打造机器人领域的 SoftBank | Andrew Kang

Tom Lee raising $18 billion for an asset that went down 50%. Imagine if you had an asset class that was actually going to go up substantially. I would propose that you could potentially raise maybe an order of magnitude more.

Hi, everyone, and welcome to another episode of the Delphi podcast and the Emerging Manager Series. I'm José Medu, your host, and today I'm thrilled to have with me Andrew Kang.

1. José welcomes Andrew Kang to the Emerging Manager series

I've known Andrew for about 8 years. We came up together in crypto, and I genuinely think he's one of the best traders and investors crypto has ever seen. For those in the industry, you'll know him as a legend who's called everything from THORChain early on to Tom Lee's ETH thesis, to calling the alt top in August 24th, when it was a very unpopular thing to say, and many other great public calls.

Then he did the thing that almost nobody in the industry does: He took the playbook that made him successful in crypto and ran it in a completely new field. He famously put $19 million into Figure before he'd even spoken to the founder and before the big markup. Today, he runs Robo Strategy, a public vehicle for humanoid robotics that trades on Nasdaq under the ticker BOT.

Full disclosure: I'm an investor, and it was fully a bet on Andrew himself. What I want to get into today is the method underneath all of it: how he finds these bets, what framework he's using for pricing them, how he sizes them, and how he tells the difference between being contrarian and being the sucker at the table, basically.

Andrew, welcome, man. I'm really happy to be having this chat.

Thanks for the kind words, José. It's great to be here.

To start with, for those who are listening for the first time and don't know your background, can you tell us a bit about yourself?

2. Managing personal capital from Mechanism Capital to Robo Strategy

Sure. I'd say I've been professionally investing for the past 6 or 7 years. Prior to starting Robo Strategy, I was managing my personal capital through Mechanism Capital. We were active both in public markets and in venture markets.

Two or 3 years ago, I started getting seriously interested in the robotics industry. The robotics industry back then reminded me of maybe AI in 2022 or crypto back in 2014 or 2015, in the sense that it was really underappreciated by the general investment community.

It really seemed like we were at an inflection point in which the broader acceptance of the industry was going to change dramatically. The adoption of the products within the industry was also going to change dramatically.

I want to dig into that because I went back through all our chats for this and found a bunch of nuggets. I remember in June 2022, you told me, “Crypto’s done for a while. Cash is king,” which was very accurate.

3. Humanoid robotics as a multi-decade, trillion-dollar industry

But by early 2024, you'd put $19 million into Figure, and you were fully robotics-pilled. We talked about that a lot, which we'll also get into, but I'm curious what happened in those 2 years. What were you doing, reading, and thinking about that led you to humanoid robotics?

It seems like there were a lot of other ways you could have gone. You could have gone into the lab companies, the inference stack, or drones. What was it about humanoids? Did you go through a journey like that, or was it just obvious to you immediately?

I definitely could have gone down the AI path or the data center path. To me, though, there was a learning curve. There were a lot of other people already playing in that field. That doesn't mean it was too late. It was still very early to get involved there, but it didn't seem like it was clear that I would have a ton of edge at that point.

I was still in an exploration phase and a learning phase for a year or 2 after ChatGPT came out. But when I started to look into humanoid robotics and evaluate some of the companies in the space, it was clear that this was almost a pre-ChatGPT phase for these robotics companies.

Of course, it's better to invest earlier than later. The maturation of the industry was still very early, and it seemed like the general understanding of the industry was also a lot lower.

The opportunity and the amount of change that robotics could inflict on the world was just as big as what AI could. It seemed like a really compelling place to start going deep on and to focus all of my time and the company's time on.

Even though to us it seemed like robotics, or physical AI, was going to become a multidecade, trillion-dollar industry, I don't think it was obvious to most. It was understandable because robotics had never produced really massive winners or massive venture-scale outcomes.

4. Finding alpha in multidisciplinary tech fields

The biggest acquisition of recent memory was probably Boston Dynamics, where they were acquired by Soft Bank around 2020 for around $1 billion. Given the skepticism and the lack of takeoff, that's what made it interesting.

It was also so interdisciplinary, or multidisciplinary. You had to understand so many difficult fields, from electrical engineering to mechanical engineering, robot learning, and how deployments work. There was just so much to learn, and I think that makes it interesting both from an intellectual standpoint and from an edge standpoint.

The more difficult an industry is to understand, the more alpha you can generate.

This brings up a question I wanted to ask later, but I think we can discuss it now because it's apropos of what you just said. You recently published this piece, “The Exponential Horizon,” that I really liked, where you say we should abandon short-termism, that trying to time this market is foolish, and that this is not the time to trade.

5. "The Exponential Horizon": Why short-term trading pales next to structural upside — The 2022 DeepMind breakthroughs and signs of true robot intelligence

You wrote, “This is the largest upside risk the world has ever seen, and the gap between the EV of trading versus investing will grow larger than ever.” In your answer just now, you said robotics was interesting to you because it was sort of a pre-ChatGPT moment. AI has had its ChatGPT moment. It's very much in the deployment and growth phase, whereas in robotics, I'd argue we're still before that.

6. What Andrew got wrong & changing views on open-source robotics models

Maybe you could say that the Figure demo a month ago was the ChatGPT moment. I'm curious how you think about timing because it is also dangerous investing before the ChatGPT moment. In a way, you could have waited for the ChatGPT moment for AI and invested in all the labs, and you would have done really well. Arguably, you would have captured most of the IRR rather than investing before that.

There are many industries where people invest before the ChatGPT moment and it never really pans out. You can see people being too early in the late 1990s for dot-com, and there are many historical examples, too. How do you think about being pre-ChatGPT moment, valuations, and timing too early versus being too late?

Autonomous vehicles are another relevant example. People have been talking about them for 20 years, and they've taken way longer. How do you think about timing?

Of course, I think that was part of the initial process: understanding how long it would be until we had humanoid robotics revenue in the billions versus tens of billions versus hundreds of billions.

Based on the research that we did, you can't get it down to the exact month or even the exact year and be 100% confident. But I think you can get it right to the exact range: In terms of tens of billions of dollars of revenue, this could occur within the next 2 to 4 years. That by itself would make the industry really attractive to invest in.

And if we're talking about hundreds of billions of dollars of revenue—

You think tens of billions in the next when?

Somewhere between the next 2 and 4 years.

Across the industry?

Yeah. You just think about it from a quantity perspective, right? You sell 1 robot for $50,000. How many robots does it take to get to $1 billion of revenue? 20,000 robots. It's actually not that much.

For $10 billion, how many robots is that? 200,000 robots. It's not very much at all. I think we can name a lot of things in the world that we make more than 200,000 of.

When we think about cell phones, cars, or PCs, we make those in the hundreds of millions or billions in terms of quantities. That speaks to how underappreciated the scale we could reach within robotics is, and also how quickly we can reach some pretty interesting numbers from an investment perspective and an industry-growth perspective.

