🔎The humanoids are coming.🤖
The latest demos are exciting and scary at the same time. Some companies are even selling their bots already. These are more toys than task performers. But progress is rapid. iPhone moment soon?
TLDR Summary
To become commercially viable, humanoids don’t just have to outperform humans. They also have to outperform specialized automation solutions. Humanoids will always be less versatile than a human and less efficient than a specialized machine. But if they get good enough, the versality/efficiency trade-off may become attractive enough to make them work at scale.
From Unitree’s impressive kung fu show to XPeng’s human like movement and Figure’s allegedly fully autonomous living room clean-up, various robotics companies are showing tremendous progress right now. Some of them are even already selling their humanoids to paying customers.
These are still more toys than actual task performers. To really succeed, these companies need to prove that they haven’t just build demos, but actual scalable computing platforms for which new applications can be built quickly and easily. Things are moving rapidly. An iPhone moment could be in the making.
In my opinion, the final structure of the industry will look similar to the smartphone industry. There will be companies providing the hardware and a generalized software platform. And there will be third parties that will develop new real-world AI applications based on these platforms.
Once humanoids scale commercially, they will cause another boost for computing demand. AI will be liberated from cyberspace to real space. Then it won’t just target white collar jobs, but blue collar jobs as well.
Real World AI Computing: The liberation of AI from cyberspace to real space.
In June 2022, close to the bottom of the post-pandemic bear market and six months before the launch of ChatGPT, I published the article below where I argued that Real World AI (RWAI) computing would emerge as the next big innovation platform.
I argued that the biggest wealth creation catalysts are major innovation events. These have two charming features: Firstly, you only have to identify them once correctly and then you can ride them for a decade or more. And secondly, you usually have a lot of time jumping on the right train, even after the innovation is unveiled and shows initial commercial success.
This was true in personal computing after 1990, in mobile computing after 2007 and in cloud computing after 2010. There are several innovation themes with the potential to impact the economy and the stock market in a similar way. One of the most promising candidates is RWAI computing.
The three innovation themes mentioned above have one thing in common: Virtually all action is happening digitally, i.e. leading companies are shipping bits and bytes around to create enormous economic value.
One obvious next frontier would be the liberation of AI from cyberspace to real space, i.e. make programs/applications move through our physical world like they are moving through the digital world. The Matrix in reverse.
RWAI Computing has the potential to unleash an enormous innovation cycle that could quickly outclass the impacts of the personal/mobile/cloud computing revolutions. Its scope will be so vast, that it will be an exercise for many parties, not just the company (or companies) developing this new computing platform.
Think about: Virtually any economic activity will over time become an application, whether it is digital or physical. Just like streaming and banking apps have disrupted cinemas and bank branches, RWAI applications could soon be available for waitressing, cashiering, cleaning or managing warehouses.
What I envisioned was a mix of AWS and the Apple App Store on steroids. There will be companies providing the hardware and a generalized software platform to operate said hardware based on which third parties can develop new RWAI applications.
Special purpose vs. general purpose automation
When people assess the potential of humanoids, they typically pit them against humans. There’s nothing wrong with that framing. To be commercially successful, humanoids will have to outperform humans on the tasks they are designed for.
But that’s only part of the truth. Humanoids don’t just have to outperform humans. They also have to outperform more specialized automation solutions.
Whether it’s done through a general purpose robot or a special purpose machine, any automation comes with fixed upfront costs. Therefore, the larger the number of tasks that can be automated with one solution, the more attractive it is to develop that solution.
The total number of tasks that a solution can automate is a function of two variables: the number of task types that it can perform and the existing repetitions of each task type.
Humans can perform a large variety of tasks due to their innate intelligence. But they get tired, bored and hungry quickly which is why there is only a limited number of repetitions they can do.
So far, machines aren’t nearly as intelligent as humans. Therefore, each machine has only been able to perform a very narrow range of tasks. Your dishwasher can’t properly clean your clothes. You need a separate washing machine for that. However, these machines don’t get tired or hungry, which is why they can perform a huge number of repetitions for the narrow range of tasks they were designed for.
In short: Humans are optimized for versatility. Machines are optimized for efficiency. Whether you hire a human or buy/build a machine depends on the nature of the task: Is versatility or efficiency more important? The more you automate a task, the more you are moving right on the versatility/efficiency frontier.
Humanoids can only be a commercial success if they move this frontier up and to the right. There will always be some trade off. A humanoid will never have the same versatility in the kitchen as a human cook and it will never be as efficient in welding as a specialized robot on the factory floor. But the versatility to efficiency trade-off must be attractive enough for a sufficient number of tasks that it’s worth the cost of developing and deploying it.
