Aurora Innovation, Inc. (AUR) Earnings Call Transcript & Summary

Okay. Great. Thank you, everybody, for joining us. My name is Mark Delaney, and I have the pleasure of covering Aurora for Goldman Sachs. With us from Aurora, we have Chris Urmson, the Co-Founder and CEO. Really appreciate you joining us.

Glad to be here. Thanks for having me.

I thought, Chris, given your extensive background in the autonomous space, I wanted to get your thoughts around AI, and how Aurora is pursuing this? You've talked in the past about using, what you call, a Verifiable AI for your product development. Help us better understand what Verifiable AI is? How is that similar or different than a full end-to-end approach?

Yes. Yes. Well, thanks. And for those of you who don't know the company, we're working on making trucks drive themselves, ultimately driving all kinds of things. Our mission is to deliver the benefit of self-driving technology safely, quickly and broadly. And we're operating trucks between Dallas and Houston today driverlessly and anticipate by the end of the year to be operating between Fort Worth and Phoenix, Fort Worth and El Paso, and El Paso and Phoenix. So that's kind of the kind of 30-second version of what we do. Verifiable AI is the way we think about bringing kind of modern AI techniques to a safety critical space. There's been a lot of excitement, of course, in the large language model space and the visual language model space. These kind of techniques are exceptionally compelling in that area. But in a world where we can't afford to have a model kind of doing the proverbial, use glue to hold the cheese on your pizza, we need to do something a little beyond just kind of train and pray. And for us, that's this is verifiable approach, where we combine the best of AI machine learning with guardrails. What does that look like in our system? It is -- one, it make sure that we actually can understand the representations that are being expressed by our models and using those as kind of a way to contain and constrain the communication of what's learned by the action part of our system. It also allows us to put in place constraints on how the system can respond that can't be kind of jail broken, if you will. Foreign constraints, for example, we want our vehicles to stop at stop signs. If you just build a trained model, it turns out something like 11% of people actually stop at stop signs. So if you actually trained off of that data, more likely than not, you're not going to have a stop at the stop sign, which we think is actually a pretty important part of building a safe system. One thing that's actually really interesting to see is how the story around this -- around end-to-end models has involved. A couple of years ago, when ChatGPT was new, and large language models were having the kind of introduction and large moment, as well, you've got to be doing completely end-to-end stuff. That's what those guys are doing. It turns out that if you actually look how these models are now beginning to work, they look much more like the approach we've been taking for a while because as these models move from curiosities and chatbots into things in industry and businesses are relying on, they're moving to federated models and constellations of models. They're moving to learner ranker approaches to make sure that the thing that comes out the end of it is something that makes sense and is valuable. And so it's great to see that kind of this approach that we've spearheaded is starting to be the way that other people are starting to think about this as well.

As you think about that approach to your product development using Verifiable AI, what does that mean for your ability to add new features to your stack? Is it accelerated, slowed down?

It accelerates to make sure we get it right, right? For example, by the end of this year, we intend to be operating in range. If you came on road in 1 of our trucks today, you'd be like, "Chris, why have you not launched this feature." It works. It works well almost all the time. But when we're talking about an 80,000-pound thing, moving at 70 miles now down the road, almost is doing a lot of work in that sentence, and we want to make sure it's really right. And so for us, being able to understand and evaluate and validate the system, avoid the challenges that come along with kind of a compounding exponential of set of complexities that come from both understanding the world and figuring out what to do in it and be able to differentiate that and verify the sub-elements of it, means, we can have really strong conviction that the thing that we put on the road at the day is going to be safe.

You mentioned before about starting to operate between Dallas and Houston. That was your launch lane, you began in late April. Talk a little bit more on some of the key learnings from that? And how has customer feedback been?

Yes. It's been kind of what we expected, right, in that the vehicles are out on the road, they're operating. As of last week, we've done 50,000 driverless miles on public roads. So this is kind of a big illustration of kind of -- that's starting to accelerate and pick up. The technology works, the validation process we put in place, we believed in it. We had conviction in it. And what we're seeing is it behaves the way we expected, right? The performance on road matches what we wanted, and that's really what you want at the end of the day. And what it's meant for us as a company is, one, we've believed for a long time, we were ahead of the game. And now it's kind of demonstrable, right? We're the only company in the world that can do this. And we feel like with the approach we put in place with Verifiable AI. We're actually continuing to accelerate relative to any competition that might be. But two, when we go and talk to customers, it goes from a hypothetical. What this could mean for your business if self-driving trucks are a thing to now self-driving trucks are a thing. We have them. They're coming to a freeway near you in the not distant future. And it's really caused a significant uptick in enthusiasm from potential customers.