How do we get there? How do we know that this is going to be happening soon and not in 50 years?

Well, in 2022, there were already what I would say were the initial-generation robot foundation models. They came out of Google DeepMind: RT-1 and RT-2. You had signs of life and robot intelligence.

For LLMs, the beauty was that you could speak to them and get answers. They were generalizable. They wouldn’t get it right 100% of the time, but they were able to understand your intent and the world from a language perspective, as well as from an image perspective. You started to see a lot of the same early signs of life in these RT-1 and RT-2 models.

7. Generalizability vs. specific programming paths in traditional robotics

For example, one of the tests they did involved placing a bunch of pictures of celebrities—Taylor Swift, Kanye West, and whoever else—in front of a robot. They asked the robot, “Pick up this ball and put it on the picture of Taylor Swift.” It was able to pick up the ball, identify it from a random location, recognize Taylor Swift rather than another celebrity, and drop the ball on her picture.

That was one of the first signs of generalizability for a robot foundation model. You could start extrapolating: if it can pick and place and identify items, then maybe it can start doing more complicated tasks over time. Being able to identify different items and different drop locations is really useful. A lot of work in factories today is just people moving items from one place to another.

8. The Taylor Swift vs. Kanye West task: Evaluating early vision-language backbones — Moving objects in factories: Breaking down the massive "pick and place" market size — Transitioning from industrial code to the human form factor

It seems like a huge market. If we could tap into that, then we could start taking on increasingly more difficult tasks that are still not that hard. Pick-and-place, packaging, and other simple tasks could be enough to get into the billions or tens of billions of dollars in revenue.

Yeah, that’s a great point, and it’s very clean reasoning. I guess the question I would have is this: you could have seen a Boston Dynamics demo 10 years ago, or even longer ago, where the robots were doing all sorts of crazy things. Someone might have been equally impressed and said, “If they can backflip or whatever the hell they were doing, or dance, surely they can move things from one place to another in a factory.”

Even with autonomous vehicles, we’ve had vehicles cross states for almost 2 decades or whatever, but getting them to the point where they’re safe and productized takes way longer than you expect. Was there something that made you think this time would be different?

The early Boston Dynamics demos weren’t really signs of intelligence. They were more demonstrations of the capabilities of the hardware itself—that it could withstand jumping, backflips, or whatever else. But you wouldn’t be able to tell the robot to identify an item, tell you what it is, and then actually interface with or act on that item.

That’s what’s important for robots: actually manipulating things in the real world and doing it under different circumstances and settings. If something moves by an inch, it should still be able to function without that becoming an issue, as it would with traditional robotics. Traditional robots are programmed for a specific path or to work point-to-point.

Maybe you could explain that a bit, because I think it’s really interesting. It’s part of what made you so bullish on robotics: the LLM breakthrough. Robots themselves obviously don’t use LLMs; they use different kinds of models, but they’re using the same principles in a way. It was a breakthrough on the robotic-intelligence side of things.

Could you talk a bit about that, how the models work, and how they differ from LLMs? I don’t think everyone knows that.

They actually leverage a lot of the important research techniques, as well as the underlying models that are important for LLMs. At least the first generation of these robot foundation models used VLMs as a backbone—vision-language models.

A simple way to think about it is that you have a model that understands language and a model that understands vision, similar to ChatGPT or Claude as a backbone. That’s how the RT-1 and RT-2 models could find a ball or a cup. They used computer-vision knowledge to understand what the item was and identify that one image was Taylor Swift rather than Kanye West.

They were using models trained on a wide variety of images to understand what they were seeing from the camera itself.

9. Math of a billion-dollar scale: Why it only takes 20,000 robots sold at $50,000

How do you go from that to taking the right action and applying the right amount of force? That’s the part that seems so sci-fi to me.

The architecture overlaid on top of these vision-language models was the ability to translate a language instruction into actions for the robot. By actions, I mean: am I supposed to grasp something, and how do I move from this point to that point?

The early, simple robots used grippers—parallel-jaw grippers, which are basically hands with 2 fingers instead of 5. The question is: how do I know when to grasp, and how do I move from one point to another?

The way it works is that the robot figures out the joint angles it needs to position its arm in a certain place. Then there are algorithms or controllers—some of which have become more AI-driven over time, while others were traditionally more programmatic—that translate this into motion.

Given a particular robot and a point where I need my end effector or gripper to be, this is where the joint angles need to be. For the joint angles to reach that position, I need to move the motors with a certain amount of force over a certain period of time. That’s how the robot’s functions are translated.

10. The hardware inflection: When traditional controllers become AI-driven joint angles — The Figure bet: Betting on the timeline to billions in physical AI revenue

Before the LLM breakthrough, were most robots’ software models basically deterministic—just a bunch of if-then conditions? Or was machine learning already being used? Did people only start exploring those research directions after ChatGPT?

For the most part, they were very deterministic, which is why things would break.

Something small would be off, right? The object you were trying to move might have moved an inch, and the robot would become completely uncalibrated. If the ground was slippery rather than hard, or if the object you were working with was deformable, you couldn’t apply the same amount of force.

You might have to apply a lot more force to hold a deformable object, such as an empty water bottle, than a full water bottle. You don’t have to use as much force to pick up the full bottle.

That was definitely the biggest breakthrough, and I think it rekindled the interest in robotics that we’ve seen ever since. A big part of your initial bullishness—full disclosure, we were looking at Figure at the same time, and you pitched me on it—was that we were very bullish on it, while I kind of mid-curved it at the time.

Looking back, my mistake was the TAM. You really saw the TAM for humanoids, while I overthought it. I started looking at industrial applications and saw that there were already a lot of robots being used in industrial settings, and that most of them were deterministic. I was left asking, “What is the market for humanoids?” That sounds incredibly dumb in hindsight, and it definitely feels dumb now.

I’m curious how you see the TAM for humanoids. The humanoid bears will tell you that, in most industrial use cases, there are specialized robots that are much more efficient than the human form factor at doing these things, and that they already work extremely well.

And most of them have been superpowered by the same AI breakthroughs that have enabled humanoid robots. So the TAM for humanoid robots is maybe sorting or something like this. But those feel like lower-value tasks, and I'm really curious how you saw it then, and whether your view has evolved at all.

Yeah. So I think it wasn't obvious, and that was maybe part of the fun in evaluating the space: trying to figure out how big this market is right now. There was no market for humanoids—not because there wasn't demand for them. I would say you can go to any company and tell them, “Hey, look, you can buy this robot. It can do the same thing a human could do, except it works 24/7, 365, as opposed to the standard 8-hour workday, 5 days a week.”

If you buy it once for $50,000 and pay a little bit for electricity and maintenance, every company would buy it, right? It's seriously going to reduce their costs, solve any labor gaps they have, and so on. It was just the problem of getting these things to work that was the challenge, not so much sizing it.

If you think about humanoids as being as good as humans, if not better from an economics point of view, then all human labor is your TAM, and that's something like $50–60 trillion.

Human physical labor.

