But it’s been two months since that initial frenzy, now what? Why are Alibaba and Tencent throwing massive events to talk about their enterprise strategy of integrating this open-source software at their cloud summits?

Tong didn’t talk about the lobsters as a consumer product. He talked about them as the forcing function that finally made enterprises take agents seriously. How could this be an opportunity for Tencent, which has been a bit slow in AI?

“AI deployment is not just an algorithm problem. It’s an engineering problem.”

That line was the thesis of his entire keynote. And it turned out to be the thesis of both summits.

Tencent thinks it can win because it has better distribution and ecosystem than everyone else. Alibaba thinks it has a better developer ecosystem and cloud service than everyone else. Let’s rewind.

Two days earlier and a thousand kilometers away, about a hundred developers, cloud architects, and a handful of investors had gathered at the glossy space shuttle-looking office building that is supposedly the “most expensive business real estate in Hong Kong” – The Henderson, for Alibaba Cloud’s “Data+AI Forward” summit, co-hosted with AMD.

Although the target audience was much more client-facing and the discussions largely focused on data storage and add-on services offered by Alibaba, some themes were the same. It was equally about building the harness, context engineering, agent orchestration, skills marketplace, and enterprise governance.

Even MiniMax, the fast-rising model startup, appeared at both events in some capacity — at Alibaba’s stage as a case study in cross-stack data support, at Tencent as a guest speaker integrated into their TokenHub marketplace. The ecosystem is so intertwined yet so cutthroat at the same time.

What was clear was that the consumer battle was halfway done, or at least they realized that ROI is not going to be there anytime soon. As OpenClaw took off, a new wave of optimism swept through the management’s ranks. This could be an opportunity to court enterprise clients. Sell tokens? Full-stack service? Cloud, the foundation of it all.

The Consumer Party Is Over

For most of 2025, China’s AI story was a consumer story; we’ve written about it here and here. It was about capturing mindshare, super-app integrations, viral chatbots, the race for monthly active users between Doubao, Kimi, DeepSeek, and a dozen others.

Li Qiang (李强), Tencent’s Group VP and President of Government & Enterprise Business, put the scale in perspective. He started his keynote showcasing a big screen indicating that there are over six hundred million generative AI users in China, that is, every second netizen. Token prices collapsed by 99% over two years, from 50–100 RMB per million tokens to a few RMB or even cents. Daily token consumption hit 14 trillion by March 2026. These numbers cannot be ignored; they mean that, in some capacity, “AI for all” 全民AI has been achieved.

As we all know, since the start of 2026, the “AI for all” narrative has been taken to the next level, as we have seen a frenzy around claw installations. And “raising lobsters” (养虾) has become both a cultural meme and a product category. Millions of users traded deep system permissions for automated productivity. Every big tech under the sun scrambled to ship one. But the commercial lesson of that moment wasn’t about consumers. It was about their limits.

The consumer agent frenzy demonstrated that lots of people wanted agents. Nobody wanted to pay for them. Personal users will happily trade privacy for convenience, but they won’t pay enterprise prices. And open-source models have made the consumer-facing AI layer almost impossible to monetize directly — a reality that loomed over both events, even when it wasn’t spoken aloud. And the new battlefield, as Wu Yunsheng put it, was moving from AI that 'can answer' (能回答) to AI that 'can do' (能做事) — making ‘好用的AI’, AI that's actually useful."

Meanwhile, at the Alibaba summit, the recent leadership reshuffle within the Qwen team went entirely unmentioned. No one addressed what happened, why it happened, or what it means for the model roadmap. Qwen was talked about as the backdrop infrastructure support. The Qwen family was presented in its full regalia — Qwen3-Turbo, Qwen-Max, Qwen-VL, CosyVoice, and the Wan video generation suite. For a company staking its cloud growth narrative on open-source model leadership, this silence was, in its own way, a telling data point showing its strategy pivot.

The unspoken implication at both summits: open-source models are strategically essential but economically punishing in their own right. Maybe at this point, the pressure to see ROI in AI is becoming too heavy to bear.

What everyone knows is that enterprise clients are where the money has to come from. China’s internet era has always given excuses for a lack of willingness to pay. But when tokens become gas and water, and ‘AI is turned on like tap water’ as Tencent’s Li Qiang puts it, then maybe now, is when companies are going to pay, or must pay.

But my question is, why OpenCLaw now? Well, enterprise agents, governed agents, agents that live inside corporate workflows and generate cloud consumption, these are the ways Alibaba and Tencent intend to collect.

The WeChat question I have

When OpenClaw first became viral, WeChat did not integrate it. Even though it felt like Tencent led the strategic embrace and was the most hyped up about it, there was a lot of caution for WeChat, the main app at first. Allen Zhang Xiaolong (张小龙), WeChat’s legendary product manager — the person who has historically guarded the platform’s simplicity and user trust with near-religious discipline — resisted.

For him, the risks were that agents with deep system permissions operating inside WeChat’s social graph raise profound privacy, security, and compliance concerns. A misbehaving agent in a personal chat is embarrassing. A misbehaving agent in a WeChat group, one that can read messages, trigger workflows, and access contacts, is a liability at societal scale.

So what happened was that the first integration was through WeCom (企业微信), the enterprise version of WeChat. Now, this was definitely a conscious move. Enterprise use cases are more controlled — fixed user populations, IT governance, compliance frameworks, and defined workflows. An agent operating within WeCom has a bounded sandbox. The company sets permissions, the IT admin reviews skills, and the security layer audits every action. Wu Yunsheng (吴运声), speaking from the Tencent Cloud product side, laid out the enterprise security architecture in detail: four principles of agent governance — visibility (see how many agents exist), traceability (audit every action to who did what), controllability (hard limits on network and command access), and trustworthiness (security screening for every external skill). Enterprises can block agents from downloading external skills entirely, restricting them to an internal-only marketplace of vetted capabilities.

And why Allen Zhang initially withheld WeChat access.

But then Lark and DingDing integrated agents — both enterprise-first platforms with consumer-adjacent surfaces, both aggressively marketing agent capabilities. The competitive pressure became untenable, and the FOMO turned real.

If WeChat, THE platform that is Chinese digital life, didn’t offer an agent layer, it risked looking like the one platform that couldn’t do AI. But ‘doing AI’ has its risks.

So Zhang relented. Qclaw connects to WeChat via a QR code. WorkBuddy allows remote task orchestration from WeChat. But notice what is still not open: WeChat groups. (see here for more on Tencent’s claw embrace) You can invoke an agent from WeChat, but you cannot yet deploy one into a group conversation. The privacy risks and liability potential remain too high, and Zhang’s instinct and protective nature of his brainbaby WeChat, which has historically been correct, is that the downside of a social-graph agent failure far outweighs the upside of being first.

This tension that is between competitive urgency and product discipline is perhaps the most honest expression of where Chinese tech actually stands with agents. Everyone feels the pressure to push out the next ‘it’ thing, but not everyone is confident they should, or at least not test it within its products that work. If it ain’t broke, don’t fix it. right?

The Harness Thesis

Even two weeks ago, I wrote about how the biggest new buzzword in the industry feels like it’s “harness.” Tencent Dowson Tong argued at the cloud summit that as mainstream model capabilities converge — and they are converging fast, with open-source models closing the gap with proprietary ones on most benchmarks — the differentiator is no longer the model. It’s the harness, echoing what we’ve been writing about. It is the engineering scaffolding of tool calling, context management, long-term memory, and workflow design that wraps around a foundation model to make it useful in production.

This is the token refinery framework — models produce crude oil, customers pay for gasoline, and the harness is the refinery. The harness is the refinery and value migrates downstream to whoever can transform raw intelligence into something usable, trusted, and repeatable. Money moves downstream; it always does.

Tong claims that Tencent is building the best refinery. The four pillars of their Agent Core are Context Engineering, Agent Orchestration, Agent Memory, and Agent Ecosystem, which can connect to CodeBuddy on the developer side and WorkBuddy on the office-worker side, with their ADP development platform providing lifecycle management. Inside Tencent, 2,000 employees have been using WorkBuddy since January across document editing, data cleaning, and business analysis. Tasks that took hours now take twenty to thirty minutes. Token consumption keeps climbing. Tong joked that soon, employees won’t just ask their bosses for raises, but they’ll negotiate for higher monthly token budgets.

The thesis is intellectually provocative, but it raises a question that neither summit attempted to answer: if every cloud vendor is building a harness and trying to build the best refinery. The harnesses converge on the same architecture, then what exactly differentiates any of them?

Same same but different

This brings me to my final point today. I felt like everything was getting to a point where it was all a bit same-same.

For two companies with such different DNA, business models, and business units, the AI monetization logic at both events felt eerily similar.

Give enterprises agent-building tools. Those tools produce agents. Agents consume tokens. Token consumption drives cloud usage. Cloud usage drives revenue. Layer on security governance and IM distribution to make the agent sticky, and the flywheel spins. Heck, Alibaba has a new business unit helmed by its CEO Eddie Wu called the Token Hub. Meanwhile, you see below on the big screen - those big words Tencent Token Hub.

Is the thinking almost too simplistic? Integrate claw = agent tools = more agents =more token consumption = more cloud revenue.

Li Qiang was unusually candid about the fragility of this logic. He compared tokens to fuel, and that highlights the risk. As switching costs for token buyers approach zero, users switch to whoever is cheaper. I don’t care where I get my gas. I check which franchise offers me the most coupons. And he said, plainly, that Tencent would not chase token volume as a top KPI - a smart message to staff, but how are you going to make money then?

The candor laid bare the tension at the center of both companies’ strategies. If tokens aren’t the margin and models aren’t the margin, the entire enterprise bet depends on the harness layer generating enough proprietary value to justify billions in R&D. Tencent disclosed RMB 18 billion in AI spending for 2025, with further growth in 2026. Alibaba’s capex trajectory tells a similar story. But neither has yet demonstrated that agent orchestration produces sustainably differentiated revenue. For now, they are building refineries on faith.

However, if you must, underneath the convergence, there are two real strategic differences worth separating from the noise.

Alibaba proudly claimed that it is now the number one cloud choice for AI usage in China, and as the company expected, public cloud demand has grown as AI applications continue to find new use cases and penetrate deeper into vertical cases. Their most distinctive thesis: embed AI inference directly within the database. Enterprise data already lives in databases — a lifecycle from cold archival to warm training sets to hot production. By running the model inside the database — zero latency, data sovereignty, data gravity — the database becomes the gravitational center of the enterprise AI stack. And gravity, once established, is extremely difficult to escape. Migrating requires moving not just data, but the trained agent behaviors, skill libraries, and memory states entangled with it.

Their new ApsaraDB Claw products give this thesis a commercial form. Tbh, part of it was too technical for me. But at a high level, DataClaw, PolarClaw, and ADBClaw are agents that live where the data lives, which covers DBA, development, analytics, and operations around the clock.

In comparison, Tencent is betting on workflow embeddedness. Where Alibaba locks you in through your data, Tencent intends to lock you in through your habits. It’s probably not an exaggeration to say that WeChat is the default platform for any social interaction in China, whether enterprise use or personal use; thus, the vision, as Wu Yunsheng put it, is that AI will become like “water and electricity” as utilities, functioning in the background and turned on with a faucet, not consciously counted.

[Translation for below: Tencent strategic pyramid: Energy —> Compute —> Models —> Agents —> Applications]

So where does value accrue now?

Schumpeter observed nearly a century ago that the economic surplus from a new technology rarely accrues to its inventors. It accrues to the people who figure out how to use it. The cloud vendors are playing the role of the inventor here, building general-purpose agent infrastructure at enormous expense. But the most interesting things I saw at either summit weren’t platforms. They were the customers.

Bilibili built an AI-powered video commercialization pipeline: Qwen-Audio and Qwen-VL analyze KOL product reviews, extract brand mentions, and match them to advertiser catalogs. Bilibili's service can be very valuable because of its content taxonomy and its relationships with advertisers, which no cloud vendor can replicate.

On the Tencent side, the case study that was shared was about the Chinese Marriott, Huazhu Hotels. The hotel chain manager deployed ‘华小二’, ‘Little Hua', an AI concierge covering 152 scenarios across 38 workflow modules. For example, when a guest says, “Send me some snacks.” Within five seconds, an agent will alert the manager to dispatch the nearest staff member robot to the room. It’s now already live in over 5,000 hotels, with 94% daily usage in pilots, and is scaling to all 12,000 properties. This is again because Huazhu knows how customers behave and how to manage hotel workflows.

The model is merely an intelligence input at this point, but the industry context is the moat. If Schumpeter is right again, the billions flowing into harness layers may end up submerging the margins of the infra providers.

The bigger theme I think we can recognize is that in the cloud era, raw infrastructure has been commoditized, and the most durable value has been captured by software companies that have turned it into systems of work. In mobile, the platforms mattered enormously, but so did the companies that transformed distribution into entirely new behaviors.

In the agent era, both Alibaba and Tencent are maybe realizing that $ will not be in infra, but in the markup lives in the harness layer — the orchestration, memory, governance, and distribution scaffolding around the model. They may be right. But if the harnesses all look the same, then wouldn’t the differentiator still be in the foundation models and capabilities?

Anyway, the bet is that the value is moving one layer further out, to the companies that weave those harnesses into specific industries, specific workflows, and specific institutional contexts, no platform vendor can replicate, which, incidentally, is why vibecoding won't replace industry-specific software — context isn't something you can prompt-engineer.

I left Shanghai thinking about something Li Qiang said almost in passing. He called AI “the greatest force of certainty in an uncertain era.” That’s a good line for a keynote. But certainty about what, exactly? Will those agents exist? Sure. That those tokens be consumed? Obviously. Will the cloud vendors building all this infrastructure be the ones who profit? That part is less certain than anyone on those stages wanted to admit.

The WeChat question might be the best lens for the whole thing. Everyone knows agents need to be everywhere. Not everyone knows if letting them loose is wise. And the companies that figure out where to draw that line — in product design, in guardrails, in knowing when not to ship — might matter more than the ones that built the fastest refinery.

The hard part is not shipping the wrapper and owning the channel where people will actually use it. Which brings us back to Tencent’s one and only WeChat touch point and how they’ve ‘won’ again.

Subscribe now

“Raising lobsters”

For the last two weeks, it felt like all of China was talking about 养龙虾 — “raising lobsters.” tphuang joked on X that OpenClaw’s marketing basically wrote itself because Chinese people already love (spicy) crayfish and the nickname did half the work. Reuters and CNA both captured the same surreal scene in Shenzhen: long lines outside Tencent’s setup event, with not just developers but also retirees and children showing up.

In my own circles, group chats and WeChat moments were full of install screenshots, side-hustle fantasies, and people asking whether this was the next internet gold rush. Lobsters were everywhere. My own mother called to ask if people were suddenly importing seafood as a sexy business, and that’s when you really know a technology story has escaped the tech bubble.

As Nicholas Gordon wrote in his latest article on Fortune, the craze has transformed China’s AI sector, pushing it to the mainstream. “If you’re in China, you can now pay someone to come help install OpenClaw on your device…and you can pay them to uninstall it too.” It was a serious hustle.

Now, despite headlines about the Wuxi government embracing OpenClaw and whatnot, I don’t think the original spread was top-down driven. With some official nods, it looked like a green light for enterprises to embrace it until federal regulators stepped in to warn of potential data risks.

But just as wrote in his awesome long post about China’s Open Source history, “The government played a minimal role for much of open source’s history in China.” This time, the initiative felt organic again.

Headlines focused on the quick adoption, warm embrace, and municipal government endorsements, but they missed why a company like Tencent would lead in this open-source technology adoption and why. Today, let’s examine how it took off and why Tencent was quick to embrace someone else’s open source technology, again.

How did it take off in China?

Well, a large reason is the timing. DeepSeek did more than prove China could produce a strong model cheaply. It changed the public’s mental model of where cutting-edge AI could come from. It pushed low-cost, open-source models to the center of the conversation and made it much easier for users and developers to accept that the best capability might come from outside the biggest incumbent stack. Then came months of demos and products, with tech firms themselves teaching the public that AI was not just supposed to answer questions anymore. It was supposed to search, plan, use tools, and do work. Moonshot Kimi-Researcher is a good example of that shift: not a chatbot, but an autonomous system built for multi-step planning, reasoning, and tool use.

The other part is distribution. China is an unusually fertile ground for this kind of product because so much digital life is already compressed into a handful of interfaces. People are used to living in chat, paying in chat, ordering in chat, and moving across lightweight in-app services without thinking twice. And superapps as an infrastructure continue to play a critical role in the diffusion of AI in the 2C space in China.

And I wrote that this was the obvious next step…

Back in October 2025, when OpenAI rolled out its Apps SDK, I wrote in OpenAI Is Becoming an Operating System that the closest product blueprint was not an app store or a browser extension. It was WeChat mini programs. At a meeting with executives from the company in December, I did ask whether WeChat would roll out a native agentic function, and a nod and a smile were all I got.

ChatGPT made the ‘chatbox’ the home screen, but only as an entry point. Lightweight services sit underneath. Payments, identity, search, notifications, and distribution are already there. In that framework, the prize was never just a smarter chatbot. The prize was owning the surface where a user says what they want, and the system can actually get it done. The more you think about it, the more it looks like the WeChat interface with its millions of MiniPrograms that users can call on.

That is why the recent reporting about Tencent building a native WeChat agent does not feel surprising to me, even calling it “top secret” by The Information seems a bit dramatic. Now, what WeChat is doing is playing its own game better than everyone else.

If the teaser and summary are directionally correct, Tencent wants an AI agent that lives natively within WeChat and can call mini programs directly. That is much bigger than adding another AI tab. It means trying to turn WeChat from a place where AI answers questions into a place where AI completes tasks. And because WeChat already has roughly 1.4 billion users and a dense built-in service layer, Tencent does not need to teach consumers a new habit. It can slide the agent into an interface that people already open all day.

Tencent has now made the same strategic move twice

Remember when Tencent’s DeepSeek integration was a big deal at the time, but in some ways it was also an admission? Hunyuan was not leading the market, so Tencent stopped pretending that model loyalty mattered more than user experience. It integrated DeepSeek into Yuanbao and then into WeChat. I wrote in AI Integration at the Speed of Light that this mattered because WeChat’s reach, search behavior, closed content graph, and mini-program layer made distribution far more important than model purity once strong outside intelligence became available.

It is now borrowing OpenClaw-style agent momentum to move from answers toward action. The bet underlying both moves is still the same: if the user already lives inside WeChat and WeChat already contains the payment rail, the service layer, and the mini-program ecosystem, then Tencent does not need to win the model beauty contest. It needs to own the place where AI turns intent into action.

OpenClaw feels like the same instinct, caution and action again, but this time applied to agents.

I think we can intepret it has an out of neccessity move again. Tencent probably did not arrive here from a position of strength. It had spent heavily on AI infrastructure, talent but still was not setting the pace in consumer AI, largely lagging behind ByteDance and Alibaba’s efforts in LLM development as well as user courtship. This time, OpenClaw gave Tencent something brute-force capex could not: a way to change the terms of competition without first winning the model race.

And leveraging on someone else’s technology, Tencent was able to come out on top again.

What Tencent actually launched

OpenClaw is powerful because it gives large models hands and feet — the ability to become agentic. It can take over a computer’s keyboard and mouse and actually do things in the background. But its weakness is just as obvious: deployment and reach is painful.

For ordinary office workers, setting up environments and dealing with command lines is enough to kill adoption before the fun even starts (a true pain point even for myself). And Tencent saw that immediately, so it did not roll out one giant catch-all product. It split the market.

• Lighthouse is the developer on-ramp. Tencent Cloud positioned it as a one-click OpenClaw deployment in the cloud, with stable online operation and control through QQ and WeCom. This is the easiest layer to understand economically. Tencent is selling the picks and shovels.

• QClaw is the consumer-facing wrapper. It takes a messy open-source agent and turns it into something a normal person can install locally. Tencent’s own write-up describes it as a local AI assistant built on OpenClaw that can be controlled through WeChat conversation and used to remotely run tasks on your home computer. That is a very Tencent form of productization. It does not ask the user to learn a new interface. It turns WeChat into the remote control.

• WorkBuddy is the office-facing version. It is a desktop AI agent for workplace tasks, locally installable, compatible with OpenClaw skills, and capable of automating tasks such as information retrieval, report generation, and content drafting. The more important point is that Tencent is not just wrapping OpenClaw and calling it a day. It is using OpenClaw to speed up adoption while building a more Tencent-native workflow layer around it.

• Another subtle move is that Tencent pushed agent capabilities into its enterprise-facing WeCom workflows. Tencent is wiring the agent into the communication and workflow surfaces people already use. It is trying to make the agent truley mold into habit.

WeChat is still the real asset

The heart of the story for me is still an ecosystem story.

The real value here is not that Tencent wrapped a hot open-source project like everybody else. The real value is that WeChat already has what everyone else is still trying to assemble: trust, traffic, payments, search, merchants, mini programs (activity loops), and habit. And on top of that, another reminder that it has over 1.4 billion combined MAU, and products like Mini Programs, Weixin Search, and Weixin Pay all living inside the same surface - connected.

Lovable’s Head of Growth Elena Verna’s framing in her recent 20VC interview stated that software is increasingly competing on trust, and not just features. In her own blog, she pushes it one step further: once basic utility is there, products start competing on connection. That is exactly WeChat’s advantage over all the other clawbot-like products that got churned out this week. WeChat already has that connection with users. It has mindshare. It has functional adjacency. People already chat, search, pay, order, and move across services there.

And I think Tencent’s caution has been part of its success strategy too. WeChat has always been slow to push out major features until the product feels ready enough for a very broad user base. That can make Tencent look late, but it also means WeChat does everything it can to protect its trust from users. That matters much more in an agent world, where the product is not just answering a question but potentially touching payments, bookings, messages, documents, and other sensitive flows. This last point is an inference based on WeChat’s product history and the services already embedded in the app, rather than a direct quote from Tencent.

So now, if Tencent can connect a reliable enough agent to mini programs, then it has a direct path from ask to done that most AI apps still do not. A user would not just ask for restaurant suggestions or travel tips. The system could, in theory, book the ride, order the groceries, reserve the restaurant, route the payment, and handle the follow-up inside an environment where those services already exist. That is a very different proposition from just being another strong chatbot. Tencent does not need to convince users to go somewhere new. The whole advantage is that the user is already there. The WeChat layer in many ways become the real (deliberate or not) strategy.

Not all agentic activities are the same

I know I’m going on a little bit of a tangent but Alibaba’s push with Qwen obviously irked Tencent too and probably nudged them to go this direction quicker. However not all commerce is equally agent-friendly and not all agentic activity is the same.

Utility commerce is easy: groceries, socks, repeat purchases, bill pay, and simple travel changes. Discretionary spending is harder. Jewelry, fashion, gifts, and beauty are not just search problems; they are expressions of taste. Agents may win convenience first, long before they replace choice.

To me, the point is simple: Tencent is not trying to win AI by owning the best model or the best open-source agent. It is betting that WeChat’s distribution and service graph are still the scarcest asset in China's consumer AI.

And I think this first wave of craze may taper a bit in a few weeks. However, the ones that do stick around will evolve their habits, and we’ll see durable everyday behavior change.

That is another reason Tencent’s position made me realize it may have a stronger chance to win over user-habit. WeChat does not need to own all commerce for this strategy to work. It already has direct paths to MiniPrograms. It only needs to own the high-frequency, low-consideration flows where agents are most useful first: local services, repeat purchases, booking tasks, payments, and routine transactions. That wedge alone is already enormous.

Some of the demand is obviously real. OpenClaw can help with reports, coding, spreadsheets, slide decks, and repetitive desktop tasks. But how much of that is actually needed by the uncles and aunties who lined up to get it installed on their devices? Probably not much. There is also a lot of curiosity, status signaling, and plain FOMO mixed in. In the mentioned CNA report, it found that some users mostly installed it to understand the hype and did not even have a specific task in mind.

The hardest question is not adoption. It is trust.

How do we explain the others jumping on the bandwagon? Well, FOMO. Once OpenClaw became culturally hot, everyone wanted a consumer-friendly wrapper hence Alibaba and Baidu’s launches of their own versions of OpenClaw.

But that is exactly why Tencent’s position is different. Alibaba has commerce rails. Baidu has search and cloud. Tencent really, has WeChat. If the game shifts from “best standalone AI app” to “best place for AI to actually get things done,” Tencent is playing a different game from the others.

The pushback I take most seriously is this: if Tencent keeps outsourcing the hardest layer — first to DeepSeek at the model level and now to OpenClaw-style products that still rely on outside model quality for execution, what exactly is the moat? Why would users stay with Tencent instead of going directly to the best model provider?

My current thinking is that users will stay only if Tencent owns the experience of completion. Most 2C users do not care whose model sits underneath. They care whether the thing works, whether it feels safe, and whether it finishes the task inside apps they already trust.

If Tencent can make the agent feel native to WeChat, connect it to mini programs, keep permissions sane, and smooth over the rough edges of outside models, then yes, users may absolutely stay. But if Tencent ends up as little more than a glossy traffic distributor for other people’s intelligence, then it risks owning reach without owning the actual relationship.

And then there is a second trust problem, which may matter even more. There is a big difference between an AI that gives a bad answer and an AI that sends a message, moves money, books the wrong ticket, or mishandles sensitive data. The more these agents touch real services and real data, the more the permissions layer becomes the product. So in this contracdiction, Tencent has the most to gain and the most to lose. Ultimately, it cannot afford to break trust inside WeChat.

The monetization story is already visible

If you look at Tencent’s potential monetization of this all.

The first layer is straightforward. Lighthouse drives cloud usage. WorkBuddy can become paid workplace software. WeCom integrations create room for governance, admin controls, security wrappers, and workflow automation. Tencent is already monetizing the infrastructure and workflow layer around agents.

The bigger monetization story sits on top of WeChat’s service graph. In my mini-program piece, I argued that Tencent historically monetized the ecosystem less like a classic app store tax and more through ads, payments, and deeper integration across WeChat surfaces. If Tencent puts an agent on top of that service graph, the leverage increases: discovery, merchant routing, transaction flow, developer tools, and eventually the default place where commercial intent is expressed within WeChat. That is a much bigger business than charging a monthly fee for an assistant.

Once LLMs are commodities and their cost collapses, monetization has to focus on the end result. People will follow the brand they trust because the functionality may be commoditized. And who has better brand value/ recognition than the penguin company?

If this thinking holds, standalone AI apps will get squeezed first. If users can ask for something in WeChat and get it done there, the cost of switching to a separate AI destination increases. Mini-program developers may also lose some direct relationship with the end user. Today, users still know which app or mini program they are calling but in an agent world, the agent becomes the front door, and the merchant becomes the backend service. That changes discovery, brand loyalty, and pricing power but it also gives Tencent more leverage over routing, ranking, and eventually monetization.

Looking at 0700 HK’s performance, it looks like investors may see the same story as I do. Or who knows, maybe they’re just punting some OpenClaw /China AI exposure.

FYI, And just a day before I pubbed, as expected, Tencent Cloud became an offical sponsor to the OpenClaw Community.

Subscribe now

Alibaba is spending RMB 3 billion this Spring Festival. Tencent is spending RMB 1 billion. Baidu, half a billion. What looks like the annual red-envelope ritual where big tech burns cash, attempts to acquire new users, a bit of a feel-good marketing campaign, and flex, this year is actually a bit different.

Honestly, I almost scrolled past it. But the more I looked at the big numbers and headlines, the more I realized it is something else. This is a fight over where daily consumption starts and which AI interface wins.

Alibaba is trying to make “ask AI” a default behavior for ordinary transactions: ordering takeout, booking a flight, buying a movie ticket. Tencent is taking a different approach. It is reaching for a playbook that already worked once, which is to use a social mechanic to force a new habit at a national scale, the way WeChat Pay used red envelopes to get people to link bank cards back in 2015.

This is why the headlines matter, but not in the way people think. It is not just RMB 3 billion vs RMB 1 billion. It is what that money is paying for: habit training and habit changing.

Why now? And what now?

Spring Festival/ Lunar New Year/ Chinese New Year is China’s largest national holiday. Everyone, I mean everyone, gets time off (well, I guess except the Chinese AI labs). In this case, it is a national-scale product lab.

