--- title: "Every Robotics Company Says It's Profitable. One Founder Admits: What We're Really Making Is Investors' Money" author: "Unique Research" sourcePublication: "Unique Research Substack" originalPublishedAt: "2026-09-09T14:09:03+00:00" canonical: "https://ffcap.cn/en/research/every-robotics-company-says-its-profitable" source: "https://uniqueresearch.substack.com/p/every-robotics-company-says-its-profitable" language: "en" --- # Every Robotics Company Says It's Profitable. One Founder Admits: What We're Really Making Is Investors' Money [![](https://substackcdn.com/image/fetch/$s_!maB6!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F703944b5-d2b9-4feb-bedb-303f37576296_2048x1152.png)](https://substackcdn.com/image/fetch/$s_!maB6!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F703944b5-d2b9-4feb-bedb-303f37576296_2048x1152.png) Compute costs less than 5% of a robot, but model companies burn millions each month; 18-month payback is the make-or-break line for industrial customers; one founder says plainly that what his company earns is investors’ money. Strip away the hype, and there are only three balance sheets that really matter in robotics. **Shenzhen Physical AI Summit · Panel Highlights** * * * The robotics industry will have truly begun the day customer revenue exceeds investor money in a robotics company's income statement. Today, a robot working on a factory floor has compute costs under 5,000 yuan—about 5% to 6% of the total machine cost. In the same industry, another company spends over a million yuan each month buying compute. Both numbers are real. The first comes from Nano Robotics, which makes industrial whole-machine robots. The second comes from Qiaojie Shuwu, which makes robot operating systems. Outsiders talk about robotics in terms of compute, parameters, and humanoid robots. But people actually working in this industry calculate entirely different ledgers every day. At a recent panel at the Shenzhen Physical AI Summit, three segments of the robotics supply chain were brought together: Wang Xin from Moxian Technology (robotic electronic skin), Song Yang from Qiaojie Shuwu (robot operating system), and Ma Yao from Nano Robotics (industrial whole-machine robots). All three are vendors. All three must justify ROI to their customers. And all three said things that go well beyond the usual talking points. Listening to the whole conversation, three balance sheets emerged as the most valuable. * * * Start with a fact many people haven't considered: today, the vast majority of robots consume surprisingly little local compute. Song Yang explains that most current robot models have only million-level parameters. Without complex perception and interaction, they don't need much compute to run. And you can't just stack more compute anyway. Because the moment you go higher, you hit a very physical problem: heat dissipation. He gave a personal example. Before coming to this event, he was at the company tuning a robot whose chest-mounted chip was overheating badly. In the end, there was no other solution: he found a cooling pad similar to what you'd put on a child with a fever, and stuck it directly onto the chip housing to bring the temperature down. In his view, many of today's robot execution and stability problems, when traced all the way back, may ultimately land on heat. Nano's numbers are more concrete. Five years ago, doing pick-and-place with vision and spatial positioning, compute costs inside one robot were over 10,000 yuan. Through algorithm optimization in recent years, plus moving some vision computation to the camera side, total machine compute costs are now controlled under 5,000 yuan—one-third of what they were five years ago, and about 5% to 6% of total machine cost. Wang Xin's electronic skin demands even less compute. The edge side uses conventional chips. Tactile data is collected and stored directly on the device—there isn't even a cloud interface reserved. So compared with pure software AI, edge-side compute costs in robotics are far less dramatic than outsiders imagine. The real compute heavyweight sits with model companies: Qiaojie's monthly compute spending is already at the million-yuan level. Song Yang judges that for robots to have full perception and interaction capabilities in the future, models will need at least billion-level parameters. At that point, some generation and planning capabilities must move to the cloud. But that's the future. At least today, what's choking this industry isn't compute. * * * Wang Xin works in tactile sensing. A dexterous hand with only vision and no tactile sensing runs into a visual blind spot: it thinks it's grasped something, when it hasn't grasped it at all. What tactile sensors provide is that last-mile judgment. The interesting part is his pricing model—calculated entirely backward. Embodied data has already formed a market rate: one hour of data, depending on dimensions, sells for a few hundred yuan. Data collection companies work backward from that: how much for labor, how much for facilities, how much for equipment. What's left for tactile collection equipment, amortized per hour, comes out to about $5. The customer doesn't care what business model you use. They give you one number: exceed this cost, and my data won't