Original · Unique Research · September 3, 2026, 5:00 p.m. · Shanghai
“Promises can be copied, and specifications can be matched. But no one can take away what you build by keeping those promises year after year.”
American high school students start school at 7 a.m. and finish at 3 p.m.
That leaves a lot of time after school. If they are home alone, wouldn’t an AI companion toy that could talk with them be a perfect fit?
That was exactly how Ye Jie, who makes AI companion toys for adult women, initially reasoned about the North American market. It sounded perfectly logical.
Then he thought about it more carefully and realized something was wrong.
First, only-child households are relatively uncommon in the United States, so teenagers may already have older or younger siblings for company. Second, American high school students typically have all kinds of activities after school. Getting out at 3 p.m. does not mean going home to sit alone; it often means playing sports, joining clubs, or spending time with friends.
The use case you imagined may not exist locally at all.
At a roundtable on taking AI hardware overseas at the Physical AI Summit in Shenzhen, the speakers included Tim of Haivivi (跃然创新), which makes AI toys and reported domestic shipments of 300,000 units last year, placing it in the first tier; Ye Jie of PandaaX, a one-year-old startup making AI companion toys for adult women; and Liu Daixuan of Saibo Chuangli (赛博创力), which makes AI designer toys by combining AI with character-based intellectual property (character IP) and cultural-creative content.
The three companies were at different stages and worked in different categories. Yet by the end of the discussion, their conclusions were strikingly consistent: the hardest part of going global is not finding an opportunity. It is whether your understanding and capabilities are sufficient.
Many teams understand going global as “going overseas to find a new market.”
Liu Daixuan shattered that illusion in one sentence: opportunity determines whether you can enter a market; capability determines whether you can stay.
“Opportunity determines whether you can enter a market; capability determines whether you can stay.”
Selling AI hardware today has long ceased to mean selling only a piece of hardware. What you are selling is the entire set of services behind it: software, upgrades, new features, and after-sales support. Taking a hardware product overseas is emphatically not a matter of replacing Chinese with English and finding an overseas distributor. Network environments differ. Certifications differ. User habits differ. Languages differ. As Liu put it, truly going overseas is more like picking up the entire shop—pots, pans, and all—moving it to an unfamiliar country, and opening for business all over again.
Tim broke the chain down in more detail. Haivivi has already been operating in European and American markets for some time. His takeaway is that going global is not something a company can accomplish by doing just one part well. Supply-chain delivery, taxation, legal affairs, compliance, product localization, operations, and customer service make it an end-to-end delivery challenge.
What is the hardest part of that chain? Tim’s answer was localization. Servers need to be deployed locally, certifications need to meet local requirements, and privacy-protection requirements are especially stringent in Europe. Ultimately, this is not merely a technical issue. It comes down to how deeply a company understands local culture, regulations, and users in the United States and Europe.
Ye Jie added another point: no product can naturally satisfy the needs of every community worldwide. In his account, European users may care more about the cultural grounding behind a product; users in China, Japan, and South Korea may be more sensitive to emotional value and companionship in response to loneliness; and Southeast Asian markets may be more receptive to novel and entertaining experiences. Some products can be adapted for different markets. Others are not suited to extensive customization, in which case the company should first identify the market that fits it best.
Get this wrong, and the entire international-expansion chain may break.
Suppose you have thought all of that through. The product is ready. Where do you find your first overseas users?
The three operators took completely different approaches, but each was practical in its own scrappy way.
Ye Jie used the most basic method. His team contacted friends, former classmates, and host families in the United States. They took the product to people’s doorsteps and nearby yards, conducting grassroots outreach in small local communities so that real users could see it, take photos, and share it with people around them.
They began with seven people across seven cities and tested the product with more than 100 users.
It may sound unsophisticated, but the logic was clear: appearance is what first makes a user stop, while AI is what ultimately determines retention. First use the product’s look to make people pause; then let them experience it. Foot traffic in the United States tends to concentrate in certain places, so the team could later run small pop-ups in major shopping malls as well.
Liu Daixuan relied on trade shows to “filter naturally.” Plenty of people visit a booth simply to look around, but only a minority are willing to take a sample—perhaps even purchase a small batch to bring back to their own country—and work with the company on testing. At that point, the deeply engaged users reveal themselves.
