Loading summary
Transcript53 lines
- [00:01]
A
Robots, they're here. They're gonna change the world. Unitree currently in pole position to discuss, we have Nico Ciminelli, longtime Semi Analysis advisor, Robot Kid venture investor and Western New York native Go Bills, as well as Rake Knudsen Robotics lead at Semianalysis. Lily Ottinger, co hosting our first Semi Analysis China Talk guests who are not Dylan and Doug. Let's see if you guys can match their. No, we're just, we're Nico. Why should people care about robots now?
- [00:39]
B
Yeah, so I think, I think we're seeing in a world where everyone's very excited about their digital AIs, you have code codex, however you kind of want to look at it, and they are helping a lot at people's productivity or what call of cognitive work. And these things are still today very, very much compliments to people. Whether or not that will remain to be true is to be seen. But the thing about robots is it's a true general purpose technology. In the same way you'd look at AGI or whatever, right? And for the subsets of tasks you can deploy a robot, you are absolutely replacing something that you would have had a person there for. And not only are these jobs usually like reasonably harsh, their pay is not phenomenal, they're also extraordinarily scarce in the West. And the interesting thing about this is, you know, it's the first time in the physical world if you, when you get a general purpose robot at some level of dexterity and capability that you know, we have written about in the past is you, you, you do fully decouple in some like very profound way capital from labor. And, and so your production capacity is deeply tied to how many robots you can make. And so in a future where you want to go spin up a small business or a manufacturing company or the US has a huge prerogative towards making X, Y and Z. If you have good strong robot capacity, you're able to do things that you just weren't physically able to imagine previously or garner the resources for. This is the first real GPT we've seen in probably 100 years in the physical world. And that that's the reason it's going to be, I think, as profound of a technology as digital AI will be in the limit. A lot of the data in this world that we want to get these really magical returns from from AI is very scarce, very physical world oriented. And if we want to make the Dyson sphere, we want to make robots and build robots, we want to make all these unique technologies. You do need general Purpose robots. And that's more or less how I do this in a pragmatic sense to fix the labor shortage, to make sure that we have higher quality of life and to make sure that essentially you can do things that the west essentially has veered themselves away from. National security oriented, sure. But the ability to actually make sure you improve the quality of life as your citizens I think is a very real possibility.
- [03:01]
C
I think we should level set June,
- [03:04]
A
late June 2026 about just how far we are away from these things being actually economically productive. Because there is this like wondrous sci fi vision. But you don't get there until there is like a market for the sort of crappy version to be made that can be like just better enough at a human, at some economically useful activity that it's like worth going from making a thousand to ten thousand to a hundred thousand. So where, where would you put, where would you put these humanoids today?
- [03:45]
B
Yeah, so I, I would almost take like a, a slightly different perspective because these things one, get really, really hard and then two, it's sort of the, the wrong framing. Not in a like by question way, but in, in like. A lot of the motivation for me to come spend some time helping semianlysis cover this space was that I think we, we look at industrial automation from the two bespoke technologies or these like very sci fi technologies. And our first article we ever wrote was levels of Autonomy and it was focused on robotics and general progress robotics broadly. In many ways these things have, you know, improved at some gradual rate over time that has kind of tracked the progress in AI. There's a few things that are very different about robots that, you know, the data is obviously harder to get, the unit economics are different and the constraints of when you can deploy are different. Like how, how performant do you actually have to be before you can deploy out in the real world? But you know, today this is not a year where we're not going to see a lot of deployments. We absolutely are. It's just going to be for a very specific subset of things. And you could argue last year and the year before pick and place saw enormous improvements and we're still seeing those things kind of pick up and how quickly they are and are not deploying in different aspects of the world. So I caution people to make this binary chatgpt moment because I frankly don't believe in it. And I think actually there were certain results that happened over the last six to nine months that if you were to coin them as the ChatGPT moment, I think you could, but I think it's. It becomes uninteresting because there have been multiple inflection points to show that certain things are scalable policies and scalable approaches. And these things will result in, over the next two to three years, a large swath, of course, manipulation tasks that will actually be able to be deployed at some scalable level. So I do expect in the next two to four years, we do see gross manipulation with mobile manipulators to be available to everybody.
- [06:15]
A
I think maybe the place to start then is the analogy of DGI going from making a drone that, like, anyone on the plant, like, people actually, like, wanted to buy, and even though it wasn't a drone that, you know, could do everything that you wanted it to do, it was enough to kind of, like, get the flywheel running such that today we now have this giant commercial drone ecosystem.
- [06:42]
D
So maybe.
- [06:43]
A
Maybe tell that story.
