Original · Unique Research · 2026-06-26
Editor's note: The first-person report and its judgments belong to the original Chinese author (hosted by Wenqin Luo; five founder panelists). This English rendition retains the narrative on edge vs. cloud, presence, hardware-as-entry, five closing lines, plus the full four-stage verbatim Q&A. All panelists, companies, and figures are preserved. Founder statements and cited numbers are source attributions, not independently verified findings.
AI Industry Observation
Models get stronger by the day — why are these people grinding on hardware?
The model does the "god's-eye view"; edge intelligence does the "smoke of everyday life"
"
AI's true popularization isn't teaching everyone Prompt Engineering, but letting that 95% who have never used AI feel AI's convenience without feeling AI's presence.
ChatGPT updated again, Claude got better, domestic large models' parameters hit new highs.
But you may not know one number: about 95% of people worldwide have never used a large-model product.
This number was shared by a founder who moved from AI Infra to hardware in a recent industry conversation. He did infrastructure upstream in the supply chain, then Agents, and finally chose to root in hardware. In his own words: "if you can't be a Token maker, be a Token seller."
Honestly, this choice is quite counterintuitive. Models get stronger and cloud Agents are nearly omnipotent; why drill into the "edge"? Why not do the sexiest application, and instead touch the hardest hardware?
Recently we held an industry conversation in Singapore; five founders — AI voice wearables, smart glasses and rings, pet smart collars, AI emotional jewelry, and smart-city IoT platforms — talked for a full two hours.
By the end, they reached an unexpected consensus: large models do "god's-eye-view" intelligence, while the edge does "everyday-smoke" intelligence. AI's true popularization isn't teaching everyone Prompt Engineering, but letting that 95% who never used AI enjoy AI's convenience without feeling it.
Behind this consensus lies the industry's biggest structural opportunity.
Making AI "Land" Is Harder Than Making It "Smarter"
Hua Kun's story is probably the segment that best represents this industry turning point.
He did AI Infra, did Agents, and finally chose to make a voice wearable. Why? Because upstream in the supply chain he saw a cruel reality: models iterate too fast, and Infra and Agent moats are refreshed every day. "You have to find a little barrier."
So he chose hardware — not because hardware is sexier, quite the opposite, hardware is hard, but hardware has one advantage a cloud Agent can never match: it can physically appear in the user's hand, crossing the digital divide.
What's his user profile? Not the Silicon Valley tech elites chasing the newest model every day, but "old heads" — older people who can't use complex software, even ordinary people in remote parts of Brazil and Central Asia. These people have real business needs — record meetings, organize information, generate summaries — but ask them to open ChatGPT and write a Prompt? No way.
"To really use some of today's large models well, for many elite users in Silicon Valley and Singapore it's very easy," Hua Kun said bluntly. "But for many 'old heads,' it's actually quite hard."
This sounds like a joke, but it speaks a truth the industry long overlooked.
Hardware naturally lowers the usage threshold. A user buys a pair of glasses, a ring, a voice recorder; holding it physically gives them psychological security. One-button operation, no need to learn Prompt, no need to understand what an LLM is; AI is right by their hand, working naturally. Hua Kun says many users feel "Amazing" after use — not because the tech is advanced, but because it's so convenient they don't realize AI is behind it.
That's true popularization.
Being a Token maker is the giants' job; being a Token seller is the founders' opportunity.
The Edge Isn't a "Downgraded Cloud" — It's "On Site"
If you think edge intelligence is just moving cloud capability local, you're very wrong.
The edge's real value isn't "replacing the cloud," but doing what the cloud can never do — being present.
Gu Yan's story makes this especially clear. He's the former R&D VP at Huami, who left to found SATELLAI, an AI pet smart collar. His own dog wears it now.
"We started by thinking, a cloud Agent knows everything, so why do edge?" He shared a key cognitive turn. "Then we figured it out — the cloud's biggest problem is 'not on site.' We're solving the problem of an animal that can't speak; the intelligence it needs must be 'closest to the ground.'"
Pets can't talk, can't use an app on demand. It runs on the grass, zones out on the sofa, paces anxiously at midnight — only the device on its neck can capture these behavioral data. However smart the cloud, it needs data as input, and edge hardware is the always-present data collector.
