive edge.

• Moonshot AI: Reduced marketing efforts for "Kimi" to concentrate on improving model training and performance, addressing previous service challenges, and aiming to compete more effectively with leading models like DeepSeek's R1.

• Zhipu AI: Focused on strengthening its enterprise sales, aiming for an initial public offering (IPO) to fund growth, and continued developing open-source AI models to enhance its offerings.

These strategic adaptations reflect the dynamic nature of the AI industry, which is moving fast. As incumbents dominate in capital-intensive areas, startups must continually evolve to align with technological advancements and market demands and find their niche to dominate.

Share

LLM Consolidation Trends

The consolidation of LLM ecosystems is not unique to China. According to Morgan Stanley's March investor insights, software companies globally aim to create larger operating systems powered by machine learning, natural language processing, generative AI, and decision-making algorithms. These systems could prove valuable across consumers, enterprises, creators, and advertisers.

In China:

• Tencent monetizes through ads and value-added services.

• Alibaba focuses on cloud-based revenue streams.

However, industry leaders like OpenAI CEO Sam Altman argue that subscription-based productivity tools may offer greater long-term value than advertising models. Altman believes consumers will pay for superior productivity-enhancing products rather than rely on ad-supported systems during an interview with Stratechery’s Ben Thompson. In contrast, tech strategist Ben Thompson suggests that OpenAI could leverage its data trove for targeted ads, and as it does not have an existing business to protect, it can just yolo a bit. A strategy incumbents Google has had to approach with much more caution due to risks of cannibalizing its Search business.

One thing that is clear is that for startups without existing platforms or cloud monetization capabilities, innovation lies in creating entirely new solutions tailored to niche markets. As consolidation continues among hyperscalers like the Magnificent Seven in the U.S. or the Terrific Ten in China, incumbents remain well-positioned to dominate the value chain.

Made-in-China Apps 2.0

As a sequel to the Made in China AI Apps piece I wrote, I thought I’d provide an update based on this super comprehensive list a16 put together: Top 100 GenAI List.

According to a16z's list, DeepSeek ranked second behind ChatGPT among web-based GenAI apps by monthly visits as of March 2025. Other notable Chinese apps include ByteDance’s Doubao, Moonshot's Kimi chatbot, MiniMax's Hailuo and Talkie Kuaishou’s Kling, Cutout.pro, monica.ai’s Manus, which have all made it into the top 50 rankings globally.

Chinese consumer adoption of mobile apps remains strong due to higher penetration rates compared to Western peers—a trend evident from DeepSeek’s high engagement metrics. And the reason for consumer adoption being so high can also be attributed to just simply a relatively low cost for innovation diffusion.

What’s interesting is that DeepSeek’s mobile engagement is extremely high. This goes back to my point about Chinese consumers’ adoption of mobile apps being much easier to penetrate than Western markets, given its embrace of the mobile-omnipresent nature and super-app culture.

“Consumers are happy to use anything that is free, but corporate is not going to spend even the soft costs if they do not have a line of sight to the ROI,” wrote inBernstein’s AI report, written by various analysts but led by Mark L. Moerdler, Ph.D.

This has been proven by the sheer number of downloads and engagement for DeepSeek compared to non-free peers. Consumers are less interested in geopolitics and more interested in cost-efficiency and user experience. (Enterprises, however, otherwise.)

Subscribe now

Future Outlook

As foundational models advance further:

• Tools for integrating generative AI into operations will become essential.

• Governance mechanisms will gain importance within enterprises.

• High fixed costs and economies of scale will solidify hyperscaler/ big techs’ dominance in infrastructure and model layers.

For smaller companies:

• Opportunities lie in practical applications built atop foundational LLMs rather than direct competition in resource-intensive development spaces.

• Vertical expertise will be key—startups can succeed by capturing niches that big tech overlooks due to scale limitations.

In China's market specifically:

The “four tigers” have each navigated distinct paths in the evolving artificial intelligence landscape. Their journeys encompass initial strategies centered on large language models (LLMs), pivots toward consumer applications, and recent strategic shifts influenced by DeepSeek's R1 model release.

As I’ve written about in detail here, the multi-modal model and busy UI/UX preference adopted by Chinese big-tech AI apps have been a foreshadowing of what Doubao, Yuanbao, and so on have planned for the future. Consumer applications heavily rely on ad revenue and value-added services. It’s about increasing touchpoints, finding friction points, and prolonging engagement = the proven playbook that internet mobile apps made their money with, but that’s for the big tech with an existing massive user base to tap into.

Startups, on the other hand, should probably focus on creativity and vertical know-how to differentiate themselves. Their game must be in vertical expertise, and doing essentially big tech cannot be bothered to try to capture. The market won't be as big, but it will probably still be big enough.

Subscribe now

This Isn’t Another DeepSeek Moment

Last night China-time Alibaba announced that it entered into a strategic partnership with the star of the week, Manus AI. This comes as Manus has said to be leveraging Anthropic’s Claude 3.7 Sonnet model, which cannot be accessed within the reach of China’s Great Fire Wall, and thus, the AI agent was only able to be tested outside of China. The partnership with Alibaba and, therefore, its Qwen AI models will boost its capabilities and usability in China. Leveraging the internet behemoths’ abilities also means that Manus AI will be able to manage its traffic surge better and not experience something that DeepSeek faced, which was its server being unable to accommodate the amount of interest coming in.

Manus AI has taken the internet by storm and is even being dubbed another “DeepSeek” moment for China AI, but I don’t think it is. This is not fundamentally changing anything we believe in AI like DeepSeek did. DeepSeek challenged the status quo belief that more capital equals more training equals better models. The V3 and R1 releases had a much more significant and wider industry impact.

In this case, Manus AI is introducing to the world that the AI agent can operate like a digital employee and execute independent tasks with little prompting or guidance - to really have the potential to replace a knowledge worker. It’s been compared to OpenAI's Operator, Anthropic's Claude 3 Opus, and lesser-known AI agents such as Cognigy.AI and BabyAGI. In many ways, it is what we have expected of AI and its capabilities, but it’s just come sooner than we expected, becoming more than just a research tool or chatbot. Impressive still. To get more details on what it can do, see Tony Peng’s writing here about how it was initially dismissed as a “wrapper,” or you can check out MIT Tech Review’s review.

But as I wrote last month, AI monetization and innovation are truly shifting from the obsession of bigger and better LLMs to applications and specific use cases of AI (as LLMs all somewhat become ‘good enough’).

Now I wrote about this when it first launched, that because it is still in its invite-only product trial phase many people questioned its technical capabilities as information was largely spread through second hand information, and mostly reviewed by independent media, which led to speculations that it was more of a marketing and PR stunt than technological breakthrough.

However given its been a few days, and more tech people have gained access to the product trial, the demo and product has been widely received positively across the board.

The Information wrote, “It’s possible that the hype around Manus stemmed from the rising interest in Chinese-made AI products, thanks to DeepSeek.” Adding that the experience was not as smooth as the demo showed when their own reporter tried it out.

But - “To be fair, Manus is still a work in progress. In its YouTube video, Ji, the cofounder, described last week’s launch as an “early preview. No matter how Manus created a somewhat competent computer-using agent, there’s no doubt it has moved the ball forward in showing what’s possible. That’s great for OpenAI, Anthropic and their ilk.”

Subscribe now

What’s interesting to me is that this phenomenon and media frenzy proves again what I’ve written before, is that a part of this excitement is fueled by this new wave of fanfare and patriotism for young Chinese (AI) entrepreneurs. Similar to Unitree Robotic’s Wang Xing Xing, the Manus AI CEO Xiao Hong was also born in the 90s and was domestically educated - which differs from a lot of entrepeneurs from earlier generations that were educated in the West and returned to China to start a business. And compared to even earlier entrepreneurs born in the 60s and 70s - who’s ambitions were often just “let me make the same business that exists in the U.S. but for the Chinese market.” The 80s/90s entrepreneurs are young and creative and really leaning into “doing the hard thing” and '‘leading in innovation” as their mottos - as DeepSeek’s CEO Liang Wenfeng has encouraged China’s youth to do.

The Founder and CEO Xiao Hong (肖弘) and Co-founder & Chief Scientist, Ji Yichao (季逸超) [the man in the demo video] are serial entrepreneurs and ’s ChinaTalk dug out some of their old interview quotes here. See how his vision for the future has been largely driven by his ideals and ambitions of freeing up humankind to be able to focus on culture, physical health and leisure again (I mean for the privileged, yes) and not be tied to a desk job. Again, proving that this generation of entrepreneurs are much more visionaries than just pragmatic businessmen.

See ChinaTalk’s translation below:

“I think that the white-collar lifestyle may be a detour for mankind. If you look at it in terms of a curve or over a longer period, say thousands of years, or even the ten-thousand-year span of human history — it's actually quite rare for people to sit in one place and engage in intense mental work without much physical activity. This is probably only a phenomenon of the past hundred years.

For a longer time in history, maintaining physical health and developing spiritual civilization have gone hand in hand. In ancient times, people also needed spiritual and cultural development, but that involved physical labor as well, which helped strengthen their bodies.

In the past hundred years, however, issues like diabetes and high blood pressure have become widespread because people work in this sedentary way. If we look at humanity as a whole, sitting and working for eight or more hours a day is an anomaly.

If AI can take over these tasks, then people can work fewer hours and go back to living more like they did in the past — focusing more on spiritual and cultural enrichment while also taking better care of their physical health.”

China’s LinkedIn Boss Zhiping Introduces AI to Recruitment

A quick note on BOSS ZHIPING (translates to Boss-direct-hire) (NASDAQ: Kanzhun), China’s leading recruitment platform which announced its earnings yesterday. The online recruitment platform connects job seekers with employers through its mobile app. During its earnings call, the managememnt announced a series of AI products and explained how AI has affected its business which I think can be a preview of how LinkedIn and Indeed may be adopting AI in the future. This is another good example of vertical AI innovation.

