This week’s podcast is about Unitree, which is about to go public.
Plus, there is a quick intro to Tencent’s Robot Brain (Tairos).
You can listen to this podcast here, which has the slides and graphics mentioned. Also available at iTunes and Google Podcasts.
Here is the link to TechMoat Consulting.
Here is the link to our Tech Tours.
Here are my 5 take-aways:
- A good growth story in robodogs and humanoids. Rapid growth. Highly diversified. In both product categories.
- Unitree has frontier-level expertise in robot bodies and agility.
- Unitree was late to embodied AI capabilities.
- The hands don’t appear advanced. Only 10 DoF?
- Operationally, Unitree is still a relatively small company
- 516 FTEs, versus 1,500 at Agibot
Disclosure. I have had a paid consulting relationship with Tencent in the past twelve months.
———-
Related articles:
- AgiBot’s Race to Embodied AI Robots at Scale (1 of 2) (Tech Strategy)
- Why I Like AgiBot (Podcast 290)
From the Concept Library, concepts for this article are:
- Robotics
- Embodied AI: World Model
- Embodied AI: VLA Model
- Embodied AI: VLM Model
From the Company Library, companies for this article are:
- Unitree Robotics
- Tencent: Tairos RobotX
———-transcribed Below
00:05
Welcome welcome everybody. My name is Jeff Towson and this is the tech strategy podcast from Techmoat Consulting and the topic for today five takeaways from the unitree IPO filing Plus a quick intro to ten cents robot brain as a service really which is called Tyros So it’s really important actually, but anyways unitree is kind of the big story right this week really in the next day
00:32
You know, they’re going public Shanghai star market. looks like market caps in the nine to $10 billion range. Okay, I don’t talk about valuation and price. That’s not my thing. This is not investment advice in any way, shape or form, but I do talk about business and strategy and strength of a business A versus business B. That’s what I do. Okay, we finally saw some numbers for unitary in the filing. So pretty interesting reading actually. So I’m going to go into sort of what my takeaways were from that that I didn’t really have at the front of my brain before that.
01:01
So I the numbers are pretty interesting actually. It’s in Chinese, these days you can translate stuff easily. It’s worth reading. They are kind of the most high profile robot company. I actually think Agibot is sort of the one I’m paying the most attention to, but yeah, obviously super important. Anyways, I’m going to go through that and that’ll be the topic for today. Let’s see, standard disclaimer, nothing in this podcast, my writing a website is investment advice. The numbers and information from me and any guests may be incorrect. The views and opinions expressed may no longer be relevant or accurate.
01:30
Overall investing is risky. This is not investment, legal or tax advice. Do your own research. And with that, let’s get into the topic. Now the concepts for today that matter are, I’ve talked about these before, but really world models, VLA models, which VLM models as well, embodied intelligence and sort of how those models work.
01:57
which I’ve gone into some in some depth. I’m going to do a lot more this week. For those of you who get the email, I’m going to send you some pretty technical breakdown of the Embodied AI that Tencent is working on. It’s pretty, I don’t know, I’m torn. I don’t know if it’s that helpful. It’s really detailed and pretty technical. I found it super helpful, but it may be the least interesting thing I’ve ever sent out. It’s more like a technical breakdown. So I’m not sure, I’m torn on whether that’s helpful or not, but.
02:26
I think understanding how these models work, it’s super important. So anyways, that’ll be the topics for today. VLA, world models, and we’ll talk about unitree, but I’ll go through Tyros Tencent’s Smart Brain as well briefly. I’ll send that mostly out by email. Okay. So what’s sort of going on here? You know, the robot thing, it’s sort of in high gear.
02:54
Right, the robots are advancing super fast. Companies are popping up everywhere. In the last month or so, I’ve written about six of them in pretty good depth. You know, like Deep Robotics, Unitree, Agibot, Linkerbot, Booster, and sort of breaking down 70 % of my focuses on the business model. What’s the use case? What’s the market size?
03:20
who has a stronger business, because I kind of want to know who’s going to win. And then maybe 30 % is more in the technical stuff. How do the actuators work in the knee joints? And how do you get the wiring to work in these encapsulated joints and things like that? Which I’m doing a lot of studying on that. I think it’s important. But that’s 30%. Ultimately, it’s the business model that I focus on in competition. So anyways, within that, you kind of get this question of, all right, what’s really going to matter in these robots?