Deng Xudong's AI note ring logic is more direct: the ring is the AI Agent's shortest interaction path. Double-click to record, long-press to converse, single-click to mark key points — the whole interaction friction is near zero. "The device must first be a good device, with you 16 to 24 hours; the simpler and more invisible the better."
Grace's AI jewelry extends "presence" into emotional guardianship. Her target users are single women and the elderly — UN data shows one-third to one-quarter of future households will be single-person. When a user faces an emergency, the bracelet is the most direct lifeline.
But Grace mentioned an impressive detail: if an emotional-AI reply takes over 1.5 seconds, users basically give up. To feel like a real human conversation, it must be under 200 milliseconds. Today's tech basically caps at around 400 milliseconds — that 0.2-second gap is the chasm edge and cloud must cross.
Imagine: a lonely elderly person wakes at midnight and chats with AI to kill the loneliness, waiting 1.5 seconds for each reply. That loneliness is instantly amplified. But if the reply comes within 200 milliseconds, it feels like someone is truly beside you — that's the irreplaceability of edge low latency.
From the platform side, this logic is simpler and cruder. Tuya Smart does lots of smart-city and enterprise projects; Tu Xiwei says: "security, healthcare, eldercare — once these scenarios go to the cloud, they face network latency and instability. Where efficiency and security suffer, the edge is a must-have."
So edge intelligence and cloud Agents are not competitive but complementary. The cloud does "omnipotent" general intelligence; the edge does "always present" dedicated intelligence. Together they make true AI landing.
Hardware Is Just the Acquisition Entry; "Understanding" Is Where the Money Is
You may ask: hardware has value, but hardware margins are so low, how do you make money?
These founders' business thinking represents the industry's most cutting-edge model exploration. Interestingly, their chosen paths differ greatly, but the underlying logic is strikingly consistent — hardware is the door, service is the room.
Hua Kun's answer is most direct: users' biggest essential demand is instant gratification. Most voice-to-text today is offline processed — wait until recording ends then generate — because real-time Token cost is too high. But he thinks this logic breaks soon. "About 10 large models worldwide compete with each other, which is good for us application companies. My goal is to ship real-time capability, drive subscription fees extremely low, so ordinary users in Brazil and Central Asia can afford it — users don't care which model you use; they want the best experience and best value."
This is a "thin margin, high volume" global logic — not profiteering on hardware, not harvesting users with subscriptions, but using extreme value-for-money to acquire mass users and finding profit in scale.
Gu Yan's business logic is more interesting. The pet collar is also "hardware + subscription," but he frankly says: "what we really want to earn isn't communication fees, but the insight value of 'understanding pet behavior.'" The exercise, health, and behavior data the collar collects, after AI analysis, can tell the owner: your dog didn't exercise enough today, should eat a bit more; recent sleep quality dropped, maybe pay attention. These are ordinary users fundamentally unable to get on their own.
"The profit order is usually women, children, then men. When I did digital products I'd break BOM cost down to the chip level; it's hard to make hardware/software money off me. But when I buy things for my dog, I suddenly understand that irrationality women have buying for themselves or their kids." Men are stingy buying digital gear for themselves, but spend on pets without blinking. "Men spend less than dogs" isn't a joke; it's a validated business rule.
Grace's entry is also clever. She didn't first do "smart hardware"; she first did "accessory jewelry." "Choosing women, the highest willingness-to-pay group, first it must be beautiful, emotionally valuable, and decorative." AI jewelry is first a beautiful bracelet, then a safety-guarding device. Pay for "beauty" first, then renew for "smart."
Deng Xudong's ultimate judgment pulls the widest frame: "In the PC era typing wasn't extra-charged, and calling in the mobile internet era was cheap. But which apps spent most in those two eras? In China it's Didi, Dianping, Ctrip." His judgment: the ultimate business model isn't "selling features," but "shout and it's done" — ride hailing, hotel booking, making PPTs, sending schedules, all one sentence through an AI Agent. Then hardware is just the service entry; the real value is the whole service ecosystem the Agent connects.
Worth noting is B2B and B2G. Tu Xiwei says directly: "As long as the pain point hurts enough, clients are willing to pay for many years. Government and large-enterprise sides still have lots of edge-intelligence business opportunities." These scenarios are less sexy than consumer hardware, but demand is real, budget stable, and willingness strong — actually one of the steadiest business paths now.
