物理AI领域的 Nvidia：Applied Intuition 150亿美元业务内幕

When you look back 25 years from now, I think physical AI companies are going to be the ones that really dominate everyone's mind. Applied Intuition is a physical AI company. We take intelligence—we take AI—and put it on machines to make those machines smarter. Our mission is to make 1 billion machines intelligent.

Dana is our new agentic platform for physical AI. We're drastically reducing the barrier to entry for building physical AI and deploying it in the real world.

So anybody out there—starting with engineers, but ultimately really anybody—can develop robots. This really is the culmination of pretty much everything we've worked on over the last decade.

I know a lot of the story we're going to tell today is going to be about a single business, Applied Intuition, but it's also really a story of the physical AI market and how far autonomous technology has come. You two see this across as many industries as about anyone. Maybe just describe the state of the physical AI market, how the whole landscape feels to you now in 2026, and maybe some of the important key landmarks on the timeline since you started the company in 2017.

In Applied Intuition's case, our mission is to make 1 billion machines intelligent. One simple example you could think of is self-driving cars. Those are intelligent machines, but that's 1 example; Instagram is an app on the phone, and there are many other apps like that. Physical AI is this intersection of AI and hardware, typically in the real world.

Humanoids fall into this category as well. The particular technical challenges of physical AI are quite different from digital AI, like LLMs and information retrieval systems—chatbots and stuff like that. You have the constraints and safety criticality of the real world. Often, when we're talking about moving machines, they're moving in time and space with humans, and suddenly that becomes something where you really have to think about safety.

The real-time nature of the problem is another challenge. If you ask a chatbot, “Tell me about Peter Ludwig,” it can take 20 seconds to process that information. But when you're flying down the highway, or if a humanoid is making a decision, there are very hard real-time constraints. Probably an underreported aspect of physical AI is the dollars involved. You need to put this on machines and silicon that are affordable within the use case you're talking about.

You don't just throw endless compute at processing something. You have to do it within a compute envelope—not only a time envelope, but also a cost envelope.

I'd just say it's useful to also think about the separation of digital AI and physical AI. Digital AI typically means what you're using on your desktop or mobile phone, where there's a screen showing you the result of the AI. Where that crosses into physical AI is when anything in the real world is actually moving. I think this is where the real impact on the economy will happen, when you talk about all of the industries that fundamentally have moving things.

Think about anything from industrial companies and manufacturing to use cases in health care and energy. There are so many different fields where, in order to get the benefits of AI, you actually have to impact these physical systems.

Just to add on to that, by putting self-driving and intelligence on machines, we're really making some of the worst jobs on the planet easier. In digital AI, there's a lot of teeth-gnashing and hand-wringing about what's going to happen to accountants and maybe even podcast hosts.

[laughter]

In the physical AI world, it's very, very different. The AI can't get there fast enough. If you look at farming, the average American farmer is 58 years old. We have record shortages in long-haul trucking. Mining, as an example, is 1% of the world's workforce but accounts for 8% of work-related fatalities.

When you peel that back, why do people not want to work in mines? These are difficult jobs, and people are choosing not to be away from their families in unsafe circumstances. I think the impact can be very, very quick and very, very significant in areas where there's a lot of demand. That's 1 of the important things to keep in mind: physical AI is very different from, frankly, digital AI.

I think when we look back 25 years from now at this particular phase of business and technology, we're going to see that we're going through a lot of changes very, very quickly. My hunch is that I don't think we're going to only be talking about code-completion products. Those are important, and you can see how big an impact that use case is making as a proxy for how big an impact AI can have on society.

I think when you look back 25 years from now, physical AI companies are going to be the ones that really dominate everyone's mind.

And, Qasar, I know you have this whole notion that the physical AI market is going to be orders of magnitude larger than the digital AI market. Unpack why you think that, and just help us appreciate the size of this market.

When you look at industrials as a category, a huge category of the economy, it's roughly 5% of GDP. Automotive is the largest of the industrials, at 3% of global GDP, which is an astronomically high number. When automotive—for definition's sake, personally owned passenger vehicles—becomes intelligent, the impact on all the people who interact with cars every day, which is essentially everybody on the planet, is really, really big.

If you're sitting in an airport and look around the gate, how many people have interacted with a car versus how many people wrote software that day? Then you go into other verticals. Applied Intuition plays in all of these verticals. As we said in the introduction, we are the only company on the planet that does this, including in the Chinese ecosystem and elsewhere.

The other verticals we're in are commercial trucking, defense, construction, mining, agriculture, and robotics. Each of those verticals is very, very large in itself, whichever way you want to define it: pure GDP numbers, growth numbers, or the number of people employed in those sectors. It's true in almost every measurable way.

