Original · Unique Research · 2026-04-21
Editor's note: The first-person report and its judgments belong to the original Chinese author. This English rendition retains the opening essay, four founder narratives, and the full panel transcript. Product claims, technical descriptions, supply-chain accounts, and market comparisons are source or speaker claims, not independently audited findings. Company, personal and product names are transliterated where official English forms remain unverified. The source is dated April 21, 2026.
Unique Awards
On AI Hardware Going Global, They Told Some Truths That Don't Make It into the BP
AI hardware is sexy, but what truly determines whether you survive is often not model parameters, but agronomy, tariffs, aquatic plants, chip prices, and whether a button should be cut.
"
Some have been skinned alive by overseas tariff reviews and persistently high chip prices; others find the domestic scene too complex and treat landing overseas as a safe harbor for dimensionality-reduction strikes.
These two extremely polarized situations are happening right now in the AI hardware global-expansion track.
Large models are devouring everything, and software moats are getting thinner. So entrepreneurs are turning their gaze to hardware, trying to retain users through physical devices in this AI carnival. Wang Chaochao, partner at Unique Capital, said something very precise at the opening: "AI hardware pulls us from the virtual world on the screen back into the real physical world. It is very sexy, but the mortality rate is extremely high."
At a Hangzhou AI WEEK roundtable on "AI Hardware Globalization," four CEOs from agricultural robots, companion holographic hardware, voice Agents, and open-source on-device hardware sat down together.
To Avoid Being Remodeled, We Chose to Farm in Europe and America
If you were a startup making agricultural robots, would your first stop be domestic or overseas? Intuition tells us we should first get the model working domestically.
But David Sun, founder of Demeter Robot, gave a counterintuitive answer: "We went overseas precisely because abroad we don't need to do remodeling."
The reality of Chinese agriculture is the classic "large country, small farmers" model—especially in South China, where orchard cultivation is not only chaotic but also small in scale. If you want harvesting robots to work the fields, it means doing a great deal of "standardization remodeling" on existing domestic agronomy. In reality, this is almost impossible to push through.
In contrast, Europe and America have centuries of agricultural history that have polished farms into highly standardized industrial assembly lines. "Flat planning of 100 acres or even over 1 hectare is the perfect breeding ground for robots." David's team directly bypassed the complex unstructured domestic scenes and aimed at the European and American markets from birth. They are calculating the most straightforward business calculation: "We are simply using China's PPI (Producer Price Index) to benchmark against Europe and America's CPI (Consumer Price Index)."
In this track, running fast has never been the first principle. They currently can harvest one apple in two seconds. "Wanting to go faster is simple too—just add arms from 4 to 8." But real physical agricultural scenes don't listen to geeks talking about algorithms. Morning light, continuous outdoor rain, the fragility of fruit skin—all are forcing the team to take a step back. "Harvesting nearly one ton of crisp apples per hour, what if they get bruised and can't make it to the shelf?"
Ultimately, this becomes a compromise on "robustness." The hardware is produced using China's most extreme supply-chain ecosystem, but what is delivered overseas is actually RaaS (Robot as a Service). Farmers don't buy cold machines; they buy that intact, undamaged apple.
AI Is Too Dopamine-Heavy; Young People in Europe and America Need "Cyber Worry Beads"
While B2B robots are calculating efficiency, B2C companion hardware is calculating "human nature."
"In the AI age, what is least scarce is dopamine and all kinds of pleasurable stimulation." Wan Yi Roy, CEO of ALLTIMEWanwushi, discovered that while big companies are competing on smarter Agents and more realistic voices, users are actually getting increasingly anxious.
They aimed at a softer entry point: healing.
Roy's team made a holographic life hardware device containing a green microalgae IP called "Qiujun" (Marimo-kun) swimming inside. It sounds a bit like a cyber potted plant for urban white-collar workers. Would foreigners buy into this anime-style healing aesthetic in the European and American markets?
