Subscribe
Share
Share
Embed
Deep conversations with the founders, investors, and operators building real-world AI - robotics, automation, industrial systems & AI infrastructure. Past the headlines, into how these technologies are really built, deployed, and scaled. Hosted by Bogdan Cristei, venture partner and former systems engineer.
Bogdan Cristei (00:00)
This is the Optim Update. I'm Bogdan Cristei, and today I'm talking to Mike Xia, co-founder and CEO of Anvil Robotics. Mike is building what he calls the foundry for physical AI - the hardware, software, and data tools that let robotics teams go from zero to training without spending months assembling their own stack. They've shipped over a hundred robots, they manufacture in Taiwan, and they just raised a six and a half million dollar seed round.
We get into why the existing robot hardware was built for a completely different era, what it actually costs to build and ship a five thousand dollar arm, and where this infrastructure layer is heading. Without further ado, here's my conversation with Mike Xia.
Bogdan Cristei (00:45)
Mike, welcome in. How are you?
Mike Xia (00:49)
Doing well, thanks for having me, Bogdan.
Bogdan Cristei (00:51)
Are you ready to talk about robotics, physical AI, and everything in between?
Mike Xia (00:55)
Always ready to talk about robots.
Bogdan Cristei (00:58)
Cool. All right, so I'm super excited for this. Here's what I want to cover today. I want to start with the problem - what the physical AI infrastructure that exists right now was built for, which was maybe a different era, and what that means for teams trying to build today. Then I want to get into what you've actually built at Anvil - the hardware, the software, the manufacturing, the economics. And then we'll ask you some hard questions about open source, moat, and hopefully close with some hot takes. Does that sound good?
Mike Xia (01:28)
Sounds good to me. Let's do it.
Bogdan Cristei (01:30)
Cool. Let's get into it. Help me understand what teams are actually dealing with today. If I'm a physical AI startup, I want to go from zero to training my first model. I have these SO-100 arms, open arms, I think I need or I think I want. I'm trying to train my first model. What does the process actually look like?
Mike Xia (01:52)
I first want to paint a picture for everybody. If you look at robots in the factories, people usually think of Kukas and Yaskawas and these large industrial robots moving really quickly. And then you have these leading physical AI labs like Physical Intelligence, and you're wondering - why does the robot look so different? It's kind of small, black colored, totally different worlds.
It's actually because a lot of the hardware and infrastructure was built for a totally different market. For the last 10 years, industrial robots did a lot of automation in factories. And that whole technology chain - the hardware, the software, the controls - all headed in that direction. The needs in the factory are really different because you're talking about very fast throughputs. There's usually no humans around, so you can have a lot of inertia, a lot of mass in the arm, because you're not worried about hitting anybody. You usually have some laser gauge or cage to keep the robot away from people.
But that looks really different for physical AI - robots in your homes, robots in your kitchens, serving food. The needs and concerns are super different, and as a result the hardware and technology stack have to follow. We're actually seeing a reset in the robotics technology foundation for this new industry because the requirements are so different. If you're trying to build a physical AI model or startup or application, you usually don't start with these Kukas and Yaskawas because you end up fighting the hardware and the software in order to get what you want.
High-speed controls are super important for physical AI because you need to be responsive. You're doing closed-loop controls - seeing, planning, acting on a kilohertz loop. Whereas a lot of the existing off-the-shelf robots or cobots are generally doing seeing, planning, acting maybe once a second or once every five seconds. You're picking up an object - see it, grab it, put it there. So it's a very different paradigm completely.
This means teams have to build a lot from zero. Actually, Bogdan, I see you have an SO-100 in the background. How long did it take you to set that up and train your first model?