sell. So Wang Xin's glove pricing is anchored right there. The gloves are consumables—they contain electronics and they wear out. The model they've formed with customers: promise a 200-hour service life. 200 hours times $5 equals a price around $1,000. Normal damage within 200 hours gets replaced free. Beyond 200 hours, after-sales is no longer covered. The customer pays a fixed amount. The longer the glove lasts, the thicker the profit. In Wang Xin's own words: if you guarantee 200 hours and it ends up lasting 1,000 or 2,000 hours, the extra is your profit. So the one thing the company really competes on internally is: keep pushing product lifespan upward. This year he's also pushing data collection gloves—what the industry calls body-free data collection. A person puts on the gloves and works, and tactile, position, and spatial data all get collected—no actual robot needed. Wang Xin analogizes this model to data labeling in the large language model era: in the future, you could entirely organize large numbers of people in Southeast Asia or Africa to do collection. Before robots enter factories, what enters factories might be this new type of "human-flesh data labeling" job. That image is worth pondering. * * * Nano's ledger is the most direct, because its customers never beat around the bush when calculating. Nano's robots mainly do pick-and-place: phone glass, back covers, Mac panels, semiconductor wafers. From receiving a requirement to on-site delivery, it generally takes 30 to 40 days. Whether a customer will pay comes down to one line: 18-month payback. How does the calculation work specifically? One robot can replace 4 to 5 workers. The customer calculates the wages of those 4 to 5 workers over the next 18 months. If the robot's purchase price doesn't exceed that amount, and the equipment can last five years or more, they place the order. If you can compress the payback period to 15 months or even 12 months, the customer gets even more aggressive. As for gross margin, customers don't care—but Ma Yao himself draws a clear distinction. In 3C scenarios, an operation success rate above 99.5% is acceptable, with gross margins of 40% to 50%. In semiconductor and lithography-related scenarios, there are cleanliness requirements. Success rates start at 99.99%—what the industry calls "four nines"—and gross margins can reach 70% to 80%. When the company's annual recognized revenue reaches over 10 million US dollars, it can be profitable overall. Industrial sites don't pursue generalization. Fixed motions, fixed scenarios—the core is high completion rate. In Nano's terms, that 99.99% success rate is something many so-called general embodied robots still can't match. Moderator Wang Chaochao raised an observation: the real automation level in many 3C factories is two worlds apart from what the media calls "embodied intelligence." Nano's response was even more direct: just look at the employee counts of global manufacturing giants. Some have over a million employees, some 700,000 to 800,000. Employment numbers that large themselves prove that production lines are still packed with people. * * * Beyond the three ledgers, there's a fourth—one that Qiaojie calculates about its own company. Qiaojie Shuwu has been profitable since its founding in 2023. Across the entire embodied intelligence industry, that should count as rare. In 2024, a single project brought in hundreds of thousands to one million yuan. After industry heat picked up in 2025, average project prices rose rapidly. Now projects basically don't go below 3 million yuan, with most in the 4 to 5 million yuan range. Delivery cycles for large overseas projects are booked months out. But Qiaojie calls this state "low-quality profit." The reason isn't something he intends to hide: project-based work can't scale. Put more bluntly, what the company earns isn't even entirely money from robot body companies. In a sense, it's earning money from the investors behind those body companies. They're quite clear about this themselves. A startup CEO, in front of peers and a live audience, dissecting the source of his own profits to this layer—this isn't common in the industry. His solution is platformization. On June 29 this year, Qiaojie's general robot development platform released its official version, and began commercialization just over a month later. There's a cultural tourism customer operating amusement parks across several provinces, putting on many robot performances every day. A traditional enterprise like this can't possibly maintain a technical team to "wait on robots" every day. Now, through this system, robots can dance and greet guests at the venue every day. What Qiaojie is doing is compressing projects that used to take one to two months to deliver and cost several million yuan into deployments completed in a few minutes. Compressing fees from several million yuan to a few thousand yuan per year in licensing. He even judges that as more machines connect, it could drop to a few hundred yuan per year. From 5 million yuan per project to a few hundred yuan per year. What stands between them is the entire distance of this industry moving from "circling money" to "running a business." * * * On what stage the industry is actually at, Song Yang had this to