第016期——Unitree 的演进对机器人行业意味着什么（机器人）| Jordan Nanos、Reyk Knuhtsen、Niko Ciminelli

Hey everyone, this is Jordan. Welcome back to Semi-Analysis Weekly. We're jumping right into it with Niko and Reyk this week to talk everything Unitree with their upcoming IPO, humanoid robots, and a few of the previous robotics articles that we've done on levels of autonomy and the quadruped state of the market. Hope you enjoy.

In terms of how many deployments are actually real in the real world, it seems like robotics is just getting started, on the cusp of taking over entire manufacturing plants, or—

Yeah, not even, but yeah. Even in their industrial deployments, I think we gave them a generous framing for the fact that their improvements have been very material. But with the burnout rates, the payload, and the internal accuracy of their hardware, we’re in the baby days, to put it lightly.

1. Robotics Is Still In Its Infancy

I’m actually not entirely certain, even among the partners. We’re trying to get metrics on what this would look like. I think they’re defining industrial deployment very broadly. Most of these robots are showing people around places and whatnot. At best, we’re talking about the frontier of deployment being teleoperation to pick up boxes and stuff of this nature.

I don’t think that’s true for general robotics as a whole, but the research playbook for them has been huge. So, industrial deployment is true for robotics, maybe as a whole. The pace of progress at Unitree is on a bit of a different axis.

I think the general point we’re trying to make is that it’s hitting a more capable set of hardware form factors really fast. But, yeah, I agree. There’s Unitree, there’s the research market, which was something nobody took seriously, and then there’s robotics as a whole, which is starting to get AI capabilities. It’s very early days, but we are seeing it.

Then there are form factors that I would see as industrially useful, but those are more sophisticated and more expensive.

Yeah.

Yeah.

2. Unitree's Pricing Power

To be fair, revenue’s healthy in the sense that they were charging a lot for these robots before, right? Quite a bit. Even in our BOM, it’s like, at a $27,000 pre-tax price, these are still 67% gross margins, which is absurd.

It’s mainly because there aren’t that many players that are going to bring down the price yet. It’s kind of just Unitree, and they’re gouging the living hell out of the market while they can. Presumably, these prices just keep dropping.

That’s a big reason for the revenue being so large. You could charge $54,000 for these robots that would burn out in 5 minutes, and now it’s $30,000. So it’s huge drops.

They’re also able to produce at scale. They’re able to produce the robots at a unique scale. Their iteration cycles and their improvements on the engineering are super fast, such that they’re trying to burn the bridge behind them a little bit.

The American bots might be trying to go all the way from 0 to 100: perfect dexterity in the hand, strong payload, make all the new innovations. Their bet is, “I’m going to make an extraordinarily cheap robot.”

By the way, we’re also all trying to figure out the software for how to solve the control problem and how to do interesting AI models to make robots useful. It turns out that, for experimentation, you need hardware that’s cheap and useful, because it’s going to break. You need good serviceability. These things need to have experimentability.

They’ve increased their capabilities on the hardware side as they’ve grown as a company. So it’s a bottoms-up approach to growth in a way that is very amenable to DJI’s capabilities, et cetera, even though it’s not the greatest or strongest drone company.

3. China's Scale Advantage

I understand the comparison to BYD and DJI because of the “burn the bridge behind you” sort of thing that’s described, which seems like a reasonable competitive tactic if you want to take the whole market and you have the ability to do so.

But in both of those cases, it seemed like there was a healthy market for drones and cars before those entrants came in. Maybe EVs or autonomy was a different part of it, but there were a bunch of players.

Robotics, specifically humanoid robots, is different. It’s not like there’s an incumbent that they’re trying to compete against. They’re pretty much defining the market and growing with it.

That makes me think about solar panels and how China has 9 of the top 10 players in solar panels, I believe. Do you believe that they will have competition in China, and that they’re just pulling up the ladder against the US, basically?

I think the article’s core purpose, in all seriousness, is a few things. One of the main motivators here was trying to communicate the level of importance of economies of scale in this market.

