Summary

The clearest consensus from the closed-door discussion with ABB, Topstar, Unitree and Xingdong Jiyuan was that the industry is headed for a shakeout. Asked whether any leading embodied-AI companies were overhyped, the host later summed up the response as everyone but Unitree raising a hand. William王浩然 provided the structural math: the top 10 companies account for 40% of industry funding, while the roughly 200 others may average only tens of millions of yuan to RMB100M each; companies that cannot raise money “may find it very difficult to survive.” Six or seven leaders have already surpassed RMB10B in valuation and should have no trouble surviving, while the winter is visible to the naked eye below the middle tier. 南立新 believes the truly brutal competition may only just be starting, because too much money has entered the sector.
The second cross-camp consensus was that industrial deployment will precede home use, and traditional industrial robots augmented with AI may move faster than humanoids in the near term. ABB’s 郑锡亮 argues that factories are highly ordered and their tasks have already been decomposed; a narrow application such as container welding at CIMC does not require complex technology. Unitree’s 谢一鹏 added the sociological case: “People have to accept new things in some environment, and that environment may be the factory.” The participants broadly agreed that hardware gaps have narrowed relatively quickly; the bottleneck is software and generalization.
On funding, the transition camp was more pessimistic than the newcomers. 郑锡亮 said he had heard that a single round of humanoid simulation and synthetic-data training could cost $20M; making a system handle general-purpose tasks will expose corner cases, and “that cost will definitely rise exponentially,” meaning RMB1B on the balance sheet could disappear quickly. 南立新 expects no funding shortage in 2026; “the real shortage may come in 2027 and 2028.” Commercialization will not arrive in 2026 either—the industry is still waiting for its “AlphaGo moment.”
Unitree is using a product-company model to avoid heavy project-based delivery and build a scale flywheel. It delivered more than 6,000 humanoid robots last year, largely centered on one model, and is targeting 20,000-30,000 this year; manufacturing scale should support margins, which can then fund internal technical iteration. 谢一鹏 said the scarcest resource is not money but time: he personally measures parts on the production line and performs sorting motions at logistics sites. “There’s no time to carry the robot over and try it—I have to do the worker’s motion myself.”
Xingdong Jiyuan offered the sharpest framework for choosing applications: the value of having a robot take the risk for a human is greater than simply replacing a human. In airport bomb disposal, an operator could directly control a robot to dismantle the bomb; if the robot explodes, “buy another one.” Pneumoconiosis, spray-painting and similar applications do not require high precision. The newcomer’s catch-up list includes CE, UL, PSE and CCC certifications, cross-border data compliance, warranties and financing—all routine in traditional industries but still unfinished work for embodied AI. Its proposed data flywheel has European customers fund the work, China develop the technology and the Philippines handle teleoperation, at a monthly wage of RMB1,500 versus €1,000-2,000 in Europe.
The industry’s progress-bar consensus is around 30%, placing embodied AI in a “Nokia era before the iPhone.” Robin estimated that China has roughly 500 robots per 10,000 workers, among the world’s top 3, while South Korea leads at about 1,000; industrial robots are around 30%-40% of the way through a 50-year development cycle and have already entered single-digit growth. Cross-company data reuse needs to be evaluated by layer: visual and environmental data may transfer, but teleoperation trajectories—Xingdong’s L7 is 1.7 meters tall—and force-control data may not. A de facto standard could emerge from the data format behind an especially convincing industry application, potentially faster than expected.
The home market’s endgame is a bet: 郑锡亮 and his colleagues are wagering on whether humanoid penetration in households can exceed 1% within 10 years. William’s bearish counterpoint is that household generalization could cost “more than you can believe”; the biggest competitor may be a Filipina domestic helper earning HK$4,000 a month. But robots have scale economics: once a model is trained, “10,000 robots can all perform the same action,” allowing the system to scale up. The key accelerator is whether an embodied-AI OS emerges—an OpenClaw-style agent that can call each company’s VLA model. The host compared mass adoption of smartphones: beyond an iPhone moment, the industry also needs Android’s open ecosystem.

