Hangzhou Alibaba field trip. Five AI business products. Who will stand?

ecomm, agents, supplier mgmt, smart cockpit, and more

I accidentally lived in Hangzhou for a few months during COVID. I was supposed to go up for a 3-week work trip (of which 2 weeks were in quarantine) that turned into nearly 10 months of bouncing around Hangzhou, Beijing, and Shanghai. Since then, I hadn’t really had a reason to be back over the last five years.

As I was prepping for my day trip from Shanghai, I was dreading the hour-long taxi ride from the bullet train station to Alibaba’s Xixi campus that I used to have to take during my stay there. But to my surprise, a former colleague told me that there is now a new bullet train station, Hangzhou West, conveniently located near Alibaba’s main campus. Trimming down that hour-long taxi ride to 10 mins.

Alibaba Cloud: Green by Design

It takes about an hour on the train from Shanghai Hongqiao to Hangzhou West, and when the doors opened, I couldn’t help but chuckle. Is this the uniform or the epitome of the Chinese techfit? Everyone on the platform was wearing a Tumi backpack or laptop case. It was quite clear that everyone looked frankly more “expensive” in their outfits getting off at that station, and there was a clear urgency in their footsteps.

I got into a taxi and started chatting with the uncle driving. If I hadn’t been back in five years, he told me, then too much has changed. New highway bridges, rows and rows of new development, and of course, the completed new campuses for Alibaba.

So I asked him, “Does that mean housing prices have gone back up with the AI boom?”

“Oh no, no”, he said, only a few are buying up luxury apartments, but they work for semi companies. He added that he heard kids working in AI are making 2 to 3 million RMB a year, which seemed like a crazy number to him. But those kids working in AI are so tired that they get into his car and pass out, and he has to wake them up at their stop. In awe of their talent, but also worried about their health, as one would expect a typical Chinese uncle to be.

Now, I was flabbergasted at the response from this uncle, frankly, who could barely speak Putonghua clearly, but had the sophistication to know that only the semi companies, their investors, and everyone along that supply chain have made a lot of money this year, and that the rest of tech and AI is still mostly waiting for its turn.

In some ways, that taxi driver in his 50s in Hangzhou had a more accurate understanding of the cycle than a lot of professional analysts.

Before, I felt like Hangzhou hosted Alibaba. As in you know by default, the dominant payment system is Alipay and not WeChat in the city, but it still felt like Alibaba was part of it, rather than the other way around (at least in Hangzhou East).

But getting in from Hangzhou West, it doesn’t feel like a city that hosts Alibaba anymore; it feels like an Alibaba campus with a city attached to it.

You know how tech companies have sprawling campuses? This whole city is an Alibaba campus. The train station itself was reportedly built in part to make the Shanghai commute more convenient for Alibaba staff, and apartment complexes went up around the new developments Alibaba established. The spending is crazy, and it cuts both ways. On the one hand, it shows confidence in the company’s direction; on the other, when you speak to investors, there is a level of skepticism because they are wondering how much of that money should have been spent versus how much should have been returned to shareholders.

With that, I will guide you through my day, my observations, and thoughts in chronological order. This is from a day spent meeting with four (five) business units: the Qwen App team, Accio, the Qoder team, and a brief hello with Banma and (light touch) Alibaba Cloud.

Qwen: The closed activity loop is real, but who is willing to pay for it?

The Qwen app is extremely impressive in how it connects all the commerce activity, even though we wrote about how it hasn’t fundamentally changed how users engage with the idea of shopping yet.

But this is Alibaba’s edge if tokenomics eventually comes down. Think about it, everything can be completed within its own platform. E-commerce, same-day delivery, grocery, maps, ride hailing, bookings, payments, all integrated, agentic, agent to agent. This is something we’ve already seen OpenAI couldn’t really figure out with its reliance on third-party partners. It’s easy to push a user to DoorDash and then to Stripe after they’ve shown intent to buy, but it’s much harder to have the agents link up so the full loop, from car booking to airplane to hotel, actually completes.

In this case, Alibaba owns the rails, so the loop closes all under its control. There is good and bad to that. The good is that this activity loop could even exist; the bad is that you bear all the liability, right?

The representative I met repeatedly reminded me that any purchases or transactions still require human verification to prevent mistakes, but if we are to push for an agentic commerce future, maybe that step will eventually be removed?