Unlike traditional platforms such as LinkedIn, Boss Zhipin distinguishes itself by facilitating direct connections between candidates and hiring managers. The platform's slogan, "Talk to the boss directly," emphasizes its mission to eliminate intermediaries in the recruitment process. Boss Zhipin employs three key strategies: AI-driven recommendations delivered through personalized newsfeeds, direct real-time chat functionality, and a mobile-first user experience. This approach has effectively "Tinder-fied" job hunting with swipe-based browsing and instant messaging features.

AI-Powered Recruitment Solutions

Boss Zhipin serves both white-collar and blue-collar workers, with significant attention to the latter segment. The platform has innovated to address blue-collar recruitment challenges, where traditional methods often prove inefficient due to high candidate volumes and the need for rapid screening.

The company's upcoming AI chatbot initiative, scheduled for release within 1-2 months, aims to automate initial interactions in "big batch recruitment" scenarios common in blue-collar hiring. This system will handle resume collection and contact information exchange after an initial recruiter greeting, managing the process until successful engagement or rejection. The reported 90% acceptance rate among job seekers suggests strong potential for adoption.

Blue-collar recruitment presents a substantial opportunity due to its vast market size, which could potentially exceed white-collar recruitment several times over. Historically, this sector has been hindered by inefficient one-on-one communications and inadequate initial screening processes.

Future AI Developments

Boss Zhipin is preparing to launch another new AI recruitment product after extensive testing. This technology promises to increase matching efficiency by over 60% while reducing recruiter time investment by 40%. Such advancements align with broader industry trends, as AI-powered tools become essential for enhancing recruitment efficiency. With commercialization on the horizon, this technology could significantly transform recruitment processes by boosting productivity across various industries.

The platform's AI-driven approach to recruitment exemplifies how technology can streamline hiring processes through automated resume screening, predictive analytics, and personalized job matching. As Boss Zhipin continues to innovate, its strategies may indeed preview how global platforms like LinkedIn and Indeed might adopt AI technologies in the future.

FYI for Part Two, where we dive deeper into - Why China’s AI Strategy Differs from that of the U.S., see here.

AI is a solution in search of a problem—a general-purpose technology that has yet to carve out its place in the economy. With hundreds of billions of dollars in investment spent on its development, there is growing pressure for AI firms to turn a profit.

Over the past several months, divergent approaches to commercialization have emerged in the U.S. and China. In the U.S., AI has been largely diffused as a way to empower enterprises, with tech companies embedding the technology into their cloud services, business software, and productivity platforms. In China, by contrast, AI is increasingly woven into consumer-facing super-apps, enhancing e-commerce, social media, and digital payments. This divergence reveals how technological diffusion is shaped not just by the capabilities of AI itself but by the distinct market structures, business incentives, and even national institutions in each country.

In this post, we’ll try to articulate these differing approaches to AI commercialization. This is no easy task, as companies are continuously experimenting with new models, making it a moving target. On top of that, systematically quantifying how AI is monetized is tricky. AI, after all, isn’t a standalone industry but is embedded into existing applications and workflows, serving primarily as an enhancement rather than a distinct product. As a result, it is difficult to estimate what portion of a company’s revenue comes from AI.

There are a few exceptions to this. Some firms are entirely dedicated to AI. OpenAI, which generated $3.4 billion in 2024, and Anthropic, which earned $200 million, can obviously be fully counted as AI revenue. However, these firms represent only a sliver of the overall diffusion of the technology. Microsoft—the only of the giants to explicitly break out AI revenue in their income statement—posted $13 billion on AI alone. So, while firms like OpenAI and Anthropic are at the frontier of model development, it's the tech giants that are capturing the lion’s share of the financial rewards from commercialization.

This pattern is no different in China. Model development companies like DeepSeek remain unprofitable—nor is becoming profitable their goal, at least not in the short term. Instead, firms like Alibaba, Tencent, and ByteDance, with their vast digital infrastructure, are best positioned to monetize AI—and they are already among the most aggressive in doing so. With this in mind, this post will compare how AI is diffusing by focusing on the efforts of tech giants in China and the U.S.

(In part one of this series, we lay out our analyses as to how the leading players from each country are positioning their businesses in this new AI sector, and in part two, we’ll explore potential explanations as to why such a divergence has occurred.)

Subscribe now

Commercialization Taxonomy

So far, we’ve discussed how AI commercialization can be either enterprise- or consumer-facing. However, within each of these categories, firms can pursue two distinct strategies: directly monetizing an LLM or monetizing downstream applications that leverage an LLM—a framework we refer to as "monetizing LLM value-add."

Keeping track of all this can be confusing, with the constant wave of new products hitting the market. To help structure the different approaches, we’ve created a table that serves as a taxonomy of how firms are commercializing AI. This should serve as a helpful reference, as all the products discussed in the next two sections will fall within this structure.

The U.S.’s Proven Enterprise Cloud Advantage

In the U.S., the top backers of AI—Microsoft, Google, and Amazon—are also the cloud giants, and so it isn’t a surprise that these firms are monetizing AI through the cloud. The way they’re doing this is by including various large language models (LLMs) into their suite of cloud offerings and making it available to the millions of enterprise customers that already rely on their cloud services. Giving customers access to LLMs is just the tip of the iceberg. On top of that, cloud providers also allow customers to customize (e.g., fine-tune), deploy (e.g., inferencing), and manage the life-cycle of these models—all of this runs on top of existing cloud infrastructure.

Microsoft’s $13 billion in AI revenue is primarily driven by this approach—offering access to foundational models such as OpenAI’s, providing customers with the ability to fine-tune these models, and providing the infrastructure for managing inference workloads at scale. Interestingly, Microsoft’s AI growth is largely fueled by inference workloads, as enterprises increasingly rely on its cloud services to deploy and operationalize AI models in real-world applications.

What’s striking about this is that the focus isn’t on AI as a direct consumer product but rather on enabling other businesses to integrate AI into their own workflows and operations. Microsoft CEO Satya Nadella emphasized this point, stating that AI’s “real demand” is in “the enterprise space or our own products like GitHub Copilot or Microsoft 365 Copilot.”

Amazon and Google’s approach to AI commercialization is similarly built around cloud-based AI services. And like Microsoft, their strategy prioritizes enabling businesses rather than developing standalone consumer AI products, focusing on building AI into their own business processes. What’s noteworthy about this approach to commercialization is that these firms provide access to a marketplace of over a hundred pre-trained models for customers to select from. So, while Microsoft’s preferred cloud model is OpenAI’s ChatGPT and Amazon’s is Anthropic’s Claude, cloud providers aren’t locked into a single model. Notably, DeepSeek’s model is available on both Amazon’s and Microsoft’s cloud.

At the core of this strategy is the direct monetization of the AI models themselves, whether through licensing access, charging for customization (fine-tuning), or deployment (inferencing). Fundamentally, this approach is about selling AI to businesses, enabling them to integrate and customize models to enhance productivity, automate workflows, and drive efficiency across their operations.

(We’ve compared the U.S. and Chinese cloud computing sectors here)

Among the giants, Meta is the exception. While it rivals the cloud giants in capex for AI infrastructure, it itself isn’t a public cloud provider and can’t rely on enterprise adoption for growth. Instead, its AI commercialization strategy is tied to AI-enhanced advertising—the company’s primary business line. Meta hasn’t reported AI-generated revenue separately. However, its recent boom in advertising revenue is attributed to AI-driven optimizations. This includes generating countless variations of the same advert, analyzing and testing them in small batches to determine which works best with the target audience, and then flooding the market with the highest-performing ones.

Unlike the other giants, Meta also has the best shot at popularizing a consumer-facing AI app with its control over WhatsApp, Facebook, and Instagram. Mark Zuckerberg has been one of the loudest in advocating for a consumer-facing AI agent, even stating that Meta will be behind a 1-billion user-serving application.

So far, consumer-facing AI apps are largely dominated by firms like OpenAI and Anthropic. However, as we've discussed, this represents only a small fraction of AI’s overall commercialization. The real financial gains remain firmly in the hands of big tech, where AI is primarily monetized through enterprise services that rely heavily on cloud computing and the vast compute infrastructure required for AI’s diffusion.

China's Hopeful Consumer Strategy

In contrast, China’s tech giants have focused on monetizing AI through applications rather than directly commercializing the model layer. This reflects a broader challenge in China’s economy: individual users and business customers are far more reluctant to pay just for access to AI models. Even in the consumer chatbot market, where OpenAI and Anthropic have had decent traction in attracting paid users, their Chinese counterparts, such as Moonshot’s Kimi or Baidu’s Ernie, have struggled to convert users into paying customers.

As a result, the prevailing strategy in China has been to leverage AI to deepen engagement within existing consumer platforms. This includes AI-powered recommendations, content generation for e-commerce, and interactive features designed to enhance user retention across existing apps. Numerous consumer-facing AI applications are also being rolled out for entertainment purposes, from Meituxiuxiu's AI-powered image editing to Talkie's AI-generated avatars to role-play character companions. Such efforts, however, have resulted in little financial payoff, with tech leaders in China even saying they aren’t sure what monetization will look like.

Fundamentally, there is more of an expectation for AI to be free in China, making monetization through subscriptions more challenging. Part of this could be because there was simply no outstanding LLM until DeepSeek (maybe except for Alibaba’s proprietary Qwen). Then, when DeepSeek was launched, it was offered for free, setting the expectation that LLM access should be free anyway.

DeepSeek’s breakthrough in model performance has also served as a wake-up call for China’s tech giants, forcing Tencent and Baidu to reassess their approach. Both companies have come to recognize that their heavy emphasis on application-driven engagement has come at the expense of advancing their own foundational models. In other words, they’ve rushed to commercialize subpar AI models, integrating them across China’s internet ecosystem before achieving true model competitiveness.

