A Dialog with Chris Urmson, CEO of Aurora

Yeah, thanks for having me.

Weather. [laughter] No, no. You know, I had you know, when I was in high school my dream had been go to a Canadian university, go to a top-tier US university, and then become a faculty member. And so, you know, I had the privilege of getting into Carnegie Mellon. I think they made a mistake on the admissions. I got to spend some time there, and then ultimately got to the point where I had a faculty job, and it had been after doing the DARPA Grand Challenge and Urban Challenges, which are ancient history at this point, 20 some years ago, but they were these at the time these big idea was you could get a vehicle to drive across the desert and nobody had done it before. And that was exciting, and we ended up doing quite well at that as a team, and then I got approached by Sebastian Thrun, who was at Stanford and Google at the time about, hey, we should do something together, and then ultimately that evolved into, hey, Larry would like to do self-driving cars. Well, except that wasn't the word, automated cars or something. We didn't have the word self-driving car at the time. At Google, and you know, my wife and I had just bought a house, we'd settled in Pittsburgh, great friend network, you know, amazing job at the university, and it seemed like a big risk. Right, this is 2009. So, at that point I think the iPhone had launched a year before, and Android had launched that year. And Google didn't have Chrome yet. It was just a search. So, the idea that this webpage would want to do automated vehicles seemed crazy. But it was you know, it was an alluring opportunity and my wife and I thought about, you know, okay, if we're going to go try something like living with the crazy people in the Birkenstocks in California, this was the time in our life to go do it. So, we packed up and moved over there.

And it was you know, I look back at my career now, and while that was a very difficult decision going in the forward time direction, the backward time direction was clearly a no-brainer. And being able to go to a place where you know, the university environment at Carnegie Mellon was phenomenal, but where we had the resources to really do what I was excited about, which was take this technology and translate it into something practical. You know, it was just an incredible opportunity. We lost a professor, but we got an AV industry coming out of this, right? And I really, you know, I couldn't say enough thank you to Larry, Sergey, and the Google folks, right? Because if we again go back 20 some years now, the Defense Department through DARPA funded the DARPA Grand Challenges, and it really created a lot of energy in the space, but DARPA's mandate is to get stuff to 10 to 20 years out, and then hand it off to industry. And at the time, you know, 2009 was not a great time for industry to be investing long-term, right? It was a bit of a financial mess. And so, for Google to be willing to kind of take a flyer, and you know, invest what they did, and initially it was you know, for their budget a trivial amount of money, but for a university or most places it was a significant amount. And then it actually did, you know, it just kind of between the investment DARPA made and the investment Google made, I think it created a whole industry, which is phenomenal.

Yeah, well, at some point it's time, right? I had been at Google for many years, and it was a phenomenal opportunity. I got to work with some of the best people I've ever worked with in my life. Still have great friendships with those folks. And awesome resources, but at some point I didn't agree with some of the decision-making that was happening at the company. And I have a belief in business, you know, you have three options. You can try and fix something you don't like, you can get in line, or you can get out of the way. And I tried very hard to fix it, and you know, it was hard for me to kind of buy into some of what we were thinking at the time there, and the company and the team deserve somebody who did. Right? That we, you know, at the time I had 600 and something people reporting to me, and if I didn't believe what we were doing, I needed to get out of the way. And so, I ended up stepping away from Google and spent some time trying to figure out what to do. It was very flattering, right? I had a lot of people that were interested in hiring me, and ultimately thought about what would minimize regret, right? I had been doing this great stuff at Google and loved it. I thought, you know, I had conviction that automated vehicles were going to be a thing. And I didn't want to be telling my grandkids, you know, I could have been a contender. I could have been part of that. And thought, okay, I can go and try and start a company where it'll either go really well, and it'll be a happy ending, and I'll be excited about it, or it go really well and I don't enjoy it, in which case I can just get out of the way again. Or it go badly, and you know, I'll know, right? And I won't have regret of like what could have been. And so, ultimately ended up starting this. Had two amazing co-founders in Sterling Anderson, an MIT grad, though we try not to hold that against him. And Andrew Bagnell, my colleague from Carnegie Mellon. And you know, it's been an amazing, I guess now we're at nine and a half years.

Yeah, something like that. Yeah, right around then, yeah.

