Summary

The humanoid-robot boom at CES has entered the “mass-production promise” phase, but 傅盛 puts the line for viable commercialization at customers paying and measurable on-site productivity gains—not units rolling off the factory floor. Of 38 humanoid-robot exhibitors, 21 were from China; Boston Dynamics says Atlas will enter production and ship to its first customers in 2026, with annual capacity reaching 30,000 units by 2028. Citing Tesla’s supply chain once floating 100,000 units the following year, then Tesla saying 6,000 units or several thousand in early 2025 before a new manager reportedly cut that figure in half, 傅盛 put it bluntly: “If you build it but nobody uses it, it doesn’t matter how good-looking it is.”
The clearer near- to medium-term opportunity is not the full humanoid, but a combination of wheeled bases, robotic arms and clearly defined tasks. 傅盛 estimates that “at least half the money in a humanoid robot is in the legs”; factories are generally accessible to wheeled equipment, and if a half-price solution can deliver more than 95%–98% of the efficiency, customers will not pay for bipedal legs. Geekplus already has a base that can move 200 kg, while Sunday uses wheels and two arms to handle wine glasses and dishwashing—both pointing to the same ROI logic: “That’s using a cannon to kill a mosquito. Why not use a mosquito swatter?”
The real commercial bottleneck in the physical world is the final 1% of reliability, not the first 99% of demo performance. Even a 99% success rate at grasping cups and plates means frequent breakage over sustained operations; robots still make real contact in 3D space, with consequences that are harder to constrain than those of 2D autonomous driving. 傅盛 relayed Musk’s judgment: “The hard part isn’t the 99%; it’s the final 1%.”
Robot demos mostly show how much optimization a team has poured into a fixed task; they do not directly prove generalization. Sharpa’s paper-folding routine contains 30 steps, so a 99% success rate at each step yields an overall success rate of roughly 70%—closer to a research challenge than a product; the paper is also laid out in a fixed position, and one robotic arm costs about $50,000. That is still far from a product constrained by cost, service life and failure rates. 傅盛 cautions that trade-show demos show vendors’ “best tuned performance”; real-world deployment requires a massive haircut.
Robot costs will not fall according to Moore’s law for chips; joints, reducers and reliability create a much harder cost floor. UFACTORY, which 傅盛 invested in and later acquired, brought robotic-arm prices down from UR’s former $10,000–$20,000 range to roughly $3,000–$5,000; it is profitable and says overseas markets account for 70%–80% of sales. But a seven-axis arm needs seven reducers, and every additional degree of freedom raises cost, power consumption and failure points. Unitree reaching roughly RMB100,000 already represents a major cost reduction, and 傅盛 expects prices to stop falling as steeply as the market imagines.
The US is the most attractive paying market for robots, but not one that can be replicated quickly. High labor costs create demand for vertical tasks such as food delivery, napkin folding, filming, lawn mowing and pool cleaning, but elevator retrofits, compliance, channel adaptation and “the last meter” can all take years; in Japan, modifying a single elevator may cost roughly RMB100,000 upfront plus RMB4,000–5,000 per month. Chinese brands’ edge has shifted from low prices toward product quality, supply chains and high-frequency iteration. The real risk is having no local channels, business networks or ecosystem.
General-purpose robots remain an important long-term technical direction, but the 3 guests separated the technology vision from the current commercialization path. 泓君 points to GPT-3 and in-context learning in 2020, arguing that someone should keep tackling the “hard bone” of generalization even without spectacular results today; 傅盛 says he would not build it now, but genuinely hopes for “the GPT moment of robotics.” Once a general model can generalize across different hardware platforms, humanoids will benefit—but a half-price wheeled-base-plus-two-arms solution could also “take off.”

Deep dive

1. Chinese exhibitors and mass-production narratives are driving CES’s robot frenzy

泓君 observed that the most significant change at CES over the past 3 years has been the rapid increase in Chinese exhibitors and companies. This year, AI ran through hardware and automobiles, while the robotics section was one of the show’s biggest magnets for foot traffic.
Her official figures: Chinese companies accounted for 22% of roughly 4,000 exhibitors, making China the second-largest source of exhibitors globally. Among humanoid-robot exhibitors, 21 of 38 were Chinese. The US-China-Korea comparison rests partly on Korea’s Boston Dynamics, now owned by Hyundai Motor Group.
泓君 said the queue for Boston Dynamics’ robot section in the West Hall stretched beyond 40 minutes. 徐皞 watched the demo but saw nothing particularly distinctive. 傅盛 also noted that attending CES for several consecutive years reveals how many supposedly new concepts were already being discussed the year before.