03:49
Is it all about innovation? You know, there was a time when refrigerators were super advanced tech. Nobody talks about the tech of refrigerators. We might look at one versus another one company for a second. We don’t talk about the tech at all. It’s just kind of standard. You know, a lot of what the robot story is going to be, it’s just going to be standard stuff soon enough. We can already see that in EVs. You know, outside of a couple key things like autonomous driving,
04:15
the manufacturing base that creates these, but a lot of the standard features, it’s kind of pretty basic. So a lot of that’s going to happen with robots. This whole idea of like, this robot can do a flip and it can dance and do kung fu, which is pretty cool this last year. That’s just going to be standard soon. You’re not going to win there. So when I kind of look at these, I think, okay, what’s really going to matter? And if you look at humans, because a lot of these are following humanoid models, what matters in humans, really?
04:45
The physical body of the human is not that big of a deal. It’s really not that much more advanced than any other animal on this planet. In fact, it’s subpar. We don’t run as fast. We’re not that strong compared to chimps or most animals really. Our bite force is nothing. mean, most of the sort of mechanical body stuff, we are substandard. But what do we have? We have three things that mean everything. We have a brain, which not as just
05:15
as an individual thing is very important, but it allows us to work collectively, which is super important. We could call that an ecosystem of intelligence, really. We have hands, which are super important, super complicated, and we have vision and perception in certain light bands. I think it’s kind of the same for robots. think a lot of the… If you move into certain use cases like industrial robots and things, the body, the chassis,
05:44
becomes more important, but a lot of them, pretty basic or they’re getting basic real quick. So if you look at those three aspects in a humanoid robot, okay, who is the best at making the brains of robots right now? It’s not Unitary. They got into embodied intelligence side fairly late, 2024, 2025. On my short list out of China, Asia, there’s ones in the US I’m not following. For the brain, I’m looking at Tencent.
06:12
And I’m probably looking at Agibot and maybe Galbot’s also on my list as, you know, on the frontier of that. If we look at hands, that’s Linkerbot, which I wrote about last week. They are, if you actually look at what Unitree talks about, Unitree is vertically integrated in that they do their own LiDAR systems. They do their own hands in-house, but their hands, you know,
06:38
As far as I can tell, have 10, 12 degrees of freedom. Well, linker bots up there are 26. They would argue 40 in some cases, the human body, let’s say 26. Elon Musk, Tesla, his optimist robot 24. So when you look at, and they have like six different types of hands, really advanced cable tendons, linkage joints, direct drive. I wrote about all this a week or so ago. So linker bots way ahead in hands. And there’s others that are good, but
07:07
That’s the one I’m studying to sort of keep my knowledge of. Okay, Brain, Tencent, maybe Agibot. I’m sure there’s others. These are just the ones I’m following right now. In terms of sight, I would probably say it’s Hikvision, which is not a company I talk much about, which is, you know, they do security cameras. They’re everywhere in the world. I’m going to write more about them. They’re on the entity list, which is an issue, but…
07:35
they have moved fairly aggressively from, we’ve got cameras everywhere with video camera to, hey, we’re building our own LLMs, really vision language models, on top of those that can see everything and understand it. So they’re really becoming experts in sort of visual perception because they have the data sets and they’ve got cameras everywhere in this world. So I would put sort of HIC Vision on the site. Now, if we’re just talking about the body, the chassis, then I think it’s probably Unitree.
08:05
I think they’re far ahead than most companies. There’s a couple exceptions, but you got to give them credit. Humanoids and robot dogs, the quadrupeds, they’re probably the best chassis actuators, flipping, you know, all of that. Okay, but I’m not sure how important that is. I’m pretty sure the other three things are important. Brain, hands, vision. Depending on the use case, the body of the chassis may or may not matter that much. We’ll see.
08:34
I might be wrong about that. Anyways, that’s kind of where I am. And we’re seeing sort of a Cambrian explosion where all the pieces are now in place, the brains, the eyesight, the hands, the reels, the whatever. So there’s all these new combinations being put together very rapidly, new types of robots for new types of UK’s cases. If you fall in the ocean, you know, there’s life preservers that fly through the air and come drop next to you drones, but kind of robots, know, deep robotics, which I’m going to go visit them in a
09:04
weeks in Hanjo, they do industrial quadrupeds that you can put into power stations and you can have go through the mountains to watch the electric lines and things like that. Industrial quadrupeds are pretty cool. So they have all these sort of new types of robots, new types of use cases, all of that. I kind of think about Agibot for that because there’s a couple companies that are playing across the board.
09:31
every market use case, every type of robot, everything from cleaning droids to robot dogs to quadrupeds to what, know, them in particular, they’re playing across the board for every type. So that’s kind of the Cambrian business model it’s called. Anyways, that’s kind how I’m thinking about it. And within that framework, let me sort of switch gears over to Unitree and really my fight takeaways for what I took out of the filing.