Five Lines for Entrants
At the end, each founder said a "heartfelt word." These five lines are for everyone thinking about entering edge intelligence.
"Supply chain is the survival bottom line." (Hua Kun)
His overseas order demand is huge; he's sold out twice. "Supply chain is whether you stay alive, not something to ignore first. You can differentiate on Agents or other experience, but first you must nail the supply chain." When demand comes and you can't fulfill it, users turn around and buy someone else's. In hardware, supply chain isn't a plus; it's life and death.
"The device must first be a good device, with you 16 to 24 hours." (Deng Xudong)
The ring must first be a comfortable ring, glasses a good-looking pair of glasses, the bracelet a good-looking bracelet. The simpler and more invisible the better, don't overcomplicate. A device you want to take off the moment you put it on is useless no matter how strong the AI.
"Don't force smart for smart's sake; think clearly what goes to the cloud and what goes to the edge." (Gu Yan)
The Pearl River Delta already has very mature hardware solutions; the hardware threshold is dropping. What you must think through is: which functions really need to be pushed to the edge, and which are fine in the cloud? Don't cram unused features just to look "smart." Stacking features to tell an AI story, users won't pay and costs rise. Good products subtract.
"Start from humans' deepest needs, not just efficiency." (Grace)
Many AI hardware focus on efficiency and precise capture, but humans' deepest needs are safety, companionship, emotional connection. Get human needs right first, then talk tech.
"B2B and B2G have real scenarios; when the pain hurts enough, clients pay continuously." (Tu Xiwei)
Founders who find C-end too involuted should seriously look at government and enterprise sides. Less sexy, but real.
Closing
For most ordinary people, maybe not one uses a Prompt, not one cares what a Transformer is, not one knows what a Token is. But they're all using AI.
It doesn't make the smartest AI; it makes the most "present" AI. It doesn't chase the biggest parameters, but aims to let that 95% who never used a model enjoy the convenience technology brings.
Large models do "god's-eye" intelligence — omniscient. The edge does "everyday-smoke" intelligence — on your finger, on your wrist, beside your pet, quietly present, just right.
Both matter equally. You could even say, with only god's-eye intelligence, AI forever belongs to that 5%. Only when everyday-smoke intelligence is also built can AI truly become this era's infrastructure.
So, rather than anxious that models iterate too fast and you can't catch up, ask: in that 95%, in the scenarios the model can't yet reach, what can you build that makes them say "Amazing"?
That may be your opportunity.
More Conversation Detail
Panelists: Kun Hua / 华琨 (Wavenote Founder & CEO); Tissia Tu / 屠熙蔚 (Tuya Smart Marketing Innovation Director of APAC); Domingo Deng / 邓旭东 (Gyges Labs (Vocci) Co-Founder); Yan Gu / 顾岩 (SATELLAI Co-Founder & CTO); Grace Xiong / 熊楚伊 (Veryloving.ai Founder & CEO)
Host: Wenqin Luo / 罗文琴 (TMTPost VP & Partner)
Stage one: introductions and edge-scenario entries
Wenqin Luo: This is probably today's only forum combining software and hardware, with physical hardware on display. You all brought your own products. Let's start, thanks to Unique Research for setting the stage. In early May, Mr. Wu and I chatted in Singapore about what topic for a full day; I leaned toward AI edge-intelligent hardware landing.
From our tech-media view, in January Huang (Jensen) preached physical intelligence at CES. And Sanwen data: the whole edge-AI market should grow from last year's several hundred billion — about 300-some billion market cap — to a trillion in five years, near 40% CAGR.
We also feel: last year many investors told us, as long as it's a big-company middle manager or exec backed by a supply chain, they're guaranteed funding. AI plus hardware was that hot last year. But from each founder's micro feel, I don't know how it is. So today we invited these five, representatives of the whole edge-intelligence ecosystem in varied forms and entry paths, to discuss from the micro angle: what's the feel, where's the opportunity, where's the risk.
First, quickly have each introduce the company and how your product enters the "edge" scenario. Mr. Hua first.
Hua Kun: Hello, glad to join. I'm founder of Wavenote. Our wearable is voice-related. Actually yesterday I attended SuperAI Summit; Perplexity was a gold sponsor, they did very well. We entered this track because we saw the demand is very, very large.