Demand spikes across essentially all things consumer — travel, food, entertainment, and gifting. Everyone is coordinating with family and friends. 串门, visiting neighbors and relatives, is also part of the ritual. Communications and touchpoints are at the highest level. People are moving fast and making decisions under time pressure. That’s when defaults can shift, because convenience suddenly matters more than “best choice.” And that is why these companies are trying to tap into this opportunity to introduce their AI apps.

We have seen this with prior primitives. GPS did not just create apps. It made ride-hailing and delivery feel obvious, and Meituan and Didi became infrastructure. Cameras plus bandwidth did not just create a format. It changed discovery and attention, and Douyin and Kuaishou became the front door for entertainment and then, later on, commerce. And let us not forget Hallmark created allllll those ‘holidays’ for consumer consumption.

The same logic applies here, but it’s more direct. The current consumption path via internet apps is user-driven: open the app, search, compare, place an order, pay, and wait.

The AI-native version compresses that chain or even completely removes it. You state the intent once, and the system executes the rest, including ordering, payment, dispatch, and after-sales. Less switching. Less browsing. Maybe less comparing. You can also say less decision-making, which is a bit terrifying.

If 2015 changed how money gets paid in China, this round is trying to change where commerce begins. From opening an app to asking an AI. From user-operated to system-executed. Once the entry point shifts, second-order effects follow.

Advertising will look different if discovery becomes more conversational and less browse-driven. Checkout looks different when it occurs within an agent flow. Payment becomes more like plumbing if it is executed in the background. And the “app install base + user habit” moat becomes less permanent if users shift from thinking in apps to thinking in outcomes. It’s not clear who has the better strategy, but the competition is once again on between the forever rivalries.

Alibaba’s bet: convenience and the bundle becomes the moat

Alibaba’s RMB 3 billion feels less like a traffic play and more like an ecosystem integration play.

The red packets are the hook to get you into the ecosystem. The real product is training users to complete transactions through Qwen across Alibaba’s own services: Ele.me for food delivery, Taopiaopiao for movie tickets, Fliggy for travel, plus the rest of the commerce stack.

If you want an agentic loop to work in real life, it cannot stop at “recommendation.” It has to complete the loop. The easiest way to do that is to own the pipes.

This is also why Alibaba’s approach is structurally different from what OpenAI or Google are trying to do with AI shopping. In comparison, what we’re seeing so far in the West is that cross-platform chains are long and brittle. You can generate a list, compare specs, surface links, and then the user still has to click around, log in, confirm, and pay. The handoff moments are everywhere, and each one is a friction point and also an interception point.

Alibaba has fewer handoff moments because it owns more of the loop. If the AI recommends a restaurant and you say “book it,” the booking and payment can happen inside the same flow.

Tencent Pony Ma’s public comment captured the tension well. He praised Alibaba for combining its internal ecosystem and working hard on integration, and then pointed out the obvious risk: users may not like a full bundle, and not every internal service is best in class.

That is exactly the best. Alibaba is betting that “good enough” across ten services wins if the experience is smooth and end-to-end. The bundle becomes the moat because it reduces switching inside the flow. In an agent world, the cleanest loop often beats the fanciest individual feature.

If it works, Alibaba does not just own transactions. It owns the moment before the transaction, the intent layer, and then everything downstream.

Tencent’s dilemma: protect WeChat, but do not miss the habit shift

Tencent’s position is more complex, and it has always played more cautiously. It's always been famous for winning at "second-mover advantage," often executed through a "fast follower" strategy - see payment (WeChat Pay vs. Alipay), short videos (Douyin vs. WeChat Channels).

WeChat is a consumer infrastructure in China. A real AI failure inside WeChat is not “a bad feature.” It is reputational damage at scale: hallucinations causing real-world mistakes, privacy concerns, and scams amplified by automation. Trust breaks slowly and rebuilds even slower.

So Tencent’s caution makes sense.

But caution has costs. Tencent knows that, too. Pony Ma has acknowledged that Tencent moved slowly and has reiterated that AI is the only area truly worth continued investment at this stage. That is honest. It is also a signal that Tencent is balancing two imperatives that conflict: move fast enough to shape new habits, but not so fast that it compromises the core product that generates cash and trust.

When I tested Yuanbao Pai this week, the product idea is interesting, but it also highlights the key problem: any extra step that pulls users out of their default WeChat rhythm is not trivial.

You have to quite literally send an invite to a friend/ group chat to ask them to 1/ download Yuanbao, 2/ enter the invite code, 3/ join the specific chat to interact. In a world where the fight is about defaults, distribution is not just “how many users you can reach.” It is how few steps it takes to become muscle memory.

In general, I don’t find it super helpful yet. Like I had to @元宝 the chatbot, and then ask for reminders, summarize chat, or offer suggestions. However, I could see it being more useful through the WeChat Enterprise interface.

But the question for Tencent now is not whether it can build an AI chatbot, it is whether it can turn AI into a daily-life execution layer without jeopardizing WeChat’s core trust.

Meanwhile, ByteDance is not waiting. Doubao’s scale and Spring Festival visibility are a reminder that consumer AI is now a standalone battlefield, and it does not fully map to the old “social vs commerce” lines. It secured itself a position as the official Spring Festival Gala partnership. The gala is broadcast across the world to over 700 million viewers. Where everything from baijiu to EV makers to home appliance brands to robots all try to snatch a highly sought-after spot as a sponsor/ advertiser. Think F1 sponsorship level exposures.

It’s all hands on deck to capture the 老百姓 (literal translation: hundred last names), the average Joe, the normal civilian, the everyday person’s heart and mindshare now.

Business plus AI, or AI plus business?

The cleanest framing I have is this.

Alibaba is treating AI as the front door to its existing commerce ecosystem. The business came first. AI is the interface that compresses the loop.

Tencent is treating AI as a capability that must earn its way onto the most sensitive consumer-facing surface in China. The product came first. AI is the enhancement, rolled out in a way constrained by trust and safety.

It’s actually the same vocabulary, but with a different ordering. Biz + AI? AI + Biz? The ordering implies speed, risk tolerance, and who wins if this transition is real.

In years to come, when we look back on the inflection point when the Chinese big-tech AI landscape truly changed, we may trace it back to this Spring Festival. Happy year of the horse, everybody~

Last year’s Spring Festival put DeepSeek V3 on the map. This year might be about something less flashy but more impactful: who owns where commerce and activity start.

1) Personal reflections

A few days ago, I was at a coffee chat with about ten women in AI. Some were building consumer-facing startups, some worked at mega-cap infra companies, some were ex–Mag 7, some were investors, and so on. The conversation was casual, but it drifted fast from how AI might reshape education to how we preserve our humanity: finding time to shut off our phones, make art, read a book, and finish a puzzle.

Where we kept getting stuck was the same question everyone keeps circling: is human intelligence really replaceable? We ended up debating whether a model is a product or infrastructure. We also debated whether the Manus deal is a good example and path forward for Chinese AI companies trying to go global, or just a situation. Many of us were young mothers, anxious about what AI may mean for education and our children’s future.

One person shared a joke she’d heard from an AI researcher in Silicon Valley. If AGI is achieved, what will humans do? He replied calmly: “We should exercise and strengthen our bodies.” The irony. But jokes aside, there’s something true in it. When new technology develops, some jobs get replaced, but new ones also emerge that we couldn’t have imagined. When cars were invented, horse carriage riders probably lost their jobs, but drivers as a profession emerged. As self-driving cars go mainstream, drivers may be displaced again, and we don’t yet know what replaces that work. Ten years ago, there was no gig economy nor food delivery riders. I guess my point is: there needs to be a cautious embrace, but also optimistic hope that new technology does not just wipe out all of our purpose. For a temporary period of time, some may be displaced—but then balance within society will resume.

To be honest, I think of it more as we are teaching our children how to eventually use these tools, just as my parents taught me how to type or search the web. I’m in the camp that traditional credential signaling (fancy degrees) may mean less, but agency cannot be replaced. How does pattern recognition help us understand the intricate relationship between people, their emotions, and the way people make decisions irrationally? How do machines ever piece together sensory things like lighting in a room, the mood of a crowd, smell, and color that can evoke emotions?

In a recent AI summit held in Beijing, Alibaba’s Lin Junyang said, “Personally, I think technological progress is linear, but human perception is exponential.” That’s exactly it, isn’t it? Even as technology removes “friction” like cars, planes, the internet, AI in whatever form, and even our fear, our greed, our lust, our love… in their rawest form, these haven’t really evolved or changed for centuries. It’s just now we cry over texts instead of pigeon-couriered handwritten letters, maybe.

So no matter how AI develops, I believe the truest form of intelligence is within our ability to connect the dots and make sense of the world of noise, sensory input, and emotions. It makes me believe more and more that no matter what my children do one day, I will encourage them to read literature, read non-fiction, and study liberal arts in school. Vocational skills can be picked up in jobs. Execution, more and more, can be done by agents. But the ability to think critically will not be replaced, and if anything, will matter more than ever.

That’s my little opening rant.

And then, like these conversations always do, it snapped from philosophy back to economics. Because underneath the “will AI replace us” anxiety is a much more practical question: what exactly is AI trying to solve, and what’s actually holding it back?

I want to double-click on Alibaba’s Qwen AI Assistant App launch today and then share some notes on my thinking from rereading the AGI Next Summit transcript. The event featured some of the leading voices in China's AI industry: Alibaba’s Lin Junyang, Tencent’s Yao Shunyu, and Zhipu/Z.ai’s Tang Jie on the bottlenecks and trends in China’s AI diffusion.

2) Where the deployment bottleneck is

The model-versus-product debate sounds abstract, but it’s becoming a heated one for a reason. On one side, you have labs committed to research and not caring much about user numbers; on the other, you have “wrappers” betting on usage and distribution, which also means betting their fate on someone else’s roadmap. Then you have the in-betweeners, and Manus is the obvious example, where the claim is that the raw material (the model) matters, but how you “cook” it is what separates a meh restaurant from a Michelin-star experience.

If the model is the product, the logic is straightforward: benchmarks, rankings, “best model wins,” and everyone else becomes a substitute. If the model is infrastructure, then “best model” is only the beginning. The value migrates to distribution, integration, and the interfaces people already live inside, because that’s where behavior is and where humans actually interact. That’s where switching costs become emotional, not technical.

Yao Shunyu, Tencent’s latest high-profile hire from OpenAI, was unusually direct about what he thinks the real bottleneck is. Not parameters. Not the model alone. Education first:

“It’s not about humans being replaced by AI. It’s about people who know how to use tools replacing those who don’t. Instead of obsessing over model parameters, it’s more meaningful, at this stage in China, to teach people how to use Claude, Kimi, Zhipu, and other tools effectively.”

Then deployment:

“From my experience in To-B companies, even if the model doesn’t get any smarter, just deploying it better in real-world environments can bring 10x or even 100x returns. Right now, AI’s impact on GDP is still under 1%, so the room for growth is huge.”

If you take those seriously, the “model is the product” story starts to wobble. Because the hard part is no longer only building intelligence. It’s getting intelligence to show up where it matters, inside workflows, inside organizations, inside habit.

And yes, there’s a cynical read here too: it’s also a neat way to justify moving from OpenAI to Tencent. If you believe deployment is the real bottleneck, then Tencent, of all companies, should be able to win simply by embedding AI into the default interface of daily life.

3) What AI is even trying to solve

The AGI discussion largely revolves around benchmarks and performance, but in conversations with practitioners in real life, it’s often a more practical question: can the system help you solve problems that don’t come with a clear recipe?

This is where the Manus thinking fits naturally. Lin Junyang framed it by borrowing Peak’s take (Manus co-founder), and it’s one of the clearest definitions I’ve heard of what “general agents” are actually trying to do:

“Just because I work on foundation models doesn’t mean I get to be a startup mentor for this. I’ll borrow something Peak (Manus co-founder) said: The most interesting part of building general agents is solving long-tail problems. That’s also the most compelling aspect of AGI, helping people solve problems that have no clear solution elsewhere. Popular problems (like top product recommendations) are easy. True AGI means solving those edge-case issues that no one else has figured out.

So, if you’re a great product wrapper that has solved an issue better than the model company itself, then you’ve got a shot. But if you lack that confidence, model companies will likely win out. When they hit a wall, they can just “burn GPUs” and train deeper models. That’s a low-level advantage, and it’s hard to compete with. So really, it depends on your approach.”

I like this because it drags AGI out of sci-fi and back into product reality. If “long-tail problems” are the point, then the bottleneck isn’t only technical. It becomes taste.

Which long-tail problems matter? Which are valuable? Which are solvable? Which are too regulated, too messy, too expensive? Which benchmarks do you chase, and which do you ignore? Those aren’t neutral choices. They’re incentives. They’re business models. They’re judgment calls.

Zhipu’s Chairman Tang Jie just experienced its IPO hype, and during the public engagement, he seemed also very alert to commercialization, describing what separates survivable agent companies from disposable ones, in a way that’s both harsh and useful:

“There are three key factors that will shape where agents go:

• Value: how valuable is the problem the agent is solving? Early GPTs failed because they were too simplistic, just prompt wrappers. Agents must tackle real, complex problems to be useful.

• Cost and boundaries: here’s the paradox. If a problem can be solved with a quick API tweak, the foundation model will absorb it. So applications need to find their niche before the base model catches up.

• Speed: This is all about timing. If we can go deeper into a use case and polish the experience ahead of the curve, even by six months, we can survive. Speed is the name of the game. If our code is solid, we might just outpace others in coding or agent development.”

That’s not “wrappers are dead.” It’s also not “wrappers are the future.” If the foundation model lab pushes out a similar feature and can swallow your feature overnight, you have no business. That’s why the startups that have proprietary data in niche use cases or have doubled down on a vertical can have a proper shot, and your moat will be your expertise. So for startups, they’ll need to either solve something valuable and hard enough or niche enough to resist absorption, or you outrun the absorption clock.

4) China’s current AI landscape at a glance

That brings me to where the Chinese AI landscape looks right now, at a high level. I’d still break it down into big tech versus startups, and within big tech, companies look similar at a distance, but in detail, they’re very different.

Big tech in China is mostly doing some version of “models plus ecosystem,” but their strengths and paths to monetization diverge.

Alibaba started with an enterprise-first mentality. Whether it was open-sourcing Qwen or selling API services, all roads led back to the cloud. They’ve been China’s cloud anchor for a long time, and the model strategy has always been tied to that.

More recently, there’s been a conscious move to make the consumer surface look like the model itself, renaming the app to Qwen, signaling: here’s China’s best open-source model, directly accessible. But the bigger play isn’t the app in isolation. It’s what the model can call inside Alibaba’s ecosystem: Fliggy (travel), Gaode (navigation and ride-hailing), Ele.me (food delivery), and, of course, Taobao. If OpenAI is trying to invent “mini programs” and shopping assistant features from scratch, Alibaba already has the services, the data, and the payment rails inside one world.

Tencent’s path has been more consumer-native and more distribution-led. They were early to integrate DeepSeek into WeChat—pretty bold in a market where many incumbents were still guarding their platforms and sticking to single-model pride. But the thinking was simple: Yuanbao, frankly, lagged behind on model performance, and distribution is Tencent’s unfair advantage.

That move mattered because it showed the power of the ecosystem and functional adjacency. Opening WeChat and giving people a chat interface to interact with AI is natural; MAU rises quickly when you don’t ask users to change habits. But the deeper question is what happens after the first wave of curiosity. Distribution can give you the first hundred million. It can’t automatically give you retention, usefulness, and monetization.

ByteDance is the third pillar. Doubao (both app and model family) has the most sheer consumer volume in China right now. ByteDance has also gone all in on multimodality. If this is a consumer crown race, they’re not leaving it to anyone else.

Then you have the startups: the “four tigers”: Zhipu, MiniMax, Moonshot, Baichuan. Each has had to make a strategic choice under tighter capital conditions and tougher competition from big tech. Baichuan has leaned into healthcare. Moonshot is still associated with frontier research and Kimi. Zhipu is pushing model-as-a-service and a developer/coding angle. MiniMax has been strong on multimodality. It’s crowded, and the reality is that startups are tighter on resources while big tech has distribution, compute access, and more room to absorb mistakes.

If anything, that pushes startups toward verticals. Not because vertical is trendy, but because horizontal consumer distribution is a bloodsport and the incumbents own the on-ramps.

5) Agentic and the ecosystems of Alibaba and Tencent

Much of the AI discourse remains obsessed with replacement. But the more honest lens is: what gets taken off our plates first, and where does AI show up without asking us to relearn life? For much of the last year, we’ve been writing about the super app ecosystem that exists in China, which is very different from the US. This isn't just as simple as a walled-garden moat or something; this is integrating all the data, the user, the experience within one interface, and truly creating an AI operating system organically as the company pushes this direction top-down.

Just today, Alibaba’s VP Wujia, who oversees the Qwen chatbot app and Quark AI (the enterprise product) announced the Qwen assistant app. The company says it believes that everyone will have an AI assistant in their pocket one day. Qwen 任务助理 Assistant 1.0 will be free to all to use - think of an agentic assistance in your palms.

It is said that it can support almost every task within the digital infrastructure. You can go straight to Fliggy (Bookings.com) and Amap (Google Maps + Uber) to plan your itinerary and book your tickets for travel. Book entertainment tickets through Damai大麦 (ticketmaster), pay through Alipay, and order goods via Taobao. So consumers can actually access everything from the Qwen app. This is exactly what I predicted. Everything can be completed within its ecosystem through its chatbot interface. This is what OpenAI is trying to do, but it has to onboard all kinds of external partners, essentially, where Alibaba could connect it all in the backend.

The key is all that “context” that Qwen has, which is the part that quietly decides whether consumer AI becomes a default interface or just a feature. However, the “context,” which is the data, will also be the hardest to incorporate and translate across different apps for Qwen to dissect.

Right now, most people don’t use AI as they would a new operating system. They use it like a turbocharged search box. That matters because it means the leap isn’t “bigger model.” The leap is what happens when the model becomes relevant in context, not just competent in abstraction.

That’s why Tencent Yao’s deployment line is so sobering: even if the model doesn’t get smarter, better deployment can bring “10x or even 100x returns.” That’s a diffusion argument. At this point, capability is not the bottleneck. Integration is.

China has a distinctive setup here: the distribution layer. If your “agent” can sit inside a superapp and call existing services, payments, bookings, commerce, and customer service, then the agent doesn’t need to invent the ecosystem. It just needs to orchestrate it. That’s why the agent story in China is so tied to WeChat mini programs, and why Alibaba’s ecosystem is interesting, because it can call Taobao, Ele.me, Gaode, Fliggy, all inside one world.

Yao touched something sensitive here, but it’s central to the consumer story:

“If the model knows it’s cold out, which neighborhood I’m in, what my spouse likes to eat, or what we talked about last week, well, that changes everything. Imagine forwarding your WeChat chat history to a bot like Yuanbao. Suddenly, the model has much richer context to work with, and the value to the user increases significantly.”

The nuance matters: forwarding implies intent and consent. That’s the difference between creepy and useful. It’s also why existing platforms: search bars, chat interfaces, super apps, are such powerful diffusion engines. They already have behavior and context surfaces. They just need permissioned pathways to make that context actionable.

Yao summarized the thesis cleanly: “In short, leveraging extra context, paired with a strong model, is key to unlocking better consumer AI experience.” And he explained why Tencent believes it has a structural edge beyond just “having a model”: “Startups often rely on externally labeled data or APIs. Our edge is that we already have rich, real-world use cases.” He also pointed to internal data loops as a moat: “We can tap into the actual development and debugging processes of our thousands of engineers—that’s way more valuable than synthetic data.” Then he closed the loop: “For To-C, it’s about going deep on context. For To-B, it’s about leveraging internal use cases and real-world data.”

And then, beneath all of it, there’s a deeper bottleneck that doesn’t show up on leaderboards.

Yao said it most starkly: “The real bottleneck is imagination. We understand what reinforcement learning can do; we’ve seen things like O1 solving math problems. But with self-learning, we haven’t even defined what the ‘target task’ looks like.”

Tang Jie from Z.ai describes the same wall through economics. When you pour in resources and stop seeing gains, you hit “efficiency bottlenecks.” “What we need is a new paradigm to measure return: Intelligence Efficiency.” And his conclusion is a shot across the bow: “That’s why I truly believe a paradigm shift is coming in 2026. And we’re working hard to make sure we’re the ones leading it.”

Lin adds another under-discussed constraint: AI is sliding into daily life, but “we lack a unified metric to judge how well it’s working.” If you can’t measure it, you can’t govern it. If you can’t govern it, you can’t deploy it widely.

So if the last two years were about “how smart can the model get,” the next two may feel more like “where does intelligence actually become real?” In interfaces. In deployment. In education. In a permissioned context. In power constraints. Whether we can define the right tasks and measure whether they’re being solved.

In the U.S., we know there are stricter regulations around opening up functionality and anti-trust concerns, so there is no super-app. But what OpenAI is trying to build is resembling more and more like a superapp, and what Chinese big tech has, even though with somewhat less capable models right now, is a crazy ability to call on all of its apps and functionalities across its ecosystem in one interface.

The companies that can deploy well will outpace the ones that can’t. And the most enduring human edge may not be raw intelligence; it may be taste, judgment, and the ability to decide what problems are worth solving in the first place.

FYI: A ton of other big tech coverage, and details of how each BAT is running their AI strategy: AI Proem Big Tech

And guess what else is back? DeepSeek with its breakthrough announcement. The original V3 came out around Christmas last year, and now the highly anticipated V3-2 is out. explains its implications here.

Geopolitechs

What DeepSeek V3.2 Really Tells Us

On December 1, 2025, exactly three years after the launch of ChatGPT, DeepSeek dropped a major surprise at this symbolic moment as the V3.2 model officially went live. The technical report is available…

Read more

5 months ago · 8 likes · 1 comment · Geopolitechs

ByteDance’s Second Act in AI

ByteDance’s Doubao surpassed DeepSeek’s MAU numbers in September, becoming China’s most popular native AI app, according to QuestMobile. Who’s behind this head-to-head competition? It’s Zhang Yiming.

When he announced his “retirement” in 2021, not even 40 years old, a lot of people assumed he’d moved to Singapore for a quieter, more relaxed life. But according to ByteDance insiders, he’s the orchestra conductor behind all of the company’s AI initiatives – and in the last two years, “relaxed” is probably the last word you’d use to describe his relationship with AI.

QuestMobile’s ranking shows Doubao now in first place with 172 million monthly active users, with DeepSeek second at 145 million. Combined, that’s around 320 million people, roughly a third of China’s internet users. The speed and scale of consumer AI adoption in China still feel a bit unreal, even when you stare at the data long enough.

Share

Did Zhang Yiming ever really leave?

We wrote before about ByteDance’s internal talent program Seed, which hires directly from top PhD programs. That talent-first mentality is now showing up in the leadership bench, too.

Insiders say Zhang has been personally overseeing AI hiring – poaching from Google, Alibaba, and up-and-coming labs, a few notable ones to mention:

• Wu Yonghui, former DeepMind VP, to lead Seed’s foundation model work.

• Zhou Chang, former head of Alibaba’s Tongyi Qianwen (Qwen).

• Huang Wenhao, previously pre-training lead at Kai-Fu Lee’s 01.ai.

And it’s not just “hire a few big names and call it a day.” One Chinese-language report quoted an industry insider saying Zhang has been reading OpenAI papers late into the night to the point of forgetting to eat and sleep. This does not sound like a founder who walked away. And thus, it is clear that he is pulling the strings behind the scenes.

Some have questioned why ByteDance has seemed rather quiet while Tencent and Alibaba have been pushing out apps and announcing big capex numbers over the last year. Well, according to Chinese media who have spoken to leadership internally, they say his vision was always clear – he was just waiting for three things to line up: the tech, the ecosystem, and the talent. And that inflection really started in 2023.

Spending like a fourth hyperscaler

Because ByteDance is private, its AI capex has been a bit of a black box. One sell-side estimate from Zhejiang Shang Securities in 2024 put ByteDance’s AI spending at around 80 billion yuan (over US$11 billion) already – roughly in line with the combined AI capex of Baidu, Alibaba, and Tencent at the time. The same report forecast ByteDance’s capital expenditure could reach 160 billion yuan in 2025, with about 90 billion of that earmarked for AI compute alone.

If those numbers even roughly hold, ByteDance is effectively behaving like China’s fourth hyperscaler – but with a P&L that still leans heavily on ads and short video.

The bet is that this infrastructure won’t just power a single model or app, but an entire stack:

• consumer AI (Doubao, Douyin tools, TikTok experiments),

• Volcano Engine as a Model-as-a-Service (MaaS) platform,

• and a growing enterprise (API sales) footprint.

This is where the story shifts from “Doubao the app” to “Volcano Engine the platform.”

Volcano Engine: the MaaS challenger

In the old Infrastructure as a Service (IaaS) era, Alibaba Cloud was miles ahead, Tencent Cloud staked out gaming and video, and Volcano Engine was the new kid trailing behind. The Volcano engine is ByteDance’s cloud computing and AI platform that provides its underlying technologies to external enterprise clients. It offers a range of services, including AI-driven tools like large language models (Doubao), video generation, and recommendation engines, as well as cloud infrastructure services.

Now, in the Mobility as a Service (MaaS) era, the game is different. The competition is increasingly about three things: model quality, latency, and cost per token.

And on this axis, Volcano has moved much faster than people expected. A Chinese long-form piece on the AI cloud race walks through the shift pretty clearly:

• In 2024–25, cloud spending started to migrate from training to inference. A lot of weaker model teams fell away; the center of gravity moved to applications and MaaS.

• Gartner’s 2025 Magic Quadrant for AI Application Development Platforms put Volcano Engine in the “Challenger” quadrant, with its “Ability to Execute” ranked above Alibaba Cloud and Tencent Cloud in China, and its multimodal capabilities scoring fourth globally.

• At the same time, Alibaba Cloud was the only APAC cloud vendor to land in the “Leader” quadrant of Gartner’s GenAI technology innovation guide – a nice reminder that infra and application platforms are being judged slightly differently.

Although ByteDane does not have Alibaba’s cloud infrastructure lead. Volcano has been quietly making very deliberate moves over the last year and a half, and making strides. In May 2025, it essentially started a price war in China’s LLM space. Pushing LLM prices into what local media call the “厘时代” – essentially driving per-1k-token costs down to tiny fractions and forcing everyone else to respond.

Then, in June 2025, it revealed what is called “Input length interval pricing,” so what it does is that instead of a flat rate, longer contexts get priced more intelligently, which directly matters for code, long documents, and video workflows. By October 2025, Volcano launched a Model Router that sits on top of Doubao and a bunch of open-weight models. Developers don’t have to commit to a single model; the system routes each call to the best option on cost/performance.

Under it all, the flagship Doubao-Seed-1.6 model supports 256K context and combines adaptive reasoning, multimodal understanding, and UI/control capabilities; their Seedance 1.0 Pro video model has tested at or near the top globally on text-to-video and image-to-video benchmarks.

By September 2025, Volcano Engine disclosed that the Doubao model family was handling about 30 trillion tokens per day, up roughly 25,300% from launch levels in May 2024.

According to IDC’s 1H25 report on China’s public-cloud LLM services, it is estimated that Volcano Engine held a 49.2% share by API call volume, i.e., almost one out of every two tokens generated on China’s public clouds came from Doubao or other models served by Volcano. We see that, unlike Alibaba, which is focused on selling its full-stack service, the API is only one aspect, specifically its Qwen3 Max. ByteDance is actually making most of its AI revenue in API sales.

However, as we mentioned in the recent Alibaba Cloud piece, Omdia’s broader China AI cloud report for the same period still shows Alibaba AI Cloud leading with 35.8% share, because it includes IaaS/PaaS and strips out some of Volcano’s private deployments and narrow NLP services.

So, depending on which denominator you choose, you can see that these two are leading: Volcano Engine is #1 in MaaS/API volume, and Alibaba Cloud is still #1 in overall AI-related cloud revenue.

One more data point: local reporting suggests Volcano Engine’s 2025 revenue target is around 20 billion yuan, roughly half Tencent Cloud and about a fifth of Alibaba Cloud, but closing in on Baidu Smart Cloud with a shot at overtaking it in 2026. Volcano Engine becomes the external face of ByteDance’s AI, while Doubao – the consumer app – is the most visible proof-point that this model family can scale.