- [06:45]
B
Yeah. And rig. This is actually where. I think Rake does a better job of describing this in some ways than I do, but this is kind of where unitree comes in a lot of. Again.
- [06:55]
A
Well, then shut up, Nico. Okay, Rake, let's go.
- [06:59]
B
Hey, go ahead.
- [07:02]
C
Thank you for the lead, Nico. Yeah, the DJI story is a bit of a. Bit of a tale to look back at when we're looking at what's going on with Yoon Tree right now. So when we look at the kind of drone market from before and the drone market to now, a lot of people will look at right now and say, oh, yeah, like, DJI owns a drone market, right? Some 70% market share. Everybody, you know who uses a drone probably uses a dji. Every filmmaker does. It kind of just seems ubiquitous now, and most people take it for granted without knowing kind of how it got there. And so when you look back at what the drone market used to be, it used to basically not exist, or the consumer drone market, at least, Right. If we look before the first Phantom from DJI, like, back in 2013, you could only buy a drone for like. Like, a competent drone for, like, $20,000. Otherwise, you would have to buy the components and build one yourself for a couple thousand dollars. Then, like, you know, you'd have to attach your own camera. It would be, like, unstable. It wouldn't really work. Your flight time was poor. The drone wouldn't be stable. It just wasn't really functional. It would crash. And it's just exp. It was a mess. And so along came dji, who just shipped a like, moderately better drone. Right. Like, it wasn't exceptional. No, like 4K camera, it still wasn't very stable. Like, the flight time was super short. It's like 10 minutes. But it was like a fraction of the price. And that was more than enough to get people to just start buying them. Right? Because it worked enough, you could use it to film something. You could use it for new things that people weren't able to do before. And that was enough to basically like jump the company from. They were at like 4 million in revenue before. And then you see like there's like this moment in 2013 where the Phantom one came out where they jumped to like 130 million in revenue. It's like, oh, okay, like that's. You're watching the creation of like the consumer drone market there, right? And so like now you've got this whole market that. No, that really didn't exist before and, you know, nobody would have considered it because it wasn't on the table. And now it's there for the taking. And DJI just kind of takes that money that they make and they continue to scale it into their drones. They continue to improve them. You know, they add better stabilizers, their flight times get longer, their cameras get better, they get easier to use, they get cheaper. As the, like, production scales, all these things, these economies of scale start compounding and, you know, next thing you know, like, we're here today where everybody knows what a DJI drone, or not everybody knows what a DJI drone is, but most people know what a DJI drone is. And everybody kind of acknowledges that this is just the drone to buy. But there's a huge kind of gap in how that came to be. That's super interesting where the creation of the drone market out of thin air, basically because DJI just made something good enough where people would buy it and fund the creation of the market.
- [10:02]
E
So the Unitree's IPO perspectives kind of indicates that they're a growing section of their customer base is now like commercial users. And it seems to me that the majority of those are like, you know, they're deploying the robot to dance in an entertainment setting. So not exactly like what you are imagining robots do in the techno optimist future, but is that going to be enough, do you think, to like set off this flywheel effect? Is this the birth of the consumer robotics market is, I guess, what I'm wondering?
- [10:36]
B
Yeah, I think it's unclear if the current form factors are that. I think the. In many ways where I was trying to describe the AI sector in this regard, and when, when Rick And I tried to look back and say, you know, how do we oftentimes get surprised by, by China's manufacturing capacity? It's like these things come in waves. It didn't happen with like a complete unlock. I mean Unitree's revenue is very analogous towards, towards unitrees in this regard. But I think yes, we will see this trajectory continue. This has been very heavily. They started in research. So I wouldn't even necessarily say that this was like the commercial. You know, everyone's buying something for their home, is buying things for fun. These things weren't really at that cost level or quality level to do something like that where you felt safe enough to do so. But the industry does have R1 which is like about as cheap as a decent computer. And I think that's the really mind boggling moment where you're like just like actually not that high quality of a robot, but if you have, you know, decent spending power as an upper middle class white collar worker. This is like kind of a funny thing. And it doesn't necessarily mean it's going to go in every house because they're like interested in being a developer or anything like this. Maybe they're not there yet. One, it's just impressive they got there. And two, small businesses may utilize them for the things that Rake has described. It becomes a bit of a meme, it becomes some small amount of usefulness and it's a novelty. Like these things are kind of things that capture the public eye in ways that we have always kind of talked about all the way back like ancient Greece. This is not a new thing. So the cost thing really does drive a lot of the.