Anyway, what I found is that the Qwen app is more nuanced than a checkout bot. The app is memory-based, so when I want a dress, it already knows my size, my body type, and my usual preferred style, and unlike GUI-driven agents, it doesn’t need to operate a screen because it calls directly on stored memory. Again, this is something unique to Alibaba, having the Taobao purchasing data.

It behaves like a shopping guide rather than a fulfillment engine. The PR said, if you want to buy a jade bed because influencers on RED are saying it helps with chronic illness, they will tell you not to. It has built-in authenticity, and I guess, in some sense, integrity.

Mention at night that your cat isn’t sleeping, and it will suggest it might be an illness or a calcium deficiency and ask if you want to buy something for it. This agentic shopping companion is its future vision.

The team’s framing, which honestly I buy, is that China already app-ified every booking and ordering need over the past fifteen years, and it is much easier to AI-fy a mature internet infrastructure than to build agentic commerce on top of fragmented rails.

The payment step stays deliberately human, so the final purchase has to be personally confirmed, except for utility goods like the weekly cat kibble or toilet paper, for which you can choose to authorize a recurring order. Shopping has been live since May, and consumer usage is reportedly growing 2x month-on-month; overall data is not publicly disclosed yet as it’s still in its relative infancy stage.

One example that really impressed me, and that I would personally pay for, is planning a day trip from Shanghai to Hangzhou. All you need to do is tell the agent the time you want to arrive and your preferred transportation, and it will plan out your time, your snacks, your coffee, and everything in between, frankly, even more meticulously than a human secretary.

But right as I was getting excited about this, it dawned on me: how can you justify the cost of token usage for such simple activities?

I would pay for an app to manage my itinerary, but how much would that token cost, and would it be worth it? I guess it really depends on how expensive your own hourly rate is.

And there is a second problem that pricing can’t solve, which is liability. If there is traffic and you miss your train, will you blame the AI or your own judgment? Who bears that risk, Alibaba, the airline partner, or you?

The bigger question here isn’t really an Alibaba question at all. It’s a question I have for the whole industry: how much is a consumer actually willing to pay for agentic support?

Just as Azeem’s team wrote in the recent report The State of the AI Economy, “The AI economy is bigger and faster than any technology wave before it, yet still small enough to be early.”

That is how it felt. Qwen, in many ways, felt ahead of the curve. But maybe the infrastructure cost-benefit just isn’t mature enough to justify such a use case yet? Because the stock price has not been a friendly reflection of this vision. That same report shows that AI is scaling three times faster than any prior IT wave, but much of that is from the enterprise level. Who is bearing the cost for consumer usage?

This is why OpenAI’s pivot last year back to enterprise rather than consumer shows a bigger revelation in the industry. Right now, there is only willingness to pay in enterprise because AI is fundamentally a productivity-gain technology, and only businesses really prioritize productivity, plus a small percentage of people whose hourly rate justifies it. Look at Doubao; its MAU dropped the moment it started charging.

The rest of the consumers are still looking for value for money, really, which is exactly why PDD and Temu, despite the hate, are doing so well. How much is the consumer agent really going to be contributing to GMV? And if Alibaba does want to monetize the transactions themselves by charging a premium on each agent-completed service, how much is justifiable before the consumer just does it themselves?

For now…I think there is still a lot to grow, iterate, and unpack as tokenomics matures and AI user behavior evolves. This is a watch-and-see app for me… no conclusive verdicts yet.

An interlude at Hupan

Between meetings, I visited a newly opened exhibition hall where they have installed a replica of the Hupan apartment, which is where Jack Ma and the original team started. The PR shared a funny anecdote: because there were so many people crammed into this little apartment and only one squatty potty, people joked it was “Lundun,” which means “taking turns squatting” and obviously sounds just like London.

Alibaba Xixi Campus + Street View

Jokes aside, it was very nice to see how much respect is paid to the heritage of the company’s founding. And a special Chinese corporate thing is to have showrooms open to guests.

Accio: AI UX on a twenty-year supplier data moat

TBH, my real goal in going to Alibaba this time was to learn more about Accio, and it was the highlight of the day, a business I find very undervalued. I read about it a year ago, but at that time it was still merely described as a merchant support AI tool.