say. The companies getting the brightest spotlight in today's market—I don't think they can even truly be called industry leaders yet. The reason is simple: companies like Huawei and BYD, with enormous industrial resources and engineering capability, haven't truly entered the field in a comprehensive way. His judgment is that today's market is to a large extent ripened prematurely by policy and capital. Hardware, bodies, models—all still need ten or even twenty years of accumulation, especially real, effective data accumulation. From this perspective, the effective data in today's industry "can even be approximately understood as roughly equal to zero." He even admits that among the many scenarios Qiaojie serves, a considerable portion are still pseudo-demands right now. This sobriety is also written into the expansion posture of all three companies. Wang Xin expanded his team from a few dozen to over 100 people in 2025, but has clearly slowed down this year. The principle: no disorderly expansion, don't hire infinitely just because the market is hot. Qiaojie has fewer than 80 people total—full-time employees, pending hires, and interns combined. His exact words: a startup is best run as a company that isn't so anxious. Nano goes further—it hasn't taken institutional funding to date, mainly spending its own money, with a team of 70 to 80 people. Every year it focuses on only four things: whether core technology has advanced, whether niche market share has grown, whether component costs have fallen, and whether technical capability has iterated. * * * It's been almost ten years since AlphaGo won that Go match. Boston Dynamics' robot dog has been doing backflips for many years too. Technological breakthroughs, looked at in retrospect, all happen in an instant. Living through them, you realize it's all long accumulation. The real watershed for this industry won't be found by watching launch events and dance videos. Look at the financial reports: the day customer revenue truly exceeds investor money in a robotics company's income statement—that's the day this industry has truly begun. Until then, the cooling pads will have to stay on. * * * **Panel | From Showmanship to Productivity: The "Last Mile" of Embodied Intelligence Commercialization** **Guests:** - Wang Xin, CMO, Moxian Technology - Song Yang, COO, Qiaojie Shuwu - Ma Yao, Founder, Nano Robotics **Moderator:** - Wang Chaochao, Partner, Unique Capital * * * \*\*Wang Chaochao:\*\* Our three guests today sit at different segments of the robotics supply chain: Mr. Wang does perception, Mr. Song does underlying robot models and operating systems, and Mr. Ma directly makes whole-machine robots and industrial delivery. We'll discuss roughly four questions today. First, please each briefly introduce yourself and your company's business. \*\*Wang Xin:\*\* I'm Wang Xin from Moxian Technology. We mainly make robotic electronic skin. Our current applications in embodied intelligence go in two directions: one is tactile sensing on dexterous hands, and the other is data collection gloves. We started working on electronic skin in 2023, though the company was founded earlier, in 2021. Our team comes out of DJI Innovations. Our headquarters is now in Songshan Lake, Dongguan, and we have our own factory. \*\*Song Yang:\*\* I'm Song Yang from Qiaojie Shuwu. The moderator just said we make the "robot cerebellum"—that's actually the industry's affectionate misnomer for us. We've always felt that using medical terms like "cerebrum" or "cerebellum" to define an AI model company is somewhat imprecise. Our self-positioning is a general robot operating system developer. Since our founding in May 2023, over three years ago, we've served more than 50 domestic and overseas body companies and over 70 robot models. We mainly provide general robot motion control capabilities. Our company is based in Shenzhen. \*\*Ma Yao:\*\* I'm Ma Yao from Nano Robotics. We mainly focus on production logistics, including composite robots for CNC scenarios, robots for PCB/SMT scenarios, and also humanoid dual-arm robots for packaging and assembly. Simply put, we make whole-machine robots, mainly delivered in industrial production and production logistics-related segments. * * * \*\*Wang Chaochao:\*\* You can't talk about robotics without compute. Training needs compute, inference needs compute, going from simulation to reality needs compute. But compute ultimately is cost. So my first question: starting from the most daily, most frequently executed action of your product—say grasping something, perceiving something, or having a robot complete a motion—how much compute does that actually consume? Converted to money, what percentage of total machine cost or service cost is this? How do you balance "enough compute" and "controllable cost"? Let's start with Mr. Wang. \*\*Wang Xin:\*\* We're somewhat special because we're hardware makers. For example, our data collection gloves store data mainly on the device side. For customer privacy and security reasons, our devices themselves don't reserve backend or cloud interfaces. When customers use them, data mainly stays on their own edge side. Tactile data is actually just one dimension in Physical AI training data. Beyond tactile, there's now a lot of egocentric—first-person