Doubao the app, the MAU leader

It’s easy to forget now, but Doubao came late. Its official launch was August 2023, when OpenAI had already redefined expectations globally ,and domestic players like Alibaba and Baidu were already on their second or third model iterations. It felt like it led in China in the 2C race for a while, and then one after another, Kimi, DeepSeek, and Yuanbao took over, and Doubao became a side narrative.

For a while, it felt like Doubao had been left behind. Internally, people talk about mini “panic” meetings around that time – not because ByteDance couldn’t build a model, but because they were risk-averse about creating something that didn’t feel “TikTok-level smooth.”

However, those were all temporary. Doubao has now moved up to the most popular 2C AI app in China again in September. It is back as the largest native AI app in China by MAU, with 172 million users versus DeepSeek’s 145 million. QuestMobile’s combined 320 million MAUs for the two also tells you something: there’s room for multiple large consumer AI apps, at least for now.

The product itself is no longer “just chat.” It’s a front-end into ByteDance’s strength zones:

• Video-first workflows – script writing, storyboards, mini-dramas, filters, auto-editing.

• Creator tooling – batch creative generation for ad performance, templated hooks for short videos.

• Education and Q&A – where the 256K context and image-understanding show up in things like photo-based homework help and step-by-step solution explanations.

That’s the “OS of entertainment” thesis I wrote before: ByteDance may not own utility payments or government services, but it owns a large chunk of China’s (and the world’s) attention graph.

Doubao on the phone

The most controversial chapter in this story is the Doubao smartphone AI assistant.

On December 1, ByteDance released a preview of a Doubao phone assistant deeply embedded in the operating system of a ZTE Nubia M153 sample phone, selling at ~3500 RMB (~500 USD).

Instead of just living in an app, this assistant can be triggered via voice, earbuds, or a dedicated AI side button, provide an “AI screen” as an overlay that can read whatever is on your screen and can hop across apps to do things like compare prices on different e-commerce platforms, place orders, book restaurants, control your car, schedule meetings, and batch download files. It even offers a “Pro mode” with stronger reasoning for more complex cross-app tasks, such as planning a trip and booking tickets based on your social-media preferences and, with user permission, stores screen information locally as memory so it can personalize future tasks.

Local media framed this as ByteDance “formally entering the operating system” (much like Tencent and OpenAI?) because it upgrades AI from passive chat and QA to active orchestration across real-world tasks at the OS layer.

Chinese business media Yicai wrote that the partnership with ZTE resembles Huawei’s collaboration with automakers. And apparently, it’s not the first. Alibaba is reportedly in partnership with Oppo Mobile Telecommunications and Honor in a similar capacity as well. But it seems like this agentic layer is much more advanced in user experience.

The smartphone maker ZTE’s A-shares hit their daily limit up, and its H-shares jumped more than 13% on the day of the demo. So I guess that’s a boost for them. However, I’m not sure this is the smartest strategy for ByteDance unless they can guarantee this is such a winning OS that people will switch over from leading phone brands such as Huawei, Samsung, Xiaomi, and, of course, Apple. Given that any of these and likely all of these are working on their own version of this AI-powered operating system.

That said, ByteDance has said publicly it does not plan to build its own phones, and instead wants to partner with OEMs. The Nubia developer phone with Doubao pre-installed is priced at 3,499 yuan and is meant as a reference device for app and ecosystem partners. So maybe it’ll be focusing on lower $ markets? But that’s not the usual ByteDance style. TikTok has always been the “first-tier” city product, while Kuaishou took on the “third-tier” city demographic in the short-video era. So let’s wait and see what the game is here.

Subscribe now

The OEM and super-app pushback

As mentioned above, leading phone makers will build their own versions, but they will have the upper hand of already owning distribution and a large reach. A recent Morgan Stanley note makes similar points: turning Doubao into a true system-level agent requires deep OS modifications and business-model concessions from OEMs, and that directly touches their bargaining power.

While for super-apps like WeChat, Taobao, and Meituan, a cross-app assistant raises the fear that user traffic will get rerouted into generic AI workflows rather than staying inside their ecosystems. But I guess similarly to what we wrote about in the WeChat Mini Programs piece, they could keep their transaction within their apps?

Morgan Stanley’s China internet team is pretty blunt in their note: while Doubao’s phone assistant is technically impressive, they see the rollout as “challenging” because both big OEMs and super-apps have strong incentives to prioritize their own AI agents. They argue that applications remain the best way to play AI in China’s internet space, and they reiterate Overweight ratings on Tencent (best AI app proxy), Alibaba (best AI infra), and Meitu (multimodal / last-mile beneficiary).

In other words, the Nubia phone is a very visible pivot into the OS, but it doesn’t automatically guarantee Doubao becomes the default assistant on the biggest Android distributions, let alone iOS. However, even if this phone doesn’t take off, this announcement still serves as a demo of what Volcano + Doubao can do for enterprise and app partners.

Is ByteDance’s strategy coherent?

Yes. I’ve spent the last two weeks looking into this and have been surprised to find it makes sense now. If you strip away the noise, the spine holds.

Zhang Yiming didn’t so much retire as step back from the day-to-day to plot the long arc, and you can feel that founder-level obsession in the way Seed has hoovered up talent from DeepMind, Google, Alibaba, and the newer labs.

The spending profile reads like a fourth hyperscaler, but the P&L is still anchored in ads and short video, which is precisely the point: they’re funding a model-and-infra bet with products that already work, then repackaging those same primitives for others to buy.

Volcano Engine is not trying to be an IaaS. It’s a MaaS-first business with sales attention to go hand in hand with it, and internally, you can tell it’s where the resources are going.

Doubao plays a dual role here. First, as the 2C flagship app and live sandbox, the MAUs aren’t just a vanity scoreboard, but rather the ongoing data and distribution loop that hardens the model and feeds the enterprise funnel.

Second, it’s trying to integrate hardware and software. Compared with peers, ByteDance is simply being bolder at the OS edge right now with this new phone partnership. While everyone is talking about agents, they actually shipped a system-level demo that can see the screen and act across apps. Even if the components are becoming common and will likely be overtaken, the courage to run that experiment in public is not. And it’s essentially signalling to enterprises that it is open to partnerships.

Open questions to end on

Can ByteDance persuade enough OEMs to allow a Doubao-level agent to sit above their own assistants, when doing so clearly dilutes their leverage with users and carriers?

The same question exists for super-apps: will WeChat, Taobao, Meituan, and others tolerate a cross-app butler that can intercept intent before it hits their native flows, or will they push harder on keeping transactions and orchestration behind their walls?

There’s also the pricing gravity no one can ignore. As in, if DeepSeek-style open weights keep driving effective model costs toward zero, the margin stack in MaaS gets squeezed, which would push ByteDance to lean even more on first-party surfaces (Douyin/TikTok, CapCut, ads) to pay back the capex.

To conclude on an attempted witty summary: Bytedance is trying: own the model, own the infra, ride the creators, and then see how far that stack can push into the rest of daily life. So far, a coherent strategy. But my bet is still on Alibaba Cloud to be the enterprise’s first choice.

Singapore was T-shirts, sneakers, and founders talking about “go-global” AI apps. Hong Kong was the Ritz-Carlton, harbour view ballroom, and a sea of suits - including my own. And, it was very clear, even from the choice of clothing, that the people Alibaba is targeting are enterprises and businesses. In Chinese, I would say the event was also 干货满满 - full of substance.

Now, Alibaba Cloud has always positioned itself to serve enterprises, but this event made me realize how it is really able to capture that better than others. It’s not just in its cloud capabilities because other hyperscalers also offer cloud. Let me dissect why it’s able to solidify its top-of-mind 2B AI position in this article, and what it’s selling beyond cloud and services.

Anyway, I’m not complaining about the Alibaba Cloud event here. I was able to geek out over AI, talk to enterprise adopters and in-house developer teams, noms on some Har Gow, and play around with all the toys at the Cha Chaan Teng (茶餐厅/ local dim sum cafe)- themed exhibition.

Share

Thoughts on 2B AI

Before we get into Alibaba, I think this 2B revelation isn’t new tbh. Much of the investor and founder focus has shifted recently in China.

Just two days ago, I listened to an interview with the former Zhipu COO, now CEO and founder of 元理智能(Yoolee AI ), Zhang Fan. He said that the initial wave of interest in 2C AI applications in China made sense because in China, there is a shared view that enterprises aren’t willing to pay, but that is in the internet era. But after the initial gold rush has passed, founders and investors are realizing there may be more opportunities 2B.

The physical infrastructure and digital infrastructure have evolved drastically since the first generation of internet companies was founded. Today, it actually makes less sense to try to make money in the 2C AI space because there isn’t actually new hardware or a device breakthrough, and the internet ecosystem is also saturated and market share taken by incumbents. Opportunities now lie within 2B use cases.

Zhang also had a pretty solid quote: “一个好的产品但是没有闭环的生意不是一个好的公司.” As someone who comes from a product background and is now a founder, his comment was very insightful.

He essentially said a good product without an ecosystem (flywheel business/ closed-loop business) is not a good company. Referring to the many AI startups that are focused on fundraising and marketing their product rather than creating a sustainable business model. And at AI Proem, what we do here is not to demo the best new product, nor is it to dissect the technicalities of new models. There are people who can do that much better than I can. What I focus on is understanding the strategy and business of tech and AI firms.

(I also realized I met him last year at BEYOND, so I guess he left Zhipu shortly after this. See selfie here.)

Ultimately, he points out that whether an enterprise pays for a service depends on whether the good/service actually generates more revenue for the business, because that is what any company's core objective is. And he said - there is no need to overcomplicate things. His thinking reminds me of this article that I read by Greylock, which also says that with the rise of LLMs, more vertical use case startups can tackle bigger problems, and “with reduced barriers to building AI applications on LLMs, data is arguably the most important currency in building a differentiated position.”

I’m providing the context as to why Alibaba has an edge in enabling enterprises to adopt AI.

Anyway, taking it further, Zhang says AI is actually not a tool, and cannot be compared to SaaS. The 2B market has potential as companies are willing to pay for results, just look at the BPO (Business Process Outsourcing) businesses. So he believes that the enterprise will be willing to pay for AI if it actually acts as a colleague and thus generates more revenue. And more importantly, an AI enterprise product shouldn’t be a one-size-fits-all solution, as every business has different needs and operates in different environments.

As models have reached a certain level of intelligence (not just good enough, but actually better than the average human across advanced science subjects), it is more important to actually have different kinds of models to serve different purposes, and then the ability to fine-tune and adjust the models for specific verticals. Then, more importantly, have the agent have the intelligence to actually adapt to different demands of different businesses - that is where the potential business use case lies.

With that, let’s see how Alibaba is capturing enterprise business and helping enterprises adopt AI.

Subscribe now

Alibaba Cloud Summit HK: The 2B AI Playbook with Qwen, Wan, and Enterprise Stack

From Tongyi Lab to pharmacovigilance and smart cockpits. Blending large models, data platforms, and real-world use cases.

Alibaba Cloud’s 2B adoption

The Hong Kong summit was framed very explicitly around enterprise use cases, and most attendees were clients of theirs who already use Alibaba Cloud in their workplace.

Maxim’s, the F&B group, was one case study among many. Their story is almost the archetype of AI 2B diffusion: highly interdependent POS and app systems, traffic spikes during promotions, rising security and data privacy risks, and no interest in building an in-house LLM lab.

With Alibaba Cloud, they’ve gradually instrumented stores and kitchens with IoT sensors, streamed that data into the cloud, and layered forecasting models on top so staffing and prep can be adjusted hour by hour. It’s not glamorous, but it’s what keeps a store from breaking when the front page of a delivery app points everyone to the same outlet.

Alibaba Cloud is not positioning itself as a solutions system in hardware and software. And that's where Tongyi comes in.

Tongyi, Qwen, and Wan: the model backbone

The most memorable speaker for me was Xu Dong, who was extremely articulate and logical.

Xu Dong, General Manager of Alibaba Cloud’s Tongyi business, laid out the model story in a way that made sense against this 2B backdrop. Internally, Alibaba now talks about two main model families living under Tongyi:

Qwen (通义千问) is the GPT-style large language model family: text, reasoning, coding, agents. Wan (通义万象) is the Sora/Veo-style video and world-simulation line. While Wan has been around the whole time, it is only gaining more attention with the rise of Sora.

“Open source is our core strategy.”

On Qwen, the scale numbers are already big. As of November 2025, Qwen has more than 300 open-source models released, over 700 million downloads, and around 180,000 derived models built by the community. By Alibaba’s own count, this makes Qwen the world’s largest open-source model family. Whether or not you accept that exact crown, it’s clearly one of the deepest ecosystems out there.

And guys, I’ve really been writing about this for over a year now. Alibaba, in a way, has made it quite clear that its strategy was always cloud sales first. They’re not selling API, they’re selling the foundational infrastructure. Joe Tsai confirms that in an articulate way for us all. And thank you to AI深度研究员. Here’s a clip in English.

Qwen 2.5, their current flagship, is not just about answering questions in a chat box. One capability they emphasised was the ability to ingest long text and automatically generate mind maps and posters – the kind of text-to-visual structuring that tools like Nano Banana have made popular. In a 2B context, that matters because most knowledge work still ends up as slides, diagrams, or briefs, and if a model can jump straight from documents to those client-ready materials, it saves a lot of tedious grunt work.

Qwen-VL is the vision-language branch. Here, the goal is to move beyond “this is a cat” to real scene understanding: 2D/3D spatial awareness, reasoning under partial occlusion, and low-light conditions where even humans hesitate.

Xu Dong kept describing this as the shift from 看见 (seeing) to 看懂 (understanding). They also used the word “Omni” quite deliberately. For his team, an omni model doesn’t simply support multiple modalities. But it means that the models can see and listen at the same time, and respond with a mix of text, images, and speech, without feeling like a pipeline of separate systems. In that sense, it's more like how humans absorb information. You observe your surroundings, you hear sounds, you smell scents, and then you digest the information you’re provided with, and that is how you form nuanced opinions.

The trade-off they freely admitted, well, as you make the model more multimodal, its apparent “IQ” often drops - it’s distracted in that sense, I guess. Aligning text-only models is hard. But aligning models that must reconcile conflicting visual and audio cues is harder. A lot of Tongyi’s engineering effort right now is focused on regaining instruction-following and reasoning performance within that more complex omni-setup.

On the other side of the house sits Wan, Alibaba’s video and world-simulation model. The latest version, Wan 2.5, is their answer to Sora-style generative video. The demos were exactly where you’d expect the frontier to be: movement and physics that look much less floaty than the previous generation, sound that matches onscreen motion more closely – but still obvious flaws if you look closely, like a pianist whose fingers don’t quite match the keys and melody, or multiple characters whose lip movements drift out of sync.

Xu Dong’s own assessment was that last year’s video models were basically toys; this year’s are finally usable for production tasks like ad creatives, short drama transitions, and filler footage.

So in 2016, we’ll likely see them continue their pace of pushing out new models. You can expect a new model every quarter, and tweaks and upgrades every ~1 month or so. The 2B angle is that Tongyi Lab is shipping out a portfolio of LLMs, vision models, and video models that Alibaba Cloud can then slot into different parts of a customer’s stack, from call-centre assistants and pharmacovigilance workflows to smart cockpits in EVs.

[Btw, it was nice to hear directly from the team that it’s pronounced “Q-when” as I initially suspected, since it’s the abbreviation of the Chinese pinyin 千问 “Qian Wen”. But it seems like English speakers have adopted ‘Qwen’ (like Gwen) in the mainstream, so I’ve started saying that too.]

The four quadrants: how Alibaba maps model applications

One slide from the summit captured Alibaba’s mental model for how these large models diffuse into the real economy: “Four Quadrants for Large Model Applications.” It’s worth unpacking because it’s essentially their go-to-market map.

On the left, they grouped what you can think of as front-end experiences and devices. This is where models surface directly to consumers: AI TVs that understand context and content, AI mobile phones with built-in assistants, robots, AI-assisted healthcare scenarios, autonomous driving and smart cockpits, AR/AI glasses, AI personal assistants.

Below that, they put AI-generated education summaries, AI-generated PPTs and music, AI-generated images, search, and document writing. Taken together, this is the “what people see and touch” surface area, which you can see as the AI layer that sits on devices, in car dashboards, in classrooms, and in content tools.

On the right, the slide shifted to enterprise brains and back-office work. At the top were what they labelled as new business-model activities: AI research, product design, code generation, marketing copywriting, business process optimisation, and smart warehouse replenishment. In other words: using models to invent new products, optimise how they get built, and change how warehouses or factories are run.

Underneath was the world of repetitive but heavy knowledge work they believe models will simplify: enterprise knowledge management, massive contract review, video security checks, production quality inspection, intelligent Q&A and training, unstructured information extraction, massive data labelling, AI-powered office collaboration, and AI business services.

• If you slot in some real cases from the summit, the quadrants become less abstract. Maxim’s demand forecasting and staffing falls into the “business process optimisation / repetitive operations” area: take real-time data and use models to smooth the spikes.

• Pharmacovigilance sits in the knowledge-intensive repetitive work quadrant: thousands of near-identical reports that must hit a regulatory standard, where a model can draft 90–95% accurately, and humans clean up the rest.

• Smart cockpits, AI glasses, and robot use cases fall into the device and experience quadrant on the left. And when they talk about AI-assisted research or product design, you can see the link to Qwen’s ability to consume huge corpora and generate structured diagrams, code, and design variations.

• The centre of the slide maybe says more than any of the boxes. They drew a blue clover of four overlapping circles: innovation, product strength, labor force, and “prolificacy,” with arrows pointing up to “new business model (new quality productivity)” and down to “simplify repetitive tasks (anti-involution).”

Alibaba’s AGI (ASI goals) chart: X, Y, Z

When Xu Dong talked about AGI, he didn’t jump to sci-fi. He drew a simple 3D coordinate system that connects back directly to the rest of the summit. In Alibaba’s own language, they call it the Artificial Super Intelligence (ASI).

The X-axis is knowledge, which in practice is language. For both humans and models, language is how knowledge is stored and transmitted. Train on more book-equivalents, code repositories, and multimodal data, and you move further along X. Here, the work is about data diversity and token efficiency – learning more from fewer tokens because each token costs GPU and electricity.

The Y-axis is smartness, defined very specifically as the ability to use tools. Humans aren’t special because we can talk; we’re special because we can pick up tools, chain them together, and adapt them to new situations. In the model world, that translates to an agent that can call ERP systems, CRMs, search APIs, workflow engines, and know whether it has chosen the right tool and used it correctly. This is the difference between a pretty demo and something that can actually sit inside a bank’s middle office or a logistics network.

The Z-axis is omni-modality: seeing, reading, listening and speaking in a single coherent loop. A text-only model can be powerful, but it will never fully inhabit the world we live in, which is full of screens, cameras, microphones, and physical environments. That’s why Alibaba is spending so much effort pushing Qwen-VL and Wan forward even when it temporarily hurts benchmark scores.

Viewed through that coordinate system, the rest of the summit content becomes easier to place. Qwen and Qwen-VL push hard along X and Y, and increasingly along Z. Wan pushes on Z with some Y. The four application quadrants are just slices through that 3D AGI space: some use cases live close to the “repetitive text work” corner while others lean heavily on perception and action.

Underneath it all: the data platform play

And finally, the data layer. Essentially, the initial GenAI frenzy is over, and the challenge for many is now to bridge the gap between potential and production. The next stage of AI integration in the enterprise is actually cloud services that can deeply integrate with enterprise data.

What we’ve seen over the last few years is that LLMs still largely operate on static, pre-trained knowledge. The challenge now is not building another model in isolation but being able to connect the model to the live data and environment where business decisions are made.

So what Alibaba Cloud is proposing is serving the enterprise growth with both cloud and models in sync.

[Illustration done with NotebookLM]

The last part of the summit that tied the story together was the data and database narrative from AliCloud’s Chief Data & AI Architect and Head of Database Architecture. The company’s view is that by 2028, roughly 80% of GenAI business applications will be built directly on top of existing data platforms, according to Gartner. The future of enterprise AI isn’t going to be another isolated tool, but is somewhat likely to be built into the data platforms companies already trust.

In their words, large models by themselves are kind of useless. The moment you care about what happened to your orders in the last five minutes, or which warehouse is out of stock, you have to plug the model into your databases and event streams.

[Illustration done with NotebookLM]

So AliCloud is reshaping its database products into a unified AI data platform. The building blocks are familiar to those who follow the infra layer closely. (I for one am not an expert here) But at a high level, Snowflake, Databricks, or MongoDB have focused on building cloud-native databases like PolarDB that scale elastically and stay online through traffic spikes, while lakehouse-style architectures blur OLTP and OLAP.

Retrieval-augmented generation is plugged into the platform and vector storage, and semantic search, so that you can retrieve “things like this” rather than just keyword matches, so what we’ll see is that increasingly, multimodal storage so that the same system can index text, images, and video.

[Illustration done with NotebookLM]

The philosophy shift is captured in one phrase they used: move from “bring data to the model” to “bring the model to the data”.

Instead of exporting data into a separate ML sandbox, you embed the model into the database and analytics environment. Data ingestion pipelines automatically vectorise documents and events; a “Smart Search / 智搜” layer sits on top, and then DB ops and BI both get their own agents.

One example shared was Bilibili, the Chinese YouTube, if you must simplify it.

Its ability to understand what’s happening inside each video and its danmu comments, and then monetise that with more contextual ads, hangs on this stack. So do the pharmacovigilance workflows and contract-review use cases.

[Illustration done with NotebookLM]

If this strategy works, Alibaba Cloud’s business mix changes meaningfully. The company stops being just a provider of CPU and storage to other people’s applications and becomes the place where models live, data converges, and agents operate.

And because the models (Qwen, Wan), the platform (Tongyi, OpenLake, PolarDB), and the vertical solutions (pharma, F&B, logistics, media) all sit inside the Alibaba ecosystem, there is a clear flywheel: more workloads and data lead to better models and better models justify more customers consolidating on the platform. And we’ll see the flywheel much more apparent now that mass adoption of AI has moved from chatbots to agentic. This brings us back to Alibaba’s original -sell cloud, get data, add on services, improve models, and then sell more cloud thinking.

From the outside, it’s easy to get distracted by headline model releases or one-off demos, and it’s impossible to keep up tbh.

What the Hong Kong Cloud summit clicked for me was that Alibaba Cloud has always actually played the enterprise strategy, just one card at a time, and it’s now becoming clearer to everyone that the company is using those models as leverage to become the default 2B AI solutions provider.

Share

Quick take of the week

And as OpenAI is also trying to find its next enterprise fit. I think as we move towards 1) more R&D, as Ilya Sutskever talked about with , we’ll see 2) more diffusion in the real economy, in forms beyond chatbots. And we’ll also see more 3) enterprise courtship by all the leading labs right now (like BABA). And lastly, we’ll likely see 4) some consolidation in the LLM space.

See additional slides from the Alibaba Cloud Summit, FYI

I am moving about a bit before the year-end. I just returned from Shenzhen after attending the Xin Conference over the weekend, and I’ll be in Singapore for the next few days. Once I return and digest all the meetings and chats with founders, builders, and investors, I’ll write up some of my key learnings.

And for those in SG who want to catch up, feel free to message me here or DM me on LinkedIn!

Toodaloo~

This was first published on last week as a guest post, but I’ve made some minor updates in the Commerce and App sections since then.

Ahead of Earnings: What to Note and Expect?

Alright, the good ol’ Double 11 has just wrapped up last week, and it’s only a few days before the company’s next earnings announcement. Let’s take a good look at what Alibaba has been up to and what to expect.

Share

“This quarter, our strategic focus on consumption and AI + Cloud delivered strong growth……Looking ahead, we remain committed to investing in our two strategic pillars of consumption and AI + Cloud to capture historic opportunities and drive long-term growth,” said Eddie Wu, Chief Executive Officer of Alibaba Group.

That was the opening statement Alibaba’s CEO Eddie Wu said during its June Quarter 2025 earnings call. Stating a clear message: Alibaba is making a comeback, and it’s approaching AI with two strategies. So, for this quarter, the main question I have now is whether the consumption flywheel can continue to turn, and whether AI and cloud can translate into durable revenue drivers.

Subscribe now

What’s Alibaba doing? Well, pretty much everything. So today, we’ll break it down.

1. Alibaba’s capex spending is leading in China, but lags significantly behind U.S. peers.

2. Commerce business is making a comeback and integrating AI across all steps from intent to transaction.

3. Alibaba Cloud, ranked 4th globally, 1st in Asia, is doubling down on its enterprise strategy.

4. Open-weight Qwen models and the Tongyi ecosystem continue to bring businesses in. Tongyi rebranded to Qwen for the 2C app.

5. AI is pretty much embedded in every 2C and 2B touchpoint. Real-life business data is their true advantage in the agentic era.

AI Capex

First, let’s look at spending. The company hasn’t adjusted its RMB 380 billion​​ (~ USD 53 billion) capex spend over the next three years since its initial announcement in February. As we wrote half a year ago, Alibaba’s massive capital allocation has been heavily spent on building its chip capabilities and growing its cloud capacity. However, it pales in comparison to those in the U.S.

Actually, for context, if you look at the total capex amount, Chinese hyperscalers’ combined capex in 2023, 2024, and 2025, as expected, will be ~ USD 124 billion, which is 82% lower than that of its U.S. peers at roughly USD 694 billion, according to Jefferies’ latest China Tech report.

Over the past six months, the build-out of AI infrastructure has contributed more to the growth of the U.S. economy than all consumer spending. In the past three months alone, the magnificent seven have spent more than USD 100 billion (!) on data centers and related infrastructure.

According to IDC’s April forecast numbers, global IT investment in AI is projected to reach ~USD 815.6 billion in 2028 (THESE ARE MAD NUMBERS). The generative AI market is expected to reach USD 284.2 billion by 2028, accounting for 35% of the overall AI market investment.

Comparing the China numbers, the U.S. hyperscalers are definitely spending way more, and are projected to be creating a much larger market.

Specifically, IDC highlights China’s continued leadership in the APAC AI market, with its spending exceeding 50% of the region’s total. IDC further projected that China’s total AI investment would surpass USD 100 billion by 2028, and its generative AI market would exceed USD 30 billion in the same year.

Commerce

The 17th Double 11 just concluded. Starting on Oct 15, the big sale lasted for nearly a month. Although the festival felt a little underwhelming for Alibaba compared to its heyday, when it even had T-Swift leading live concerts and countdowns, we cannot deny its impact on retail as a whole. Nonetheless, it is essential to note that it achieved the best growth rate over the past four years. If anything, it shows some revitalizing of consumption intent.

What is evident is that the era of competing solely with flash numbers and traffic wars is now behind us. a16z recently suggested that the traditional “search-compare-buy” model is being replaced by AI agent-driven purchasing—essentially, search on steroids.

At the same time, Precedence Research projects that the global AI-enabled e-commerce market will grow from USD 725 billion in 2024 to USD 64.03 billion by 2034, over the next decade.

In Alibaba’s own corporate blog, it boasted that “LLMs have been integrated to core search and recommendation engines,” supposedly leading to “double-digit improvement in result relevance for complex queries and a double-digit lift in click-through for certain recommendation scenarios.” But frankly, I’m not sure how that is measured.

Kaifu Zhang, Head of AI Initiative at Alibaba International Digital Commerce, said, “Useful AI” is the benchmark for Taobao’s AI applications. And we’re really seeing this recent Double 11 serve as the first major showcase for this quiet revolution, where Taobao (the leading e-commerce platform in China) has applied AI comprehensively across traffic distribution, consumer experience, and merchant operations to drive efficiency and improve the user experience for transactions.

AI is now no longer a supplementary tool but a core variable embedded into every环节 (link/ step) of the e-commerce system, driving smarter operations, growth, and GMV.

The evolution from traditional to multimodal recommendation systems represents a fundamental strategic upgrade that transforms a key part of the digital platform experience, evolving it from a reactive utility into a proactive growth engine.

E-commerce

Traditionally, recommendation engines have acted like junior sales clerks, relying solely on historical purchase data to suggest items. This approach severely limited business strategy by failing to effectively promote new products with no sales history (the “cold start” problem) and by missing opportunities in niche, “long-tail” markets.