I previously did AI Infra, then Agents, but found models iterate too fast, so we need a little barrier, hence combining hardware. The starting point was also to let people worldwide use AI universally. In plain terms, we can't be Token makers; we're more Token sellers. Our vision is to let ordinary people use AI.
I saw data: now globally about 95% of people have never touched a large-model product. Our wearables' main users aren't the AI elites or heavy software users we usually see, but rather older users. In business meetings they need to record important information into what they want, but their UX bar is very high, so hardware-plus-software has lots of work; current penetration is low.
Wenqin Luo: Can you show the product?
Hua Kun: This is one of our wearables. We actually have many products, including card-type, and are developing a new product for different groups. This kind of wearable, we found overseas university students, self-media workers, and doctors use it more; card-type is more widespread, mainly business SMB users.
Wenqin Luo: In one sentence, Mr. Hua went from cloud to AI apps, then to edge consumer hardware, bottom-up into the recording scenario. OK, now Mr. Deng.
Deng Xudong: Hello, I'm Domingo Deng, co-founder of Gyges Labs (W Gadget). We do Wearable AI, integrating all resources for next-generation display and interaction-computing platforms.
In 2025 we launched our first co-released product, Holiday Glasses, a smart glasses with display, very light.
Wenqin Luo: Is it the pair you're wearing?
Deng Xudong: I just put it below to charge. This glasses' display is very discreet, letting the wearer present display solutions conveniently. Glasses solve "display"; also in 2026 we made an interaction ring. If next-gen phone-replacing devices need display, interaction, and compute, we may only be missing Mr. Huang's compute. So I'm glad to learn how edge computing fuses with wearables.
In one sentence: we want to integrate resources for next-gen display-interaction-compute platforms and launch good products. Let consumers stop relying on phones for everything, and let Agents be within reach very invisibly. So our ring is the AI Agent's shortest interaction path and shortest entry.
Wenqin Luo: Thanks, no need to summarize. Down to Mr. Gu.
Gu Yan: I'm Yan Gu, co-founder of SATELLAI. Simply, we're a pet smart-hardware company. Our main product form I can't wear myself; it's for our little dogs. First-gen products are various collars.
I have a beloved dog now wearing it, just couldn't bring it to Singapore this time. Our form is pet smart hardware, but what we truly want to solve is connecting pet owners — or more broadly, between pet families and pets — building a channel of connection. We hope to help owners better understand pets' behavior and what those behaviors mean. Over their ten-plus year lifecycle, help owners grow with their pets. Our brand is SATELLAI (SL).
Wenqin Luo: Thanks. In one sentence, Mr. Gu differs from Mr. Hua. Mr. Gu came from Huami, hardware and wearables. He starts from the edge to get the closed loop running, then to pet data platform, then to cloud coordination. Now Grace.
Grace: Hello, I'm Grace, founder of Veryloving.ai. We make a series of products for people living alone, from emotional connection to safety guardianship.
This on my hand is AI smart jewelry. The people it helps: UN research predicts one-third to one-quarter of future households will be single-person. In emergencies, we can promptly contact family, with safety guardianship and alarm functions.
Wenqin Luo: Thanks Grace. Last, the center seat; you're last for a reason. Introduce yourself.
Tu Xiwei: Hello, I'm Tu Xiwei from Tuya Smart, responsible for whole APAC marketing and innovation. Back to the topic, I differ a bit from other panelists because Tuya is a platform. So on the edge we look more at playing a B2B2C or B2G2C role, bringing whole AI applications into smart cities, smart buildings, and larger scenarios, forming a closed loop.
Why, at this moment, with models so strong, grind on the edge?
Wenqin Luo: You can see today's roundtable guests occupy different niches in the whole cloud-collaboration ecosystem; these different cuts will bring more sparks.
First question: whether going up from Infra or radiating from edge to cloud, why at this moment — models getting stronger, Agents seeming omnipotent — do we still emphasize "edge intelligence"? Starting with Mr. Hua?
Hua Kun: As I said, let more ordinary people use AI by their habits. We feel this deeply ourselves. A wearable AI device has two parts: hardware and software. Hardware distribution is traditional, mainly through e-commerce and offline channels.
Globally this involves many countries and regions. E-commerce and offline, though traditional, can reach more, more remote areas, like Brazil, even Central Asia. They have demand, but the most advanced models, due to user habits or other reasons, they can't use frequently or can't use at all; some countries can't directly use due to policy.