Sometimes the markets get so big that they're almost hard for people to grok and put their heads around. But you can just use robo-taxis as 1 instantiation of self-driving cars. Waymo is being valued at $126 billion by fairly sophisticated investors. They're not valuing Waymo at $126 billion because they just want a high valuation. It's because the impact can be very, very large, and that's 1 part of 1 of those markets.

Before we go any deeper, Applied Intuition, in my sense, almost has this Palantir mystique around it, where it's actually very hard to describe what it is and what you do. Maybe just orient us and literally describe as simply as you can what you guys actually build and sell.

Applied Intuition is a physical AI company. We take intelligence—we take AI—and put it on machines to make those machines smarter. Whether that's having them drive themselves or enabling you to interact with them and have intelligent interactions, just like you would with your phone, but in the context of the real world. There are a lot more sensors and a lot more information, and then we do all the things you would think you need to do in order to get there.

We develop our own models, train and deploy those models, evaluate those models, and make sure that they're safe. That's 1 half of the company: putting models on machines. That's much more complex than even what I just said because the machines have a huge amount of diversity. When you're putting something on a phone or laptop as a developer, you have great operating systems that abstract hardware away from software.

If you're writing software for a combine, that's very different from writing software for a car, which is very different from writing software for a humanoid. We've done a lot of that hard work over nearly 10 years, abstracting all these different types of hardware away from the software. Peter was 1 of the early engineers on Android Automotive. Android is basically known for this: it runs on thousands of different devices, and it's the same operating system. We do that, except we're doing it with the intelligence on top.

The other half of the company is all the off-board software—the development tools that you would use to develop that intelligence. We're a B2B company in the sense that we're an enterprise company. We sell to other enterprises, and enterprises meet us in 1 of those 2 ways, sometimes both ways. They're buying the development environment from us so they can make their own intelligent machines, or they just buy the models and put them directly on the machine.

So it’s a technology provider. In terms of Palantir, when the company was very young, it was funny: people would say, “Oh, you guys are kind of like Palantir for automotive and trucking, or Palantir for defense,” which is weird because Palantir doesn't do defense. It’s not exactly a one-to-one comparison, but we’re different in the sense that we really are kind of 2 big product areas.

We do have some forward-deployed engineers, or what Palantir calls them—we call them something else—but the vast majority of our workers are in a product company. In that way, we’re much more like a silicon company. We’re a technology provider. It’s almost like when you think about chips: they go into all these different machines and can do all these different things, but they’re kind of a platform, and we’re kind of like that, except our platform isn’t silicon. It’s intelligence.

Imagine that you run a company that makes some kind of machine. Again, this maybe is in transportation, maybe it’s something in robotics or something in healthcare. You want to make this machine intelligent, right? The machine, let’s say, has sensors and actuators, and you want it to do something intelligently. How do you actually do that?

If you want to develop the technology yourself, you’re going to need a really strong tooling platform to actually do that development. And so, at Applied Intuition, we make and sell that tooling platform, which can be used by engineers at the company that’s building that machine. Or maybe, as the maker of the machine, you actually want to purchase more of a complete solution that can make the machine intelligent almost out of the box.

We also make more of those complete solutions, which we then sell to those companies as well. And so we have that spectrum from tools to solution. And then we license this technology out to the industry.

I was in San Francisco last week, and basically every other car now is a Waymo or some other autonomous vehicle. It’s kind of cool to see the explosive nature of that technology. But there are also a bunch of other technologies, whether it’s drones, humanoids, robotics, mining technology, farming technology, et cetera, that, if we had them, would be amazing, and we could immediately see how valuable the potential would be.

But it’s very hard to actually predict how long the timelines for these things will be. And you’ve been able to develop tools across a bunch of these different technologies. I’m curious how you were able to stay flexible and build tools and systems for these technologies, which are hard to predict.

Before Applied Intuition, I was the COO at Y Combinator. It was in the era when Sam Altman was the president and I was the COO; I ran the firm. It was the era when OpenAI was created, and when we funded Cruise, Scale AI, and a bunch of other great companies that are in the space now.

I give that context because the most important thing, especially for founders who are listening, is timing. Timing is everything. If you build a technology that’s maybe 2 years too early, the market isn’t ready to consume it, and you burn a lot of money waiting for the market to mature, which, by the way, I think is the default failure case. Most companies fail because they’re too early. Rarely do they fail because they’re too late.

The opposite, though, is that you can also be too late, where it’s just very competitive, there are lots of players, and margins in the market are kind of being destroyed. I’m an engineer, but I also did a graduate degree at HBS, an MBA, and that aspect of market dynamics is sometimes underemphasized as well.