Ironically, DeepSeek’s success has also triggered a new race to commercialize AI. As Paul Trilio, Partner and SVP for China and Tech Policy Lead at DGA Group, notes: “Chinese AI firms have struggled to come up with ways to monetize AI model deployments, and the emergence of DeepSeek [...] means that the application side is critical: quickly leveraging an advanced model that is deployed across multiple platforms and drives down costs is now seen as the goal.”

Tencent—lacking both a leading model and a working monetization strategy—is capitalizing on this shift. It has integrated DeepSeek into its consumer-facing applications, notably embedding it directly into its super app WeChat, where it can optimize user experience across the super app’s over 1.5 billion monthly active users. Similarly, Baidu has also dropped its closed-sourced approach to integrate DeepSeek into its search app—a more lackluster product but backed by similar thinking.

ByteDance has thus far refrained from integrating DeepSeek’s model into its applications, opting instead to stick with its proprietary LLM. Before DeepSeek, ByteDance’s Doubao was China’s leading LLM chat app, but its success was driven more by “cool features” than by the strength of its foundational model. These features included love therapy chatbots, English tutoring, and even a bot that mimicked Elon Musk—all designed to generate revenue through embedded pop-up advertisements rather than through direct monetization of the model itself. (See more about Doubao here.)

In just two months of its launch, DeepSeek surged to 22.2 million daily active users, surpassing Doubao's 17 million users. This highlights that foundational model capabilities remain the key driver of competitiveness.

Among the big tech companies in China, Alibaba and, to a lesser extent, Huawei are the only ones to adopt an enterprise-first strategy similar to the U.S. cloud giants. Alibaba has a consumer-facing app, Kuake, which has tried to challenge Doubao’s dominance, but it remains a second priority for the company. Instead, Alibaba is focused on commercializing AI through cloud sales with internal goals of allegedly 10 percent year-over-year revenue gains for next year, which it achieved and even exceeded during this past quarter.

Mimicking the U.S. cloud giants, Alibaba CEO Eddie Wu made the analogy during the most recent earnings call that the company will be the grid provider (cloud infrastructure) for the electricity demand (AI). It has offered its open-source Qwen model on top of or bundled with its cloud service, posting a 13 percent year-over-year revenue growth for its cloud business during its most recent earnings announcement. Like Google and Amazon, however, revenue directly from its cloud AI services is not reported separately and can only be inferred at best.

Huawei, also more focused on AI commercialization through enterprise cloud, has trailed behind Alibaba, finding success mainly among its existing SOE client base, largely due to its lauded safety features. (Grace has written about Alibaba and Huawei and their AI strategies)

What’s clear is that there is no consensus over how best to monetize AI, but everyone is choosing a lane that makes the most sense with their existing business to race in.

Implications of Each Approach

We’re still in the early days of the commercialization of generative AI, and a breakthrough “killer” application from one side could quickly push the other to follow suit. It's also possible that no such application ever emerges, and the excitement around AI remains just that—hype.

However, it is becoming increasingly clear that the top firms in the U.S. and China are thinking quite differently about their plans to monetize AI—one focused on enterprise adoption and white-collar productivity, the other on consumer engagement and digital consumption. These divergent approaches will shape not just who profits from AI but also how AI reshapes the economy itself—whether as a tool for workplace transformation or as a driver of consumer spending and engagement.

In the U.S., AI commercialization is largely enterprise-driven, with a strong focus on enterprise adoption. This approach takes the pressure off big tech firms to drive AI innovation alone, instead allowing the millions of businesses that rely on cloud infrastructure to integrate AI and explore new ways to use it on their own terms. This likely means more emphasis on enhancing worker productivity as AI tools become embedded in enterprise software.

By contrast, China’s AI strategy is heavily consumer-focused, prioritizing consumer applications that directly shape how users interact with digital platforms. This means that AI diffusion is largely driven by its impact on user engagement and retention, with companies racing to embed AI into super-apps, e-commerce platforms, and content ecosystems. As a result, the effects of AI in China are more likely to be concentrated on stimulating domestic consumption rather than transforming workplace productivity.

Regardless of whether AI is monetized through enterprise adoption or consumer engagement, from where we’re looking at it now, it is the tech incumbents who are best positioned to reap the financial rewards. Unlike the internet era—where new consumer apps could scale without owning the underlying fiber infrastructure—the AI era is far more capital-intensive. This gives a significant advantage to the firms that can afford to build or rent vast amounts of compute infrastructure and even secure dedicated energy sources to power these data centers. Whether through enterprise cloud services in the U.S. or AI-powered consumer platforms in China, the firms that own and control such infrastructure will likely dominate its commercialization, leaving smaller players dependent on their ecosystems.

Nevertheless, this divergence in commercialization strategies reflects a deeper divide in the tech ecosystems of the U.S. and China, shaping not just how AI is monetized but how innovation diffuses in each political economy. These differences are not accidental—they are rooted in structural factors that have long influenced how technology is adopted and scaled in both economies.

In Part 2, we’ll explore the deeper factors driving this divergence. We’ll examine why enterprise software (e.g., SaaS) never gained traction in China and how this created a path dependency that continues to shape AI adoption in enterprise products. We’ll also take a look at how the economic and labor structures of these two countries have influenced their distinct AI commercialization strategies.

As I wrote last week, Chinese companies are upping up their games in this AI application race. While value creation is now shifting from the model infrastructure layer to the application end, businesses are finding more creative ways to implement AI into application use cases. Now, largely driven by the success of DeepSeek, the Chinese AI industry is garnering a lot of international attention - people are trying to learn more about promising Chinese AI startups. So that leads me to Hangzhou, the home to Unitree Robotics, DeepSeek, and other household names that have been put on the international map.

Citing SCMP, “[Innovation] relies on a trifecta – fundamental research, applied technology, and financial support – all of which are essential,” Zheng Yongnian, dean of the School of Public Policy at the Chinese University of Hong Kong (Shenzhen), adding that “Hangzhou is excelling in this regard.”

When comparing innovation hubs in China and the U.S., Zheng noted that the U.S. has two key regions —one stretching from Silicon Valley to Texas and another from Boston to New York. China, similarly, has two main clusters: the Yangtze River Delta region (Hangzhou at the center) and the Greater Bay Area (Shenzhen at the center, which I’ve written about before).

Today, we dive into Hangzhou, which is home to the widely talked about - “six AI dragons.”

Hangzhou- “Paradise on Earth”（人间天堂）

Most of you probably have only heard of Hangzhou (deemed as a second-tier city by China standards, has 12 million people inhabiting it, and ranks 9th in GDP in China) because it is home to Alibaba, NetEase, and Geely (some of China’s largest publicly listed companies). Now, the city is quietly cultivating a whole generation of leading AI firms.

The city actually has a few nicknames in Chinese: “Paradise on Earth”（人间天堂）, “Home of Tea”（茶叶之乡）, “City of Silk”（丝绸之城), or “Eastern Venice” (东方威尼斯）. You can visualize it. It is a culturally rich and GDP-rich city.

The historically culturally rich city of West Lake has always been a center for scholars, artists, and thinkers, contributing to advancements in literature, philosophy, and the arts. It is also known for its craftsmanship in premium silk and china, and it was largely the epicenter of Chinese artistic work exported to the West.

In more recent days, it’s cultivated mega-heavy weights like Alibaba and Geely. Thus, by nature, it has a robust e-commerce and retail ecosystem boasting a series of local DTC cosmetics and fashion brands. Some say it’s the most fashion-forward city in the country now, as well as having one of the most advanced manufacturing industries. In this juxtaposition, it is an industrial-heavy, manufacturing-strong, and creative hub with strong ancient art and architecture as the foundation.

That ends my tangent on Hangzhou. I truly enjoyed my ~ nine-month stint there, which was originally intended to be a three-week work trip to Alibaba’s XiXi campus during the peak of China’s COVID restrictions (the quarantines were brutal, but freedom tasted better than anything when let out free). It gave me an opportunity to explore this city that, frankly, otherwise I might have never visited.

Hangzhou’s AI “Six Little Dragons”

There are six AI and robotics firms headquartered in Hangzhou that are drawing investor and industry professionals' attention right now. Together, they’re nicknamed the “six little dragons.” (I feel like China’s AI space is making a 10-season series of Crouching Tiger, Hidden Dragon these days with all the “Four Tigers” see here for an article on the apps backed by the four LLM startups and the “Six Dragons).

Closer look at them…

You may ask, what makes Hangzhou so special? Or was it all a coincidence?

TL;DR

1) Strong government support; 2) Strong talent pool; 3) Entrepreneur and hustle culture; 4) Strong tech and manufacturing ecosystem; 5) Proximity to Shanghai and distance to Beijing

Just a week ago, the Hangzhou government announced that it would further support AI and robotics firms by providing a financial “boost.” The municipal fiscal fund increased from 49 billion yuan last year to 50.2 billion yuan. More support is given to future industries, such as general artificial intelligence and humanoid robots, with the hope of cultivating more innovative enterprises similar to DeepSeek and Unitree Robotics.

Subscribe now

Beyond the government’s active support for cultivating the AI ecosystem in Hangzhou, as mentioned above, Hangzhou has a business-friendly nature to it. It is far enough from the regulators in Beijing, but it is close enough to the consumers and capital in Shanghai. It distinguishes itself by nurturing businesses of all sizes, not just industry leaders. The city implements a "responsive yet non-intrusive" governance approach, allowing companies to focus on technological innovation and have access to a mature market for go-to-market trials without bureaucratic hindrances.

Alibaba’s founder, Jack Ma, once notoriously said it is a blessing for young talent even to have the opportunity to work for a company that has such tremendous growth potential that “996 (working 9-9, 6 days a week) is a luxury. With that ingrained in the talent pool locally, the city boasts a vibrant entrepreneurial culture - characterized by the sleep-is-for-the-weak mentality and dedication to innovation (for better or worse, it takes that hunger sometimes to build something amazing). I had my fair share of 9-11-6 during my Alibaba days.