vision data—as well as position, spatial, and other data. Our downstream customers fuse different dimensions of data, then do simulation and training. So for us, what we really bear is mainly hardware cost, and hardware wear and tear during use. \*\*Wang Chaochao:\*\* To be more specific, within your hardware cost, what percentage does the edge-side chip or compute chip account for? \*\*Wang Xin:\*\* We actually don't demand much chip compute. The chip is mainly responsible for data collection and analysis of tactile sensing points, so the edge side uses relatively conventional chips—no need for very high-compute platform-level chips. \*\*Song Yang:\*\* This question is actually very important for the embodied intelligence industry. At the company, I'm usually responsible for daily operations—put more directly, responsible for how money is spent and earned. So compute does occupy a fairly important position in our monthly expenses. But from the robot side, today the local compute consumption of the vast majority of robots actually isn't as large as everyone imagines. Why? Because many robot models today still have only million-level parameters. At this stage—without extensive perception, without complex interaction—local compute first doesn't need to be particularly high, and second actually can't be particularly high. Because the moment compute goes high, you immediately run into a very realistic problem: heat dissipation. Before coming to this event today, I was still at the company tuning a robot. The chip on its chest was overheating quite badly. In the end, there was no other way. I found a cooling device similar to what you'd use on a child with a fever, and stuck it directly onto the chip housing to help it cool down. Many of today's robot execution problems and stability problems, traced all the way back, may ultimately land on heat. So at this stage, robot local compute can't be stacked infinitely. But the future is certainly different. If robots are truly to enter thousands of industries and thousands of households, million-parameter-level models clearly won't be enough. We judge that if robots are to have relatively complete perception and interaction capabilities, models may need to reach at least billion-level parameters. Of course, compared with large language models, that's still several orders of magnitude behind. At that point, you can't simply rely on local compute. Some model generation and complex planning capabilities must move to the cloud. So we're already building a general robot motion control development platform, hoping to offload some generation and planning compute through the cloud. From the company's own perspective, we're a model company and need to buy a lot of compute. At this stage, our monthly compute spending is already roughly at the million-yuan level. But overall, because model parameter counts are still limited today, compute hasn't yet constituted a particularly large operational difficulty. As platform scale expands in the future, the real question will become: after I buy compute in large quantities, when exactly can I sell it? \*\*Ma Yao:\*\* Many of our robots are delivered in workshops with confidentiality-level requirements, so data doesn't go to servers—it's basically all computed on the robot itself. About five years ago, when we did grasping, because it involved vision, spatial positioning, and other computation, compute costs inside one robot might have been over 10,000 yuan. In recent years, algorithms have continuously optimized, and we've also moved some vision computation to the camera side. Now the cost of the entire compute portion can basically be controlled under 5,000 yuan—roughly one-third of what it was five years ago. Compute cost as a percentage of total machine cost is now roughly 5%–6%. \*\*Wang Chaochao:\*\* So compared with pure software AI, today the compute cost in robotics or embodied intelligence, as a proportion of total product cost, actually isn't as high as everyone imagines. * * * \*\*Wang Chaochao:\*\* Next let's talk about scenario closure. Each of you, talk about your currently most mature product—in plain language: what work is the robot actually doing? In what scenario? From deployment to actually starting work, how long does it take? What does the customer ultimately care about most when paying you? \*\*Wang Xin:\*\* We currently mainly have two areas: dexterous hands and data collection. On the dexterous hand side, starting around 2024, more and more dexterous hand manufacturers have been adding electronic skin—tactile sensors—to the hand. Why add it? A very direct problem: if there's only vision and no tactile sensing, when a dexterous hand or industrial gripper executes a task, it runs into a visual blind spot. It might "think it's grasped something," when actually it hasn't grasped it at all. What tactile sensors provide is that last-mile judgment capability. The second area is data collection. In the past, when people did embodied data, a lot was still collected using real robot bodies. That meant the body and dexterous hand themselves also needed to collect tactile data. This year we're working on another solution: data collection gloves. They contain both tactile and position/spatial information. The industry