Ambarella, Inc. (AMBA) Earnings Call Transcript & Summary

Okay. Mic'd up. All right. So thank you, everyone, for joining us. For those of you who don't know me, I'm Aileen Smith, I work alongside John Murphy on BofA's U.S. autos team. We are going to have Ambarella kick off our second track officially up here. The Ambarella is covered by our U.S. semiconductor analyst with [indiscernible], but from our perspective in the automotive industry, it's one that we are keeping an eye on, it's increasingly relevant to our space. Specifically, Ambarella is a fabulous developer of low-power processors for applications where high bandwidth sensors like cameras and radars are collecting data, their systems-on-chip's technology, integrate compression, advanced image processing and deep neural network processing to enable perception, fusion central processing and to extract valuable data from high-resolution video and image radar streams. From Ambarella, we are very excited to welcome back Feng-Ming Wang, who is the company's Co-Founder and Chief Executive Officer. He is a seasoned company leader. He's an entrepreneur, a video compression technology expert, and he holds several digital video patents. And we also have from Ambarella, Brian White, who is the Chief Financial Officer. He's a couple of weeks into the job, so we won't grill him too much up here. But Brian, thanks for joining us as well.

So Fermi, welcome back. I think a really good place to start, and we've asked this a couple of times at this conference in prior years, but to help set the stage for our audience, can you talk a bit about what is computer vision technology and why are cameras so important to efficient data collection?

Right. So computer vision is really about capturing video, then you analyze the video real time to understand the environment, right? In the past, in fact, the computer vision has been there for many, many years, traditionally, people doing it, for example, if you want to do face detection and license plate detection, you have 2 totally different algorithms and make it almost impossible to build a signal processing platform that general enough to cover all of the different type of object segmentations. So computer vision was very segmented for a long time. Until the deep neural network approach was published in 2012. In fact, neural network was not a new thing either. Why was the PhD student in Colombia in 1987, neural network was there. The problem was that there was no hardware fast enough to implement neural network until 2012 and then people started publishing great result based on neural network for computer vision. That's where people realize that using a neural network approach, you can do any kind of object detection with the training data and using 1 network can probably can differentiate multiple different objects at the same time. So that's where it becomes possible to build silicon architecture to address all the possible different computer vision tests, including object detection, because segmentation for other high level of digital video processing. When that happens, that make interesting for a company like Ambarella that we can create unique single -- silicon architecture to take advantage of neural network approach and deliver the best performance and best power efficiency that we can for the customers.

So when we think about the applicability in particular, for Ambarella and the space within autos that you guys occupy, active safety, advanced driver assistance systems, Level 2 autonomy to ultimately Level 4 fully autonomous capabilities. Can you talk about why computer vision is the ideal solution for those capabilities?

I think computer vision definitely has become viewed as one of the most important sensor modality because it provides some information that cannot be replaced by any other sensor modality, for example, color. For example, we identify the characters and the numbers on the signs or on a license plate. There are a few things that camera can do that no other sensor modality can do. So that becomes so important for any autonomous driving vehicles to start using video. But also, it represents a huge challenges because if you look at today, Level 2+ car, people talking about 10 cameras, each one of them can be 8 mega pixels at 30 or 60 frames per second. When you multiply all of the cameras input to a chip, we can easily talk about 100 gigabits per second of video data into a chip and to process all those information to do video processing in computer vision and in real time, that really requires a huge amount of computation performance to do that.

So you bring up a really interesting point around the applicability of cameras and the sensor modality in particular. And I want to ask a question, Elon Musk, who we are all very aware of in the industry. He's been pretty vocal that he believes full self-driving capabilities will be entirely solved for by cameras. He's a big proponent that LIDAR is not necessary. He's actually pulled radars out of his vehicles. So as you think about Ambarella and computer vision technology, would you subscribe to that view generally that most of the capabilities and the technology can be solved for by camera? Or how do you think about the applicability of other sensors?

Right. Well, I do admire the achievement that Elon Musk has achieved in the last several years. And also I agree with many things that he's talking about, particularly on the video processing side. But at the same time, for Level 2+ car, I do believe you need multiple sensor modality to achieve the redundancy as well as the safety requirements required by the -- not only regulation, but just for the safety of the vehicle and the passengers in there. The way we look at the sensor modality is we do believe RADAR is the best match for cameras, for many different reasons. The most important reason is that when we look at the challenging condition for cameras, like low light or bad weather. In fact, because LiDAR using the same technology, which is optical sensors, some of the problems that exist in the cameras also exists in LIDAR. But however, RADAR definitely provides the best complementary technology to video because those problems we talk about that for the camera do not exist in the RADAR side. On top of that, I think one of the reason Elon Musk complained about RADAR performance is really the existing RADAR technology, the performance in terms of distance, in terms of angular resolution, in terms of the point cloud you can build is really weak compared to LiDAR. But however, with the latest 4D tech image radar technology that we saw and that the company we acquired Oculii, we acquired. When we saw the latest 4D image radar resolution particularly on the distance and the angular resolution, we can achieve LiDAR type of a resolution with RADAR technology. So with that, we definitely believe that the video and the RADAR combination can address the big portion of Level 2+, if you -- particularly for the segment that are cost sensitive.