Hangzhou has truly benefited from the presence of internet giants like Alibaba and NetEase. These tech giants have brought in some of the country’s best talents and built up an ecosystem in the city over the last two decades and drove the housing market up to one of the most expensive second-tier cities. Coupled with manufacturing leaders such as Hikvision in the city and many family-led manufacturing businesses in the nearby area, the city has created unique opportunities for cross-sector collaboration. (In that way, it’s smilar to Shenzhen, as it has factories nearby all around the city, it has a natural advantage in manufacturing - though SZ more in industrial parts and HZ more for textiles.)

Last but not least, many people are now saying that the role of Zhejiang University in Hangzhou's AI ecosystem is particularly noteworthy, with three of the "Six Dragons" founded by its alumni. In the West, you’ve probably only heard of Peking University and Tsinghua University, but Zhejiang University is probably just trailing after the two in international rankings.

The university has a strong entrepreneurial spirit and is the alma mater of many household names, including Huang Zheng, founder of Pinduoduo (PDD), Liang Wenfeng of DeepSeek, Zhu Qiuguo of YunShen Technology, and Han Bicheng of BrainCo.

See a full list of Hangzhou-based AI companies.

Fashion meets AI

Given all the context, you now know Hangzhou has a huge pool of high-caliber talents in tech, manufacturing, and even fashion. Thus, I want to focus on Style3D, a company at the intersection of these three fields.

Many have argued that AI can replace humans in tedious and monotonous chores (hooray!)., but what makes humans unique is their originality and creativity. However, that doesn't mean AI cannot be integrated into the creative space and increase efficiency for design houses and brands.

Style3D, headquartered in Hangzhou, China, has emerged as a global leader in integrating artificial intelligence (AI) and 3D technologies to transform the fashion industry. Founded in 2015 by Eric Liu, the company has established itself as a pioneer in digital fashion infrastructure, offering end-to-end solutions that streamline design, production, and marketing processes. By leveraging its proprietary physics-based simulation engine and generative AI models, Style3D empowers brands, manufacturers, and fabric mills to achieve unprecedented efficiency, sustainability, and creativity. As we delve into the company’s technological innovations, client base, and strategic acquisitions, we can get a clearer sense of how Style3D is making an impact on the industry and is positioning itself at the forefront of the digital fashion revolution.

Jumping on the AI bandwagon

I met with their global marketing and partnerships lead this past week and was impressed by their impressive pivot from 3D software to quickly adopting AI—call it opportunistic or future-thinking. I was also impressed that they were very aware that if they don’t become the best AI tool for the vertical, they will be replaced. So, instead of rejecting AI, they are leveraging their decades of data, experience, and expertise to capture the 2B market in the fashion business.

The company offers a variety of products, but its core is a 3D design tool widely adopted by industry leaders to help with the design and manufacturing of 3D models. This means it has had 10 years to accumulate the data points, patterns, and know-how of how material, texture, colors, buttons, zippers, and accessories all work together. With over 2,000 clients worldwide, its clients include international brands such as Anta Sports, Shein, YKK, and many other household brands/ accessories names.

Style3D’s core technology is a smart system that turns flat clothing patterns into realistic 3D digital garments. It mimics how fabrics naturally drape, stretch, and interact with virtual models in real-time, considering factors like thickness and stretchiness. A key breakthrough, revealed at a 2024 tech conference, automatically converts 2D sewing patterns into perfectly fitted 3D designs without glitches—a task that previously frustrated designers. By blending AI and advanced math, this tool cuts designers’ workload by nearly half, making the process faster and less error-prone.

Now, being AI-empowered, the company also harnesses powerful graphics processors (GPUs) to create ultra-fast, detailed simulations of clothing movement—something that once required supercomputers. Designers can now see how sportswear stretches on a running avatar or how yoga pants compress during workouts, all with pinpoint accuracy. This tech, showcased at a 2024 textile expo, helps brands ensure consistent fit across sizes and test fabrics digitally before production, saving time and reducing physical samples.

Style3D’s AI+3D+AI Value Chain represents a paradigm shift in fashion design. The company’s self-developed generative AI model, trained on vast datasets of clothing, fabrics, and avatars, enables designers to generate high-fidelity designs through text, sketches, or existing patterns.

This is an example of how experts in specific verticals are adopting AI for their unique use cases. Style3D has built its application on top of Alibaba’s Qwen support and recently integrated DeepSeek into it. With the support from Qwen’s open-source LLM infrastructure, coupled with its own database and tweaks by the technical team to ensure the agent understands every aspect of the fashion industry, it has become a leading choice for AI assistance for fashion designers.

According to case studies from its LinkedIn Newsletter, fashion retailers and design houses use Style3D Cloud to manage digital collections, reducing sample production by 85%. Manufacturer OEMs can utilize 3D/2D connections to synchronize patterns with virtual samples, cutting development cycles from 6 weeks to 10 days. Fabric Mills, like Italian mill Marzotto, employ Style3D Fabric to digitize 20,000+ annual fabric designs, enabling clients to visualize materials on aviators before ordering.

These modules support text-to-design, sketch-to-design, and AI-powered editing, enabling designers to modify specific garment areas (e.g., collars and sleeves) while maintaining structural integrity. For instance, a designer can input “bohemian summer dress,” and its proprietary tool, iCreate, generates multiple 3D renditions with adjustable textures and colors, slashing conceptualization time by 70%. iWish can help generate campaign-ready tools using existing avatar models in 2D and 3D for brands, cutting costs and dependency on photoshoots.

International Ambitions

Chinese AI companies are also quickly snatching up international firms with synergy to complement their international expansion goals. Especially for Style3D, Christian Hu said their products are increasingly popular in APAC, U.S., and Europe. In 2023, Style3D acquired Assyst, a Munich-based leader in the field. This was largely to adhere to data sovereignty requirements for their European clients, as well as for 2D-3D combined solutions and an overall strategic business decision. Due to GDPR-compliant Cloud requirements, Style3D Cloud hosts data in Europe for its European clients.

Style3D’s fusion of AI and 3D technologies has redefined every stage of fashion creation, from ideation to consumer engagement. By digitizing workflows, slashing waste, and democratizing high-end design tools, the company addresses the industry’s urgent needs for sustainability and agility. Strategic acquisitions like Assyst have fortified its global presence, while R&D investments in GPU simulation and generative AI ensure continued leadership. As Style3D pioneers 3D textiles and blockchain-enabled supply chains, it stands poised to lead the industry into an era where digital and physical fashion coexist seamlessly, empowering brands to create responsibly and consumers to experience garments beyond the limits of imagination。

This is just the beginning for many of these AI startups that looking to go global. In the Unitree piece, we explored how Chinese robotics companies are leading in global sales. Following Manycore, the first of the six dragons to go public, quite a few of these mainland AI companies are being courted to establish corporate headquarters in Hong Kong - a neutral pitstop for them to access the rest of the world.

Subscribe now

DeepSeek Myth Busters

As we speak of Hangzhou, let’s take a look at its pride and joy again. Last month, DeepSeek's fundraising caused a media frenzy. We will quickly wrap up what is happening with DeepSeek now as the dust has settled.

1. Rumors about Alibaba’s investment in DeepSeek circulated a few weeks ago, which the tech behemoth quickly rebutted. Baidu VC's arm claimed that it tried to invest in DeepSeek but was rejected because it was not fundraising.

2. There have been rumors that the company has been looking to raise funds for a while now, but The Information just confirmed this.

3. There are rumors that the company will start investing in R&D on chips themselves – not confirmed

4. The reason founder Liang Wenfeng bought a ton of GPUs wasn’t because he’s some AI “prophet,” which is what some of the narratives have been pushing. He bought a ton of GPUs because he had a big $$ year and wanted to use the chips for his quant trading and his hobby in training LLMs then. As he realized quant investing required very few GPUs, he quickly pivoted, doubling down on LLMs. (specifically 10,000 Nvidia A100s in 2021)

5. There’s been a lot of interest in the talent behind DeepSeek’s R1. Some argue that China is experiencing a brain drain. In contrast, others say that Chinese talent is not seeking to pursue higher STEM degrees in the West anymore, especially given the geopolitical tensions in recent years. His team is indeed mostly young, China-trained engineers. As Liang said, China may not have the top 50 AI engineers, but it can easily find the top 500. These hungry young engineers are willing to work and honestly believe that non-hierarchal is the most efficient way to achieve scientific breakthroughs. Many reports have shown that Liang gets his hands dirty and does the coding himself.

6. See the download comparison of DeepSeek and its US peers.

7. According to Quest Mobile (similar to Sensor Tower), within nearly a month of its launch, the DeepSeek mobile app has accumulated 32.9 million daily active users (DAU) in China. It surpassed Doubao (ByteDance) within 15 days post-launch and now has twice the DAU of Doubao. With 32 million DAU, DeepSeek ranks 51st among apps in China currently, for reference (slightly trailing behind WeChat Enterprise and above Meituan Dianping).

8. Since the advent of the AI era in 2016, the AI industry has undergone several rounds of reshuffling. With rounds and rounds of iteration and evolution, we witnessed the rise of the “four tigers” and now the era of the “six dragons.” DeepSeek’s R1 breakthrough has also changed the Chinese AI ecosystem in a seismic way. It seems that, among the ten, only Moonshot AI, Zhipu AI, and StepfunAI are purely focused on large - model technological innovation and the pursuit of AGI.

Full disclaimer: Style3D is a sponsor of AI Proem. All editorial decisions are promised to be made independently. However, during my engagements with various companies, I make sure their business and vision align with AI Proem’s and thus naturally sometimes find their business and innovation as inspirations for stories. All views are of AI Proem’s.

Just last Thursday, Tencent's AI app, Yuanbao, announced its integration of DeepSeek's model, accelerating the domestic race to mass AI adoption. I thought I was fast enough to write an article on the importance of integrating DeepSeek into Yuanbao only two days after its announcement. I even predicted that WeChat might integrate DeepSeek directly. But kudos and jokes were on me, as I woke up to the news that broke the Chinese internet. Kudos, because I called it. Jokes, because it’s a lovely Sunday afternoon, and I’m now chowing down a chicken bahn mi while frantically trying to capture this wild news: TENCENT JUST INTEGRATED DEEPSEEK DIRECTLY INTO ITS WECHAT APP.