Deep dive

1. 徐勐 Sets the Frame: Why Shenzhen, Overseas Expansion and Embodied AI

徐勐 offered three reasons: the Yangtze River Delta and Pearl River Delta have the most complete supply chains, drawing a large concentration of embodied-AI companies; many embodied-AI startups “have overseas DNA from day one” or are targeting foreign markets; and these companies can iterate products rapidly as AI advances. The sector is entering an explosive growth phase—“this year is the year we walk into the sites, meet the customers and enter real-world practice.”
Host 卫诗婕 asked the sharper question: “Isn’t it also because embodied-AI companies currently have relatively large cash balances?” 徐勐 admitted the question was “a little awkward,” but said the relevant issue was the outlook. More than 80% of the companies at the foundation of SAP’s customer pyramid are small and midsize businesses; Linde, Siemens and Bayer were not particularly large when they first partnered with SAP.

2. SAP’s 30/70 Anxiety Split: AI Is Rewriting the SaaS Standard

Responding to claims that the previous generation of industrial-AI companies belongs in the history books, 徐勐 gave a “30/70” answer: 30% anxiety—“the capital markets may be more anxious than we are”—and 70% excitement. “The AI in future SaaS companies will grow out of data and applications. That is the standard for the next generation of SaaS.”
SAP sees three ways to create value for embodied-AI companies: driving cost reduction and efficiency gains as B2B production scales, while supporting overseas compliance; supplying the “nutrition” embodied AI most needs—context-rich data generated by customers; and connecting its Business AI solutions with embodied-AI companies so agents can penetrate from the physical world into the digital world and back again.

3. ABB Rejects the “Old Guard” Label: Paradigm Shifts Leave No Room for Path Dependence

Robin郑锡亮 opened by joking that “my senior colleague immediately put ABB in the old-timer category,” then laid out the credentials: ABB developed its first fully automatic, fully electric robot in 1974, has spent 50 years in robotics, and began working on dual-arm robots in 2015. “We should say we are pioneers in embodied-intelligence robots.” Anxiety and the instinct to survive are what keep the company humble.
His core view is that the industry is going through a paradigm shift. “You can’t have path dependence. You have to use new methods, just as with autonomous driving.” ABB is proposing “autonomous multifunctional robots” that move beyond being programmed by humans to execute tasks: the robot plans automatically, perceives, decides and executes on its own, while switching across multiple scenarios. He sees himself as an internal entrepreneur—“from veteran to rookie”—with anxiety, but more excitement.

4. The Four Players’ Confidence: Industrial Robots May Lead in Single-Point Applications

Robin ran the degrees-of-freedom math: industrial robots typically have 4-6 degrees of freedom, or 7-8 with an external axis; humanoids can reach more than 80. The two types therefore have different breadths of application. Industrial deployments, however, demand stability, reliability and precision. Industrial robots have a natural advantage in narrow use cases because they do not need to adapt to many scenarios at once; they can go deep and narrow in a specific application.
His emphasis was explicit: “In single-point applications, I personally think industrial robots may move ahead of embodied intelligence.” The scale of existing deployment in machine vision and industrial vision is the evidence.

5. Topstar’s Transition and “Secure the Home Front Before Going Abroad”

Ryan王琪 reviewed nearly 20 years of evolution: Topstar started with peripheral injection-molding equipment in 2007, entered Cartesian robots in 2010, moved into 4-6-axis industrial robots in 2013, launched a wheeled humanoid robot in 2025 and released a wheeled quadruped robot in January this year. The company has remained “oriented toward customer needs.”
SAP’s role is to “secure the home front before going abroad.” Topstar has multiple business lines, and planning, orders, warehousing and logistics, finance and other functions need standardized operations. SAP provides the operating foundation, “so we can expand outward with peace of mind.”
Robin described ABB’s global coordination with SAP through the restaurant analogy: ERP logic is “one global chessboard,” but local operations must adapt and coexist. “It’s like an international restaurant: the menu is broadly the same, but it has to fit local tastes, and it has to keep introducing new dishes.” SAP helps ABB remain flexible and agile between global coordination and local adaptation.

