FoloToy / Hunban Technology: Is AI Hardware the Next Consensus Trade?
Summary
- FoloToy is betting that AI will let users define hardware, rather than add a chat feature to standard devices. Wang Le wants products to ship when they are 65% complete, potentially falling to 45% over time, with users finishing the software and services themselves; the hardware and industrial design still need to meet consumer standards. His core thesis is that as AI lowers the cost of making software, demand for personalized hardware will keep growing: “Everyone is a developer; everyone is a creator.”
- AI Passport’s key assets are shifting from the device itself to its users, use cases and creation loop. The community has more than 1,300 use cases, while the company has built only about 10 simple demos; creators account for roughly 5% of registered users. Wang concedes that the hardware has “no advantage” on specs alone. Copying a circuit board is one thing; replicating a community willing to create, share and promote on its own is another. He points to Nintendo to explain the difference.
- FoloToy’s business model is taking shape around one-time hardware sales, free consumer services and B2B event purchases. Wang attributes the low price to DTC distribution, which cuts out channel costs, and lets users choose their own software and servers. He declines to disclose gross margin, and the company is still working out how developers can make money. Event passes, employee devices and hackathons already provide real buying occasions; the product is more than an “all-purpose little card” for hobbyists.
- Context capture is the hardware value participants most agree on, but privacy can undermine the quality of that context. Wang says recording devices are the ones he has kept using; Raymond argues devices should be worn unobtrusively, without requiring a daily decision to bring them along. Yet they acknowledge that knowing someone is recording makes people more guarded, especially around cameras. Zhen Di says his team chose not to give parents full access to children’s chat histories, sharing only weekly keywords—a sign that more complete capture is not always better.
- The hurdles for a general-purpose AI phone come from both closed platforms and the attention-based economics of the internet. Mark Tang questions whether a device meant to handle every scenario will get blocked by access restrictions across food delivery, social and other platforms. Wang adds that Agents save time and skip ads, which conflicts with platforms’ goal of keeping users engaged. Browsing has value in its own right, too; the discussion’s view that JD.com is easier for Agents to displace than Taobao rests on that distinction.
- Hunban Technology divides embodied AI into high-tolerance experiences and low-tolerance tasks, prioritizing B2B use cases that can generate repeat purchases. Branded robots can tell jokes and give tours; robot dogs tackle “3D” tasks that are dull, dirty and dangerous. Zhen Di says the customs project went through an extended testing POC, with robot dogs used for port patrols; one customer bought 5 for a single site. He sees customs applications as replicable nationwide. That does not mean general-purpose household robots are ready.
- Companionship does not depend on language ability; asked whether Labubu should talk, Zhen Di’s answer is “no.” Touch, the desire to care for a character and room for imagination can stand on their own; adding AI may actually damage the experience. IP partnerships also involve long-term legal work and questions about how a character is defined. Wang says FoloToy has prioritized tools for stories, dictation, time management and guided tours; the emotional satisfaction of creating a product is an “extra.”
- Both companies have clear plans to expand overseas, but Wang sees it as “starting a second company.” FoloToy plans to prepare its overseas business in Q4 that year, enter the US with a differentiated AI Passport product and return to CES in 2027. Hunban Technology plans to exhibit with overseas partners while navigating geopolitical restrictions already emerging in embodied AI. China’s supply chain offers a fast prototyping base; overseas sales, deployment and ongoing operations still need to be built from scratch.
Deep dive
1. AI hardware is shifting from manufacturer-defined to user-defined
Asked why consumers need new hardware when phones already have so much AI, Wang Le reframes the product: “The point isn’t to run AI inside the hardware. AI should define or control the hardware.” FoloToy calls this approach “user-defined AI hardware.”
Wang recalls that when he spoke with Zhen Di in 2024, he was already working on self-hosting: supplying the hardware and software while leaving configuration to users. The business logic has not fundamentally changed; AI has simply made it easier for ordinary people to modify and customize products.
Raymond asks whether more customization also raises the barrier to getting started. Wang acknowledges that standard products will remain, but says people who once could not hack a product can now modify it with a sentence. The pool of users seeking personalization is growing with them.
2. A 65%-complete product leaves room for users to create
Wang wants products to ship at 65% completion, potentially moving from 95% or 85% to below 50% over time—say, 45%. Users would finish some configuration and setup themselves. Individual developers, small businesses and large companies could then turn the product into a solution tailored to their needs.