- [12:18]
C
Yeah. And like to kind of expand on what Nico's saying, like you, you start to see like more and more people are going to buy them as well for like whether it's entertainment purposes or research purposes or the very few people like deploying them in industrial settings for tasks that are like, you know, actual job based tasks. The fact of the matter is that they are selling more, right. And they are slowly eking into actual like job like settings. And so this can keep like happening incrementally. The question is, is like, like Nico said, is this the right form factor? Have we found the right task for this? Maybe. Is the AI there? But as long as this keeps happening like this trajectory, the robot will get better, it will get cheaper and it will become more accessible for a lot of these markets to start to attempt to use, which might begin to kind of improve the bot in other ways that Might make it more useful for this task in specific or unlock other tasks or even fuel a further research flywheel that might make the kind of AI unlocks easier and make the robots more competent generally.
- [13:25]
B
One thing that was kind of the core kernel of our last paper, which it's not to just like, you know, harp on this notion of, you know, gradual change. It's just simply that Unitary was a quadruped company no more than a handful of years ago and people just don't tend to remember. Our memories are so short. So much is going on in the political economy in AI and everything. People don't remember that unitreat was just a robot dog company. And when I say just, it was thought as just a robot dog company almost pejoratively, as if that's not like a market in itself. And in the rapid pace of progress where they get to even things like the H2 which is from payload capacity a much more performant robot, they're showing that they have much deeper ambitions to go ground up and get to the cheaper consumer robot is clearly within their line of sight. It's also going to be very clear that they're going to try to compete on the robots that are going to do economically valuable work. The G1 is this a little bit in between robot that I would consider a research robot. It's used for entertainment, but we'll see much better entertainment robots and we'll see them expand. The H2 is, you know, okay, but impressive that they got there so quickly. But there's, there's just no, no reason to believe they're going to stop. Yeah, some people point out their, their strategies on entertainment a lot these days. But their, their iteration speed is, is what people need to focus more on.
- [14:53]
A
Yeah.
- [14:53]
C
And yeah, it's going to be kind of the trajectory was what. The trajectory was what we were mainly trying to highlight as well. Like we don't have to make the case that these G1s are, you know, that these are the ones you put in the factory right now, that these are the ones that are going to be it, that these are the ones that will do everything. It's like no, no, no, the, these bots didn't matter before really. Like effectively they were just research tools or toys before. Now they're. People are trying them out. Like actually they're taking them very seriously and trying to deploy them. And the kind of pace in which that happened is remarkable. And then the H2 comes more performant bot. It can be really exciting if this trajectory continues is kind of, you know, how to say it, as safe as we can. Without giving too many implications. Can you talk a little bit about
- [15:41]
A
the dimensions on which these robots have gotten and will continue to get better? I mean, like, how long they can go without needing a human to sort of fix them up. How long they can, like, hold a thing at like £10 or whatever without, you know, getting tired or overheating. What are the sorts of things besides, like, I don't know, more kung fu moves that you would need to start actually unlocking places where you can think about, you know, having this fill in for actual human being you have currently doing the job
- [16:20]
C
just to kind of, yeah. Comment on, like, the overheating stuff. Right. Like, you go ahead and look at the, you know, maybe a year or two ago at the original G1s who were being used in or which were being used in the research tasks where they could only maybe carry a box of a couple kilograms for five minutes, right. At most, and then it would kind of overheat and you'd have to let it sit in the corner for 30 minutes or like. Yeah, 30 minutes, sometimes even a full hour if you want to do the full five minutes again. Right. This is something that like, clearly they're iterated on, as we showed in our paper, as we, like, spoke with people, where you can now get around 5 minutes of work with 10 minutes of rest. You could even get up to 10 minutes of work with 10 minutes of REST or 15 minutes of REST. Varies, obviously, on the tasks that people are doing, but this is a vector that they very clearly improved on. Other ones are going to be interesting because. So the way we look at it right now is admittedly, and we. We kind of point this out in the paper too, is that a lot of what we're talking about is very specific to this task where it's like the task is basically you pick up a box and you move it from A to B. Right. We're not doing anything crazy. We're not doing, you know, any tactile manipulation or, you know, anything funky like plugging in stuff and doing anything really difficult that requires an understanding of force. It's just moving a box. So in this regard, you could see maybe battery life getting better. You could see maybe stability getting better. You could see maybe the kind of general controls getting better, such that maybe, I don't know, it functions better in the environment and it's more safe around people or the items. But for other tasks, if you wanted to move outside of that, you get into a couple of different questions depending on what tasks you're looking at. Right. And I think Nico, if you want to expand on kind of what the other task could be with more capabilities and dimensions unlocked, by all means.
- [18:21]
A
Yeah. And let's talk about hands. Like how do we get my data center electricity?