So, with that context, how did we navigate nearly a decade ago? Peter and I went about making this company in an extremely intentional way. We didn’t just, I would say, guess our way into a part of it because we’re old. We’ve done companies before and had led large engineering teams at Google and other places. We have technical experience, but the other part is also just being very, very intentional about putting these constraints on.

Initially, what we envisioned was—the first time we talked about working at a company together, even before I was at YC, when we were working together at Google—“Hey, we should maybe do a robotaxi company.” We concluded at that time, in the early teens, that the technology hadn’t really been figured out, which meant we were going to be too early. We were going to spend a lot of money waiting for the technology to manifest itself into a production product.

Secondly, the business model hadn’t been figured out. How is the robotaxi going to be a really profitable venture? I ended up at Y Combinator; Peter stayed at Google. Then, fast-forward, we funded Cruise at Y Combinator, and then Cruise was bought by General Motors.

I went to undergrad at the General Motors Institute, I worked at General Motors, and Peter’s father and grandfather worked at General Motors. We’re both Detroit guys; our family roots are very deeply in the automotive industry. So, in 2016, when Cruise was acquired—it was acquired by none other than General Motors—we started talking again about what was happening in this industry.

This industry at the time specifically was automotive. But where automotive goes, honestly, that’s where defense goes, and that’s where construction or mining goes, and that’s where agriculture goes, because the ways that you build a haul system—if you’re Caterpillar or Komatsu or a combine like John Deere—it’s actually kind of like a cousin product to a car. Or if you’re General Dynamics and you’re building an infantry squad vehicle, a troop mover, or something like this.

When we said, “Okay, well, where is this industry going?”—this industry being automotive—we were like, “Well, there’s going to be kind of like the Tesla-fication of this industry.” These machines are going to get smart, they’re going to be software-first, and then you’re going to have all these tools that are going to actually enable that to happen, from fleet management to updating software to literally testing the software to make sure it’s dependable.

Again, from first principles, just enumerating this for people who are going to start companies themselves, we thought, “Okay, well, if we make software right now and try to sell it to the manufacturers, they’re not going to consume it from us, the little young company, because they’re safety-critical systems. You need a lot more track record and heft, and they’re frankly very complex systems. A team of 50 people cannot build an autonomous vehicle. There are too many subcomponents and complexities.”

And so we started with tools. Today, one thing we’re here to talk about really is our biggest product launch in that fundamental category of the company, which is a product called Dana, an agentic platform in order to do everything we’ve been doing for the last almost 10 years, but in a much more AI-first way. But that’s kind of how we started.

Since you brought up Dana, maybe it’s helpful context for everyone if we trace the evolution of the business. You mentioned that you started with tools instead of the vertical-integrator route of building autonomous vehicles. So, you started with tools, you built the OS, now you have the autonomy stack, and now you have Dana. Peter, maybe it’s helpful for you to walk us through the history of that evolution and how it all fits together.

Yeah. Firstly, I would say something that we knew when we started almost 10 years ago was just that the technology was still going to change a lot, right? There’s a lot of advanced engineering and research work that goes into this entire field. And so a way that you can be part of that but not be, let’s say, overly exposed to any specific implementation is to think more horizontally.

For us, that meant initially really focusing on tools and then building tools in such a way that we could continue adding on to the platform, while recognizing that the technology itself, when we talk about physical AI and advanced autonomous systems, changes constantly. Almost every 2 years, there’s some sort of breakthrough that changes how you have to think about these things.

And I think if you're not dynamic enough to understand how to adapt that latest technique, that latest breakthrough, then you can become almost obsolete in that sense. That's been baked into Applied DNA. It's almost like this internal disruption that we have to do to ourselves to make sure that we stay on top of things.

Tools were a great way of doing that initially and of doing that horizontally across all these industries. But what happened after a few years of working on tools is that you actually hit a point when deploying this technology onto machines, and the problem is much larger than just the tools. You have to think: What are the bottlenecks? What are the rate-limiting factors?

Again, with this North Star of wanting to have a big impact, we want to make 1 billion machines autonomous. We want to do this safely and efficiently. And so you hit this point where, all of a sudden, the operating system actually becomes our bottleneck: deploying this technology onto the vehicle itself.

That sort of forced us into that business, and we had to build a really good solution that allows you to deploy software onto machines, update that software reliably, and have all the right diagnostics. Running advanced models—I mean neural networks—on machines is extremely complicated. I think sometimes it gets trivialized, and people think, “Oh, it’s just about the model.” But there are about 1,000 different problems you have to solve to make this all work, and the operating system piece is a really big part of that.

And so we had to solve that. Once we had those 2 components—the tooling platform and the operating system platform—we had to start thinking about creating more of that full solution. That's what brought us really into the vertical autonomy stack, doing more of these models ourselves and then making those available to customers. And now we have a really complete and very compelling offering in a lot of these areas.