The research results surprised them enormously: European and American users' acceptance of anime-style art was actually higher than that of Japanese users.
In this cross-cultural collision, product positioning naturally diverged. The same product—Japanese users see it as "pet companionship" at home; while European and American users are more likely to place it on their office desk as a stylish "stress-relieving desktop aesthetic."
Faced with the same user-experience tradeoff, Hua Kun, CEO of Wavenote (focused on AI voice hardware), chose another extreme kind of "stubbornness."
What do users buying an AI recording card actually buy? They buy the "seamless flow" that eliminates all cumbersome operations. Hua Kun discovered that if the device has two buttons, there will always be some users who can't figure out which one to press. From a technology-first team's perspective, this could completely be solved with a manual. But in pursuit of a geek-level minimalist experience, Hua Kun invested extremely high R&D costs to turn two buttons into one.
This is also the destiny of the hardware industry: to smooth out that tiny usage threshold, the entire production line is frantically tuning behind it. "Sometimes, in pursuit of this kind of experience, when the second-generation product comes out, the first one quickly has to be discontinued. The cost of this tradeoff is extremely expensive."
"Memory Chips Are More Expensive Than Damn Gold!"
But as long as you make hardware, the first test of going overseas is never far away—it's in Shenzhen or Dongguan.
When it comes to supply chains, Wu Wei (SipeedSipeed), who started out running developer-community operations, feels it most deeply. They earliest did computer vision (CV) and small development boards; now their minimalist Agent project PicoClaw is surging on GitHub, racking up 30,000 stars in a short time.
But behind the surging code commits, the physical world taught them a lesson. "Memory chips right now are genuinely more expensive than gold," Wu Wei said half-jokingly. Over the past three years, the impact of geopolitical friction and tariff wars has cut to the bone.
Online developers are extremely cost-sensitive. If components rise across the board, the team—even under pressure from declining profits—has to stock up in advance on components that might rise in price, to smooth out price fluctuations in development boards. Even more absurd is the pervasive suspicion. They once saw unfriendly noise on YouTube, where someone deliberately took their hardware and software to find nonexistent security vulnerabilities, hyping up the "China threat theory."
Forced into the center of this storm, Wu Wei's team had to accelerate product transparency, open-sourcing, and patching speed. "It was completely forced learning and growth."
On this point, Hua Kun, who just recently experienced a sales surge, deeply relates. Their first recording hardware went out of stock three months after launch. "Hardware out of stock—this is a very painful lesson." After coming back, Hua Kun directly stationed people long-term in Shenzhen, grinding hard on these mid-to-high-end supply-chain resources.
Roy's team's solution, however, was as light as an accident. Their companion hardware contains not only electronic components but also real living aquatic plants. Electronic components can cross oceans, but plants can't pass customs. What to do? Their approach: manufacture the hardware shell in China and ship it to local warehouses overseas, then find local aquatic-plant vendors in the target country to supply. What buyers receive is a "Chinese shell + local soul" hybrid platter.
No matter how the world changes, and however different product forms may be, all Chinese AI global-expansion teams hold one common bottom line: technology may determine how high you can fly, but China's mature supply-chain system determines your floor and the capital to survive the winter.
More Conversation Details
Unique Awards · Hangzhou AI WEEK Trends Roundtable Panel
"From Product Innovation to Supply-Chain Collaboration in AI Hardware Globalization"
Guests:
Demeter Robot — Agricultural Robot — Founder & CEO — David Sun
ALLTIMEWanwushi — CEO — Wan Yi Roy
Wavenote — Founder & CEO — Hua Kun
SipeedSipeed — PicoClaw Operations Lead — Wu Wei
Moderator: Unique Capital — Partner — Wang Chaochao
Wang Chaochao: Just now I said our hardware industry is both sexy and brutal. Why do I say that? Because currently AI hardware is pulling us from the virtual world on the screen back into the real physical world, so it is very sexy. But the mortality rate of AI hardware globalization is actually extremely high—whether in compliance operations or supply-chain localization, there can be many challenges at every link. Today we have invited four guests, all in different tracks. Next, please introduce yourselves and what you do.