But hardware, through mature logistics, lets more ordinary people access these products. Because they have the scenario and shopping habit, they buy hardware, but actually what they use daily is more the software. Hardware must get simpler, basically one button. But what really delivers value is the software function; AI can do so much now.
So we focus on letting these users enjoy AI's convenience through AI Agents, making AI easier to use. It essentially records lots of offline things, then can do much. It can become a PPT, a self-study scenario, a test, a script, etc. With hardware as the carrier, ordinary people use it and naturally, through our design, use these functions in one go. Many users feel Amazing, so many convenient things usable, and the operation is very simple.
To really use some of today's large models well, for many elite users in Silicon Valley and Singapore it's easy, but for many "old heads," it's quite hard. So that's our opportunity — design a good hardware, put it in his hand, let him one-go and naturally use AI's many capabilities very well in their scenarios.
Wenqin Luo: You mentioned your background in cloud, apps, then hardware. Hardware makes AI easier, but must it sacrifice many functions? From cloud to edge, does model capability get sacrificed? If there's an "impossible triangle," what's your order of tradeoffs?
Hua Kun: I think it comes back to target users. Our target users are generally older, different from Silicon Valley young people. As I know, like Plaud, young users in California are few. Why? These elite users use the most advanced software daily; maybe in three months they won't even use phone apps anymore.
But the broader global users only use a few scenarios, and don't use them much. So I think this is a very design-worthy direction. Though models iterate fast, most users don't need so many complex, advanced features. And these people are the vast majority; like Apple smartphones, many Apple users don't use all the complex features 100%.
Another important proposition: we must ensure UX is simple enough, giving them the latest capabilities through our tradeoffs around their scenarios. I think many complex features they can't use and don't need.
Also how to lower cost and lower the threshold; that's what we study too.
Wenqin Luo: The cost you mean is product hardware Cost, or the user's learning barrier?
Hua Kun: Both. Mainly usage cost, i.e., learning threshold low enough. Another is, per country/region, we must lower their money Cost. Users feel they bought hardware, especially China and Japan users. We found Japanese users are very frugal; they even ask us to filter out useless audio, so they don't spend Token money.
I used to do Infra; I feel obligated to lower user Cost. If ChatGPT is $20 a month, why would he pay another $10 or $20 for other software? I think this cost should be lower. That's what we'll do; we're also pushing Token globally, obliged to make experience better and cost lower.
Wenqin Luo: On experience, I want to ask Mr. Deng. You mentioned Wearable AI capability; first-gen smart glasses, second-gen a ring. Why choose the ring form factor?
Deng Xudong: Our company positions next-gen display and interaction-computing platforms. Display matters, so first product was glasses. But while doing glasses, we found many consumers' essential need is collecting voice information in conversation. Collecting voice doesn't necessarily need glasses; many devices do it already, pendants or clips on phones.
The ring happens to be our natural capability. When developing glasses we paired a ring; consumers asked: why must recording and voice interaction go through glasses, is there a better interaction? So we extended from display to interaction.
From interaction logic, the ring is indeed a good carrier. We were among the earliest, launching the world's first AI note ring; from CES launch to now, the buzz and feedback earned mainstream media and consumers' recognition. This form factor has many advantages: almost invisible wear, accompanies you all day, and very low interaction friction. Interaction cost — as Mr. Hua said — must be very low. I double-click anytime to start; in a meeting one tap to Mute; long-press to interact with the Agent. Across Agent invocation and execution, this carrier is most convenient. So we found positive feedback from user experience and first-gen product communication, started the project last year, and launched this new product at CES in January 2026.
Wenqin Luo: But here's a question: conventional thinking says the ring's most convenient scenario is health monitoring, like Mr. Gu's. Why do you make office and voice transcription the ring's core scenario? These serve completely different people.
Deng Xudong: First, company positioning: next-gen display, interaction, and compute. Health is a very big need and good market; Oura fully validated it as a multi-million-unit annual shipment market. But each company has its own mission and the consumers it serves. We can't expect one product to serve all consumers, including all pets; that's unrealistic.
Any company relies on its own talent. Whether starting with Infra, hardware, or market/need, all combine natural talent with market need to make a product solving a specific problem for users or pets in a specific scenario; that's the company's commercial value. So back to the topic, health matters, but it may not be our dish. We're stably taking business scenarios, productivity tools, and convenient Agent invocation as our stage goal.