Subscribe now

To my own credit, I did call it the "most likely outcome" at the end of the article given Tencent's departmentalism culture.

Basically, while I was sleeping and my U.S. readers were trying to digest the Yuanbao article, WeChat began grey-scale testing of direct DeepSeek integration overnight.

Tencent's Grey-Scale Testing of DeepSeek in WeChat

Said in the most cliche way but no other way possible: Tencent’s decision to integrate DeepSeek’s R1 model directly into WeChat has sent shockwaves through China’s tech industry.

The grey-scale testing phase — where AI Search within WeChat is tested with a limited group of users — has already revealed some transformative capabilities. Users can now use DeepSeek R1 to perform real-time searches across WeChat’s vast ecosystem, including Official Accounts and user-generated content, both of which are walled gardens within the Tencent ecosystem. AI Search also enables intelligent summarization of articles, personalized recommendations, and enhanced search functionalities.

WeChat Grey-Scale Testing - Left: “AI Search” is integrated into the search bar on top; Right: within “AI Search,” can toggle between “fast answer” and “deep thinking” models (which is based on DeepSeek R1); Photo credit Xiaohongshu

1.3 billion monthly active users

Let that number sink in. With over 1.3 billion monthly active users globally in 2024 and more than 840 million users in China alone, WeChat is by far one of the most widely used apps in the world ever. By embedding DeepSeek directly into its interface, WeChat could enable hundreds of millions of users to access advanced AI capabilities almost instantly (let’s wait and see). Unlike what ByteDance had to do with Doubao, there is no need to market this app. Pretty much everyone has WeChat, even my 93-year-old grandma, but no kidding, everyone is on this app already.

This scale of deployment is unprecedented; within days of a full rollout, DeepSeek could become an integral part of daily life for a significant portion of China's population. Such rapid adoption would dwarf the reach of standalone AI apps, even like ChatGPT (and, unfortunately for Tencent’s Yuanbao team, no one will remember them again).

By embedding advanced AI features directly into WeChat, Tencent is positioning the app as a potential AI super-app. WeChat’s unique role as a combination of a social media platform, a mini-program hub, and a professional network makes it an ideal candidate for such an integration. This move enhances user experience and sets the stage for Tencent to dominate the consumer-facing AI space in China.

The GPU Question: Can Tencent Afford This?

The integration of DeepSeek into WeChat comes with significant computational demands. According to estimates from 共识粉碎机 (Consensus Crusher), an excellent tech investing content creator in China, Tencent has recently ordered 100k to 200k H20 GPUs (export restrictions complaint) to support this initiative. Based on Consensus Crusher’s estimate, each H20 GPU can handle up to 30 concurrent users running DeepSeek models.

This means 100k to 200k H20 will enable Tencent to support 3 to 6 million simultaneous/concurrent users. Note that since not all DAUs will use the AI functions simultaneously (the current ChatGPT DAU is 20x concurrent users), 3 to 6 million concurrent users will mean 60 to 120 million DAUs, which should be enough for WeChat’s initial AI user base.

Financially, according to Consensus Crusher, the cost of these GPUs is estimated to be $2 billion upfront, with an annual depreciation cost of $400 million (assuming depreciation over five years, similar to Amazon’s depreciation policy). The $400 million annual cost is less than 2% of Tencent’s net profit last year. Deepseek’s efficient model has made costs low enough for tech giants like Tencent to experiment with new AI features with a massive user base. Thanks to DeepSeek, Tencent has found a cost-effective way to scale its AI ambitions without straining its financial resources.

What Comes Next?

The integration of DeepSeek into WeChat is truly a strategic move that could reshape the AI landscape in China and beyond. Here’s what I think the future could hold:

Future for Tencent

1. WeChat as an AI Super-App: With DeepSeek embedded directly into its ecosystem, WeChat is poised to become an all-in-one platform for communication, content discovery, and productivity tools. Potential features include personal knowledge bases, AI-powered group management in WeChat Work, and even AI agents for summarizing subscribed content or crafting personalized greetings based on chat history. These advancements could redefine how users interact with the app, making it indispensable in both personal and professional contexts.

2. Expansion Beyond Messaging: According to Consensus Crusher notes, WeChat’s vision likely extends beyond efficiency tools to building ecosystems that enable everyone to provide creative services. This suggests that Tencent could leverage DeepSeek to introduce tools for content creation, marketplace functionalities, or even browser-based AI agents. The seamless integration of these capabilities would further solidify WeChat’s dominance as a super-app.

   Share

Potential Impact on Competitors

Tencent’s bold move will undoubtedly put pressure on domestic rivals like ByteDance and Alibaba to respond swiftly and make a move. I have some feelings that…

• ByteDance: To remain competitive, ByteDance may integrate DeepSeek models into its Doubao app. While Doubao currently relies on ByteDance’s proprietary LLM, DeepSeek's growing capabilities might force ByteDance to adopt a more flexible, multi-model strategy.

• Alibaba: Alibaba could expand its Qwen ecosystem by embedding similar AI functionalities into Taobao or Alipay. This would align with its focus on e-commerce and financial services, potentially creating an AI-enhanced shopping or payment experience. BUT, this wouldn't be the most intuitive AI chatbot use case for what we know now.

Broader China AI Implications

1. Accelerated AI Adoption: The success of this integration could serve as a blueprint for other industries in China to adopt AI at scale. From e-commerce platforms to education providers, the ripple effects of Tencent’s move could inspire widespread innovation.

2. Potential Effects on Hardware and Infrastructure: Tencent's deployment of 100k–200k H20 GPUs is already boosting demand for semiconductors and cloud infrastructure in China. This surge in demand will likely accelerate advancements in domestic semiconductor manufacturing and cloud services. We’ve seen Chinese data center operators like 21Vianet and GDS enjoy stock price rises of 143% and 65% year-to-date, respectively.

3. AI Ecosystem Maturity: Tencent’s use of DeepSeek within WeChat could catalyze the development of more user-friendly AI applications across China. By embedding cutting-edge models into everyday tools, Tencent sets a precedent for how AI can seamlessly integrate into daily life.

Conclusion: The Race Is Just Beginning

Not to sound repetitive, but Tencent’s integration of DeepSeek into WeChat marks a turning point in the evolution of consumer-facing AI applications in China. By leveraging its massive user base and unique ecosystem advantages, Tencent has positioned itself as a leader in the Chinese AI race.

However, this isn’t like the finale. In fact, we’re probably only in Season 1, Episode 2 of this whole showdown. The competitive landscape remains dynamic, with Alibaba and ByteDance likely to respond with their own moves. Will ByteDance double down on Doubao’s proprietary model or pivot to embrace DeepSeek? Can Alibaba leverage its Qwen ecosystem to challenge Tencent’s lead?

As technology advances at breakneck speed, the next wave of breakthroughs could come from anywhere. For now, all eyes are on Tencent, which is pushing the boundaries of what an AI-powered super-app can achieve.

See why it was a huge deal that DeepSeek was integrated into Tencent’s Yuanbao/ WeChat and what Tencent’s advantages are now in winning the domestic AI application race.

Last week, Tencent dropped a big one on the Chinese AI world. Because the news came out on Thursday night, it didn’t pick up much media momentum. But it really could be a big deal. I didn’t want to wait until next week to share my thoughts, so here they are.

Please share, like, and subscribe if you find my analyses useful/ insightful.

Subscribe now

TL;DR

Tencent just shook up China’s AI scene by integrating DeepSeek, the country’s top AI model, into its chatbot app Yuanbao (元宝). This makes Yuanbao the first major Chinese chatbot to support external models, giving a kick in the face to competitors like ByteDance’s Doubao and Alibaba’s Qwen.

Why it matters:

1. Multi-Model Power: Yuanbao combines Tencent’s Hunyuan model with DeepSeek, solving performance issues and boosting user experience.

2. WeChat Advantage: Tencent leverages WeChat’s massive content ecosystem, creating an AI super app potential.

3. Fast-Follower Strategy: Tencent’s siloed teams and “copy-then-innovate” culture let them adopt DeepSeek without internal resistance.

DeepSeek’s success seems to have suddenly given ammunition to China’s AI, putting it officially on the global map. ByteDance released an updated version of its flagship AI app, Doubao, with sparse MoE architecture. Only a few days later, on February 6, 2025, Alibaba’s Qwen 2.5 Max was released. Then Baidu announced it would release its Ernie 5.0 model with multi-modal capabilities. Chinese domestic AI competition is on fire.

But I think the real game-changer is Tencent’s mic drop. After staying hush-hush while the rest of the big tech companies tried to out-bling each other, Tencent came out with a disruptive app that has the potential to become a killer app.

This past Thursday, it announced that it has integrated DeepSeek into its own, much-unnoticed AI chatbot app, Yuanbao (元宝, which literally means “gold nugget” in Chinese). It would be as if ChatGPT decides to integrate Gemini models besides OpenAI’s models.

What makes it so meaningful is that Yuanbao is the first chatbot app backed by a major Chinese tech company that supports models other than its proprietary model. Before DeepSeek, Yuanbao only supported Tencent’s internally developed LLM, Hunyuan (混元), just like Doubao only supported ByteDance’s own LLM.

Share

Why do I think this is a significant development? Let me explain my thinking:

Tencent Yuanbao’s Advantages

A. Uniquely positioned to be multi-models

I wrote before that one of the key advantages of “GPT wrappers” like Perplexity is that it can use multiple LLMs, each for what they are best at (for example, use Claude 3.5 for coding, use OpenAI o3 for reasoning). On the other hand, OpenAI would not open up its ChatGPT to other models, including Claude, Gemini, or DeepSeek. In Tencent’s case, before DeepSeek, its Hunyuan model has been the least competitive out of the Chinese big techs, which made its app Yuanbao the least competitive (who has heard of Yuanbao before today? If you read my pieces before would have known that Doubao has been the dominating consumer-facing AI app in China).