Yeah, I think it's also worth highlighting this concept of a horizontal company versus a vertical company. A vertical company is kind of easy to understand; Tesla is a vertical company. A horizontal company is like NVIDIA. It's a company that sells this technology across a broad base of customers, who then package it together into something and take it to market.

It was important that the stuff that we built for one vertical could be used in other verticals. And so we also put that constraint on. Again, for the founders at home, it's not enough that you have ambition in building a company. Your ideas also have to be correct.

Almost 10 years ago, the conversation always was, “Well, tools is a bad business, and why be horizontal? Vertical is the right answer.” I really implore everybody to think from first principles, and our results speak for themselves.

In the company's history, we raised about $1 billion, and all of that is in the bank. We've never used any money we've ever raised, and there's not an AI company like us in the business that is financially stable and healthy. I attribute that to our technical strategy: putting the stuff that we make in automotive onto defense, putting the stuff we make in defense onto construction and mining, and so on.

I know you guys have this grand vision of getting—and you guys mentioned it—to 1 billion intelligent machines over the next decade. We're here to talk about Dana and how that's going to enable and unlock that possibility. Maybe you should talk about what Dana is and how that's going to help us get to that future.

Yeah, so Dana is our new agility platform for physical AI. This really is the culmination of pretty much everything we've worked on over the last decade. It makes developing these systems so much easier than it has been in the past. And it's important to understand why that's important at the outset, though.

Building physical AI is extremely complicated. If you ask, “Why don't we have intelligent robots and intelligent vehicles everywhere today?” it really comes down to the fact that building this stuff is really hard, and that is the limiting factor. We know how to make the chips. We know how to make the hardware for these systems.

It's more that actually developing all of the technology and getting it to work is very, very difficult. Our engineering tools over the last 10 years have been addressing parts of this. But now, with modern AI and this new Dana platform, we're really drastically reducing the barrier to entry to building physical AI and deploying it in the real world.

With the lowering of the barrier to entry, I think it's going to make it far easier to build a very large variety of solutions and really supercharge our customers.

Let's use an example. If you're building an app for an iPhone, high school kids can do that now, because there are all these things that exist and make it easy. And with new coding platforms, like vibe-coding platforms, it's easier than ever to make a web app. It's super simple.

It's very hard to do that in terms of robotics. If you wanted to build, let's say, a delivery robot for college campuses or a little vacuum that cleans your house, it's a pretty daunting thing, even for hobbyists and computer scientists. You have to patch together lots and lots of disparate products and tools. Then you have to somehow figure out how to deploy that software onto the physical machine.

Dana really is that, along with the fact that it brings a lot of that agentic power in writing software purely for web applications. The thing that Peter's really emphasizing, and the thing that we really want to do, is just lower the bar.

Anybody out there, starting with engineers but ultimately really anybody, can develop robots. And I think that really takes us much closer to that mission. Frankly speaking, I think it wasn't possible a few years ago because we didn't have the intelligence—the models, literally—that would help us create Dana and then for end users to use Dana to actually create intelligence.

How much of the drive to build Dana was about building where the puck is going? Or how much of it was these customer pain points or friction points, where we could enumerate some of them and then go ahead and build the solution for them?

This is just the next evolution of the stack that we're building, right? Tools, operating system, autonomy stack, and now we have Dana. How much of it was outside-in versus inside-out?

Both. I say both because we use our own tools to develop autonomy as well. We're our own customer, and those are different parts of the company. I would say our most aggressive customer feedback comes from internally, where there's very little patience for anything that doesn't work quickly and on the first try.

But we're an enterprise company, and for almost a decade we've been deploying tools to customers. We get feedback from many thousands of engineers who depend on us, and they're also using things like Claude. So then it's very natural, if you're using Cursor and Claude, to say, “Hey, where is the Claude-Cursor thing in physical AI?”

We're the company to develop that. This is, frankly speaking, our bread and butter. If you're in the space, it's very obvious. It should be that easy, just like it is to use one of the major coding platforms.

Help us understand why Applied Intuition is uniquely positioned to build this, rather than another company. As you said, there are other generic AI assistants or coding agents that could potentially do some of this. What's so different about Dana, and why Applied Intuition specifically?

Yeah. It's important to understand a bit more about the technology itself. The general-purpose models that come from companies like Anthropic or OpenAI are great, and they're very useful for very general-purpose tasks.

But when you're dealing with things that have a very deep safety-critical component—things where lives are literally on the line based on what is being developed—and they require a very complex development toolchain, just the model is not enough. That's 1% of the full solution.

The Dana platform itself is built on top of everything that we've built over the past 10 years. Everything has an API and has been rearchitected to work in a model with AI agents at the forefront.