“Semi-finished” does not mean skimping on hardware quality. Wang wants consumer-grade design and hardware, but sees no need to preload every service, from Ximalaya to Tencent Music. The product can leave interfaces open so users can connect Baidu Netdisk for their family’s or children’s needs, then use AI to process the content.
The pitch initially struggled to land with some B2B manufacturers, so FoloToy tried selling directly to consumers. Wang says the results showed the model could work. Zhen Di adds that AI is evolving so quickly that manufacturers who aim for a 100%-finished product may struggle to keep up.
3. Everyone can build—and older parents are joining in
On the popular claim that this is the “last Programmer’s Day,” Wang prefers a different formulation: “Everyone is a programmer now.” Programming is no longer confined to a profession or industry; more people are making the things they need themselves.
Wang points to an example inside the company: at an event in Beijing, an on-site support person with no development experience—a product manager—was able to handle technical support. He says that would not have been possible a year earlier.
On the AI-native generation, Wang says sixth-graders are making 3D games and other projects themselves. He estimates that a child might consume 5 billion tokens over a weekend; this is his estimate of children’s usage, not a consistent weekly measure.
More surprisingly, buyers are getting older. Wang says people older than he is are buying the products and learning to use Agents and coding tools to build things for their children. Compared with younger users’ natural adoption, they may be “more eager to change things,” and get a strong sense of accomplishment from finishing a project.
4. More than 1,300 use cases have turned a hardware card into a creator community
At the time of recording, Wang says the community had published more than 1,300 use cases. The company had built only about 10, mostly simple demos. User-created content has pushed the product far beyond what the company could design on its own.
Creators make up about 5% of registered users. Wang calls that share “very small,” while Raymond says it is already high, close to the single-digit creator share seen on content platforms. Wang insists on calling the community “100% AIGC”: users participate in the creation, but AI produces the work.
The community also breaks creation into steps: see what others have made, try it, copy a prompt to build a version of your own, then modify it in more depth. Wang wants users to enjoy both the act of creating and the satisfaction of having something worth showing.
The example that surprised Wang most was a user-made rhythm game in the style of Dance Dance Revolution, which he felt was close to Nintendo party games in polish. He contacts the creator directly and even encourages them to release it on iOS or Steam rather than stay on FoloToy’s platform.
5. Limited-purpose devices can hold attention more easily
Wang argues that the iPhone 4 already had enough hardware integration to meet many digital needs; today’s phones, however, are loaded with ads and complex systems. Small devices may be less powerful, but can make users feel “this was made for me” and focus their attention on one purpose.
Wang explains users’ sense of flow as a matter of subtraction: “No ads, no games, no push notifications, nothing.” Instead of juggling several AI tasks, choosing one of many use cases to stick with can restore sustained attention.
Physical devices also change how creations are shared. A website or phone app may still require a server or App Store listing; once loaded onto a small device, it can be handed directly to a friend or child. Raymond compares it to a McDonald’s Tetris handheld: the same game is available on a phone, but a single-purpose physical object still has appeal.
6. Hardware can be copied; communities take years to build
Asked whether Huaqiangbei could copy the product, Wang makes no claim that the circuit board is a moat: “On performance or specs alone, it has no advantage.” He asks whether Nintendo’s consoles have the best hardware. Product value is not determined by specs alone.
Wang sees FoloToy more as a place where users create content with AI, with the hardware serving as the medium. A competitor making a similar device would still have to pay for promotion; without users who share, participate and contribute over time, commercialization remains difficult.
On a competitor cutting prices in half, Wang says it would take heavy subsidies and does not rule out rivals with 10 or 20 times the funding. He cares more about whether the company keeps growing quickly month by month and year by year, and whether it can keep shipping products and stay in the game 3 to 5 years out.
7. DTC and user choice support low prices; gross margin stays under wraps
Raymond thinks AI Passport is very cheap and says that, in the US market, he might even double the price. Wang does not see a price increase as inevitable. He refuses to disclose gross margin: “That’s our trump card.”
Wang’s math is that DTC sales eliminate the need to leave room for intermediary channels, while users can choose whether to pay for software subscriptions and other services. After cutting out those costs, he sees no reason to raise the price simply to earn a little more.