Bogdan Cristei (04:10)
So it was actually very annoying, believe it or not. When I first set it up, for some reason some of the wires wouldn't connect, so I would lose one joint and then everything downstream would be lost. I had to unplug and plug everything back in a bunch of times. Then I recorded my first five episodes and realized the camera wasn't positioned right. Then I did 50 episodes, but I kept interrupting by forgetting that I can't put my hand in there. It's not trivial to get this stuff done. It helps to have the infrastructure.
Mike Xia (04:59)
Did you 3D print it yourself or did you get a kit?
Bogdan Cristei (05:07)
Good question. The first arm, I ordered it. The second one, I ordered the motors but I 3D printed the parts. Just so I have a reason to buy a 3D printer. So now I'm trying to set up the bimanual stuff.
Mike Xia (05:21)
Well, first off, I think Hugging Face and LeRobot have already done a really good job of making it easier, because it could have been 10 times harder than what you experienced. Not to say that what you've gone through is at all easy. But that's for a small toy arm that can do 200 or 300 grams of payload.
Imagine the additional leap in complexity when you want to build a system that can do two to three kilograms. Then you're talking about sheet metal parts, actuators that can do 20 to 30 Newton meters, CNC pieces. If you disconnect a cable and these are QDD drives, instead of a little bit of plastic going limp, you have this five kilogram thing just plopping down onto whatever workspace or table you have. The problems and the risks and the pain compounds as you get closer to real payloads and closer to production.
Teams today are doing basically what you had to do but with a lot more pain. If you buy a Yam arm, or an I2R, a Galaxea A1, or an Agilex Piper, or any of these arms, you're effectively buying a high-risk piece of metal. You have to go write the controllers yourself, write your inverse kinematics yourself, put together safety and non-collision - things that are effectively basic non-negotiables. Making sure you're not colliding into your own arms or bumping into your camera stands. When you control them, are you controlling at 100 hertz or 1,000 hertz? Is your camera frame rate at 30 or 60? Are you dropping frames? Are you dropping CAN bus connections? Is there jitter in your CAN bus?
It ends up being - the way I would put it - everybody is selling a component, but really what's needed is a system.
If you think about the system as a whole, all the pipes, all the loops, these things actually start mattering. Teams are doing that themselves, and it just becomes a huge pain in the butt for everybody. It's not the area you want to invest time into. You want to get the model to pack boxes. But you spend the first six months tuning your PID and your controls and your gain values just to be able to get the arm to move smoothly and not do crazy wobbles.
AI teams want to build the application but they find themselves building the infrastructure, which is not a good place to be.
Bogdan Cristei (08:01)
So if I'm a team and I come to you tomorrow as a robotics team, what do I get that I just cannot get on day one by myself?
Mike Xia (08:10)
Thanks for setting me up for this one, Bogdan. Anvil is specifically trying to solve this problem. We're positioned as a platform company. If you buy a dev kit from Anvil, effectively we've solved all the inverse kinematics, the low-level controls, the reliability in the cabling, the reliability in the CAN bus, the reliability in the camera frames, getting all that data into a common format like MCAP, and then having pre-built ETL scripts for converting that into LeRobot format, and then integrating that with LeRobot so you can train ACT, diffusion policy, pi0. All of these things, with an Anvil DevBox, you could basically do in the first two to three days.
We're trying to compress this tremendous amount of gradient descent and pain and trial and error into a really polished, easy-to-use dev kit that you can just skip a lot of the pain and get straight to validating your model or your application.
Bogdan Cristei (09:12)
You mentioned your sensor backbone runs at kilohertz rates. Classical robot control is orders of magnitude slower. Explain what that speed actually unlocks. Give me a concrete example of a task that fails slow, works fast, and how you guys solve that.
Mike Xia (09:19)
The right way to think about it is slow controls kind of dulls your robot's senses. When you see robots move in really fluid motions, the reason it becomes fluid is because at the lowest level when you're sending commands to the joints, it's running at a really high speed. You're basically saying every second, here's a thousand different poses you're going through. That creates these really fluid motions and lets the robot do very human-like actions. My arms are being controlled at probably 30 kilohertz or whatever by my brain. If you cut that down to 10, it looks a lot like those SMT machines where they pick and place parts on a PCB.