The new paradigm of multimodal AI shatters these limitations by acting like a master sales expert who can actually see and understand your product catalog. By analyzing product images and reading descriptions, the system grasps the intrinsic qualities of items, allowing it to intelligently recommend a new restaurant based on its appealing food photography or connect a user with a unique culinary interest to a particular dish they’ve never tried before.

This directly enhances business strategy by accelerating the successful launch of new products, unlocking new revenue streams in specialized markets, and creating a more intuitive and satisfying customer experience that fosters loyalty.

By building this product intelligence and connecting it to user intent, Alibaba achieved direct bottom-line results, including a significant increase in orders and total sales volume. Looking ahead, the move toward generative recommendation will further deepen this strategic advantage, enabling the system to explain its reasoning in plain language, thereby building unprecedented trust and personalization that can dramatically increase a customer’s lifetime value.

Sidenote: Today, we won’t delve into the nitty-gritty of a side commerce business that showed an unexpected twist - that is, seeing Ele.me and AMap, two struggling businesses, revive and not only resurrect but thrive. What Alibaba has done is rebrand these businesses into a new category called Taobao Instant Commerce. You can check out who explains it well here.

Now that’s the consumer side of the story. The other half is about Cloud + Qwen.

Cloud

What is Alibaba’s game? Well, I’ve written extensively about how it’s actually selling the full stack. It’s capturing value across the ecosystem. (What Alibaba is not doing is chasing the AI empowered consumer super-app. That’s Tencent’s game.)

In the June Quarter of 2025, Alibaba’s cloud intelligence revenue grew by 26% with AI-related product revenue achieving triple-digit growth for the eighth consecutive quarter, primarily driven by public cloud revenue growth and an increasing adoption of AI-related products, the company stated. There has been a shift in sentiment, where the company is now pursuing a long-term strategy to capture new valuation creation not just in the present, but also considering its strategic position in the next ten years.

And at the Alibaba Cloud flagship event Apsara, the company announced plans to open data centers in Brazil, France, and the Netherlands, as well as facilities in Mexico, Japan, South Korea, Malaysia, and Dubai in the coming year. With existing service centers in Germany and Indonesia. Going big on its international expansion and ability to serve customers worldwide.

What we can see is Alibaba Cloud doubling down on its enterprise-first strategy, and we’re really seeing the model+cloud bundle playing out to be a smart strategic move.

Investment bank Jefferies recently wrote in a note that they’re expecting to see triple-digit growth in AI-related revenue for Alibaba in the September quarter (which is the upcoming one), that’s BOLD.

Adding that, as Alibaba pursues its “one cloud with multi-chips” strategy, the company is likely to see revenue boosted by (1) AI demand by customers, (2) inference growth, and (3) an increase in training demand for proprietary data across sectors.

And as the company continues to ship open-source models, it has organically brought in companies also seeking cloud infrastructure support. Its openness has proven to be a successful bet as there are now about 80% of the tech companies and half of the LLM labs in the Chinese market all running on Alibaba Cloud. There are nearly 7 million developers in its ModelScope community with 10k models available. Let those numbers sink in.

According to Omdia, the AI cloud industry growth is expected to grow at ~150% YoY in 2025. The research group highlighted that Alibaba Cloud ranked first, again, capturing nearly 36% of the AI cloud market share in the first half of 2025. A dominant leader, taking up more market share than the share of #2,#3, and #4 combined.

And finally, taking a step back and looking at China’s cloud sector as a whole, we can see that China’s cloud revenue only represents a 10% market share in global cloud revenue currently, according to Jefferies. There is still significant potential for that to grow.

Qwen

What runs on the cloud? Well, the Qwen models.

At the most recent Apsara event, CEO Eddie Wu said, “We remain committed to open-sourcing Qwen and shaping it into the ‘operating system of the AI era,’ empowering developers around the world to build transformative AI applications.”

Qwen is continuing to lead across benchmarks in open-source/weight models, even surpassing Meta’s Llama as the world’s leading open-source model family. What it has built is a cost-efficient, scalable, and deployable ecosystem. Even Brian Chesky from Airbnb has admitted to building on Qwen. And beyond text, the stack now includes open-source video gen (Wan 2.1), which helps pitch full-stack multimodal to developers.

DeepSeek’s big win nearly a year ago, proving that ‘good enough’ can be ultra-cheap, has also been a big booster for Qwen. Enterprises in China, in particular, now expect lower-cost, model-agnostic AI, putting Alibaba Cloud’s win condition squarely on price/ performance and Qwen efficiency.

The reality is that the intelligence score gap across is becoming smaller and smaller amongst the top frontier LLMs, but the cost efficiency is quite different.

According to the company’s own figures, as of September this year, the company has reported over 600 million model downloads of its Qwen and Wan series, approximately 170,000 derivative models, and more than 800,000 AI agents built on Model Studio.

While the performance of China’s best AI model - the latest MiniMax M2 is only 10% behind the best US AI model GPT 5, according to Artificial Analysis. The cost of achieving that is significantly less than what it cost OpenAI. You can probably conclude that Chinese AI labs - including Alibaba- are currently much more mindful of the price-to-performance cost-benefit.

Apps and Agents

But models don’t sell themselves. They need surfaces. That’s where Qwen (the app), Accio, Quark, and Qoder come in—agents wired into real APIs, not just demos.

Tongyi - Qwen

The most interesting news emerged last Thursday when the company announced that it would be renaming the consumer-facing app Tongyi to Qwen. For context, Qwen in Chinese is 千问, which actually means a thousand questions.

This will be a game-changer because, thus far, there has not been a clear 2C AI app absolute dominant leader yet. ByteDance’s Doubao is hovering around 172 million MAU, while DeepSeek’s app is around 145 million MAU. For an economy with over 1 billion internet users, there are no clear winners yet.

With Alibaba throwing its hat in for the 2C race, its ability to connect Taobao, Alipay, and Ele.me all through the app Qwen and leverage its Qwen models could be a complete game changer. It’s as if Amazon actually had a leading frontier model and also owned PayPal, as well as DoorDash. In this case, you can easily then tell Qwen, “hi, I’m hungry, I want Thai food”, and that search, purchase, transaction, and even delivery will be completed within the app.

Agentic Future

Beyond its 2C app, the company has released an array of products that are AI-powered, from Accio, its AI tool for merchants that helps them build campaigns and run marketing materials much faster, to Quark, its internally built AI agent, and the newly launched Qoder, its agent-based coding tool.

While many believe there is significant potential in Alibaba’s AIGC creation tools and value for merchants, driving a 12% increase in marketing ROI. What I’m realizing is that the edge for these in-house built agents isn’t actually just the infrastructure access, but it is the high-quality, real-life business use case data Alibaba could provide them.

Since “context is king,” having real-life business use-case data will actually give Alibaba agents a significant advantage over companies trying to build agents from scratch and those attempting to purchase data or partner with companies to obtain real-world use-case data.

What we know is that Alibaba is the only Chinese big-tech with both a scaled cloud and a leading open-weight model, now moving into the agentic era, their real-life business data will be another key differentiator.

Subscribe now

Some Closing Thoughts

Ultimately, the big tech race on AI is only just heating up. Now, even food delivery apps like Meituan are shipping their models. However, in China, Alibaba and ByteDance are emerging as obvious leaders in AI infrastructure, proprietary models, and app integration.

As Chinese AI labs such as Qwen continue to lead in open source models and focus on improving model efficiency to cope with the hardware constraints, driving down token cost, thus API pricing, we can see that with already API pricing being the lowest globally. China (and, in this case, BABA) will continue to diffuse the technology across real-world use cases at a significantly faster pace.

As for Alibaba’s cloud investment, it seems to be finally paying off. The company’s strategy of integrating AI across the full stack – as a service, as an infrastructure, as a tool – is allowing it to capture value and find synergy across it all. So now, let’s see, will this coming quarter be a home run? (like Jefferies’ anticipated triple-digit growth kind of home run?)

Some Relevant Reads

• Alibaba Q3 FY25: Forget About ROI, Let’s Talk About AGI

• China AI Capex: State vs. Corporate Capital Injection

• Alibaba Cloud’s Hypergrowth, China Industrial Prowess, AI Video Editing Wows

• China and the US are Running a Different AI Race

• Why Tencent’s Integration of DeepSeek Into its AI App is a BIG DEAL

[Ok, I first pushed my original analysis out Nov. 14, but it seems like people are catching up on my analysis of Alibaba’s big move into 2C, so I updated this piece on Nov. 17 with some snippets from LatePost’s interview with Alibaba directly and some coverage by 36kr.]

Subscribe now

In the U.S., OpenAI’s ChatGPT has become synonymous with consumer artificial intelligence, with an estimated ~800 million monthly active users. People do not say, “I am going to use an LLM.” They say, “I am going to ask ChatGPT.”

In China, the world’s second-largest economy, that obvious consumer “seat” has yet to be crowned.

No single app has reached similar dominance. According to QuestMobile, Doubao is currently in first place with around 172 million MAUs, with DeepSeek’s app ranking second at about 145 million. Combined, they reach roughly 320 million people, which is about one-third of China’s internet users. For context, that is huge, but still nowhere near a ChatGPT-like, default position in people’s daily lives.

The thing is, DeepSeek only allocates a tiny fraction of its resources to consumers and does not actually care about reach. It is focused on training and infrastructure. Its consumer product lacks polish in its interface and has no real distribution rails of its own.

Tencent, on the other hand, has been the obvious winner in distribution with more than 1 billion users and an all-encapsulating app, but it has not had a true frontier model.

ByteDance is doubling down on models and has integrated Doubao into TikTok/Douyin, but at its core, it is still an entertainment-first business, not an operating system for daily life.

But hold on, now that Alibaba has properly stepped into the race with Qwen, my call is that 2026 will be the year of the 2C AI battle in China.

Big picture

China’s generative AI user base more than doubled to 515 million in the first half of 2025, according to the China Internet Network Information Center, representing about 36.5% of the country’s internet population. Yet this growth masks a key reality: unlike the U.S., where ChatGPT and Google’s Gemini have captured broad consumer mindshare, China’s market is split across enterprise tools, niche apps, and integrations inside existing ecosystems.

As we said, there has not been a clear 2C AI app that is an absolute dominant leader. ByteDance’s Doubao is hovering around 172 million MAU, while DeepSeek’s app is around 145 million MAU. For an economy with over one billion internet users, that is all a bit meh. No one has hit that exit-velocity number yet.

Actually, if you look at DAU numbers from QuestMobile, in October 2025, Doubao had about 54 million daily active users, DeepSeek had around 29 million, and Tencent’s Yuanbao had roughly 6 million. No one in China has stably broken through the 100 million DAU line, which is probably the minimum threshold for a true “national” AI app. So in the LatePost interview, Alibaba confidently said, “We are not late” — that is the backdrop. The consumer race is crowded, but no one has actually won it yet.

BAT: different superpowers

Tencent: think Apple, not Meta

Tencent’s original bet was simple. It did not have to own the best frontier model. It already owned the rails.

WeChat has more than a billion users. For people outside of China, WeChat is often lazily described as “like WhatsApp” or “like Instagram.” That framing is wrong. Tencent is not Meta. It is much closer to Apple.

WeChat is not just a chat app or social media. It is its own operating system. Inside WeChat, you chat, pay, shop, book doctor appointments, pay for your utility bills, call a ride, run your side business, and consume content. The real moat is not any single feature. It is the die-hard loyalty and habit. WeChat is where your life is anchored.

Historically, Tencent leaned on that distribution moat. Its own model, Hunyuan, exists, but it has not been seen as a top frontier model. (Though the next Hunyuan will see meaningful improvement, said mgmt today).

The thinking was that the distribution beats the model (or they just haven’t figured it out yet). You can always swap in a better model later or integrate someone else’s. Behind the search bar, the user does not really care which model is answering. They just type into WeChat.

This is where the Apple analogy helps. Apple is already working with external model providers to power Siri and Apple Intelligence on device. The brand and the OS remain Apple. The user does not think “now I am using someone else’s model.” They are just using Siri.

Tencent could easily end up in a similar position. It might not ship the absolute best model. What could be a win-win deal is that they could pay Minimax, Moonshot, DeepSeek, or others to sit inside the WeChat experience. And that could still be a very good outcome for Tencent.

The open question is whether that is enough in a world where Alibaba is now coming in with both a strong model and very powerful life touchpoints of its own.

This is also why Tencent looks a lot quieter than Alibaba on the AI infrastructure front. Capital markets would love to see another “China hyperscaler” story and just project Alibaba’s spending pattern onto Tencent, but the business reality is different.

Proven again in its recent earnings announcement. Tencent’s three core engines, social, games and ads, are still growing nicely and have pretty good visibility. AI is already seeping into every layer of those products, from content recommendation to ad targeting to game operations.

For Tencent, the current AI strategy is more about squeezing better performance out of every unit of compute and lowering deployment and inference cost, not about building the biggest, loudest GPU farm.

There is also a very practical scar-tissue example sitting right in front of them. Meta is another social giant that decided to go very hard into AI infra and now has to explain to investors why capex looks like a black hole.

Seems like Tencent does not want to wake up one day and realize it built a hyperscaler balance sheet for no good reason. So it stays more disciplined.

Alibaba: Qwen and “intelligence direct”

Up until now, Alibaba in the AI story looked a lot more like Amazon or Microsoft than like Tencent or ByteDance. It did not have a WeChat or Douyin-style national traffic firehose, so the instinct was to stand behind the scenes, buy cards, build data centers, and sell cloud plus AI to every industry.

Most of its AI progress showed up in the infra layer. It took stakes in several top Chinese model startups, cranked out its own Qwen family that quietly beat other open models on a bunch of benchmarks, and pushed Alibaba Cloud as the way to catch “AI for a thousand industries.” The consumer experiments felt more like side quests than the main plot.

Alibaba’s “ah ha” moment on the consumer side came when it decided to stop hiding its model behind enterprise branding and bring it straight to the front.

The company renamed its consumer-facing Tongyi app to Qwen today. In Chinese, Qwen is 千问, which literally means “a thousand questions.” The name is not an accident. It matches a pattern that Stratechery has written about in the West.

OpenAI’s consumer app is ChatGPT. The underlying models are GPT. Google’s models and consumer app line are called Gemini. You get the gist. You, as the user, feel like you are talking directly to the “intelligence,” not some generic assistant name that is separate from the model.

Alibaba is doing a version of that. It has been very vocal about how far its Qwen family of models has come and how it now believes it has one of the best, if not the best, domestic models. By putting that same brand on the consumer app, it is telling users: this app is the model. You are speaking to Qwen itself.

This pivot did not come out of nowhere. All year, Alibaba has been firing off big AI moves every few months. In February, right before its FY24 earnings, it pledged more than 380 billion RMB over three years to build cloud and AI hardware infrastructure.

In May, ahead of Q1 FY25, it officially launched the Qwen3 model family. In September, The Information reported that Alibaba was working on a new AI chip. Internally, by the summer of 2025, the debate shifted from “do we have the model and the infra” to “what is our consumer entry point.” Over the October-long holiday period in China, senior management basically decided: Alibaba must have a native AI super entrance on the C side, not just sprinkle AI across existing products.

People inside the Qwen team describe the timing as “just right” for two reasons. First, they think Qwen3-Max is finally at the global frontier level, so you are not trying to sell a half-baked brain to consumers. Second, the agent ecosystem, both inside Alibaba and in the wider developer world, is mature enough that the model can actually do things, not just chat. They have spent a lot of time just on data plumbing and authorization so that Qwen can safely call Taobao, Amap, Alipay and so on when needed.

What it is reminding us is that it is not just a cloud-plus model company. It has Taobao for commerce, Alipay for payments, Ele.me for food delivery, plus travel, local services, and entertainment. If it can wire all of that into the Qwen app and give users the feeling that “I can ask for something and it just gets done,” that is a different type of consumer proposition.

The analogy I like here is:

Imagine if Amazon had a leading frontier model and also owned Amazon.com, DoorDash, and PayPal at the same time.

In that world, you could just tell the Amazon AI app “I am hungry, I want Thai food.” The AI could search restaurants, decide which one is likely to please you, place the order, apply the right coupons and complete payment. All inside one app, backed by one identity and wallet.

Inside Alibaba the mantra is “one AI native super entrance on C, plus every old business gets AI-ified,” so Qwen is not replacing Quark or 1688. It sits on top of them and then gets embedded back into them again, like a loop.

That is what Alibaba is hinting at with Qwen. Instead of AI living off to the side as a cute chatbot, it becomes the orchestration layer across Taobao, Alipay, Ele.me, and whatever else they plug in.

If this works, then the consumer AI story in China becomes less about “who has the best standalone chat app” and more about “which super app has the most powerful, most trusted AI concierge behind it.”

Alibaba’s play now is to let users access commerce, food delivery, payments, ride-hailing, and entertainment, all powered by its own frontier model plus its proprietary business data. Watch out, Tencent.

For years, the narrative was that Tencent had the unbeatable social ecosystem. But Alibaba has its own ecosystem too, just anchored in commerce, logistics, and payments, and now it also has a very credible frontier model and a consumer brand to match.

ByteDance: model plus entertainment OS

ByteDance sits in a slightly different spot.

On paper, it has both pieces. It is doubling down on models, and it has entertainment distribution that is second to none. Doubao is already integrated into Douyin and other products. Users can use AI directly for chat, creative writing, and code, but also inside their content creation flows.

The reality today is that ByteDance is still an entertainment-first business. Douyin and TikTok own the mindshare for short video and discovery. The life touchpoints are narrower. You are not paying your utilities bill inside Douyin. You are not doing most of your boring daily errands there.

Where ByteDance shines is in multimodal and creative use cases. It has the best R&D capabilities for video AI in the world, with real user data and a giant ad machine sitting behind it. AI in the ByteDance world is there to help you create better videos, scroll smarter feeds, and buy things in a more entertaining way.

ByteDance can absolutely carve out the “AI plus entertainment and creator stack” slot, and that is still a huge opportunity. But I am just not sure that it will become the one app you go to for everything in the way WeChat or a future Qwen super app could.

Given that it is the only non-listed company out of the three, it is harder to get information on the company’s exact moves but I am working on a strategy update piece.

Why startups will not own the mass consumer slot

This brings us back to the independent AI startups.

Startups like DeepSeek, Minimax, Moonshot, and others are critical in pushing model quality, efficiency, and open-weight innovation. They are the reason the pricing for LLMs in China is already under pressure. They are also where a lot of the interesting research is coming from.

But when we talk specifically about the mass 2C slot, they all run into the same three bottlenecks:

1. Money. Training and serving top-tier models is insanely expensive.

2. Distribution. They do not own a super app, a cloud, or a payment network.

3. Expectations. DeepSeek has conditioned the market to expect strong models that are cheap or free.

So even if they ship very good consumer apps, they still have to buy traffic from WeChat, Douyin, Taobao, and the app stores. They are always, in some sense, tenants on someone else’s land.

This is why I think in the long run they are more likely to be model suppliers, infra providers, or owners of specific vertical agent products, rather than the “one icon on your home screen” that dominates consumer AI in China.

And that icon, again, probably lives inside Tencent, Alibaba or ByteDance.

2026: the OpenAI slot is still up for grabs

In the US, the consumer AI story already feels settled in people’s minds. OpenAI is the default app you think of when someone says “AI.” Even if the competitive landscape shifts in the long run, it already has the brand, the distribution, and the cultural reference point. It’s reached its exit velocity when it hit 500 million MAU.

China is not there yet. There is no single app that owns the consumer AI relationship in the same way. DeepSeek blew open the model conversation. Doubao, Yuanbao, and a bunch of other apps have hundreds of millions of users. Tencent and ByteDance are both embedding AI deeper into their ecosystems. But no one is yet the obvious first stop for “I have a question, let me ask [Insert name]” across the entire population.

That gap is what makes 2026 interesting. The “OpenAI slot” in China is still up for grabs. It might not belong to a pure model company at all. It might go to whoever wires a good enough model into the densest graph of daily life touchpoints, but I think it’s really now just a matter of whether that is WeChat, Qwen or Doubao.

One Alibaba exec put it more bluntly - If you define a real AI super app as something with over 100 million DAU, he said it will show up very soon in China. “Not 2025, then the first half of 2026.” In other words, the OpenAI slot is still empty, but not for long.

Go to AI PROEM BIG TECH for more on Alibaba, ByteDance, and Tencent. Toodaloo~

When OpenAI rolled out its “app-within-app” model and the new Apps SDK in early October, a lightbulb went off for me: this looks a lot like WeChat’s Mini Programs.

To be super clear: I’m not saying OpenAI and WeChat share the same business model. I’m using WeChat as a mental model. In my head, ChatGPT is the surface where intent lives; OpenAI’s agents/apps are the Mini Programs—lightweight, in-flow capabilities the surface can discover, invoke, and compose.

Think about when you open up Chrome or Safari, and you can access the whole internet on your laptop/ edge. WeChat is like that browser entry point for you. It’s not just a social media tool. My point is that OpenAI could be the operating system for the future, and you will access the world of AI agents through OpenAI.

We will start by explaining how WeChat became an operating system, essentially with its payment functionality as a foundational infrastructure and its rollout of mini programs. Then, we dissect the business strategy for Tencent and the incentives for third-party app developers to develop a mini program and onboard onto WeChat. We then explain what OpenAI recently revealed, and then the most exciting bit- tie it all together, and showcase how WeChat’s Mini Program may serve as a roadmap for OpenAI’s future as a “super agent” commander.

Subscribe now

Quick Refresher on WeChat (for friends outside China)

If you haven’t lived in China or used WeChat, it’s easy to think of it as “WhatsApp with stickers.” But it’s much more than that.

WeChat has roughly ~1.4B monthly active users today, according to its latest earnings filings. Tencent doesn’t publish a DAU/MAU ratio, but it is estimated by analysts that over ~1.2 billion are DAUs. What matters for this piece is that ~90% of those users are Mini Program users, which tells you how sticky the platform already is. And in fact, it is said to have grown by “teens %” YoY in 25Q2.

Key takeaway here is scale and stickiness - we’ll come back to this later.

WeChat as “OS,” Payments As the Flywheel

WeChat, as you know, is not just a messenger application. In reality, it behaves like an operating surface, more browser/OS than chat app. And the payments rail is the secret sauce. And this isn’t just me saying this. Even as early as 2023, a16z Partner Connie Chan (my girl crush, omg) was saying this.

And the beauty of WeChat is that it has built a sophisticated payment rail that allows businesses to build on top of it. How this all got started was that WeChat launched a red envelope function that allowed users to send (Cantonese lai see/Mandarin hongbao) during CNY or special occasions as a token of good luck.

It all started as a game, but then the red envelope set off a flywheel, prompting people to think about using the balance they’d accumulated from collecting red envelopes to withdraw it from their bank accounts. Soon after that, the idea evolved into using the collected money for goods and services, and that is how the payment ecosystem was built up.

This is vital to the future of mini program proliferation because it means the transaction element was already in place when mini programs were introduced. Today, hundreds of millions of transactions occur through mini-programs each year.

Eventually, payments became the flywheel. Key theme here: irreplaceable and convenient underlying infrastructure foundation.

Share

Mini Programs - the enabler of businesses

What most of you may know is that WeChat is a “super app” that lets you do all sorts of things within it. But how does that work? It’s because of the Mini Programs.

Mini Programs launched on Jan 9, 2017, led by WeChat’s founding product head, the OG Allen Zhang (Zhang Xiaolong), as a WeChat team initiative. They run inside WeChat, and do not require a separate download, nor do they require an additional layer of verification to be plugged right into WeChat Pay, login, search, QR, and notifications. They’re runtime-specific to WeChat, not portable as standalone iOS/Android apps. (Nico Han just did a handy WeChat Mini Program primer for AI Proem readers if you want a quick tour.)

When Zhang was creating Mini Programs, most people didn’t get it (duh, visionaries are so cool). What he said was that “Mini Programs are a kind of app that doesn’t require downloading to be used, and they will realize the dream that apps can be accessible everywhere. Users can scan (search) to open the app and use it anywhere, anytime, without needing to download it. “

And soon after, outside China, a comparable product was launched called Android Instant Apps, but WeChat’s identity + payments + distribution made the format go mainstream in a way Google never quite did.

So when MiniPrograms were launched, we saw that it perfectly connected the moving parts: payment, Official Accounts (公众号, and Channels (视频号), and created an ecosystem that seamlessly connected three elements: social, discovery, and payment. This ecosystem was able to then onboard MNC brands, SMEs, personal brands, and more, because this meant that they were able to lean into the whole WeChat ecosystem while being able to have full control of their mini application’s interface, user experience, transaction, and customer service.

And a reminder again, that is all without asking you to install anything.

Only a year after its launch, despite initial skepticism, usage exploded. By then, it was said that there were ~580,000 Mini Programs live and ~170M DAU engaging with them (a huge awareness boost came from the viral “Jump Jump” mini-game late 2017).

And within 22 months of its launch, the ecosystem had passed 1 million Mini Programs with third-party trackers pegging ~200M active users at that point.

Now, for those who are familiar with Chinese super apps, you may say, payments and mini programs are not unique to WeChat. Alipay, the OG digital payment app, soon adopted the gamification element of money transfer as well and launched its own version of red envelopes.

Then the concept of “apps-within-apps” was adopted by Alipay as well, which now also offers embedded applications that let users order food and rides, scan for physical bike rentals, and charge batteries.

In many ways, both WeChat and Alipay now function like an operating system; it’s often a matter of preference in using a Tencent-camp product vs. an Alibaba-camp product (which is what prompted the whole antimonopoly internet crackdown in the first place).

The key difference is that WeChat has the crucial network and social element that makes it more powerful for businesses onboarded.

Subscribe now

The Business Strategy

So, if you are like me, you’re probably asking: “what is the incentive for Meituan or JD to launch a miniprogram?”

Well, to start, a hugeeee advantage is what I mentioned earlier, the social element.

, an absolute expert in China social media marketing, explained to me that when Mini Programs were first launched, “they were a bit of a flop.”

Allen Zhang wanted mini programs to be a tool, not something that would addict you (like many Western apps), so they need to be very value-driven, she added.

Although it’s not for every kind of business, for the F&B sector, it is an excellent use of it. You order through the mini program and collect loyalty points and more, without having to download or open a separate browser on your phone. It also works well for luxury brands that want to provide VIP services or do a special pop-up, and you need to scan the Mini Program to enter the pop-up and become part of their membership loyalty program.

And it seemed like there was an internal company vibe shift. Plotnick said, “Since the launch, it feels like WeChat has begun this evolution towards a much more utility-focused platform rather than ‘social media’ as we lump it.”

And businesses noticed and began incorporating Mini Programs into their business strategies. By 2019, transaction volume within Mini Programs became proof, as GMV (Gross Merchandise Value - Sales) hit RMB ~800 billion and annual transactions exceeded $400B USD by 2021 (YoY up ~70%).

Adoption went from experimental to inevitable in under two years. And once people understood the strength of the ecosystem, brands followed the GMV.

It eventually enabled many businesses. The most famous case is probably Pinduoduo, which used WeChat distribution and social mechanics early on, before it grew into a giant on its own app. The initial goal was to tap WeChat’s large user base and leverage its social features to drive traffic and sales through group buying. The business has since spun off into its own platform and is now a ~ $ 182 billion business listed on NASDAQ.

Business Incentives and Monetization

The lesson: distribution + payments + low friction can incubate real businesses.

But beyond that, it was these key factors that made it a no-brainer for businesses to launch Mini Programs:

1) Additional entry point for businesses: Mini Program adds an additional entry point to their businesses (which takes a lot of work to convince someone to download), which potentially means more users. But how that’s calculated towards their MAU depends on the independent company disclosure rules

2) Native transaction: In terms of monetization, Tencent does not take a cut of the transaction because the transaction is still done within the business’s platform. But you’ll see below how Tencent makes $ from additional transactions.

3) Social sharing amplifies purchasing desire: The ease of use and ease of sharing for the business is actually a huge value add for businesses. Launching a Mini Program lets you share the item link in WeChat Moments, individual chats, and even group chats with a single click. Its social element and commercial activity are more organically embedded in the user’s purchasing behavior. (the strength over Alipay)

4) Ecosystem: and lastly, it leans on the WeChat ecosystem, where embedding transaction links directly into Channels or Official Accounts all make it more seamless.