- [18:27]
B
Yeah. So I mean the framing is you have two main axes you have to improve on. Right. Your hardware has to improve. That needs to be things like your cooling, your actuators need to become more performant, your internal tracking and as Rick mentioned, things like your battery life. This is things specifically we're talking about for unitry. Right. Because all robots have a lot of different trade offs people are making in their designs. And then on the AI side part of this is just like we are not at the data scale yet that we're going to get very dexterous tasks unless you're making more custom models. Now to talk about hands and data centers hands, we have guys like Wuji Sharpa, a few people in the US like Tesla's doing some interesting things. We're making a lot of progress on hands. Now the important question is do hands matter for the data center electrician and I would argue very discreetly, no. So in the same way that there's a handful of startups, there's quite a few projects that are not startups that are becoming increasingly interested in data centers. The cost of data centers are going up super non linearly especially if we consider with this build, continue with this build out. You can make a custom end effector in a custom model that does execute on this. Well. One, two, you can teleop in when the data center is not fully brought up even in gen AI data center because gen data centers may or may not be air gapped always, but they often are and you can't always use teleoperation then. Right. So we will get when we are getting the robots as an electrician, at least for things like cabling already as early as this year. And it will be probably deployed in a much more material way in the next few years. So again we will have this very jagged frontier where even though that's a very dexterous task and actually does have a lot of catastrophic failures, there's a lot of ways where clever engineering in the same way you would have considered in the gen AI you have a lot of harnesses and a lot of safety backstops, a lot of extra code surrounding the model that you will essentially get a lot further than you think. One of the ways you can think about this is we talk a lot about it in AI tool use. Right. But there's a lot of these, like I shouldn't say a lot. There's a few very clever application layer Companies that are 3D printing tools if you will to surround specific say like your spoon or some other object you have to grasp where a two finger end effector doesn't hold it very easily. And this allows you to kind of jump the dexterity scale massively. Right. And so these are things where when we say ah, these things are going to be a bit jagged and a bit more gradual or a bit more depends is because there's a lot of ways to get a lot further than the general model offers you naturally. Or there are a lot of things where in bespoke domain custom data just is really all that matters and distribution data is all that matters. That doesn't mean that the general models are not going to come and they're not going to elicit massive value, but we're going to get robots deployed via many means over the next half a decade.
- [21:48]
A
Speak Speaking of custom data, I just got pushed an ad for someone to clean my house for free in New York City because they were like wearing a headband with a GoPro on it and apparently this is data. Like how is this economical?
- [22:06]
B
I don't know but frankly these things, these startups do exist and some of them have more interesting unit economics than others. But we're all going to be collecting data, doing a lot of stuff pretty soon. So this is, this is the beginning, not the end.
- [22:21]
C
Yeah, yeah, we're, we're, we're, you know, we're soon to be all data labelers. It'll be, it'll be a really good time when we have all the robots taking everything for us.
- [22:33]
A
So let's come back a little bit to the sort of unitree arc and maybe let's do the BYD analogy a bit. So how are they trying to sort of do the playbook of like internalizing
- [22:48]
C
more and more of their bill of
- [22:52]
A
goods and what you need to make a robot not rely as much on outside suppliers?
- [23:00]
B
The most important thing to note here I think is that BYD is a proof point on the fact that this is an underrated company by the west for a really long time and we didn't look at their quality at all. We assume the reason we point them out is because it's a lot more of an intelligible example than what, what China's also done in cobots. And so I want to point this out because this is the Real reason we use BYD as an example. China has arms that for a long time were not considered high quality. China BYD had cars that were not considered high quality for a very long time. And these things really, really tend to sneak up on you because they have extreme economies of scale. And this is where you get to kind of deeply close the loop when you're still in an R and D phase. Right now the vast majority of robots still in China, they're hand assembled, right? Like these things are not at scale. This is still a very high mix manufacturing business. And one of the reasons why unitry benefits so much from in housing their production of actuators is because these designs are not set. We talked a lot about how they get to the R1 and the H2. That decision very early on to make sure you're not going out to contract manufacturers until it's a very high order, a high volume design which they have used CMS before. Like I believe the G1 has a CM now for specific things and et cetera. It's that you don't get the internal feedback loop. You don't have the ability to see all the tiny details if you don't control every given level of the stack. They basically go from like rare earth materials to full robot at full pace with their suppliers down the street. Their iteration loop has allowed them to improve at that rate. And it tends to be that once you get to some revenue scale, the west demarcates you as something. And we mentioned the quadruped to G1. They were demarcated a quadruped company. Right now they're being demarcated as an entertainment robot company. And it's more a warning shot sign of. We've seen this before. Look at byd. We went like a decade looking at BYD and saying China makes crappy cars. Now they're extraordinarily cost competitive at higher quality than most if not many western vehicles. And like that's, that's the core point. And they do this via extreme verticalization in a very dense external ecosystem. And yeah, we can talk about their actuators if you'd like, but that's, that's just the core kernel to this.