They move in these really straight lines really quickly. That's basically what the whole industrial robotics paradigm looks like - going from point A to point B as fast as possible, because in the factory throughput is what matters most.
But with physical AI, throughput is not the thing you're optimizing for. Generalizability is the thing you're optimizing for. Obviously you want some degree of throughput, but in order to navigate complex environments, the straight line no longer works. Chipotle's not going to modify their back kitchen. McDonald's not going to modify their frying system to turn it into a factory in order to use classical robotics. Instead, we want to make generalizable systems that are intelligent, that can live and move fluidly in all these different diverse environments.
Bogdan Cristei (11:19)
So I remember you saying something interesting - that today's actuators force a choice between high payload and force compliance, and you say that's a false tradeoff. What made you say that?
Mike Xia (11:34)
Going back to Physical Intelligence, you can see they have these kind of small arms, and the actuator inside is called QDD or planetary. It's really different from what's used in industrial robotics, which is typically cycloidal, harmonics, or belt drives. The primary difference is the amount of reduction ratio or gear ratio.
Harmonics are known for 1 to 150 or 200 gear reduction ratio. So you're able to have a very small motor spinning really quickly, and that converts a 1 Newton motor into a 150 or 200 Newton actuator - a tremendous amount of force. It's almost magical. But what ends up happening is when you have that really high gear reduction ratio, it creates a ton of friction in your gearbox.
So when you're moving, you're kind of blind to the world. The example I'll give is if I'm putting my hand on a table and I close my eyes and I have no sense of feeling, I can bump into the table and just crash right through it. That's why in classical robotics, if you go to any YouTube videos of robots on factory floors, they're always moving through air. There is nothing in their way, and that's by design, because if something was in their way you would have 20 kilograms of metal going at 200 Newton meters with a ton of rotational inertia - you'd crash right through it. If it was a person, they would probably get really hurt.
Now, with physical AI, you have very low gear reduction ratios, usually 10 to 20, maybe 40 at most. And the motors are also equally small, so you're getting maybe 20 Newton meters. What this unlocks is you can actually, from the motor's current and the output encoder's data, tell if you bumped into something even without fancy force torque sensors. If you can do this at kilohertz - every second you're reading a thousand times: how much current am I pushing? Am I bumping into something? - you're actually able to build relatively intelligent and responsive systems for contact detection, whether it's a person or a table, and to some degree even tell how much force is acting on the actuator. You can tell how heavy the thing you're holding is. Did I bump hard into something or just a little bit?
One really cool thing we're going to be releasing at Anvil is we have a little bit of this in the gripper now - a small RL model that detects whether or not you're gripping really tightly. With a hundred-dollar actuator, you're able to grab things like eggs and grapes and not crush them. So to some degree that technology is going to make it into these very affordable robots.
Bogdan Cristei (14:35)
I'm so happy that's the case. I really think vision alone is just not enough. I went to the dentist recently and the entire bottom half of my mouth was numb. I was literally looking in the mirror trying to drink a glass of water. I had all the visual inputs I needed and I still couldn't do it. The water was just running down.
Mike Xia (15:15)
Do you ever get it when your arm falls asleep? You wake up, your arm's asleep, and you're trying to reach over. You can move it for some reason, but you can't feel anything, and it's got this really uncanny quality. I have a feeling that's what these current generation of models is experiencing. You're just kind of going for it.
Bogdan Cristei (15:35)
Yes. So it's not quite tactile haptic feedback, but it's a clever way of adding another vector of information to the visual cues.
Mike Xia (15:54)
I think it's actually really important. One example is if you close your eyes and you have a water bottle, you can open and close the water bottle just fine without using your eyes. You can kind of feel the contours of it. I think this is a hill I'm willing to die on - force, maybe tactile, our senses are going to be really important for physical AI.