There are obviously a ton of AI tailwinds. This is the obvious point. People had bought cars for a while before BYD, right? EVs are a different situation, but this is an obvious thing that’s separate and different. Drones are similar in this regard.

This is definitely an AI demand pull. Language models have become extraordinarily powerful economic vehicles and tools, or whatever you want to refer to them as. Robot models, from a research perspective, have shown their early signs of life, and I don’t think anyone wants to wake up one day and say, “Oh, wow, we missed that boat.”

China’s been a huge leader in industrial automation in general. Their robots per worker are higher than in the US. This is more classical industrial automation: single pick-and-place systems for things like mobile electronics, et cetera.

Even their cobots are getting much stronger over the years and getting much cheaper. They’re still not as reliable as some of the European and US ones, but they’ve improved rapidly.

This has been a really big mandate in China: extraordinary amounts of automation. They’re well positioned because of what they’ve done in mobile electronics, consumer electronics, and automotive markets. They can really see this as a serious thing, given the markets they’ve done well in previously.

Between the fact that this is within their core advantage—the things they’ve grown heavily in, in consumer electronics and automotive—and the fact that the AI tailwind is very clear, they’ve been very forward-thinking about how to build the hardware ecosystem early on in the progress of AI.

I kind of want to add here, too. Jordan, I think you’re poking at a good one with the DJI comment. The fact that DJI sold into the drone market is important, but in the paper we really try to point this out: there was no consumer drone market. This wasn’t a thing.

You go around in the early 2010s, and it’s a bunch of dudes in their mom’s basements building drones for maybe a couple thousand dollars. Or you go and buy the $20,000 one that’s basically military-grade. There was no sector for this to begin with.

Then DJI comes in and brings out a pretty mediocre product now, but at the time it was groundbreaking because it was affordable, functional, had a camera, and was stabilized enough to be a useful drone for anybody who purchased one and wanted to use one.

While you’re talking about DJI, I’m going to throw this on screen so we can take a look at this drone.

Yeah, yeah.

And the numbers you’re referring to: $4 million in revenue—

Yeah, there we go.

$4 million in revenue—

$4 million to $130 million. Crazy, right?

It’s not to say that Unitree is booming and originating the whole humanoid market right now, because it’s so small. It’s hard to totally declare that.

If it happens, we kind of pointed it out here: this is a pretty competent company. They're creating robots that are becoming moderately useful at a reasonable price. In the past, this has been really successful for initially getting a customer base, then scaling upward, growing and growing, and getting better. Now it's like—I don't know—DJI is just everywhere.

I think it's not a super-galaxy-brain take to say that, just like with drones, people had a very strong interest in them. They were novel, extraordinarily expensive technologies that were basically licensed to governments or people who could afford them. There are a surprising number of people in the market, both within small companies and among hobbyists, who, again, to Reyk's point on the drone side, see this as a cost problem. Unitree has essentially doubled down on the fact that when they made their quadruped cheap enough, people started taking the quadruped market very seriously.

You can make the industrial argument that there are enormous amounts of progress-tracking, security, and other use cases that I think people deeply underrate, but that's not the main thing quadrupeds are selling themselves into. People want to buy a robot dog and see what they can do with a robot dog, right? Whether or not these people are extraordinarily nontechnical and are playing with Claude Code and running experiments, you see social media influencers doing it, but also reasonably competent engineers throughout Europe, the US, and China who are just trying to see what this technology really is.

If it comes down to a reasonable price where someone on a white-collar wage can save up to buy it, there's a surprising number of people who want to go over and play with this in the same way that they would have with drones. I think the deeply peculiar part about all of this—and this is actually a pretty awesome chart to bring up around this time—is that a few years ago, this was a quadruped company and people really didn't care about them.

I didn't really make this point yet, but one of the motivations of the article is that this was a quadruped company, and the economies of scale from finding a market where people really want the product have allowed them to bring the cost down, improve their hardware, make their engineering more reliable, improve quality across the board, and make a product that people love. That has given them the mandate of heaven to make the next product. Because they have the mandate of heaven to make the next product, they've been able to go up the stack of capabilities, shrink the cost again, and open up a wider and wider market every time.