Wenqin Luo: Thanks Mr. Deng. Now Mr. Gu. This question differs a bit for you. The first two talked from Infra/software to edge, sacrificing model capability to prioritize different regions, ages, habits. But you go from a very concrete, clear edge scenario, reverse toward a software-side "pet health data management platform." In this process, what do you think is the biggest technical evolution or the biggest blocker?
Gu Yan: Good question. Actually when we first chose pet smart hardware, we repeatedly thought about the problem we solve: in a pet's whole lifecycle, what insights can we provide, and what services based on those insights.
AI is certainly the most effective tool we can't escape this era. We thought about existing model capability, e.g., more cloud Agents or cloud intelligence; it's truly an omniscient intelligence, but its biggest problem is "not on site."
Our chosen approach is the edge. Though middle-tier hardware intelligence is limited, for the special category of pets, a more important capability is "on site." When positioning the landing path, the first thing was to land intelligence first. We casually say inside the company: on landing, our hardware is closest to the ground. Because it's always worn on the pet's neck or body.
Our problem-solving differs from other companies. Most Agents solve human problems; we solve an animal that can't speak, and what human intelligence it may need. So across the whole chain we started from hardware, slowly learning all the pet's behaviors and emotions. From that we found insights and fun things. In this process we slowly found models that better understand pets, then put the whole thing on the cloud; that's how we came.
Wenqin Luo: Thanks. Now Grace: I understand you're more of a native software-hardware collaborative intelligence company, both sides synchronous rather than one before the other. Introduce how you think about cloud-Agent vs edge-intelligence.
Grace: Hardware is just a carrier, a terminal collecting data. Worn on hand, users interact better. Since our user type is a protected group, I started from my own pain — single, living alone, solo, and a woman who loves to explore and travel — extending to nursing homes. For example, we connected to some single elderly in 30 nursing homes. We combine AI compute and cloud, storing their daily emotional memories. Hardware is just a data-collection terminal.
Wenqin Luo: Thanks Grace. Basically the hardware panelists all say they're not limited to one hardware form; hardware is just a data terminal. Now Tuya (Mr. Tu) — you're a platform connected to many edge-intelligence forms. From your broad data samples, which scenario has the strongest demand for invoking intelligence?
Tu Xiwei: From my view, because we do many large government-infrastructure projects and serve many enterprises, including eldercare and medical scenarios Grace mentioned, their edge need is: once on the cloud, network latency and instability. For some scenarios, that instability is intolerable. So it needs edge-local deployment or offline Agent processing. Scenarios where efficiency and security suffer are must-haves.
Wenqin Luo: Like security, cameras — these less sexy-looking fields are actually the must-have scenarios.
Tu Xiwei: Yes, exactly.
Grace: I can add, there's a trade-off. Because we do emotional user data, it needs very strong privacy, especially sensitive medical, emergency, and health data. On invocation efficiency, we tested: if emotional AI's reply speed exceeds 1.5 seconds, users basically give up quickly. We try to be under 200 milliseconds for a more real human conversation feel, which users accept better. But now due to efficiency, tech basically caps around 400 milliseconds.
Wenqin Luo: This is also the technical landing blocker and efficiency point everyone faces on edge-intelligence and human experience.
Stage three: edge-intelligence business models and subscriptions
Wenqin Luo: Everyone just touched on scenario pain and cloud-collaboration blockers. Mr. Hua mentioned how to consume Tokens at lower cost, including subscription beyond hardware. This brings us to the unavoidable closed-loop business question: do we just let users buy hardware, or (of course not satisfied with hardware alone) how do we get users to pay subscriptions? In edge-Agent implementation, what experience, methods, or insights can you share? Mr. Hua can continue your topic, since you mentioned different regional preferences.
Hua Kun: You mean how to get users to pay. First, everyone here wants to sell hardware globally. Globally, many countries/regions have decent payment habits, though purchasing power differs.
For example US users spend generously, strong purchasing power, directly buy our tiered packages. But back to our product, there are still many unsolved pain points. Users' biggest essential need is turning offline-captured voice records into real, accurate text. Today's experience isn't perfect; why? Because OpenAI, Perplexity etc. actually provide an offline experience, i.e., users wait.