Meanwhile, for Tencent, it’s not a bad deal either. There are some less visible commercial activities happening. Let’s break down how Tencent makes money from Mini Programs in plain terms.

1. Advertising Revenue: Mini Programs and Mini Games can run banner, pop-up, and rewarded-video ads, and Tencent shares that revenue with developers (the split varies by format—roughly 50% for ordinary programs and up to about 70% for certain creative games under set thresholds).

2. Transaction Fees: Payment transactions run through WeChat Pay, and thus Tencent collects merchant processing fees (often cited around 0.6% domestically, with rates that vary by sector and cross-border use). The scale behind those fees is significant—industry trackers estimate Mini Programs handled over $400B of transactions in 2021 (about RMB 2.7T across roughly 3.7 million Mini Programs), firmly putting GMV in the “hundreds of billions of U.S. dollars” range and creating a flywheel where payments and ads reinforce each other.

3. Ad-stack Integration: Mini Programs are plugged into WeChat’s broader ad stack integration. Moments, Video Accounts, Search, and Official Accounts can be streamlined, so brands can buy across the surface and drive users straight into a Mini Program to convert.

Product Ethos

Another reason WeChat Mini Programs took off is their design. We’ve written before that the super apps in China often also meant UI/UX feels “messy” and that feel has bled into some AI apps, even, such as ByteDance’s Doubao, but not so much in DeepSeek or Kimi.

Just as Apple’s designer Jony Ive has said that product design should be about making things that work perfectly and are intuitive to use, emphasizing that “There is beauty when something works and it works intuitively.

Allen Zhang is just in this camp, and luckily, Tencent gave him the time and autonomy to embrace this philosophy rather than pushing him toward quick monetization with WeChat and later Mini Programs.

He famously said, “When you see friends, you dont wave around transactional sheets or ask to borrow money immediately; you make small talk, you bond, you build trust, and when emergencies arise, your friends will help you.” That’s the relationship WeChat should have with the users. In a way, taking a stab at those ad pop-ups and sponsored content forced onto your feesd.

It is reflected in the products’ design. When you open up WeChat, you’re not bombarded with advertisements, and you are only receiving direct messages from contacts you’ve accepted. When you use Mini Programs, you’re able to seamlessly go between your chat, payment function, moments (feed), and the application. You can also slide the mini program application to the side and access it more easily later with a single click.

And he always goes back to the core principle: focus on the user experience. That is core to WeChat’s operating system. Because WeChat isn’t trying to extract every dollar from users, the user experience is elegant and seamless. It is communications, payment, entertainment, and now an operating system, and that is something sometimes overlooked in the US, and it goes beyond the idea of just a “super app.”

The user-first, low-friction, less-aggressive monetization philosophy is also exemplified in Mini Programs. There is no pop-up ad assault, there are clean transitions between chat ↔ pay ↔ Moments ↔ Mini Programs, and one-tap return to the mini app you slid aside earlier.

Does any of this product design ethos remind you of the sleek ChatGPT interface? The not-in-your-face kind of $ strategy? I mean, you can choose to pay 200 bucks for premium service, or you can use the free model at this point. But what it is is, it’s not IN YOUR FACE.

Key takeaway here: Super agents?

What did OpenAI Ship?

So now, let’s go back to what OpenAI’s in-app strategy is.

OpenAI’s “app-within-app” push turns ChatGPT into a place where developers can build in-chat apps that show real UI, think forms, tables, and visualizations right inside the conversation.

Those apps can call external APIs, chain multi-step tasks, read and write the files you share, and keep state across a session. Permissions work like a modern mobile OS: the model asks for clear scopes such as “connect to your calendar,” “read this sheet”, and you grant or deny.

Distribution happens inside ChatGPT on web and mobile, with discovery via search, categories, and recommendations; on the enterprise side, admins can rein things in with allowlists. [Hello?!]

This shifts ChatGPT from “clever chatbot” to operating layer - essentially an operating system. So when you’re booking a flight, reconciling a ledger, or whatever other tasks you’ve asked the machine to do, it can all live in a single, continuous thread where the agent actually does the work. And that’s how STICKINESS IS CREATED.

This is also why its first platform attempt with plugins had to go. The first platform attempt was brittle because of prompt-heavy logic, inconsistent renderers, and a discovery surface that felt like rummaging through a junk drawer.

GPTs improved control, but you still had to prompt your way through business logic—basically building a backend out of vibes. Apps are the reset. It means less prompt alchemy, more deterministic tool-calling, and a much easier experience (fewer clicks and typing - God, we humans can be so lazy), and does this remind you of the user experience of WeChat? Seamless, easy to use, and user experience first.

And lastly, much like what we’ve been talking about until now with WeChat. OpenAI also wants to own the interface, and with it, distribution - the whole ecosystem you conduct your activities in.

When you own the entry point, you set discovery, monetization, and then the end activity - you own the whole experience. While some are saying this is the App Store playbook. I think it’s more than that. It’s actually the relationship between WeChat and Mini Programs.

As for the social element, maybe Soras-like products are the future - even though I honestly did not enjoy the product. Seems like the target user group is middle-aged finance bros - because my husband wouldn’t stop annoying me with videos of him scoring a three-point field goal or him riding a dragon flying on the moon.

Anyway… Let’s now hone in on the mental model and summarize the key points below.

How This Mental Model Predicts OpenAI’s Future Strategy

ChatGPT is becoming a “super-agent” entry point, a default front door on web, phone, or desktop that captures what you want (by text, voice, or even vision) and routes that intent to the right agent or app built on OpenAI’s stack.

Like WeChat’s single login that seamlessly covers Pay, Mini Programs, and content, a verified ChatGPT identity can become a similar trust layer for businesses and users as OpenAI fleshes out inbound and outbound payments and distribution.

Mini Programs live inside WeChat’s own runtime and use platform APIs. OpenAI’s version is an agent runtime inside ChatGPT, where third parties expose tools and actions that the agent can call so multi-step tasks feel native.

And just as WeChat’s AI search can take you from a query straight into a Mini Program, it can hover on your screen in the corner while you go back to your chat. OpenAI’s flow aims to take you from intent straight to the right tool, and you can do whatever you want in the background. At the same time, the agent carries on with the job.

Share

The “Super Agent Commander”

Taking this all to the next level.

I used to think model APIs are pure commodities. In reality, stickiness creeps in at the workflow level. Sure, you can swap the model for a single step, but building evals and re-verifying outcomes across many steps takes time and money. As more of those steps live inside a single surface (ChatGPT) and its agent runtime, the switching cost rises—very similar to how distribution + payments + identity made switching away from WeChat costly for merchants and devs.

Just as WeChat created an irreplaceable, easy-to-use foundation and ecosystem, OpenAI could be building a similar flywheel for its future strategy.

Think about China’s Android world: because Google Play wasn’t available, you had a maze of third-party app stores. Unlike Apple’s App Store, those Android stores often differed in what they carried, and even which version you got. Mini Programs cut through that fragmentation. Developers ship once into WeChat’s runtime, iterate fast without juggling multiple store uploads, and users get the same, up-to-date experience everywhere.

Now, OpenAI, given its lead in user adoption, will likely do the same for everyone. Instead of jumping between various LLM interfaces, it will be easier for users to go straight to the default platform to access the tools they need unless a need is incredibly niche or expertise-specific (which also explains why many LLM companies are doubling down on industry-vertical builds—but that’s another discussion).

A long post today as I’ve finally found some energy amidst my third-tri brain fog and was able to put my thoughts on paper. Hope you find it valuable/ entertaining.

Alibaba Earnings: AI, the Growth Propeller

A few days late to the Alibaba June Quarter 2025 Earnings, but I want to double-click on this. Alibaba’s stock price dipped as low as USD 60 two years ago, but it has since made a remarkable comeback. As I wrote on August 21, ​​the biggest cloud winner in China’s AI boom would be Alibaba Cloud. We have now seen it play out. Now let’s go into the details a bit more.

Big Picture - BABA

Alibaba’s financial results for the quarter ended June 30, 2025, were pretty darn good, beating expectations and mainly driven by AI-fueled growth and their “AI+Cloud” strategy.

Since Jack Ma’s return, he’s basically pardoned his young prodigy Jiang Fan, who was temporarily sidelined and pushed to Singapore to run the company’s international commerce business after a high-profile affair/ scandal that was made public right when the company was going through some major PR crises - notably being accused of having a horrendous ice-breaking culture. (Though I must say, while I worked there, I really didn't experience it, but it could be because I was working in the international department that mostly recruited from established MNCs/ investment banks/ international media and exhibited similar international standards of professionalism. However, the domestic/ more local scene could be quite different.)

Anywho, with Jack back, he’s also pulled quite a few of the 元老s - the OG co-founders and leaders that fought alongside him and that he had groomed back to the center of the power orbit, and the company is said to be reinvigorated with energy again after a decade of average performance and bureaucratic management compared to the newer disruptors like ByteDance, Meituan and PDD. From e-commerce to the recent food delivery war to AI and cloud, Alibaba is all in on combat mode. And honestly, it shows. [For details of the company’s financial performance, see here]

With that, during the earnings call, Eddie Wu, CEO of Alibaba Group, reiterated the company’s now-defined top-down strategy (and I dont blame you if you can’t keep up because there’s been so many restructurings and re-positionings over the last three years).

He said, "Looking ahead, we remain committed to investing in our two strategic pillars of consumption and AI + Cloud to capture historic opportunities and drive long-term growth.”

AI and Cloud

Zooming in on the AI vertical, well, it is a rich company’s game. During the software era, the highest cost for companies was brainpower and talent, and businesses can then scale with little additional cost (maybe spend on marketing and BD). However, with AI, despite crazy remuneration $$$ headlines, the cost of talent is negligible in comparison to the capital expenditure required. And who in China has more money to spend on infrastructure as well as talent? Well, the BAT, as we’ve reiterated time after time.

And who has been playing the enterprise strategy from the beginning, offering cloud services with added AI capabilities? Alibaba.

On Sept 1, Alibaba’s share price reflected its strong performance. Driving ADRs + 13%, which was mainly due to the stronger-than-expected performance in cloud revenue and record high quarterly capex despite the chip constraints that have stolen headlines over the last few months. The renewed narrative for Alibaba Cloud’s growth potential is solidified by its clear dominance as China’s largest cloud hyperscaler.

Key highlights:

• For the quarter ended June 30, 2025, revenue from Cloud Intelligence Group was RMB 33,398 million (~USD 4,662 million), an increase of ~26% year-over-year, compared to 18% in the previous quarter.

• AI-related revenue maintained triple-digit growth for the eighth consecutive quarter, driven by demand for compute, storage, public cloud services, and tech support.

• Strong training and inference demand across sectors and Alibaba Cloud’s continued strategic investments across model training, customer base, and 2B demand will maximize its AI cloud scale leadership.

• AI-related product revenue continued to maintain triple-digit year-over-year growth for the eighth consecutive quarter.

• Despite U.S.-chip (Nvidia) constraints, the adoption of a multi-chip approach, leveraging domestic options, has helped to alleviate earlier concerns about the impact of cloud capabilities.

• As for capex, the company will continue to play out its 3-year plan to invest RMB 380 billion on cloud and AI-related products (see China AI capex analysis)

• Qwen updates continued: Qwen3-4B-Thinking-2507 (Aug 6); Qwen-Image-Edit (Aug 19), multi-modal Wan2.2 AI Video Generation (July 28), Qwen3-Coder (July 22)

Beyond Cloud

Alibaba is pushing out a parallel narrative to its cloud story, which is becoming an AI + everyday consumption app leader, think ecommerce Taobao and Amap (Gaode Map - which a few years ago was pretty meh tbh but is definitely having a moment now) for navigation (which I previously called out in July how Gaode will have a huge advantage in wearables for Alibaba). [side rant, the whole thing with Gaode was that it was really good for driving directions in China since Google Maps does not work in China, but for a while Alibaba tried also to lure users to book rides within Gaode to compete with Didi and it was just a really shitty experience. They seem to have now found a new way to use this previously sidelined business as a clear advantage in its whole AI business ecosystem.]

The management said during the earnings call that they expect the market concentration in AI cloud to be significantly higher than traditional cloud, as developers require more technological support. Notably, Alibaba Cloud leads in not just traditional cloud computing, storage, databases, and big data, but also in what we have repeatedly emphasized: AI deployment needs, including AI infrastructure, AI LLMs, and open-source capabilities.

Public cloud demand is expected to grow as AI applications continue to find new use cases and deeper into vertical cases. Alibaba is ready to take a large share due to its full-stack capability and developer-friendly open ecosystem (EVERYTHING IS COMING TOGETHER).

Furthermore, Alibaba highlighted that developing AI agents will require models with much larger context windows so they can manage more complex tasks and longer reasoning chains. Since agents often need to orchestrate multiple tools, or even chains of tools, management stressed that this trend creates new opportunities for Alibaba to deepen its role as an infrastructure provider.

The relatively new product, AgentBay, is a new sandbox environment designed specifically for agent development. Executives noted that coding capabilities will become a major differentiator as the market evolves. Importantly, Alibaba has already embedded AI-native features into its core platforms—Taobao (e-commerce), DingTalk (enterprise collaboration), Amap (navigation), and Alipay (payments). These upgrades position its enterprise clients to develop and deploy AI agents directly into existing workflows, a strategy that management sees as a key lever for driving both adoption and monetization going forward.

Some additional industry thoughts/ brain farts

1. During the most recent BG2 podcast, Bill Gurley goes into depth sharing his thoughts on his recent trip to China (and how people gifting him or lending him various types of EVs for test driving). A lot of it is not new to keen China-tech analysts/ watchers but I think his key messages are put quite succinctly so worth a listen. He makes a point about talent, which is that the brightest Chinese business leaders are learning about the U.S., but the U.S. is not doing enough in reverse. And there’s a talent flow reversal that is happening right now as well, which is being overlooked. This is the whole point of AI Proem: to provide nuanced business analysis to the world and bridge that knowledge gap. Read my AI Proem manifesto if you must.

2. Dan Wang’s new book, Breakneck, is the talk of the town, from Bloomberg’s Odd Lot interview to Bill Gurley referencing the book during the episode mentioned above. I’ve been a big fan of his work since his blog posting days, and while he was an analyst at Gavekal Dragonomics. His analysis of the Chinese political economy is deeply nuanced, and I think that comes from his background, really.

   Or it might be that I’m biased because we share a similar upbringing, being raised in Canada to Mainland immigrant parents, and having worked/studied in Hong Kong/ Mainland and the U.S.

   I think being local but never local, and being bicultural/ bilingual, really makes you 1) confused at first when you’re young, but 2) able to really see things in a less singular way. And being raised in Canada is actually quite different from being raised in the U.S., where the identity by race is a lot less in-your-face. When I first moved to the U.S., I was shocked by how many people would introduce themselves as “XX hyphen American,” whereas in Canada, I don’t think anyone really asked me “where I was really from” growing up (or maybe that was just because it was too obvious?). The geopolitical influenced competition narrative is also not prominent in day-to-day conversations, so I never felt the two sides of my identity clashed or that I was forced to choose a loyalty side. Then working and studying in both “worlds” as “wanna-be intellectual adults” allowed me to be more open-minded and blend in locally in all the places mentioned above, given the privilege of sounding native in both languages, to see the dynamics from different perspectives.

   But aside from me projecting this similarity, which Dan doesn't know about because I’ve not spoken to him directly despite having a few mutuals, I think the framing that he proposes that is most relevant to our research here at AI Proem is his study, which suggests that China’s approach to industrial pursuit is also the foundation of its technology strategy. And its top-down economic planning model has laid the foundation for this generation of hardware, energy, and infrastructure prowess. Anyways, excited for my copy that is coming through with Amazon Prime in a few days.

3. AI video & audio editing (for an intro piece on Kling, see here)

   I wanted to save this surprise, but I have “zero chill”, as my brother likes to put it. The news is that I am working on a podcast, which will be launched in the next month, and I have a stellar guest lineup! (SO EXCITED - spoiler alert - one of my guests might have just been on Odd Lots as well, and another is essentially the voice of GenZ on AI culture in SV vs. China, **wink wink)

Anyway, as I tried out various recording, podcasting, and video editing tools, I was shocked by the sophistication of Riverside (which I had previously only used when being interviewed). However, this time, after paying for the premium version and operating from the backend, I’ve been genuinely impressed by its ability to AI edit audio and video, as well as the ease of use.

Throwing me back to years ago when I was working in broadcast, where we would spend hours just sifting through b-roll, and then hours transcribing and typing up subtitles, only to spend hours then editing out the exact time frames. The AI features in Riverside- just wow. For those who are looking for an easy-to-use editing tool for vlogs and podcasts, try out Riverside (I’m not being sponsored at all). Which resulted in a text to my husband that he won't stop making fun of me for - “OMG SHORT ADOBE.”

Despite taking a complete semester of courses at CJS and NYFA each on using Final Cut Pro, Audition, and various editing tools, I still struggle with editing and hate it. I think it was the biggest mental hurdle for me even to start a vlog/podcast over the years - so now that’s been resolved.

Expanding on the creative industry use of AI, there should be more interesting news coming out from Kuaishou’s Kling and ByteDance’s CapCut. I’m watching this space and agreed to attend some closed-door events hosted by Kuaishou at the end of the year, so I will keep y’all posted.

I want to learn more about real-life AI diffusion and deployment standards vs. academic/ technical measurements. Could someone please share some high-quality work on this topic?

If you have found value in my work, I would appreciate you sharing it with more people. It is word-of-mouth referrals like yours that help me grow.

And luckily, I was able to join a closed-door WeChat/ Weixin event for a few analysts in Hong Kong at the Tencent office just two days after WAIC concluded, so I combined some takeaways from that as well into this post.

Tencent Leadership’s Open-Mindedness to Open-Source

I think the most interesting point was hearing firsthand from Tencent employees on the decision-making behind the company's integration of DeepSeek into its Yuanbao/WeChat and its own LLM Hunyun in parallel. It was confirmed that there were no financial relationships between DeepSeek and Tencent, rather a high-level mutual admiration-driven decision by Pony Ma himself and Liang Wenfeng. I think that shows 肚量 (stomach size) - open-mindedness + acceptance.

Tencent’s leaders have always been more low-profile compared to other prominent tech executives in China. Its PR function is mostly embedded into each business group/ unit and is self-driven rather than having a top-down strategy. This is due to a few reasons: 1) a less demanding KPI for high-profile PR; 2) the company functions in a way that product leaders call the shots and own their P&L. Showcasing again its less-centralized power set up.

For those who missed it, see here from February: Why Tencent's Integration of DeepSeek Into Its AI App Is a BIG DEAL.

Many Chinese LLM labs are choosing open source/ weight, and that is because it allows them to work on top of each other like an open-weight remix that can accelerate improvements. Essentially, using one model to distill another, and in some ways, it might mean there will be less likely to be a monopolistic breakout winner in the long run. Still, for the big techs that are more trying to stay relevant and capture value in the distribution channels of their existing apps, it makes more sense to be open source than not. Letting the real frontier model startups do their thing.

In that case, Tencent has been the most 心大 “open-minded” to embrace not only open-source but also this philosophy above… so today, with that…

Share

LLM: Hunyuan

Leaning on its Hunyuan large model, Tencent has rolled out what they call a "AI Family Bucket" - like the KFC bucket you get it - product matrix, aiming to turn AI capabilities into accessible productivity tools. This strategy is anchored in Tencent's self-developed cloud technology and the Hunyuan large model.

Tencent said it is trying to cover a full-link technology architecture "from cloud to edge". Following Hunyuan Large and Hunyuan-A13B, Tencent's foundational models will further open-source more small and medium-sized models that also demonstrate excellent performance, which will likely further accelerate adoption by developers.

• The Hunyuan-A13B is the industry's first 13B-level MoE (Mixed-of-Experts) open-source hybrid inference model, has a total of 80B parameters but activates only 13B, achieving performance comparable to leading open-source models with similar architectures, while being faster in inference and more cost-effective. It excels in Agent tool invocation and long-text capabilities.

• The Hunyuan Turbo S is Tencent Hunyuan's new-generation flagship fast-thinking model, which has shown significant improvements in code and science capabilities, and complex instruction following, maintaining a leading position in global large model rankings. Its response speed has significantly improved, with first-token latency reduced by 44%.

• The Hunyuan T1 is its deep-thinking model, further enhancing inference capabilities and excels in project-level code generation, high-difficulty mathematical reasoning, and text writing.

A breakthrough is its Multi-Modal AI Model Matrix.

This 3D model, version 1.0, is the industry's first open-source 3D world generation model that allows immersive roaming, interaction, and simulation. It combines panoramic visual generation with layered 3D reconstruction technology.

The Hunyuan 3D World Model (First Release & Open Source) is its debut 3D world model, focusing on 3D scene generation for game developers and 3D enthusiasts. Said to support text-to-3D and image-to-3D world generation.

The 360-degree scene is an exportable mesh, which means the results can be edited and used in mainstream game engines/modeling software. Its interactivity and editability allow developers to foreground objects, and separate background terrain, ground, and sky for secondary editing. Its roamability means that users can move around the scene, not just view it from a static point.

Subscribe now

AI Infrastructure

Although not the biggest, Tencent Cloud usually ranks within the top three for cloud providers in China. Tencent Cloud's infrastructure spans 21 countries and regions globally, with 56 operational availability zones, over 1 million servers, and more than 3200 acceleration nodes. This forms a "technical backbone" supporting a wide range of AI applications. In May, we covered its CapEx projections to be RMB 27.5 billion (~USD 3.75 billion) for the first quarter, and compared to peers like ByteDance and Alibaba, it seemed quite tempered.

So far, we’ve seen the company play out its planned expansion in AI integration in 1) Agentic AI; 2) AI in WeChat; 3) AI boosting ad and gaming; and 4) Cloud, as it said it would during its 2025Q1 earnings.

The company has developed infrastructure, including high-performance HCC clusters, high-speed storage, and the Xingmai Network. HCC, designed for large model training, incorporates advanced technology and is used by 90% of the leading large model vendors in China.

AI Platforms

Similar to what Huawei has been doing with its “AI in a box” strategy. Tencent introduced five major AI productivity platforms to bridge the "last mile" of AI industrial applications.

• The Tencent Cloud Intelligent Agent Platform (ToB-focused) was upgraded from the large model knowledge engine, which helps enterprises build stable, secure, and business-compliant intelligent agents by integrating their exclusive knowledge and scenarios. It includes industry best practices, leading RAG algorithms, and supports canvas-style workflow orchestration. It offers various development frameworks like LLM+RAG, Multi-agent, and Workflow, with a one-stop toolchain for application configuration, debugging, evaluation, and release. The new upgrade brings comprehensive Multi-Agent capabilities, leading RAG capabilities (connecting to mainstream databases and Tencent Docs), and intelligent Workflow capabilities with global-view Agents.

• Tencent Yuanqi (元器) (ToC-focused): An all-in-one intelligent agent development platform that provides convenient intelligent agent building and distribution for C-end users, leveraging WeChat Official Accounts, Tencent Docs, and WeChat Pay MCP ecosystem advantages. Upgrades include stronger Official Account agents (Multi-agent mode, MCP integration, matrix account authorization), Tencent's unique ecosystem integration (Official Account articles as a knowledge base, Tencent Docs, WeChat Pay for monetization), and C-end traffic distribution via platforms like Tencent App Store.

• Lexiang Knowledge Base (乐享知识库) is an enterprise-grade knowledge management tool from Tencent, offering capabilities such as knowledge accumulation, updates, collaborative management, permission settings, AI Q&A, and effect insights. It helps over 300,000 enterprises build knowledge-based organizations.

Entering into Robotics

A significant announcement at the event was the Tencent Embodied AI Open Platform (TAIROS).

This is a purely software platform for the robotics field that Robotics X Lab officially launched. Tencent's overall strategy in robotics is to be a "titanium screw" – a partner to all robot manufacturers, rather than replacing them in hardware.

TAIROS includes embodied AI large models, cloud services, and supports rapid model verification through simulation environments. It provides three core model capabilities:

• Planning Large Model for complex task planning,

• Multi-modal Perception Large Model for understanding the environment and body state,

• Perception-Action Joint Large Model for connecting "seeing" to "doing" with safe and precise action commands.

The platform also covers the entire R&D chain through cloud services, enabling developers to perform simulations, training, and data management, while flexibly integrating hardware via SDK/API. It supports cross-body adaptation, modular support, and standardized products and emphasizes success rate and stability for real-world deployment.

Subscribe now

For context, it’s not entirely new that they’re working on robotics; it’s just being noticed by the public now. Tencent Robotics X Lab was established in 2018 and is led by Dr. Zhang Zhengyou, focusing on advanced robotics research and application, aiming to develop next-generation robots for human intelligence enhancement, physical augmentation, emotional care, and human-machine collaboration.

Robotics X Lab has developed various mobile robots (Max, Ollie), manipulation robots (TRX-Hand, TRX-Arm), and industrial robots. Their latest human-residential robot, "The Five" (小五), launched in 2024, integrates a quad-legged wheeled design, large-area tactile skin, multi-finger dexterous hands, and safe human-robot physical interaction for complex tasks in human environments, including assisting with elderly care.

AI for Consumers: Leaning into the “AI for everyone” theme

Tencent's "AI Family Bucket" has evolved to encompass various “Life All-rounder," "Workplace Secretary," "Travel Planner," and "Content Creator," thereby creating an "AI Friend Circle" for users. AI capabilities are integrated into social platforms, tool applications, and cultural entertainment products, extending to every aspect of user production, life, travel, learning, and entertainment scenarios.

• Yuanbao (元宝) is its proprietary chatbot tool that is the main interface for how WeChat users can access its Hunyuan LLM and DeepSeek. The AI assistant focuses on improving work and study efficiency. As mentioned, it supports Hunyuan and DeepSeek dual models for stable, smooth, and unlimited use. Yuanbao connects to Official Accounts and authoritative web sources, can read images, analyze files, create charts, and assist in logical reasoning, office work, knowledge learning, creative endeavors, and general life inquiries. It can be added as a WeChat friend, allowing users to interact with it like an AI friend. Yuanbao also offers AI video calls, utilizing the Hunyuan voice model for lower latency, more human-like, and emotional calls, with screen analysis capabilities.

• ima: An AI workbench product centered around a knowledge base, integrating Tencent Hunyuan and DeepSeek R1 models. It helps users collect and interpret information, engage in AI Q&A and interaction, and generate content, creating a personalized knowledge base. ima supports various formats (Markdown, PPT), offers large storage capacity (up to 30G free), folder management, and allows sharing knowledge bases with others, making it a "shared brain" for teams. It is available on PC, mobile, and mini-programs.

• QQ Browser (like a kind of outdated Chinese Safari): Upgraded to an AI browser with QBot, powered by Tencent Hunyuan and DeepSeek dual models. QBot offers five main functions: AI search, AI browsing, AI office, AI learning, and AI writing. It can also invoke Agents to perform complex tasks.

• Consumer-end, this is probably the most well-known function and a big game changer for Tencent. The WeChat integration of Yuanbao can be added as an exclusive AI friend. Hunyuan large model integrates into WeChat's social ecosystem, bringing rich content services and smart experiences, including natural language mini-program generation (allowing users to create mini-programs without coding), AI search (in "搜一搜 search bar, integrates Hunyuan and DeepSeek), and AI WeChat Reading (AI Q&A within books for themes, summaries, key plots, knowledge points). Btw, Zichen Wang wrote a deep dive on the importance of WeChat for the newsletter Pekingnology.

AI Agents

Since OpenAI’s o3 and then Manus’ AI agent rollout, essentially every AI company is trying to build a better and more user-friendly agent AI tool. Many are leaning into their existing industry know-how or relationships to make more tailor-made tools.

As I mentioned earlier, Alibaba owns a business called Gaode Map, which is one of the most downloaded/ preferred navigation tools in China for drivers. This gives it a significant advantage in its wearable glasses. Tencent actually has a similar business called Tencent Maps. It has also rolled out a series of relatively niche products that, frankly, probably won't be major revenue generators, but are nice to have.

• Leaning on that, Tencent has launched a Travel Planning Agent which offers a one-stop intelligent travel experience. It integrates travel guide generation, nearby search, attraction introductions, mini-program ordering, route planning, and editable itinerary saving. It's the first in the industry to link "map area + guide" interaction and support "save and re-edit" functionality for guides.