Yeah. All physical labor in the world. But it doesn't mean that humanoids are going to be the form factor that does all physical labor, because there will be a market for special-purpose robots. It just won't be everything.

It's in the same way that you look at other general-purpose devices, like your smartphone. It replaces your camera, your calendar, your watch, your timepiece, your alarm clock, your calculator, and so on. Those are all still existing instruments that you have in the world, but it's more convenient to have something that does all of it at once.

Maybe if you're producing 10,000 of them a year, it's not very economical. But because it's so useful for so many different circumstances and so flexible, you can make billions of them, and it actually becomes a lot more economical than a special-purpose device.

11. Hardware economics & why general-purpose scale dominates niche tools

It's the same thing for GPUs, right? You have GPUs as the biggest computing product in the world, but you also have ASICs, FPGAs, and custom silicon. To understand why general-purpose products really are such a big market and almost dominate special-purpose products, I think you have to appreciate the economies of scale, which some of the naysayers don't fully appreciate.

Then you have to appreciate the business circumstances that some people have. Why do so many factories today, even though industrial automation has existed for so long, still have so many humans? Why do BMW, Volkswagen, and BYD still employ hundreds of thousands, if not a million-plus, workers?

It's because, yes, they have industrial machines, and they're making things at high SKUs. They don't need millions of humans for that. But there are so many things that change on a factory floor from quarter to quarter or year to year. There's a lot of work moving things around, unboxing things, moving and collecting machines, and handling the machines themselves. It doesn't make sense to have industrial automation for all of that unless you're doing it at scale.

There are so many things in the world that exist at not-million or multimillion-unit SKUs. Maybe they're in the hundreds of thousands or tens of thousands, and they're still really big markets, especially when you have them all together.

12. Case study: Path Robotics and the $10B+ automated welding opportunity

Okay, interesting. Even if you think about what the equivalent of the coding use case for LLMs would be—the one that really inflects first and takes this to tens of billions in revenue—do you have one in mind that you think is a killer app for humanoids, whether it's something in industry or something else?

I think tens of billions isn't trivial, but you're going to have it across so many different industries. Just within industrial automation itself, that's easily tens of billions of revenue.

Is there a specific part of industrial automation that you think humanoids are particularly well suited for, or where you're seeing initial traction?

For industrial automation, it's going to be a mix of humanoids and industrial arms, or basically wheeled humanoids as well. You don't necessarily need legs for everything within a factory.

For example, mixed-case palletizing, where you need to move boxes from one place to another and organize them in a certain way, and maybe the way you organize them is different from load to load or from season to season. For machine tending, in some use cases you may just have an industrial arm manage a big industrial machine, but sometimes you need to move parts from place to place, so you need something that's mobile.

There's also stocking shelves within a grocery store, a pharmacy, or whatever. There are hundreds of thousands of these types of stores around the world, or millions. The home is also going to be a tens-of-billions market.

Physical AI and robotics are going to touch every single industry that you can think of, and I think so many of them individually can be so big, which is why the space is so exciting. Space robotics itself is probably going to be tens of billions as well. It's going to be really difficult and uncomfortable for humans to start industrializing Mars, the moon, or whatever planet, and it's going to make sense for robots to do that first.

13. What VCs miss about Figure and physical AI evaluation

Okay, makes sense. When it comes to Figure, you've said it's the company you're most bullish on in the space, and that it's a multitrillion-dollar future company. You've also said that there are essentially zero humanoid companies in the U.S. close to catching up to them, except Optimus. I don't know if you still believe that—that was maybe a few months ago.

But I'm curious, because when you speak to traders, and I know you know this, they really talk about Figure. Maybe it's cope from missing it because there weren't many of the big-name traders in the round. You've told me that you think a very small percentage of investors truly understand how to evaluate and underwrite robotics companies right now. Why do you think there's such a disconnect between your conviction, and that of a few others, and the VC consensus? What is everyone missing?

I think what's funny is that it's changing. After Figure's livestream a month or two ago, we started to see a huge influx of VC interest and demand across basically all of our portfolio companies and non-portfolio companies in robotics—to do follow-on rounds, new financings, and to get exposure to the space. Perceptions are shifting, and I think they're shifting because people are actually starting to see these robots do real work that the robots weren't capable of doing before.

To some extent, people just have to see it to believe it, right? Before, people probably didn't do their research. They just didn't put a lot of thought into it. This happens all the time: We talk to founders and other VCs, and our team members here say, “Hey, look, you guys are some of the most knowledgeable people about robotics compared to everyone else we've spoken to.” A lot of these conversations are just education. We're educating other venture capital investors about how the industry works.

I think the level of education and sophistication is just so low, which is part of the reason it was missed. Maybe part of it is that, as I said earlier, there wasn't proof of concept for massive adoption yet, like you had with ChatGPT. It was the same thing with AI: There was interest in AI in 2020 and 2021, but it went up 100× or more after ChatGPT came out, because there was more widespread, obvious evidence for it.

Some of it may also come down to things like Brett being a showman. Within the research community, which some VCs might consult with, some engineers or researchers are big fans of people who are very showy.

Well, most VCs were skeptical of SpaceX and what Elon was doing, and of the way he was promoting the progress with Tesla. But he did it, and I don't think that is a reason to believe a company will not do well. Of course, you need to be able to read between the lines and understand what is an embellishment and what is actually real. That does take a lot of work, and there's a lot of nuance involved there.

But I think understanding that is what separates a good investor from a bad investor. Very cool. I want to spend a lot of time now on the robotics industry.

14. Launching Robo Strategy (BOT) on Nasdaq to challenge traditional VC

But before that, I'm really curious how you got from Figure and getting bullish on robotics to deciding to run RoboStrategy as a publicly traded closed-end fund. You've done really well already, so why subject yourself to the stress of being a public-market CEO? Why not just keep running proprietary capital like you were with Mechanism, or even raise a venture fund if you wanted some sort of external capital leverage? What made you want to do this?

I think to do it at the scale that was interesting would have been difficult. Being able to deploy tens of billions, right, and raise that in the private markets means you have to have a track record and an organization as impressive as a16z. They raised, what, $15 billion in their last series of funds. But then you go out there and see Tom Lee raising $18 billion in 8 to 9 months, and it was really just him carrying all that weight himself for an asset that went down 50% and a thesis that wasn't that good. That really opened my eyes to the opportunity in the public capital markets to raise a huge amount of money for investment purposes.

Imagine if you had a thesis that was really great and an asset class that was actually going to go up substantially. Then I would propose that you could potentially raise a lot more—maybe an order of magnitude more. I think that opportunity was really exciting for me: the opportunity to create this new model for venture capital. I think it was a compelling case that you'd be able to essentially achieve the scale of some of the biggest venture funds in the world, or even larger than the biggest venture funds in the world, and do it in a very short period of time using the public capital markets versus the private capital markets.

It's just thinking from first principles: Why haven't people done this before? Is there a good reason besides the fact that it's hard? Well, not really. People have always done things in a certain way, and we're not afraid to think differently.