Bogdan Cristei (16:15)
100%. I agree with you. All right, let's move on. I wanted to ask about the factory. You manufacture in Taiwan. You ship custom configurations in one to two days. I'm super curious about the economics. What does a $5,000 arm actually cost you to build and ship?
Mike Xia (16:42)
I think this is something that, unless you've lived and swam in the supply chain for long enough, it's really difficult to grok. I highly recommend anybody who's serious about robotics to - if you can't go to China, go on Taobao. Ask ChatGPT for the Mandarin search term and then just browse. But if you can, you should definitely go to Shenzhen. Go to Huaqiangbei and try to soak it in and see what it's like over there.
There is just an immense amount of infrastructure. It's basically walking through the Amazon.com of every component, every connector, every vendor that you could possibly imagine, all in one place. This creates a lot of competition, which drives pricing down. It also creates a lot of second-order effects where talent - people who make tools, who do assembly, who do careful things like wiring, cabling, or tapping.
Something I learned recently was the screw holes and the threading - a lot of it is human labor. It's using something called a tap where a person actually has to spin this tap around and figure out if they hit the limit. It's a bit of a technique because you're trying to screw a hole into something, and if you have some shavings you have to take it out, pour it out, then do it again. Imagine trying to do this in the US - thousands of people holding this thing and being like, did I get it right? That's what gets solved almost by default in Southeast Asia, Taiwan, China. You have large amounts of trained skilled technical labor as well as supply chain infrastructure for building physical things.
It makes building a $5,000 arm very affordable. On our end, we probably keep maybe 20 to 30 percent gross margins at the end of the day. This is accounting for all of the software we do, all the R&D, all the customer service attached to it. For our volumes - we've shipped maybe 150 robots to date - it's actually really good for a low-volume business. That lets us do all these things that normally a company would have to go to an ODM for. How do you customize the actuator? How do you change the kinematics? Can you develop a new product relatively quickly? Can you ship just three of these with positive unit economics?
These are things that make it actually possible for us to build a company like Anvil. Whereas typically you'd have to raise a hundred million dollars, go to an ODM, and convince them to build one SKU for you, because nobody wants to build a hundred robots - they want to build a hundred thousand robots. So we fill this middle gap in the current stage of the market for physical AI.
Bogdan Cristei (20:01)
So you can do low volume, but if you need to 10x or 100x the volume, you can probably do that too.
Mike Xia (20:19)
Our volume is growing. I've been asking our founding mechanical engineer and factory head how much we can get to. I think with our current setup, we can probably get up to about 200 robots a month. We have enough room to run this way without needing any major overhaul of our sourcing, supply chain, assembly, and QC processes.
Bogdan Cristei (20:49)
So you basically have a factory where you can drive 20, 30 minutes to somewhere, pick up a part, try something, come back. The turnaround time for anything that could go wrong is probably hours, not days.
Mike Xia (21:08)
That would be true if we were based in Shenzhen. But because we're in Taiwan - and there's a long and important reason why that's the case - we actually incur about a three to four day lead time, because things have to get shipped overseas and go through imports and customs in Taiwan. So not as fast, but still pretty damn fast.
Bogdan Cristei (21:30)
Probably much faster than if you're in the US. So you've said all your customers to date came kind of inbound. Is that still true?
Mike Xia (21:48)
It's mostly inbound. I guess word of mouth also counts as inbound, but a lot of YC founders use Anvil systems. They talk to each other - they're in the same batch. Once you see one person with a robot with the logo, you're like, hey, where'd you get that? I think we might need one pretty soon. That's been super helpful for the company so far.
Bogdan Cristei (22:10)
I'm curious about retention. Of the teams that came to you, what fraction has come back for second or third orders?