As they grow as a company, they've been able to make products that are more and more capable. I tell this to Reyk pretty often as we work through a handful of the articles we've already collaborated on: economies of scale are China's scaling law. We see this time and time again, with the oversupply strategy working really well within China. We've seen it several times now with enormous businesses that started in funky, weird ways and surprised people by the fact that they just don't stop iterating and they just don't stop innovating.

Yeah. Let me try and summarize two things quickly. One is, you guys both seem very convinced that demand for humanoids is obviously going to be there when the cost drops and the quality improves.

Quality of both AI capabilities and hardware, yes.

Yeah.

Yeah.

Reyk?

This is the argument in our levels-of-autonomy paper. I'm going to harken back to this one for all the readers who remember it. The point of levels of autonomy was to show that you don't need to have the highest capability in order to be a useful robot. You can be doing very basic tasks.

This is why we had the quadrupeds paper. Unitree has good quadrupeds—great. What does that actually mean? It's just a dog walking around, so who cares? But then you can actually find some use cases with it. You can have it do scanning at construction sites, which is a very expensive job when you have to do all the captures. You can maybe have it do delivery in locations that are really constrained by size, where it's difficult to get a car in, so it's economically challenging.

You don't need my humanoid to be perfect. It really doesn't. It just needs to be able to do a few things that I want it to do, and it needs to be able to break even on a cost basis with another human doing the task. That's where we get to that whole heat-map scenario, where we show that we're not telling you this is a phenomenal robot. We make it very clear that this isn't the perfect, cream-of-the-crop robot right now.

Even with all of its challenges—even if you assume 100% teleoperation, a mean time to failure of every 20 minutes, and 5 minutes to repair—when you compare it to a human doing this task, it's actually just a little bit better. It's good enough that you can put it in a warehouse and have it do something.

We're not having it do the craziest task. We're not telling you, “I'm moving 200 things a minute out of a box, and I have to think a lot about how to sort them.” It's just, “I'm taking this box and putting this box right there.” Very basic, but that's a whole job. That's a whole job.

Yeah.

Yeah.

Okay, you're making me think of an analogy to the ChatGPT moment in 2022, maybe the Claude Code moment a couple of months ago, 6 months ago, or a year ago. It seems obvious to me that demand is there, but we're also going through the experience where there's an indication that this is going to be incredibly useful in the future, even if it's just really narrowly scoped right now.

You don't think there's anything obviously limiting the hardware from improving in reliability and the software, or general operation, from improving in quality so that it can do more economically valuable tasks more reliably over time?

I think there are a few things to this. Not to rewind, but touching on what the ChatGPT moment is, I tend to take this as a bit of a misnomer. The reason I say this is that when we got Sonnet 3.5 or 3.7, depending on how religious you are about when Claude Code came online, I don't like to talk about when code is solved, because code has become extraordinarily more capable through language models. I don't think many people who are serious would call it solved.

But it's gotten very, very useful. Robots, in order to deploy, need to reach a certain level of nines of reliability. You're not really a complement for very long—an economic complement, right? You're trying to replace an individual unit of labor. I wouldn't necessarily say you replace a person, but you're adding to capacity that would otherwise be provided by someone who would have done that task.

What does that actually look like in practice? Well, that might be legitimate teleoperation, where it's a cheaper person from another geographic region controlling the robot.

That's ideally a very high level of autonomy, potentially corrected by a person. Hopefully, that person is helping correct multiple robots at once. These things don't have to be just humanoids, right? We see a lot of 2-arm manipulators on wheelbases, and to Reyk's point, these things kind of scale up over time, and the hardware will improve over time.

The claim we're trying to broadly make is not that there's not a lot of work to go, right? People debate whether or not there are going to be tendon-based arms or whether you're going to do what Wuji and Sharp are doing, which requires enormous precision and great machining—a very, very difficult problem to solve on the hand, still today. Whether or not you need a hand depends on the use case. We have a long way to go on the software and a long way to go on the hardware.