Another big piece of it was that we were okay with investing pre-seed through Series A and B before, because that's where we felt there was the most asymmetry and you could achieve these 100x returns. Traditionally, at the growth stage, you'd be looking at a 20% to 30% average IRR if you're doing pretty well, and that was just not as interesting to us.

But things changed when ChatGPT came out. AI started taking companies to a whole new level of power law and a new level of scale, so you could start to invest at multibillion-dollar valuations and expect almost a seed-level outcome. You could still expect a 50x, 100x, or 500x return investing at a $2 billion valuation. It's not going to be every company, but if you're investing in the right companies, such as one of the biggest companies in humanoid robotics, I think it was definitely possible.

You could put billions, if not tens of billions, of dollars to work and achieve extremely substantial returns. People did that with OpenAI; they did it in robotics as well. It finally made sense to operate at a larger scale when it hadn't made sense before.

Those are, I would say, the key reasons why we started RoboStrategy. Of course, there's also this problem that we're solving, which is accessing the public capital markets for high-quality robotics and AI investments. I would say that's a secondary goal. Our motivation is to win, establish a new capital-markets model, and become a new challenger to SoftBank. You're going to win by tackling some of the biggest problems, and we've perceived this public-private-market disconnect as being a really big problem.

I love it. It was such a smart play, and I think almost no one really knows about it or realizes what you've done here. I keep telling people in traditional VC, and they're kind of like, “Yeah, that's cool,” but I think there are so many people sitting on these very large, illiquid private positions. You've shown this playbook where you can take an illiquid private position and not only make it liquid, but have it trade at a premium—at a 3x premium to the underlying value, to the underlying NAV. It's kind of insane.

Do you think there are other RoboStrategy-shaped opportunities out there for people to do this? What do you think is the next interesting opportunity that someone could take a private strategy, make public, and see similar success?

15. Public vs. private market asset valuations and premium to NAV

I would say that understanding a premium to NAV and a premium to true value are 2 different concepts. I think it's a common misconception that NAV is equivalent to true value. The way in which a fund marks net asset value is that we look at valuations that have occurred in the past, most likely the prior round of financing. That could have occurred 6 months ago or 18 months ago, and maybe we apply some small discount as well. That is how NAV is marked.

In reality, that is a point in time in the past. Some people could say this is the value of the assets today, and that's how I believe fair value is. But I would argue that if the public markets are valuing these assets at these prices, then that is also a fair value—a fair marking of value.

You have to understand that public markets and private markets value assets differently because they have a different set of market participants. The set of market participants in the private markets is substantially smaller than the set of market participants in the public markets. This isn't something crazy. Public companies go public all the time and then trade at higher valuations than they would in the private markets.

The reason why companies like SpaceX or Anthropic, or whatever company decides to IPO or SPAC, sometimes do so is because they run out of capital in the private markets. If your set of market participants is much larger, it's not a surprise that public-market prices might be higher as well. It can also happen in the reverse direction: Some companies can be worth a lot in the private markets and then not be worth as much in the public markets.

People have to understand this nuance because it occurs in every market, including real estate. If I were looking at a building or a house and saw a transaction for it from a year or 2 ago, I wouldn't necessarily say, “The market value is the value at which it last transacted.” I would have to ascribe a new value to it, do an appraisal, and go out to the market to see what buyers are willing to pay. Sometimes that value is different. If I'm going out to a larger group of buyers, more likely than not, I'm going to find a different price or a higher price.

There are a few companies out there, like TransDigm and Constellation Software, whose whole business model is predicated on acquiring companies in the private markets that trade at a 3x to 8x earnings multiple. In the public markets, they would be valued at something like a 15x to 40x earnings multiple. That's what the parent company is valued at.

They can spend $100 million, and then once the acquisition is complete and it's on their books, those cash flows are worth $400 million. They have this almost immediate $300 million accretion, and that is the whole driver behind how those businesses work. They've been able to scale that to many tens of billions of dollars, and we're applying a very similar concept to venture capital.

It requires that combination of financial engineering and true venture expertise, but I'm curious to see who else does this. I definitely think there's clearly public-market demand for a lot of these assets, but they need to be packaged properly and have the right thesis behind them.

Tom Lee—I don't know if he was the right guy, but he definitely proved you can raise money. Obviously, Michael Saylor or Cathie Wood, with the thesis behind them, are examples of people who could do it. I think there are a few people who could, but let's get into robotics and the market.

I'm curious if we can start with you walking through how you see the robotics landscape. I'll let you take this however you want. Maybe you have a mental model of the stack, or maybe you see it in terms of form factor. I'm curious how you think about the sectors and portfolio construction within robotics, and which sectors you're most interested in.

16. Supply chain bottlenecks & identifying durable moats in hardware

We're most interested in the companies that have the greatest returns.

That could be across general-purpose robotics, application-specific robotics, or within the supply chain as well. But it doesn't mean every single general-purpose robotics company is going to be a big winner, and it doesn't mean every single company within the supply chain is going to be a big winner. There are a lot of components in the supply chain. For example, if you're talking about robots, you have PCB boards, wires, battery components, torque sensors, actuators—there are so many different components there.

If you look at the iPhone, for example, or the GPU market, there are a lot of companies in the supply chain. Even though the iPhone is a massive product that Apple sells and makes billions and billions in revenue from, some of the companies in the supply chain aren't worth very much. They might do really high volumes, but maybe they're operating at 50 bps of margin or 1% margin. That's because what they're doing is kind of commoditized.

At the same time, within the GPU market, you do see some really large supply chain players that have almost as much margin or more than NVIDIA. Those are the cases where they have a real competitive advantage and a way of operating that's unique to them, something that other people can't copy. We look for companies like that, not just companies that are emerging these days saying, "We're going to build actuators—the same actuators as everyone else—but the only difference is we're making them in America."

I don't think that's a true long-term moat. In the short term, it can be, because people prefer products made in America, but once it becomes a big enough market, you're going to have more players come in and the margins are going to be competed down. You really have to have something special, whether it's in your technology—for example, rethinking motors or actuators from first principles and having an IP license or real IP protection over what you're building—or whether you have a really large contract or a unique supply chain advantage with sourcing, like magnets if you're building batteries. There are ways to build really large, scalable businesses within robotics and within the supply chain, but it is very nuanced in the way that we evaluate these companies and their long-term, 5-, 10-, or 15-year potential.

Is there one that you've seen like that that you think has the potential to really benefit from this? Obviously, like you said, some of them are commodities. But even for commodities, if demand inflects sufficiently violently and supply takes a while to come online, you can still have these commodity companies—like memory right now, where at least some of it is a commodity, DRAM.

You're still seeing very violent price movement because it's just hard to bring supply online fast enough to deal with the ramp-up in demand. Are there similar trades that you've seen that you think are interesting in robotics? Obviously, I don't think we're at the point yet where the bottleneck is interesting, because we're just not producing that many of them, but I'm curious if there's anything you've identified like that.