The next big thing is already here. That's Samsung would say.

Yes, I guess so. Yes. We're [ going to look ] at that tagline for sure. Let's go with it.

I think it was a smartphone advertising campaign, they ran when they were starting to compete more against the iPhone, if I'm remembering correctly and [ dating ] myself.

I don't know if I want that as our metaphor, yes?

You recently expanded to nighttime driving. With that change, how many miles per week is 1 of your AVs typically doing fully autonomously?

Yes. So the way to think about it is where this is heading, right, is that we expect trucks to be able to do somewhere between 200,000 and 250,000 miles a year, if not a little bit more in certain situations. And being able to go from operating in 8 to 12 hours a day of daylight to be able to operate and night, means, we unlock that potential. That translates to something between 4,000 and 5,000 miles a week roughly. And so yes, we're -- that's kind of what we're expecting.

And the company, as you mentioned before, is planning to address rainy and windy conditions as the next step in expanding it's operational design domain. Can you provide an update on how the validation, testing and safety case for these conditions is progressing, and if you're on track to do that by the end of the year?

Yes, it is moving along well, right? We continue to expect to be able to address those conditions by the end of the year and begin operating in them. And again, it's really important to understand there's a difference between the performance of the system and the validated performance of the system. Today, we take you and put you on 1 of our trucks in the rain, and it will feel like it works. If you're unlucky, what you'll see is that in certain conditions, we'll sometimes think that some of the vehicles adjacent to us are a little bit bigger than they really are. And that may cause us to think that they bumped it to us, right? And that causes them to pull this out of road and stop. It's rare, but it's a thing that you don't really want in a product. And so as we continue to pull the additional data that gives us conviction that the rest of things, that will work the way we will, and we make the minor refinements to the system. We're feeling really good about the timing.

Very helpful. Company is also working on adding driverless operations for it, Fort Worth, El Paso, Phoenix Lane. How are you progressing with starting to add those lanes to your map and overcoming some of the more unique features like having to go through a custom stop?

Yes. And I'm going to sound a little bit like a broken record. And that, yes, it kind of works that we're on the road today. We have those -- we're operating for customers Fort Worth to El Paso and on to Phoenix. Phoenix is particularly interesting because at that point, it's 1,000 miles, about 16 hours of driving. It's -- Fort Worth to El Paso, it's right around the hours of service limitation for a human driver. Fort Worth to Phoenix is beyond it. And so companies like Werner are super excited about that and the opportunity for their business. And again, we're feeling quite good about it. As you said, there's a couple of things that are a little bit novel. So the prevalence of cattle on the freeway is slightly higher than it is between Dallas and Houston. Again, not a thing I think we've ever seen, but we're making sure we're good with -- we don't want to be hitting someone's steer. And then there's this inland border patrol station. And again, we've put in place a first-of-its-kind partnership with customs border patrol agencies -- agency. We were -- we've been testing that and working that for a couple of years. And now it's just let's go through the final validation and call it ready to go.

Great. That's good news. As you start adding some of these different capabilities, right, going to a different city, going through a customs and border patrol stop, how much incremental R&D is needed to add those sorts of features? And how long does it take to validate those things?

Yes. Well, I guess you can be looking at the clock, right, that we got to the initial launch. It took us about 8.25 years to be able to get there. By the end of this year, we expect to be operating in the range we expect to have unlocked 3 additional routes. So relatively modest. And part of that is because of the investment we've made, not just in the technology, but in the process and tools that allow us to have conviction in the technology. And I really think that important to understand. If you look at a company like SpaceX, in fact, let's look specifically at SpaceX. The reason why SpaceX is lapping Boeing is not because Elon is brilliant and he very well maybe, but it's because they can validate their flight software in a week, and that takes Boeing a year to 3 years to do that. That means they get 50 to 150 more shots on goal than Boeing does. And if you read any article on business, any article or any book on business or on technology development, fast iteration cycles done safely is the key to this. And so for us, we've invested heavily in this VNV tooling and framework. And that means that not only do we have a product that's on the road and safe, but that we can continually add to it and improve it and do that rapidly and with conviction.

And that's something you've developed in-house.