In the same way, when code came online, we had a long way to go. We had a long way to go for autonomous research. We have a long way to go even for things that have to do with white-collar work that's not in code and is in less verifiable domains. In the same fashion, you will have some things come online for robotics that take a lot longer due to the reliability, but you get this interesting way to look at the problem where you say, "Hey, you do get enormous demand shocks that are extraordinarily powerful and give enormous amounts of capital to the ecosystem that I think people don't deeply internalize."

Can you explain the demand shock you would foresee happening?

As the price has come down and as the capabilities increase, some go-to-market opportunities will offer enormous amounts of pull into the market—capital into the market, talent into the market, and investment, too—

Yeah, but what's the—

From the customers.

I mean, for example—

Do you have anything?

For example, we highlight a few of the use cases in the autonomy paper where you're at the point now where the robot is capable on a few axes: throughput, reliability, and the failure tolerance of the task itself. In the autonomy paper, one of my favorite ones is the cooking robot, where it just uses the 2 arms to stir some onions. It's super basic. I can't really botch that, right? I'm a robot; I have a timer in my head. I can see how hot the pan is, so I'm not going to burn anything.

Now you've got a robot that can cook, right? How many line cooks are there globally? This is a huge pool that just opened up because the robot knew how to stir the onions. The big demand shock that Niko was talking about comes from something like that, where it's, "Oh, it's doable, I guess."

Well, and again, the fun part also—shout-out CloudChef. I think the fun part about this is what people kind of forget in Western markets, particularly, right? We're having enormous attrition in some industries. So when Reyk points out, why do we have to be only a little better than a person? What does that really mean?

Let's say the robot's still not at human speed. It depends on what the requirements of the task are. Is it a high-throughput task? Is it a low-throughput task? Is it a high-mix task where you're doing a bunch of different types of things, or are you doing something where the space is still reasonably constrained? There are different ways to break that down, whether it's relative to whether the failure mode is catastrophic. When you consider catastrophic failures, are you hurting someone? Are you breaking something that's expensive, et cetera?

You can break down these tasks in a few ways to determine what's ready and what's not ready. But we want to look at something called the loaded cost of labor, which is not just how much you're paying someone. If it's a minimum-wage-oriented job—which many of these robot applications are not just for minimum wage—these things tend to scale linearly with inflation and regulation or whatever, but it's relatively linear over time.

The loaded cost of labor is something that has been somewhat nonlinear in places like the US and Europe, where people are just quitting at higher and higher rates. Hiring gets more expensive. You have to invest more into it. That makes the cost of onboarding someone, skilling someone up, and finding people really difficult for the business.

So if you can get a robot that works at half the speed but get it to work 2 shifts, or maybe if you get it to 70% as fast as a person, maybe it doesn't really matter anymore because your output ends up still higher. Or maybe you don't need rework, right? Maybe the robot doesn't make as many visual mistakes. It might be slower and it might make manipulation mistakes, but maybe it doesn't make visual mistakes. That's usually what you tend to see is the case.

The unlock for an application happens through many axes and is very domain-specific. When we talk about these demand shocks, they're going to come in many forms. They're going to come through businesses that are entirely robot-oriented and were service-based businesses. They're going to come from things like cooking. They're going to come from things like kitting and insertion tasks, which are going to come online reasonably soon, depending on, again, the level of catastrophic failure.

We'll see a lot of things in logistics. Data centers are the really fun one that a few folks are going after, which I'm pretty ecstatic about, because the cost of electricians—talk about nonlinear prices, right?

Yeah.

You guys have called this out a few times. A lot of the AI fast-takeoff folks are like, "Everyone's going to turn into an electrician. We're all going to turn into electricians," which is going to be fun. That sounds like a really awesome way to spend my day, just unplugging and plugging in server racks for all of society, and we're all going to have to compete on this kind of stuff.

The thing is, the market for that is massive, right? Whether or not it applies to the neoclouds is uncertain, just because of what their scale is, how fast they're building out, et cetera. But with the massive infrastructure build-out, even on bring-up, that's sure to be an enormous value add.

People underrate maintenance and how sticky that revenue is going to be, right? Data centers are sometimes in remote places, and sometimes it's hard to find a very skilled electrician. These are serious things. The robot's advantage in a task is not always just pricing on the labor; sometimes it's an enormously high-value-add use case where the business really highly rates it, and they'll pay a lot for it.