- [25:55]
A
So, so actuators, what are they and why do they matter?
- [26:02]
C
The, the kind of simplest way to put it is it's the, the part of the robot that generates motion.
- [26:07]
B
Right.
- [26:08]
C
And this can be pretty easily overlooked as something, you know, that might be simple but you know, you have to kind of figure out the speed at which you want to Move your arm perhaps? How do you actually move your arm appropriately? How much torque do you want to put? If I hit something, should I stop? How much? How, how do I stop? Right? Do I have to reverse the whole thing at the same time? There's kind of a number of configurations of how you design an actuator, but all of it is to just basically generate the motion accordingly.
- [26:42]
A
Why is everyone freaking out that we don't make any good ones in America? So like, so what?
- [26:47]
C
Yeah, so the, the kind of other part of the actuator is obviously it generates motion, but at the same time it's the largest part of the robot's bill of materials. And so you look at, you know, we've, we've done internal bomb costing for a number of humanoids. The G1 is just the one we put out. You run into around like 50 to 70% of the bomb being just sheer actuator cost. And without these, I mean, without these, obviously you don't generate motion, but in essence your robot doesn't function right. And then if you want to go a step further, you, you know, you want to source it externally, then you're going to have to pay for all the extra costs. On top of it, it's going to be shipping costs, it's going to be people taking a margin. On top of everything, there's going to be, you know, some value based pricing going on. And your robot ultimately, say in the US where it is challenging to manufacture actuators, it will be structurally more expensive than if you had the same supply chain that maybe a unitary has or many of the Chinese robots have. And so you're kind of stuck in a very strange situation where if you want to kind of compete on cost, you're going to have to compete on the same supply chain for these actuators now. And this is extremely difficult to do. This was also what our first paper touched on a lot, was the many components of these and how they are hard to manufacture but also just not really manufactured in the U.S.
- [28:19]
B
well, we frankly don't have the capacity to spin these things up very easily either. They just don't exist. What does that mean? I think a lot of companies in the US do winding of their actuators, but all the components are coming externally. All the aluminum is coming externally. The magnets are coming externally. Magnets are rare earth oriented and I think that's sometimes not understood. Neodymium, we have some capacity, but we don't process it. The processing has to go back to China. The chemicals to do the processing are in China. There's some steps that are like multi decade build outs that serve other industries for China that they have that network effect, if you will, that if we wanted to go over and bring this back home, there's really a absurd amount of investment that would have to happen today or even before to even a quarter catch up in a decade. And, and it's your only baby. It's not another place that this happens. Malaysia does some of this processing, Australia has some of these metals, US produces some of these metals. But there's a very complex, very sparse supply chain outside China and China's really the only place that can do this currently, today, at scale in a true sense.
- [29:37]
A
So sometimes I hear the argument like oh, we've got better software, like our robots will be smarter, maybe we'll have to pay that extra 20% to build these, but they'll have an OpenAI GPT7 brain and that's something that your unitry won't. As Nico shakes his head I just
- [30:00]
B
don't know what it means. I think there's a lot of extrapolation in AI that I think is obnoxious. Um, and yes, we're in a world where progress is happening really quickly, but hardware system integration is just a different thing and hardware production is a different thing in the same way that the west said we don't actually need a manufacturing base and we'll, we'll just have good consultants and good banks and like we can just do services. I think it was an extraordinarily naive thing. If you believe in AI progress, then you actually deeply believe that you need production capacity. If you are to deflate the cost of human cognitive labor, then production capacity becomes more vital because your ability to produce chips, produce cars, produce the things that you need to continue to elicit value from your intelligence is all deeply in the manufacturing domain, all might be a strong word, but it's not very strong. You know, you can discover drugs, but if you cannot manufacture drugs, turns out you are frankly out of luck. And so there's, there's that. And two, what happens when China no longer offers their supply chain at all? Because I don't know why they would. And this is not something that is going to come for free. We're not giving them chips in some regards under certain regimes and certain policies, I don't know why they wouldn't negotiate with robots. This is an enormously powerful technology. In the limit 1 and 2 on the AI argument, I don't know why your robot needs to be Einstein Level intelligence if it's 15x cheaper. And granted this is potentially sounds like an over claim, but if you look at our margins, Unitree can go pretty low from where they are today while making quite a bit of money. And if you zoom in on that, China's not known for operating at this level margin capacity currently. So early on in a company they're very willing to go over and operate at low margins. We're not willing to operate at low margins. That's why you see people claiming my humanoid is going to be $100,000. It's like, well, is that going to be really be true and Unitree is going to come out with a robot that's going to be like 30 grand at 80% of the capacity. Well, that doesn't seem like it's going to check out. One and two, if they don't want us to make robots, we can't make robots.