David Sun: Hello everyone, I'm David, founder of Demeter Robot, focused on agricultural harvesting robots. The word Demeter comes from the goddess of harvest in Greek mythology. Why did we enter this track? It is a relatively slow but very deep vertical track. Currently there are roughly over 300 companies in the global agricultural robot landscape, most concentrated in crop protection and weeding; harvesting is a very difficult piece. This scene is very hard to do, so our team combines scientists with Chinese and American backgrounds and people with commercial backgrounds. We were founded in California last year, born as a global company, and recently preparing to set up in Hangzhou. Our engineering prototype is already out and has been validated domestically. Next, commercial deployment will gradually roll out in Europe, America, and Australia.
Hua Kun: Hello everyone, I'm Hua Kun. We make AI Agents centered on voice. We originally always did software, then discovered first that there is demand for voice; second, that software moats are getting increasingly unclear because large models are devouring everything. We added hardware in the second half of 2024. I think the hardware moat is a bit thin, but it can still leverage the Made-in-China advantage. Essentially, we have now entered an era where large models or AI define hardware. We also need to make a good hardware product, because users first buy hardware, and what they use daily is still software—this is a very interesting combination, and demand is quite large.
Wang Chaochao: What is our hardware mainly used for?
Hua Kun: We are centered on voice—simply put, we take recordings of large volumes of offline conversations and use large models to unlock their value. We need to observe large model capabilities; in text and voice they are already relatively mature, so this is a field worth deeply cultivating. As for video, several large model vendors still have uncertainties; Sora has actually been paused because its costs are too high. Essentially, when we do applications, we are still leveraging the foundational capabilities of large models.
Wan Yi Roy: Hello everyone, I'm Roy, founder of ALLTIMEWanwushi. We are alumni of MiraclePlus' 2023 cohort. We are currently making a healing-style AI holographic life hardware, targeting urban youth and white-collar workers aged roughly 22 to 35 in North America, Japan, and South Korea. It may be different from the AI companion products you imagine. We position this hardware as a home natural-aesthetics product, hoping to provide urban people with readily accessible healing and relaxation. We feel that in the AI age, what people lack least is dopamine and pleasurable stimulation; what they lack instead is calmness. AI often brings people anxiety, and our product brings everyone the power of healing. We have an IP called Marimo-kun, a cute green algae plant, with nearly 100,000 fans globally. Our product will meet everyone this year.
Wu Wei: SipeedSipeed has unknowingly become a company nearly 10 years old. It was founded by a group of young people with very strong hands-on skills who wanted to do some AI-related things. Sipeed earliest did camera-related CV, CNN image recognition. Early users were college student makers and overseas hardware enthusiasts. From the beginning, AI was Sipeed's first keyword; the second keyword was globalization. Sipeed started trying to go overseas very early, with extensive developer interactions on X, and multiple different product lines on Reddit and Kickstarter. After 2020, Sipeed added a new keyword: RISC-V. This is a very open instruction-set standard for global makers, with many interesting chips—especially RISC-V plus AI chips. Sipeed played a pioneering role in this wave. Recently during Spring Festival, we got another new keyword and identity: PicoClaw. PicoClaw initially also leveraged AI code generation, built on Go language, a very streamlined AI Agent that can run on various embedded small devices, like making things with OpenClaw etc. Because Sipeed has so many global developers gathered, in a short time it got 30,000 stars, over a thousand PRs processed and submitted, forming a very active community. We believe PicoClaw is now gradually being pre-installed by domestic manufacturers; it may truly enable the vast number of embedded or edge very small devices to access AI, giving them new life.
Wang Chaochao: Simply put, it's on-device or on-device Agents on consumer hardware.
Wu Wei: Right, like Raspberry Pi, set-top boxes, even very old phones can be installed.