There are a lot of tasks where the business is really going to pay a lot of money for this. That's not just the case with data centers. It's also the case with other tasks in construction and logistics. We'll see new businesses spawn up on this end where people do things in a robot-native way, not just an AI-native way. I'm quite excited for things like this.

That's not just a Unitree thing, right? You'll see other form factors and other companies.

4. The Shenzhen Supply Chain

Yeah. We've been covering a lot of ground on this so far, talking about not just the humanoids, but obviously—

Exactly.

—quadrupeds, all sorts of robotics, which I think is all related. But maybe we can go back to one of the points from earlier. If we're comparing the humanoid market specifically to previous markets that a Chinese company has entered and then dominated, whether it's drones, solar panels, electric vehicles, or anything else—consumer electronics—you say that economies of scale are a key critical component of that. I'm curious if you can talk a little bit about the Shenzhen consumer electronics ecosystem that powered some of the existing ones and how that might come into play here.

Specifically, I'm just going to show this picture on screen that I love from the article, where you can see—

Yeah. It's so cool.

—you can just visually see the supply chain.

The scale of this thing. Yeah.

Yeah.

It's amazing. Yeah. It's Huaqiangbei, or I'm totally botching that name, but yeah.

So I'm taking it for granted a little bit, maybe for the people who are audio-only, but this is a picture of a—

I think this is Huaqiangbei. This is an electronics market in, I believe, Shenzhen, where I think it's a 7-story building of just consumer electronics parts. The entire supply chain for consumer electronics is just in this tower.

It's unbelievable. You can go in there, buy your microcontrollers, any field-oriented controller you're looking for, any camera, any IMU—whatever you want. You show up with your yuan or whatever, pay, throw it down, and then walk out with every single part you need to build a drone—all within this single building. It's phenomenal, and it's just one part of the whole Guangzhou, Guangdong, and Shenzhen area—the Pearl River Delta. It's one part of it; everything is in this region. I'm totally not answering your question on the economies-of-scale thing, but yeah.

No, but let's tie it into what I thought was the coolest graphic in the Humanoids: Unitree article, which is where you guys go through the BOM of one of these humanoids, right? There's the arms, waist, head, torso, and legs, and you've got all of these individual components. Conceptually, I can zoom in on just one of these—the torso, let's say, right? There's going to be a battery system, a CPU board, and an NVIDIA Jetson Xavier NX. Maybe that one's a little bit different. Then, in the legs, there are gearboxes, motors, joint drivers, linkage bars, and bearings.

If you look at all of these components, can you talk about the supply chain in the humanoid context and how this compares to what China has done with consumer electronics? What components are shared? What components already exist? Is the whole supply chain already done? What's left?

Yeah. So I think—for me, this is a bit of an overstatement, but I think an okay portion of this was already helped out by the original consumer electronics market, right? Most of your standard controllers are going to come from the original supply chain there. It's like there's—

Injection molding for your plastics. There are going to be motors, which are pretty standardized nowadays. The point is that a lot of the base components are pretty common throughout China now.

What's kind of interesting with the Unitree and humanoid case specifically is that you're watching an ecosystem form around Unitree right now—or not Unitree specifically, but the Chinese humanoid market—where people are now making a lot of these planetary gearboxes, which is what goes in Unitree's arm to make it move correctly. People are making a lot of these now in the right specifications and sizes, which wasn't really necessary a few years ago. Drones don't really use gearboxes; they're mostly just high-speed motors. These are cropping up. It's like every province now has somebody who cuts your gears, basically, and it's like, well, nobody actually needed these gears before.

Then you look around and it's like, oh, well, I think there's that one article saying there are 200 humanoid companies in China now. All these kinds of companies keep cropping up, and the massive number of them showing up is what eventually drives all of these new suppliers to come in and be like, “Hey, listen, I got a gearbox for cheap. I'll sell it to you, that's fine. Be my customer. We'll both do this together. We're going to get on the humanoid wave right now.”