6. Unitree’s Differentiation: Use High-Pressure Scenarios to Force Technical Progress

Peter谢一鹏’s one-line positioning is a focus on high-performance legged and humanoid robots, with industrial partnerships dating back to 2017. Inspection was the first industrial “small closed loop” to work: high-value, asset-heavy continuous-production factories that are extremely sensitive to downtime have already felt value being transferred and created.
The host’s read is that Unitree knows how to use high-pressure platforms to its advantage. Appearances at the Spring Festival Gala continually force technical progress. After working closely with performance specialists, Peter said, “the robot can learn a new performance in a few hours or a few days.” Unitree is pursuing the same depth and value creation in industry. Asked about 王兴兴’s changing timeline for humanoid deployment, he only said that the speed of technical development and the breadth of scientific exploration have reached unprecedented levels, and “everyone can keep looking forward to what comes next.”

7. Xingdong Jiyuan: A Stand-Up Comedian Selling Robots, with 52% of Orders Overseas

William王浩然 disclosed that he was a full-time stand-up comedian before entering the industry, with an academic background in mechanical engineering and automation—“so it’s not as though I entered the wrong field.” Xingdong Jiyuan’s three strengths are world-class brains and dexterous hands; internationalization “written into its DNA”—founder 陈建宇 graduated from UC Berkeley and roughly 52% of the company’s orders come from overseas; and a logistics-sorting partnership with SAP. During Singles’ Day, 618 and America’s Black Friday, warehouses can overflow, and even three 7×24 shifts cannot clear the work.
He compared the partnership to a stage performance: he is like the director, responsible for putting on a show for the audience; Xingdong Jiyuan must remain ready at all times, like an actor. The two sides will connect interfaces, APIs and SDKs, refine the solution and product, and ensure SAP can immediately put Xingdong’s products in front of partners.

8. Convincing Factories: POCs Plus a Data Flywheel, with Teleoperation in the Philippines

William’s methodology comes down to two terms. First, POC: sign an NDA, have the customer send videos, quickly validate in a demo room whether the task can be done, and deliver an MVP. Second, the data flywheel: an industrial operator wants to see returns immediately after investing, while conventional model training can take months or even a year.
His cross-border model is for European customers to provide the budget, China to refine the technology, and the Philippines to handle data collection and teleoperation. One person costs RMB1,500 a month; three people working three shifts around the clock cost RMB4,500, versus perhaps €1,000-2,000 for one person in Europe. Customers get immediate output, while the videos can continue feeding model training.
Peter’s answer was more basic: value creation begins with a deep understanding of real needs. Teams must “go deep into the real site, real motions and real tasks,” then find areas where older technology could not meet the requirement but new technology can cross the threshold. “There is a huge amount of work and deep thinking” embedded in that process.

9. 南立新的 White-Paper Finding: The Industry Is Still in a Fragmented Phase

创业邦 and SAP interviewed more than 50 embodied-intelligence startups. The background figures: more than RMB52B in disclosed funding last year, over 500 financing events, and many undisclosed rounds. The core finding was a “fragmentation”: logistics, inspection and other enterprise applications have begun to see some batch deployments and real use, but “most are still in the POC and small-batch validation stage.”
Her framework is that entering the commercial world requires not only physical intelligence but also business intelligence. Robotics companies are responsible for breaking through on physical intelligence; business intelligence is where SAP can participate in co-creating data. The entire industrial chain still needs to be connected.