Wang says the consumer service is entirely free; for now, the company mainly makes money from hardware sales. It is exploring ways for developers to earn money and hopes one use case might become a user’s “first bit of side income,” but has not described a mature revenue-sharing system.
8. The cyber badge turns B2B procurement into an interactive experience
Wang recalls that a creativity conference around August chose AI Passport as its event pass, where the product also made its debut. Hackathons are one use case, but the basic function is a badge: turning a static credential into a device that can run software.
Around the pass, the team built check-ins, games, points, prize redemption and effects synced to a large screen. Wang wanted to “turn walking around an exhibition into something that feels like playing a game,” rather than simply adding a screen to a conventional badge.
At WTCC, TiDB also bought devices and issued one to each employee for a hackathon. Other major tech companies bought passes and interactive experiences for their events, showing that the same hardware can be funded from event organizers’ B2B budgets.
9. Hardware is starting to be designed around what AI needs
Zhen Di sums up the old product model as human enhancement: glasses, headphones and phones that augment human capabilities. He is now more interested in “embodied intelligence”—giving an AI assistant or cyber companion a body and perception to create a “human-AI loop.”
In that framework, audio, vision and other sensors supply input to the Agent. “We can’t remember everything,” but AI can. As software is increasingly designed for AI, hardware should help it gain context from the physical world.
Zhen Di recalls that when the team worked on a POC for an AI toy in 2023, it was still dealing with ASR, TTS and latency optimization from WebSocket to RTC. Wang argued at the time that the future should be fully native and unembellished, with direct output—a point that left a strong impression on Zhen Di.
10. Recording is the use case that sticks; unobtrusive wear determines daily use
Wang has bought glasses, bands, watches and rings, but recording devices are the ones he has kept using because they build context. He even records conversations with his child and asks AI to assess whether he handled them well, including whether he was “especially preachy.”
He finds Feishu’s recording pod relatively easy to use, but from his own perspective, does not think a subscription is necessarily worth it: if users can connect their own server, the cost may not need to be so high. That reinforces his view that devices should let users define their own services.
Wang’s ideal device would have long battery life, recognize who is speaking and work continuously all day; he admits he has not found one that fits. Raymond says capture devices should blend into daily life so users do not have to decide each day whether to bring them. For people who already wear glasses, glasses may be a natural form factor.
Raymond points to the AI glasses he wears: removing the camera made them lighter and less intrusive, while they can still capture communications and conversations and provide navigation. He also likes that users do not have to rely entirely on the official app; they can code an OS on the web and add their own configurations.
11. The awareness of being recorded changes what people will say
Raymond turns the question around: once people recognize a recording card, pod or AI glasses, they know they may be on tape. The person recording may feel they are simply saving moments from daily life, but the person being recorded may hold back.
Zhen Di puts it plainly: “When you record others, you’re documenting your life; when others record you, you have to watch what you say.” Wang says he is generally reluctant to talk much around devices with cameras. Mark Tang agrees that knowing someone is recording makes you guarded and can make some topics impossible to discuss openly.
Wang says that telling people before a work meeting that you are recording to help with a later summary is usually acceptable. Recording someone you do not know well in a private, non-public setting can make the conversation superficial. Claims that anti-recording devices are selling well and that major companies have designated anti-recording rooms were presented by participants as hearsay or secondhand reports.
12. The children’s product shares keywords with parents, not full chat histories
Zhen Di says the team specifically considered whether parents should receive a copy of children’s conversations on Jijifei. Although full transcripts could help with review and safety, the team ultimately removed parents’ access to all conversations and shares only weekly keywords, helping parents spot interests or concerns.
In product design, Jijifei removed the camera and eye display, while keeping storytelling and flapping wings; the team also considered battery life and compute. Wang says his children’s products likewise have no camera. Some require users to manually start recording so they know when the device is listening.
13. Small companies can serve the niches large players leave uncovered
Wang does not want to spend too much time studying big-tech strategy: “Even if they want what I’m doing, I can’t make it.” He would rather serve a niche used by 10, 20 or 100 people, filling the gaps between the “soybeans” of big tech with “little green beans.”
Zhen Di thinks the gap between large and small companies is not as wide at the starting line in this phase. Capital and talent density are advantages, but they also bring execution and organizational costs. Both companies have B2B businesses, so working with big tech is not simply a matter of competing against it.