I think this wait is very unreasonable. Why should he wait? Because cloud cost is too high. Offline voice-to-text consumes fewer Tokens than real-time interaction. So within today's payment capacity, a trade-off was made to offer offline service. But I think this must change; users want instant gratification — say something, see it immediately.
I think we should ship real-time capability soon. Token cost will keep dropping. Luckily there's more than one large model worldwide; actually about 10 large models compete, which is good for us application companies. Competition keeps lowering cost, and we choose the best cost model meeting user experience, giving users the best ROI.
Users don't really care which model you use; though we offer manual choice, those hardcore users are few. Most middle-aged/elderly users don't know. They want the best experience and best value.
So here are a few things to repeat: first, reduce wait time. Global inference capex this much is mainly to strengthen inference speed. We also use large-model acceleration to give users the best instant experience.
Second, through our global scheduling, get best value. Our priority is clear: first, fast UX; second, Cost Down. We Cost Down not to profiteer and hang subscriptions high. My philosophy is to lower cost so users' subscription is very cheap. He feels it's great and affordable. Because I think more of ordinary users in Brazil, Central Asia, not the California tech elites, so they can afford it too.
Wenqin Luo: This resembles cloud-era peak-shaving and usage scheduling?
Hua Kun: A bit. Because I did Infra; first I negotiate resources, second I use tech to schedule. But to users, latency and performance are OK, and I lower cost. Though we're doing apps, I found scheduling very important, directly about user wait time and payment ability; it's a real problem.
Wenqin Luo: Good. Same business-model/willingness question to Mr. Deng. I understand your glasses and ring serve similar business-elite people. After buying your glasses, why buy your ring?
Deng Xudong: I think this can be thought bigger. When a new compute or interaction paradigm appears, how does the business model change? In PC and phone eras, PC typing wasn't extra-charged; calling in mobile internet was cheap. But which apps spent most in PC and mobile eras? In China, Didi, Dianping, Ctrip; in the US, DoorDash or Uber.
So strictly, the model of "directly charging consumers on a feature" may not reach ultimate value. What business model has ultimate value? I think it's still being explored.
Five or ten years later, when people don't reach for the phone so often, which apps give consumers stickiness? Shout and it's done — making a PPT, sending a schedule, booking tickets, hotels, rides. I think this may make the business model cleaner. We're exploring, because we want to make a whole set of Wearable Devices serving consumers' varied scenarios, enough to replace some or all phone functions.
Wenqin Luo: After the ring, what's next? New form factors?
Deng Xudong: From interaction, glasses have their advantage. First-person info displays perfectly, closest to the mouth for full voice capture; consensus. The problem is in the impossible triangle, weight and battery life are inevitable issues. So most display smart glasses now ship maybe tens of thousands to 200k units; except Ray-Ban Meta selling millions via channel advantage, there's no true hit ecosystem yet.
So glasses matter, glasses are on the way; the ring matters, very convenient interaction. What's next? It must be something that conveniently collects info and interacts. We're watching; anything wearable on a person has value. If you follow, see you at CES next year, only six or seven months away.
Wenqin Luo: OK, everyone will go to CES. On that, I want to ask Mr. Gu. You also launched a phone-side app tool beyond hardware at CES, right? So far in your revenue, what's the split between hardware (camera) and app-service tools? Do people prefer buying hardware or the whole cloud-service toolset?
Gu Yan: This depends more on our business model. Our product broadly counts as wearables, though not human wearables but pet wearables. But human wearables are almost all offline (like Bluetooth to phone), except kids' watches. Kids' watches are real-time connected, roughly a small phone. Our whole product line is 24-hour real-time connected, so it inherently generates carrier communication fees.
Our initial approach: besides hardware profit floor, users choosing this product must subscribe from the start, because communication service is itself subscription-type. But what we truly want isn't earning user money from communication. From actual user feedback, we have monthly and yearly subscriptions; the vast majority choose yearly.
What we truly value is the "subscription" value for pet users. Personally I think enterprises shouldn't directly pass cloud or AI cost to users; in fact most enterprises currently cover this large-model fee for users. We want to provide users capability and value.
In our direction, how to understand all kinds of pet behavior data is a dimension ordinary users can't do alone. He really wants to know what his dog is doing, how it did today, eat more or less. Traditional "food-clothing-shelter" products (like smart feeders) are more quantitative, timed-amount feeding. We, through getting the pet's exercise and health data, tell him eat more today, less tomorrow, what to supplement while growing. In this dimension we help users; we hope to charge from these value-added services.