That was how one of the company’s early overseas customers emerged. The customer took the product home and supplied extensive feedback about local language use, environmental differences, and many other issues. The country places a strong emphasis on education, so its compliance requirements for educational content are particularly strict. The customer also had many requests for new features. The two sides kept talking; the product improved, and the service system evolved with it. Only after serving a customer like this well can the company replicate that capability in the next market.
Tim took the opposite approach and began with strangers. His team warmed up the market on social media, publishing content in advance through official Instagram, YouTube, and Facebook accounts; gathered seed users and built a community; launched crowdsourced beta testing early, both to validate the product and retain those users; and supplemented this work with in-depth, in-person interviews at trade shows in Europe and the United States.
Their products were priced between US$149 and US$199, so the strategy was clear: focus deeply on Europe and the United States, and for now go especially deep in the U.S. market.
The greatest advantage of Chinese hardware is rapid iteration. But when a customer buys a smart hardware product, they are also buying the company’s promise that it will continue to support that product over the years ahead.
How can a company iterate quickly without “betraying” the owners of its previous-generation products?
Liu Daixuan gave the answer I liked most in the entire session: upgrading the hardware does not itself amount to betraying users. What does? Launching a product and then, three months later, leaving no one at the company responsible for it. That is the real betrayal.
“What really betrays users? Launching a product and then, three months later, leaving no one at the company responsible for it. That is the real betrayal.”
Their current approach is to reserve hardware capacity in advance for future features. The first generation may not expose all of that capacity, but the company can gradually release it later through over-the-air updates. At the same time, the team is building a product ecosystem in which the earliest and newest generations coexist. Different devices can even talk to one another. An older product does not lose all of its value simply because a new one has launched.
Ye Jie poured cold water on the question from the perspective of repeat purchases. Based on his company’s own data and its analysis of peers, he said that unless an upgrade is “truly necessary,” it does little to stimulate repurchases. Slight changes in appearance or a little more battery capacity do not persuade users. Even many co-branded editions may not generate much real repeat purchasing.
What counts as a necessary upgrade? Tim’s first-generation product used push-to-talk interaction: users had to hold down a button to speak. The second generation introduced end-to-end natural conversation, allowing users to interrupt at any time and enabling multiple people to talk with the device. An improvement that users can perceive directly counts. Other examples include moving from Wi-Fi to 4G, or changing an exterior that could not be replaced into one that is detachable.
By contrast, a new internal chip may make the specification sheet look better while leaving the user experience entirely unchanged. It is like someone telling you that your car now has a stronger screw when pressing the accelerator still feels exactly the same. That kind of upgrade has little meaning.
Put simply, there is one test for whether an upgrade is necessary: can the consumer directly experience something better?
Ye Jie’s principle is to make every over-the-air update as compatible as possible with existing sensors and functions, and at a minimum to keep meeting the expectations the customer had at the moment of purchase. Even if services such as repairs, cleaning, or trade-ins carry a fee, giving users a clear channel is better than leaving them to find a solution on their own.
Sales channels generate a large volume of data: conversion rates, search terms, reasons for returns, and reviews. How can these fragments be turned into product insight?
Liu Daixuan gave an example. One user left only a single comment: “It responds too slowly.”
If you look only at that sentence, all you have is an emotion. Dig deeper: is the problem the network, the backend pipeline, or the model? Which part of the chain is actually causing the delay?
Search terms reveal what users really want. Conversion rates show whether you have clearly communicated the product’s value. Channel data is not there to create an attractive report about how much you sold or how much you grew. Its real value is in guiding the next generation of the product.
Tim’s team holds a monthly cross-functional product meeting. If software can solve an issue behind a negative review, the team improves it through an over-the-air update that same month. Areas that receive especially strong positive feedback help determine where the next generation should go. In addition to data from Amazon, TikTok, and the company’s independent online store, the team speaks in depth with mid-tier and smaller key opinion leaders. It gives these KOLs new products for early testing; they film unboxing videos and compare the products both horizontally with competitors and vertically across generations. Because they are serious power users themselves, their feedback is highly valuable.
Ye Jie’s team goes even further. Its community has more than 1,000 members, and the operations team has a KPI: conduct at least 25 in-depth interviews every month, each lasting more than half an hour. The team actually has to send the device to a user, let them try it, and then talk in depth afterward: How did it feel? What problems did you find? What do you most want it to do? What score would you give the product today?