• Smart Cockpit Service Assistant Agent: Integrates AI Agent capabilities with the WeChat mini-program ecosystem to provide services like dining orders, calendar meetings, queue number booking, and flight/hotel reservations, becoming a trusted travel service assistant for drivers. It uses long-term memory to understand user habits and offers natural, human-like AI interaction.

• Learning Assistant Agent (QQ Browser QBot): Supports multi-modal input (text/image/voice) for problem-solving, explanation, key point extraction, 1-on-1 tutoring, and personalized learning plans.

• Health Management Agent (Tencent Health): A visualized AI agent built on the Hunyuan large model, acting as a "health sentinel". It interprets physical examination and lab reports, dynamically analyzes health status, identifies potential risks, and provides targeted health management plans. It can also track health indicators and act as an "action coach" for long-term health benefits.

• AI Gaokao Tong Agent (QQ Browser): The industry's first college entrance exam Agent. It leverages QQ Browser's 7 years of accumulated authoritative college entrance exam knowledge (nearly 3000 institutions, 1600+ majors). It helps students quickly get comprehensive information, intelligently generate personalized volunteer schemes (with three tiers), and provides a detailed report within 3-5 minutes. It also offers AI Q&A and AI essay tutoring for exam preparation.

AI for Enterprise

At the offline WeChat event in Hong Kong, I was quite intrigued by the enterprise AI functions that the team has rolled out. Because given its ubiquitous nature, many people use WeCom (微信企业) as the default tool for Zoom-like calls and a file sharing cloud platform that is offered by WeCom (partially why ByteDance’s Lark and Alibaba’s DingTalk did not do that great in capturing marketshare. There is no one clear winner like Slack for China.

Knowledge Worker/ Enterprise Productivity Tools

• Enterprise WeChat - Smart Tables: Features a new AI engine that revolutionizes traditional data processing. It introduces "AI Fields" with four breakthrough capabilities: AI Smart Classification (for customer grading, feedback categorization, risk tagging), AI Smart Summary (extracts core conclusions from long reports), AI Image Understanding (for patrol inspections, financial invoice info extraction), and AI Content Generation (for marketing copy and scripts).

• Tencent Meeting - AI Assistant Pro: An intelligent agent designed for meeting scenarios. It helps users quickly review and summarize past meeting information, supports real-time Q&A, analysis, and summary during meetings, and intelligently generates meeting minutes and post-meeting to-do lists. It also offers automatic captions, camera tracking, multi-person screen sharing, and smart noise reduction.

• Tencent Docs - AI Document Assistant: Fully integrated into Word, Excel, PPT, PDF, Smart Docs, forms, and mind maps. Users can use "one-sentence" commands for intelligent office work and efficient creation. It offers smart analysis, data visualization with AI algorithms, document summarization, Q&A, and outline generation. It also supports one-click PPT beautification and cross-category information processing, such as generating PPTs from mind maps or summarizing PPTs into Word documents. Tencent Docs serves over 200 million monthly active users and 1.2 million enterprises/organizations.

• Qidian Marketing Cloud Agent: Based on large model inference, multi-agent architecture, marketing knowledge base, and MCP technology, it automates tasks like audience extraction, persona analysis, product optimization, content generation, and activity review for marketing.

• TCDataAgent: A data analysis agent from Tencent Cloud, enabling users to use natural language for complex data queries, report generation, and trend analysis on structured and unstructured data. It ranked third globally and first in China on the BIRD-Bench NL2SQL leaderboard.

• A.I.G Agent (AI-Infra-Guard): A comprehensive, intelligent, easy-to-use, and lightweight AI infrastructure and MCP service security vulnerability scanning and automated Prompt security evaluation agent.

• Wukong Code Security Detection Agent: A high-efficiency, accurate, and intelligent automated code auditing, vulnerability mining, and fixing agent based on multi-agent collaboration.

• TWeTalkAgent (Voice Intelligent Agent): A software-hardware integrated AI dialogue solution for smart hardware, focusing on voice agents and real-time audio/video communication. It integrates Tencent ecosystem resources (WeChat calls, WeChat Pay, QQ Music) and is applied in AI companion toys, robots, smart wearables, and more.

• TWeSeeAgent (Visual Intelligent Agent): A new generation Visual Agent solution that transforms traditional video surveillance into an intelligent system with active perception, semantic understanding, smart interaction, and scenario closed-loop capabilities through "edge-cloud collaborative multi-modal perception + scenario-based decision engine".

• CodeBuddy Agent: Tencent's self-developed AI-assisted programming tool for developers, powered by Tencent Hunyuan + DeepSeek dual models. It offers AI technical Q&A, the new development intelligent agent Craft, code completion, unit testing, code diagnosis, and intelligent review. It supports team knowledge base management, custom intelligent agents, multi-model integration, and enterprise account integration. Tencent internal engineers use it for over 85% of their coding, reducing average coding time by over 40% and increasing AI-generated code by over 40%.

• Qimiao Digital Human (奇妙数字人): Tencent Advertising's "Marvelous Digital Human" (奇妙数字人) first appeared at WAIC. It's a commercial AI solution combining "Miaosi" (commercial content AIGC production) and "Miaobo" (AI livestreaming e-commerce) platforms. It offers over 3000 industry digital human images, supports 1:1 real-person customization, and can restore real voices. It has been applied in the catering, apparel, tourism, and education industries. It boosts material generation efficiency by 50%, per capita efficiency by 300%, and reduces livestreaming costs by over 90%.

Industry Verticals

Various other tools include digital flight simulator tools that were co-developed with China Southern Airlines, Game Visual generation using Hunyuan, which is said to be the first industrial-grade AIGC game content production engine, educational tools, and so on. Lastly, to note, beyond Tencent’s social media and gaming business units, what is often not as well known in the West is that it also has a big entertainment segment, which is categorized under the Digital Content category.

• Tencent Music Entertainment: Showcases AI song creation (generating songs from text or images), the world's first all-AI singer "AI Leehom" (developed with Wang Leehom's studio, featuring AI-generated voice, visuals, and derivatives), AI musician (converting popular songs into different instrument sounds like piano, guzheng), and AI singing (recording 30s of voice to "sing" any song in high quality).

  

• Dark Zone Breakout (FPS Game): Features F.A.C.U.L., the world's first FPS AI teammate that understands complex player commands and has scenario recognition ability. It can provide intelligence feedback, resource sharing, and tactical discussions.

• Holographic Cabin: Based on Tencent's audio/video algorithms, it converts photos into various artistic styles (anime, 3D, watercolor, illustration, clay, jade) with optimized human portrait effects.

• Smart Claw Machine: Uses Tencent Cloud Real-Time Communication (TRTC) for remote control, offering stable service, clear picture, and low end-to-end latency (under 300ms) for an immersive experience.

• AI Newsstand: Uses YouTu Lab's self-developed image-to-image algorithm to transform images into various styles (anime, 3D, watercolor, illustration, clay, jade).

Conclusion

Anyway, the list goes on, and it’s quite overwhelming. If you look at the list of tools, the most important ones I think to note are its WeChat integration tools and its WeComm enterprise tools.

Other tools are pretty niche, and as Tencent is known to all work in silo (product managers to product managers do not vibe/ share info), it’s often an internal competition. For example, the WeChat team is located in Guangdong, and the headquarters is in Shenzhen. Tencent Games and Tencent Media teams are based in separate cities/ offices.

It's not entirely apples-to-apples, but similar to Meta, where WhatsApp, Instagram, and Facebook operate independently with separate corporate and product teams, Tencent’s main businesses like Games, WeChat, and others also function independently. Then there is a group-level corporate team, such as the legal, IR, and PR teams, which consolidates the businesses under the same P&L for the listed company that is then overseen by the C-levels.

Thus, even people internally say that sometimes it feels like there is a lack of synergy or high-level strategy within Tencent that you are seeing implemented by Alibaba and ByteDance top-down. But what that translates to is that you can probably expect product managers to kill for resources and try to compete with each other within the firm, much like teammates within the same F1 team, to try to develop the next killer tool, much like Zhang Xiaolong did with WeChat in the OG days.

If you have found value in my work, I would appreciate it if you shared it with more people. Word-of-mouth referrals like yours help me expand my reach, so please feel free to tag me in your post so I can see/share it.

And reach out to me:

Email: aiproem@substack.com

New to AI Proem? Start here

LinkedIn: www.linkedin.com/in/gmzshao/

Instagram: graceshao_proem

Following years of covering the fashion and consumer sectors for leading publications, Yaling founded the consultancy ApertureChina, which helps global brands with their market research and PR strategies in China. Following the Yuan needs no introduction. It is a leading China Consumer newsletter on Substack.

We each bring our expertise and interests to this coverage, and will demystify RED 小红书 for you to understand its history, relevance, and future AI strategy.

Following the Yuan

Get on-the-ground 'color' in China beyond the earnings calls and market reports.

By Yaling Jiang

There have been rumors that RedNote 小红书 is revisiting the idea of an IPO, but this time for real. After two failed attempts to go public in the US in July 2021 and then a pivot to HK in December, and a challenging few years for the whole Chinese internet industry. Four years later, RedNote is said to be courting investors in Hong Kong amid the hottest IPO market in the city since 2021.

So first, why is RedNote relevant? In many ways, it’s become a weathervane for consumption trends and confidence across China’s top-tier cities. RedNote’s rise is the classic case of having the right idea at the right time—and with the right people.

Origins and Early Growth

Co-founders Mao Wenchao and Qu Fang, both from Wuhan, founded RedNote in 2013. Mao had just completed his Stanford MBA, while Qu was at Bertelsmann in Shanghai. During China’s early consumer upgrade, they saw a shared challenge for global Chinese shoppers: no one knew what to buy on overseas e-commerce sites or when traveling.

During the early wave of China’s consumption upgrade, they identified a common pain point among global Chinese consumers, including themselves: many people didn’t know what was worth buying on overseas e-commerce platforms or when they traveled abroad. Their solution was creating seven PDF shopping guides targeting popular destinations like Japan, Korea, Thailand, Hong Kong, Singapore, the USA, and France, geared mostly toward women and titled “Little Red Book.”

Thus originally, the app focused on building communities, spaces for people to meet others who were in the same physical cities or interested in the same topics. As it grew, it evolved.

What began as a niche need soon gained momentum, especially after 2015, when they introduced algorithms to their recommendation features. The hyper-accurate tagging is known to be ‘molecular’; it boosts users’ stickiness as well as efficiency in businesses, communities, and consumers finding each other.

Today, RedNote has evolved into a lifestyle guru and search engine for over 300 million users, with a gender split of 3:7 of male and female. And over 50% of users reside in first-tier and second-tier cities.

RedNote in the Global Spotlight

Many of you may have first heard of RedNote during the #TikTokRefugee wave, when American social media users, bracing for a potential TikTok ban, flocked to the app earlier this year. That migration helped it become the No. 1 app not just in the U.S., but in 80 markets worldwide.

It has often been regarded as a localized adaptation of Instagram in China. Its Chinese name, Xiaohongshu, technically translates to "Little Red Book" in English, but is usually referred to as “RedNote”. The app is more than Instagram, and we’ll explain why below.

Today, RedNote is a global platform that caters to a diverse audience, comprising China-based users, overseas Chinese speakers, and English-speaking audiences.

Business-oriented communities and brands see it as a highly valuable channel to efficiently lure in customers with relatively high spending power. This has been largely tested during Covid as RedNote-native businesses were born ahead of certain lifestyle trends, and saw exponential growth. One standout example is outdoor experience brand Dare Glamping, which sourced 60% of orders from the platform in 2021 and secured an angel round of RMB 10 million in the same year— a time when financing in the consumer sector dried up. Of course, whether startups can sustain that momentum by using RedNote’s taste-maker, trend-amplifying capability is another story.

But despite the hype, RedNote hasn’t had much luck in e-commerce for a long time. It launched its e-commerce store “Welfare Community (福利社)” within the platform as early as 2014, but in 2017, a new import tax policy cast a shadow over cross-border e-commerce, and the platform took a major hit. Perhaps partially inspired by the RedNote-native businesses’ success during COVID, RedNote launched an outdoor e-commerce platform, "Little Oasis （小绿洲)” in early 2022. RedNote closed both of these unprofitable, asset-heavy e-commerce stores in late 2023, yielding space for third-party sellers instead.

Subscribe now

It is said that recently, it widely known in the e-commerce industry and among consumers that RedNote is no place to shop. Consumers Yaling has spoke to complain about poor customer service, and what they often do is sift through product reviews on it before heading to their familiar e-commerce platforms to shop.

The B2K2C Business Model

Anyway, going back to the original business model. RedNote was created to provide tips and shopping guidance for Chinese consumers traveling abroad or who want to purchase foreign luxury brands. But it has made a few slight pivots and now evolved into this ubiquitous app. Its uniqueness lies in combining social media, content, and e-commerce.

According to a leading Chinese tech news publication, as of early 2022, RedNote reported over 200 million MAUs, with over 43 million content creators, and amongst them, 72% were born after 1990. More than 50% live in what are called first-tier and second-tier cities (basically, think of the crazy cosmopolises with a population of around 20 million, such as Beijing, Shanghai, Shenzhen, Chongqing, Chengdu, and so on).

So what does B2K2C mean? It’s an abbreviation for “Business to Key Opinion Consumer to Customer,” so businesses often sponsor and pay creators that then sell to consumers, in some ways, similar to how KOLs operate on Instagram now, but it’s more institutionalized. RedNote offers tools and functions for creators to directly list themselves as taking sponsors, and brands can directly reach out to creators, and the platform even pushes out notifications to creators and small-scale competitions to incentivize creators to take on sponsorship and help the two parties matchmake.

In fact, RedNote has multiple revenue streams:

1. Advertising is its main revenue driver, making up 80%-90% of its revenue, based on industry estimates. The platform helps push out advertisements, search ads, as mentioned, brand and content creator matching to maximize brand exposure to users. It uses CPM/ CPC to charge brand fees.

2. E-commerce is another relatively large revenue stream as the platform charges commission, technical support, logistics service, and other value-added services charges. Eventhough it’s not been doing amazing, it’s still a major revenue driver.

3. Consumer data service: The platform also offers low-key reports and trend analyses to brands to better help them understand consumer behavior and trends.

Now, with this kind of business model, there come certain risks and issues. First is the KOL business that some content creators are doing what is called “grass-planting” 种草, but the recommendations are sometimes no longer genuine, rather driven by commercial gains. Then there is the issue with RedNote heavily relying on brand sponsorship and marketing, and thus again making users question the authenticity of some of the content, and lastly, the purchase-conversion rate is still relatively low, given that people do not go on the application with a clear intent to shop unlike blatant ecommerce platforms such as Alibaba’s Taobao.

However, despite these business model challenges, RedNote is charging ahead and quickly integrating AI into its strategic long-term business model.

Similar to what Grace has written about ByteDance’s TikTok, there has long been a legacy of using machine learning to better recommend quality content to users, creating a positive process of “better content → more exposure → more traffic → more creators.”

RedNote’s AI Strategy

Now, after understanding who RedNote is, what it is about, we zoom in on its AI strategy. In 2024, the company poached Dai Lidan, a former capital at a leading VC firm in China, and appointed her the Chief Strategy Officer to oversee the company’s strategic (cough* AI) developments. A Peking University computer science major who was a product manager by training, obtained her MBA at Harvard Business School, and then made a pivot into investment and was made a partner at Today Capital in 2022. Her experiences make sense to have led her to this role to oversee strategic M&A and new product innovation.

So this is relevant because she is said to be reporting directly to CFO Zhang Ziqi, who is known to have led investments for RedNote into LLM companies such as MiniMax and (attempted investments) into Moonshot. RedNote is seemingly taking a more active role as an investor in the space than actually integrating AI into its core functionalities.

As mentioned, RedNote had been an active investor in the AI space, but little do people know they actually also have a research lab called “hi lab” (Humane Intelligence Lab) that is focused on ensuring models are better trained on image and text on the platform.

In fact, it’s focus has been outlined as three priorities: 1) implement multimodal technology in internal scenarios and leverage its vast database of image and videos; 2) help creators reduce cost in production and boost efficiency with AIGC tools - which they’ve aggressively rolled out across RedNote; 3) priotize cost control and ensure strategic M&A in the space serves their long-term goals.

As mentioned earlier, RedNote started off as a community-building tool, not an intentional platform, but on July 8, it changed its tagline from “your lifestyle guide (你的生活指南)” to “your life interest community (你的生活兴趣社区).” It has broadened its ambitions and reach through this. To us, it shows how they strive to return to their core value and have stopped wasting energy on things they aren’t good at. They have come to terms with the fact that they do not have to copy other tech or e-commerce companies in the latest e-commerce or AI push; instead, bringing together people and communities has been their true strength, something no one else can compete with.

From that angle, it’s clear why RedNote has not gone all in on AI like some of its peers. As a pre-IPO company that only turned a profit in 2023, they cannot afford to follow suit with big techs because every investment needs to focus on things that show potential for future investors. In addition, they do not have the know-how at the top level to make the next viral AI product, so the AI products and initiatives we see are enhancing their existing community-driven purpose.

The collaboration with Fudan University’s philosophy department is such a classic RedNote move. It demonstrates their ambition to prioritize humanity and shape AI ethics; it’s their answer to the discourse of AI replacing humans, which we’d call staying true to their original goals and playing to their strengths.

1. Humanities-First Approach

Compared to other e-commerce behemoths (although RedNote isn't really that comparable to them yet) like Alibaba or even ByteDance, RedNote has been relatively conservative and low-key in its large language model push.

Since the inception of “hi lab” it has been heavily recruiting researchers and academics in the humanities fields to collaborate with AI algorithm engineers in post-training AI models. The ultimate goal is to enhance the AI capabilities of humanity through its social sciences literacy and behavioral consistency, according to an exclusive interview by leading Chinese financial news outlet Yicai.

2. AIGC Tools

The thing is, AI-generated content in many ways disrupts RedNote’s core offering - content, so they have to do it carefully. They’ve introduced a series of smaller functions that enhance content creation but do not replace human creation.

1. Davinci was launched in late 2023, which was an AI chatbot built based on MiniMax’s foundation model. But the chatbot did not perform very well, for the purposes the company hoped it would, such as providing food recommendations, life tips, or being able to answer questions on topics like geography and such. Thus, this feature has not really received much notice beyond content creators within the RedNote ecosystem.

2. Trik was then launched, offering a Chinese-style painting tool that was launched by their team. It can also help content creators generate cover images and doodle-to-image features.

   

3. Diandian (点点) was then released in early 2025. RedNote launched a chatbot agent on the platform that, if you “@” diandian, it will then reply to your queries similar to how Grok’s chatbot works on the platform.

Keeping it Human-first

Taking a step back, we have to acknowledge that, end of the day, RedNote is nowhere close to the size or influence of one of the BATs yet. However, it is carving out its unique playbook, adopting AI and paving the way for what social+commerce+AIGC may look like in the future.

On the occasion of RedNote’s 10-year anniversary in 2023, co-founder Mao outlined the three principles that guide the company’s future: “Development first, human-centered, small is big (发展第一、以人为本、小即是大).” He pointed out that the origin of Chinese business lies in the mom-and-pop shop, and every great brand starts with a single good product. Reiterating that they’re not here to be the “biggest and fastest”, rather, humans are at the core of the RedNote experience.

Yaling says she loves that perspective because that is what she largely advocates for: understanding China through its consumers. The human-centered approach is so…not China. Look at Chinese tech and e-commerce companies that are deemed successful, such as PDD, Shein, and JD.com. In terms of business model or work culture, they foster a dog-eat-dog mentality, which partially shapes societal attitudes.

We are also pleased to see RedNote flourish, as it reminds us that the humanities have a place in the business world. It’s also a personal reminder to stay true to the core values we bring as humans.

This has been a long time in the making. It is part of a series of in-depth examinations of Chinese big tech firms and their AI strategies, beginning with Alibaba, followed by Huawei (parts 1 and 2), and a few shorter pieces on Tencent. Without further ado, our protagonist today is ByteDance.

In January, we wrote that Bytedance’s AI application Doubao showed that inference demand could indeed go up 1 billion times. Since the release of DeepSeek’s R1 model, the global spotlight has been on the startup lab, and ByteDance seemingly took a quieter approach compared to its aggressive consumer conversion attempts before this moment. First-time AI users swarmed to Tencent’s WeChat as it integrated DeepSeek and its own Yuanbao LLM into its ubiquitous social media platform. The functional adjacency and existing 1+ billion users helped propel the consumer usage of Tencent’s platform to the lead quickly. But that lead remains in text models, with ByteDance offering video and multimodal AI, it is quickly catching up in consumer usage.

However, ByteDance is playing a long-term game, and its edge stems from a decade of recommendation-system expertise, now channelled into a fast-iterating LLM family (Doubao) and an enterprise cloud layer (Volcano Engine).

ByteDance’s legacy of algorithmic advantage

Before Doubao, ByteDance had already widely adopted AI. But it’s not what you think. ByteDance’s short-video apps TikTok and Douyin, as well as its news aggregator Jinritoutiao, leverage a sophisticated recommendation algorithm. At its core, it utilizes machine learning technology to best personalize the "For You" page (FYP) for each user. What it does is it analyzes user behavior and content characteristics to predict and deliver content likely to be engaging. And this algorithm is the secret sauce to its success, outcompeting other social media platforms like Shorts, Threads, and many other products that tried to emulate that success.

And because of this special edge, the company has reportedly stated that it would rather shut down the app in the U.S. than sell it, even under intense scrutiny and pressure from the U.S. government over the past few years.

What's so special about TikTok's algorithm? | REUTERS

As it has remained a non-publicly listed company (despite some of that not out of choice), it has been able to enjoy the anonymity, or at least been able to avoid the pressure to disclose certain business and financial details of the company. With that comes a perk, where they can build in stealth.

Subscribe now

The Evolution of ByteDance’s AI Research

While most of the Chinese research labs started really working on LLMs starting in the late 2010s to the early 2020s, it was said that ByteDance didn’t really prioritize this “business line” until the ChatGPT explosion - and after the CEO Liang Rubo held a big meeting for the company management expressing disappointment in the leaders lack of vision and innovation. With Alibaba releasing its Qwen models on open-source platforms and Tencent integrating DeepSeek into its existing platforms, ByteDance seemed somewhat sluggish externally in this regard. However, they have actually built up a formidable team in the quiet.

Many say that ByteDance founder Zhang Yiming’s involvement in AI research at ByteDance was similar to the role Sergey Brin played at Google. Despite having kept a low profile since 2021 (the Chinese internet crackdown) and passing over the reins to the current CEO Liang Rubo, Zhang remained active in the actual business decisions, similar to Jack Ma/ Joe Tsai’s roles with Alibaba. Similar to Alibaba, where the top leadership baton was passed to long-term, trusted, and loyalists, Liang is the co-founder and former university roommate of Zhang. The founders, and many would say the real visionaries, are the anchors and leaders steering the internet behemoths amid AI disruptions and new innovation.

Doubao, a Side Project Turned Flag-Bearer

As mentioned, the AI lab was launched well before Doubao, as early as documented in the mid-2010s, and has undergone a few name changes and priority shifts. It started off mostly focused on natural language processing, and LLM research wasn't truly prioritized until ChatGPT’s 2022 moment. Soon following the release of OpenAI’s consumer application, Chinese big tech and the original four tigers one by one released their versions of chatbots.

In 2023, Doubao was launched, and a team focused on large language models quickly rose in importance internally. The LLM and the consumer app are both named Doubao, which means Bean-Paste Bun (a sweet bao). It now has over 60 million monthly active users, though lagging behind DeepSeek, it’s far exceeding most domestic rivals.

According to recent data, Doubao’s paid token traffic rose 22× in twelve months while pricing fell 60%, which obviously is not welcome to all users. The flagship model, Doubao 1.6, was recently released and now supports 256k-token context windows and multimodal reasoning. Its video model Seedance overtook Google’s Veo on the Artificial Analysis leaderboard, which was a bit of a shocker for many.

Talent-first approach

How did Doubao come about? Let’s start looking at Seed. This team was created in early 2023 as an umbrella research group to focus on frontier AI research work, which covers NLP, CV, and Speech Lab, and the team is spread across the headquarters in Beijing, as well as Singapore and Palo Alto. It heavily recruits from PhD programs through what is internally referred to as the “Top Seed” program. Though it mirrors efforts by Alibaba and Tencent, it is less discussed externally, and it is said that there are probably a couple of thousand people part of this program, which is more sizable than many expect.

The strategy behind Seed is to build a strong team that can pivot when needed and is focused on the most frontier model research. The team lead joined Seed after 17 years at Google (although news literally came out today that he potentially has been demoted/ let go for personal reasons - misconduct in the workplace scandal - not confirmed yet). It’s ByteDance’s solution to Alibaba’s Damo Academy and Tencent’s Tencent Research Institute.

Actually, before Seed, it was said that an AI lab was started in the mid-2010s, and I found ’s article here very helpful in providing more context. As he wrote, by mid-2023, ByteDance released its first LLM called Skylark (which was later rebranded to the same name as its LLM model, Doubao). Still, its GPT-like bot Doubao was received with little interest, only reaching 1.3 million DAUs by November, compared to ChatGPT’s 300 million weekly users already by then.

https://x.com/deedydas/status/1936990542669459552

Share

Doubao: The Flagship AI Product

Doubao was actually first launched as Skylark in 2023, designed for tasks like text generation, image creation, speech synthesis, and reasoning. The Doubao large model has evolved from a single language model into a comprehensive model family system, covering multimodal and multi-scenario applications, and has become one of the most successful cases of large model commercialization in China. The technology enables e-commerce sellers to create studio-quality promos in minutes, significantly reducing both time and economic costs. The technological iterations and market expansion of the Doubao model family fully demonstrate ByteDance's "technology-driven product" development logic and rapid market responsiveness.

In terms of technological evolution, the Doubao large model has undergone several major upgrades.

In May 2024, ByteDance first released the Doubao model family. Subsequently, in third-party FlagEval evaluations, the Doubao-Pro-4k model ranked second in the "objective evaluation" of closed-source large models with a comprehensive score of 75.96, trailing only GPT-4 and becoming the highest-rated domestic large model; it also secured second place in "subjective evaluations."

By August 2024, Doubao had significantly upgraded its visual and speech capabilities, with its text-to-image model gaining a better grasp of "Chinese aesthetics" and its TTS (text-to-speech) achieving precise emotional expression. By 2025, the Doubao model had iterated to version 1.6, featuring enhanced reasoning, multimodal understanding, GUI operation, and front-end programming capabilities.

The technical architecture of the Doubao model family is mainly characterized by three key aspects: 1) Full-modality support, 2) vertical scenario optimization, and so-called 3) Chinese cultural reinforcement.

1. For full-modality support, Doubao covers not only basic modalities like text, images, and speech but also extends to complex ones such as video and 3D.

• For instance, Doubao's video generation model supports multi-shot narrative and multi-task video generation, producing smooth, detail-rich, and cinematic-quality 1080p HD videos.

2. For vertical scenario optimization, Doubao offers specialized models tailored to different applications, such as the UI-TARS model for graphical interface interaction, vector retrieval models, and role-playing creative models.

3. In terms of Chinese cultural reinforcement, Doubao's text-to-image model demonstrates precise comprehension of Chinese elements—including figures, objects, dynasties, geography, cuisine, and festivals—enabling high-quality generation of culturally authentic images.

The main publicly available models include:

• Doubao 1.6: The flagship next-gen model with comprehensive upgrades, excelling in knowledge, coding, and reasoning.

• Doubao 1.5 Lite: A lightweight language model matching or surpassing GPT-4 Omni and Claude 3.5 Haiku in benchmarks like MMLU_pro (comprehensive), BBH (reasoning), MATH, and GPQA (professional knowledge).

• Doubao VideoGen (Seedance 1.0): A multi-shot, multi-task video generation model.

• Doubao Vision: A multimodal model with advanced visual understanding and reasoning.

• Doubao Real-Time TTS: Generates ultra-natural, high-fidelity, personalized speech by predicting text emotion and tone.

• Doubao Voice Clone: Achieves 1:1 voice replication in 5 seconds, supporting cross-language migration.

• Doubao Text-to-Image/Image-to-Image: Specializes in Chinese cultural elements, offering 50+ style transformations and creative extensions like outpainting/repainting.