What is autonomy really?
The final form of a humanoid would not differ much from a human when you instruct it. For example, you would show this robot a piece of laundry and tell it to go through the house, collect all laundry and place it in the laundry machine. A robot with such a capability would naturally be extremely useful.
The most primitive form of a humanoid would be 100% teleoperated where it would simply copy the moves of a human operator. Such a robot might still be useful for certain tasks, for example if they are dangerous. But there would be nothing intelligent about it. In fact, a humanoid company showcasing mostly or even fully teleoperated demos without clearly labeling them as such would in my opinion cross the line to misrepresentation and fraud. I fear there are many such examples out there.
There are several levels of autonomy between those two extremes. For example, a robot might move autonomously (i.e. without a human operator), but follow a very strict preset routine that cannot be easily changed for other tasks. Instead of programming the robot to find and collect laundry, it might be programmed to go three meters in a certain direction, pick up a red shirt and throw it into the laundry bin to the left. This might be considered autonomous movement. But it would be challenging to develop this idea into an actually useful robot for household chores.
There is one key ingredient to move closer to the final form outlined above: understanding context. The humanoid must be able to contextualize audio and video signals just ChatGPT, Gemini and Claude do with text. I don’t see a reason why it eventually wouldn’t. The different amount of available training data is in my opinion the main reason chatbot technology is more mature than humanoids. Another reason is programming efficiency which has greatly increased with AI.
The sections below include a number of videos which I found very impressive when I watched them and you might be impressed by them, too. But one aspect is important to acknowledge: We don’t know the degree of autonomy of these robots or whether they are autonomous at all. If we did know for sure, the story about the company’s ability and integrity could change drastically. A healthy amount of doubt and skepticism remains as long as we don’t.
Some companies are selling their humanoids already. We should pay attention if and when customers are able to present demos that are as impressive as those presented by the manufacturers.
I have divided the company profiles into four categories: Chinese start-ups, Chinese established companies, US start-ups and US established companies.
Chinese start-ups
Unitree
Unitree was founded in 2016 in Hangzhou, China, and continues to be privately held. Their latest funding round values them at $1.7bn. They are allegedly targeting a $7bn valuation in a potential IPO.
In contrast to most competitors, Unitree’s humanoid franchise is at commercial stage already. You can go on their website and order one of their robodogs or humanoids. In 2025, they shipped more than 5,500 humanoids which is more than $100m revenue and makes them the number one humanoids manufacturer globally. This is a great vote of confidence in their technology. The company is comfortable with unvetted third parties testing it.
Based on what I can observe online from customer reviews, these are still more toys than actual task performers. A convenient programming platform seems to be missing (or at least is not advertised widely). But the pace at which this is developing is rapid.
For example, check out the Kung Fu show at the Unitree Spring Festival in the video below if you haven’t already. I’m still not sure whether I should be impressed, excited or scared when I watch it.
Compare that with the same show from just a year ago, which looks like straight out of the middle ages in comparison.
Ubtech Robotics
Ubtech was founded in 2012 in Shenzhen and is the only publicly listed start-up with a humanoid focus. They are listed in Hong Kong with a $7bn market cap. They have several humanoid models for industrial and commercial applications, which are being tested with major Chinese manufacturers such as BYD, Foxconn and Geely. The company is currently operating at a revenue run rate of more than $200m, most of which is not from selling humanoids though.
Additional players in the Chinese humanoid start-up space include Fourier Intelligence, Agibot, Engine AI and Kepler Robotics.
Established Chinese companies with humanoid programs
Xiaomi
Xiaomi is originally a manufacturer of smartphones and other consumer electronics. They have shown progress on humanoids since 2022 when they introduced CyberOne. Since then they have occasionally presented newer models, all of which remain concepts at this point.
XPeng
XPeng’s EV business is an ideal starting point to play a role in humanoids. Their model is called Iron and is at development stage for now. Iron is famous for its natural movement which made people question whether it might be a human disguised as a robot. The company even cut it open on stage to prove it’s real.
Various other Chinese technology companies such as Tencent, Baidu or BYD are researching and developing robotics technology, but have not announced a humanoid (yet).
US start-ups
Figure AI
In the US start-up space, Figure seems to be leading. They were founded in 2022 by one of the cofounders of the eVTOL company Archer Aviation. The company is run by alumni from Boston Dynamics, Tesla, Google and Apple. They are currently valued at $2.6bn with investors including Microsoft, OpenAI and NVIDIA.