These very complex workflows to actually build physical AI maybe would have required switching between 20 different tools in the past for different tasks, and deeply understanding the precise flow of information between all of those things to accomplish your end goal.

To actually make a physical AI system work, there are many, many layers of that technology stack. For example, why couldn't you use Claude to build the Linux kernel? Well, because the Linux kernel is actually very, very complex.

Now with Dana, all of those complex workflows can be very seamlessly orchestrated from an agentic interface, where you're able to write things in plain English, receive answers in plain English, and do very detail-oriented things that are very deep in data science and production deployment of this type of AI.

I think, using that example, let’s use an autonomous lawnmower. What are all the things that you need? First, you need some sensors. You need some compute. Then you need to create a software package that understands, hey, this is the yard I’m going to work in. This is the physical space I’m going to work in. Don’t go in other places. So now the sensors have to understand that physical space.

Then you want to create scenarios that the lawnmower has to successfully pass in simulation. So you need a simulation framework. How do you simulate your backyard? That also is complex. Simulation itself is a massive industry. There are companies worth tens of billions of dollars that only do simulation. That’s where we started; our bread and butter was in simulation, in the world-model universe.

Now that you have a simulated backyard, you have scenarios in a simulated backyard that you can run again and again, and you have to do that in the cloud. There’s a whole orchestration that needs to happen. Ultimately, once you’re performing at a certain level of efficiency and fidelity, you’re going to deploy that first version onto the physical lawnmower.

Then a bunch of things aren’t going to work. You have to figure out why they didn’t work. Why didn’t the actuation happen as you thought it would? Why are the control systems maybe not behaving as you thought they would? So then there’s a whole feedback loop. You can’t do that all in an LLM. It’s not made for that. LLMs are made for a different environment.

We keep hammering things like the hardware interface or safety criticality, but we’re, frankly speaking, underemphasizing how different it is to build a web app versus an AI product in the physical world.

And also in that process, right? You’re doing model training and evaluation. You’re doing data collection and post-processing of that data. These are very complex systems, but you can apply that exact analogy to any kind of machine or any kind of robot, and those same things apply.

Mhm. Your team, while we were prepping, actually described this very beautiful loop where you can take information and data that you’re getting from, let’s say, a tractor or an underground mine, and then that’s relevant to, let’s say, a drone or an autonomous vehicle, which is also relevant to an autonomous sea vehicle. They all feed into this broader platform and inform how the broader platform gets made.

I would love for you to describe that loop and how that helps build a general-purpose, broader platform.

An AI system really is always two big components. One is the actual platform where you develop the intelligence, and the other is the actual intelligence. It’s almost like the world model and then the actual intelligence that you’ll deploy on the machine.

On that second half, on the intelligence that you’re deploying on the machine, the way to think about it is this data engine. It’s a feedback loop. As the car is in the real world and consumes data—that is, it consumes the world around it and creates data packages—it also isn’t successfully navigating specific scenarios. You can almost mark, “Hey, the car had difficulty doing this.”

How do you help the brain on the car navigate that scenario that it wasn’t able to navigate last time more successfully next time? You can do it a couple of ways. You can expose it to lots of scenarios that humans may have already literally driven, so it imitates how humans handle that scenario. You can create a synthetic environment where you show it, “This is how you would navigate this type of scenario.” There are a bunch of techniques.

The macro point is that there’s a data loop. It’s just feedback. The machine interacts with the scenario, and it figures out what it can and cannot navigate.

Now, step back. Don’t just make that a car. Make that any type of machine. It can be a drone. It can be a mining dirt mover. It can be a combine. The same thing happens.

The interesting thing we’ve learned in our development of intelligence and models is that, as we take scenarios from, let’s say, a drone—we run autonomous trucks right now, L4 trucks in Japan—and take data from those trucks, it actually makes the performance of models in fairly different environments better. What’s really happening is the model is getting a sense of physics in the real world.

This should elicit some corollaries in the chatbot universe. Chatbots used to be very specific. Transformers happened, and general chatbots can now perform really well. The same thing has happened in self-driving. We benefit a lot from that because we see this diversity of data in all these different use cases, and so it feeds the data loop and the data engine.

It also gets to more of the distinction between digital AI and physical AI, right? In digital AI, you’re thinking about these general-purpose models. Those are oftentimes trained on the internet, plus maybe some extra data that the model company has built, and that’s usually text data.

But in physical AI, almost all of the data is actually proprietary, right? It’s data that we ourselves are collecting through our own vehicles and partnerships that we have with our customers, collecting that data because you’re ultimately building these models on data that’s just not available on the internet.