- [32:25]
C
And I think this is, this is
- [32:26]
A
an interesting contrast to this sort of like, oh, okay, so I guess we'll just ban all Chinese EVs and like we'll be fine. Like America is still getting cars, right? Like there is, you don't need, you don't, you don't need BYD materials to like be able to drive around your country because you already have, you know, these hundred year ecosystems and yeah, you know, you're paying more but like, like this, you're not like sort of like leaving an entire like vector of economic productivity on the table by not being able to buy Chinese EVs. And if what you're, you guys are saying is true and this is like an entirely new like economic paradigm which is set to unlock, then you're missing out on all of that. Not just like cars that are quiet and you know, a little cheaper.
- [33:14]
B
Well, the one thing that is, is constantly not brought up is everyone's so excited about this like memory thing, right? And this is one going to be even crazier with robots. But that's like my fun side comment is there's a lot more memory going to go on these bots than there has been in the past. But this was from when you got a new capability in AI. You realize you have some absurd amount of demand, right? You have a, you have a shock to supply, you have a shock to demand. However you want to kind of look at it. You know, you can model it both ways, but that may actually not be fully true. But the point of the matter is when you have this new capability, there's extraordinary amounts of demand. The thing is in semiconductors, you at least did have a call it mature supply chain. Not an AI mature supply chain, but a very mature supply chain. The robot supply chain is extraordinarily nascent. And you say, what do you mean? We have nuptesco, we have Harmonic Drive Corp. These things are made in harmonic drives. That's not the same way you make quasi directive drives, which is the core thing we're pointing out, that have surprised everyone about unitrees. It's one of the reasons people thought they were truly just a dog company, because they were using an actuator that was considered not performant. And then their iteration cycles, their ways of making it bigger, making it more cool, adding new cooling mechanisms and just generally being better designs has allowed these robots to be more performant at lower cost. That's like, it's really the reason they've done so well and the reason this stuff gets scary is if you don't invest in that early, you won't have it later. Right? We can't really catch up that easily to some of the ecosystems China has for its electronics and etc. It's just completely economically not viable for some things. And so the belief that we're going to AI our way out of this tends to be, I just think, enormously naive in a way that just doesn't apply the way it has in the past where we have fsd. Well, it's like now China has FSD too. I don't know that they were like, oh my God, I can't catch up. Since when are Chinese researchers not good at deep learning, number one? And so I think that's a bit obnoxious. And two, we don't have the industries to do the data collection in mass the way they do. Many of our companies that are doing data collection are doing data collection in Southeast Asia. So I don't always know where these arguments actually come from. I do think the data quality currently is actually quite a bit lower in China. But that's not to say again, these things don't just improve at rapid paces over time, they will catch up.
- [35:53]
E
Do you think America has a chance? Like, do you think there are things that can be done? What are those? And Jordan has said before, as soon as China starts doing zero to one innovation at the cutting edge in some field, like suddenly the United States will get really good at process espionage the way that China has. And I love that they're selling these robots. You know, like we can disassemble them, we can figure out what makes them good and If Unitree has to spend all this money doing zero to one innovation, is that an edge for US companies that can learn to be fast followers? I don't want this to color the entire way that you answered this question like what other things can we do to fix this problem and to give us a fighting chance And I guess what happens if we don't get it together?
- [36:35]
B
Yeah, I mean it does touch on the same point to last time. It is we need to find a way to utilize our allies to make a supply chain and we need to make sure we become much better hardware engineers. I think we do have phenomenal AI research and we do are very good at deployment but this is a real thing.
- [36:57]
C
Yeah, the we've kind of. Nico and I have already worked on a paper that we hope to get out soon looking at like what it would take to kind of how, how difficult the supply chain actually is in the US like what the actual state of this is. It's quite difficult. It's. I mean there's a lot of holes right. If you look at pcb, I don't think like you know, the copper foil. You don't make copper foil in the US anymore. All your CNC machines typically becoming imported. Most of your bunch of regular industrial machinery that nobody wanted to think about before. Like now we can't really make it without importing it from somewhere or you know, just straight dependence on Chinese product at a certain point. I think rebuilding this supply chain is a much more daunting task than most people realize. I, I understand that. You know, maybe we do a tear down of the G1 and we figure out, you know exactly how to build it correctly. And it's like great guys, who wants to build it right? Like who am I, who am I going to pay to build this? Like I'm, I'm. You know most VCs probably won't look at the kind of the like CNC shop company and try and invest in it because it's not going to give them you know like SaaS returns. Right. This is a, it's a hard bet to make for them. And so you're kind of left with this weird pocket where these businesses aren't sexy by any means but they're completely necessary and nobody or not a lot of people really want to fund this or participate in it quite a lot. It's a lot of fun to shout like you know, US re industrialization. Like it's a lot of fun to talk about. It's a lot of fun to you know, really Hope for and I hope we, you know, obviously I hope it happens, but it's much more complex than just kind of a motto or a mantra. And I think, yeah, you're going to need some amount probably of industrial policy to force this to happen in some way because regular capital incentives won't be, probably won't be enough for it.