now calls this approach body-free data collection. Why are customers willing to pay for this? Simple: because collecting data with real machines is too expensive. If you do body-free collection, it's a bit like data labeling in the large language model era. In the future, you could entirely organize large numbers of people in Southeast Asia, Africa, and other regions to do collection. Including some internet companies that naturally already have extensive data collection scenarios—they only need to add data collection equipment to people, and they canlarge-scale collect data back, then do simulation and training. So what we provide to customers is essentially an important dimension in robot training data. If robots are truly to move forward in the future, tactile data will definitely be an indispensable link. \*\*Wang Chaochao:\*\* Mr. Song, you make OS and embodied models—how does your product actually land? \*\*Song Yang:\*\* Our customers mainly fall into two categories. The first category is customers developing robot bodies. Within that, there are two types: one is relatively leading robotics companies in the industry; the other is traditional manufacturing companies that havelarge-scale entered the embodied intelligence field starting from 2024, including automotive, mobile phone, home appliance, industrial robot, and collaborative robot companies. What we provide to this category of customers is underlying infrastructure and system capabilities. The second category is enterprises that actually use robot bodies. This category covers a much broader range of scenarios, including petrochemicals, power operators, logistics supply chains, and so on. There indeed exist extensive high-frequency, reusable robot demands in these industries. But I'll also be frank: today the vast majority of them may still be at a very early stage, and even a considerable portion are still pseudo-demands right now. It's just that the combined push of policy and capital has led legged robots and embodied robots gradually entered these scenarios. So what Qiaojie does is provide a cross-scenario infrastructure capability. Whether you're developing a body or using a body, you don't need to reinvent the wheel. Through a set of standard APIs, you can call our underlying robot capabilities. Later we'll further release "cerebrum" capabilities, gradually merging the so-called cerebrum and cerebellum into the same model. Ultimately we hope users don't need to understand the underlying technology at all—don't need to know what configuration this robot is, what its spatial model is. Just like using a PC or phone today. You don't need to know what chip is inside or how much memory—you only need to know: what function do I want it to complete? So what we provide is inherently cross-body, cross-scenario capability. \*\*Wang Chaochao:\*\* The auto industry used to have the "soul doctrine." One type of company makes both the body and the soul itself; another makes the body itself and buys the soul externally. You're equivalent to "selling souls." Among your current partner customers, which type is more common? \*\*Song Yang:\*\* Indeed, many investors analogize Qiaojie to a "soul-selling" company in the robotics industry. This analogy isn't entirely precise, but it's somewhat similar. To answer directly: the second type is more common now—that is, body and underlying intelligence capability are procured separately. \*\*Wang Chaochao:\*\* Mr. Ma, what specific work are your industrial robots doing? \*\*Ma Yao:\*\* What we do most currently is still pick-and-place. Mainly various workpieces in the 3C and semiconductor industries. For example, phone glass, back covers, Mac panels, and some wafers in the semiconductor industry—these pick-and-place tasks are what we do most. From receiving a customer requirement to ultimately being able to go on-site for deployment and delivery, it generally takes about 30–40 days. Customers calculate very directly: using this robot to replace labor, can the cost be recovered in about 18 months. This is a relatively typical calculation method for industrial customers. * * * \*\*Wang Chaochao:\*\* That leads nicely into ROI. All three of you share a common identity: you're all vendors, sometimes even vendors' vendors. Today customers, investors, and startups are all calculating ROI. So please each of you calculate a specific ledger. Using your most mature scenario as an example: how much does the customer invest, how much do they get out, how long to recover cost? Also, how do you as suppliers price? Where is your own positive gross margin node? \*\*Wang Xin:\*\* Many of the customers we serve are doing embodied models and world models, and upstream of models there are extensive data collection companies. Data in this industry has gradually formed some relatively clear prices. For example, one hour of data now, depending on data dimensions, might sell for a few hundred yuan. Then data collection companies work backward: how much is the labor cost inside, how much for the venue, how much for equipment, and finally how much is left for tactile collection equipment. Currently a relatively clear cost constraint customers give us is that tactile equipment, amortized per hour, is about $5. They don't care what business