09:58
All right, first thing is if you’re not familiar with Unitree, they’re moving fast, right? But they’ve been around for a decade plus, but they’ve really got one, two, three, four, five, six types of robots now. The humanoids are the ones that are kind of getting a lot of video attention this year, although it was the robot dogs last year that were doing the flips and stuff. So the full size one, the H1, that’s about 180 centimeters. That’s literally my size. That was launched in 2023.
10:28
Did okay, but they were really the quadruped company back then. Then the G1 comes, which is more little more mid-sized. That’s 2024. That’s really the one where they’re showing pretty rapid growth in their units sold and their revenue. So it was, they were sort of a robot dog story a year and a half ago. Now it’s about 50-50 humanoids robots, which was a surprise. I didn’t know they were moving that many so quickly. But really you look at the G1, that’s the one.
10:55
They also have a smaller one called the R1. This is 2025. These are more lightweight models. mean, they’re kind of, historically the way they’ve sort of positioned themselves is we do the training on SIM and then we download it into the robot and then you control it with a controller and you trigger the programs. And they’re sort of a general robot platform that others can then build their programs on. And the R1, the smaller lightweight one,
11:24
That’s really a robot you would use to try your new algorithms and train and do research. You’d buy the smaller one. And a lot of companies have a smaller version for this approach. Okay. Within humanoids, you also have the H2, which is larger. This is October, 2025 released. Full size, 31 degrees of freedom for the whole robot. You hear these numbers a lot, 31 degrees of freedom for the whole I don’t really know what it means at the whole robot level. I know what it means for hands.
11:54
But this is the one that’s kind of most positioned, as far as I can tell, for embodied intelligence. Instead of this program, it seemed to re- know, and then download it and trigger it. No, this is getting more into embodied intelligence on the robot. Now they do have a new one, which they, you probably saw the videos. It’s that big, huge robot that you stand, you get inside like Sigourney Reever in that movie Aliens, where at the end she gets into that, whatever you call it, crate loader.
12:23
and she fights the alien. It’s like that. You can see pictures of it. It’s the GD01. It’s the big one. Not really sure what that’s used for, but it’s pretty cool. That’s our humanoids, one, two, three, five humanoids, Quadrupeds, kind of a lot here. Three different series, the Go series, the A series, the B series. Basically small, regular size, and then industrial, which can carry basically more loads. And you…
12:52
You don’t really use these for loads. You can put gripper arms or cameras or sensors on the back. This is the one when I was out there a couple of times last year. You can stand on these and it will stand up with you on its back like it’s nothing. You can carry kind of a lot. The little ones, the other ones which are much faster are smaller. And then they have their robot arms, the Z1, which you can mount on the back of these robot dogs. And that, okay, that’s kind of their…
13:19
portfolio. There might be some other things in there, but that was in the filing. So there might be a couple of things there that I didn’t know. Five takeaways. Number one, the growth was more than I thought. And it was both in robot dogs and in the humanoids. Okay. I didn’t know they were growing that much and I didn’t know it had diversified so quickly since I was there twice last year and it was mostly about the robot dog still. So
13:48
If you look at sort of their revenue in the filing, humanoids are about 51 % of revenue, about 850 rem and B. So that’s a hundred plus million dollars. This would be annual 2025 revenue. Quadruped was about 40, 45 % in that range by revenue. So 700 rem and B. And then component parts, which is a lot of robot arms, about 6%. And that gets you up to…
14:17
about 1.67 billion renminbi. Divide that by seven, that gets you US dollars. So 200 to 300 million USD. so compelling. I thought that was good. And if you actually look at the rate of growth, just talking about revenue, pretty fast. So 2025, 1.7 billion renminbi revenue. 2024, 392. Yeah.
14:46
Okay, they’re moving. Like that’s impressive. And you go one year before that, 2023 is about half that. So 159. Okay. It’s, it’s now anytime you see big growth numbers like that right before IPO, like 23 to 24 was they about doubled their revenue. 24 to 25. It goes up by a factor of four. Okay. Anytime I see a big revenue growth one year before an IPO, get, Oh, okay. You’re, you’re going for the big valuation. get it.
15:15
Is it sustainable? We’ll see. And the story, as far as I can tell for the growth, had a lot to do with the G1, the humanoid, which took off 2024. That’s your mid-sized humanoid. I actually looked at the price. I think it’s about 160,000 renminbi. So $20,000 to $30,000 in that range, right? Break down the sales by geography, 44 % in China.