Peter, maybe you could talk about the quantum of data that you’re able to collect. It’s almost unimaginable across all the different vehicles and machines, industries, and applications. I’m thinking of this mega brain or giga brain, in a way. Maybe you could talk about the data that you’re able to collect, and then the actions that you’re able to take on top of that data that maybe no one else is able to do.

Yeah, so we do have an enormous amount of data. That is a fact. In terms of moat, it’s very meaningful because, first off, it’s just expensive to do, but it’s also very difficult. The actual tech stack required to do reliable, high-quality data collection is surprisingly deep and complex, and there aren’t that many companies around the world that really have a very high-quality tech stack for doing data collection for physical AI. That’s, I think, a pretty fundamental moat and long-term advantage that we have.

There are also all these other very deep things that are unlocked based on that data. There’s this combination of imitation learning with reinforcement learning, which we’re very deep in, and I think this is really the critical unlock to scale physical AI.

A lot of the talk right now in autonomy is about end-to-end models, where you basically take data that’s been collected and train a model off of that data using something called imitation learning. That then allows a machine to effectively mimic what was being done in that training data.

That’s great, and it’s been proven that it’s very effective, but oftentimes it doesn’t actually get you to a fully productionizable solution. What we’ve now added, and really innovated on in a big way—and I’ve done a lot of research and actually published a lot on as well—is reinforcement learning.

You take that base imitation-learning model and complement it with a really powerful simulation environment using highly performant reinforcement learning. That can actually smooth out a lot of the problem cases that you would get with pure imitation learning.

We see this as the technical path toward large-scale, widely deployed physical AI, and I think we have some pretty unique advantages across the spectrum right now.

Are there other rate limiters that we should discuss? Obviously, data’s a huge unlock for Peter. You were mentioning how the operating system is a rate limiter in terms of how hard it is to design, develop, build, test, and analyze all these different systems. Dana’s now going to be able to go ahead and do that.

Are there other rate limiters, whether it’s anything from the chips to the sensors to the actuators to materials? Power’s a big issue now. Is there anything around the actual technology that you’re building that worries you?

The diffusion of this technology will be at very different rates and in very different ways. If Fable comes out, it can work on your phone and your laptop because those environments are quite standardized because of the browser, the operating system, and a bunch of other things—app stores, payment methods, and stuff like that. So it’s very easy to consume that intelligence as an end user.

There are just impediments to diffusing this intelligence into physical machines. They could be manufacturers, or they could be the operators of the farm and the mine. There are a lot of other things that get in the way. Just as importantly, there are the actual dollars.

When people talk about self-driving cars, I always like to use passenger vehicles because everyone can understand them. It’s maybe a little harder to grok ports. In terms of your personal vehicle, let’s say you just bought a Honda Accord yesterday, and then tomorrow self-driving is available for free.

Well, you still own that Honda Accord. Over half of Americans live on fairly small savings accounts. When they buy a car, it’s a big deal and a big purchase, and they’re going to use that car for 10, maybe 15 years, regardless of what other product is available in the market, just because of the nature of economics and how much money they have.

The diffusion of this intelligence into machines has a lot of different complexities that you won’t see on a desktop or a phone. But I think those are also moats. Once you figure out how to make a mine autonomous or a farm autonomous, we, as a technology provider in that ecosystem, are really advantaged because we’re really in there. In the same way silicon is so sticky, once you’re a chipmaker and your chips are in a bunch of machines, that’s a really deep moat. So, we have both the disadvantages and advantages of those realities.

This is a business podcast, so we definitely need to talk about the actual business. But I’m curious how you would break down the revenue for Applied Intuition. There could be different buckets. One could be software, which has one margin profile. Services could be another, or consulting. Maybe just break down the different components of revenue.

We are a very classic product business. The way we make money is licensing. It’s really a straightforward relationship with our customers.

We do a weekly live all-hands inside the company, and we’re a little over 1,000 engineers, to give some scope of how big the company is. I always say that we want to be very innovative with our technology, and I want to be very boring with our business model. I think when it’s the other way around, that’s when you get into trouble—when you have a very boring product with very innovative ways to account around it.

Our last round was BlackRock, and Fidelity was involved before then. These are very traditional, conservative investors who do actual diligence. Not to say that venture investors don’t, but I think why I bring that up is that, as a founder and as a company—and as I’m speaking to other founders here—it should be really easy to understand your business. Your customers should have a very clear understanding of your incentives and motivations, how and where you make money, where you don’t make money, and what you don’t want to do.

For us, it’s: let’s make your products—that’s our end customers—better. We make them better by putting some intelligence into them.