Wang Chaochao: So first software then hardware, first hardware then software.
Wu Wei: At the very beginning we did hardware, because many developers asked software questions, so we gradually invested in software personnel to do some software. Because college students need to do AI learning, later we actually did training services like MaixCAM, doing fine-tuned models, and gradually software services increased.
Wang Chaochao: All guests are in different tracks: agricultural robots, recording devices, holographic companion hardware, and the both-software-and-hardware developer ecosystem. Going back to the source, our theme is "From Product Innovation to Supply-Chain Collaboration in AI Hardware Globalization." Let's start with product innovation. We all know that what we now call AI hardware globalization is never one product sold globally—that's not realistic. The question is: when you guests made your first stop overseas, how did local users or local demand force you to iterate and improve your products? From another dimension, how do you balance localization and globalization, and how do you achieve differentiation? Starting with David.
David Sun: This question happens to be an opposite answer for us. We are currently doing fruit harvesting in agriculture. Abroad, agronomy may have five or six generations or even centuries of history and is relatively standardized. After being introduced to China, large-scale planting only began after the reform and opening-up with the household responsibility system. In China now, especially in South China, orchard planting is relatively chaotic. This means harvesting robots need to do "standardization remodeling," which is unrealistic in many places. Conversely, European and American countries may have centuries of planting history, and their agronomy after over a hundred years of iteration is now very standardized. Second issue: in China most situations are called "large country, small farmers," with relatively small planting areas; in developed countries many are 100 acres or even over 1 hectare, which conversely is more suitable for robot operation. So for us, going overseas actually doesn't require remodeling; agricultural robots overseas are actually the most suitable scene. People ask why not do it domestically? One is that domestic labor costs are still relatively low; the second is that a lot of standardization remodeling is needed. Except for large farms developed in recent decades in Xinjiang or Ningxia where robots can be used, going overseas is actually a more suitable scenario for us—we were born global.
Wang Chaochao: Because Chinese agriculture differs quite significantly from Europe and America, you should all have experience. Hua, let's hear your explanation.
Hua Kun: I think it's several layers. The first step of localization is still building trust. We are a consumer-grade product facing many consumers; some foreigners, because they have been hurt by certain merchants before, inherently have a guard mentality. This process requires quite a lot of energy for localization: language, content, advertising, product design all need to be done very carefully, making them feel this is an international product that respects and understands them. After doing this well, the product gradually sells out, and at this point customer service needs to be done well. At the beginning we really didn't understand users much; all differentiation is not for differentiation's sake, but focusing on selecting a group of users and doing their needs well—and in the process of doing it well, a certain differentiation actually forms. For example, some competitors much stronger than you may have already captured many users; you need to see where penetration is still insufficient. For example, we found their coverage of young consumer groups may be slightly weaker, and this may be our audience. So whether in product ID design or basic functions, we will do different designs, then deeply communicate to get the functions right, and only then gradually form product positioning and differentiation.
Wang Chaochao: So it's differentiation derived backwards from user needs. B2C products have this characteristic—repeatedly communicating and aligning intent with user groups.
Wan Yi Roy: I think user demand is a very important point for us to find PMF. We released a product concept domestically last year, using the Marimo-kun IP at the time, which was a healing anime style. Now we are doing a 3D upgrade, making it a cooler holographic structure. Subjectively, it's not because users asked us to do this, but objectively, at the time we had a concern: would European and American users like this anime style? We sent survey questionnaires in China, Japan, South Korea, Europe, and America, and found that actually European and American users like this style more than Japanese users. Europe and America rarely have this kind of particularly refined style, so I think all products ultimately return to basic human needs—to a certain extent, some pursuits transcend borders. For example, Japanese users will define it as pet companionship and place it at home; European and American users have a higher proportion treating it as a home decoration, placing it on an office desk. Although everyone has subtle differences, the underlying layer is the same—everyone needs this kind of healing, cute, natural product to soothe them. As long as you grasp the human nature resonating with this era, you can definitely make products people like.