You can make the case that literally every single component within a robot might become a bottleneck. Even for relatively commoditized items like copper wiring or PCB boards, if demand for robots goes from hundreds of thousands per year to hundreds of millions or billions per year, then it's going to cause a huge strain on even those very simple products, just because of the amount of increase we'll need year over year.

But that doesn't mean there will be a lot of long-term value capture. There could be a lot of short-term value capture through margin expansion among the current participants in the ecosystem. Maybe their margins go from 1% to 10%, and they're 10x-ing the profitability of their business. But is that going to last 2 years? Is it going to last 5 years or 10 years? That's important.

17. Form factors: Humanoid vs. wheeled manipulators in industrial & home settings — Vertically integrated full-stack companies vs. model-only approaches

I don't think it's clear that it would last for more than 5 years, and if that's the case, then I don't think that would be an investment that's interesting to make. But if there's a strong case that this is going to be a 10-year-plus competitive advantage or shortage for whatever they're making, then yes, that is very interesting.

Okay. When it comes to form factor, let's go back to that. People have very strong opinions on form factor, and you have positions across the stack. Obviously, your core positions are in humanoids with Figure and Apptronik, but you have exposure to medical microrobotics with Endiatx, soft robotics with Alonic[?], and drones with Purple Rhombus[?]. You have mobile robots, and there are also mobile manipulators, quadrupeds, and others. I'm curious: how do you think about form factor and time for each one? Or just how do you think about it, I guess?

I wouldn't differentiate TAM for humanoid robotics versus cobots or just the general 6-axis industrial arms. There could be differences—maybe one market is 10 times bigger than the other—but they're still so small today compared to what they're going to be in the future that nitpicking that difference just doesn't matter. For something that becomes very niche, like robots that tie your shoes or robots that create pottery art for you, then yes, we do need to think about the differences in TAM for those robots. But generally, I think those are pretty easy to tell apart.

I'll give you one example: welding. I think welding is an underappreciated market, and that's one area where we made an investment in a company creating application-specific robots. That company is called Path Robotics. They build robots for welding, and welding today is something like 80% manual and 20% automated.

That's really surprising, because if you look from a distance, welders are just moving a torch from point A to point B, but it's a lot harder than you think. You have to work with variations in the metal that you're working with—not just the type of metal, but the shapes and sizes of the metal. You also have to work with differences in your environment. Are you doing this on a big ship, or are you working with smaller parts that are fixtured in place?

You have to make sure the weld is going well and evaluate the reasons why it isn't. It's always been difficult to automate unless you're doing this at multi-million-unit levels, which is why welding itself is an underappreciated market. I think the world pays welders something like tens of billions of dollars per year, and there's a shortage of welders as well, even though the job pays decently. We simply cannot find enough welders in the world.

Path has collected hundreds of thousands—maybe more—of welds. I can't say the exact number, but it's a very large number. Collecting welding data isn't easy. It's not like people typing on their keyboards or videos of people folding laundry; to capture welding data, you actually have to capture not just the successes but the failures.

The failures are the key to getting these robots to be really performant. That's one of the key pieces of robotics data collection. What does getting failed robotic welds mean practically? It means that you have to scrap the pieces of metal or whatever you were working on, and that's really expensive. If I have to work with 2 sheets of metal and, because it didn't go well, I have to throw them away, that's a significant cost.

Then you have to work with the shipbuilders, the plane manufacturers, or whoever else, get into their actual process, and see how they work with that data. It's constrained in those silos and isn't open to the public. That's an example of a business that can be both very large and focused on a specific application.

We've covered this a little bit, but I want to get the sound bite and give you a chance to actually respond to the form-factor thing, because I do think people obsess about this a little bit—the humanoid haters. I want to put the bear case to you, which is the strongest bear case that I've heard, and I want to hear your answer to it.

What people will say is that humanoids are overkill for 95% of industrial tasks. Anything you'd point a humanoid at can already be done by a mobile manipulator or a specialized robot that's cheaper, easier to integrate, and already 99.9% reliable. On safety, an 80-kilogram humanoid robot that falls is throwing 100 to 150 kilograms of force, which is a non-starter on a factory floor.

And the deeper claim here is just that the humanoid people say that the world is built for humans, so we need human-shaped robots. But that's backwards, actually, because industrial environments have always adapted to the machine: floor layouts, path markings, fixturing. Where does that break? And where is the task that genuinely needs a humanoid and not a wheeled torso with two arms?

But you know what's funny is that the goalposts are changing for the critics, because a wheeled mobile manipulator like Mobile ALOHA is essentially a humanoid with wheels instead of legs. Before, maybe they wouldn't have classified that, right? That's a general-purpose robot. That's not a robot that's made to do a specific thing, right? It's not a robot that's made to put bottle caps on bottles.

Yeah.

That's a robot that's meant to do literally hundreds of different tasks. And so that's an example of a general-purpose robot. It's an example of a humanoid, in my opinion. Right now, it is a little more difficult to program or utilize a humanoid with legs, and it's a little more expensive than one with wheels, but that is going to change over time, especially when these things get into really high quantities and the improvements around AI get better. The control systems get better for utilizing legs, and that's going to happen.

I would agree to the extent that, for industrial environments, it probably does make sense to use humanoids, mobile manipulators—whatever you want to call them—with wheels instead of legs, because it's easier to balance if you're carrying heavy items. But anywhere outside of the industrial environment, like the home, which is going to be an even larger market or just as big as industrial use cases, something with legs definitely makes more sense. You need your robots to either go up and down the stairs in your house or, if you're in an apartment, at least leave your apartment and go out and do tasks for you, or take out the trash, et cetera. So I think there's no argument against having a humanoid within your home with legs. That's definitely the best form factor.

Nice. I like that answer. Let's move on to robotics foundation models. This seems like a really interesting field. Obviously, there have been billions of dollars poured into it, and as I understand it, there are a bunch of competing approaches. There are companies like Physical Intelligence building these hardware-agnostic VLA models—you know, what we talked about before, basically LLMs that output motor commands. There are companies like Skild AI, which argue that the VLA approach lacks physical common sense and instead train on these massive physics simulations. And then there are full-stack companies like Figure, betting that the real value is a data flywheel from having your own robots deployed in the real world.

You've invested in Dyna Robotics too, which is kind of your bet on this layer. I'm curious how you think about the different approaches and which one you think wins, or whether there is space for multiple. Would you say the right analogy is that Physical Intelligence and Skild are kind of like OpenAI—the model captures value—and then Figure is sort of the Apple, the full-stack approach? I'm curious how you see this whole sector.

Yeah, we've concentrated a lot of our bets in companies that are building from a vertically integrated perspective. They're building the intelligence, the hardware, and the software, and they're also managing their own manufacturing or working with partners that do. I think that's interesting from an investment perspective solely because you have more shots on goal, right? If the model layer becomes monetized, then at least you have the hardware layer to differentiate on, or the manufacturing layer to differentiate on. And so risk-reward is maybe the key principle that everything folds up into.