  Share

  ---

Zooming in on the three models that were recently released in June 2025:

(1) Doubao-seed-1.6 is an All-in-One model that supports 256k contexts in deep thinking and multi-modal capabilities. Developers can use automatic thinking mode selection and achieve token savings.

(2) Doubao-seed-1.6-thinking supports 256K long contexts, deep thinking in coding, mathematics, and logic.

(3) Doubao-seed-1.6-flash supports deep thinking, multi-modal, 256k contexts with low latency time. Its visual capabilities are comparable to global peers in the market.

Volcano Engine: The Enterprise Platform

Volcano Engine is ByteDance's cloud computing and AI service platform, launched in 2021 to provide enterprise clients with advanced technologies like recommendation algorithms, data analytics, and artificial intelligence solutions. It focuses on enabling businesses with tools for intelligent applications, visual and data intelligence systems, and multimedia technologies, drawing on ByteDance’s expertise from apps like TikTok and Douyin. Unlike traditional enterprise solutions (e.g., CRM or ERP), Volcano Engine emphasizes AI-driven capabilities, including large language models, cloud infrastructure, and multimodal AI services, to support industries such as e-commerce, automotive, and entertainment.

It serves as the primary platform through which Doubao’s capabilities are commercialized and delivered to external enterprise clients. Doubao models, including versions like Doubao Pro and Doubao Lite, are integrated into Volcano Engine’s ecosystem, enabling businesses to leverage these LLMs for applications such as AI chatbots, content creation, and smart cockpits. It also supports Doubao’s infrastructure, offering model fine-tuning, inference, and evaluation through its Volcano Ark platform, and has initiatives like providing trillions of free tokens for academic research. This synergy allows Doubao’s AI advancements to scale across industries via Volcano Engine’s cloud services.

Commercialization and Market Impact

As we’ve written before, Alibaba has taken on a roadmap more similar to American big tech, offering infrastructure support and AI as SaaS, while Tencent is taking a full-on consumer-facing approach through its existing 1+ billion active DAUs across its applications.

According to the latest QuestMobile, Q1 2025 saw Doubao reach 110M MAUs (a pretty drastic improvement from last year), ranking second among AI apps (after DeepSeek's 190M).

By April 2025, Doubao surpassed DeepSeek to claim second place on Apple's App Store free chart (behind ByteDance's short-drama app Hongguo), reflecting strong user growth momentum, partly due to traffic funneling from ByteDance’s super-apps like Douyin, which integrated Doubao’s AI features and promoted its contact interface in user message lists.

For commercial applications, Doubao has penetrated diverse industries. Data from Volcano Engine shows average enterprise token usage grew 22x since its May 2024 launch, highlighting rapid adoption. Key use cases include:

• Social/Entertainment: Role-playing, interactive storytelling, chat assistance.

• Education: Personalized tutoring, Q&A.

• Customer Service/Sales: AI agents, sales pitch generation.

• AI Search: Semantic understanding, precise answers.

• Marketing: Ad copywriting, content creation.

• Hardware Assistants: Device interaction, voice control.

Anyway, daily usage on Doubao has exploded since May 2024, like I mentioned, even though temporarily it felt a bit irrelevant when DeepSeek was rolled out and Moonshot’s Kimi had its 15 minutes of fame. The platform now serves more than 16 trillion tokens a day—roughly 137 times its year-ago throughput. ByteDance cut prices just as aggressively, dropping input and output rates to RMB 0.8 and RMB 8 per million tokens, respectively. That makes the current 0–32 k-context tier about 63 percent cheaper than both Doubao 1.5 and DeepSeek R1, signalling a deliberate “price down, volume up” land-grab.

The product itself has evolved in tandem with it. Doubao’s new DeepResearch workflow allows the model to think, search, and stitch together answers autonomously, turning tasks that once took analysts days into 5-to-30-minute tasks to complete.

A multimodal backbone expands the addressable market, as e-commerce sellers can now verify merchandise images in real-time, autonomous-driving teams can label sensor data at scale, and security customers gain faster anomaly detection. The front-end has been shown to be equally ambitious as a graphical interface orchestrates agents, web browsers, and off-the-shelf software so smoothly that booking a hotel room becomes a one-step process.

On the video side, Seedance 1.0 Pro raises the bar for generative quality. Its seamless multi-camera narrative, blended actions, and natural camera movement sit at the top of the third-party leaderboard, Artificial Analysis. What people are highlighting is also that the pricing is disruptive. At RMB 0.015 per thousand tokens, a five-second 1080p clip costs roughly ~RMB 3.7. A marketing team with an RMB 10,000 budget can crank out about 2,700 high-definition videos on Pro, or close to 9,700 clips on the lighter Seedance 1.0 Lite. And its no secret that ByteDance is already pitching the model to e-commerce studios, film-and-TV producers, and game developers. (whether that is good for copyright issues is another discussion)

Audio is moving just as quickly. Real-time speech models enable live singing, voice impersonation, and performance effects, while the new podcast engine can generate entire shows from a prompt, a web page, or even a Word document that can be completed with natural back-and-forth dialogue between virtual hosts.

All of this sits on top of a growing toolkit for third-party developers. ByteDance now offers a dozen “AI-native” services that span the build pipeline: an upgraded IDE (TRAE), a model-control panel (MCP), PromptPilot for rapid prototyping, the open-source learning framework veRL, and an internal knowledge-management layer. Big-data specialists utilize multimodal labeling tools and automated data agents; infrastructure teams can leverage AgentKit, TrainingKit, and ServingKit. The company has said that it has built in security that leads to encrypted computing modules and an AI-aware firewall.

The common thread is ByteDance’s conviction that enterprise users will soon rely on agents to complete complex, end-to-end workflows, and that those agents must be cheap, multimodal, and natively integrated with the wider cloud stack to be adopted.

Competitive Landscape and Hybrid Approach

From Tencent’s WeChat to Baidu, from Meituan to Xiaohongshu, major Chinese internet apps have integrated DeepSeek one after another. Yet Doubao remains committed to its self-developed approach. But Doubao’s efforts were not announced until DeepSeek; it was, in fact, the number one used AI app in China, and in the latest estimates, it was said that the company spent over 1 billion USD (80 billion RMB) in marketing spend to convert users.

But after looking at ByteDance’s AI strategy holistically, we can see that it is taking a hybrid approach. Similar to the internet era, there were only 1-2 apps that were widely known by people, and many highly popular applications were never marketed as a ByteDance product, such as CapCut for video editing, Lark, the wanna-be China Slack, or now Coze, the no-code agent builder. It’s all a product-led strategy.

Ultimately, it’s the Doubao large model family that serves as the core product vehicle of ByteDance's AI strategy. Interest and expectations for ByteDance are high. And we can see that, at last week’s product launch, where ByteDance introduced its upgraded and updated Doubao LLM 1.6, it was absolutely packed. According to this WeChat 虎嗅 (Tiger Sniff) account, people lined up well out of the conference halls. Its strength is in its high-quality, engaging products.

To sum it up, you can think about it like this. Seed supplies the science, Doubao supplies the models and user scale, and Volcano Engine supplies the monetization and enterprise credibility. With the new video generation and deep-thinking agents, ByteDance is pushing ahead of its leadership in Chinese AI.

Even looking at Doubao, its success has primarily been rooted in ByteDance’s powerful ecosystem. Don’t forget how Meta and Google leverage their access to walled-garden data. ByteDance has something similar. It’s not just about the quantity of data but more about the quality of data, especially as we start shifting focus to video-generation models. It’s about the labeling of data, of where it was shot, how it was shot, what equipment was used to shoot the video, and, of course, a detailed description of the object being shot. All of these quality data points enhance the data used to train video models. And who has more video data than ByteDance?

Today’s post is an update snapshot of China’s 2B and 2C AI adoption and where the major players are all at with their AI game. Given that Tencent and Alibaba just announced their earnings last week, I will give them a little special attention, and then we’ll take a look at how the BBATs are all doing on the AI front.

Subscribe now

Where’s the CAPEX?

While Tencent highlighted “Agent AI” as its next race track, as written in its earnings coverage, it aims not only to build general-purpose intelligent assistants, such as Yuanbao and Ima, but also to create “unique-type Agents” that are deeply integrated into the WeChat ecosystem. We’re seeing more and more that the existing ecosystem control is critical to the adoption of AI at scale, similar to Google and Microsoft (which we’ll explore in another piece).

Leveraging its exclusive network of social relationships, the mini-program system, walled-garden content, and chatbox interface will really propel AI adoption at speed at later stages when technology and interaction become more seamless. On the other hand, deploying “Game AI” in large-scale competitive games remains its biggest opportunity at the current stage.

That largely sums up what we wrote about last week - the so-called “consumer winning formula”.

But hold up, just a day after Tencent announced its earnings, its arch nemesis, Alibaba, also announced its 2025 March Quarter earnings results.

One noteworthy thing is that Tencent seems to be clocking in ~some~ AI ROI, or at least has a more straightforward strategy on that front. Meanwhile, its arch nemesis, Alibaba, while seeing high double-digit growth in its cloud business this quarter, was largely expected in many ways given its positioning as Asia’s leading cloud provider. However, it has still not given a clear AI go-to-market strategy…?

The biggest puzzle this quarter is that both companies' capital expenditures have decreased significantly quarter over quarter. Has that loud tout of $$ seemed to have been hushed this quarter? (I mean yes, yes, we’ve moved on from the capital discipline talk from 6- 12 months ago, but where is da monehh?)

Alas, it seems like the pre-planned spending was largely on GPUs, but as Nvidia’s H20 ban goes into effect, the gold is now sitting there without a newfound way to deploy for both BABA and Tencent.

Based on 信息平权 Information Democratized, a trusted WeChat blog’s calculation, “that the $5.5 billion impairment provision by Nvidia, when converted into the terminal sales price of H20, would result in an income impact of approximately $15 billion, which is about 100 billion yuan. Deducing the impairment provision by Nvidia is essentially the expenditure saved by BAT (Baidu, Alibaba, and Tencent). And it further estimated that if 40% of it belongs to Tencent and Alibaba, about 40 billion yuan would have occurred between January 31 and April 30 (Nvidia’s fiscal quarter),” which basically offsets the missing capex questioned above.

And on that note, most recently, at Computex 2025, Nvidia CEO Jensen Huang cautioned against writing off the entire Chinese market in Taiwan. He noted that “50% of the world’s AI researchers are Chinese.” If U.S. tech is absent, “they will build on Chinese platforms—and they will build excellent things.”

He estimates China’s AI market could reach $50bn as early as next year, underscoring its importance from both a demand and innovation standpoint.

But until China comes out with better domestic solutions, big tech is looking at software solutions. [Huawei Ascend explainer coming next week]. To DeepSeek’s role in all of this, it seems like it could potentially lead to finding an innovative solution in model + chip + system optimization.

Strategic Disruption to China’s Internet BBAT

From 2023-2024, the competitive logic of Chinese AI labs, from big tech to the startups, was based on several assumptions: 1) The stronger the self-research and development capability, the more solid the moat; 2) The number of parameters is related to capability, and performance is won by piling up large models; 3) It is necessary to build a "self-controlled" closed loop (ecosytem) of model + applications.

However, that has primarily been challenged by DeepSeek’s R1 release early this year, as its open-weight model has allowed everyone to start expecting models to be accessible and reshaped the landscape.

Tencent was the first to integrate DeepSeek into its applications and recognize that the battlefield wasn't in LLMs for them. The “four tigers” all somewhat pivoted into vertical use cases.

But not all disruptions were bad. They simply changed how the game was going to be played. Probably comforting to the startups is that a recent Jefferies China Software report highlighted that startups with deep vertical know-how, industry-specific data, and knowledge of foundation models are likely to succeed given the strong 2B demand, so the pivots made by startups like Zhipu, Baichuan, and so on seems to be now proven wise.

Despite concerns that we have written about China’s enterprise adoption rate and willingness to pay, Jefferies wrote in the same report that China AI’s 2B demand and willingness to pay are actually - surprisingly high. After the success of DeepSeek and Manus on a global stage, both 2B and 2C-end appetite in China has increased. That is based on the caveat that the applications are high-performance and low-cost solutions.

In one of our previous articles, we wrote that sometimes China’s SOEs do not prioritize efficiency gains, and that partially explained the low SaaS adoption rate, but it seems like that can be scrapped for AI. I guess I was wrong.

The current AI 2B willingness to pay is driven by fear of losing competitiveness, rising budgets in tech spending, reduced AI costs, and SOE/government top-down KPIs. The top-down encouragement of AI adoption is rallying SOEs in finance, telecom, and healthcare to want to be AI-empowered, particularly, and SOEs in these industries, well, their ability to pay (their pockets are deep) is high (was just a matter of willingness).

And in 2C apps, whether in search, office, travel booking, or companionship, have shown early signs of good demand, wrote Jefferies. I believe these are areas where the big internet players can leverage their ecosystems and existing mega-numbered MAUs.

AGI and Applications

The BBATs are now racing toward AGI, with some saying that ByteDance’s Seed Edge team is low-key pushing boundaries in research and development. But not a lot of public information on this so far.

For Alibaba, the recent results showcased that customers include not only early adopters such as internet, fintech, education, and EV but also other traditional sectors such as manufacturing. Customers' need for AI continues to rise, thus the demand to migrate to the cloud. The services this bundle needs to include are not just APl calls, but also post-training using specific internal data. On top of that, we’re seeing the need to lease GPUs and laaS services increase to satisfy various demands. At its core business, Al has been a positive booster for its commerce business, especially for online shopping with enhanced user experience, including search and ad recommendation, increased merchant productivity, and new forms of interaction (check out Alibaba.com’s 2B Accio).

As the company pushes ahead with its “AI-first strategy”, Alibaba CEO Eddie Wu predicted that 50% of global GDP will be based on or reliant on AI in the future. This bold and aggressive claim has become a guiding belief internally to push forward its flywheel: 1) cloud + model; 2) AGI.

My understanding is that Wu has repeatedly said that AI will become infrastructure, and the company is there to be the provider of that demand.

Because in many ways, models are now commoditized, and focusing on improving the models solely will be like how the internet went from 3G to 4G to 5G. As in, the experience can get better, but it shouldn't be a bottleneck to application innovation. So what is happening now is that the R&D can continue, but the adoption and innovation CAN now be applied. So, whether it’s Tencent or Baidu, they’ve fully embraced that point and have made it clear that they’ll integrate whichever model is the best into their applications.

Open-Source Era

Baidu’s pragmatic approach shows that the company is quite self-aware, given that their models really haven't made any major splash since their initiation. Now, embracing open source seems like a pivot that Robin Li reluctantly had to make, as prior to the February announcement of opening up its models, he repeatedly said that “closed source ensures technical control and business models; open-source is simply a fluke/ gimmick.”

Alibaba seems to be softening its tone of only using its Qwen series as its base infrastructure as well, but so far, no major moves have been made. However, Alibaba has the advantage of continuing to be loyal to Qwen. AI Proem recently wrote about its Qwen2.5-Max’s popularity amongst open-source developers.

ByteDance seems to be the most conflicted internally. On one hand, it’s spent massively on its LLM Doubao and the namesake application, hoping to continue its closed-loop ecosystem dominance in algo-driven applications like its Douyin, Jinritoutiao, and so on. However, to be the leader in AGI, it currently doesn’t have the frontier LLM capabilities. According to internal discussions, whether to integrate DeepSeek or not has been broadly discussed, but the attitude so far is “what’s the rush?”

Note that since DeepSeek’s release, ByteDance has stayed relatively low-key. Given that it’s the only non-public company in the BBAT, it has the luxury of not disclosing and addressing investor concerns quarterly.

All in all, open-source/ weight seems to be the trend forward. Under the impact of DeepSeek-R1, major companies have been re-examining their relationship with AI, and subtle changes in AI strategies are coming to life. This change will not stop: in the rapidly evolving technological wave of AI, to be honest, no one should be bogged down by legacy ways.

Until Next Earnings Season, Y’all

Do the BBAT still have an advantage in ecosystem moats? It seems like yes. And are startups even competing in the same tracks as the incumbents? It seems like no.

On the superhighway of AI technology advancing at breakneck speed, the most dangerous thing is not falling behind by one step, but it may be more important to be open-minded to continue to iterate and adapt quickly.

Tencent announced its 2025Q1 earnings yesterday, and beat expectations all around. As we last wrote about and predicted, many things were aligned. Its consumer-facing AI strategy continues, WeChat’s integration of AI functions continues, and AI-empowered ad sales, value-added services, and gaming use-cases boosted its overall revenue.

On a macro level, despite investor concerns over Nvidia’s H20 ban, the company’s president said it holds a significant stockpile of AI chips that will protect it from U.S. restrictions, and mentioned the options existing in the Chinese domestic market going forward. However, he also added the caveat that the alternative software innovations should optimise chip efficiency (alluding to Huawei’s Ascend chips and the CANN software that is known to be more power hungry than Nvidia’s system).

Subscribe now

Beating expectations, the growth was largely driven by strong performance in gaming and advertising, with both segments benefiting from AI enhancements. (I pulled out some interesting quotes from the earnings transcript, see below)

“During the first quarter of 2025, our high-quality revenue streams sustained their solid growth trajectory AI capabilities already contribute tangibly to business such as the performance advertising and evergreen games. We also stepped up our spending on new AI opportunities such as Yuanbao application and AI in Weixin. We believe the operating leverage from our existing high-quality revenue streams will help absorb the additional costs associated with these AI-related investments and contribute to healthy financial performance during this investment phase. We expect this strategic AI investment will create value for users and society and generate substantial incremental returns for us over the longer term.” — Pony (Huateng) Ma, Co-Founder, Chairman & CEO’s opening remarks.

Ramping AI Spending & Services

“AI capabilities already contributed tangibly to our businesses, such as performance advertising and evergreen games. We also stepped up our spending on new AI opportunities, such as the Yuanbao application and AI in Weixin,” Tencent’s earnings press release wrote.

Tencent said its capital expenditures for the first quarter totaled RMB 27.5 billion (~USD 3.75 billion), up 91% from the same period last year, and that the majority of that was continued investment into artificial intelligence-related functions/products/ infrastructure.

Key AI-related highlights that demonstrate the company’s future AI strategy: 1) Agentic AI; 2) AI in WeChat; 3) AI boosting ad and gaming; 4) Cloud

• Agentic AI was a key focus as management said there will be two types of opportunities in this space. 1) One is the unique agentic AI that will be able to connect functionalities within WeChat’s ecosystem, such as information to assist in transactions and/or more use-cases. 2) The other is in general AI native products, such as its Yuanbao and IMA, which may be tweaked and further enhanced to be more similar to what Manus AI looks like.

“On the general agent AI, I think we are creating that capability within some of our AI native products such as Yuanbao and Ima, over time, as these AIs continue to evolve to increase in terms of their capability. So in the very beginning, these AIs actually answer questions very quickly. So those are the sort of quick response.

And then over time, they include -- they start including the chain of thoughts, a long thinking reasoning model and you can answer complicated questions. And over time, the capability can actually allow them to start doing more complicated tasks. So they start evolving to have Agentic capability, and they will be interacting with all other apps and programs and external APIs to help the users. So that would continue to evolve.” — Martin (Chi Ping) Lau, President of Tencent.

• AI in WeChat: It was a big deal that WeChat integrated DeepSeek earlier in the year, completely changing the rules for AI consumer-facing monetization. Management said rolling out AI services within WeChat, such as integrating Yuanbao as a Weixin contact, and powering the search function with Hunyuan models and DeepSeek, and even adding in AI coding assistance for developers to reduce Mini Programs development time, will make WeChat much more competitive as well. I think this is huge. (see here why Tencent's integration of DeepSeek was such a big deal)

  ---

  

• AI investments: The company highlighted that integrating AI has boosted gaming and ad revenue, and the management expects operating leverage from existing business to help absorb the high AI spending. Specifically, AI could continue contributing to better ad recommendations, improved ad generation, and higher click-through rates. And for gaming, there are opportunities in AI deployment for content creation improvement and enhancement of the multiplayer game experience.

“Now at the same time, there's a sort of separate opportunity to deploy more generative AI into more content-driven games, and by doing so, perhaps, one would be able to facilitate initially faster creation of content by the game studio itself. And then over time, some degree of user-generated content, varying quality of experiences and ultimately, dynamically created content where the game itself is spinning up new environments if the players choose to venture off the preset map. But all of that is, I think, to be explored in the years ahead. And as of right now, the sort of bigger, more tangible opportunity is around utilizing what we'll call game artificial intelligence within the big competitive multiplayer games.” — James Gordon Mitchell, Chief Strategy Officer & Senior EVP.

• Tencent Cloud has focused on "AI + Cloud” as its strategy, and cloud revenue saw slight growth. Still, it remains a lower priority for the company as a whole, especially compared to Alibaba, which is doubling down on cloud and AI infrastructure support for 2B clients.

Overall, Tencent is embracing AI with full speed. Its moat is in its walled-garden content (like Google with YouTube), but tapping into its over 1 billion MAUs on WeChat and building in AI-empowered functions is much more like Meta’s. Don’t forget what we wrote about before: its strength is also in WeChat’s functional adjacency in the chatbox user experience, and actually in its shoulder-shrug attitude to LLMs. Why try to compete on frontier models when you can just integrate DeepSeek?

Other related Tencent AI or China Big Tech articles for context:

• Why Tencent's Integration of DeepSeek Into its AI App is a BIG DEAL

• AI at the Speed of Light: Tencent WeChat's DeepSeek Integration

• China's AI Application Super Powered: Open-Source Momentum Meets Super-App DNA

• Why China’s AI Strategy Differs from that of the U.S.

• China’s ChatGPT Doubao shows that inference compute demand COULD actually go up by 1 billion times.

According to headlines, the Big Four’s most recent quarterly earnings all seemed to have “beat expectations” or performed " better than expected.” Today, we take a closer look at their AI spending, revenue, innovation, and everything else related to AI in a more nuanced way.

Subscribe now

A few months ago, there was some noise about the unclear AI ROI causing investors to want to pull back, but that fear seemed to have dissipated. Comments about AI infrastructure in excess and oversupply also quieted down.

Alphabet, Amazon, Meta, and Microsoft, as the incumbents, are in some ways the most prominent players in AI in the U.S. right now, from owning the data-centre footprints, model talent, to the distribution channels that have set the pace for the entire generative-AI economy.

Their first-quarter 2025 results don’t just tell us who sold more ads or cloud credits. The results show whether the unprecedented cap-ex splurge that began in early 2024 is translating into usable GPU capacity, paying customers, and durable economic moats. It seems that spending continues to go up. Despite headline fears of excess supply for data center build-out, demand is exceeding supply, and the consequence on the environment is larger than ever. On the business side, the strategic gaps between the four giants are widening fast.

Macro factors matter, but CapEx spending continues

Google’s (Alphabet) business chief Philipp Schindler said that despite the company’s size, it is “not immune to the macro environment,” adding that President Donald Trump’s decision to end the de minimis trade loophole next month will “cause a slight headwind to our Ads business in 2025, primarily from APAC-based retailers.”

The Trump tariffs humbled the Magnificent Seven, Fabulous Five, Big Four, whatever you want to call them, as their stock prices all revised back to pre-Jan/ Feb rally. However, now removing the hype created by AI optimism, we can look at how the companies are revising their strategic plans.

According to CNBC’s report, Amazon plans to boost its capital expenditures to $100 billion in 2025, mainly on AI. Google parent Alphabet said it expects to invest about $75 billion in capital expenditures this year. Microsoft said it planned to spend $80 billion in fiscal 2025 on building out data centers to support AI workloads. Meta said it will spend as much as $65 billion on capital expenditures as it works to construct more data centers and computing infrastructure.

However, no one thinks that macro tensions are a reason for slowing down investment in AI. Almost every CEO on these calls has described AI as a “once-in-a-lifetime” platform shift in some way or form, and the numbers back up the rhetoric. Combined capital expenditure for the quarter hit roughly ~$80 billion, and the Big Four's full-year guides now total more than ~$240 billion. For context, ~USD 250 billion is approximately Sweden, Belgium, or Switzerland’s annual GDP.

No sign of slowdown on AI infra spending, highlights:

• Alphabet spent $17.2 billion in Q1 and reaffirmed a record $75 billion guide for 2025, almost all of it on servers and data centres to support Google Cloud, Search, and DeepMind workloads.

• Microsoft laid out $21.4 billion of capex and stressed that half of that spend is on long-lived real estate for data halls that will monetise “for the next 15 years and beyond.”

• Amazon’s cash capex was $24.3 billion, “the majority” of which was devoted to AWS infrastructure and custom silicon like Trainium 2. Even larger deployments arrived in the back half of the year.

• Meta raised its full-year guide to $64–72 billion, explicitly to “bring capacity online more quickly” for generative-AI workloads.

  Share

Company Breakdowns

A. Amazon: Betting on Vertical Integration

AWS closed the quarter at a $117 billion revenue run-rate, but CEO Andy Jassy now says the Gen-AI slice is already a “multi-billion-dollar business growing triple-digits”.

Amazon’s pitch is simple: if inference costs don’t fall, then mass adoption will stall. Thus, we’re seeing Amazon pushing ahead with its custom AI chip - Trainium 2, which claims to offer 30–40 % better price/performance than comparable GPU instances. Meanwhile, AWS’s Bedrock platform is expanding its model menu (Claude 3.7, Llama 4, DeepSeek R1) and showcasing its own Nova family. The Nova series includes the speech-to-speech Nova Sonic and the action-oriented agent Nova Act that aims to lift browser-based agent accuracy from 30 % to 90 %. Every element of its focus right now comes back to cost discipline: cheaper silicon, more efficient agents, and cap-ex that scales only when utilisation is proven.

As Mark Mahaney at Evercore ISI wrote, generative-AI “should produce a material narrative shift,” positioning Amazon as “a pure play on robotization and AI cloud services.” If Trainium 2 can work as advertised, AWS can then truly protect margin even as GPU prices rise and undercut rivals on inference pricing, which is precisely the lever that made EC2 dominant in the last compute wave.

B. Microsoft: Doubling down on Co-pilot

Azure revenue grew 33 %, and AI services contributed +16 percentage points of that growth. Under the hood, the commercial remaining performance obligation (RPO) soared to $315 billion, 40 % of which converts to revenue within 12 months. The company processed over 100 trillion tokens last quarter and already has 10,000+ organizations building agents on its Agent Service.

Despite the cloud gross margin falling three points, the company’s management reminds investors that half of the Q1 spend was on 15-year assets and depreciation lags monetisation. Azure’s AI attach is not a side business; rather, it is now the single largest incremental growth driver in the Microsoft Cloud. The backlog suggests that demand will stay ahead of capacity well into 2026, even if GPU supply loosens.

But tbh, I’m not fully seeing the grand winning vision here beyond cloud-driven growth, given that the co-pilot’s experience is still quite sub-par. Tapping into SaaS AI is probably the smartest way to capture vertical use cases and leverage its enterprise know-how and distribution.

C. Alphabet: Innovation and Margin Squeeze

Alphabet rolled out Gemini 2.5 Pro and Flash, Imagen 3 for images, and Veo 2 for video in a single quarter, while open-sourcing Gemma 3 (140 million downloads) to keep developer mindshare. CEO Sundar Pichai highlighted that AI Overviews in Search already reach 1.5 billion monthly users, and early adopters of the full AI Mode type twice-as-long queries. This datapoint could translate into materially higher ad prices once monetisation toggles on. The flip side: server depreciation is climbing 31 % year-on-year, and CFO Ruth Porat says that growth rate will accelerate through 2025.

Google is betting that richer, AI-generated answers lift retention and ad yield enough to offset a near-term margin squeeze from the heaviest cap-ex programme in history. Investors need to watch early monetisation tests closely. But as Brent Thill at Jefferies wrote: “The cloud division’s struggles are a microcosm of Alphabet’s broader problem: growth is slowing just as costs are exploding.”

However, there are more concerns and doubts surrounding Google’s ability to keep its search business relevant in the long run as GPT usage takes over search engines. Google’s strength in being cash-rich and having walled-garden content on YouTube, as well as an array of diversified products that would allow AI embodiment, may be where the next strategic shift will go.