Wang Chaochao: Simply put, some first impressions may differ, but the essential needs of human nature are universal; Japanese-style products hitting Europe and America are realized here with you.
Wu Wei: Sipeed's situation is relatively much simpler and cruder. Because we are B2C facing various developers, and developers are the most active group on the internet. If the product is not good to use, they will directly complain online, so we can very intuitively receive problems from different regions. At the same time it itself has great viral potential; many users buy development boards and do fun things, posting them on Bilibili or YouTube, naturally helping us spread. But in recent years, after geopolitics rose, we can still feel resistance. First is tariffs, various costs—memory chips right now are genuinely more expensive than gold. Additionally, we have indeed observed some unfriendly voices on YouTube, looking for security vulnerabilities from our hardware or software, hyping up the "China threat theory." Because we do have influence, in this regard we were forced to strengthen rapid response and security vulnerability patching, accelerating transparency. This was also forced learning and growth, with very great efforts made.
Wang Chaochao: So, the core logic of localization and globalization is whether there are large differences in demand scenarios between regions. For holographic companions and maker hardware, some aspects are similar; but for recording and agriculture, the differences are relatively large. In product iteration, how do you make tradeoffs and balance among technology, scenario, and function?
David Sun: Industrial prototypes and engineering prototypes are already out, and have been iterated three times. The biggest difference between agricultural robots and industrial robots is that their scenes are unstructured. Morning light differs from afternoon light, and outdoor operations also encounter rain. Our main tradeoff is between efficiency and stability. Currently our harvesting efficiency is in the first tier, averaging two seconds per apple—it is a tracked multi-arm harvesting robot. Wanting to improve efficiency, changing from 4 arms to 8 arms would do it. But we step back to look at stability: will multi-arms collide under spatial constraints? We have a motion algorithm that calculates how many arms is most efficient. Second, harvesting nearly one ton per hour—since apples are relatively crisp, how to avoid damage? So we are doing electronic fruit vibration and collision tests. Because not only must we harvest fast, but later we also need collection and transport, coordinated with how many AGVs. We mainly consider the robustness and stability of the entire harvesting system, rather than single-mindedly pursuing efficiency. We must ensure the system operates stably, and the harvested fruit meets commercial grade—able to go on shelves without damage.
Wang Chaochao: To summarize, it's high stability, low loss, making tradeoffs under this core goal. Hua.
Hua Kun: What we consider more is the efficiency of user experience. Previously recording was very hard to use—after recording you couldn't do much, and operations were very cumbersome. Now with the combination of large models, this seamless feeling is what users pursue. In the face of extreme experience, users are willing to pay costs, so the current stage is about how to make user experience sufficiently efficient. For example, originally the device had two buttons, one of which was a mode button, and many users couldn't figure out what it meant. We could blame the users and strengthen the manual, but we felt that was too troublesome, so we solved it technically. We spent a lot of R&D effort, and now we've turned it into one button. The cost of this kind of geek-experience tradeoff is very high—in pursuit of experience, iteration must be done quickly, and when the second product comes out, the first one gradually stops being produced.
Wang Chaochao: Overall it's efficient product experience.
Wan Yi Roy: Many people with technical backgrounds tend to make the product good enough before launching it. But we are very close to consumers, and there is a working method called "Amazon Working Backwards" that is particularly useful. Before making a product, you can imagine what the final press release will look like and what the selling points are. Working backwards from the result makes you more focused on the core value delivered to users. The team internally continuously discusses P0 to P2 priority core functions, while going out to do user interviews to confirm. Sometimes a function you think is very cool, users don't care about at all. Offline is the most efficient place to validate—you can see what kind of product is truly useful to them, and their consumption decisions are very fast. Startup teams have very limited energy and resources, and should focus on polishing the core value.
Wang Chaochao: It's finding the greatest common divisor of customer needs. Wu, how do you make tradeoffs and balance in technology, function, and product?