15:43
Overseas sales, 55%. USA was 13 % in 2025, which is interesting because of the recent robot ban coming down the pike in the US. But generally speaking, that’s a pretty compelling growth story. Diversified a lot across a lot of products. Geographically diversified, the revenue’s real. This is not, hey, we’ve got a lot of growth online in users, but no revenue. This is revenue-based, it’s robots.
16:13
So I thought that was decently compelling. Within that, all these valuations they’re getting are based on the sort of what is the actual market for these things. And I don’t think it’s as obvious as everyone says. You kind of get this story that, well, humanoids are just going to replace humans, so we make them look, you know, all those jobs. That’s your TAM. I don’t buy most. I don’t buy like 75 % of the use cases I hear for robots.
16:44
Okay, what are these robots dogs going to do? They’re going to do inspections of facilities. I think deep robotics has a compelling use case. Industrial robots that do inspections of tunnels, power lines, big utilities out in the cold and wind and rain. Okay, compelling. Is that the only use case for robot dogs?
17:06
And I’ve been asking this question for a year. What’s the use case? Now, I assume military and police and security are part of that, but no one ever seems to say that part out loud. Okay. What about, you know, humanoids? Okay. Retail. You’ll hear this a lot. Retail is going to be a robot thing where you put a robot in a 24 seven convenience store and you can keep it on all the time and it just sleeps there and comes alive.
17:36
Is that really a big upgrade? mean, most convenience stores are really efficient. And if this is a big upgrade, it of starts to sound to me like a vending machine. Is a vending machine really a major disruption? What’s the difference between a robot-run convenience store or one of these robot cubes and a vending machine? I don’t know. Anyways, the use cases I like are
18:05
Basically factory, assembly work, logistics, warehouses. And there are some consumer use cases I like, I think. Within that retail story, there’s some cool ones. Cleaning robots are pretty good. The inspection robot dogs are pretty good. Drones, which do aerial inspection, are quite compelling. So within those use cases I buy, but a lot of them I’m like, okay, I don’t quite see it, but you know, I’m open to it.
18:34
Anyways, within that growth story, interesting. All right, so that’s kind of takeaway number one. Look, the growth story looks good. Takeaway number two.
18:45
Unitree seems to be singularly impressive in the expertise of robot bodies and their agility. They seem to be at the next level from almost everybody. If you actually go through the filing, which I would recommend, the most interesting part for me was looking into their sort technical innovation and what they’re focusing their R &D on.
19:14
And most of that has to do with sort of vertically integrating the entire robot manufacturing and R &D. So they’re not outsourcing and using supply chains. They’re building everything in-house. And when you actually read about how they’re building their chassis and how they’re building sort of these, you have to build a pretty impressive density of three things within parts of the robot, which are power,
19:43
perception and compute. In a hand, you need enough power, enough energy, you need enough perception, which is often sensors, but also usually a camera in the wrist for some of these. And then you need enough compute within there to make it all work. That sort of density of different functions within parts of the robot is really impressive. And I think they’re very good at that. I think their robots are still, I’m sure there’s others that are maybe up at their level, but it’s hard to see anyone who’s ahead of them.
20:13
in just sort of agility, dexterity, speed of the robot body. And most, a lot of companies, not most, a lot of companies sort of avoid bipedal robots and they do wheeled chassis because it’s much simpler. Or robot dogs are simple, but no, mean, Unitree is doing bipedal. Their robots have four legs and wheels and their robots jump up on their back legs. They do double flips now. I mean, it’s really quite impressive. The other…
20:41
R &D areas which are interesting are they’re what they call the joint, the integrated joint modules that house the actuators. So they have these, you the actuators are in certain parts, elbows, anywhere there’s joints, right? And they have these sort of integrated sealed up modules where everything’s within there, which protects it from dust, water, things like that if you’re going to have these outside. And I assume it also makes them easier to replace.
21:10
that is a key subcomponent. That might be the one thing I add to my list of what is really important in these robots. Brains, eyes, hands, maybe these joint modules for actuators where they can generate a certain amount of power in these key components. So that’s kind of what they’re doing that I was most interested in. But also their robots do things like they talk about shock absorption.