Yeah, so we continue to be building a general-purpose driver, but the application we're choosing to go after first is trucking. And so, talk a little bit about the background of that. So, we started working with light vehicles, frankly because it's easier. Right? That just operationally, a car costs a lot less than a truck. You can fuel it all over the place. You can fit in smaller spaces. You know, you don't have to have a very special driver's license to operate them. You know, it was just operationally easier. And then added to that at the time we didn't think the technology existed to see far enough down the freeway to truly do level four driverless automation. And this is borne out of some of the experience I had at Google, where most of the time when you're driving down the freeway it's quasi steady state, right? We're all moving about the same speed, and so you don't actually have to see that far to interact with the vehicles next to you. The problem comes if there's something rare that happens, like a vehicle stopped in the freeway and you're the first to encounter them. And there you have to see that at a very long distance to be able to react with a rational amount of braking or avoidance maneuver.

And so, when in Google we had done experiments on this, and we kind of come to conclusion like, ooh, we don't know how you can see that far yet. And of course there's been some breakthroughs. And so, at Aurora we said, okay, it's easier to operate light vehicles to begin with. We ultimately want to operate all vehicles, but now we don't have to solve that. So, let's focus on building the core competence. Meanwhile, I spent a couple of years with one of my colleagues here trying to find the technology that would actually unlock high-speed driving. And we ultimately found this in this little company in Bozeman, Montana called Blackmore. And what they had been working on for 10 years, 15 years before that on frequency modulated continuous wave lidar. And we can get into that at some point, but basically it unlocks the ability to see further down the freeway. And at that point we said, okay, we've got this basic driving capability. We now think we have a unique ability to see far enough down the freeway to solve this problem. And the business of trucking was a much clearer and more compelling business than the ride-hailing business. And I credit Sterling for that. He was a big champion for like this is actually the place where we can go build the most interesting business. And so that's where we've been.

Right. Yeah, yeah, because we thought we had kind of conviction that trucking was going to be a big and important market and we should go win that market. But we didn't believe at the time, this was 2017, 2018, that you could literally see far enough robustly to trust a 70,000 pound thing driving at 70 miles an hour. And so this technology was a fundamental unlock for us to go attack that space.

Yeah, it's a big deal because I mean it's interesting we talk about the technology, but I can tell you about the product advantage is we can react to things somewhere between 11 and 20 seconds sooner than you can relying on conventional lidar. And that is the difference between an oh crap and oh look at that. And it's at the point now where we are seeing things, particularly at night, the lidar is allowing the vehicle to react to things before our operators on board the truck are able to see them. And so originally it was why is it making that lane change too? Okay, it's moved over. What was that? Oh, there's a guy walking down the side of the freeway or there's a vehicle parked on the side of the road and we've just moved over and made space for it before the person in the truck is able to react to it.

Yeah. So if you go back and look at some of our white papers from 2017 that we published, we talked about the importance of having a data-centric approach to developing this technology. Right? That you can build engineered systems to allow you to bootstrap to the point where you can gather data from the environment, but you need to be purely, to the degree possible, purely data-driven. And that it's clear that AI and ML are going to play a huge part in the success of this technology. And I think we've maybe under-marketed ourselves on how we've approached that. But where we are, the system we've developed today, we've branded verifiable AI. And the idea here is that in many ways what's happened in the large language model field is, well, it is just fundamentally mind-blowing. Right? It's incredible what's happened there. And things that I wouldn't have believed possible five years ago are now certainly real. But those applications have the benefit that in general there's a person that can intervene between the generation and the execution. The canonical example from a couple years ago is if you ask the LLM, I think it was Gemini, how do I keep the cheese on my pizza? It told you to put glue. Right? And any rational person would say that's a dumb idea. Okay, try again. The equivalent at 70 miles an hour will be a disaster.

And even today with the incredible advances that these systems have made, the last version of Gemini was released where they were proud that they had dropped the hallucination rates to 40%. Right? And when you think about the diversity of what we encounter in the real world, we just can't bank on that by itself. And so with verifiable AI there's two important concepts. So one is we decompose from one kind of homogeneous model into, in our case, fundamentally two models. A part that's seeing the world and a part that's responding to the world. And then we create an interface between the two that defines how do we want to understand the world and then how that's expressed to the part that's reasoning about it. And this loses information. Right? Because the expressivity of the first model is limited by that interface that you've created. But importantly it forces the action to respond to things in the world that matter. And it allows us to create a testable system.

And so maybe let's talk about that. If you think about these two parts, understanding the world and reacting to the world, they're both exponentially complicated. And if you smoosh them together into one system, you now have the product of exponentials. In contrast, if we are able to thoughtfully decompose them, we have two exponentially complicated systems. Which verifying and validating that is literally exponentially easier by definition. And so that's an important concept because we're not just trying to make a thing that looks like it drives or seems to drive well, we want to before we put it on the road have conviction that it drives well. And that means that designing for testability is actually an important part of this.