10. Rapid-Fire Q&A I: Is RMB1B Enough? The Transition Camp Stresses the Funding Burden

Ryan believes “the real embodied-intelligence race is only just beginning.” Data, scenarios and talent all require heavy investment, and “no amount of money is enough.” Robin added a number: from the perspective of humanoid simulation and synthetic-data training, he had heard that one training run costs $20M, so RMB1B can disappear quickly.
Peter and William raised their hands for “yes,” but shifted the focus. Peter said “time is the most valuable resource in this field”: he discussed a solution in the morning, manually measured parts on the production line that afternoon, and personally performed sorting motions at a logistics company the following afternoon. “There’s no time to carry the robot over and try it—I have to do the worker’s motion myself.” William acknowledged that a large operation also has large expenses; RMB1B is a lot, but building a good product requires more resources.
南立新 believes funding will become scarce. Current POCs do not fully exploit embodied robots’ 3D characteristics; many tasks can still be done by conventional 2D robots. Large amounts of capital will be needed when the technology reaches real-world deployment, especially mass production.

11. Rapid-Fire Q&A II: Overhyped? The Host Says Everyone but Unitree Raised a Hand

Peter did not directly assess any particular company. He said the people he sees in work and daily life are all working hard, and he believes many results unknown to the public are still being incubated.
Robin gave two reasons for skepticism. First, “even humanoids today have not really achieved embodiment”; they are still mostly hardware and not fundamentally different from industrial robots. Second, technology is changing so rapidly that the situation could look very different in the future. “Whether they deserve the name is a question mark,” and the sector will certainly be reshuffled.
William added that the industry is still in a hot-market phase and may be in a state of “mud and sand flowing together.” Some degree of shakeout will come later.
南立新 warned that the battle is far from over: “Today’s leaders may not be tomorrow’s leaders. The truly brutal competition may only just be beginning, because too much money has come in.” Her punchline was that the computer industry is being redefined: “It is still too early to say which companies are mud and which are sand.”

12. 2026 Outlook: The 1% Home Bet, Two Paths and the AlphaGo Moment

Robin reiterated that industry will come first: a robot can handle a scenario by adding dimensions such as force control on top of motion. He also raised a bet with colleagues: can humanoid penetration in households exceed 1% within 10 years?
Ryan offered the demand-side logic. After visiting dozens of leading manufacturers, he sees an aging population combined with young workers’ preference for flexible jobs such as ride-hailing and food delivery, meaning “the labor shortage will keep growing.” On the supply side, there are two paths: conventional industrial-robot workstations augmented by AI and agents can address existing workloads, while tasks such as moving items through the “last 10 meters” are “the vast ocean of stars left for humanoid robots to conquer.”
南立新 delivered three judgments: there will definitely be no funding shortage in 2026, while “the real shortage may come in 2027 and 2028”; commercialization and commercial value will not arrive in 2026, and the industry is “waiting for an AlphaGo moment”; and most robots this year are still on fields and stages, while other companies may still be in the laboratory. William said he had bought SAP shares: “The embodied-AI industry has only just entered the Age of Exploration. Who the Four Emperors of the future will be is still undetermined.”

13. The Closed-Door Opening: The Core Battle Is the Commercial Loop and Self-Generated Cash

Ryan stated the criterion for a shakeout plainly: “Whether a technology can develop over the long term comes down to whether it can form a commercial loop—whether the business can generate its own cash.” Topstar’s approach is to focus resources and talent “like a laser,” connect one scenario end to end, and then replicate it. Once the first battle is won, morale, confidence and the ability to secure resources will follow.
The numbers provide the backdrop. Before 2021, revenue grew at a compound rate above 40%. The company later cut smart energy, environmental management and other businesses, narrowing its core product and business cluster to embodied-intelligence industrial robots, injection-molding machines and peripheral equipment, and CNC machine tools. Several million people in China work across the upstream and downstream chain surrounding injection-molding products. Last year, Topstar delivered a workstation incorporating an AI agent.