21:38
that if you’re going to jump robots, have to be able to handle the shock, durability, weather protection, obviously. So anyways, when you get into how they’re building their bodies, that was the part of the filing that I think impressed me the most was the hardware sort of innovation, mostly in these areas, not so much within the brain, the hands or the eyes. Actually, I’ll talk about that. And they’re doing it vertically. I mean, they’ve vertically integrated into this, which is the same approach BYD took, which is why
22:08
They’re so cheap because they’ve basically integrated up into batteries. That’s what Tesla has done. So that’s interesting. So that’s number two. Takeaway number three. They really seem to be late to the game to embodied intelligence. Doesn’t really seem… This company’s been around for, I don’t know the exact, This robot dog started as a…
22:34
really almost a college project of the founder and then went out into work and kept at it and eventually started his own company with this. There were other founders involved, but they’ve been in robots for a long, long time, really early on. But the embodied intelligence aspect doesn’t seem to be until 2024, 2025, which is late. As I mentioned, like when I was looking at LinkerBot, which does these hands, which you have to have embodied intelligence.
23:03
They started that in 2019. So I think that’s issue. Now, one of the ways you can get around that is you can have other companies port their intelligence into your robot. And Unitree does work with Tencent, and Tencent is building basically embodied in… Tencent wants to be the Android of robots, where anyone can make smartphones and you can just download Android.
23:32
Most of it’s open source and free, some of it’s not. And you can just make your smartphone. Well, this will be the same. You can just download the operating system, the nervous system, and the brain into any robot, and then you can train it where you want. So they do work with Unitary, although I’m not really sure what’s in that deal. But yeah, I think that’s a problem. Now, if you actually look into what they’re doing in terms of embodied intelligence, Unitary, they have a world model.
24:01
One, have weird names, Unifo, LM, WMA, zero. So Unifo, I assume LM is language model, WMA world model action. So it’s an action world model. That was released September, 2025. That’s really not very long ago. And these world models are super important. The other one, Unifo, LM, VLA. So that’s their VLA model.
24:29
Also, the way I’ve explained these before is you kind of run both of these models and you can build something halfway in between them as well. But a world model is like putting a camera on a billiards table and the world model sort of projects various future states. If you hit the ball like this, the cue like this,
24:56
what will be the various potential future states and then you choose the one highest probability that works. So to do that, have to, the model has to look at the pool cue, because that’s your action. It has to look at all the balls. It has to look at the table. It has to understand physics to know how they’re going to move. And then it projects the future. Say that’s a world model. A VLA model is more like what comes in the camera
25:24
in terms of the vision of the robot, what it sees with its eyes, and then it responds to something very directly in the physical environment. So that would be like if you hit a ball, a baseball, and you got your hand out to catch it. The VLA model, that’s what it does. It sees the ball, it moves the hand very directly, it catches the ball very quickly specifically. It doesn’t have to model the whole environment, but it’s got to catch the ball.
25:51
So VLA to me is like, okay, can it catch a ball if I throw it to it? And world model is it, it project if I hit the pool cue, what’s going to happen and what’s the best way to hit the pool cue? Anyways, but you can actually do stuff between those, which are kind of halfway and Tencent’s actually doing that. So those are the two from Unitree. There’s another one, is Unifo LM, the X1, which is, I think they’re building specialized models for various industrial situations.
26:21
I think they have one for industrial, which is the X1, and I think they have one for the home care, which is the Omni-A. I’m not really sure what those are. Okay, but yeah, mean the VLA model came out in January. It’s August. They embodied they’re not on the frontier. So that’s, yeah, that doesn’t necessarily mean it’s a bad strategy. Maybe it’s better to specialize in what they’re doing and just partner. Maybe.
26:52
Okay, and just get to scale and manufacturing and get usage. Who knows? We’ll see how it plays out. All right, that’s number three. Number four, the hands don’t appear that advanced. The number in the filing is 10 degrees of freedom for the hand. That seems very low. I’m not 100 % sure, but that’s the number that was there. Yeah, I would have expected something like 20s, 25. And yeah, so that’s a concern.
27:22
Now again, you could partner with that. You could do a deal with LinkerBot and LinkerBot only makes the hands and they’ve got a really good use case. I wrote about this last week. They’ve got maybe the best robot use case I’ve seen, which is, you know, they go to people who have factories. I know Samsung, Siemens, all these factories have robot arms that they’ve been installing for 15 years, but none of them have hands and they just sell them hands that you can put on the end of your
27:50
existing robot arms in your factory and dramatically increase what they can do. That’s a lot easier than deploying humanoids into a factory that’s not built for humanoids. So the growth story and the use case for linkerbot is pretty great. Take a look at that article I sent out if you haven’t. All right, that’s number four. Number five, last one.