2. “Mass production” only matters when customers pay or productivity is real

Boston Dynamics announced that the production version of Atlas will enter production and ship to its first customers in 2026, with broader deployment and annual capacity of 30,000 units by 2028. By comparison, the program cited plans of roughly 5,000 to 10,000 units for AgiBot and Unitree.
傅盛 recalled that when Tesla first began working on humanoid robots in 2021–2022, its supply chain was reportedly preparing for 100,000 units the following year. He thought Tesla had underestimated the difficulty. By early 2025, Tesla had cited 6,000 units or several thousand; after a change in leadership, the figure was reportedly cut in half first.
His test is straightforward: a large company can force through several thousand units, but “either customers have to pay, or the robot has to genuinely deliver efficiency in real-world settings.” Otherwise, 30,000 units is just a capacity number. 泓君 also questioned Boston Dynamics’ cautious wording that customer pilot projects would begin in 2027: “Why not 2026? Why not this year?”

3. Legs consume at least half the cost while rarely changing factory efficiency

Watching Boston Dynamics’ warehouse-handling demo, 徐皞 initially saw nothing special, then added that the robot was indeed more flexible and versatile—though he was unsure whether that amounted to a fundamental difference. Factories and warehouses already use large fleets of wheeled transport equipment; a humanoid solution has to prove incremental efficiency, not merely that it can move.
傅盛 cited Geekplus: its lift-equipped base can carry 200 kg, and two arms can be mounted on top. Arms are useful for handling goods, but factories are generally already accessible to wheels; legs mainly serve a limited number of staircases. “That’s using a cannon to kill a mosquito. Why not use a mosquito swatter?”
He estimates that at least half the cost of a humanoid robot sits in its legs. If a wheeled machine priced at half the cost can deliver 95%–98% or more of the efficiency, while carrying 100 kg and incorporating its own cargo bay, buyers will not pay a premium for a human-like appearance.
Asked by 泓君 why top-tier companies are all building humanoids, 傅盛 turned the question around: “Maybe because building a humanoid robot makes you look top-tier.” He believes Musk reignited the boom. Honda worked on bipedal robots for 30 years before discontinuing them, while Boston Dynamics was once sold by Google and acquired by SoftBank.

4. Robots that can work will grow out of specialized forms

In 2017, 傅盛 built a two-armed robot for a birthday performance: engineers had it lead guests, strike a match and light a candle. The result was a one-off routine produced through the team’s secret overtime, not something that generalized; the two arms also sharply increased power consumption, weight and cost.
Sunday demonstrated a household-task setup combining a wheeled base with two arms, picking wine glasses off a table and placing them in a dishwasher. Someone familiar with Figure visited and was “deeply impressed” because the stable base freed the team from spending its effort on legs, allowing it to focus on the hands that actually do the work.
傅盛 stressed that robot comes from Robota, with the emphasis on labor rather than appearance. Sweeping, mowing, warehouse transport, and wheeled bases paired with robotic arms are the path by which products gradually expand. “All good products and technologies grow into existence”; they are not conjured fully formed as perfect general-purpose machines.

5. The final 1% turns a 99% success rate into an unusable product

傅盛 used cups and plates to illustrate the reliability threshold: even at 99% accuracy, a robot handling multiple objects every day will keep breaking things over time. Home and commercial customers will not accept a steady stream of minor accidents simply because average performance looks good.
He relayed Musk’s judgment: “The hard part isn’t the 99%; it’s the final 1%.” Real deployment means contact with physical space. Corner cases are no longer merely wrong software answers; they can cause injury or damage to equipment and property.
The fact that autonomous driving has still not truly reached Level 4 after years of work shows how difficult tail risks are. Autonomous driving mainly handles a 2D plane, while humanoids must judge, grasp, apply force and make contact in 3D space. 傅盛 sees enormous industry progress, but the commercial passing grade remains high above the ground—and nobody knows when the industry will clear it.