model you use. They only tell you: this is the cost I can accept—any more expensive, and my data won't sell. Because model companies, beyond buying data, also have compute costs, and overall training costs are already very high. So our pricing is calculated backward. For example, the data collection glove itself is a consumable. Even an ordinary glove, with long-term operation, will definitely wear out. Our gloves also contain electronics, and different scenarios and tasks will have different degrees of wear. Now we've gradually formed a model acceptable to both sides with customers: I guarantee your usage cost is roughly $5 per hour. For example, we promise customers a 200-hour service life. 200 hours × $5 = $1,000. Then the price of this set of equipment might be in that range. I guarantee that within 200 hours, if normal damage occurs, I'll replace it for free. Beyond 200 hours, I no longer bear this part of after-sales cost. So what we really need to do internally is continuously push product lifespan upward. Because the money the customer pays has already been calculated backward. How long the product actually needs to last to cover R&D, pre-sales, after-sales, and still have its own profit—that ledger becomes very clear. \*\*Wang Chaochao:\*\* So for you, the real pressure is making a pair of gloves more and more durable. \*\*Wang Xin:\*\* Right. If you guarantee 200 hours but it ends up lasting 1,000 or 2,000 hours, then the extra is your profit. \*\*Wang Chaochao:\*\* Mr. Song, how do you price this "soul"? \*\*Song Yang:\*\* We still distinguish between R&D customers and usage customers. Frankly, the entire industry is still very early. Most of Qiaojie's current revenue actually comes from customers developing robot bodies. I can share a data point: since our founding in 2023, the company has always been profitable. Across the entire embodied industry, that should be relatively rare. In the early days we were mainly project-based. In 2024, a single project charged roughly hundreds of thousands to one million yuan. After 2025, the entire robotics industry's heat suddenly rose, and our project average prices also rose rapidly. Now projects basically don't go below 3 million yuan, with most in the 4–5 million yuan range. We even have large overseas projects whose delivery cycles are already booked months out. But I myself would call this state: low-quality profit. Why? Because project-based work can't scale. Put more bluntly, what we earn isn't even entirely money from robot body companies. In a sense, what we earn is money from the investors behind those body companies. We're quite clear about this understanding ourselves. So why must we consolidate capabilities into a platform? Because only with platformization can the business model truly potentially move from project-based to scaled. Our platform's official version was released on June 29 this year, and just over a month after release it had already begun commercialization. For example, we now have a cultural tourism customer operating amusement parks in different provinces, putting on many robot performances every day. For a traditional cultural tourism enterprise, it obviously doesn't have the capability, and can't possibly maintain a large robot technical team itself, specifically "waiting on robots" every day. After seeing our system at an exhibition, they can now, through this system, very simply have robots dance and greet guests at the venue every day. This model can be replicated to other scenarios. Going further, we've also started serving many real operation-type demands, including factories, petrochemicals, operators, and other scenarios. What we're doing now is using systematic products to compress projects that used to take one to two months to deliver and cost several million yuan into deployments completed in a few minutes. At the same time, compressing fees from several million yuan to a few thousand yuan per year in licensing. Even as the number of connected machines grows, we judge that in the future it could drop to only a few hundred yuan per year. This is the truly scalable model. \*\*Ma Yao:\*\* Our customers are relatively direct. As I just said, 18 months is the limit for customers calculating investment recovery cycles. For example, one robot can replace 4–5 workers, then the customer calculates how much wages those 4–5 workers will need over the next 18 months. If the robot's one-time purchase cost doesn't exceed that amount, and the equipment can be used for five years or more, they're willing to buy. If you can achieve a recovery cycle of 15 months, or even 12 months, then customer purchasing becomes even more aggressive. As for our gross margin, customers actually don't care much. We have large differences depending on different application scenarios. For example, 3C, including some Apple supply chain customers, although their requirements for equipment operation success rate are also high, they're relatively less extreme than semiconductors. Achieving above 99.5%, many scenarios are already acceptable. For this type of equipment, our gross margin reaches about 40%–50%. But semiconductor and lithography-related scenarios have stricter requirements such as cleanliness, and operation success rates may require above 99.99%—that