Users make all kinds of unexpected requests, and the product team has to decide for itself which ones to pursue. Beyond interviews, the team also examines comparable products: which other SKUs has this user purchased, what keywords have they searched, and what types of KOLs do they follow? When a user asks for a feature, the team first checks whether competing products offer it, whether those companies advertise around it, and whether they have bought the relevant search terms before deciding whether to follow suit.
One final question: as hardware specifications converge, why should users keep choosing your product?
Ye Jie was direct: hardware can be copied quickly, and software even more quickly. In Shenzhen especially, differences based purely on hardware can be closed with ease.
That is why the most important task for a startup in its first stage is to communicate the product’s value in the lowest-cost, most direct way possible. The moment users scroll past a video, they should be able to imagine what life would be like if they owned the product. That determines whether the company can survive and win its first battle.
Only after surviving does a company earn the right to think about the long term. What matters most over the long term? Ye Jie’s answer was character IP. Its importance in consumer industries has been demonstrated repeatedly, and it is something users perceive directly. Change the chip, change the storage, or train a model with however many billions of parameters—the ordinary consumer may feel nothing. But character IP can directly create the hunger of “I really want this,” the pride of showing ownership, and a sense of identity.
Intellectual property also takes time and carries enormous uncertainty. It is genuinely not something that money alone can guarantee.
Tim agreed that hardware gaps will narrow, but he is betting on two other things: product positioning and a brand’s digital assets. Does the product solve a real pain point in a specific use case, and which audience does it serve? Has the company built its own communications channels through public relations in mainstream media, a social-media matrix, and influencers who speak about the brand consistently? Follower counts and user relationships are difficult to buy in a short period of time.
At the end of the roundtable, Liu Daixuan offered a three-part formulation:
“The product is how we fulfill our promise to users. The channel is how we communicate that promise. The brand is the result we ultimately earn after fulfilling that promise over the long term.”
As the session ended, I kept thinking about that sequence. Many companies reverse it: first they spend heavily on branding, then they lay out their channels, and only at the end do they go back and repair the product.
What these three people were really saying was the same thing: promises can be copied, and specifications can be matched, but no one can take away what you build by keeping those promises year after year.
Tim, Head of Overseas Brand, Haivivi (跃然创新)
Ye Jie, Founder, PandaaX
Liu Daixuan, Co-founder and CTO, Saibo Chuangli (赛博创力)
Huang Jingrui (Jerry), Vice President, Unique Capital (非凡资本)
Editor’s note: The Chinese source identifies Ye Jie’s company as PandaaX, while the official event agenda lists 努努AI. This rendition preserves the name used by the source and does not infer whether or how the two names are related.
Huang Jingrui: This discussion is not only about how to sell products overseas. It is also about how Chinese hardware companies make the transition from “taking products overseas” to sustained global operations. Before we begin, I would like each of you to briefly introduce yourselves.
Tim: Hello, everyone. I’m Tim from Haivivi. We mainly work in AI toys and robotics. Last year, we achieved what we consider a strong result: our domestic shipment volume reached 300,000 units, placing us in the first tier. What problem does our product solve? Put simply, it lets the stuffed toys in your home that could not previously speak come to life and talk. It connects to large language models, including Doubao, and can keep children company and give them a happier childhood. We are also expanding overseas and have already achieved some encouraging results in Europe and the United States.
Ye Jie: Hello, everyone. I’m Ye Jie from PandaaX in Hangzhou. Our team mainly makes AI companion toys for adult women, with an emphasis on full-ecosystem companionship—from hardware and software to portable, desktop, and other form factors. We are also a startup founded this year. We have brought a demo to the area outside the venue today; anyone interested can try it and get a more direct sense of what we are doing in companionship.
Liu Daixuan: Hello, everyone. I’m Liu Daixuan from Saibo Chuangli in Beijing. We mainly make AI designer toys, combining AI with character IP and also with cultural and creative content to produce related cultural products.
Huang Jingrui: The first question is fairly broad. Does the globalization of Chinese hardware begin when a company discovers an overseas opportunity, or only after it has already validated a globally capable organization? Many teams understand going global as “going overseas to find a new market,” but establishing a stable presence abroad actually tests a comprehensive set of capabilities. What factors should a hardware company consider most seriously when going global? Tim, let’s start with you.