Then you think about it from a more practical perspective: if I'm developing intelligence, I need to collect a lot of data. Yes, I can use egocentric video, and then you get the question: What kind of data do I need to collect? Of course, egocentric video data—more like internet video data—is important, and simulation data is important, but what is absolutely necessary is robot rollout data as well.

If I were learning how to play basketball in a 7-foot-tall body and then immediately transformed into a 5-foot-10 basketball player, I'd probably be all wobbly and uncoordinated because I was trained using data from a different embodiment or form factor. When you have embodiment-specific data that your robot or model is trained on, then it's obviously going to run better on your own hardware. That's one important piece.

And if I need a lot of that data, then I need a lot of robots to collect that data. If I'm not building my own hardware, then I have to rely on these third-party vendors and/or manufacturers. Can they make all the robots I need? Are they going to give me the customer support I need? These robots that you buy from third-party vendors break all the time, right? Maybe they have overheating issues. Maybe their actuators have quality issues and they'll break from time to time. Maybe you have software bugs. You're almost reliant on these outside parties that may make it more difficult for you to develop the best model that you can. Maybe they have supply chain shortages and they just can't deliver as much as you need.

When I think about simulation as well, right now simulation is mostly used to train locomotion policies, which are the policies—the models—that help robots walk around and balance well. That type of data is best collected using robots that you have a high degree of confidence in, in terms of what joint torques your actuators are producing, which are not entirely transparent or could be highly variable if you're using a third-party vendor. You have to have all these details down. I think you have to co-develop both your model and your hardware to get the best results.

And that's what we're seeing, right? Some of the most impressive results are coming from companies like Figure, which is vertically integrated, and Path Robotics, et cetera. When you look at some of the companies that are model-only, they're actually now starting to try to develop their own hardware. I don't know if it's going to be successful or not. Maybe it will, but it requires a whole different level of expertise and knowledge, and there's going to be a lot of time spent iterating, because you can't simulate a lot of what you need to do in hardware. You have a lot of bottlenecks, and things take a lot longer than they do for software or model development. The companies that have already put that time in are going to be at an advantage.

Interesting. So do you think there's a Mercor- or Scale AI-shaped opportunity in robotics? Because it seems like maybe the answer is no: you actually need to be vertically integrated and generate the data yourself. You don't think this egocentric data is kind of a meme, or this world-model approach is too far out, and you actually need to do something like have robust, live-in-production data, or teleoperation, or something like this, or a human in the loop?

Do you think there's going to be—because a lot of investors have been pattern-matching to that, right? It's like, I want to bet on robotics. There's this data layer that's kind of like an infrastructure layer. There were these Scale AI and Mercor outcomes, and I want to bet on that. Do you think there are interesting outcomes there, or do you think it's going to be more the vertically integrated companies—the hardware companies—basically generating this data themselves?

18. Why egocentric video data collection is a short-term cash play

You know what's interesting is this is almost like a meme within the robotics research community. There's the Buzz Lightyear meme where you see one Buzz Lightyear, and then it zooms out in the background and there's 1,000 of them. There are just 1,000 of these companies trying to be the Scale AI of robotics.

Then you talk to them and try to ask them: How big do you think the market is? What types of data do you think are going to be needed? How might the space evolve over time? What happens if the robots become really great at continual learning and self-play becomes a dominant means of training these robots? They don't really have good answers to this, and I don't think they've thought it through clearly themselves.

It feels like an easy path to go down, right—collecting robot data that people can pattern-match to. I think in the short term there will be a lot of cash generated by some of these businesses. It’s still an unanswered question how many hours of egocentric video data—which is the primary means of data these companies are collecting—are needed for these models.

Is it in the millions? Is it in the tens of millions of hours? Is it in the billions of hours? It’s unclear. One, it’s unclear, and, two, I think it’s unclear what happens when the models become almost like humans and they can learn by doing themselves—as in, the robots don’t need an extraneous source of data to communicate better and better. And I think that is the end goal of all these model companies, right? It is—

For the robots not to be taught, but to be able to learn themselves—to be able to self-improve continuously. That’s why I think this is a short-term cash cow: you might have a short-term spike in revenues and valuations, and then it starts plummeting maybe 2 years, 3 years, 4 years, 5 years down the line.

That’s not the type of business that I think we’re interested in. Some of these companies could become neo-integrators, which I think is maybe interesting, but we’re too far away from that to tell. Their business model becomes, instead of collecting data, deploying robots into different companies or enterprises.

Yeah, you could say the same thing about Mercor and Scale. I guess it’s unclear how long it’ll last before the models can teach themselves, or before the pre-training paradigm—you’re generating synthetic data or whatever—but they still had massive outcomes. So I guess it’s interesting. There might be some big, at least short-term, outcomes in robotics, and the long term is TBD.

Yeah, I mean, the question is: are you going to get liquidity? I think for Scale, the acquisition was more around talent and know-how than it was around the business itself. I haven’t diligenced Mercor extensively myself, so I can’t speak to that company.

Mercor is a crazy one. I think they’re doing a couple billion dollars of ARR at this point, so it’s kind of a crazy one. Every round has looked expensive, and then they’ve just grown an insane amount—just accelerated at an insane clip.

Yeah, and then you might have—I guess the one risk I see is you have these valuations explode. Is there going to be liquidity at the peak, or when you want to sell? Or is it just going to be a market where there’s nobody else buying, and then the revenues plateau and the growth stalls, et cetera?

I’m not saying that’s going to happen, but that is a risk that I see. That’s not our interest. You can play that game. It almost feels like trading as opposed to investing. I want to invest in businesses that are durable over decades.

19. U.S. vs. China: Manufacturing scale, software margins, and national security — First-principles investing and risk/reward in outlier companies

For sure. So, for the last question on robotics, I want to move to investing philosophy in general and then finish up. One thing we haven’t touched on, which is a natural thing to talk about when we talk about physical systems and robotics, is China.

The consensus narrative seems to be that the U.S. leads AI, but China leads robotics. You spent a bunch of time in China touring Chinese robotics companies, and you came back with a very different view on this: China is actually behind on the software piece. There are only 1 or 2 Chinese companies that are close to solving the software piece. They might be very far ahead on the hardware, especially on the scale manufacturing of the hardware, but they’re actually behind on the software piece.

I’m curious because there’s been a lot of progress in the Chinese robotics companies. There’s AgiBot and Galbot and all these companies. I think there are around 100 Chinese humanoid robotics unicorns. Have you updated your views on this, or where do you sit now? What do you think the endgame of the U.S. versus China in robotics actually looks like?

Yeah, I’ve updated my views a little because I’ve been impressed by the research that’s come out of Robbyant, which is a subsidiary of Ant Group. They’ve released open-source models that are kind of at the frontier of robot foundation models. They were one of the first groups that released a robot foundation model based on a video-model backbone, which is kind of where the space has transitioned to and what is understood to be the future of robot foundation models, instead of having a VLM backbone.