- [39:04]
B
To talk about what Ray can I really mean by the supply chain I think, I think is important. Yes. We don't make tools in the US and you'd be like, yeah, but like, doesn't Germany, doesn't Japan, like we have Okumo and DJI Mori for CNC and ES and like, you know, a bunch of other companies, let's pretend those are applicable. Which I would like pretty deeply argue are oftentimes not cost structured performance for a bunch of reasons. We can, you know, go into that. It's, we aren't even good at getting decent yield out of these things. So we don't have the tacit knowledge to run these things at scale. This stuff takes, manufacturing takes. Tacit knowledge is like a buzzword these days. But it's, it's, it's, it's very deeply true. And so, and let's go to the materials society, you know, materials processing side, right? You, let's, you know, you need rare earth magnets, neodymium, dysprosium, boron, a handful of other things. You need to be able to process them. You need the chemicals to process them. The machines that utilize this as well are also in the same way, refer to CNC machines. Very complex. We don't have a very deep knowledge about how to go over, make sure, you know, the details work. We talk about this a lot in transformers and AI and etc. The details matter, right? Some, some layer norm blows up and the infrastructure doesn't work at this scale for X, Y and Z configuration. This is true for these types of machines as well. And so the difficulty is the component supply chain for us. Everyone talks about we're funding humanoids, we're funding humanoids, we're funding robots, we're funding arms. These arms are useful. These robots aren't useful. The difficulty is we don't have an ecosystem and venture capital is funding a lot of robotics. That's true. There's enormous excitement around the space and some things can be vertically integrated. Again, you can wind your actuators, et cetera. The problem is we don't have the second order, right? We don't have the second order companies. And this is what Rick and I are kind of harping on, and this is what our next paper is going to be more focused on is when we talk, everyone talks supply chain and talks about building their own actuators. That's like not fully true oftentimes. And they do it to the best of their ability. We're doing our best. But we likely need economic zones where there's a special environmental policies or economic zones where you have very favorable tax situations or very favorable permitting policies where you can build things very quickly and you can make more experimental attempts on how to do this. And we need to be able to utilize Malaysia, Vietnam, Australia, Japan for the things that they're good at. And we need trade agreements there. And the only way we're going to do this is, is by the people that are outside of China who are also worried about their industrial capacity. That very much understand and would bet that they would benefit from the US's commercial like viability and our commercial strength. And it's kind of the only way up is through like we need to essentially just find the ways to dampen some of the pain and utilize other people's strengths. Because this is not going to be an easy thing to get the tacit knowledge. It's not going to be an easy thing to get people to buy in and really do some pretty serious investment in infrastructure that's just not present today.
- [42:18]
A
I think the most proximate Washington debate right now is around imports. Where are you guys on the. To what extent should whole unit trade robots or pieces be allowed to flow into this ecosystem?
- [42:40]
C
I'll comment on the bots really quick. It just, I think, you know, banning the imports of the unitary bots is. It's going to be challenging in a way that I don't think people realize in which the reason humanoid or like the main driver of humanoid progress in the US for the past, you know, year or two or you know, a couple of years maybe even has been because people had access to the unit trees right for the most part before you couldn't really buy these robots that were like reasonably priced, they were reasonably accessible. They weren't like entirely bespoke. They were, you know like unitrees are fairly standardized across their, you know, from G1 to G1. If it breaks you can just get another for the regular old, regular same old price. Back in 2024 you didn't really have many options. It was kind of a saving grace that unitree entered the market to create all of this humanoid research. And now it's not like there's really Been a another company to come in and also supply these humanoids. That isn't Chinese. Right. I mean you've got a few companies in the US that are making like open source bots which are nice as well, but these also typically rely on Chinese parts as well. If you look at the bill of materials, of course. But all that's to say is if you ban these bots, you kind of remove the platform of choice for the researchers and so you're going to inevitably stall the progress on humanoid research.