Tim: Let me start with the conclusion. I think going global presents an enormous opportunity for Chinese hardware companies and Chinese brands. China has very strong production and manufacturing capabilities, and it is also at the forefront of global competition in AI. All of this creates many opportunities for Chinese companies. But not every company can seize that opportunity, because the international-expansion chain is extremely long. It involves supply-chain delivery and the construction of organizational capabilities, including taxation, legal affairs, compliance, product localization, operations, and customer service. This is not a matter of doing one part well. It is an end-to-end delivery challenge.
Huang Jingrui: If you had to choose one, which part of that chain is currently the hardest to solve?
Tim: At present, I think the harder issues are product localization and understanding the local market. Overseas users think quite differently from domestic users. If a product is going to the United States, for example, the company first needs to address cultural and perception differences. Servers may need to be deployed in the United States, and the relevant certifications need to meet local requirements, including the tax system and laws and regulations. The same is true in Europe. Requirements in areas such as privacy protection are particularly high in these markets. So this is not merely a technical or product issue. More importantly, it depends on how deeply the company understands the culture, laws and regulations, and users in the United States and Europe.
Huang Jingrui: Mr. Ye, what do you think?
Ye Jie: Based on our own product design and some of my previous experience living overseas, I think the most important link is scenario fit.
A product cannot naturally satisfy the needs of different communities, ethnic groups, and social environments around the world.
When we were discussing the North American market, we used an example. We thought: American high school students start school at 7 a.m. and finish at 3 p.m. Don’t they have a lot of free time after school? Our companion toy could keep them company.
It sounded perfectly logical.
Then we thought about it more carefully and realized something was wrong.
First, only-child households are relatively uncommon in the United States, so teenagers may already have older or younger siblings for company. Second, American high school students usually have all kinds of activities after school. Getting out at 3 p.m. does not mean going home to sit alone; they are more likely to play sports, participate in clubs, or spend time with friends.
So we often unconsciously use our own lifestyles and the consumption habits of domestic users to reason about overseas markets, and that can easily lead us astray.
Determining how a specific product can work in a local use case requires time for research and analysis.
For example, European users may care more about the cultural grounding behind a product; users in China, Japan, and South Korea may be more sensitive to emotional value and companionship in response to loneliness; and some Southeast Asian markets may be more receptive to novel and entertaining experiences.
Every market asks something different of a product.
Some products can be adapted for different markets and ultimately become truly global. But some products are not suitable for extensive customization. In that case, you should first identify the market that fits your product best.
I think this is one of the most time-consuming parts of the globalization chain, but it is also something that must be done correctly. Get it wrong, and the entire international-expansion chain may break.
Huang Jingrui: Mr. Liu, what do you think?
Liu Daixuan: My view is this: opportunity determines whether you can enter a market; capability determines whether you can stay.
Selling AI hardware today no longer means simply selling a piece of hardware. It means selling the entire set of services behind that hardware.
Those services include software, subsequent upgrades, new feature updates, and after-sales support.
So taking a hardware product overseas is emphatically not a matter of replacing Chinese with English and then finding an overseas distributor.
Network environments differ overseas. Certifications differ. User habits differ. Languages differ. There are problems at every link in the chain.
That is why I think truly going overseas is more like moving your entire product and service system abroad and validating it all over again.
It is not a matter of copying a system that already worked domestically and dropping it into another country unchanged.
Huang Jingrui: Put simply, you are not just extending the sales chain; you have to bring the entire system behind it overseas as well. So here is the second question: when you first launch abroad, how do you find your first customers? Let’s start with Mr. Liu.
Liu Daixuan: For us, the first overseas users do not necessarily have to be the largest customers. We place greater value on users whose needs are especially clear and who provide extensive feedback.
We met one of our earliest overseas customers at a trade show.
After taking our product home and using it locally, the customer gave us a great deal of feedback, including comments on local-language use, environmental differences, and many other issues.
Because that country places a strong emphasis on education, it has particularly stringent compliance requirements for educational content. The customer also had many requests for new features.
The two sides kept discussing them.
In the process, we did not merely improve the product; the entire service system evolved along with it.
Only after you truly serve a customer like this well can that entire capability be replicated for other overseas customers.
Huang Jingrui: But many people at trade shows are only there to look around. How do you judge who is genuinely worth investing in as an early user?