I explained on the ~100-investor JPM call on Friday that it really undersells it, especially now after a year of iterations. On the surface, it looks just like any other agent service for merchants, and the interface even looks like Claude. But it relies heavily on Alibaba’s sprawling supply network, which is its ultimate advantage.

Accio is less an “AI product” and more an AI UX on a twenty-year supplier data moat. This is the clearest enterprise 2B use case Alibaba can push out right now, and frankly I don’t even think it’s getting enough attention internally at this point.

The value of Accio isn’t just the AI agent. It’s that the AI sits on top of 1.5 million verified suppliers, 400M+ SKUs, and decades of transaction and rating data that Alibaba owns because it is the B2B marketplace. There are already millions of MAUs.

And there is also a reason that there is no Accio equivalent in the West, because the supplier graph simply doesn’t exist there for anyone to build on. No Western company has that asset. Even Amazon can’t do this.

For a merchant, it is essentially everything in one; it shocked me how it removed the barriers for business/ drop shipping and all the tedious search and due diligence work that is needed to start a brand. Look, I even toyed around with the idea during COVID because I was bored, but it was much more work than expected to source the right product and manage the distribution channels. The best way to understand it is to walk through what a merchant actually does all day, because Accio covers nearly every step of the workflow. Most users are startup entrepreneurs, brand owners and e-commerce sellers, brick-and-mortar retailers, and service providers, with 30 to 40% coming over from alibaba.com.

It starts with supplier finding. The agent sits on the supplier graph and can find, filter, and talk to vendors for you. Autochat can speak to vendors through one interface, say, for sourcing just the lid of a bottle, run the first round of filtering against your core requirements, and then hand off to you to speak to the shortlist directly. Sometimes it is literally Accio talking to Accio on both sides of the negotiation. And the sourcing guidance is real know-how, not chatbot filler. One buyer didn’t know the measurements or the requirements for their product, and the agent guided them through the safety requirements, the weather conditions, and what those mean for raw material choices- the kind of professional judgment that lives in the network and twenty years of transaction data, frankly. It’s not more intelligence but more know-how. [Industry-specific use case: AI tools are the future is what we’ve been writing about!!]

From sourcing, it moves into pricing and margin. The agent negotiates pricing with vendors behind the scenes over multiple rounds, then benchmarks against industry comparisons and hands the merchant a rough margin before they’ve committed to anything. This is the part I keep coming back to, because it means a small merchant gets the pricing intelligence that used to require a sourcing office in Guangzhou.

Btw, I know I’m like super positive about this product. Swear to god, I got zero dollars from any of them. I self-paid for my trip and all.

Then the operational layer: supply and inventory management, plus the back-office work around it. Accio Work claims to handle VAT filings across 100+ markets, multi-round supplier negotiation, and logistics ops, though how much of that is truly AI-executed end-to-end versus AI-assisted with a human in the loop is one of my open diligence questions. On the storefront side, it covers CRM and platform management, lead generation, content management, and SEO and GEO optimization. It connects to Mailchimp, WordPress, and the like, where some opened up access to Accio, and some skills were built directly by the Accio team, so the experience is smoother, and it is opening up WhatsApp and Telegram so your vendor conversations get summarized straight back into Accio instead of dying in a chat thread.

And then distribution. Like HELLO. Plugins are already live on Amazon, Shopify, Etsy, and Shopline, and you can swap the 1688 agent for an Amazon seller toolkit depending on where you sell. The data runs both ways, too, internal Alibaba data from Alibaba.com and 1688 alongside external data from Amazon. A marketplace’s self-run toolkit only works inside its own marketplace; Accio pushes to every channel the merchant sells on.

The dropshipping honestly would be perfect, and it is incredible. If I were without kids and had more mental bandwidth, I would SOO create my own loungewear brand with this. (During COVID, I had ideas of setting up my own Shopify shop, but the process was actually much more tedious and consuming than I expected. I toyed with doing dropshipping and creating a pet product storefront, opening up a bagel shop in HK but realized you can’t mess with that here, there’s enough drama already, and/ or a luxury loungewear brand inspired by my West Coast upbringing living in leggings and Birkenstocks)

Anyway, back to our main topic. So you can give a verbal prompt. Accio provides the sourcing, then provides visuals, then technically can communicate the pricing behind the scenes given that you’ve given it permission, then provides an industry comparison and a rough margin, and then it creates the Shopify store link and can run the whole website with product descriptions and everything, and helps with social media and image banners. It is sourcing to storefront to marketing in one conversation box.