That's something absolutely we developed in-house. And talking to the team, you want to -- my theory is, if you think we have 3 major assets for the software in our system, we've got the actual code itself, our data sets or our process, you told me you had to delete 1 of them. I'd tell you delete our code. Because with the other 2, like please -- by the way, do not delete any of them. They're all super valuable. But if you had a gun to my head and told me you had to delete 1 of them, I'd say delete the code because with the infrastructure we have and with the data we have, we can quickly go and ultimately recover that if need be.

And just clarify, the VOV, what is that?

VO -- VNV.

Oh, VNV.

Verification and validation. Sorry, I pronounce it more clearly. This is basically the process for making sure that the thing that we've built actually does what we think it should do, and what we should do is actually what it needs to do to be out in the world and be safe.

That's a lot of simulation and software capabilities...

Yes, it's a combination of simulation. And importantly, simulation is not enough, but actually understanding how to use the simulation to test, right? Because you can create a virtual world, and it's pretty easy to understand, did the thing crash into something in the virtual world. Understanding whether it didn't just not crash into something but it drove in a smooth, predictable way. It drove in a way where it's not like jerking and hitting the brakes, there's a lot of nuance in actually assessing whether the behavior is good or just didn't crash into something. And clearly, you want not crash in something, but you need to do more than that. And that's part of some of the interesting infrastructure we've built.

Very helpful. You've spoken about eventually driving to and from customer endpoints. Talk a bit more about what the key steps are the company needs to accomplish in order to be able to enable this and any sense on the timing?

Yes. This is 1 of the things where I think we probably shot ourselves in the foot a little bit in telling our story that driving to a customer endpoint is driving -- like driving 1 of our terminals. I just saw some analysis recently that said that 50% of distribution centers across the U.S. are within 1.5 miles of freeways, and 80% are within 5 miles of freeways. This is kind of where our stuff is today. There's -- we literally in going to our terminal site in Houston. We drive past 1 of our customers' terminals. And so it's been expedient for us to operate between terminals to date. Believe it or not, it's actually good for our customers at the senior executive level of our customers like you need to go to our terminals. If you're the operations person on the ground in that terminal, you're like, "No, do not put that f*** truck in my terminal right now while you're still learning", right? I have to get [ end ] trucks out of a day, and I've got to turn this over. I don't need some new technology messing that up. Now that it's actually working. Now that the customers can get it, we're ready to go, and we expect in early 2026 to begin operating to customer terminal sites -- customer sites, to be clear.

Maybe you could speak about the partnerships with the trucking OEMs. You've announced partnerships with Volvo and PACCAR. When does Aurora plan to start using trucks with redundant steering and braking included from those OEMs. And how was your view informed by the association to receive 20 development trucks from Volvo by the end of the year?

Yes. So we use trucks with redundancies from our OEM partners today. By -- in 2027, we expect our OEM partners to be producing trucks off of their line that have those redundancies in them. We've been working with them for several years to collaborate with them to ensure those, what they call autonomy enabled trucks, meet the requirements that we're going to need to ultimately operate them safely. We've also been working with them to line sight and install the Aurora hardware kit. And for those of you not as familiar with the story, we have hardware today that we've effectively built ourselves. We tested extensively. We're confident in the safety of that on the road. In 2026, you'll see us bring to the road hardware that's been manufactured by Fabrinet, that will have a major step down in cost relative to what we have today. And we're excited to see that come to life. And then in '27, you'll see the hardware that we're developing with Continental, which is again a huge step down in cost from where we are today. We'll produce that at tens of thousands of units a year with Continental -- or able to produce at tens of thousands of units a year with Continental. And what's really exciting about that is that hardware kit is going to be paid for through a hardware as a service model. So Continental is investing or spending $300-some million to codevelop this with us, set up manufacturing for it and ultimately finance this hardware. And then we will pay them back on a per mile basis. So if you're a customer wanting to use the Aurora driver, you're going to buy a truck that basically looks like a normal truck, and it's going to cost very similar to what a normal truck costs. And then you're going to pay a subscription to Aurora to operate that truck for you to be the driver. And that's going to include the hardware costs. So there's not some giant hurdle that a customer is going to have to overcome to actually bring this into their fleet. And of course, these customers are used to turning over their vehicles regularly, right? I think it's about 3 to 4 years in most fleets where they're going to move 1 truck out and sell it off to a second party and bring in a new vehicle. And so there's a really natural refresh cycle we see for our customers over time.