14. Unitree’s Industrial Map: Turn with Manufacturing as Manufacturing Turns

Peter decomposed the word “handling”: high-cycle-rate unloading from multi-cavity injection molding; oil contamination and sampling inspection for machined parts; random-order palletizing and depalletizing; loading and unloading trucks; packing and unpacking; and tray loading. “They have different names because the underlying logic is different.” Many of these scenarios also require high precision.
His scenario-selection logic is macro. Manufacturing is adapting to uncertainty: an automaker’s final-assembly line may become “one final-assembly island after another,” located closer to end markets; large die-casting and composite-material technologies are still evolving; and black-goods products in this “traditional industry” now have product cycles of only 6 weeks. The more important use of new technology is to change the operating model: less inventory, faster response and faster product iteration.
The key positioning is that Unitree is a product company. It will spend its main effort understanding what solutions customers need and how to update and improve the product, rather than putting all its energy into heavy delivery work. It aims to improve reliability, stability and adaptation speed across electromechanical systems, AI and chips, using manufacturing scale to support margins and accelerate internal iteration. Unitree delivered more than 6,000 humanoid robots last year, largely centered on one model, and hopes to reach 20,000-30,000 this year.

15. Xingdong’s Screening Process: 300 POCs in Hand, with High Value in “Let the Robot Take the Risk First”

William’s triage framework starts with roughly 300 POCs. He evaluates them on economics, technical feasibility, company strategy and brand value. Logistics is the first priority, industrial applications second, followed by inspection, perception, navigation and dynamic path planning. In sorting, a good application is one where the main portion stays unchanged and only a small portion needs to change. He also acknowledged that the market is still large: “It is nowhere near zero-sum.”
He identified a class of scenarios suited to the current stage: airport bomb disposal. Training a bomb-disposal expert takes more than 10 years, and a human may still die after a bomb explodes. No model may even be necessary; direct teleoperation can dismantle the bomb, and “if the robot explodes, buy another one.” Other examples include pneumoconiosis, fiberglass, spray-painting and toxic or hazardous-material handling—tasks that do not require extreme precision, where “someone just needs to get the job done.” The conclusion: having robots perform work humans do not want to do or that endangers them, taking on the risk first, may create more value than replacing humans first.
He described the latecomer advantage this way: rebuilding Beijing’s CBD is less attractive than “flattening a piece of rural land and building a new city.” The New York subway has existed for 100 years, but its signal management and planning have failed to keep up with usage, leaving the system overloaded. First movers carry technical debt; in this new, model- and dexterous-hand-dependent industry, “we may only now qualify as the old hands.”

16. ABB’s Entry Logic: CIMC’s Steel-Plate Tolerance Shows That One Narrow Scenario Can Be Enough

Robin would not judge the choice: “Whether to pursue social value first or commercial benefit first is a company’s decision.” ABB’s principle is to choose deployment opportunities in industries it already serves. “If you have never served an industry and do not even have the know-how, how are you going to serve it?”
He used CIMC’s container-welding case. Containers differ from automobiles: automotive inputs generally meet six-sigma standards, with high CPK requirements and tight tolerances. The two steel plates entering a container line, by contrast, may have a large gap between them, and that problem has not yet been solved. “This is a very narrow scenario. It does not require very complex technology; just solve this small gap.”
Asked whether every niche was already crowded, Robin poured cold water on the premise. Some areas have not just several players but “possibly hundreds or thousands,” while machine vision has been driven down to a few thousand yuan. His building analogy: first make sure the foundational modules on the first floor are complete; deepen capabilities in a few areas, then build the second floor and eventually reach the tip of the tower.

17. Cash Burn Will Rise Exponentially: Screwdriving Corner Cases and Investors’ Expectations

Robin used screwdriving to illustrate the cost curve. A robot can already drive screws in a specific setting, but universal screwdriving requires establishing an initial methodology, collecting data from many companies, testing it and retraining the model. Teams must force out the corner cases, and “that cost will definitely rise exponentially.”
William gave two reasons the industry may run short of cash. It has not reached its GPT moment, so continuously updating data requires substantial funding. And investors buy expectations: “It’s like buying a stock. If you fail to meet expectations, the share price naturally falls.” If technology misses a key milestone by one step, continued fundraising may stop.