On the market’s future shape, Zhen Di currently expects applications to be less likely to produce a single, highly concentrated breakout hit; instead, many products may flourish. Underlying sectors such as GPUs may remain concentrated. He also thinks scrappy products from small teams may be no more likely to disappear than big-tech projects that go through frequent reorganizations.
14. A general-purpose AI phone first runs into ecosystem access limits
Mark Tang raises doubts about all-in-one products such as Doubao Phone. Once a product is a phone, users expect it to cover food delivery, ride-hailing, shopping, documents and every other scenario. But Chinese platforms have not opened up fully, leaving a gap between what the device can do and the tasks it can actually complete.
One example discussed: a phone claimed to integrate Meituan, but the integration was with Meituan Errand and could not place a real food-delivery order. If each platform builds its own AI, ecosystems and data may not connect automatically. Mark therefore thinks the all-in-one product could look like “a somewhat false boom” in the short term.
The view allows for change: a vertical platform might partner with an OS-level phone maker to gain a competitive advantage, loosening ecosystem constraints. The discussion does not claim that today’s closed platforms can never open up.
15. An Agent that saves time may not replace browsing people enjoy
Wang says the internet both helps users get things done and makes money from ads and time spent on platforms. AI that saves users time and skips ads runs into that business model. Taking control of the phone OS does not mean other platforms will surrender their position.
The pleasure of browsing is another counterexample. Some people enjoy scrolling through Taobao or browsing content on Meituan and Dianping; placing an order for them directly could be a “negative optimization.” Raymond reasons that JD.com, which is built around maximum efficiency, is easier for Agents to replace than Taobao. That is a projection from the discussion, not an outcome that has already occurred. Mark adds that Meituan and Dianping have content ecosystems and traffic monetization models.
16. Hunban Technology has focused on applications and deployment, not building the robot body
Zhen Di says the starting point has not changed from Jijifei to robotics: Cyber Partner means giving AI a body in the physical world. What looked like a consumer product early on was also a B2B collaboration with an IP company. Content and IP contributed value alongside sales and retention.
Enterprise customers are closer to today’s early adopters of embodied AI products: robots can cost hundreds of thousands of yuan, while robot dogs cost tens of thousands. By working with IP owners and robot manufacturers, Hunban Technology has built expertise in bringing AI into real-world settings and creating value.
The bodies of the two robot dogs on site come from different manufacturers. “Lin Shao” is the upper-body platform and control layer developed by Hunban. The humanoid robot “Mo Xiaosu” also uses a major manufacturer’s base body, then receives IP styling to become the branded character of Magic Space.
Zhen Di compares Micro Duck with AI Passport: both use products that are not fully defined yet to attract early users willing to participate. By contrast, he says people understand that a humanoid robot that does all the housework is “not something we can do yet.”
17. Different tolerance for failure separates companionship from dangerous tasks
Zhen Di divides the team’s work into two categories. Companionship has a high tolerance for failure and is relatively low-hanging fruit, but competitors have rushed in. Some products still have problems with the actual experience, retention and returns; the fact that they can all chat does not mean they have proven their value.
Mo Xiaosu represents a higher-tolerance robot application: a brand character, tour guide or purveyor of cold jokes. Users can accept that it cannot do many practical tasks because what they are buying now is interaction and display.
Robot dogs are moving into lower-tolerance tasks. The team is focused on “3D”: dull, dirty and dangerous. Zhen Di acknowledges that the team may be better suited to companionship, but says the B2B business grew gradually and gained some ability to fund itself early on.
18. Labubu should not talk just because the technology can make it talk
Raymond asks whether Labubu should speak now that small devices can connect to AI at low cost. Zhen Di’s answer is “no.” Cuteness, touch and emotional comfort are complete sources of value in their own right; the fact that technology can add a feature does not mean the product should have it.
Zhen Di compares it to a pet: when he is stressed, hugging his dog feels good because of the touch and companionship. “I don’t want it to talk to me.” Some products should bring out a desire to care for them, rather than turn them into better question-answering tools.
IP deals also involve real legal and timing issues. Zhen Di recalls that early contracts went through many revisions because no one had defined this kind of partnership before. Getting a positive response from Disney, Toei and other organizations does not mean the product is ready to launch immediately.