Figure’s goal is to create a single general-purpose robot that can replace many specialized robots (in warehouses, logistics centers and factories) und eventually expand the addressable markets for robots greatly by making them household appliances.
Their latest model is Figure 03, a 5'8" humanoid weighing 61 kg with a 20kg payload. This bot is run by Helix which is Figure AI’s generalist humanoid Vision-Language-Action model. They uploaded a fascinating video a few days ago where a bot tidies up a living room. Management claims that it’s doing that ‘fully autonomously’, which would be very impressive.
Agility Robotics
Agility was founded in 2015 by academic researchers from Oregon State. Their humanoid is called Digit, designed to take on tasks on manufacturing floors and warehouse spaces.
It’s developed for a narrower use case than Figure which may ultimately limit the commercial potential. However, it does allow for a faster and more reliable commercial rollout. Their most important customer is Amazon which has been testing Digit since 2023 and is also the most notable investor in the company.
Apptronik
Apptronik was founded in 2016 by University of Texas researchers. Their humanoid is called Apollo, which is primarily designed for manufacturing and logistics applications. Their main customer is Mercedes-Benz which is testing the robot in its factories.
Sanctuary AI
Sanctuary AI was founded in 2018 in Vancouver. Their humanoid is called Phoenix The company sees its main competitive advantage in hand dexterity, one of the most challenging aspects of humanoid design.
1X Technologies
1X was founded in 2014. They focus on humanoids designed to work in homes. It looks pretty sleek in their ad video below. However, most of the things it’s doing in that video seems to be remote-operated.
Boston Dynamics
Boston Dynamics was the first to publish amazing demos and became a synonym for progress in robotics. Now owned by Hyundai, they are still leading from a technological perspective. Their robots move with extraordinary agility and stability and they seem to be getting closer to really commercialize their products.
US established companies with humanoid programs
Meta
Meta hasn’t revealed a humanoid robot yet. A media report from February 2025 is the only indication that they have a humanoid R&D program. Such a program might be axed anytime if it fails to show sufficient progress. However, I believe Meta is a promising dark horse that could enter the industry successfully.
Under Zuckerberg’s leadership, they have repeatedly proven their willingness to bet courageously on the future, even before it’s fashionable to do so. They launched supercomputing into the commercial age long before the launch of ChatGPT. And they have been developing augmented reality (the most promising smartphone successor) for more than 10 years.
Secondly, on top of having the willingness, they also have the ability to pour capital into moonshot projects. Developing and commercializing a general purpose humanoid will be very expensive and take a long time. Hardly anyone can match Meta’s capital expenditures.
And thirdly, developing a humanoid robot requires an understanding of the real world. Tesla claims that the knowledge gained from their autonomous driving project can be transferred to humanoid robots. But how closely related are cars and humanoids really? Aren’t AR headsets a better starting point? Helping humans navigate the real world should overlap a lot with doing the same for robots. By developing the smartphone successor, Meta gathers information and builds knowledge for another breakthrough product.
I wrote more about Meta’s humanoid potential last year in the article below.
Tesla
In spite of very limited actual tangible progress, Tesla still rules the media and investor mindshare on humanoids. But I‘m not sure they will actually play a meaningful commercial role. Elon is famous for his audacious claims which occasionally become reality. But it’s getting out of control lately.
For example, when he claims Optimus and space AI will vastly exceed Earth GDP.
Or when he says that Optimus will work as a surgeon.
“Optimus will be a incredible surgeon, for example, I imagine everyone had access to an incredible surgeon. Of course, we make sure Optimus is safe and everything. But I do think we’re headed for a world of sustainable abundance.”
Elon Musk, 3Q25 Tesla earnings call
Sure, there are surgeon robots already and their importance will only grow. But they will obviously be specialized machines given the tasks are so complex and the stakes are so high. There is no incentive for anyone to trade the efficiency for versatility in that field.
Frankly, Elon talks like a grifter who likes to jump verbally on the latest topics without actually intending on delivering. Optimus might have the same fate as the Tesla Semi, the Tesla Roadster, Tesla Dojo, Autobidder and so many other Tesla initiatives that never made it from concept to serious commercial adoption.
While Optimus primarily lives as a rendering in AI videos, others are busy actually moving this nascent industry forward, outperforming Tesla just like BYD has outclassed them in EVs.
Sincerely,
Rene
great framing on the versatility/efficiency frontier. the real signal will be when customers show demos as good as the manufacturers -- thats when we know its not just teleoperation theater.
also cant stop thinking about the compute angle: every humanoid learning to fold laundry is another GPU-hour sold. second demand curve on top of LLMs.