Can you talk about the actual customer base? I was reading that 18 of the top 20 automotive manufacturers are your customers, and you guys expanded into—I think you were talking about this—we’re not just land autonomy anymore; we’re sea, we’re space, and a lot of other industries that you guys are going into. Maybe just give the audience a sense of the different customer buckets that you work with. To my understanding, it’s also quite global, so give us a little bit of a feel for the different countries that you work with, too.

Our customers typically, but not exclusively, are manufacturers. They’re people who make physical machines. I say “not exclusively” because we also work with folks like a mining operator or somebody who runs a port. They’re automating a heterogeneous mix of machines, and those machines have to talk to each other and work with each other.

We provide either Dana, the platform that is the tooling side, or the actual intelligence that a manufacturer would use and then embed—literally embed—into their machines to make them more intelligent.

In terms of the verticals, it’s all the big verticals that make machines and deploy them in the real world. It’s automotive, commercial trucking, defense, construction, mining, and agriculture. In the short horizon, it’s robotics, humanoids, and space—anywhere where there’s a physical machine moving around with people, goods, or information.

It’s frankly fairly evenly split. A lot of times, people think we’re an automotive-only company because I went to the General Motors Institute. It’s frankly a minority of our business, so we’re quite evenly split. We’re also quite international, as you mentioned, so we really work across the globe.

Our first international offices were opened almost right when the company started. That’s also part of the founding story. I’ve lived in Japan and Germany, so obviously opening offices in Detroit, Japan, and Germany as literally our first 3 offices made sense. Then, as we got into defense, going to D.C. was a fairly logical thing. We’ve always liked to be close to our customers, and that’s a good reason to have international offices.

There’s also a lot of engineering talent, frankly speaking. This is not just, as Peter mentioned earlier, about knowing AI and optimizing models. There’s a lot more to our technology, so we find people around the globe who can help us succeed in our mission.

From an outside-in perspective, it’s actually quite hard to pin down exact direct competitors. There are synthetic-data providers, and you could have big platforms like NVIDIA. You mentioned Tesla, which is more vertically integrated—they’re full-stack operators. I’m curious how you think about competition and whether there are some companies that you feel are more aligned with or in your path.

In conversations like this, it’s also important to define even the word “competition.” There are a lot of companies that play in, let’s say, self-driving, but that doesn’t necessarily make them competitors. Waymo is an example. We’re both ex-Googlers. Is Waymo a competitor? Not really, mainly because they don’t take money out of the bucket that we’re taking money out of.

We’re selling to manufacturers; Waymo is doing a robotaxi for consumers. If we did a robotaxi for consumers or Waymo sold to manufacturers, then we would be more direct competitors. But by any definition, it’s not really a competitor.

You’re correct that there isn’t really an Applied Intuition out there, but there are many companies that compete with portions of our business. There are companies that make something in construction or mining, or something in automotive.

From our perspective, our own team gets asked these questions all the time about how we should think about competitors and things like that. I fall into the classic YC model here, which is: you should be aware of your competitors, you should fight them aggressively, but you can’t let them dictate your future because they are a different company with different skills, and these markets are really, really, really big.

Competition becomes really important if you’re in a small town and there are 1,000 people who live there, and those 1,000 people are going to go to 1 shoe store or 2 shoe stores. Then competition becomes really important because it’s a little bit of a zero-sum game. There’s a finite number of shoes they’re going to buy.

In our business, the market is growing so rapidly and so aggressively that, let’s say you wrote down all the subcompetitors for Applied—all of them could be successful, and Applied Intuition could be successful, because the markets are so big.

One way to think about this—I’m talking to founders here of young companies—is that when you have kids, you see the image of the universe where they show the Sun, Mars, and Earth and stuff, and everything kind of looks close together. It’s just there to show the Earth in relation to Saturn and Jupiter. If you actually have that at scale and the Sun is the size of a basketball, the Earth is many tens of feet away and is a little pin, like the head of a ballpoint pen. All this vastness is black, empty space.

Markets are kind of like that. People focus a lot on how close these companies look to each other, but the markets are so vast and so big that they actually don’t really impact each other’s gravity. I very much fall into the view that if we don’t succeed, it’s because of us. If we execute, we’re going to do fantastic, and I think this company can be, honestly, certainly 10×, if not much, much bigger than it is.

Before, when we used to say things like this—that there would be a multihundred-billion-dollar physical AI company—people would say, “Well, it doesn’t really make sense.” Now you see how big companies like SpaceX, Anthropic, and OpenAI have gotten. They’re hard-tech companies that are really focused on one thing, and you just see, “Wow, these markets really, really are big.”

Peter, there does seem like there’s a real renaissance of people building physical and hardware companies. You have Bezos and Prometheus. You have Travis Kalanick and Adams[?]. I think literally just today, a company called TerraFirm[?] or something that’s building robotic mining technology.