Simply, part hardware, part communication. But we'll also fully fold communication into hardware, giving users a "one-off." What we truly want to earn is the extra understanding and insight value beyond basic functions.
But this reminds me, our track is a bit different. We often joke, on market willingness, the profit order is usually women, children, then men.
Like their digital products; I'm personally a digital-hardware geek, buying many products I even tear down to chip BOM cost, see which function is no good, DIY one myself. So it's hard to make hardware/software money off me. But when I buy things for my dog, I suddenly fully understand that irrationality women have buying for themselves or kids. That's the pet market we serve.
Wenqin Luo: So your track is a niche more willing to pay premium and service. OK Grace, time, quickly, your future business model.
Grace: I deeply understand Mr. Gu, because I also had a dog. Later my little dog died in a car accident, which started this entrepreneurship. I also strongly agree that women, when emotionally resonating, pay for "emotional value," so I focus on the women's market.
First, when we defined this, we didn't rigidly define it as smart hardware; we thought from "accessory jewelry" first, choosing women, the highest willingness-to-pay group. First it must be beautiful, emotionally valuable, decorative. Our smart jewelry can be worn on hand, body, and ear. Because this position monitors closer to the heart, it can track physiological cycles and emotional swings; I think there's big imagination. The business model is also hardware plus value-added service and emotional value.
Wenqin Luo: Understood; this market's ceiling is high enough, depending on how hardware and service forms show. Last question this round; I ask the center seat (Mr. Tu) to respond. Everyone mentioned different edge products call intelligence differently; from your platform view, in handling different edge products calling intelligence, is there tiered management? Which common scenarios don't need expensive models or high intelligence? How do you handle tiered applications?
Tu Xiwei: As a platform, we have an advantage. Because we mainly do smart cities, IoT as urban-infrastructure projects, helping government land solutions. In these big projects, many edge devices don't need high-level AI; even some basic sensors and transmission devices only need to upload data to edge-gateway devices with processing ability.
For these most basic devices, the first requirement is stable and low-power. Then at the gateway and edge layer, we do basic data processing. Of course the cloud has the strongest, most complete base compute and continuous iterative optimization. So our platform architecture clearly splits into these three layers (edge, side, cloud) for tiered application and resource scheduling.
Stage four: "heartfelt words" for entrants
Wenqin Luo: Clear. Thanks. Time, one line each. I know many AI entrepreneurs are here, both software and hardware. If each guest gives one suggestion for entering edge-intelligent hardware, what's your heartfelt word? One line. Starting with Mr. Hua.
Hua Kun: I think first make a usable hardware; supply chain matters a lot. Our overseas order demand is large; we've sold out twice. So I now spend lots of time running supply chain. I think for hardware, supply chain is the survival bottom line, a necessary condition. You can differentiate on Agents or experience, but first you must nail the supply chain.
Deng Xudong: Continuing Mr. Hua: on the device side, the device must first be a good device. As Grace said, the bracelet must be a good bracelet, glasses good glasses, ring good ring. To blend into users' daily carry, be a natural thing that accompanies you 16 to 24 hours. The simpler and more invisible the better, don't overcomplicate.
Gu Yan: On hardware, China's whole Pearl River Delta (Shenzhen) already has many mature solutions to learn from. For entrants, think: in future scenarios, should cloud intelligence and edge intelligence both be crammed into hardware? Make tradeoffs by very concrete scenarios. Which truly go to the cloud, which, under constraints of size, power, and cost, optimally pack into the edge. Think through the split; don't force smart for smart's sake; look at your real scenario pain point.
Grace: Borrowing Ms. Luo's stage to throw an olive branch. I hope everyone joins us to create. We now have Shenzhen's best supply chain, LV-level genuine molding, our own factory. On AI we focus more on human connection and emotion computing; we can detect 48 different emotions. Many AI hardware focus on efficiency and precise capture; we hope to explore from humans' deepest needs. Hope you join; this is a recruiting post.
Tu Xiwei: One short line. I see on government and large-enterprise sides there are still many real scenarios. As long as the pain hurts enough, clients are willing to pay continuously for many years. B2B and B2G still have lots of edge-intelligence business opportunities; everyone can pay more attention.