If we just look at LLM capabilities, Alibaba’s Qwen has been leading, with ByteDance’s Doubao trailing behind and then Baidu’s Ernie. By integrating DeepSeek’s R1, which arguably leapfrogged Alibaba and became the best LLM in China today, Yuanbao’s capability and user experience suddenly became a top-notch product.

You may ask, I saw that DeepSeek has become the number one downloaded app globally, so why can’t people just use DeepSeek directly? That is because the demand for DeepSeek is so high that the company has been having trouble keeping up. Frequent users of DeepSeek would know that very often, DeepSeek will tell you that “servers are busy, please try again” as they are not able to serve all the inference demand. As a result, user experience in DeepSeek has been very inconsistent, and I would not be surprised to see if the users churn out quickly (even when I got impatient and just went back to Perplexity).

Now, DeepSeek’s inconsistent capability to keep up with demand does not affect users that use DeepSeek via Perplexity, which hosts DeepSeek’s model in the U.S. Which has done an incredible job of increasing its DeepSeek inference servers to keep up with the demand (its CEO mentioned that it was doubling servers for DeepSeek every day).

However, in China, there are no major venture-backed wrapper companies like Perplexity. Big tech firms’ chatbot apps only support their own proprietary models. This gave Tencent an opportunity to re-vamp Yuanbao as a “DeepSeek wrapper”. As the inference demand on Yuanbao will be served by Tencent cloud, Yuanbao can offer DeepSeek’s model capability with far more consistency, which should allow it to acquire a large number of users quickly. We will go over in a later section why, among all the big tech firms, Tencent is able to host external models alongside its proprietary model.

B. Walled garden content/data

Just like how Google Gemini prides itself on crawling and scraping YouTube’s data better than others and providing the most comprehensive research, leveraging its access to all the data under the Google sun. Yuanbao’s competitive advantage here is its full access to content within the WeChat ecosystem. And I’m not talking about your chit-chat history with your parents. In China, WeChat is beyond just a messenger app; its content features are similar to a combination of X and LinkedIn, where users generate many long-form and short-form content. In fact, there are over 20 million self-run content accounts on WeChat as of 2023, based on online estimates.

That means that Tencent can access billions of articles and videos circulated on WeChat. If you were to use Baidu Search (China’s Google), you would not easily find articles published within WeChat. Such data and content advantage, now coupled with the most frontier model, DeepSeek, should allow Yuanbao to offer a superior experience.

C. Functional adjacency

A third advantage, although quite subtle, that Tencent has is user intuition. As mentioned before, WeChat is not just a messenger app, but it also has a lot of content with an embedded search engine within WeChat. Users already use the search bar at the top of the app for news, tutorials, and even direct links to other programs: some called “mini-programs,” which are APIs plugged in from other apps. So, for an AI chatbot to be embedded into WeChat as a new chat window or search engine makes complete sense to users.

In comparison, its main domestic consumer-facing competitor, Doubao, wouldn't have such “functional adjacency” right now. ByteDance’s main app, Douyin (China’s version of TikTok) and Doubao, have little “synergy” in user experience. It’ll be a weird situation where you’re scrolling through 20-second videos and suddenly want to switch to an AI platform to look up something. It's not completely unlikely, just not as seamless of a transition of all being embedded into one app - and, to my points before, an AI SuperApp.

Functional adjacency refers to the degree to which a new feature or functionality is contextually aligned with the primary purpose or use case of an application. When users associate an app with a specific function, their expectations are shaped by that core purpose. Features that align closely with this primary function feel more "natural" to users, while those that deviate significantly may face adoption challenges.

[A minor comment: Yuanbao’s main interface is quite clean, like ChatGPT, but it allows you to zoom out of the window and go to a page with various “specialized” apps for different purposes (similar to the App Store on iOS). The multi-app format is similar to what Doubao has so far. See here for more on China’s AI app strategy.]

Okay, that’s my take on why Yuanbao has the potential to be a massive win for Tencent. Below is my take on why Tencent is the only big tech that is doing this. How was this possible? Why would Tencent allow Yuanbao not to be a loyalist to Hunyuan LLM?

Subscribe now

Tencent’s Unique Culture

A. Internal competition and departmentalism could have its benefits

It is widely known to China tech insiders that within Tencent, the internal teams work in silos (山头主义, or “departmentalism”) and the company has always been notorious for its internal-competition (赛马, or “horse-racing”) culture.

When ByteDance sets clear, top-down objectives to develop the best AI app, Doubao, everyone is expected to work together to achieve this goal. The internal LLM team would work closely with the Doubao team.

Tencent, on the other hand, has their different businesses in silos. WeChat is one separate team (WeChat’s HQ is in Guangzhou, while the Tencent HQ is in Shenzhen, nearly 90 miles apart). Tencent’s LLM Hunyuan is from a separate team. Yuanbao is from another team.

One consequence of different teams working in silos is that they do not leverage the entire firm’s resources and advantages. For example, Tencent’s ads business has much lower conversion and monetization rates than competitors like ByteDance and Kuaishou, despite having the most personalized data on users from WeChat. The reason is that WeChat is run like a separate business within Tencent. Some in the industry even say that the WeChat team is not sharing its full data with the Tencent ads team. Can you imagine Meta’s ads team not having full access to Instagram and Facebook data?

Back to AI. Rather than having all teams working together to develop its frontier LLM and AI products, Tencent’s LLM Hunyuan has been developed in silo, and in many ways unnoticed and not very impactful. Hence, when DeepSeek came out, the separate Yuanbao team could independently decide to integrate DeepSeek’s model in its products without caring too much about internal politics and “hurting the Hunyuan team’s feelings”. Departmentalism has its benefits sometimes.

B. Roots in fast-following

Another unique thing to Tencent is that it is often not the innovator but a very strong fast-follower. It would wait for competitors to battle it out at the frontlines, watch the winners come out one-by-one with their proven strategies, and then strike with fast copying.

In fact, today, Tencent's crown jewel, WeChat, is a result of its fast following. In 2007, China’s state telecom, China Mobile, launched a messenger app called Feixin, which was gaining traction. In 2011, when Feixin's “winning formula” matured, Tencent rolled out a copycat, WeChat (Weixin), which eventually took over the whole market.

Another example is Alibaba, which launched its payment app, Alipay, in 2004. In 2013, Tencent launched WeChat Pay, which is a payment app integrated into WeChat. Despite being late to the game, WeChat Pay leveraged WeChat’s massive distribution reach to become China’s number two payment app quickly.

Once more, ByteDance’s Douyin was released in 2016, and Tencent launched its copycat WeChat Video Account in 2020. Again, leveraging WeChat’s massive reach, it is now China’s number two short-video app.

You get my point.

You can say Huanyuan is another attempt to compete with ByteDance’s Doubao, which has shown the potential to become the first killer AI app in China. If Tencent could leverage DeepSeek’s model to displace Doubao, it would be the most “Tencent” thing everrr.

Tencent is a perfect example of what Uber founder Travis Kalanick said: many Chinese tech companies were extremely good at copying and copying, then innovating.

But what’s so impressive with Tencent is that they’ve always been able to be #1 or #2 in whatever they do (copy). In gaming, they are the largest gaming distributor globally. WeChat, its super app, is the most used app in the country by far. Basically, in everything else they do (copy), they’ve gotten to the 2nd place, and they have been okay with that. I mean, you don't need to be a winner at everything.

Parting Words

We have discussed the many ingredients for Yuanbao’s potential success (the DeepSeek model, walled garden data/content, and functional adjacency) and why Tencent’s unique culture has played a key role in cultivating Yuanbao.

Now, the question is: Will the Yuanbao app actually take off (like how DeepSeek has become the number-one downloaded app in more than 140 countries)?

If it does, will Tencent formally integrate Yuanbao into WeChat, making itself China’s AI super app (like Meta aims to integrate Meta Assistant into Facebook, Instagram, and WhatsApp)?

Or, alternatively, if not through Yuanbao, would WeChat directly integrate DeepSeek and embed an AI chatbot into its functions? Given the departmentalism culture, this is probably the most likely scenario.

Looking forward, it will also be interesting to see how the arms race plays out and whether ByteDance’s Doubao will soon integrate DeepSeek models to remain competitive with Yuanbao. Yes, clearly, ByteDance wants to use its own LLM, but when faced with competition, maybe keeping Doubao as the number-one AI app in China is more important.

One thing we know for sure is that, thanks to DeepSeek’s open-source model, the Chinese domestic AI race is on — like, really on.

See follow-up analysis of DeepSeek’s integration into WeChat directly: Integration at the Speed of Light: Tencent's WeChat Embeds DeepSeek

Appendix: For the sake of the Chinese domestic AI race, in case you were getting lost with the various apps and how they’re all related. Let’s just look at these three big tech ecosystems. The three companies below make up the BAT: ByteDance/Alibaba/Tencent.

Subscribe now

TL;DR

China’s incumbent big tech have 1) access to “good enough” frontier model DeepSeek; 2) crazy amount of user data; 3) multiple touchpoints across various apps in their own ecosystems; 4) less stringent regulatory control of data sharing internally; 5) hustle work culture (unlike U.S. big tech, no one’s got time for J2, J3s, they’re all barely staying afloat in J1); 6) crazy sales force; 7) innate commercialization DNA; 8) ability to copy, copy, then innovate.

There is a chance that Chinese tech may lead in consumer-facing AI applications. I will dive deeper into the U.S.’s lead in enterprise AI application in a separate piece.

R1’s Significance to China’s AI: Open-Source Momentum

China’s LLM models have often been perceived as lagging behind U.S. giants like OpenAI, Anthropic, Meta, and Google. However, with DeepSeek’s breakthrough in the large language model, China’s AI development is finally being taken seriously by its Western peers and investors.