Liu Daixuan: It is actually quite obvious. Many people come to a trade show to learn about a product, but only a minority are willing to take a sample—perhaps even purchase a small batch to bring back to their own country—and work with you on testing.
By that point, a group of deeply engaged users has effectively filtered itself out.
Ye Jie: Our method is relatively simple.
Because our product is a toy, with a stuffed toy as its physical form, we have always believed that appearance is the first thing that attracts a customer, while AI is what later determines retention.
At the beginning, we contacted friends, former classmates, and host families in the United States and asked them to help us conduct some very basic tests directly.
We literally took the product to their doorsteps and nearby yards and did something like grassroots outreach in small local communities. We let real users see it, take photographs, and share it with people around them.
At first, we found about seven people across seven cities and tested the product with more than 100 users.
Later, we also conducted online tests based on the feedback from this group.
In addition, foot traffic in the United States tends to concentrate in certain places, so we also considered running small pop-ups in major shopping malls.
Our logic was to use the appearance first to make users stop, and then let them experience the product.
That was how we acquired our early users.
Huang Jingrui: So during the initial launch, you still relied on some local resources overseas.
Ye Jie: Yes.
Tim: We took the opposite approach and started with strangers.
Seed users are extremely important because they can help you complete early testing before the product officially launches. After that testing, you can move into formal shipment.
We generally use several methods.
The first is social media.
For example, we publish warm-up content in advance through official Instagram, YouTube, and Facebook accounts. We first gather a group of seed users, then build and operate a community, and collect their opinions and feedback.
Second, we can launch crowdsourced beta testing in advance.
The testing itself helps us validate the product’s feasibility while also retaining a group of seed users.
Third, we use overseas trade shows, including events in the United States and Europe, to conduct direct and sometimes quite in-depth interviews with users in person.
Those are essentially the main methods.
Huang Jingrui: Based on your own experience, which countries’ trade shows deliver higher conversion rates?
Tim: It depends on the product’s price range.
Our products are priced at roughly US$149 to US$199, so for now we are mainly concentrating on Europe and the United States.
Within those markets, our current priority is to go especially deep in the United States.
In the U.S. market, user acquisition is currently still centered on social media.
Huang Jingrui: The third question has to do with “China speed.” One of the biggest strengths of Chinese hardware is rapid iteration, but when users buy a smart hardware product, they are also buying its compatibility over the next few years and the company’s commitment to service.
So when a product is iterating quickly, how do you make sure you do not “betray” users of the previous generation?
Tim: That is a very good question.
Our product has now reached its second generation.
The first-generation product used push-to-talk interaction.
For example, a child could press and hold the product, ask it a question, and have it tell a story. It could tell the story in voices modeled on particular character IP. If the child had a question in the middle, they could press and hold again to interrupt it and then begin the next round of conversation.
So the first generation was still essentially push-to-talk.
With the second generation, we made a relatively major hardware upgrade and moved to more natural, end-to-end conversation.
You can interrupt it at any time, multiple people can talk with it, and it can carry out more tasks.
The more important upgrades in this generation came at the hardware level.
For software upgrades, we rely more on monthly OTA updates.
That is our basic approach: hardware upgrades need to create real differentiation.
After the first-generation product launched, we inevitably received a large volume of user feedback, both positive and negative.
For problems that could only be solved through hardware, we made changes in the second generation.
At the same time, large language models have continued to improve, and we have made deeper optimizations to the product as overall model capabilities have developed.
Ye Jie: My thinking is fairly similar to Tim’s.
Our own principle is that compatibility must be done well.
Hardware upgrades are unavoidable. They are also a necessary path for the continued evolution of both a product and a company.
What you can really do, therefore, is make every OTA update as compatible as possible with the original sensors and functions, so that existing users can continue using the product.
At a minimum, you have to meet the expectations users had already formed at the moment they originally bought the product.
Only then can you maintain the relationship between the brand and its users.
You can also provide users with additional value-added services later on.
Even if those services are paid—for example, repairs, cleaning, or trade-ins—you are at least giving users a clear channel, which is better than leaving them to find a solution on their own.
So software compatibility must be done well, while the hardware still has to continue evolving.
Huang Jingrui: Then how much impact do hardware upgrades have on repeat purchases?