Now you have this whole supply-chain convergence onto certain architectures that work really well for humanoids, which just doesn't exist outside of China—not at meaningful scale, really. You're watching this build in real time, and Unitree is actively benefiting from it. I can really gush over the BOM, but I'll withhold for now. But yeah—

Well, maybe we can extend this a little bit to how this influences the development of the next versions of the systems, and specifically the reliability and thermal problems that people are seeing right now.

Yeah. I think one thing to harp on, too, is that it's not just the mobile side, right? Again, we've mentioned not just DJI, and not just the fact that they have a few phenomenal phone companies in the country that have grown at extraordinary rates over the last decade or 2. But it's their automotive industry, right? The contract manufacturers for a lot of these robotics companies are the same contract manufacturers—or adjacent talent from the same contract manufacturers—that allowed the automotive market to grow to the size it is today, and it's why it's still growing so massively.

To point to the electronics market, the reason it exists is because they have an extraordinarily diverse ecosystem of many, many, many small players. I'm forgetting the term, and it's a huge shame because this is just going to be misquoted. There's a friend of mine who's talked about this as a “sea of a thousand bosses,” right? There's some guy who's the best in the world at making a specific component, who's been doing it for 20 years for some number of customers. He knows so well how to get the right yield on his machines, how to improve his machines, how to understand how to make the part perfectly, and how to make it super quickly. In fact, you get it the next day.

They're all competing. He's got some absurd number of customers on a relative basis to the US, and if he can't make that part at the level of quality, precision, and reliability, with unit economics that work for him and allow him to continue lowering the price—because they're super competitive with one another on price—then he can't survive. Their competition allows consumers of the different vendors to all benefit from an extraordinarily price-competitive market. Their quality has basically come from the pressure of the internal community to survive on their own.

This comes from the fact that you have several markets that are all hardware markets, all basically benefiting from the level of technological complexity required to build things like humanoids and other form factors as well. But the fact that they have a diversity of technology to experiment across the spectrum is where they get these ecosystems from. This is not a new ecosystem for them; it's a derivative ecosystem.

I think that's the thing that people really don't get: when our production for Apple went over there, when we built up their supply chains from the US, they haven't stopped. They've made newer, stronger, and cheaper products, and their manufacturing processes have created this massive second-order effect of businesses that have serviced all these large companies as they've grown. They've become extraordinarily competitive markets where they've just become really, really great—a wealth of domain and tacit knowledge that has allowed them to thrive.

So yeah, these things matter a lot.

5. What Could Break The Thesis

Yeah, makes sense. Okay, I got one last question, then we can move to wrap up. We covered it a little bit at the beginning, but maybe you can do it in more concrete terms. The whole theme of this podcast so far has been quite positive, and there's a lot of excitement around humanoids and Unitree specifically, but there are big differences between deployment reality and hype right now.

I know we started out talking about that a little bit, but what are the—let's say there are 100+ humanoid companies and interprovincial competition in China. Can you make the bear case for Unitree, where somebody comes along and outcompetes them for the entire humanoid market, or they have less success than you're currently expecting? What would that look like? What are the challenges they still need to overcome, do you think?

Yeah, I think this is difficult, right? I'm not going to sit here and say I have a magic ball, a crystal ball, to see how this plays out in perfect form. There are many players emerging. There's a new humanoid company in China every week, it feels like, and they're building robots. It used to take several months to spin something out that looked remotely okay, and this gives a lot of credit to the supply-chain aspect of things.

You'll see a new company come out and say, “Oh, yeah, in 2 weeks we just made this robot and it's walking, and it does a triple-axis backflip,” and all these crazy things. Now, granted, the dancing isn't that big of a deal, to be honest. Talk about a robot that can—

Really? Okay, what's your favorite demo you've seen so far?

I'm boring, right? I want things that are hard for a robot to do, which is repeatable precision things that require—

So, screwing onions? What are you—

I mean, to be honest, I’m a big fan of data centers. I want to see assembly. I want to see these things put together a bike. I want to see force and torque and jerking things around. I want to see real dexterous manipulation that requires force and torque understanding, right?

Right.

Unitree is not on a clear path to having that directionally figured out just by scaling up what they’re doing today.

Yeah. What’s your favorite demo?

The Spring Gala.