18. Disorder and Data: Bringing the Garlic-Speculation Playbook to Robot-Speculation

Robin said bluntly that “industrial robotics already has a lot of disorder; you can see it very clearly in the financial statements.” The narrative is to seize territory first and develop later, “using the internet playbook.” William’s historical view is that “there is nothing new under the sun”—the industry is simply importing the experience of speculating in property and garlic into robot speculation: find investors, build a prototype, demonstrate it in several places, and perhaps ignore whether there are real orders. He nevertheless supports data collection and data factories, saying the industry genuinely lacks high-quality data. Peter said he has no time to think about any of this; the priority is simply to make the product good enough to survive.
Robin’s technical breakdown of data reuse was that cross-company data “seems unable to be used directly.” He highlighted the difficulty of annotation and measurement: autonomous driving is roughly at the meter level, humanoids may need centimeter-level precision, and industrial robots may require millimeter-level precision. At that resolution, vision alone may not be enough to judge whether a motion is good or bad.
William answered by layer. Visual data and environmental data can be reused. Human teleoperation trajectories may not be: Xingdong’s L7 is 1.7 meters tall, and teleoperation data collected on a 1.7-meter robot may not transfer directly to a 1.8-meter or 1.2-meter robot. Bottom-layer motion control and grasping-force data may also fail to transfer because algorithms and force-control systems differ.
Peter discussed standards. China’s Ministry of Industry and Information Technology is pushing standards development, with Unitree participating, while the company is closely tracking IEC, IEEE and other international organizations because the outcome directly affects product certification. Nvidia’s “universal data” and “unified action space” remain primarily at the research and paper stage. His view is that if “a set of particularly convincing industry applications” emerges, its data format could become the de facto industry standard. The process can sometimes be shorter and faster than he expects.

19. The Shakeout Math: The Top 10 Take 40%, and the Winter Below the Middle Tier Is Visible

William walked back his use of “winter” and called it a shakeout instead. The top 10 companies account for 40% of all industry funding; the remaining roughly 200 companies may average only tens of millions of yuan to RMB100M each. “The average has no meaning for any individual company.” Companies that cannot raise money may find it very difficult to survive. Among the leaders, he believes 6-7 have already surpassed RMB10B, including Xingdong Jiyuan; those companies should survive. Below the middle tier, down to ankle level, the winter may be visible to the naked eye.
William also urged small companies to ask themselves: “Am I moving in the right direction? There are still so many unmet needs in society—have I truly met even one of them?” From the perspective of industry evolution, Robin said robot hardware is already fiercely competitive. “If everyone keeps competing this way with the old path dependence, no one has a way out.” Embodied AI offers “another level of competition.”

20. Why Industry Comes Before the Home: A Factory Is a Designed Mini-Society

The host challenged the premise: American home-robot demos go from cooking to folding blankets and “seem to work,” so why insist industry will come first? Peter gave a two-sided answer. Appliance companies must break down household needs across the world, even asking whether 200 labels are enough or whether they need 300. Factory environments, by contrast, are designed and tasks are controllable. More importantly, “a factory is also a small society”—people need to encounter new technology first in a factory governed by safety rules. “It is not that households must come very, very late. People ultimately have to accept new things in some environment, and that environment may be the factory.”
Robin added personal experience. More than 10 years ago, he spent RMB15,000 bringing a fifth-generation iRobot back from the US; after “one week,” it was put away. The distance from a demo to comfortable everyday use is still long. Factories are highly organized systems in which tasks have already been decomposed, unlike the randomness of a home.
Ryan believes the hardware gap has narrowed relatively quickly. “The core issue is soft capability and generalization.” Homes contain enormous numbers of scenarios, while each factory process is relatively standardized and enclosed. “Complete the small closed loop first; you do not need to take such a large step.” Industry has its own hurdle, however: the factory manager immediately asks about cycle time and stability.