19. AI toys can become content products—or remain tools
Zhen Di’s experience is that doing companionship well can push a company toward becoming a content, entertainment, IP or design business. Jijifei can hold a conversation, but stories still account for a large share of its content; AI is more a way to deliver that content.
Wang stresses that FoloToy has never positioned itself primarily as a companion product. Its plush toys tell bedtime stories, its desktop products manage time and dictate spelling, and its museum devices provide guided tours. “Our products all have very clear tool uses.”
He therefore benchmarks the products against study devices costing a few thousand yuan and time-management tools costing a few hundred, rather than only against plush toys and collectibles. Wang acknowledges that parents feel satisfied when they define a product themselves, but treats that emotional payoff as an extra on top of its use as a tool.
20. Customs patrols validate the combination of a mobile platform and sensors
Asked about repeat purchases, Zhen Di says the team screens for repeatable use cases before running tests. He gives a scaled repeat-purchase example in patrols: one customer bought 5 robot dogs for a single site. The customs project went through an extended testing POC, using robot dogs to replace or assist with some port patrol work.
The port presents specific challenges: container stacks change quickly, conditions are exposed to sun and wind, and cargo opened after a month at sea may have all kinds of odors, including potential drugs or hazardous chemicals. The platform needs autonomous planning and navigation. Zhen Di stresses that it does not necessarily have to be a “dog”; the form factor can change.
The “Lin Shao” control layer handles monitoring, fleet control and connections. The upper body adds different senses: vision and LiDAR for observation and recording; audio for voiceprint recognition and measuring noise and construction decibels; gas modules for detecting drugs and hazardous gases. Different tasks require partnerships with different robot-body and module manufacturers.
In tests in Xi’an and with public security agencies, robot dogs can also broadcast messages, effectively giving them a “mouth.” Zhen Di sees the customs case as replicable nationwide. The commercial focus is on these specific tasks, not a general claim that robots can already do anything.
21. Going overseas means building the business again, not just shipping the product
Zhen Di plans to attend CES with overseas partners and says geopolitical restrictions have begun to emerge in embodied AI. He still believes the company must go overseas, but says value should be measured against the tasks being solved and local factors such as labor, social structure and social responsibility.
In Zhen Di’s view, China’s hardware advantage is iteration speed: after prototyping, teams can bring the product back at any time to update and adjust it. They may pilot in China first, then seek a balance between domestic and overseas markets when scaling. Overseas pricing and gross margin will also differ.
Wang attended CES in 2026 and plans to return in 2027. FoloToy ran an overseas sales push at the end of the previous year, with a good response, but after considering the whole business, he concluded that “going overseas is like starting a second company.” The domestic and overseas businesses need to be built in parallel.
Wang wants the domestic business to keep growing quickly while preparing for overseas operations in Q4 of that year, with AI Passport as the entry product for the US. He says the product for the local market will be differentiated, rather than a straight copy of the domestic offering.
22. A 1.8-meter sunflower can find a paying customer, too
Wang’s partner Guo Xinhua studied automation and previously worked at FANUC. The team made a roughly 1.8-meter-tall sunflower for trade-show display; several units were bought, at prices high enough to surprise them. The team later received orders for large custom mascots.
Wang does not present this as a breakthrough in replacing human labor: “Our large creations aren’t meant to replace people or work in dangerous settings; they’re really just decorative.” Their value is that they look good, are fun and draw visitors to the booth for photos. Zhen Di adds that these products do not need task intelligence; movement and interaction alone can meet a real use case.
23. Cyber Studio lowers the barrier to taking part in robot development
In addition to “Lin Shao,” Hunban Technology has released Cyber Studio. Zhen Di wants people without the budget, time or opportunity to buy a robot to DIY their own cyber companions on the platform and start training them in simulation.
Zhen Di divides robot capabilities into motion intelligence, interaction intelligence and task intelligence. Task intelligence is the biggest current gap and a long-term challenge. Different products can serve different customer groups, while still sharing parts of the underlying technology stack and team capabilities.
The team has also entered heavyweight boxing competitions, and URK2 has just made the round of 16. Zhen Di sees the competitions, platform and deployment tools as part of the same effort: enabling more people to participate in, create with and use AI. That is also the common thread running through the conversation between both guests.