6. Beautiful demos prove task-specific optimization, not true generalization

泓君 asked multiple technical people whether a company’s victory in a robotics competition could establish that it had the best technology. The answer was always no: competition requirements and environments are fixed, so the result says more about how much time a team spent optimizing for that task than about its ability to handle variation.
Sharpa’s paper-folding demo contains roughly 30 steps and was treated as a long-horizon task. At a 99% success rate per step, the overall success rate would be about 70%. The robot also dealt cards for Blackjack, took photos with a camera and caught ping-pong balls. Dealing cards is a fine-motor task; catching a ball tests reaction, latency and split-second decisions. 徐皞 was not especially moved by the photography scene, while 泓君 found it somewhat funny.
傅盛 agreed that folding paper was much harder than catching a ping-pong ball, but kept the key objection: the paper had already been laid out in a fixed position, and the task had been heavily optimized. A few years ago, reproducing ALOHA’s egg-frying routine achieved only about 80% accuracy within a constrained range; pushing materially higher remains extremely difficult.
Cost changes the conclusion further: one Sharpa arm costs about $50,000, before accounting for service life and failure rates. 傅盛 compared it with the match-lighting demo once described as the “moon landing” of robotics in 2017—throwing resources at a problem can generate shock and awe, but that does not make it a replicable product.

7. Joints, reducers and service life form robotics’ hard cost floor

Because robotic arms were too expensive, 傅盛 invested in and later acquired UFACTORY, pushing down the cost of six- and seven-axis arms. A UR arm once cost roughly $10,000–$20,000; UFACTORY’s current products are in the $3,000–$5,000 range. The company says customers include Google and Stanford, that overseas markets account for 70%–80% of sales, and that it is already profitable.
Unlike chips, motors, harmonic reducers and mechanical joints can only improve incrementally. A seven-axis arm needs 7 reducers, each of which used to cost several thousand yuan or even more than RMB10,000. Even though the relevant structural patents expired long ago, the industry continues refining the same architecture because no new design has emerged at scale that balances load-bearing, emergency stops, reduction and precision.
Figure uses harmonic reducers: the upside is high load capacity, the tradeoff is slower movement. Some Chinese robots use planetary reducers and can adopt other architectures for performance scenarios that do not require heavy loads. More degrees of freedom mean more joints, higher appearance-related costs and more failure points. “Looking good” itself carries a price.
傅盛 considers Unitree reaching roughly RMB100,000 “an extremely, extremely large” cost reduction, but does not believe prices can fall indefinitely. The bigger commercial problem is that arms wear out quickly, have short usable lives and lack consistent warranty standards, while real customers expect equipment to go 1–2 years without a major failure.

8. Robotics still lacks unified software-hardware layers, dedicated chips and middleware

泓君 divided the industry into software, hardware and the development platforms connecting them. 傅盛’s answer was “none at all”: vendors are each building hardware or software, but the industry lacks unified software-hardware interfaces, development platforms and a unified foundation model that robots can use directly.
The market already has chips for autonomous-driving systems, gaming, and large-model training and inference, but no mature robotics chip. Robots must source components from other industries. Chinese mid-range chips are already highly usable, but chips are not the main cost; mechanical structures and full-system manufacturing remain the expensive parts.
泓君 summarized the situation as a chicken-and-egg problem: without industry scale, suppliers will not invest in wear resistance, quality or mass-production specifications; without mature components, system makers cannot achieve scale. She sees an element of well-intentioned “hype” in the current boom—get more people to invest first, and the supply chain may eventually be built up.

9. 2–3 years of delivery-robot tuning reveal the real scale of deployment

A defined use case is not automatically an easy one. Domestic food-delivery robots have reached end prices of roughly RMB10,000 and still do not sell well. Even if they could theoretically break even against service-worker costs, restaurants will not train staff and change existing workflows unless the value proposition is compelling enough.
傅盛’s deployment example was not flashy: at an Italian deep-diving pool, the kitchen was far from the restaurant, and 2 robots handled only the back-and-forth transport. Even that simple motion took at least 2–3 years to resolve collisions with tables, people’s feet, outstretched children and lost positioning.
To keep the full-system price under control, the team could use only a single-line lidar costing roughly RMB500. Dense tables and chairs, glass, solid-black objects or KTV fog could make the robot lose its way. The engineering fix was to place infrared-readable reflective markers invisible to the naked eye every 3–5 meters on the ceiling, requiring a dedicated tool for applying them.
These products are not solved by a single AI breakthrough. They require more downward-facing sensors, repeated hardware adjustments and a steady effort to cover corner cases. 傅盛’s summary: “The devil is in the details.” A navigation system that looks unexciting still takes 2–3 years to become deliverable.