Ye Jie: Based on the data we have seen so far, including some analysis of peers, an upgrade has little impact on repeat purchases unless it is “truly necessary.”
This category has only really taken off in the past year or two, and we can now see products that have gone through two or three generations.
If the only changes are a slight modification to the appearance or a little more battery capacity, these small upgrades provide very little stimulus for repeat purchases.
The same is true of many co-branded editions: actual repeat purchases may not be particularly strong.
Huang Jingrui: So how do you define a “necessary upgrade”?
Ye Jie: Take the example Tim just mentioned: moving from push-to-talk interaction to genuine end-to-end natural conversation is a very direct upgrade.
Another example would be moving from an ordinary Wi-Fi version to 4G.
Or perhaps the product’s appearance could not be changed before, but is now detachable and replaceable, immediately creating many more ways to use and play with it.
These are necessary upgrades because users can perceive them directly.
But if you merely replace a chip internally and the specifications look higher while consumers perceive no difference at all in actual use, then I do not think that upgrade is as meaningful.
So the most important question is whether consumers can directly experience something better.
Huang Jingrui: Mr. Liu, what do you think?
Liu Daixuan: I do not think a hardware upgrade in itself amounts to betraying users.
What really betrays users?
A product launches, and three months later no one at the company is looking after it.
That is the real betrayal.
In our early days, we also had push-to-talk interaction in plush toys, and later upgraded to a base-mounted product.
When we design products now, we reserve some hardware capabilities in advance for functions that may be needed in the future.
Not all of them necessarily need to be enabled in the first generation, but those capabilities can be released gradually later through OTA updates.
In addition, we are now building out the broader product ecosystem.
For example, our earliest-generation product and our latest product can be placed in the same ecosystem.
Different devices can even talk with one another, and several products placed together can communicate as a group.
Through this approach, we hope an earlier-generation product does not completely lose its value as soon as a new product comes out.
Huang Jingrui: Which overseas countries are your main markets at the moment?
Liu Daixuan: Mainly Japan, South Korea, and Southeast Asia at present.
Huang Jingrui: Do you make different hardware or software adaptations for different regions?
Liu Daixuan: Yes.
For example, server deployment may differ between Japan and South Korea and Southeast Asia.
For Southeast Asia, we may choose a Singapore node, and we adjust the entire network path and server setup according to the market.
Language is another very specific issue.
Different regions have accents and dialects that can create large differences in speech-recognition performance, so we optimize for those local languages.
Users in Japan and South Korea may care more about character IP, anime culture, and whether the voice truly matches the character.
They care a great deal about one question: is this the voice I imagined for that character?
In some other markets, users may care less about those things. What matters more to them is whether the product can understand their questions accurately and then answer them accurately.
The central concern can be completely different from one region to another.
Huang Jingrui: Markets such as Indonesia and Malaysia also involve compliance issues.
Liu Daixuan: Yes.
I just returned from Malaysia a few days ago.
We discuss using local models and local cloud services directly with local companies, keeping the data local as much as possible.
That is because data compliance has to be considered in this area.
Huang Jingrui: So repeat purchases of hardware should not be built on “locking users in.” They should be driven by the product’s continued evolution, so that users willingly choose it again.
Huang Jingrui: The fourth question is about channels.
Platforms actually provide a great deal of data: conversion rates, search terms, reasons for returns, reviews, and so on.
But this consumer feedback is usually highly fragmented.
How do you turn channel data into genuine product insight?
Tim: Data integration across many online channels is already relatively mature.
On Amazon and TikTok, as well as on a company’s own website, you can see data such as click-through rates, conversion rates, purchase conversion, impressions, and traffic very clearly.
Positive and negative reviews are especially important.
Every month, we hold a cross-functional product review meeting.
If an issue reflected in a negative user review can be solved through software, we can optimize it promptly through an OTA update.
If positive reviews cluster particularly strongly around certain points, or if a competitor performs very well on certain metrics, that can also help us decide how the next generation of the product should evolve.
The development of AI tools is also helping a great deal with this work.
For example, WorkBuddy and Codex can help us integrate data from different sources and then explore it in greater depth.
That can provide extremely useful guidance for subsequent product iterations.
Huang Jingrui: Of all the fragmented data the channels provide, which kind is most valuable for product iteration?
Tim: I think it ultimately comes back to users’ real voices.