28:13
It doesn’t seem to be that big of a company in terms of their operating capabilities. Now the sales, they argue that they’re number one for robot dogs, which the numbers I’ve seen is they’ve sold 23, 25,000 robot dogs in total, which is the leader by far. They’ve said that they’ve sold, I think JPEG was going around from then that they sold 15 to 18,000 humanoids.
28:41
Now I’m not sure about that one, but they’re definitely big. Like they’re selling a lot. I know the growth numbers, kind of went through those, but when you actually look at the operating capabilities behind that, they have 516 employees. That number surprised me probably more than any other number. You know, that’s, that’s 2025, but 516 FDs in 2024, they had 328.
29:10
Okay, LinkerBot has 400, agibot has 1500. If your game is, we’re going for scale, we want to be the technological leader, this would be a standard strategy for anyone doing hardware, right? We want to be the technological leader, edge of the frontier in terms of capabilities, which they’re trying, and we want to get to scale and manufacturing before anyone else, because that’s how we’re going to drop the price, we’re going to get easy, right? This is a standard China hardware game.
29:39
Now think in this case there’s another piece which is like, are you competing as we’re a hardware manufacturer, that’s BYD, or are we competing as an embodied intelligence ecosystem? That’s Tesla. Tesla is never going to beat BYD in manufacturing, but their AI could. which game are you playing here? Well, I just kind of said like Unitree doesn’t have a lot of embodied intelligence, okay, maybe they’re playing the hardware game. And then I’m like, oh, they’ve only got 500.
30:08
employees. That seems relatively small. I would have thought they would be bigger. Now they’re using the IPO funds apparently to build a robot manufacturing base right now. So we’ll see. These are early days. This could all be very different in 12 months. But yeah, they were significantly smaller than I thought. And when you break down the scale of the 516 FTEs,
30:36
35 % were in R &D. Okay, that’s pretty cool. 30 % were in sales. Fine. 23 % were in production.
30:47
Okay, that doesn’t look to me, that looks to me more like an R &D hardware innovation company than a powerhouse manufacturer.
30:58
But who knows, these are early days, everything’s moving fast. But yeah, that’s kind of my fifth takeaway that they didn’t have the scale, at least on the manufacturing and operating side that I thought they might have. I thought it would be better. Anyways, those are my five takeaways. Let me switch gears and sort of go over to Tencent. It’s a pretty interesting contrasting story. Now I’m not going to go too much into this because it’s…
31:25
It’s pretty technical. As I said, look, the email is really technical. So I’m not sure that’s how many people are going to like that. Well, me going through the technical bit on a podcast, well, that would be even less interesting. So I’m not going to do it, but I’m going to kind of hit the main points. There was a Tencent forum back at the World AI Conference in Shanghai a month ago, which was really cool. And I’m writing up, I know, probably five or six articles about it right now, about what kind of they’re focused on within AI.
31:55
What are the highest priorities? They’re doing a lot, but what are the big priorities? Because they’re very capable. They’re a super capable company. So when they focus on something, stuff happens, right? Anyways, the number one takeaway I had from that forum was like, they are way more focused on robot brains and body AI than I thought they were. I knew they were doing it. I’ve seen their sort of T-Rose Robot X joint venture, but it was way more of a priority than I think I’d appreciate it. So I’ll talk about that.
32:23
Second takeaway not related to robots. They have, they basically have what they call buddies. They have six buddies and you can literally walk into their exhibition or whatever and they’ll say choose your buddy and there’s six different buddies. Some of them are AI assistants, some of them are full agents and it’s pretty cool. And they’re focused on sort of making these tools super easy to use, especially within the Tencent ecosystem.
32:53
So they have CodeBuddy. Only coders are going to use that for the most part. The one they’ve been talking about most that’s an agent is WorkBuddy. You’ll see ads all over Shanghai like Metro for this. It’s going international right now. WorkBuddy like Claude Cowork is impressive right in the West. This like Tencent’s WorkBuddy is the China version of that that I’m watching. And then actually Quoter Work by
33:22
Alibaba is the other one I’d watch. This could be your enterprise agent. And I think I’m betting that Tencent’s going to move faster here. Anyway, so they have six buddies. I’ll write about these and I’ll send out the email. But WorkBuddy’s sort of their flagship AI agent product right now. It’s totally worth looking into. I’ve written about it I’m going to write a bunch more. The third one, they talked a lot about AI marketing. And that was really…
33:53
When you look at like who’s going to be the best at combining e-commerce with AI in this world? Because AI totally changes e-commerce. I would argue it’s probably going to be Alibaba or Amazon. Obviously they have both businesses, so you put them together and they’re leaders in AI and they have major e-commerce. So you would expect that. Who’s going to be the world leader in AI meets marketing?