21. The Progress Bar: Industrial Robots at 30%-40%, Embodied AI in the Nokia Era

Robin estimates that China has roughly 500 robots per 10,000 workers, placing it in the world’s top 3, while South Korea leads at about 1,000. After 50 years of development, industrial robotics is “probably at 30%-40%” on the progress bar; it has passed the rapid-growth stage and entered single-digit growth. Humanoids are just getting started, but the existing industrial-robot base could make iteration faster.
Peter said industrial customers care most about reliability and algorithmic convergence. “Being extremely accurate most of the time and then suddenly flying off one day is something industry hates.” Unitree has been sensitive to continuous production since it began inspection work in 2017. Firefighting applications such as water-cannon robot dogs also show society that the technology can genuinely save lives.
William also put progress at roughly 30%, comparing the industry to “the era of Nokia, Motorola and the Chinese brands before the iPhone arrived.” Each company’s data and interfaces are different, and the sector is waiting for an embodied-AI iPhone or GPT moment. He also listed the newcomers’ catch-up requirements: CE, UL, PSE and CCC certifications—there is still no genuinely unified certification standard—cross-border data transfer, warranty and repair, financing and lending solutions, and dealer networks. Traditional industries handled these long ago; emerging companies still need to catch up.

22. Tennis Videos, Multi-Decade Adoption Cycles and the Filipina Domestic Helper as Competitor

On 银河通用’s tennis video, Ryan said it “probably was not” teleoperated because the reaction speed was extremely fast, and teleoperation may actually be harder than real-time movement. The court’s grid, dimensions and height are standardized, making it fundamentally similar to an industrial setting. He used the opportunity to identify the real blue ocean: industries with many product types, small batches and high changeover and commissioning costs still have very low robot penetration. That is the terrain for AI plus automation and AI plus humanoids. Peter said he had read at least 2 papers using Unitree robots in similar demos: first train specialized models for visual tracking and trajectory prediction, connect them into a closed loop, and only then consider fusion—following autonomous driving’s path from modular systems to end-to-end models.
William put the technology on a historical timeline. The first computer appeared around 1945 and did not begin widespread adoption until around 2000. The first mobile phone appeared around 1973 and reached mass adoption around 2010. Industrial robots appeared in 1974 and became widely adopted around the 1990s and 2000s. Humanoids bring Honda Asimo to mind, but he acknowledged that the timing of the next mass-market breakout remains unknown.
On the 1% household-penetration bet, William did the math. Generalizing across homes means collecting data for every case from “a bath towel thrown onto the toilet, the bed or the luggage rack,” and under current modeling methods the cost could be “more than you can believe.” “The biggest competitor may be a Filipina domestic helper earning HK$4,000 a month.” Robots do have scale advantages: once the model is trained, “10,000 robots can all perform the same action,” allowing it to scale up. The program also suggested abstracting household use cases first into eldercare or hospitals, where needs may be more standardized. If a robot can administer medicine, provide rehabilitation or work as a caregiver, its use value becomes clearer.
The endgame is an ecosystem. Peter imagines deploying an OpenClaw-style agent on robots that can call different vendors’ VLA models and even world models. William asked whether an agent OS can emerge and remained optimistic that an agent-like form could appear in the VLA field. The host’s closing analogy was that smartphone adoption required not only an iPhone moment but also Android’s open ecosystem: “Whether an embodied OS can emerge—if it can, the whole industry will accelerate sharply.” Peter answered the 1% bet with one line: “I will definitely work hard to help him win—we still have 10 years to work on it.”

165: GEAR 高深远: World Models, Self-Evolving Loops, DreamDojo Freda's Investment Notes Ep. 2: Tokenmaxxing and Human Connection