Every month and every quarter, we invite users to take part in relatively in-depth interviews.
We also look at every piece of feedback from real users on channels such as Amazon and TikTok.
We collect feedback from social media platforms including Instagram, YouTube, and Facebook as well.
At the same time, we have in-depth conversations with external mid-tier and long-tail KOLs.
During the early stage of a new product, we give units to some KOLs for beta testing. They film unboxing videos and make both horizontal and longitudinal comparisons with other products.
These people are deeply engaged users themselves, so their feedback is extremely helpful for subsequent product upgrades.
In short, we collect the voices of real users and opinion leaders from different dimensions and directions.
Huang Jingrui: What kind of user feedback makes you feel that you have to conduct another in-depth interview? Can you give an example?
Tim: Take the push-to-talk interaction in the first-generation product.
Users would ask: why do I have to press and hold to talk? Why can’t I just have an end-to-end conversation directly?
That question is actually open to debate.
Push-to-talk has its advantages. It gives the interaction a clear, tactile sense.
End-to-end conversation is more convenient: there is no need to press anything; you can simply speak.
For questions like this that directly affect the product’s interaction logic, we conduct relatively in-depth user interviews early on before deciding exactly how to change it.
Ye Jie: We may not yet have as much experience as Tim on this question, because our product is still in its first iteration.
But we are also conducting in-depth user interviews very actively.
Our community currently has more than 1,000 people.
We have set a KPI for our community operations team: every month, they must complete at least 25 in-depth interviews, each lasting more than half an hour.
You have to actually send the device to a user and let them use it.
After they have used it, you talk with them in depth:
How did it actually feel?
What problems did you encounter?
What do you most want it to do?
What score would you give what it can do today?
That is one of the questions we ask users most often.
You find that users really do come up with all kinds of unexpected requests.
At that point, the product team has to make its own judgment: should this request actually be implemented?
Every user is an individual, after all.
In addition to user interviews, we also look at data from benchmark products.
For example, what other SKUs has this user bought?
What keywords have they searched for?
What types of KOLs do they follow?
We analyze all of those things.
If a user makes a request, we also investigate whether benchmark products have already implemented it, whether they have run campaigns around that request, and whether they have purchased the relevant search terms.
Only then do we decide whether we should implement it ourselves.
That is roughly the internal decision chain we use.
Huang Jingrui: How do you currently look at your channel-side data?
Ye Jie: At this stage, our main domestic channel is JD.com.
JD.com has its own integrated procurement-and-sales system and is also responsible for paid traffic acquisition, so we discuss these matters with the platform team.
One particularly direct data source for us right now is search terms.
What terms do users search that lead them to your product?
Which keywords ultimately produce a purchase?
Mature e-commerce platforms generally provide fairly comprehensive reports on these questions.
My own background is also in data, so whether I analyze it myself or use AI-assisted analysis, this process is already fairly standardized internally and does not take a particularly large amount of time.
Liu Daixuan: Channels generate a great deal of data. The key is how to reconstruct it as a user problem.
Search terms, for example, essentially reflect what users actually want.
Conversion rates may reflect whether you have explained the product’s value clearly.
The same applies to after-sales support.
A user may leave only one comment: “It responds too slowly.”
If you look only at that sentence, all you may get is an emotion.
But if you investigate more deeply, you need to ask:
Is it a network problem?
Is it a problem in the backend chain?
Is it a model problem?
Exactly which link is causing the slow response?
Channel data is therefore not meant to produce an attractive report saying how much you sold or how much you grew.
Its real value should be to support the iteration of the next generation of products and new features.
Huang Jingrui: Is there any kind of channel feedback that directly prompts you to upgrade the software?
Liu Daixuan: There are mainly two kinds.
The first comes from highly engaged users whom we contact one-on-one ourselves.
At the startup stage, we establish direct contact with some users.
They can tell you about problems particularly quickly.
For example, a service may suddenly stop responding, or a character may go OOC and move beyond the characterization established by the original character IP. That feedback is very direct.
The second kind comes from channel customers.
For example, after an overseas partner sells the product to local consumers, they may tell us that a certain function is particularly well suited to the local application environment, or that local users urgently need a particular capability.
In that case, we discuss it with the channel, turn it into a new direction for product iteration, and then actually build and launch the function later on.