Mike Xia (22:24)
We're in a bit of a strange middle ground, because if you go to Anvil's website, you'll see it's a Shopify store, which gives you this e-commerce vibe. It's a very self-serve motion. As a result, it's become very easy for people to come back and say, I need another gripper because I wanted to swap something out, or I need another robot because I hired a new engineer and they need something to develop on. I'd say 20 to 30 percent of our customer base comes back and buys something again.
One interesting thing I realized recently is that my customers - all these new physical AI labs and companies - are also experimenting and gradient descending on the stack they want to stick with. One stealth company out of Switzerland bought every kit we had. We have a leader-follower kit, a Quest teleop kit, an open arm kind of table-mounted Physical Intelligence-style kit. They said, we don't really know what we're going to do yet, we've raised some money, we think all these embodiments have potential. They just bought everything to try it quickly.
It made a lot of sense. They spent less than $50,000 and saved months of development, tried a bunch of different embodiments. After two months they came back and said, we decided to stick with the Quest-controlled open arm form factor. They ended up buying more accessories - grippers, Quest headsets, our DevBoxes - which converted all their existing stuff into Quest-controlled systems. Then they were able to give them to multiple engineers and do more data collection.
That configurability is part of what we wanted to offer. You can come in, get everything, try it, and when you decide on a route you can keep 80 percent of it. The only things you might throw away would be the leader handles - you can convert it into a robot you can use. So that made us really happy to see success cases like this.
Bogdan Cristei (24:47)
Super cool. All right, a few hard questions now. Feel free to push back or disagree. Your reference designs are all open sourced on GitHub. So what stops a well-resourced manufacturer from cloning them? If open source isn't the moat, what is?
Mike Xia (25:03)
I think if anybody has a few mechanical engineers and some funding, it honestly wouldn't be hard for them to clone and reverse engineer any of the robots out there. This includes Unitree's dancing robots, the open arms, the I2R arms, the Galaxea A1s, even the whole Galaxea R1.
I really feel like the value isn't in the hardware design anymore. The value is actually in things that, in the West, we don't fully appreciate. The volume you have is really important - that's a moat. How much volume you have determines how much leverage you have on the supply chain. It moves you on a cost and performance curve that nobody else is on. OEMs want to work with you. They'll come to you and offer to customize things. They'll give you great payment terms - just give me 30 percent of your full year order, you can pay the rest as you ship, and I'll do all the NRE and work for you.
Volume is hard to earn and hard to replicate. You can't just fundraise your way into it unless you're willing to take a really huge bet. So we look at Anvil's position as: we want to drive volume. We're selling these open source dev kits. We give a lot of value away in software - the controls, even the training pipelines - effectively for free. But our goal is to reach a lot of customers, build volume, which gives us leverage on the supply chain, which lets us offer a better product for a lower price. That's the flywheel we're trying to build.
And on top of volume, there's also relationships. Asia is still very much relationship-driven business.
If you haven't sat down with the person making the decision, if they haven't gotten to know you as a person, gotten to know your company and worked with you over multiple years - not just a single project or a single PO - it's hard for them to really double down and invest in you. Similar to how co-founders and investors work in the Valley. They usually have seen founders work through multiple companies or rounds. Co-founders usually have worked together for an extended period. The supply chain is not a transaction.
This is something that's really hard for people to appreciate initially. It's very much a relationship. A big reason why I spend so much time in Taiwan is because in order to build that trust, that partnership with the people designing and manufacturing motors, actuators, sensors, cameras - you actually have to be there. It's a long-term thing, not a PO.
Bogdan Cristei (28:15)
So you're traveling a lot between the Bay Area and Taiwan these days?
Mike Xia (28:35)
Yeah, the other way around.
Bogdan Cristei (28:36)
One thing you mentioned a couple of times is customization driving early revenue. Help me understand to what extent you're customizing things without turning into a services business. Because you can't customize forever. How are you thinking about that?