D. Meta: Open source and Distribution Reach

CEO Mark Zuckerberg’s AI vision laid out in Meta’s AI road-map is actually the most clear in terms of monetization strategy and is split into five pillars: ads, content ranking, business messaging, the Meta AI assistant, and eventually hardware, which is then backed by that ambition with a $64–72 billion cap-ex budget. The company added 2,800 employees last quarter, most into AI and infra, yet still generated a 41 % operating margin. Model-based ranking tweaks delivered fresh engagement bumps: +7 % time-spent on Facebook and +6 % on Instagram in just six months. And management admits that, because even with the higher spending, the compute demand from internal teams already exceeds supply.

Meta’s competitive edge remains in its reach, much similar to some of the leading Chinese internet players too (Tencent, Bytedance, Alibaba). Meta’s DAU of 3.4 billion and its embrace of the open-source Llama 4 ecosystem in many ways are similar to Baba’s playbook (see here). If developers flock to that stack, Meta could monetise through messaging and device agents without ever selling a cloud GPU. The risk, though, is that until hardware is in users’ hands, the pay-off remains speculative.

The Bottom Line: Capacity Is the New Moat

When three trillion-dollar companies say they’re GPU-constrained, owning physical power contracts and data-center real estate becomes a strategic advantage, perhaps even more than owning the next great model. As we’ve covered extensively on AI Proem, there has been an obvious divergence in business strategies between AI companies in China and the U.S., and now the nuanced differences between the companies are also starting to play out.

Monetization ladders are diverging: 1) Amazon courts developers with the broadest model marketplace, b) Microsoft sells seat-based Copilot subscriptions, C0 Google chases higher ad yield, and d) Meta leans on engagement and devices. Expect very different revenue mixes by 2026.

Whatever they choose to move forward with in terms of their AI strategy, they are all faced with one shared challenge: the biggest bottleneck is the lag in GPUs.

Among the three hyperscalers that sell cloud capacity, they admitted they are still short on GPUs. Microsoft now expects “some AI capacity constraints beyond June.” Alphabet warned that Cloud growth will vary “depending on capacity deployment each quarter,” and Amazon said it could already be “helping more customers… if we had more capacity.” In other words, the companies with the biggest cap-ex budgets on earth cannot buy or build silicon fast enough to meet demand. That scarcity is itself becoming a moat.

For investors, the signal is clear: in an era where even the richest companies complain about hardware shortages, the victor will be the one that turns scarce computing into high-margin, high-retention products faster. Watch utilization, not just cap-ex, and the answer will reveal itself in the coming quarters.[There might be something for them to learn from DeepSeek and other Chinese AI companies now that they face even further restrictions on GPU access.]

Admittedly, I’ve been a bit distracted by the trade war, and if you’ve noticed, I actually announced some personal news last week - I’m expecting my second baby this fall, so I’ve slowed down a little to let my body do its thing. So, thank you for your understanding. (If you missed the last post, check it out; it includes a video interview with the Big Technology Podcast on China’s AI strategy)

Subscribe now

Now, I've been following the movements around the trade war and trying to understand the potential bigger tech and AI implications, on China and globally. Interestingly, I was having lunch with a German friend today, and he said the Trump administration’s recent shenanigans, from the Zelensky snub to the blanket tariffs, have made Europeans more pragmatic and really believe they need to be more self-sufficient. That is in auto, agriculture, energy, tech., military, everything (and I think this is an ah-ha moment that many North Americans/ Chinese were not as sensitive to)…

So I think, while many considered AI’s development in a way that previously had been China vs. the West, we may see French, German, etc players really take it more seriously to develop their own frontier models. In the policy space, for sure, there will be Europe’s own guidelines (of course, there already are, but it will be even more obvious that it is not dependent on how the U.S. goes).

Today’s post is not a deep dive; I just wanted to highlight a few quick updates relevant to the China AI space.

Share

1. Tencent Introduces Yuanbao Chat

It finally happened on Wednesday. Tencent introduced the Yuanbao chatbot in its WeChat interface. “信息平权” Information Democraticized on WeChat wrote: “All kinds of AI or agents, when integrated into WeChat, are more akin to a contact in the address book. This is likely to come sooner. However, WeChat's own native universal agent may take its time." Pointing out that this launch is not a competition in who has the best frontier model, rather, who has a stronger distribution advantage + walled garden proprietary data and a vast amount of it + functional adjacency, all of these are really setting Yuanbao apart from other new apps.

Select translation of Information Decraticized write-up:

“It's not about proving how good Yuanbao is, but about showing how deep the WeChat scenario is and how great the potential of the Agent is…. At present, it's essentially just a more convenient chatbot within the WeChat content ecosystem. After all, nearly 80% of everyone's daily information flow is within WeChat. With enough permissions, there's a lot of room to play around. So, what should the real WeChat agent focus on? Productivity scenarios, life scenarios, efficiency tools. While everyone is discussing MCP, mini-programs are WeChat's own MCP protocol ecosystem.”

Anyway, I think it will take some time for Tencent to fine-tune and work out the internal politics. Yuanbao is not the baby of the famous WeChat product manager Xiaolong. So it’s said that the integration isn’t as smooth as many have hoped, as there may be some edges to soften out internally. Nonetheless, this is going to hurt others, especially Doubao, where ByteDance has spent upwards of hundreds of millions of $$ on user conversion. That said, Doubao’s text-to-image and overall capabilities are still much stronger.

Looking at Yuanbao’s current abilities, it doesn’t seem to be able to match Doubao’s nor is it considering nuances. I wasn’t very satisfied with my first attempt, but if they can catch up in capabilities, then this is game over for the others.

I asked where I could get the most authentic Shanghai soup dumplings in Hong Kong, and the answer was the global food chain Dingtaifeng… Mhmm, I just feel like there are better places for authenticity. But it is definitely a game changer for how AI agents and consumer apps are going to compete.

If you can lean on an existing 1-billion MAU super app, why wouldn’t you?

2. Nvidia’s Jensen Huang Met with DeepSeek and Chinese Officials in Beijing

Pretty ballsy. Jensen Huang jetted off to Beijing right after the Trump administration announced its ban on H20 exports to China.

First some context. The whole thing can be a bit confusing, as the administration has gone back and forth on this.

In January, it was reported that U.S. President Donald Trump's administration was considering tightening restrictions on the AI leader's sales of its H20 chips, that was literally redesigned for the Chinese market (the most advanced chips that were legally being allowed to be sold to China). This idea had been brewing even under Joe Biden’s administration.

Given the DeepSeek rally in early 2025 and the fear of an H20 chip ban, Reuters exclusively reported a surge in orders for the H20 in February. As mentioned in this post, the tariffs and chip ban could have detrimental impacts on China's big tech. According to the Information’s report, together they placed at least $16 billion in orders for Nvidia's H20 server chips in the first three months of 2025.

Then, after that notorious million-dollar-per-head dinner at Mar-A-Lago, the Trump admin seemed to have temporarily backed off on this potential ban. Instead, on April 9, it told Nvidia that it would require a license to export the chips to China, which would cost the company a whopping $5.5 billion from canceled orders, the company said in a filing. And look, Nvidia was ready to take the fine and public beating, if it meant business could continue as usual. Which also led to speculation that Nvidia’s commitment to invest $500 billion in U.S. AI infrastructure in the U.S. was influenced by the administration’s temporary softening…

BUT obviously, Trump is not about predictability or keeping his words…and Jensen Huang being the smoothest talker in tech executives there is …

Just showed up in Beijing in a suit two days later.

Nvidia issued a statement saying, “The U.S. government instructs American businesses on what they can sell and where—we follow the government’s directions to the letter.”

Then, on the other side of the world, according to Chinese state broadcaster CCTV, Jensen said, "We hope to continue to cooperate with China," in a meeting with Ren Hongbin, head of the China Council for the Promotion of International Trade.

It has been hard to keep up with who meant what and what talks are being had, and whether Chinese AI companies will continue to have access to Nvidia chips or not. But one thing for sure is Jensen wants to keep the Chinese market open to him.

While on this trip, he also met with DeepSeek’s founder, Liang Wenfeng, seemingly not their first meeting. And then he made a flash appearance in Shanghai with the Mayor.

(Chinese netizens are joking that ”老黄”, old buddy Huang, is aging really fast with noticeably more white hair and really having a hard time trying to balance all this… While some nationalistic American netizens are giving him some sh*t saying he should be deported…)

In his previous trips to China, Huang has shied away from publicised meetings with high-level officials and mostly just met with entrepreneurs, as we’ve covered here before. Still, this time, he’s pulling out all the stops to make sure one of his biggest clients isn’t going to stop buying from him… Just saying, real ballsy to put a bullseye on his own back and bringing on potential PR troubles, but then the U.S. AI space is so dependent on Nvidia too, so what can the administration do to Mr. Huang?

And that’s where we’re at (for now) with the Nvidia- China- Trump drama.

Now for funsies.

3. Beijing hosted a robots vs. humans marathon (Video) today

See here for the CNN coverage: Chinese robots ran against humans in the world’s first humanoid half-marathon. They lost by a mile

This is not financial advice. I just think Tencent’s AI monetization strategy is quite smart. This is a quick wrap on its latest AI strategy from the company’s 2024 Annual Results. Please do not buy or sell stocks based on what I write.

Tencent (HKG: 0700) hit a home run yesterday. During its 2024 annual earnings results, ​​Chairman and CEO (and founder) Pony Ma attributed double-digit revenue growth primarily to AI-powered enhancements in advertising, higher engagement in video accounts, and an increase in evergreen games.

Subscribe now

AI Strategy Overview

Tencent has revamped its AI strategy, focusing on product innovation, deep model research, and increased AI-related capital expenditures. The company announced RMB 76.8 billion (~USD 10.7 billion) in capital expenditures for 2024, at ~12% of annual revenue, marking a 221% increase primarily dedicated to AI investments. While significant, this spending still lags behind Alibaba’s February announcement of RMB 380 billion (~USD 53 billion) earmarked for AI and cloud infrastructure over the next three years.

Management said that capex is expected to be in the low teens of total revenue in 2025. Morgan Stanley estimates it at ~RMB 94 billion (13% of revenue), which will mainly be spent on GPU procurement. Given that it’s not taking a cloud-first approach, capex being lower than Alibaba is also within expectations.

You could tell AI was front and center of the conversation. Tencent clearly took a cautious and gradual approach but is now all in on AI.

On capex, management highlighted:

1. Incremental capex spent in 4Q24 can generate returns in ads and gaming;

2. DeepSeek’s breakthrough has led to a significant increase in productivity;

3. GPU servers can generate returns and can offset the costs;

4. Monetizing AI through subscriptions remains challenging in China; instead, Tencent will focus on ads and value-added services (VAS) as long-term revenue drivers.

On 2C Applications

AI across all touchpoints and consumer applications - gaining traction:

1. Tencent released the Turbo S version of Hunyuan, which enhances computational efficiency and accelerates enterprise adoption.

2. The embrace of multi-modal foundation models means that Tencent will continue to invest in its foundation models to leverage its proprietary data entirely. Still, the additional use of external models will allow Tencent to leverage innovation across the industry.

3. LLM Hunyuan has been integrated into Tencent products, including ima.copilot, all WeChat verticals, games, QQ, Yuanbao, etc.

As I wrote last month, Tencent's integration of DeepSeek into its 2C application—especially its move into WeChat—was a huge deal. Unlike its peers trying to battle it out on the LLM front, Tencent made a strategic call or choice out of desperation. Whatever the reason, it chose to embrace DeepSeek, and it is now a straightforward consumer winner. On top of that, WeChat has also introduced a series of AI-empowered features, such as AI search and content generation.

AI-native app Yuanbao has proven this strategy even more successful. It is one of the first institutionally backed “model wrappers” in China, allowing users to select from multiple models. Like Perplexity, Yuanbao can process prompts in image, voice, and document formats. DAU increased more than 20-fold in February and March, becoming one of China's top three AI apps. Management believes Yuanbao will proliferate as an AI assistant in various use cases.

During the earnings call, Tencent's president, Martin Lau, explained why the company took a while to empower WeChat AI. Given that the WeChat ecosystem is highly complex, encompassing life, entertainment, work, mini-programs, payments, and more, there was a lot to take into consideration. Now, carefully positioning WeChat as a top-of-mind agent will naturally serve as a super gateway to all its offerings.

AI’s ROI

During the conference call, James Mitchel, Chief Strategy Officer, said that the capex's ROI is quite good. Already, the tech behemoth highlighted that AI has been able to help unlock the potential of existing businesses, including:

1) Advertising: LLM capabilities have supercharged advertising capabilities, allowing more accurate and faster approval of ad content, better understanding of merchandise categories, and acceleration of ad creative production automation.

2) Gaming: The use of machine learning technology in player-versus-player games and integration of LLM capabilities to high-production value games

3) Video and music: AI has improved animation creation, live-action video, and music content/ composition productivity.

2B Businesses

Tencent’s 2B AI cloud-first strategy has trailed behind its competitor Alibaba until now. The company’s management seems to acknowledge this and is addressing it head-on. AI cloud revenue approximately doubled year over year in 2024, partially constrained by Tencent's internal GPU usage on LLM training and improving ad targeting. With the capex hike to ease supply constraints, cloud services should see growth acceleration. The company is ramping up its 2B cloud business to meet external customer demand, as it has resolved supply constraints.

Investors anticipate that Tencent Cloud will benefit from this. Competition among China’s hyperscalers—Alibaba, Tencent, and ByteDance—continues to intensify in the LLM space. See tweet.

Quick Takeaways

Tencent’s strategic focus on consumer-facing applications like WeChat and Yuanbao positions it uniquely in China's rapidly evolving LLM landscape.

Tencent’s integration of DeepSeek was a brilliant move as China’s LLM sector ultimately becomes commoditized. Tencent’s multi-model approach—centered on ads and value-added services—sets it apart from competitors like Alibaba (enterprise-first cloud strategy) and ByteDance (frontier model development). This approach could be seen as similar to Meta's, with existing functional adjacency advantages, walled-garden content, and a massive user base (1+ billion MAUs). Tencent is making it clear that monetization will be on the consumer end.

However, as BABA still takes on an enterprise-first, cloud, and bundle approach coupled with some consumer-facing apps (Quake has not taken off, really), and ByteDance continues to try to compete in frontier models (which frankly rightly so if they can come out on top, but if not maybe they should try out DeepSeek?) and heavy marketing for user conversion for its 2C app Doubao. Our protagonist today, Tencent, is taking a multi-modal approach with a clear vision of making the $$ through ads and value-added services.

Glossary:

• Tencent LLM: Hunyuan

• Tencent consumer AI app: Yuanbao

• Alibaba LLM: Qwen

• Alibaba consumer AI app: Quake/ Kuake

• ByteDance LLM: Doubao

• ByteDance consumer AI app: Doubao

Disclaimer: Do not take this as stock-picking advice. AI Proem is purely for informational and educational purposes.

Subscribe now

I can't remember the last time that investors were so excited about Alibaba's earnings. All ears were for Alibaba’s ($BABA) earnings last night, and I believe the server was having a hard time handling the interest, as my call was quite glitchy in the beginning. But this time, no one was expecting jaw-dropping MAU growth numbers for its commerce businesses. At the end of the day, the company has had over 930 million MAUs for years now. How much is there to really grow when the whole population is 1.4 billion? (today, let’s only focus on AI)

Subscribe now

Leading with 13% yoy in cloud business revenue growth driven by double-digit gains in public cloud revenue (AI, duh), compared to ~7% the quarter before and ~5% before that. In fact, the company is strategically heavily investing in AI capex.

[You can find the 6K filing here]

On the earnings call, Alibaba CEO Eddie Wu said that the company would “aggressively invest” more capital into AI in the next three years (at least RMB100 billion—USD13.78 billion—per year) than it did in the last ten years in aggregate. When asked by a Citibank analyst if he could give more information on how the spending will look, Eddie said this is a “huge opportunity,” something that comes around once in decades.

Ultimately, for Alibaba, AGI is the main goal, and the company will continue to push forward the technical capabilities of its LLM Qwen. Although the monetization strategy is not clear yet, he is certain it will come through. By continuing to offer a strong cloud service (number one in China) and a leading open-source LLM, Alibaba is well-positioned to capture AI opportunities in the future. Eddie added that he believes that in the future, 90% of AI computing will be centralized and not on edge, and Alibaba has the advantage in cloud infrastructure. He concluded his answer with an analogy of AI being electricity and AI cloud computing power being the electricity grid that enables it all.

(This is not translated verbatim. It’s my notes from the call. The official transcript will be up on its investor relations page within the week).

Ok, now let’s look at the financials.

AI as the Growth Engine

Alibaba Group’s Q3 FY25 results highlight more than just strong performance—they mark the beginning of a profound shift toward artificial intelligence (AI) as the tech behemoth's core driver of future growth.

The company’s 8% year-over-year revenue growth in Q3, powered by strong performances in customer management revenue (CMR) and cloud, shows that its core businesses are back on track. However, what truly stands out in this quarter’s earnings is the company’s focus on AI and its growing potential to redefine both business operations and the broader Chinese tech ecosystem. With AI-related revenue maintaining triple-digit growth for the sixth consecutive quarter, Alibaba’s push to integrate AI across its business models is clear.

As explored in my previous articles, China’s AI Application: Super Powered, Why Tencent’s Integration of DeepSeek, and AI at the Speed of Light: Tencent, China’s leading tech companies, including Alibaba, are harnessing AI not just as a tool, but as the engine driving next-generation innovation and monetization.

Long-term AI Strategy

AI Infrastructure and Multi-Chip Design

Alibaba’s massive capital allocation reflects its executives’ belief in AI's transformative power. The company aims to position itself as a leader in AI infrastructure and a hub for AI innovation across multiple industries.

In line with broader trends in China’s AI application landscape, this infrastructure will power everything from cloud computing to AI-native applications, accelerating the development of tools and services that can drive productivity across sectors. Shaping the future of AI-driven business models.

The company further emphasized that Alibaba Cloud adopts multi-chip infrastructure capabilities, which positions it well to maintain its planned investment trajectory, regardless of potential regulatory challenges.

Proprietary AI Models and the Open Ecosystem

As written in the Alibaba AI strategy primer, the heart of Alibaba’s AI strategy is its commitment to developing proprietary AI models, such as the Qwen2.5-Max’s strength and uniqueness before DeepSeek came out with an open-source model was its technological prowess, but also because it was one of the only leading models in China that fostered an open ecosystem. With over 90,000 derivative models developed on Alibaba’s platform, the company is creating an AI-driven environment where developers, businesses, and even consumers can tap into cutting-edge technology - which then they can access the talents/innovation/data and then upsell their cloud services.

However, the most intriguing aspect revealed during this earnings call was perhaps the fact that Alibaba’s AI strategy is its long-term goal of achieving artificial general intelligence (AGI). Eddie said on the call that, based on research, 50% of the global GDP could potentially be replaced by AI once AGI is achieved, a significant shift in labor dynamics in the world today.

The CEO emphasized that this focus on AI isn’t just about innovation for innovation’s sake. It’s about creating a competitive edge in a rapidly changing market. Alibaba is using AI not only to improve the efficiency of its e-commerce operations but also to enhance the user experience and increase monetization opportunities. Consumer-facing apps like Kuake and Tongyi are seeing strong user engagement, and by leveraging the Qwen models, Alibaba is adopting AI integration across platforms like Taobao, which is helping to enhance transaction efficiency and merchant support. All of these are consumer-facing innovations on top of its grand investment plan on AI infrastructure. On top of that, it’s enterprise software Dingtalk (the Slack of China) has ample opportunity to utilize AI to help productivity and efficiency in the workflow.

Broader Impact on China’s AI Landscape

As I’ve highlighted in the past, the country’s tech giants are increasingly focusing on AI as a central driver of economic growth. Alibaba’s latest earnings report underscores the company’s shift from focusing solely on short-term ROI to pursuing long-term AI and AGI ambitions. With its massive investments in AI infrastructure and commitment to developing proprietary AI models, Alibaba is positioning itself as a leader not just in the cloud space but also in the future of artificial intelligence. As AGI looms on the horizon-ish (give it another few years and we can circle back), Alibaba is trying to build the foundation for a new era in technology—one where it may shape the future of human productivity itself.

Full disclosure: I worked at Alibaba during its most tumultuous years, and I am forever grateful for that opportunity and the friends I made there. However, in 2020, the company was in a position of weakness, and the founder, Jack Ma, despite his outlandish nature, went into a bit of hiding. The whole big tech industry was faced with an anti-monopoly probe that resulted in a record fine of USD 2.8 billion from Alibaba to the government for abusing market dominance.

That was followed by continuously losing market share to newcomer Pinduoduo (PDD) in e-commerce, which was caused by overlooking competition. But to be fair to the company, there was also an overall economic slowdown and, therefore, consumption downgrade, and as Alibaba’s Taobao/Tmall was to capture higher priced goods compared to PDD’s platform, offering low-cost white label products mostly, Alibaba was hit quite hard. Coupled with an intensifying trade war, ADR delisting risks, and a few corporate governance scandals… All of this did not contribute positively to the company’s growth over the last few years, and we saw the stock price drop from an all-time high of ~USD 310 to as low as ~USD 60. Things were kind of lackluster.

UNTIL NOW.

China Tech is back, and BAT is back. I don’t know if they’ll actually achieve AGI, but at least every retail investor auntie is getting their money back after the wild ride of the last five years.

Today, I will analyze Microsoft’s latest article, which calls for cooperation in the development of AI from a strategic communications perspective. Microsoft’s Vice Chair & President of Microsoft, Brad Smith, last week published an opinion piece on the corporate website titled “The Golden Opportunity for American AI,” please see the full article before reading further.

TL;DR

1. What is the key message?

AI is good. (quite literally tagged at the bottom of the page)

2. Who is this for (which stakeholder)?

The incoming Trump administration. Def, not Silicon Valley or investors nor the general public.

3. What goal is it trying to achieve?

Asking the government to rally resources to help/ and go easy on AI regulations.

Subscribe now

According to the company’s official bio, Brad Smith spearheads the company’s work on critical issues involving the intersection of technology and society, including artificial intelligence, cybersecurity, privacy, environmental sustainability, and more. He leads a team of roughly 2,000 business, legal, and corporate affairs professionals in 54 countries.

It’s interesting that the head legal and PR/IR/GR guy is the one to speak about AI opportunities, not a technical person or a product person. It made me think that this letter really is less of a “what is Microsoft AI strategy” or “how do we plan to monetize AI” but more of a “let me tell you why we’re doing it, (please support it)” to regulators and the incoming Trump administration (as he blatantly also spells it out in the first sentence).

He has positioned Microsoft’s AI ambitions as an American pursuit, stoking patriotism while ensuring the government that the company will put America first, rather than its own interest first, so it says.

“Not since the invention of electricity has the United States had the opportunity it has today to harness new technology to invigorate the nation’s economy. In many ways, artificial intelligence is the electricity of our age, and the next four years can build a foundation for America’s economic success for the next quarter century. “ — Brad Smith, Microsoft.

When he talks about what could be applied as a global phenomenon, the letter reiterates how the U.S. has the right, responsibility, and almost duty (to its allies) to hone AI to power its economic success and maintain its global leadership position.

I’m not mocking at all; seeing such adulate kowtowing is just a little funny. It’s good to know that even mega-tech firms like this have to abide by laws and rules and are in fear of regulators. I guess it makes sense because you gotta do what you gotta do to ensure you don't get curbed to invest and develop this critical strategic business.

And look, Microsoft is not the first company to communicate its positioning to an incoming administration. As we know, the Biden-Harris administration focused heavily on “Build(ing) America’s Infrastructure and Clean Energy Future Faster, Safer, and Cleaner,” and during their administration, renewable companies repeatedly highlighted how they are the ESG choice of the future.

Now, with the incoming Trump administration, renewable energy companies’ narrative has shifted to how they are the solution to the energy shortage amid the backdrop of Trump’s promises that he will “frack and drill on day one.”

Source: Not AI. Google Image of Trump and handmade on Word Doc

I translated the corporate PR language in his three-part vision list into normal human language.

1. “This starts with advances and investments in world-leading American AI technology and infrastructure.” - translation: let us invest and help us with the energy shortage issue that has become the main bottleneck for hyperscalers to build out data centers faster.

2. “Second, the country needs to champion skilling programs that will enable widespread AI adoption and enhanced career opportunities across the economy.” - translation: we know how we are one of the biggest cloud infra service providers, and we are in an “investor relationship” (as Satya Nadella puts it on B2G podcast) with OpenAI - the leading frontier technology provider - let us hire, build and be the enabler of this.

3. “Finally, the United States must focus on exporting American AI to our allies and friends, bolstering our domestic economy and ensuring that other countries benefit from AI advancements.” - translation: it’s actually interesting that he would state such a strong stance on diplomacy/ geopolitics. I guess it is to feed what the Trump administration wants to hear and appease the “Make America Great Again” mentality.

Overall, the piece highlights the critical role the private sector has played in ensuring America's success during each technological revolution. He chronologically lists the Industrial Revolution, electricity, computer chips, software, and now General-Purpose Technology, which isn't wrong. It's just very strategic positioning.

It’s also interesting to see that he’s pointing out the energy shortage as a key bottleneck to AI development in the U.S. and asking for the administration's support in finding solutions to speed up the process, which I’ve written extensively about for months.

I think the biggest surprise from this all is his last section on AI upskilling, which has not yet been as widely talked about by big tech. And I believe it is a genuine call for attention to the fact that the country needs to take it seriously and embed AI proficiency into education and the workplace.

He pointed out that AI skilling has been a focal point of U.S. policy since President Trump’s 2019 AI Executive Order, which is a shout-out to Trump’s good judgment here, I guess.

He talks about an initiative to push forward the need to integrate AI into educational curricula and develop apprenticeship programs, particularly in STEM fields, to prepare the workforce for an evolving job landscape influenced by AI. And companies like Microsoft are already taking steps to train millions in AI competencies, emphasizing the importance of community colleges and partnerships across various sectors to make AI education accessible.

Ultimately, the goal is to empower all Americans with the skills necessary to use AI to improve their job prospects and career success, positioning this as a national priority for future growth.

Last but not least, without pointing fingers, he wants the government to rally resources and ensure that the U.S. comes out on top of the AI arms race with China, whether that’s playing the cards diplomatically to export relevant technology to allies or extending export bans to China. However, I did earlier predict that Trump would actually bring a respite to the Chinese tech/ AI sector despite tough rhetoric.

Smith says that with the rise of generative AI and intensified competition from China's rapidly advancing AI sector, the U.S. must develop a strategic approach to promote its technologies globally. Historically, American companies have led in innovation, but China's government has successfully subsidized its tech exports and garnered growth across the infrastructure supply chain. (which I’ve written about here)

He calls for the U.S. government to help them more. To counter this, he says the U.S. needs to leverage its superior technology and foster strong international partnerships, particularly with allies in the Middle East and beyond (which Chinese companies are also courting).

He reiterated that companies like Microsoft are making significant investments to build trusted AI infrastructure worldwide. The American government should prioritize enabling its private sector to thrive without burdensome regulations while balancing security through pragmatic export controls. Optimism for American AI's future stems from a robust technological foundation, a dynamic business sector, and a strong educational system poised to equip workers with the necessary skills. These factors position the U.S. to lead in global AI adoption amid competition with China.

My takeaway: GR strategy over PR/IR strategy

• IR/ Microsoft and big tech have definitely won over investors and have successfully justified their $200 billion+ spending on building out AI infrastructure for potential future returns, as we can see through investor reactions to their earnings.

Since the release of ChatGPT (Nov 30, 2022) and the thrust of AI on the general public, the big four in the U.S. have added $4.1 trillion to their market cap…

• PR/ To the public, the AI story is so sexy that despite concerns over its harmful implications, the public is happier than not to try out the newest tools. The last and most important stakeholder to manage now is the government/ regulators.

• GR/ This is just one of many public pleads that we’ll probably see in the coming months. I think many tech companies are lobbying, rallying, and doing everything they can to get the government (which is fairly skeptical due to concerns over resource consumption, income disparity creation, data security, national threats, and more) to back them and propel AI and see the technological development as good for humankind/ society than the harm it could bring.

New 2025 Pieces

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

2. Release of OpenAI's o3 has completely changed the debate on whether AI scaling law has hit a wall

3. 2025 Predictions, New Year Plans for AI Proem, and Appreciation for Collaborators and Readers