While I was going through the demo, what really stood out wasn’t the chatbox or the interface or anything. It was the supply chain layer where the real differentiation was, built on their own network and sourcing know-how, while store management, marketing, and operational support were more similar to what other agents could do, helpful but not as unique.

This is also where the willingness-to-pay math finally works. A consumer won’t pay for twenty minutes of saved itinerary planning, but a merchant will absolutely pay for an agent that compresses sourcing, negotiation, storefront creation, and marketing into one workflow, because the ROI is measurable and the alternative is headcount.

I would love to have someone from the team join us on the podcast to talk more sometime.

Models, Tokens, Bureaucracy

The most candid conversation of the day was with someone in the cloud/Qwen unit. Qwen, he admitted, still lags well behind GLM on coding — awkward when you’re selling Qoder, because in practice many Qoder users are paying Alibaba for a managed service to run GLM. In effect, Alibaba is taking a margin on its rival’s model.

Domestically, this doesn’t matter much. There’s no dominant Cursor or Codex equivalent in China. I mean, ByteDance’s Trae competes, but Alibaba’s enterprise reach carries the product. Outside China, it is a different story, with competition obviously led by Anthropic and OpenAI.

However, I’ve heard a few voices saying that they feel startups have an advantage in leading innovation because of the lack of layers and bureaucracy, which makes big tech feel slower. In general, a lot of these big tech companies are really good at integrating AI into their existing businesses, but they are spending way too much and still having a hard time pushing out new technology. And this is not just a China thing; the same voices pointed to the U.S. peers too, look at Google and Microsoft. The bureaucracy and the management are sometimes the ones holding new innovations back.

I’ve been getting mixed responses on internal token usage, but in general it seems like Alibaba has not officially capped it for employees.

Developers can use whatever they want, while purely writing or document-focused work gets a 3,000-credit limit, though people can apply for more. They deliberately switched from tokens to credits because explicitly showing how many tokens each employee uses is too transparent, and they don’t want to disclose that. With the cost per token coming down so fast, the point is really about managing efficiency rather than rationing consumption. Even inside the company that owns the models, nobody wants the unit economics to be THAT transparent.

Beyond the technology side of things, he sees the potential issues stemming from the fact that the last generation of Chinese tech people are not international enough, in both language and mindset, whereas the AI-native companies and the next generation of AI companies are often very international and focus on going global first. Domestically, they obviously have the mindshare and dominance, so whatever they push out for B2B is a lot easier to accept because of the credibility and trust already built into the brand, but that credibility doesn’t travel.

As for what comes next, he thinks the next wave of products might be general-purpose usage agents that find vertical use cases (!!), maybe, say, building agents for content creators, with specified, tailored plugins for each use case. And within Alibaba, its Cursor-like product Qoder is already showing signs of this pivot, integrating more industry know-how and running smaller models to help with the token usage constraints and worries, much like what we wrote about before. It echoes the epiphany from the Accio demo.

Wrapping up my rant

While it’s much easier to get carried away by narratives about the consumer story, the real question is how they can monetize it.

The moat is the rails and the data, not just the model. Qwen commerce works because Alibaba controls Alipay, Taobao, and Amap, and Accio is defensible because of the supplier graph, while Qwen the model lags GLM, and it barely matters domestically.

Similarly, if you think about it, whether it’s the smart cockpit business Banma or Dingtalk, B2B monetizes before B2C, because businesses buy productivity and consumers buy value for money, and until token costs fall far enough, or someone solves the liability question, the consumer agent is a moat-defending feature rather than a revenue line.

And the reason why I’m so gung-ho about Accio is that the merchant use case is obviously chargeable and defended by a real edge.

And despite some concerns around the company’s heavy spending, one small detail that gives me some confidence in it is that the mgmt is now taking a page from ByteDance and running A/B teams apparently on various products - including the foundational model layer- to see who can outcompete and win out. Maybe that kind of “wolfness” or hunger is back at Alibaba?

---

Last but not least, I’m super happy to say we’ve been approached by some super interesting mid-size cap companies that span AI software to spatial intelligence and physical AI, and we should be releasing some interviews with them in the coming weeks. Stay tuned :) I’m really excited to share them.

---