- [44:13]
B
Many of the American companies currently, yes. Have not been like hardware suppliers. Again, the unit economics in the US are not favorable to this. US capital markets are not favorable to this. Like they're not. That doesn't excite them oftentimes. Rick, you make, you make the point about researchers first choice. This position is absolutely definitively true. Is like we're not psyched that this is probably the case in the US But I tend to agree. The main problem that comes after this is yes, we might get humanoids of our own soon. We have obviously people like Optimus Figure, happy droid, optronic, etc. There's very wildly variable levels of quality and production capacity among each of those. Like won't get into today. Yes, the supply chains, the bill of materials are Chinese. But the reason why we should be careful on this, it's like one, no one right now, yes, to Rake's Point has the cost structure. The other thing is what happens after that. We've already seen them put an export control in neodymium, so we were forced to buy their full actuator. So if we ban their robots, we're not in a high leverage position currently to have these negotiations. So we can ban a robot. That doesn't mean we can do anything though. So I think that our policymakers, I'm not saying that this is something that we should have Chinese robots that could be sending data back and all these things and giving China a massive advantage. But I think we need to be a bit more thoughtful about what is our playbook. Are we just doing this because, you know, this is, you know, we think we've put pressure on them with chips. Well, I mean we have TSMC for that. We have enormous supply chains outside of China for that. And these things are present. You, you cut off the components supply chain. We don't have robots and they've already threatened this before. They don't send us as many CNC machines already that those exports are already dropping off. Right. They have put export controls on their earths that are required for this. And we might be able to make some neodymium and we can't process it again as I've mentioned, but we don't have any dysprosium or boron. So where's that coming from? They can cut us off in more places than we can go over and pivot, especially this early on in the game. And so that would completely cut us off from our own progress. And so I think these are complex things to navigate. I don't have clean answers here, but I think we need to be very conscientious about how to deal with these things.
- [46:48]
A
Well, instead of that dour note, I don't know. Nico, what are you most excited about for these robots to be able to do in the next. I don't know, however many years?
- [46:58]
B
I'm earnestly, wildly optimistic over, over things like them doing my laundry, like them cooking my food. I think the really wondrous part about robotics is we're going to see a lot of the deflation that were pitched by the AGI people in robots. You will compress the cost of your lawyer potentially. You can get legal advice, you can get your primary care, whatever the case may be in the universes where that's true. But if someone essentially is able to have work done in their yard or their plumbing done by the robot and all these other things, there's all these like severe taxes that like people just can't afford good services. And we're going to see a wild quality of life increase from personal robots. We're going to see people to operate at healthy margins in the US in manufacturing. If you have robot labor that helps out, there are still be humans but with robot labor. And so I think there's this like wild revitalization to the quality of life of people, but also the manufacturing capacity where we can do things greener. Going to do these things cheaper. Will you do things at higher quality? And I have a, I do have an investment that's in a company that does construction robots. And their whole goal is if you can bring down the cost of construction, you can make beautiful buildings again. And because you can have the esthetics like get reborn. And I think these, these, these second order effects of bringing the cost down of labor in the physical world will increase our capacity to do really, really amazing things. And I think that's genuinely super exciting.
- [48:38]
D
Series B deck says tireless workforce deck. Don't mention nap time. Shipped in a crate mark G1 handle with care back flip in the parking lot everybody stares Data center needs a hand on the night shift crew I can do kung fu, boys, how hard could it do? 5 minutes on, 30 minutes on already 5 minutes on, 30 minutes off don't laugh. Overheat, lights blinking, I go rest in the corner Boss just sighs says should have read the warning no, no hands, just a gripper and a whole lot of ambition 3D printed claw, they called it innovation reached around the cabinet, couldn't hold the wire you ain't the wrong one loose and now a rack is on fire tell y' all guy panicking, signal cutting out server's going dark and everybody starting to shout oh, wow, this is a disaster Call the fire crew oh, wow, this is a disaster what did we just do? Spark supplying informers crying, sprinklers coming on first day on the job and the whole day to Santa's gone. Boss surveys the wreckage ankle deep in foam should I hide? A union man should have left me home but he pulls up the spreadsheet and he sighs 30 grand and 80%, boys. Oh, we'll keep going all right
- [50:47]
C
A
- [50:47]
D
human costs a fortune and I work for almost nil burn one day to send her down I'm still cheaper on the bill so he takes a sign up Trainee, please stand clear Hands me back my gripper says you're somehow still hired here 5 minutes off, 30 minutes off rehire burn the whole thing down But I'm cheaper than the guy I retired I'm the future, baby standing in the ash do the math on the damage I'm still cheaper by the fact 5 minutes on give me one more shot version 2 won't charge the whole server locked oh, I'm the future, baby learn it as I fry 5 minutes off give it time, give it time Quarterly earnings call still says successful pilot deployment.