You said a little bit around the time frames. I think you said you're using trucks from those OEM partners with redundant steering and braking already, but it comes off the line in '27. So today, if I understand quickly, there's updating happening. Is that -- you're doing that with your partners, right, in terms of the upfitting?

Yes. So today in '26, we'll be upfitting trucks. So this is where we're going to be taking the overall hardware kit and installing it on those trucks for our partners.

And you put a blog post up on your website, you talked about how important it is to have the good relationships with your OEM partners and in PACCAR had asked to have an observer in the front seat. From your conversations with them and as you think about some of these considerations like when these trucks that are coming off the line and maybe the redundancy and breaking included on the line, what are some of those things that would have to happen to take the safety of observer out of the truck?

Yes. So to be clear, we are confident in the safe operation of our trucks today. We do have an observer sat in the truck, but they're just out there. Go check out YouTube.com [ at ] Aurora driver, a quick plug to the Aurora driver live. This is 1 of our ways of kind of continuing to demonstrate transparency, industry-leading transparency. You just go see our trucks driving down the freeway, what they're doing today in this moment. And what you'll see is for the ones that are operating driverless, like literally, our drivers will be sat there twiddling their thumbs or eating a bag of Fritos as things go down the road because they are not there to observe the safe operation of the vehicle. [ We're not ] there to ensure the safe operation of the vehicles, sorry.

Yes. So from our perspective, the trucks are working well. The observers aren't needed. Anything that you think would have to happen for them to not need to be there anymore?

No, I think it's once we have those production parts from Peterbilt in particular. And then the Volvos that we are receiving this year, Volvo has indicated those were -- those have the hardware that's necessary for us to ultimately operate driverlessly.

Maybe talk about remote assistance. How often are you meeting remote assistance? I mean that's always been part of your plan. It's in the long-term plan. I mean, any surprises on how often remote assistance is contributing?

No, it's kind of going to plan, right? That we are disclosing the rates of that at this time, but we've shared that there's an objective where you hit kind of economic knee in the curve for the ratios is about 1 to 20, and we're not seeing any concern on building towards that or building up towards that ratio.

Very helpful. You spoke about Fabrinet, you're going to be using hardware. They make next year in your trucks. Can you elaborate a bit more on what they're going to be doing and how impactful that might be to the BOM?

Yes. So we really believe deeply in working with folks who know what they're doing and can do their job better than you can do their job. And so Fabrinet is a fantastic contract manufacturer. So this is hardware that we have designed. We've done the DFM with Fabrinet. They're now manufacturing this hardware kit. We have early samples from it. The big objective for this harbor kit is to reduce the build materials cost so that we can ultimately get to unit economic profitable product.

Okay. Very helpful. And then you also spoke about working with Continental and making the shift with them in 2027. They had a lot of restructuring that they're dealing with at a corporate level. Any impact on the relationship with you?

No, not at all. In fact, I think it's actually exciting because I'm a big believer the companies when run well are focused on their mission. And so as Continental spins out AUMOVIO, which is going to be focused purely on the electron components and ADAS systems for vehicles. They're just going to be even more closely aligned to our mission. They've been very clear about this -- our work together is 1 of the tentpole programs that they have as a company. And they've been a tremendous partner. We continue to see that and really just love working with Philip and the team.

I think if I'm not mistaken, Fabrinet's second gen hardware kit. Continental will be doing the third gen hardware. What's different with this third-gen kit relative to the second gen?

Yes, there's a few things that are different. One is price. That's actually the biggest thing. We're going through -- one of the benefits of working with someone like Continental is they are used to looking at a bill of materials and saying, how do I make this for less. How do I make this like a mass manufacturer. And so having that ability around the table with us is going to lead to a major, again, another step function reduction in price. Architecturally, it's a little different. So today, we're using a conventional kind of x86 GPU type architecture in our computation. As we go to that continental generation hardware, we're moving to the Thor -- NVIDIA Thor SoC. So lots of computation, lower cost, lower power. And then the other big 1 is that our FirstLight Lidar, which is this proprietary Lidar sensor that allows us to see further than we think basically else can, is going to go from discrete optical components to an integrated optics on a chip system. And that really helps with reliability, manufacturability and cost. So we're really excited about that and we're -- parts of those systems are already up, and we're bringing up today.

And I mean, I imagine optics on a chip, I mean there's other partners like a Fabrinet would still be able to be involved or somebody else working with Conti to enable those changes.