is, above "four nines." For this type of high-precision, high-requirement equipment, gross margin may reach 70%–80%. For our company, if annual recognized revenue can reach over 10 million US dollars, it should be possible to achieve overall profitability. Ultimately it's still a process of obtaining profit through scale replication and scale cost reduction. * * * \*\*Wang Chaochao:\*\* After hearing Mr. Ma's data, I have a feeling. The actual automation level in many 3C factories today, and what outside media and self-media call "embodied intelligence," or even the degree of automation AI achieves on computer desktops today—might be completely two worlds. Are there actually still many people in industrial sites, repeating very basic labor? \*\*Ma Yao:\*\* Yes. You only need to look at the employee counts of many global manufacturing giants. Some enterprises have over a million employees, some 700,000 to 800,000, 500,000 to 600,000. Employment numbers that large themselves prove that manufacturing sites still have many frontline workers. \*\*Wang Chaochao:\*\* Then do you face the problem of "robots not being flexible enough"? \*\*Ma Yao:\*\* Industrial sites have very high requirements for operation success rates. Even in 3C, a scenario with relatively higher tolerance, you need 99-point-something percent. Semiconductors may directly require starting at 99.99%. This success rate is something many so-called general embodied robots actually still can't match. \*\*Wang Chaochao:\*\* So in industry, there's conversely less pursuit of generality and generalization. In fixed motions, fixed scenarios, the core is high completion rate, high accuracy. \*\*Ma Yao:\*\* Right, high precision. * * * \*\*Wang Chaochao:\*\* The last question is about financing and operations. Over the past two years, the robotics capital market has continued to heat up, but the market has also been continuously clearing. Some companies are expanding, some are contracting. The three of you happen to be at different stages: some have raised a lot of money, some are on the financing path, and some currently mainly spend their own money. How do you balance financing and operations? How do you judge what size the company needs now? Too many people might burn the company to death, too few and you can't get moving. \*\*Wang Xin:\*\* Our fastest expansion was actually in 2025. The team grew from a few dozen to over 100 people. We're still expanding this year, but the speed has clearly slowed. Our business model also has quite a few project-based work, though it's different from pure software projects. Many of our projects arearound specific dexterous hands doing customization. This means upfront R&D investment is large, but product life cycles are also relatively long. For example, when we do a complete customization solution for some dexterous hand companies, the front is definitely losing money, because there's extensive R&D investment. But once it goes into mass production, the life cycle is relatively long, and it's not so easily directly replaced by standardized products. Generally, about 3–5 months after entering mass production, this project can start becoming profitable. So currently the company's overall profitability is still okay. Our thinking is no disorderly expansion. There are many dexterous hand companies in the market, and we can't do work for every one of them. We'll prioritize choosing customers with consistent product philosophy and future scaling opportunities. Beyond robotics, we ourselves also have different directions such as wellness and healthcare, AI hardware, lawn mowers, robot vacuums, smart beds, and so on. So the overall principle is: grow with the embodied intelligence industry,allocate corresponding people, but try to keep headcount within a reasonable range—don't hire infinitely just because the market is hot. \*\*Song Yang:\*\* I'll speak from both industry and company perspectives. First, the industry. The market now pays special attention to some star robotics companies, including some capital market moves. But from the perspective of a third party like us that doesn't make bodies at all, I conversely feel: today embodied intelligence is still at an extremely early stage. The companies getting the brightest spotlight in today's market—I don't think they can even truly be called "industry leaders" yet. The reason is simple. Companies like Huawei and BYD, with enormous industrial resources and engineering capability, haven't truly entered the field in a comprehensive way. If these companies' resource capabilities are truly invested in the future, the entire competitive landscape could be completely different. So today's robotics industry is far from mature enough to write its epitaph. It's still early from truly going into factories and screwing in screws at scale, in my view. I'd even say today's market is to a large extent artificially ripened, pushed forward by policy and capital. Whether it's robot hardware, bodies, or the so-called cerebrum and cerebellum—all still need a long time to accumulate. Maybe 10 years or even 20 years. Especially the accumulation of real, effective data. From this perspective, the effective data in today's industry can even be approximately understood as "roughly equal to