Many investors believe that Chinese big tech, now leaning into DeepSeek’s breakthrough with a frontier model, coupled with its innate commercial sense and competitive 996 work culture, might be able to create the next AI super app.

Meta’s Yann LeCun noted that R1’s success exemplifies the power of open-source collaboration. At this rate of adoption, it’s only a matter of months before China’s internet tech companies can combine the strength of their consumer know-how with the open-source model to accelerate AI application innovation. In fact, just yesterday, it was announced that Apple will partner up with Alibaba to develop AI features for the Chinese market.

The R1 model breakthrough represents a milestone in China's artificial intelligence landscape. As an open-source solution, R1 has quickly been adopted by China's tech giants, including Tencent, ByteDance, Huawei, and Alibaba, and integrated into their cloud infrastructure.

Recent developments in China’s AI sector have officially re-established its tech and AI capabilities on the map for global investors after a prolonged hiatus.

My point here is not that if China is to lead the innovation breakthrough in AI applications, then it is a loss for the U.S. I think it’ll just mark the start of the application hustle and the start of global consumer adoption of AI.

The West’s Shock to China’s Innovation

Fast-forward to the present. We’re at a pivotal point where value is migrating from models to applications. “China has always been very good at apps… the data, talent and capital is there,” said Alan Lau, CBO, Animoca Brands, and Vice Chairman of M+, at a HKU AI event last week.

His comment referred to the shock the West experienced at DeepSeek's success, especially the inquiries he received when attending Davos, where business leaders congregated. However, he said DeepSeek’s success, driven by domestic talent, did not surprise him at all. He believes that as value creation shifts to the application end now, China has an advantage in being the first.

China’s Super-App DNA

He recalled that in the early 2000s, when he was at Mckinsey, he would bring Chinese internet entrepreneurs to Silicon Valley to learn what they were doing. Still, by the 2010s, the tables had turned, and he was helping American entrepreneurs on tour to Shenzhen and Shanghai to learn what was happening at the frontier of tech innovation. In his eyes, China’s tech will very likely lead in application innovation.

In the most recent earnings call, Meta's CEO, Mark Zuckerberg, announced his recent vision of a “1 billion-user AI assistant” that largely hinges on Meta’s vast data and engagement. With that logic, this ambition could be applied to any of China’s tech behemoths, too.

Tencent (1.3 billion WeChat users), ByteDance (Douyin, which has 800 million DAUs), and Alibaba (over 900 million MAUs) have arguably even deeper integration into daily lives and could even tap into the data across platforms/touchpoints even easier with fewer regulatory hurdles.

Share

China’s Multi-Modal Models

Consumer-facing AI has primarily followed a utilitarian approach in the West. ChatGPT, Perplexity, and Claude function as productivity tools—chatbots designed for research, coding, or content creation. Their interfaces are minimalistic, and their use cases are more narrowly defined. While innovative, Perplexity’s recent update for its Android assistant still prioritizes task completion: booking rides, drafting emails, or summarizing articles. This method has its advantages—85% of ChatGPT’s mobile users are male, with a significant skew toward professional use—but it risks limiting AI’s consumer use-case potential.

While Western AI tools like ChatGPT and Perplexity remain streamlined, single-purpose “answer engines,” Chinese platforms like Bytedance’s Doubao, Alibaba’s Kuake, and MiniMax’s AI suites are pioneering multi-modular interfaces that could redefine how users interact with AI and how companies monetize it.

Echoing back to China’s internet era, super-app infrastructure was widely shared to create more touchpoints for consumers and make applications feel more personalized and integrated into daily lives. This could be similar to AI applications.

Subscribe now

Different UI/UX Preferences

In the West, we’re accustomed to minimalist app UI/UX designs. We have separate apps for each function, such as Amazon for shopping, Netflix for visual entertainment, and Spotify for music and podcasts. These tools excel at efficiency but function in silos. On the other hand, China’s mobile internet ecosystem is designed quite differently. Alibaba apps, like Taobao, feature a unique interface architecture with options for searching, purchasing, recommending, viewing videos, and using VR fitting rooms, among other things. Tencent Video boasts one of the busiest interfaces, filled with ad pop-ups, viewer commentaries, and recommendation buttons. You can even keep a virtual pet on Taobao or water virtual trees on Alipay, which can be exchanged for real tree planting by someone in the western desert regions of China if you consistently log into the app for an extended period. WeChat enables you to make video calls, voice chat, shop, pay utility bills, transfer money, and even order food or a taxi. All is to maximize touch points and engagement.

Similar trends are emerging for AI applications (see popular AI apps from China). Bytedance’s Doubao resembles a command center, featuring modular windows that serve diverse roles such as English tutor, healthcare provider, love guru, joke curator, 3D asset designer, and even a virtual Elon Musk. Each window functions as a specialized “AI agent,” with pricing tiers enabling Bytedance to monetize premium features while keeping the core query-based chatbot experience uncluttered.

While I’m not particularly fond of excessive pop-ups and multi-module features myself, I acknowledge that a more complex app with role-driven functions can enhance both business and entertainment value. When the demand for monetization arises (which it will), integrating native advertisements or paid content will likely feel much more seamless and less disruptive to the user experience in these already busy separate modules than the sleek interfaces currently provided by Western chatbots. As a regular user of apps from both universes, I can’t say one is better than the other, but if I must, Chinese mobile apps thus far have really shown creativity that is not seen in the Western ecosystem. (I.e., TikTok and RedNote that you’re now familiar with)

China’s “Copy, copy, then it flips to innovation”

Now, I don't think this divergence was accidental. In November 2024, Chinese mobile users spent an average of 7.91 hours daily on apps—mainly across the five WeChat, QQ, Douyin, Taobao, and Alipay. Compared to the U.S., where the average smartphone owner uses 10 apps per day and 30 apps each month, Chinese consumers have grown accustomed to super apps with “busy” UI/UXs that prioritize versatility over simplicity. Compared to U.S. app fragmentation, these all-in-one ecosystems (mainly due to less stringent antimonopoly laws) allow these super players to accumulate massive user bases, which they can leverage for AI apps.

Uber founder Travis Kalanick recently stated on the All-In Podcast that ten years ago, when he was trying to compete with Didi in China, he was both impressed and shocked by the speed and effectiveness of the Chinese tech company’s (Didi Chuxing) copying.

“So, but what happens is when you get really, really good at copying and that time gets tighter and tighter and tighter and tighter and tighter, you eventually run out of things to copy, and then it flips to creativity, to creativity and innovation,” said Kalanick.

But at which point does the copying turn into innovation? I think we’re starting to see that happen now in AI.

It’s the super app playbook, rooted in a decade of experimentation: make yourself indispensable and monetize at friction points.

International Investment Implications

With China’s super-app ecosystem, which presents a mosaic of touchpoints (payments, social media, e-commerce), it is clear that there is also a fertile ground for potential AI monetization.

Critics argue that China’s LLMs still trail OpenAI. But DeepSeek’s R1 model—matching GPT-o1 at a fraction of the cost—suggests the gap is closing. It’s noticeable that if China’s big tech firms are all trying to catch up to frontier LLMs, starting with Alibaba’s latest Qwen 2.5 Max, and Minimax’s Kimi K1.5, Chinese firms are pivoting to application-layer innovations, much like Tencent did with WeChat after lagging in early mobile OS wars (which was once dismissed as a WhatsApp clone).

China’s ecosystem could birth the first AI super app—a platform blending search, entertainment, commerce, and OS-like extensibility. Eventually, it will be about designing the most immersive platform the masses can use (again, for better or worse).

Ultimately, I believe the application race is not a zero-sum game. LLM incumbents like OpenAI can still maintain an advantage in their businesses across 1) model dominance, 2) top-enterprise choice API, and 3) GPT can still capture the majority of market share for the chat-bot function globally.

However, going forward, with the most frontier open-sourced models being embedded into big-tech workflows, what an AI app looks like may no longer just be in a GPT-bot form. That’s where application innovation comes in.

This essentially opens up the potential AI application market. I reiterate what Brian Wong has said here: This isn’t a win for China or a loss for the U.S. It’s a push forward for AI application innovation as a society. This means that the future AI super app, whether Chinese or American, may not matter to consumers (just as whichever model AI wrappers use also does not matter to consumer users).

Last but not least, providing some extra context: DeepSeek’s success and penetration across age groups (users under 19 to over 50) during the Lunar New Year was fueled by the right timing and some patriotic fervor domestically. What makes me think it is not just a fad but rather a signal that Chinese AI may finally be able to be a key contributor to the global innovation ecosystem is that it blew up not only in China. In fact, it topped download charts in over 140 countries within days of its launch, echoing TikTok’s global rise and highlighting a point: Consumers care less about geopolitical tensions than about accessibility, product quality, and user experience.

China’s AI and tech sector resurgence is not just about replicating U.S. models, as some would like to say, but innovating on top of them, as we saw during the internet era. DeepSeek’s founder, Liang Wenfeng, has also advocated this vision. For Chinese companies to really contribute to the global innovation ecosystem, they need to innovate in hard tech and continue to think outside the box in software (where it has the competitive advantage).

Please subscribe, share, and like if you enjoyed this analysis.

//

NEWS FLASH: **Baidu just announced they will release a next-generation model this year. **Tencent Cloud Development announced a major update on February 11, 2025, introducing a new feature that simplifies AI integration for WeChat Mini-programs. Developers can now integrate DeepSeek's large language model into WeChat Mini-programs with just 3 lines of code.

Last Friday, I was invited to a Hong Kong University event on DeepSeek’s implications for the AI industry, which was hosted by the AI & Humanity Lab and the Hong Kong Ethics Lab. During the panel session on financial and business implications, Vince Feng (Adjunct Associate Professor, Faculty of Business and Economics; Former Managing Director and Head of North Asia of General Atlantic) said that AI applications will be where most value will be created and captured.

We’re still in the early days of AI. Like many others, he also used the metaphor that AI models are similar to building fiber networks during the early days of the Internet. He added that during this era, fiber companies laid out fiber networks nationwide, which enabled internet application companies such as Amazon, Google, and Facebook later on. Even then, it took a few app iterations to get to where we are now.