I think there is definitely some great AI research talent there, but I don’t agree with the framing that one country is going to win versus the other. I think they’re both going to develop great technologies and robotics companies. It’s more a question of what is interesting from an investor point of view.

I think this is what you see with not just robotics, but also cars, EVs, and cell phones, right? Both countries can make great products. Sometimes the Chinese companies can make even more impressive products. But what are the biggest companies in the world? They’re the American companies—the Apples and Teslas of the world.

One reason is that the American capital markets are stronger. The second is that the margins are a lot higher. They can sell products that generate 20% to 50%-plus margins, so you can sell 10 times less product but be a bigger business because you’re not operating to the bone like some of these Chinese companies are.

There are so many great companies, and that’s great for society. It’s great as a consumer, and great technology becomes very affordable. But it’s not great from a shareholder perspective. There are still going to be good outcomes; it’s just not going to be the multitrillion-dollar or multihundred-billion-dollar outcomes that we’re looking for as investors.

Another piece I kind of like—there are so many paths to this as well—is that there are a few bills going through Congress right now to help legislators restrict Chinese robots within America. I think there are 2 components to this. One is the national security issue. There are a lot of people who are concerned that Chinese robots that can see and hear everything could be within American companies and homes. The other is the idea that, of course, you want to promote the domestic industry.

This is getting a lot of bipartisan support, and I think it’s a very high probability that a bill like this passes. I don’t see American companies having to compete with Chinese companies in the biggest market in the world, which is America. Maybe they will in other countries around the world, but at least they’ll have America, and that’s usually 30%, 50%, or 60% of any major market or product.

I’m not too concerned about who’s going to win. I think both are going to do great, and America is the place you want to invest.

Have you invested in any, or are you looking at any Chinese companies?

We’ve looked at a lot of them. Galvi AI.

You passed for investment reasons, or was it impossible for you to do it given that you’re publicly listed? You know—

Look, I think we could have made some argument, and they were open to exploring ways for us to make an investment, but it just wasn’t compelling enough to offset the risk of being a publicly traded, U.S.-based company and investing in technology that can be considered dual-use.

Okay, dope. Let’s move on to investing philosophy. This has been really interesting. I’m enjoying it.

I have a question about being contrarian in venture. It’s one I’ve been asking most of the managers I interview. You told me once that your edge in deep tech is the same as in crypto, which is just your ability to think about things from first principles because most people are herd investors, and all your biggest wins are from things that weren’t cool at the time.

There’s a bit of tension here, right? To find great deals, you need to see something others don’t or pick from a pool others are overlooking, which is also the easiest way to get adversely selected. How important is it for you to actually be contrarian? And how do you distinguish contrarian and right from contrarian and wrong?

I don’t really think about contrarian versus not contrarian. If you are contrarian, sometimes you might be thinking about pricing risk and reward the right way, and the risk and reward can be really mispriced. That makes those spots really attractive. But you can also be non-contrarian and invest in really big companies, and the risk-reward can still be really mispriced.

I really like that data analysis from Coatue where they looked at the probability of public companies doing 10x. You would think people always want to—they’re always like, “Hey, the valuation is too expensive.”

I want to invest in the smaller company because I think there’s more upside. But then they showed with data that it’s actually the bigger companies that have a higher probability of doing a 10x than the smaller companies. And you think about, well, why is that? It’s because there’s a reason why these companies are bigger in the first place, right? They have a lot more competitive advantages.

They’ve been able to make the right decisions, and they’re playing in bigger markets. They have founders who are maybe instilling the right culture and working with urgency, and so on. I’m of the mindset that winners win. Within robotics, I guess the field itself was kind of contrarian, but typically we’re investing in the strongest founders and teams. Sometimes those might be at higher valuations than some of the other companies that are doing similar things, but we think that’s the right bet a lot of the time. But sometimes, you know, what’s in our lane as well is misunderstood. So it’s all about pricing risk and reward the right way.

I really like that. I think the risk and reward is very simple, but I think it’s a great framework to think about investments. And I’m curious on the founders’ side, because I think that’s a lot of the mistakes that crypto investors have made transitioning over to non-crypto.

In crypto, you could often get away with investing in not the best founders in certain spots because there was just a rising tide that was lifting all boats, or there was some sort of nepotism with where they were positioned in the chain or something like this. And there’s also liquidity, right? So these mistakes are just less punishing. You’re not riding or dying with this person for 10 years. I’m curious how you’ve adjusted your founder lens and what you’re looking for in the founders you’re backing in these robotics companies.

I think investing early-stage sometimes, or in fields like crypto, almost instills really bad habits because you almost get rewarded for doing that sometimes. That same kind of outcome is not going to happen in venture. For you to get big wins, for companies to go public, you’ve got to be investing in the strongest founders who are actually creating real products that are going to have a lot of revenue and adoption in the real world.

If you look, for example, 3 years ago, when AI was becoming understood to be a very big market, you could have invested in Anthropic or OpenAI at valuations of tens of billions of dollars. Or you could have said, “Hey, I want something that is grittier and has a cheaper valuation. I’m going to invest in Character.AI or Inflection AI,” which, by the way, most people probably don’t even remember or know. It’s just because those companies have failed and they’re not relevant, or they got acquired for not-great outcomes for investors.

20. Evaluating founders: Reference checks, urgency, and execution speed

I think a lot of people are at risk of that playing out as well for robotics, where they’re just trying to invest and catch up. How do you know someone’s a strong founder? I assume a lot of them might look similar on paper. They’ve gone to great universities, maybe they’re great researchers, maybe they worked at a big company. What are you looking for to differentiate great from good?

What we care about is whether a founder can build a really great company that can develop a really great product. How do you do that? Well, I need really great people. I need to make really great decisions. And then how do you do that? Well, I need to recruit and inspire really great people.

How do I know a founder is going to do that? One is, you can see whether they’ve done that already. What is the quality of the team that they’ve assembled? Are they really great? That sounds a little bit easier than it actually is, because there could be some people who look really great on paper but aren’t actually great operators or executors, or people that you would want on the team who are getting stuff done or pushing the space forward.

To be able to evaluate that, you need a network, and you need to be able to reference-check talented executives—not just the executives, but also the people who are at the engineering level or who are doing a lot of the day-to-day work. That’s reflective of the general quality of the people at the company and the outcomes that they’re going to drive.

If we look at the example of Figure, who did they recruit? It was people from really great robotics companies like Boston Dynamics, Google DeepMind, and Tesla Optimus. But we looked into them, and they were also just incredible people at the organizations they used to work at. For example, one of the heads of AI at Figure was on some of the papers for the initial robot foundation models at MIT.

There are examples all the time of us doing reference checks and finding out, “Oh, wow, this person worked at a brand-name company before,” but there were all these issues working with them, and they weren’t delivering what they were going to deliver. There’s a lot of diligence involved—not just in evaluating talent, but also in evaluating the contracts that these companies say they have and the product quality that they say they have.