Wu Wei: Sipeed is mainly divided into two major aspects. First, we are close to developers—if a new function comes out and feels fun, we will do early user research and voting on Twitter and domestic WeChat groups in advance. After making a prototype, we do pre-sales on Kickstarter to see the response. For example, previously based on a RISC-V chip we made a very small development board SG2002. After making it, college students and developers wanted it to have a camera and AI attributes. To accommodate them, we did software matching, made an IDE, pre-trained models like YOLO, and later made a small app store. Then later someone thought it was fun, added a night-vision camera, and made a "night-vision device." Actually we don't have a rigid three- or five-year plan; developer feedback and complaints are enough to enable us continuously leapfrog from existing products to develop new product lines.
Wang Chaochao: Sipeed, as a both-software-and-hardware company, pushes backwards from the maker-developer group—making whatever is needed, gradually satisfying them. We've talked about localization, product function, and technology. Next let's discuss the second topic: supply chain. What is your company's supply-chain strategy? In the next two years, what do you think the black swan or life-or-death crisis in supply chain will look like? Starting with Wu.
Wu Wei: Sipeed itself does AI hardware design, and we are in Shenzhen. Doing hardware must be close to the industrial belt—this is definitely the most ideal place globally. At the same time, because of globalization, we also have extensive connections with distribution hubs like Hong Kong. Supply chain is a long-term process of screening and tuning, gradually forming a comfortable upstream and downstream cooperation network. The past two years, due to the tariff war plus memory price increases etc., have been relatively difficult. We can only keep some profit margin and inform users of pricing strategies in advance. Overall maintaining healthy positive cash flow. Additionally, based on industry intelligence, we will stock up in advance on components with rising-price trends to smooth out the risk of price fluctuations.
Wang Chaochao: Overall it's relying on China's Shenzhen supply-chain efficiency, being fully prepared for geopolitics, and managing cash flow well. Roy.
Wan Yi Roy: Our product hasn't reached the mass-production stage yet; what we pursue now is certainty, and we won't necessarily use overseas components just for cost reduction and efficiency. We hope to choose reliable factories to properly deliver the product to all users. But our supply chain is quite fun—the product is a combination of virtual and real, containing living plants. We can't ship living things globally, so our hardware standard configuration is shipped to various warehouses, then in the local country we separately cooperate with aquatic-plant vendors to ship together to users, forming a complete experience. As for black swan events, recently we've found that screens, sensors, and chips have all risen in price; political factors and tariffs have an impact on early-stage products. At this stage the most important thing is PMF—selling the product and delivering it well is most critical.
Wang Chaochao: Simply put, China production, overseas packaging, then direct to customer.
Hua Kun: Supply chain for hardware is too important. I just came back from running supply chain in Shenzhen. After launching in the second half of last year, we sold out within three months. First time doing hardware with insufficient experience—this is a very profound lesson. So now our first is a multi-supply-chain strategy; second, because we position mid-to-high-end, finding experienced, stable, high-quality suppliers is not easy, and relationships need long-term maintenance. We have specifically placed people in Shenzhen to build this system.
David Sun: In one sentence, we are using China's PPI to benchmark against Europe and America's CPI. To break it down: the current strategy is not selling robots but RaaS—delivering results. Hardware is definitely produced with China's supply chain, but if delivered as a service, we need localization. We will form local teams to serve these farmers and cooperate with local agricultural associations. To recap: hardware produced in China, service in Europe and America.
Wang Chaochao: China production, sales delivery and service locally. Listening to the four guests' sharing, AI hardware globalization has different paths, but overall one thing remains unchanged: no matter how technology changes, if you grasp demand and make the product, it will sell. Product innovation determines how far an enterprise can go. Relying on China's supply chain can determine the floor. Although geopolitics etc. may cause supply-chain turbulence, as long as cash flow is managed well, supply chain is done well, and customer needs are deeply understood, this business can continue to be played. Thank you to the four guests for sharing.