34:18
I’m thinking it’s going to be Tencent. Now it should be Facebook. They do a lot of marketing, even Facebook doesn’t do, I mean, they’re an advertising driven business. But if you look at breadth of products, Tencent does advertising marketing across so many different products. Messenger, video games, literature, movies, chat. mean, they are, now that’s, it’s not their main, well, it’s their number one usually revenue source.
34:46
but they’re actually pretty diversified in terms of revenue. But it’s hard to think of a more robust marketing company in the world that is also on the frontier of AI. So I’m thinking AI marketing. I’m thinking Ten Cents the Company to Watch. And I’m going to dig into that. Actually, I’ll write an article about that this week probably. OK, those are my three. Let me just talk about the embodied AI bit, which is called Tairos, T-A-I-R-O-S. There’s a joint venture they do with
35:15
some laboratories, but Tyros is sort of the Android, right? It’s an Android operating system plus other stuff that any robot manufacturer can just download and use and build their robot and it’s going to work.
35:33
So it’s sort of a tech stack, but it’s also supporting services. And big surprise, you can then use this on Tencent Cloud. So it ties nicely into their cloud business, but it’s just a really cool thing. All right, now in terms of understanding it, there’s really just one, two, three, five things to keep in mind. They have sort of their foundation models, their intelligence models.
36:02
They have some supporting tools for developers to build these things, to do whatever you want with a robot, and then to run the thing. And then they have some sort of supporting. So there are really five things to keep in mind. And I’ll just give you the list. It’s basically RxBrain.
36:21
Where is it? Let’s do, A vision language model, which they call VLM 1.0. And then it’s a vision language action model, which they call VLA, this one’s 0.5. Those are the three models. And then under that, they have basically developer tools for building these things and running these things. They call them agents. It’s not really an agent like we talk about. It’s not like AI agent.
36:50
That’s a little confusing. Anyways, one’s called TIGEROS Agent. That’s basically a, it’s a developer SDK kit and the API. any developer can download all this, the library, software kit, all of that, and you can build your robot. And then they have something called Apexio, which is, that’s what actually runs the robot. So in terms of those last two, the way I think about them is,
37:19
If you are making a video game, okay, a developer wants to make a video game instead of a developer of robots or a company of robots, know, hardware maker making, you’re making a video game. Well, you would design the game using the game engine of a company, right? That’s Taro’s agent. It’s the game engine that sets up the whole system so that this thing can function. And in this case, really what it’s doing, you
37:48
It’s the development kits, this code libraries, it’s the open AIs. You get all that from them. That’s Tairos agent. You build your gaming engine. No, basically build it to run the thing. But what it’s really running is it’s sort of running connection loops. Like you got to get everything to work. The brain has to connect very quickly with the sensors. The sensors have to have feedback loops. You need three or four loops to do real well. Well, this is setting all that up. So if this was video gaming,
38:17
we would call this building the gaming engine. Apexio is more like playing the game. Okay, you’ve built the thing, you’ve deployed it, but now you need to control the thing and run the thing. Instead of building the video game, we want to play the video game. That’s Apexio. So this is how you control the robot or fleets of robots all together and it, know, sensitive. There’s other kind of the two developer tools that I thought there’s other ones, but those are the two that…
38:47
were interesting. Now within the embodied intelligence bit there’s three as I mentioned. There’s the RX brain, VLA, VLM. The important one, well they’re all important, but the most interesting one is RX brain which is basically their world model. So this is, you couple with the visual input with language. We want to hit the pool cue and make the balls go into the hole.
39:15
that creates a sort of future imagination and state, oh, the ball could go here, you got to integrate the physics, all of that, and then it predicts the various outputs, and then it makes the best call for what to do. Now, their world model seems a bit more, it doesn’t seem like a pure world model like Agibot. It seems sort of halfway between Agibot, a pure model, and a VLA. So it’s more tied to sort of direct action.
39:44
And that’s a spectrum, but I’ll write up if you’re curious. I’ll write up like how it’s different with what they’re doing and Yeah, was I won’t go into that detail, but there are some differences in how they’re approaching this the vision language model Well the VLA which I talked about, know, that’s the robot catches the ball in the air, right? It’s a very tight high-frequency feedback loop Well, not that high frequency but higher than the world model
40:14
You know, moving the hand, seeing the ball come towards the hand, adjusting the hand, know, grabbing the ball, catching it in air, using the right amount of pressure if it feels like it’s slipping, increasing the pressure, all of that. Okay. That’s the VLA, the embodied VLA, which all of their models are HONYUN, which is, you know, their suite of LLMs. Everything’s called HONYUN, HY, HY3. Okay. This is the HY embodied suite. All of these models fall under that.