10. The US will pay, but infrastructure and “the last meter” take years to modify

Hotel delivery runs into the elevator. In China, a repair technician can be paid roughly RMB2,000 to add a board and let the robot call the elevator over Wi-Fi. In Japan, an elevator maker may charge about RMB100,000 upfront and another RMB4,000–5,000 per month. Compliance and infrastructure costs can consume the advantage created by high overseas labor costs.
傅盛 sees the US as a milestone market, with single-action robots such as napkin folders particularly promising. More than half of the profit from the food-delivery robot business also comes from overseas. Europe is harder, Japan lacks software talent and secondary-development efficiency, and South Korea performs reasonably well but has limited capacity.
Low-cost navigation can spill over into lawn mowers and pool cleaners. Chinese teams keep the chip cost below RMB1,000, with RMB2,000 considered high-end; autonomous-driving vehicle chips cost at least RMB10,000, while lidar adds several thousand yuan. The competitive edge comes from combining fewer sensors, cheap compute and engineering algorithms.
One UFACTORY chain-store customer used a robotic arm for surround photography, and the two sides spent 2 years refining the system. In delivery, the harder problem is often not movement but the lack of a handoff process between staff and robot. This “last meter” changes over years, not days or months.

11. China’s overseas expansion has moved from low prices to model leadership, with local relationships determining durability

CES booth location, floor space and build-out convinced 泓君 that Chinese brands have clearly upgraded. Insta360 sponsored yellow bags across the venue, and many booths were staffed by foreign salespeople; only after talking with them did visitors realize they were Chinese companies. 傅盛 sees the biggest change as this: “You can’t even tell it’s a company expanding overseas.”
傅盛 described overseas expansion in 3 stages, though the program elaborated only on the first 2. The first was a people-powered approach in transparent-information sectors such as apps: while a Silicon Valley company might have 5 or 6 people building a tool, a Chinese company could deploy 100. The second exported “hardworking” cloud implementation, industry applications, R&D and operations capabilities, gradually gaining an edge in iteration and supply-chain models. He did not give the third stage a specific name, instead stressing that today’s overseas expansion requires understanding and respecting local markets, channels and new growth models.
“There are the same 54 weeks in a year. I iterate 54 times; you iterate twice.” Hard work ultimately becomes an advantage in experience, efficiency and operating models. But brands cannot be “yanked into place.” Costco and Target in the US, along with the distinct channels and user habits of Japan and Europe, require top executives to put down roots for the long term.
傅盛 viewed one product delisting as an almost existential lesson: the company was growing quickly but had failed to build relationships and an ecosystem in Silicon Valley. 徐皞 also warned that a “Tian Ji’s horse racing” strategy of winning by refusing to play by conventional rules works only once. 傅盛 therefore refuses to push sales by stuffing inventory into channels and refunds unsold stock, preferring to “move a little slower.”

12. Commercialization and generalization are separate paths; the industry still awaits robotics’ “GPT moment”

泓君 recalled seeing GPT-3 in 2020, far weaker than today’s models but the first system to show her the potential of in-context learning and generalization. That is why she supports Musk in continuing to tackle the “hard bone” of humanoid robotics. 徐皞 agreed with that conclusion and also expects robotics to reach a comparable moment of generalization.
傅盛 believes Musk initially made the problem look too easy and now acknowledges that robots are much harder than cars. As a practitioner, he “definitely would not do it” today, but he does not reject the direction over the long term.
泓君 and 徐皞 both believe the 2 paths must coexist. One tackles technical, engineering and scientific problems in pursuit of a general-purpose robot usable in any setting; the other commercializes first in vertical markets to keep teams alive. OpenAI spending $1M-plus, and then several million dollars, on model training in 2018–2019 may have looked absurd, but it also shows that long-term experimentation requires patience independent of near-term ROI.
傅盛 hopes companies like Scale AI that specialize in the brains of robots—or even a Manus for robotics—can generalize across different hardware platforms. If “the GPT moment of robotics” arrives, humanoids will take off, while wheeled robots with arms will also benefit from being half the price. Autonomous driving, as one form of robotics, is already advancing rapidly along another path.

Market Overview: January 20, 2026 The AI Opportunity that goes beyond Models