Mike Xia (29:01)
That's a really good question. We are definitely not a services company. I think right now maybe less than 2 percent of our revenue is from customization services. Usually the way we do customizations is in a way that adds to the platform.
A good example: we had a few customers who wanted us to integrate a dexterous hand into the robot. Nobody else was asking for a dexterous hand, but we could see how other people might want one too. So we weren't opposed to it because we could see how it fits into the platform.
We helped them speak with four different robot hand companies. We got quotes, spec sheets, SDKs. A lot of it was in Chinese, so we helped translate to English. We sent somebody in Shenzhen to visit the factories, meet them, see samples in person. Then we sat down, evaluated the pros and cons, and they picked one. That became the one we bought and integrated, and it became something they could use.
Funny enough, about a month later another customer of ours said, we want this exact dexterous hand. They had done their own research - their team had people based in Shenzhen who spoke Mandarin. They didn't want to do the integration. So we ended up getting lucky where we could just send them the new parts and the hand, and they were up and running.
We take on the services when it adds long-term value to the platform - something we want to offer to all of our customers. In about a month or so, you'll probably start seeing dexterous hands on our website, because we're going to start supporting that out of the box for anyone who wants to try the five-active-DoF or twenty-active-DoF hands. Not that I'd recommend it - it's a very complicated machine to control. But if somebody wants to try it, we make it easy.
Bogdan Cristei (31:30)
Super cool. I may actually look at that just to learn. All right, let me do some rapid fire questions to close us out. First one: how many of the roughly 1,500 new robotics startups from the last 18 months survive for the next 24 months?
Mike Xia (32:08)
That's a hot take. Let's first separate the startups. There's going to be startups building core technologies, startups trying to do go-to-market where they're deploying something today, and we can split core technology into hardware and software.
For people building software systems and brains, I think they'll survive and do really well. There's a lot of venture funding there. It's an area where there's a lot of progress, similar to the LLM world where every few months you see jumps in robot model capabilities. I'm really excited about Physical Intelligence and Generalist and these companies, and where they're going.
On the hardware side, I think that's where we're going to see a lot of challenges. The supply chain is an area where you can't just venture-dollar your way into making it work. You have to build volume, build relationships, be very careful about where you invest R&D dollars because you have to make it manufacturable and cost effective. People playing in the hardware layer will all see challenges because the supply chain today doesn't map perfectly to physical AI. The things the brain needs - tactile sensing, force sensing, compliance, backdrivability, safety - just don't exist in the form we need right now.
Then there's the go-to-market people. If Anvil does our job correctly, I think they'll do really well. We already have companies deploying in chocolate factories, textiles, pot packing. These are the people who will benefit from both the brain and the body companies doing their jobs correctly.
If we do our jobs as the foundational pieces of physical AI, then all the people doing go-to-market will be able to do a lot better. The survival rate will be hardest in the hardware layer. Recently there's been a few companies like Cartwheel - unfortunately, they've built some really cool stuff, but volume, business model, cost - these things eat into you really quickly in the hardware layer.
Bogdan Cristei (34:52)
All right, another one. What's the most technically wrong thing physical AI teams are still doing in 2026?
Mike Xia (35:22)
I think it's in the actuators. We've tested maybe 10 different actuator vendors from Shenzhen and Suzhou and Dongguan and Taiwan. There's a lot of them, and they all look the same on the spec sheet and all look the same the first month you're using them.
But when you get into the nitty-gritty, they are so different. We look at something called breakaway torque in the actuator. Inside the gearbox you have lubricant, and gears grinding against each other inside this medium. Breakaway torque is how much torque you're applying from the motor before the gears start moving.
For some vendors, you see a very clean, tight distribution - I apply this much torque and the motor moves predictably from one direction to the other. And then for some other very popular unnamed vendors, it is a completely non-linear, random plot. It's basically random.