Cool.

If you watch that video, that video kicks so much ass, I’ll be honest. That video rocks, right?

I’ll take your word for it.

You go and look at it, and they’re doing the parkour over the boxes. That video is incredible.

Okay.

I’m a huge fan of that one.

You do like the dancing. Okay, we found a disagreement here on the podcast finally. We placed the dancing over the dancing.

No.

What about the marathon, then?

It’s not easy for a robot. It’s not easy for me to dance, but the robots are born to dance. It’s low accuracy. They can futz around. They can land in the wrong spot. It’s still going to look cool.

I’m not trying to be a hater. It’s just better than me, so I have to let my humanity through.

Yeah.

Did you like the performance in the marathon, or was that boring too?

Yeah, definitely. That’s a really big showcase of the burnout not lasting as long anymore.

Yeah, exactly.

But that’s just cool, actually. The fact that it can go even better than me, and the fact that they can run at long distances now, is scary from a Terminator perspective, but really impressive in terms of how far we’ve gone and how long our motors last. So I’ve got to give it to the guys.

It was good. We’re too practical over here, Jordan. But I will say, I do like the kind of social use cases of these things. To be clear, I don’t want this segment to be us making Unitree into a joke, but I do want to point out that these robots are pretty funny. I’m a big fan of the robots where they have them walking around the street and saying outlandish shit all the time.

These ones—I’m a big fan of these ones. Don’t get me wrong. But they’re not exactly Tetris.

Right.

People love the bots.

Yeah, people love the bots.

Come on, you know? It’s like this 4'11" dude just walking around, saying whatever, and doing dances and stuff. This rocks. I’m happy to have this guy around. Actually, that reminds me. I think I saw one of our colleagues ask one on a date a few weeks ago, yeah?

I heard about this. I heard about this.

Shout-out to Michelle. No, no.

Yeah, Michelle. Hope that one went well. Godspeed.

Okay. Sorry.

I derailed that a little bit.

Yeah. Do I think Unitree is going to have some serious competition, though? Yeah, 100%.

Yeah.

But are they in a position to benefit? Look, yeah, it’s more the strategy, right? Unitree is an emblem of what China has been able to do, and they benefit from their ecosystem. Every bot produced right now, even in the US, benefits their ecosystem because our supply chain heavily relies on them, right? So they benefit from themselves, and they benefit from us currently.

And the customer.

So, is China going to get some competition? Yes. Is Unitree going to get some competition? Yes. Are they going to be a great business? I put my money on them. If I could go into a handshake bet, I’d say they’re going to be a good business. I wouldn’t count them out. They have a great research culture.

You guys said this at the beginning, right? This is Silicon Valley-based startups raising venture capital to develop robots for all sorts of different areas, who are then taking that capital and buying Unitree humanoids to put in a warehouse in San Francisco. That’s maybe the biggest part of their business right now: R&D for future stuff.

Yeah.

I think the warehouses, as an example, were a test case for us to show that some people were attempting to go to market with them today. I’d be hesitant to say that that’s in any form the majority of their use cases. The majority of the use cases are universities. Granted, things are changing now with new governmental regulations in the US.

Yeah.

That will curb that quite a bit. But Unitree has done a really strong job of selling to researchers who need low-cost robots that are still usable, serviceable, and strong, and are good to go and train AI models with.

I am, by the way—I do say this very seriously—I am bullish on the US being able to compete in this market. It’s going to be very hard, but the US is taking this increasingly seriously to make sure that we have some form of our own supply chains. It’s very difficult. I think we’re massively underinvested.

We don’t have metal processing. Our neodymium production is too low. All the chemicals, if we wanted to set up processing plants, are still going to come from China. We don’t make a lot of PCBs here.

We can wind our actuators, but we’re still going to have them produced mostly in China. These are really big problems. But the necessity, as AI progresses, is going to encourage us to have massive investment.

Right now, China is in a phenomenally advantageous position. This article was also attempting to be a wake-up call, saying that you don’t have to believe we are at what Optimus’s dream is, what Figure’s dream is: having a full-fledged, zero-to-100-level humanoid to show that this market is moving.