40:43
fine. So they have the VLA. I’ll go through that in the email. But then they also have an additional VLM which is just more horsepower to visual processing and separating that out and then that feeds into the VLA. So okay I think the other two the VLM and the VLA look pretty similar. I’ll write out the details but there was the RX brain that I thought was I think and that’s what they talked about a lot in the presentation was the RX brain.
41:13
I think that’s where they’re really trying to be different and ahead of the curve. And then the developer tools, the Tyros agent and the Apex seal. Anyways, those are the five things to kind of keep an eye on. I think it’s really important. I think this is going to be super important in sort of understanding how these things go forward. I think a lot of the competition in why company A might be company B is going to come down to this intelligence ecosystem. Like we know hardware scale.
41:43
is going to matter. That’s the BYD model which was crushing in EVs. Their cars are just better and they’re cheaper and nobody can match their price point. Well, at least not sustainably. And then you have the Tesla model. Okay, we can’t beat them in hardware manufacturing at massive scale, but we can be better at intelligence and AI that drives the car. I’m assuming my working assumption is it’s going to be similar. within the intelligence aspect,
42:11
Tencent and this sort of breakdown I think is super important You know, it’s that’s kind of where I am. I want to go to a lot factory I want to try to start putting together the cost structure some of the companies they’re working with this is from their slide I’m not sure what these deals are but Bosch ad you bought unit tree Anyway, so some pretty big names working with them. I’m going to I’m going to see what’s what I’m going to try and find out more about this in the next couple weeks
42:37
Anyways, that’s enough for today, but that’s kind of where I am in my thinking in all this. And I think it’s becoming more clear. think one of the things I’m sort of thinking is like, China is just going to be a juggernaut in robots. I always sort of had my working assumption where like, I don’t know how you beat Chinese manufacturing and I don’t know how you beat Chinese manufacturing and their AI, which is open source and low cost. Each of those was really powerful on its own. When you put them together in robots, how do you beat that?
43:07
I think my opinion is getting stronger and stronger. This looks like a juggernaut. mean, look what’s happened in EVs over the last three years. They are everywhere in this world. Unless they’ve been banned or kept out, they are just rolling through this world. And the Chinese EVs are great. The prices are amazing. And I’m thinking it’s going to be the same thing. yeah, anyways, my conviction is strengthening, but we’ll see. Anyways, that is the content for today. I hope that’s helpful.
43:35
Two concepts, VLA models and world models, really important. And also I think hardware, the economics of robot manufacturing. I’m trying to get some numbers on that. I think that’s going to be a big, big deal. We’ll see. Okay, anyways, that’s it for me. As for me, I’m going to be on the road for the next couple weeks. I’m going to Manila in a couple days. I’m teaching the course with Alibaba to a bunch of entrepreneurs there.
44:04
That’s super fun. mean, the content, it’s kind of what I was saying earlier. A couple of the classes are about the intersection of AI and e-commerce. And man, Alibaba is so far ahead on this subject. It’s really like, I’m as much of a, cause I’m teaching my stuff, right? But I’m as much of a student on a lot of this other stuff coming from them. Really spectacular. A lot of AI marketing, how to build one company, one person companies using AI, all of that.
44:33
So that’s a lot of fun. Then going up to Beijing for the World Robot Conference, which is in a couple days, end of this week, the World Robot Conference is, that’s going to be fun. I’m not going to be there too long, just sort of run through and see what’s there. And then there’s the World Robot Competition where they fight and do track and field and all that. I’m going to go to an event there as well. So it’s going to be a fun couple of days. Anyways, that’s it for me. I hope everyone is doing well and I’ll talk to you next week.
45:03
Bye bye.
——
I am a consultant & keynote speaker on how to increase digital growth and strengthen digital AI moats.
I am the founder of TechMoat Consulting, a consulting firm specialized in increasing digital growth and strengthening digital AI moats. Get in contact here.
I write (a lot) about digital growth and digital AI strategy (3 best selling books, +2.9M followers on LinkedIn). There is a free book and email newsletter below.
My Moats and Marathons book series is a framework for building and measuring competitive advantages in digital businesses.
Note: This content (articles, podcasts, website info) is not investment advice. The information and opinions from me and any guests may be incorrect. The numbers and information may be wrong. The views expressed may no longer be relevant or accurate. Investing is risky. Do your own research.