Yes. So the chips that we're developing, these are chips that we've developed in-house, and then we worked with a fab to manufacturer. We're actually developing not just the lithography for it but also the deposition process. And even things like how do you -- because silicon is a really good passive conductor of light, but it doesn't create light. You need to actually bond it with at least 3 to 5 semiconductors that can actually produce light. So how do you flip and align those chips with -- like there's lots of cool things that are part of the tech infrastructure and tech stack that we've been building that our IP that we have at Aurora. But of course, we do work with a variety of other suppliers to enable us to manufacture stuff at scale with Continental.

And Continental has a lot of experience in the industry working with OEMs and so certainly understand the benefits of that partnership. As you think about the trucking partners starting to manufacture lineside redundant steering and braking, is that something that Continental is going to do? Or are you guys going to do that even before Continental begins to take over some of the...

Yes. So our OEM partners are going to have their suppliers for the redundancies in the braking and steering systems in their trucks, right? And the way really I think about this is there's a truck and a driver. And so the drivers, the software, the computer, the sensors that allow that driver to see the world. And then the truck is all the stuff you think of a truck steering, braking, power, engine, right? And so the OEM is responsible for those components today.

Okay. You've spoken in the past around a bigger ramp in volumes in the late '27, '28 time frame as Continental begins to come on. You're already working with Continental on the third-generation hardware kit. Do you have a sense Continental is already engaged with the trucking partners as well such that you could have this sort of faster ramp in that time frame?

I don't just believe it. We're actively involved in those conversations, right? That's -- we're the one developing and designing this hardware kit. And so we're the partners with Continental and with our OEM friends as well. So 3-way conversations in both cases.

That's great. Maybe talk about the regulatory environment. I know you've made the point many times about how -- where you're operating, it's very favorable for autonomous trucking. Speak a bit more on some of the more recent developments from a regulatory standpoint and how Aurora has been advising and partnering with some of the government officials.

Yes. Again, the environment today, we can live with and work with. We have the ability to put trucks on the road wherever we feel confident in the safety of doing that across, I think it's something like 40 of the 50 United States. So that feels great. What we are seeing is continued enthusiasm at the state level. When we launched, Governor Abbott in Texas, congratulated us publicly, right? That's not normal, right? So we really appreciate the government of Texas' support for this technology. And then we're really seeing a lot of support at the federal level. This administration has really been forward on automated vehicles, whether it's Secretary Duffy's comments in his confirmation and since, whether it's Vice President of Vance's comments around the importance of automated trucking to the United States. Just yesterday, I think former Secretary and Governor Perry had an ad that came out in support of this technology. So really, we're seeing a lot of support there. And then just recently, representative Fong, a Republican from California, put forward the AMERICAN DRIVES Act, which really is -- puts forward legislation that would create a federal framework and preemption, which we think would be really powerful in helping the U.S. stay on the forward foot with this technology relative to overseas folks.

That's very helpful. Maybe speak a bit on the market and some financial topics. You spoke about having trucks on the road and how that's been very positive as you're having conversations with customers or prospective customers. We've also seen some weakness in just the trucking market at a broader industry level in terms of the trucks driven by people. Talk a bit more around what you're seeing with the market and how is that affecting your business?

Yes. I think we're building something strategic for the long term. And there's just going to be a shortage of drivers in the U.S., right? We expect to be -- it's not we -- American Trucking Association expects to be short of 1 million drivers over the next decade. And with the current administration's immigration policy, that probably is going to be more steep. If you look at the average trades of a truck driver in the U.S., it's 55 years old and it's increasing. So yes, there's been an unusually deep kind of down cycle in trucking coming out of COVID. But as we look to the point where this technology is going to begin to scale, the benefits far outweigh any challenges to adoption, right? As we look at our customers, their ability to significantly increase their margins, dramatically increase their revenue, improve safety and really build their business. Like if you're not using our stuff in 5 years, I don't see how you're going to be competitive as a trucking company. So we're -- we love to work with these partners, and we love to help them go win.

At the 2024 Investor Day, the company spoke about ASPs for its Driver-as-a-Service model being projected in the $0.65 to $0.85 per mile range. Since then, there's been tariffs, there's been broader inflation in the global economies and in the U.S. As you reflect on some of the things you're observing in the business and some of the momentum you have with your technology, what does that say about this pricing projection from your last Investor Day?