They might be telling their prospective investors that they have a great product and all these sales coming down the line, but you have to dig in. You have to call up the customers—not just the ones that they provide, but the ones that you’re able to find through your network yourself. Yeah. Is that real? What are the issues? How do they compare to the other products on the market? What’s the pace at which they’re developing?

People say this all the time, but a sense of urgency really is important, because that could make the difference between developing at a 5-times-faster pace than somebody else.

Yes, it sounds like you’re doing a lot of diligence. I’m curious—one of the questions I had was about process. I’m really curious to ask investors about their process, because you have some who are very gut-driven, and others who are doing much more extensive diligence: reference checks, checking contracts, and so on. Maybe they need to write about it, write a memo, or talk to people about it. What’s your process like? Do you ever have a gut instinct and commit on the first call, or are you always doing a bunch of work and thinking around it before you’re in?

21. Testing physical models vs. shortcutting research

I think it’s very rare to make a decision on gut instinct. I don’t know if we’ve done anything except maybe some very small checks, and for everything larger, you have to do the diligence. We have a fiduciary responsibility to do so.

Unless it’s something like Prometheus—Jeff Bezos starting a new company, okay; Elon Musk starting a new company, okay. These guys have enough credibility, I think, to be able to shortcut some of the diligence that you need. But for most other people, it’s not the case.

It’s going to look different from company to company. If a company is developing a robot foundation model, and that’s one of the reasons why we’re really excited to invest, we have to go there and see the model for ourselves. Not just look at a video that they made, but mess around with it: move the objects around, move the table height, put it in a different place, maybe have it take a different path, or whatever. Just test the actual capabilities and performance of the models in real life.

Maybe someone is creating a completely new type of actuator, which is something we’ve invested in as well—redesigned actuators for humanoid robots. We have Scott on our team, who I consider a foremost expert on actuators. He’s been in the robotics space for 40 years, and he’s built and sold 2 robotics companies. We trust him a lot.

We also brought that company to one of the leading actuator manufacturers in the world, which is working with a lot of humanoid companies, and we asked for their opinion. They were very excited about what they were building. Then we went to humanoid companies as well and got their opinions. It’s a lot of expert validation if a company is too early for the product itself or the model itself to be tested.

That’s awesome.

All the claims that a company is making need to be validated, I think. But that doesn’t mean we’re taking 6 months to do the diligence. It’s a question of prioritization as well—figuring out the most important ways to understand whether this is a great company and whether they’re doing what they say they’re doing.

You can get really in the weeds on the research approach, which sometimes makes sense, but sometimes you can shortcut a lot of the diligence by just testing the model itself, for example.

Very interesting. Yeah. Cool. Because it definitely changes when you’re doing later-stage rounds. I think more diligence is required than at seed and pre-seed, where you’re basically betting on a founder and a plan, but that makes a lot of sense.

22. Portfolio construction and position sizing for Robo Strategy

One last question on investment concentration. I think historically this has been really important to your returns and ours, right? I think it would have been hard to get where we are without being heavily concentrated in the past. You know, on PA, you’ve run pretty concentrated. I’m curious how you think about position sizing and concentration for Robo Strategy, especially now that you’re running public money. How do you think about how far you let that run when you diversify? Also, what’s your starting position sizing versus how big you let it run before trimming?

I’m generally of the mindset of, “Let winners win. Let winners ride,” right? Because there’s a reason why they’re winning. At the same time, if a company becomes 80% of the portfolio, I think we might have some regulatory restrictions requiring us to reduce our position. Or, if a company goes public, there may be some considerations there. The value proposition of our fund is to be mostly private, but we can have some public exposure. It just depends on the percentage.

Overall, our portfolio is going to look like a barbell, where 6 to 10 companies are going to comprise around 70% of the portfolio, and then maybe a smaller 20–30% is going to be earlier-stage, pre-seed to Series A and Series B bets. The rationale behind that is I think you’re going to have something like an Anthropic- or OpenAI-type outcome play out, where there’s going to be a lot of concentration of value capture. I don’t think it’s going to be as concentrated as what happened with some of the model companies, but there’s going to be concentration, and you want to own as much of those companies as possible.

Especially if you’re going to start scaling some size, we just have to because there aren’t many other great places to put a lot of capital—billions to tens of billions of dollars. At the same time, there are going to be some breakout winners in the future that are just earlier in their life cycle right now. Maybe they’re not at the growth stage yet, but we want to identify them early.

So, 1, we can track execution over time, and how well you’ve done in the past is a great predictor of future success or future execution. And then, 2, we need to build a relationship with those founders early, so that when they start getting real traction, we’ve shown that we can be a great partner for them and have as much allocation as we need in the follow-on rounds. That’s how it’ll look: a very multistage approach.

Okay, last question, and then I’m going to let you go. We’ve been talking for an hour and a half. What’s surprised you the most, and what have you gotten wrong so far in robotics? Are there any investments or assumptions that didn’t play out the way you expected—something you’ve really changed your mind on?

I would say I’ve gotten a little more bearish on model-only companies. My view was not extremely bullish before, but it was more neutral. I’ve just gotten a lot more bullish on open source, and I think open-source models are going to start to monetize these closed-source models. You’re seeing that happen with LLMs, with Kimi getting as good as Claude on some benchmarks. I think that’s going to happen in robotics as well, so that’s one area where we were wrong.

That’s bullish China. No, if software is commoditized and they have the cheapest hardware—

Well, not if Chinese robots are restricted in America. That’s bullish for hardware companies, which we’re invested in, and companies that can do their own manufacturing and have secured great supply-chain partners.

There are maybe 1 or 2 companies that we didn’t have on our radar because they were in stealth, and we’re now interested in building positions in them, though we haven’t yet.

Names, sir? What are the—what—

That will be revealed once we’ve built positions in those companies.

It’s really just 1 company.

Yeah. I think we’ve had a very exhaustive process on the space, making sure we see everything that we think is interesting to see and evaluating all of it. We have an incredible pipeline that we’re planning on executing on—making follow-on investments or executing investments that we negotiated before—and we’re excited to keep building on that.

Awesome, man. This has been great. I’ve really enjoyed it. You’re absolutely crushing it with Robo Strategy, both on the investment side and just the structure of Robo Strategy itself. I think it’s going to be something that people mimic.

I hope this podcast helps some. I’m going to send it to a bunch of the people I know in venture to hopefully inspire more people to go this route and expand public-market access to private markets. I do think it’s one of the biggest problems that we need to solve to get people to not revolt against AI and robotics and the revolution that’s coming. Thanks so much for the time, for all the insights you shared, and I’m excited to have you back soon once we have more stuff to share.

Yeah, just some self-promotion: if you're going to send it to other venture people, if they want to work with a company that's already doing something like this, we are hiring for exceptional talent. So, feel free to reach out if you're looking for a role at a company like ours.

Amazing. Thanks so much, and thanks everyone for tuning in. See you again in 2 weeks for another Emerging Managers. Bye.