Huang Jingrui: One final question. If the technological gap between everyone’s hardware eventually becomes fairly small, why should users choose you? Does the answer lie in the brand, the channel, or the product? And which brand assets take years to build and remain difficult for competitors to replicate quickly, even if they invest heavily?
Tim: Especially in Shenzhen, where the hardware industry is highly developed, gaps at the pure hardware level can close very quickly. I think real differentiation comes more from two things: product positioning and a brand’s digital assets.
First, product positioning. Does your product actually solve a consumer pain point in a specific use case? That is critical.
Second is audience fit. Exactly whom are you serving?
Third is the brand’s digital assets. By digital assets, I mean whether you have truly built your own brand-communication channels—for example, PR in mainstream US media, your own social-media network, and sustained advocacy from influencers. Have you consistently communicated the company’s ideas and brand elements over time to the people you genuinely want to reach? That is an enormous asset.
And then, of course, there are the channels. Platforms such as Amazon, TikTok, and AliExpress are becoming increasingly open to ordinary hardware companies. Ultimately, the product, brand communications, and channels need to form a complete loop.
Huang Jingrui: What you just described was mainly the brand’s communication channels. But when it comes to something that truly takes many years to build and that competitors cannot quickly replicate simply by spending money, what do you think it is?
Tim: First, it is still the technology path. The technology path a product chooses at the outset is extremely important.
Second are the brand’s digital assets we just discussed, because digital assets really are difficult to overtake in a short period of time. Long-term brand accumulation, follower numbers, and user relationships all take time.
Ye Jie: Hardware can actually be copied very quickly, and software can be copied even faster. So I think the most important thing for a startup in its first stage is figuring out how to communicate the product’s value to users in the lowest-cost, most direct way. That determines whether you can survive first and whether you can win your first battle.
Once the company has genuinely survived and can move forward over the long term, accumulating character IP becomes particularly important. In fact, all three of our companies are building our own character IP. But character IP takes time, and it also comes with enormous uncertainty.
So the first stage is still about getting the product right. How does your product directly address a user pain point? How do you make it possible for users, the moment they encounter a video, to imagine what their lives would be like if they owned the product? That is extremely important.
Only after you execute well in the first stage and the company survives do you have the opportunity to discuss longer-term character-IP accumulation and deeper technological barriers—for example, using user data to build your own fine-tuned or vertical models, or finding cheaper storage and inference solutions to reduce costs further. Those things come later.
At the beginning, the most important thing is simply this: how do you sell the product first? We have followed this logic ever since the team was founded.
Huang Jingrui: If a company has already moved beyond the first stage, which brand asset is most worth concentrating on building next?
Ye Jie: If I could choose, I would still say character IP is the most important. Its importance has been demonstrated countless times in consumer industries.
Models, chips, and storage are certainly all forms of technical accumulation for a company. But ordinary consumers do not perceive them as directly. Whether you replace the chip, change the storage, or train a model with however many billions of parameters, ordinary users will not immediately notice.
That is why character IP and community matter more. Can you operate the community well? Can you create a sense of “I really want this” among users—or, after they own it, a desire to show it off and a sense of identity? Those feelings have to be established first.
If a character IP genuinely succeeds, it becomes a very important foundation for the company’s long-term operations, and it is also one of the hardest things to replicate. Technology can be accumulated gradually. But whether a character IP will ultimately succeed is genuinely not something that can be guaranteed simply by investing money.
Huang Jingrui: And you, Mr. Liu?
Liu Daixuan: I think hardware specifications and technologies will inevitably become more similar over time. What is genuinely difficult to replicate is the overall experience ultimately delivered to the user.
Users do not pay especially close attention to which chip or model you use. What they actually experience is:
Is it fast?
Are its answers accurate?
Is it stable?
Can they keep using it over time?
Will someone take care of problems when they arise?
Will there be further feature upgrades?
So, from this perspective, I would put it this way:
The product is how we fulfill our promise to users; the channel is how we communicate that promise; and the brand is the result we ultimately earn after fulfilling that promise consistently over the long term.
Huang Jingrui: Then, speaking specifically about brand assets, what do you think matters most?
Liu Daixuan: As we discussed earlier, I do think character IP is extremely important—especially because our industry inherently combines AI with character IP. A strong character IP does more than influence the brand; it can also expand sales channels and create more room for the entire product to extend into new forms.