Yes. When we look at the cost of labor for driving trucks, it's basically at $1 per mile at this point. And I don't see that coming down. It's just a question of how rapidly is it going to go up. And so we feel very confident in that pricing window and expect there's opportunity for that to go up. But we also want to make sure that we are good partners and helping our customers see the benefits of this technology and build their businesses.

And besides just sort of the cost of labor, I mean, you've also talked about some opportunities around insurance and fuel, right? And so I mean the broader savings you could bring to your customers beyond that $1.

It's gigantic, right? I apologize for -- for underselling it. When you look at the fuel economy benefits, we expect between 14% and 34% improvement in fuel efficiency. Given the cost of fuel for these businesses, that's big. Given the environmental sustainability benefits, that's also big. When we think about insurance over time, this will be a better driver than people or a safer driver than people. And so we expect that to drive insurance costs down. For our customers today, it's already a win, right? Today, they are responsible for their people driving the truck. If they do something wrong, they're the one who are going to write the check. We immediately kind of alleviate that risk from them. So that's a meaningful benefit in and of itself. And then the fact that we have the data that comes along with whatever events may occur on the road means that we have the ability to quickly adjudicate that and ultimately get to what will hopefully be the right and reasonable answer of whatever occurs.

An autonomous truck has more hardware that goes on to it compared to a traditional truck. Help us better understand the evolution of the BOM. I mean you spoke about Fabrinet and Continental and them contributing to improved cost structure. But is there a cost point where the BOM needs to hit for EV trucking to really take off?

I think we are on the cusp of that. Our expectation is that the generation of hardware that will come from Fabrinet will actually enable positive unit economics for us. Today, it's really about how do we set price relative to our cost. And then there's other things we have to do. We have to continue on the path that we're on with reducing the rate of remote assistance and support and get that into the neighborhood we're anticipating. We need to look at the rate of recovery on the road and continue to drive that down, but -- and then increase utilization. And the more value you can provide to customers, the more it's going to be worth to them and the more we can charge for it.

And you spoke a little bit around over time getting to 20:1 ratio on remote assistance and tracking to your prior expectations. As you think about that variable? Is that going to be tied to some of these future technology releases like second or third gen, or is it more just as you see the trucks driving on the road?

It's really going to be around the performance of the software system, which is kind of independently evolving or parallel evolving to the hardware generations. And so at this point, we've released several versions of the Aurora Driver since launch, one major release, which was the update to be able to operate at night. And so we'll continue to see those incremental improvements in both the Aurora Driver, but also in the Aurora services layer that sits above the truck, right? This is effectively how our customers ultimately will interact with that truck.

You mentioned with the second-gen hardware kit that allowing for positive unit economics. So just that's something that would allow you get to positive gross margin. Just to be clear, that's what you were referring to.

Yes.

And is there a certain miles per week that we need to have in mind as we think about that utilization where you can start to cross over?

No. I don't think I'm going to share anything more precise than we want to get to that 200,000 to 250,000 miles a year. And it's quite attainable. .

Well, it's really interesting, you start looking at some of these routes, right? I mean you just talked about the Phoenix, right, that's already longer than a human can do. And so you can actually get to very high numbers of miles traveled in a year as you start adding more of these routes and nighttime driving and different weather.

So you do the round trip from one-way leg each day from Fort Worth Phoenix of 1,000 miles. You do that every day, that's 365,000 miles a year. Now do we do it every day? Probably not quite. But you can see that these numbers are very approachable. Even a short trip like Dallas to Houston, if you do 3 round trips a day, you're already well over that threshold.

Yes. We did a big AV report a few months ago, and we dug into insurance and BOM costs across the broader autonomous space. But one of the interesting things as we're doing the modeling was how important that variable was just of miles traveled. And for trucking, right, you can see how it quickly starts to...

Massively heads up.

Yes. Just lastly, just free cash flow and capital allocation. Maybe just talk about how you think about managing the balance sheet. You've done a few raises so far. You talk a bit about. I'll just remind investors what your thoughts are before you reach positive free cash flow.

Yes. So we left Q2 with a very strong balance sheet, $1.3 billion. We feel better about the capital position basically than we've ever done. And we're well positioned. We've been clear about the fact that we're going to need to raise incremental capital at some point. We're going to continue to do that in a thoughtful way where it meets the capital needs of the business and doesn't put the long-term goal at risk, but doesn't unduly dilute our existing investor base. So we'll continue to be careful about it.

Great. Well, we're approaching the end of the session. Chris, really appreciate you joining, and thank you for coming.

No, glad to be here. Thank you.