The point of this article is that this market is moving. The technology is showing signs of life in a unique way, and the supply chains are very clearly feeding into that increasing attractor state over time, where their advantage is not going away anytime soon.

Yeah.

So, to look at them like a toy company when they were just a quadruped company a few years ago, I wouldn’t ever count these guys out because they’ve already shown enormous progress.

Yeah.

And I’m sure, by the way, we’re going to see a lot more from different companies in China. But the US is going to put quite a bit of effort into this as well.

Yeah. The fact that they’re even getting into this market at all is significant, right? This wasn’t even in conversation a year ago, basically.

I do want to harp on one thing here because this is a bit more anecdotal. I think in the piece, one part of it’s not in the piece and the other part is in the piece, but it’s a bit highlighted, or it’s a bit looked over, and that’s the scaling of the Unitree project as a whole.

We mentioned that they came into quadrupeds. They went, boom: quadrupeds work, 95% cost drop. We’re shipping tens of thousands of these things now. Excellent, right? This works, number 1.

We really wanted to draw a perfect one-to-one correlation, so the paper doesn’t say this explicitly, but in my heart of hearts, the idea was that you get your actuator for the quadruped good enough, and a lot of the system is designed for the quadruped well enough. A lot of the actual mechanics—what you’re putting into the robot, all the parts—you get good enough.

Then you can transfer a lot of the actual components toward a humanoid, which is technically what the original H1 was. In the piece, we mentioned this. It’s such a cool data point; I love the guy for telling me this. It was from some guy close to Unitree. The H1 was originally designed as a quadruped standing on 2 legs, right?

So cool.

You look at it and it looks bizarre, right? The legs are already bent like the quadrupeds are. You look at it walk, and it kind of does this thing where it patters its legs a little bit. It’s designed to be a quadruped on 2 legs, basically.

And so you go from that, and then you look at the S1, where they shipped—what was it?—400 in the beginning of 2025, and then it suddenly jumps to 4,000 over the next 9 months. Then, 3 months later, in January, they go, “By the way, we’re actually at 6,500 right now,” right? You can visibly see the scaling happening in real time.

It's unbelievable what they're doing right now. It's early stages, obviously, but this is super impressive.

Yeah.

Okay, guys, we have to move to wrap here, but this has been great. Anything you think is left unsaid?

Early stages are here. I think it's no longer sci-fi, and I think it's going to get increasingly weird over the next 4 or 5 years.

First inning.

Yeah.

Play ball.

Yeah. The game began. The game began already.

Would you get a G1 in your house, Niko?

Yeah.

Or would you get an H2 in your house?

Again, I'm a little security-concerned. I do think these things are real. I do think the geopolitics is going to get very peculiar. The U.S. needs to prioritize this kind of stuff. We need to figure out our own supply chains. Unitree's a phenomenal company. How these things develop is going to be odd, to put it lightly.

Yeah, yeah.

Would you get one, Reyk?

It'd just be kind of freaky to look at it in the hallway. That's my only problem, right? I wake up to get a glass of water, and it's this 6-foot dude right outside my door.

Yeah.

Start with a roommate.

Yeah, yeah.

If you can't handle a roommate, how are you going to handle a humanoid?

I'm excited for home robots. I want a robot folding my clothes. I'm pretty excited for home robots.

Yeah. I'm excited for the clothes-folding robot. It doesn't need to walk and look like me, but I could do a clothes-folding robot. That's fine.

Put a Hawaiian T-shirt on it. It'll be fine.

Yeah, yeah. Put some sunglasses on him. Call it a day.

Yeah.

One for the floor in the apartment complex, so it just goes door to door and folds your clothes every now and then.

I'm pretty excited about a robot as an amenity. Actually, I've been waiting for this for years. I'm pretty excited for robots as an amenity. My apartment complex has a gym. Mine has a robot that does all my stuff. That's going to be pretty good.

Give the pelt on that.

I'd be pretty happy paying an extra $50, $100 to do that, 100%.

Yeah, makes sense. Okay, guys, thanks for joining the podcast today. Appreciate it. Nice job.

No worries.

Of course. Thanks, guys. Have a good one.