The other part of this is that these large models are so complicated our ability to truly understand what's happening in them is still limited. You know, and I point to the folks who say, oh no, just put it in one model and what we'll do is we'll take heads off and we'll train to make sure that it understands the concepts of cars and pedestrians and whatever else it is that we have part of that interface. The contravening concern around that is that if you look at some of the studies on how these reasoning models operate, the idea is show your work so I'll understand what you did. What we're seeing is that they will actually not always describe how they got to the answer. They will provide a description of how they got to the answer that feels compelling, but is not actually the process they did. Right? And so the parallel here is that just because I trained heads that would get the model to have certain concepts doesn't mean, certainly doesn't require that is how it's actually understanding the world and how it's reacting to it. And so for us making sure that we know how the thing works, to a level of resolution is critical and then being able to test and validate and have conviction of how this works.

Yeah.

Yeah, so we are in a quiet period, so I'm not going to be breaking any news because we're a publicly traded company at this point. So there's only so much I can say, but what I can tell you is we started last year in April operating a truck between Dallas and Houston in good weather in the daytime. And over the course of the last year we've expanded the conditions in which you can operate. So we operate day and night. We operate in the rain at this point. We also expanded from Dallas to Houston to Fort Worth to El Paso, Fort Worth to Phoenix, El Paso to Phoenix, and Dallas to Laredo. And so it's been a really exciting year for us. And it's going to pale in comparison to how exciting 2026 is. So we've started 2026 with a handful of trucks operating on these roads. Over the course of this year we're going to end up operating a couple of hundred trucks. We're going to be operating them across the Sun Belt serving customers and hauling freight as we have been for the last several years for them and providing value from this technology.

Yeah, so maybe I've got to be a little careful here, but I can tell you how we talk about it. So when we think about the value to customers of automated trucks and our customers are the trucking companies. So companies like FedEx, Werner, Hirschbach, Schneider, Dettmar. These are companies whose job is to own trucks and haul stuff for other people. So the values we see for them. So first is a safety benefit. For every one of these companies safety is first. The fact that the Aurora driver is never distracted, that it has 360-degree field of view. Something that's underappreciated is that it's actually able to provide equal attention around the vehicle. So I think of an example. One of our trucks was driving down the freeway and we came to a point where the freeway separates. And there was an 18-wheeler parked on the gore point with the driver had got out onto the side of the road and was, I don't know, checking the load or doing something. Terrifying place to do it. And so as a person looking at this, my attention is focused on that. Like jeez, that guy's in the road. Meanwhile, at this gore point, there's a pickup truck with a trailer that decides that even though they're on the right hand side of us as the thing's splitting, they want to be on the other branch. And so they zip right in front of the truck at the last moment, do the New Jersey slide and go take the exit. And fortunately, the truck was driving, your driver is driving because it's able to both pay attention to that critical person on the side of the road and also see all the rest of the traffic. And so it has already started to brake as it anticipates this pickup truck making this move across in front of us. And so that kind of very anecdotal bit is part of that full field of view but fully attentive and sharp intellect that we get that brings safety.

So that matters to customers and that matters both in terms of the morality of being safer on the road, but also in terms of their insurance costs. Right? And the impact of that to their bottom line. A second part is fuel economy. So we expect to, and we've got a white paper published on this, to be able to save between 14 and 34% on fuel economy for these vehicles. And again, that's a big deal. Like fuel historically has been top three in the cost structure for these companies, particularly with the ongoing current geopolitical dynamics, it's going up. And so this is a big deal. And then of course that's good for the environment. Right? That goods need to move being 14 to 34% more efficient means 14 to 34% fewer emissions. And then finally, access to drivers. So anyone who is willing to drive a truck, we should say thank you to. If you look around the room you're in, the room I'm in, literally everything here moved on a truck. Our economy would cease without trucks. At the same time, we don't have enough people who want to drive trucks. Over the next decade we expect to be short a million truck drivers. And so for our customers, being able to get access to a skilled pool of drivers to complement and importantly to complement the people they have driving trucks is critical. And so we see this as a way to help them power their businesses and to kind of continue grow the economy over the coming decades.