I think, frankly, these are all potential customers for us, right? I think it’s great that many more hardware companies are starting, and so much of this has to do, again, with the barrier to entry that we talked about earlier. If building an intelligent hardware system is an extraordinarily daunting task, very few companies are going to do it. But once it becomes more achievable by a reasonable-sized team with a reasonable amount of funding, then all of a sudden you can have hundreds or thousands of organizations building all kinds of things. We can sort of imagine what those things could be, but a lot of it is going to be the creativity of humanity that comes up with these new physical AI use cases.

Just to echo Peter, I think all of these folks could definitely be customers of the company because we provide that platform in order to develop this technology. And I think, if you look back again, history is such a great way to learn about these things. Google started in 1998, when there were multiple search engines that were already public, and I think if we were having this podcast in ’98 and said there was a new company coming up, you would say, “Well, the market’s already saturated.” These markets are really, really big.

I think the instinct always is, “Oh, should you be worried that Jeff Bezos is going to start a physical AI company?” That’s like saying Jeff Bezos is starting a software company. It’s like, yeah, it’s definitely—we are also a software company, but it doesn’t necessarily mean anything as ominous and as negative as that is. [laughter]

I heard this crazy stat that you basically haven’t spent any of the money that you raised, and you raised a non-insignificant amount of capital, somewhere in the range of $1 billion, and you’ve built this company to be self-funding. So the question that begs to be asked is: Why have you raised that amount of money, and how do you broadly think about allocating and deploying capital?

Yeah, just to be very clear, we’ve tried to spend it. [laughter] We’ve been fortunate enough to grow faster than that. In every fundraise, I always start off with, “We intend to spend this money. We don’t intend to raise it and put it in the bank.” So that’s one very obvious thing.

I think, as we look at resource allocation, we want to be very thoughtful, but not so conservative that we become vulnerable to an emerging company that wants to, let’s say, be less frugal or something like that. As we look forward, we’re fortunate enough—I think partly, frankly speaking, it’s our track record. Partly, it’s what we did as technologists and engineers before we even started this company, so that we could raise very significant amounts of capital from the markets if we needed to.

I think the way we think about this is: “Hey, this is our mission. If the bottleneck is capital, as Peter was talking about, then we should take care of that. If it’s technology, we should take care of that. If it’s customers, we should take care of that, or product—whatever products we need to build. So it’s just one variable in the path, in the mission, and when we see it being constrained, we fix it.” I think also, frankly, we’re getting to the size and scale where we could deploy a lot more capital much more effectively. It’s something we always talk about and think about, but I wouldn’t say it’s the first thing I’m thinking about in the morning.

The first thing I’m thinking about is, “Okay, how do we make sure we’re making the best products in the business?” If we make the best products in the business, a lot of things take care of themselves because, unlike other businesses, the product really matters here. In a lot of businesses, the products can be kind of okay, but not in safety-critical systems.

Similar to the challenge of pointing to a direct competitor, if I force you to point to a public company or a basket of public companies that would be helpful for an investor to value Applied Intuition against, where would you point them?

Yeah, there aren’t many publicly traded companies, frankly, that directly play in this physical AI world. Companies will talk about it. NVIDIA talks about it. They’ll talk about physical AI, but I think that’s why there’s frankly enthusiasm around Applied Intuition. If you listened for the last hour and had your thinking brain on, it’s pretty easy to understand the problem and the solution. We’re the category leader in physical AI, and it’s a big market. That’s really the punchline for why, I think, we get such enthusiasm, honestly, from engineers and investors alike.

The companies that you’ve been fortunate to work with are kind of a who’s who of leaders across all of these industries that we talked about, right? You have defense, you have automotive, you have farming, industrial manufacturing, et cetera. And you have such a unique perch and vantage point. Looking 3 or 5 years out, what are the most interesting ways you think the future will be different than today?

I think the future is, frankly speaking, safer. This is not to be understated or made to be pithy. If you know anyone who’s gotten in a car accident or who’s been in a workplace accident on a farm or a mine, it’s absolutely devastating in a way that is hard to quantify because it impacts everything they do forever, for the rest of their life. And so I think that’s huge.

As we started off, you mentioned how in San Francisco you see self-driving all around. I think that’s going to be way more common in many more cities and with many more companies, not just Waymo or Tesla that are fielding those products. And then, as you go to other places, you go to college campuses, you’ll see shuttles and food-delivery robots more and more. You already see some of them, but you’ll see this at an increasing rate. Before you know it, just like having a supercomputer in your pocket is taken for granted, having machines move around you, take care of things for you, and take care of you will be taken for granted. And that’s a very positive thing.

Peter, Qasar, it’s been a pleasure. Thank you for your time.

Yeah, thanks for having us. It was fun.

Thanks for having us.