What he said resonated with me, as I have been thinking about how future value creation and value capture will happen in the application layer instead of the fundamental model layer.

Subscribe now

TL;DR

• The AI industry's value is shifting from foundational models to applications that blend multiple AI models for specialized tasks.

• Frontier AI models (GPT-4, Claude) face rapid depreciation as open-source alternatives (DeepSeek) erode their moats.

• Apps like Perplexity and Cursor, once dismissed as “GPT Wrappers”, now dynamically combine models for superior task-specific performance, outperforming single-model platforms.

• Competitive advantage lies in product execution, domain knowledge, and network effect, not model ownership.

• Value now stems from solving real problems via adaptable AI integration, not owning the "best brain."

• Frontier model providers may need vertical specialization (like cloud providers) as differentiation through generic models fades.

Investors have recently debated whether companies will and should continue to invest heavily in frontier models or whether it is time for the next phase of AI innovation, the application layer. Many argue that value creation is migrating from core model development to application-level solutions. Just like in the internet era, none of the fiber companies became trillion-dollar businesses, but a few of the application winners eventually did.

Gavin Baker, CIO and Managing Partner at Atreides Management, has spearheaded this conversation in Silicon Valley. He says that frontier LLMs will depreciate quickly and be commoditized sooner rather than later. He believes that the opportunities are really on the application end.

What are “GPT Wrappers”?

When we think of AI applications today, we often think of OpenAI’s ChatGPT, Bytedance’s Doubao, or these days, DeepSeek. Still, to many people’s surprise (my surprise, at least), beyond these, many of today's most popular AI applications do not use a single (self-developed) model to power them. They operate more like a restaurant chef - sourcing ingredients from multiple suppliers (models) while crafting unique experiences.” For example:

• Cursor (an AI-powered code editor with a $2.6B valuation) now dynamically routes queries between Claude 3.5 Sonnet (for code refinement) and GPT-4o (for documentation synthesis), achieving 40% faster developer workflows than single-model approaches.

• Perplexity (AI-driven answer engine; $9B+) maintains redundant model capacity like cloud providers keep backup generators, instantly switching between four AI providers during peak loads.

• Glean (Enterprise AI platform; $4.6B) blends six AI models simultaneously in its enterprise search product, using proprietary algorithms to weight outputs based on query context.

These were all once called a “GPT wrapper”, a dismissive term that refers to a lightweight application or service that uses existing AI models or APIs to provide specific functionality, typically with minimal effort or complexity involved in its creation, wrote Alvaro Vargas, Founder and CEO of Frontline. A common viewpoint back then was that all those wrapper businesses would be replaced by the frontier models themselves one day. For example, Cursor is an AI coding assistant built on frontier models’ APIs (in its case, Anthropic’s Claude). Wouldn’t Anthropic one day replace Cursor and help users with coding directly?

“GPT Wrappers” over ChatGPT?

The assertion that “GPT wrappers” have little moat is now being challenged. YC partners said during an episode of the Lightcone Podcast that calling a startup a “GPT wrapper” for building on top of OpenAI’s APIs is the equivalent of calling a SaaS company a MySQL wrapper for building on top of a SQL database.

In any case, the moat of frontier models is now being questioned. Since DeepSeek’s release of an open-source model that is free for everyone to use and has capabilities matching OpenAI’s o1. This really raised the question of whether billions of dollars in investment in frontier model development have been significantly impaired or whether ROI will ever make sense. Previously, people thought frontier models had moats in large capital requirements and technological leadership, but now, neither appears to be true.

Another uncomfortable truth for frontier model companies is that no single model can be the best at everything. For example, OpenAI's GPT-4 leads in voice interaction parity (87% indistinguishable from humans in blind tests), Anthropic's Claude dominates complex coding tasks (62% adoption among developers), and Google's Gemini maintains search integration advantages.

This plays into the advantage of AI application companies. Suppose I’m using Perplexity. Hypothetically, for the sake of illustration, let’s assume that OpenAI models answer physics questions best, Claude is best for coding, and Deepseek is best for writing. Then, depending on the type of question I ask, Perplexity will call on the model that is best suited for the task. As a consumer, I won’t care or even know which model Perplexity used, so long as the user experience is good (i.e, the answers are satisfactory). Now, imagine ChatGPT by OpenAI; they are limited to using the OpenAI models, which means they will give worse answers to some of the questions, which is, in turn, worse for the customer experience. All of a sudden, the “GPT Wrapper” seems to have the upper hand against GPT itself.

This is happening to many AI application companies, which are focused on solving specific issues with the best-suited model for the customers. This could be a paradigm shift regarding where value will be ultimately created and captured across the AI technology stack. This transition also means that competitive advantages or moats stem from effectively deploying AI with end customers rather than possessing advanced AI models. Success will increasingly depend on product-market fit and old-school execution rather than technological breakthroughs.

Much like during the Internet era, it wasn’t the ones who owned all the routers and optic cable infrastructure that dominated the consumer market; instead, it was the most innovative use cases that enhanced our whole experience of buying, food ordering, and taxi-hailing that won over consumers.

Similarly, the AI revolution's second act won't be about who builds the best brains but who designs the most indispensable tools with the best user experience.

Of course, the biggest caveat to the above prediction is that one frontier model company can come up with one model that rules them all, being best at all tasks and across all capabilities, and there is no way for other model companies to catch up. [Preview- I’m working on a story about China’s super app DNA x AI application]

Then, that company will build the best applications, leveraging its best model, which remains proprietary. However, we are seeing the industry moving in the opposite direction so far. For one, multiple frontier model companies are fiercely competing, and each is better at certain tasks. Meanwhile, Deepseek has quickly caught up to frontier model capabilities and even open-sourced its model.

Model-Agnostic, just like Cloud-Agnostic

During the cloud era, hyperscalers like Amazon AWS, Microsoft Azure, and Google GCP have monopolized the cloud infrastructure layer. This layer requires a massive fixed investment in land, energy, and servers. However, their moats in the infrastructure layer did not translate into dominance in the application layer.

For example, when you use Netflix, would you care whether they host your data on AWS or Azure? Probably not.

For one, I had never checked which cloud provider Netflix uses before writing this article (spoiler alert: it is mainly AWS, so you can call Netflix an “AWS Wrapper”, I guess). What is more interesting is that despite Amazon having the leading position in the infrastructure layer while Netflix has none, Netflix hands-down beats Amazon’s Prime Video.

The same is true not only for consumer applications like Netflix but also for enterprise applications. For example, Salesforce is the leader in enterprise customer relationship management (CRM) software. Do Salesforce users care or know which cloud their data is stored on? Probably not. Similarly, neither Microsoft nor Amazon came up with the leading CRM applications, despite having such a huge lead in cloud infrastructure.

So, why did the cloud application companies win? The answer lies in vertical expertise and customer reach. Salesforce has been the leader in CRM with deep domain-specific knowledge (CRM for healthcare, CRM for finance, etc.) and seamless integration with customers’ existing workflows. Moreover, Salesforce was already the leading CRM provider before the cloud era, which means it has vast customer reach and distribution advantage.

I believe the same will play out in the AI era. First, as consumers and enterprises start getting used to the AI applications they like, they will become LLM-agnostic, just like they are cloud-agnostic in most cases today (caveat: obviously, not all clouds are created equal, and there are certain cases where customers will prefer one cloud provider to the other, but you get the point I am trying to make here). Second, existing application companies with domain knowledge, distribution, network effect, and high switching cost will likely remain winners in the AI era, leveraging multiple frontier models (just like how Salesforce remained the CRM leader in the cloud era, leveraging the cloud service providers).

I am not the only one who holds this view. Andrew Chan, a partner at a16z, explains better than anyone else that distribution and network effects are still key to competitive advantage at the application layer.

Future of frontier models?

One may ask, okay, maybe the frontier model companies aren’t application winners, but becoming the AWS, Azure, and GCP in the AI era is not a bad outcome. These three are massive businesses, after all! I’d say that was many people’s ambition before DeepSeek. Before, it was widely assumed that building frontier models would require a large amount of capital, just like building cloud data center infrastructure, which means only a few large players will dominate this space. However, what DeepSeek has shown is that the same frontier model capability can be achieved at much lower cost, and more importantly, it open-sourced its models (this is like someone invented a much cheaper server, and gives it away for free!). Frontier model companies are really in a tough spot.

I spoke to , a cloud expert and former Microsoft Azure engineer, about whether comparing model-agnostic to cloud-agnostic made sense. He offered some of his wisdom, which I believe can serve as a roadmap for what the frontier model companies could do.

In the early days, cloud providers didn't have a strong vertical focus, primarily focused on infrastructure-level services. “AWS, Azure, and to some extent GCP were more or less competing on price,” he added. However, AWS, Azure, and GCP developed advantages over time in certain verticals and use cases. Now, the reason an enterprise chooses one cloud server over another varies and can be complicated.

This is a valuable lesson for today’s frontier model companies. Differentiated model products ultimately can only come from differentiated expertise. So it seems like AI application developers now have three most likely outcomes: 1) develop sector or vertical specific advantages, 2) leverage and capture existing distribution and network advantage, or 3) wait to be commoditized.

[For more on U.S. vs. China’s cloud infrastructure comparison, see an article I wrote in collaboration with JS Tan.]

Additional thoughts

Please feel free to leave a comment or DM to share your thoughts on this: AI applications, as of now, are still being accessed through what we know - on edge through web pages or mobile apps. But what if someone comes up with a completely different way of engaging with AI, not just an interface but a completely new form factor that we’ve never seen? Think Nokia; they had apps through the early “smartphone” era (I had my red/white brick phone and thought I was the coolest person on the block; anyone here with me waawhaat?). Still, Apple reinvented what a smartphone means, which became the default form factor for mobile internet. This completely changed how mobile apps are designed, and the rest is history.

Subscribe now