Yeah, we started to look at this loosely and there's two ways that it helps. So there's some places where our customers just can't get drivers and so they put it in air freight. And that's just expensive. Like it's dramatically more expensive. But yes, I think this kind of mid-range air freight is one where automated trucking is going to quickly win that business. And this comes in particularly relevant around the holiday shopping season. So you have some bit of consumer electronics that is really selling well in San Francisco and is not selling at all in Houston. You need to effectively overnight move that from your store in Houston to San Francisco, otherwise you are missing that shopping window and you're stuck with a bunch of unsellable inventory. If you stick it on a plane, which is kind of what you do today, it gets there tomorrow. But you've lost most if not all of the margin. If you stick it on a truck, it gets there in three days and you've missed the window. If you stick it on an automated truck, it gets there tomorrow. Right? And you now have most of your margin still. And so that becomes a really interesting impact there. And then when you think about the footprint of distribution centers, that's one thing where again, they're based around the human distribution network and so we expect that real estate market will shift substantially as well.

Yeah, the Aurora driver is agnostic. The practicalities of long haul trucking today are that the batteries for kind of significant range are heavy enough that it actually impacts the payload you can pull. And for certain kinds of goods, you cube out before you weight out, but for lots of goods, you're bounded by the weight of the trailer or the gross vehicle weight of the whole system. And so that's one of the real detriments for long haul trucking for batteries today. We'll see where things go with fuel cells. We'll see as we get to solid state batteries, which are potentially lighter. You know, it'd be great to have this electrified, but today it's practically not useful for long haul freight.

I can tell you what's right. It's the one we're in. [laughter] So yeah, maybe what I'll do is I'll tell you why we think. So our model is to work with the ecosystem rather than try and change the ecosystem substantially. So today we are actually a trucking company, right? We have a DOT number, we haul freight. We don't go out and bid from shippers. We basically use it as a way to kind of try before you buy for carriers. So while we have our own DOT number, we will haul freight for Werner for their customer or for Hirschbach for their customer. And so starting next year we'll be driver as a service, which means that our customers will buy a truck, they'll pay a subscription to Aurora. We'll get paid based on how we drive that truck for that customer.

So why that model? So first, it allows us to be asset light. Right? That we're going to be focused on what we can do best, which is build the driving capability, work with others, rather than having to figure out how to do all these other things because I'm a big believer if you're a company that's been around for 50 or 100 years, you're an apex predator. Right? That you have evolved over that time to be extremely good at what you do. It may not be what I do, but you're very very good at that. And so whether it's Peterbilt and Paccar or Volvo or Daimler Truck or International, right? These are companies that know how to build trucks. Let's go work with them, right? Let's go help serve their customers and allow them to be more competitive in what will be an evolving market. Similarly on the carrier side, if you're FedEx or Hirschbach or Schneider, like you've seen a lot of challenges. You've evolved your management structure, you've evolved your business model, you understand how this works. And so let's support you to continue to grow and deliver your business. And let's be proud of what we're good at, but let's have the humility to understand that we're not good at everything and then go work with others to go do that. And that's been, I think that approach has served us well. And like I said, a nice economic benefit that we end up being asset light in the long term, which I think is a good thing for us. Because frankly, we're just not that good at maintaining and owning trucks.

And other people have a lower cost of capital than we have. And then the other is with this driver as a service model, it slots into the existing ecosystem kind of naturally, right? Today, they buy a truck and they pay a driver, and going forward they'll buy a truck, they'll pay human drivers and they'll pay the Aurora driver. And so it's a very natural kind of alignment into their business.

Yeah, I think this is one of the biggest red herrings that people talk about. Like, oh, how is insurance and liability going to work? We're in America, we'll figure it out, right? We don't have a shortage of lawyers. And so most obviously there will be product liability. We're telling people we can drive this truck well for them. We stand behind that. At some point if we're not, then there'll be legal proceedings that are associated with that. When we talk to insurance companies, like you said, they've been insuring things, some of them since the Mayflower sailed across the Atlantic, right? And if they had the same insurance product today that they had then, they wouldn't be in business. And so they've evolved with the technology over literally centuries in some cases. And so when we look at it, we talk with customers and we have what I can say in few words which takes the lawyers a lot of words to say, which is if the Aurora driver is responsible for something that happens on the road that's a bad outcome, then we're responsible for it. If the base vehicle breaks in some way that was not supposed to, then that's on the folks who made the truck. And if the operator used the vehicle in a way or abused the vehicle in a way that's kind of not appropriate, then it's on them. It's kind of common sense and insurance companies seem to understand that. And then around the insurance part itself, today we pay less to insure our trucks on a per truck basis than our customers do. Because of how seriously we put safety, how we can demonstrate that it's real. My expectation is ultimately that insurance in this market will effectively collapse to the loss rate plus epsilon. Because we're going to be able to very clearly predict the loss rate. And so the margins on that will collapse like they do in anything else where you have very good insight and understanding.