You're sitting there thinking - every time you command the motor, the gearbox is behaving differently. How are you supposed to build a well-tuned, responsive machine if every command, and we're commanding at a thousand times a second, gives a slightly different response? Imagine you're typing on a keyboard and every key decides to be stickier the next day or looser the next time. You would absolutely hate typing as an experience.
This is where we are right now, and most teams are in this mode. This is why you really need a company like Anvil to build out this lower-level infrastructure. I can't see every company customizing the actuator, telling the factory exactly how much breakaway torque to maintain as QC, what lubricant to use inside, what granularity and what process to use on the gearbox, what fill rate on the motors.
The example I always go back to is Amazon has solved this for SaaS companies. They decided what interface in the data center - load balancers, S3 buckets, EC2 instances - those are the modules they provided, and that allowed SaaS to exist as we know it today. We're stuck in this mode right now where people are testing which link to use - should I use ethernet or should I use this obscure protocol I found somewhere? We're in the very early innings, and I believe you need AWS to exist for SaaS to exist. I also believe you need infrastructure and platform physical AI to exist in order for physical AI applications and the market to truly develop.
Bogdan Cristei (38:57)
That's fascinating. I could imagine a situation where you build a thousand robots, you spec the motor, and you just didn't notice that it was over-torquing slightly every time it did something. They all break - the whole fleet starts breaking at the same time and you don't know why.
Mike Xia (39:29)
If you haven't done the longevity test, if you haven't run these motors through a lot of different scenarios and sat down with all the data and asked, what did I just buy here? You're taking on a tremendous amount of risk. Even if you get a good batch today, if you don't have a relationship with your vendor, who's to say they're not going to change the lubricant a little later?
Some of them will leak. We've had ones where we ran them for a few hours and you start seeing lubricant leak out. I'm thinking - I know 10 companies who use this actuator. Did I get a bad batch? Were they in a rush because they have so many orders and decided to cut corners, didn't realize the downstream effect?
This is why it's so important and why it's not a transaction. You have to go there, build a relationship. They have to bet on you as a platform or distribution point to help them win the market. It's interesting because it spans beyond the technical into the business. You've gone through the technical parts, and to maintain that technical quality you also have to solve the business relationship. A lot of this stuff is not intuitive.
Bogdan Cristei (40:40)
If you don't do it day in and day out, it's just stuff. All right, last question. What will look obvious in five years that sounds non-obvious today?
Mike Xia (40:52)
I've got two. One is that force sensing is very important. I know a few stealth companies that are Anvil customers working in this area of doing force sensing in their models, and I think that's going to be super important. There are many reasons why we don't do it today, but I believe it will be important and we will have to do it five years from now.
The other is I believe the market will evolve in a form where there will be some vertical players doing very well - the Figures and Teslas of the world. They've built their own actuators, their own brains, their own bodies. They're going end to end, they're vertical, they're big, they have a lot of leverage.
But I also think there's going to be a big piece of the market of non-vertical players where it just doesn't make sense to recreate everything. I can't see deployments at McDonald's being like, we're just going to build our own robot. They'll probably take a brain from somewhere, a body from somewhere, and then somebody is going to be stitching it together like a solution provider figuring out how to automate industrial kitchens - like Cloud Chef, for example.
I think these three pieces in the stack will help support this other segment of the market, which we hope to be big.
Bogdan Cristei (42:29)
All right. And with that, Mike, where can people learn more about Anvil?
Mike Xia (42:35)
If you go to anvil.bot, you'll land on our website, our GitHub, our documentation, our Shopify store. Feel free to go there and take a look. Otherwise, I very much welcome people to reach out. You can find me on LinkedIn, and you can also email me - I'm mikex at anvil.bot. If anybody listening is curious, has their own hot takes, or wants to chat about robots, like I said at the start of the show, always happy to talk about robots.
Bogdan Cristei (43:05)
Awesome. Really great to have you on. Thank you so much for your time today.
Mike Xia (43:09)
Thank you for having me, Bogdan.