zero." So making any particularly definitive judgment now is still too early. Second, from the company operations perspective. At the company, I'm responsible for both financing and sales. I don't think financing and commercialization are contradictory—they should instead promote each other. Financing gives us more ammunition. Although Qiaojie has been profitable from day one of founding, that doesn't mean relying only on existing profits will necessarily allow rapid development. We still need external capital and resources to help the company cross cycles, giving us a chance to truly reach the day the industry matures. But commercialization is equally important. Because commercialization is essentially using market-based means to verify: is your model actually useful? Is your platform capability actually valuable? It's not success just to lock yourself in the company and feel the technology is good. So from day one of founding, whether developing cerebrum, cerebellum, or other capabilities, one of the first things we think about is: how do I sell this to customers? How do I commercialize it? Another point I feel is particularly important: a startup is best run as a company that isn't so anxious. \*\*Wang Chaochao:\*\* Having money in hand of course makes you less anxious. If Moxian is now over 100 people, how many are you now? \*\*Song Yang:\*\* Full-time employees, pending hires, and interns all combined—fewer than 80 people. \*\*Wang Chaochao:\*\* Mr. Ma, how do you balance financing, operations, and team size? \*\*Ma Yao:\*\* We currently haven't taken institutional financing. Up to now, we mainly spend my own money. The team is now about 70 to 80 people. So I currently mainly look at a few things. First, whether the core technical capability of our own products has continuous progress every year, and whether it can maintain leadership over peers. Second, in the niche markets we truly value, whether the market share of core customers has improved. For example, in a niche market we're focusing on this year, a total of 3,000 units were sold—then I need to see whether our own share is rapidly increasing. Third, whether the procurement cost of core components has continued to fall. Fourth, whether the technical capability we ourselves possess has continued to iterate. So our thinking is: within the scope of losses we can bear, continuously improve niche market share, while expanding scale and gradually forming scale effects. The company is currently roughly at this stage. * * * \*\*Wang Chaochao:\*\* After listening to all three share, I have a very clear feeling. Today's so-called robotics industry may actually have two markets coexisting. One is more oriented toward research, R&D, and manufacturing "embodied intelligence itself"; the other has already begun truly entering civilian and industrial scenarios. The latter is obviously closer to real marketization and commercialization. The robotics industry has indeed developed very fast over the past few years. Using the word Mr. Song just used—it may have been "ripened" under the influence of capital and policy, and has already been ripened into a fairly substantial niche market. Above the body there are various components and perception; below there are models, operating systems, and scenario integration. There are already very many roles in the industry chain. There's indeed still a lot of money in it now. Both investors' money, and money that has already begun to truly come from the market, from customers. But now it may be entering another stage: you can't only look at whether a robot can perform—you have to look at whether it can be productized, whether it can truly enter industry, commerce, and civilian use. This process doesn't necessarily have a clear critical point—it's more likely a gradual process. It might take one year, two years, or three to five years, ten years. Looking back, Boston Dynamics has developed for so many years, and AlphaGo is also about ten years ago now. Many times you feel technology has suddenly broken through, but looking back, it's actually been accumulating for a very long time. Robotics truly forming large-scale civilian and commercial products may also still need time. * * * This article is a complete English translation of the original Chinese article published by Unique Research (非凡产研) on September 8, 2026. Original title: 机器人公司都在喊盈利，但有人坦白：我们赚的其实是投资人的钱 Original source: https://mp.weixin.qq.com/s?\_\_biz=MzU5Mjg5MjQ5Ng==&mid=2247522518&idx=2&sn=78916b40edafb424 Original publish date: 2026-09-08 18:26 (Asia/Shanghai) Panel: Shenzhen Physical AI Summit · "From Showmanship to Productivity: The Last Mile of Embodied Intelligence Commercialization" Guests: - Wang Xin (王鑫), CMO, Moxian Technology (墨现科技) - Song Yang (宋阳), COO, Qiaojie Shuwu (桥介数物) - Ma Yao (马尧), Founder, Nano Robotics (纳诺机器人) Moderator: - Wang Chaochao (王朝超), Partner, Unique Capital (非凡资本) This translation preserves all substantive content, including all numbers, names, companies, and quoted statements. No substantive content has been omitted, summarized, or altered. --- Original publication: https://uniqueresearch.substack.com/p/every-robotics-company-says-its-profitable On-site reading page: https://ffcap.cn/en/research/every-robotics-company-says-its-profitable