Yes. I do want to dive into that for a second, which is the Oculii acquisition and specifically Ambarella getting back into -- or getting into RADAR. It seems like there's really 2 tenets or motivations for the deal. One is the expansion of the serviceable addressable market as you get into RADAR but then there's also the deep fusion aspect, which actually seems to be the more important aspect of the deal of using RADAR technology with computer vision technology. Can you talk a little bit about that?

Absolutely. I think you're absolutely right. A lot of people looking at the reason we acquired Oculii think, oh, you expand your TAM, which is definitely important. But 3 years ago, when we start doing -- working on CV3, which is important technology that we do a domain controller, and we try to decide whether we need to integrate a different sensor modality into the chip. And based on our algorithm study, we convince ourselves when you do a sensor fusion at a deep level, today, majority of the automo-driving car doing sensor fusion between video and radar at objects in this level because they process these 2 data type totally independently and try to merge at [indiscernible], which I think is most -- is not efficient. We propose that the sensor fusion should happen at the point cloud level, both video and radar at the point cloud level. And if you can combine this point cloud and do a training and inference on that level, you will get a much better accuracy in terms of the -- all of neural network performance. So based on that understanding, that's why we acquired Oculii and also, it's our intention that our CV3 family chips, we will offer the integrated video and radar solution at a deep fusion level to provide the best possible sensor modality solutions to our customers.

Great. And I do want to get into the CV3 family, which you unveiled at the Capital Markets Day. But maybe before then, if we think about the last time you and I had a conversation at this summit a couple of years ago, the focus at Ambarella was really AI processing for both external and internal cameras in the vehicle. But in the past several years, there's been a big expansion of the company in providing the processing power not only for the cameras, but also radar, the fusion and the planning levels down to processing on a single chip. Can you talk maybe about the expansion of the product set, how you can scale up from what was focusing only on computer vision to now a much broader suite?

Right. So 2 years ago, when we met, we just started ramping up our CV2 family products. And CV2 family is really good at video perception. And with that, we started building a lot of our customer base based on that. However, by dealing with all those customers, we also realized that video perception is important, but people need a more integrated solution for all the possible applications out there, not just auto, but also other applications like commercial slow-moving robot, even security cameras. They want to have a different sensor modality and have a chip can integrate the sensor modality and fusion and the control. So based on that knowledge, we're building a family of chips that we hope in the future, any -- not only just auto, any robots, which would require motions or moving well can leverage on the family we're doing providing both radar and video perception as well as all the high-level software stack, which is important for our customers.

Great. And maybe diving in now to talk a little bit about CV3, which was a big announcement earlier this year. And that's the next-generation SoC family and your specific AI domain controller for automotive ADAS to AV. Can you talk a little bit about the product family, the major leaps in the technology from computer vision to now an entire domain controller system and then how that fits into the commercial strategy with respect to your automotive business?

Right. So like I said, when we start looking at CV3 family 3 years ago, the biggest challenge for us is to try to understand the application, what kind of software and the algorithm we need to implement for autonomy driving. And that's how we decide we need to get integrated radar. And also we know that we need to serving all the high-level software stack. In order to achieve that, we need to really leverage what we know on the CV2. For example, let me give you an example. Rivian is using our CV2 family chip for Level 2 cars -- for the Level 2+ cars. And Rivian design, we only serve on the video perception. They use multiple of our chips for different video perception applications, but they use other components to do higher-level RADAR solution or higher-level software processing. It's our ambition to integrate all those components into 1 chip, which is CV3, and we think we'll achieve that. But in addition to that, in order to achieve so much performance, for example, our CV2 computer vision performance compared to CV3, CV3 is 40x higher than CV2 in terms of CV performance. Another important thing we improve is on the CV3 is our power efficiency. Our CV2 power efficiency is already much better than our competitors. But when you compare CV3 to CV2 power efficiency, we increased it by 3x again. We really believe that for -- not only for automotive, just for any computer vision or I should say that mobility solution that require because all of them will be enabled by the batteries and try to make sure you have lowest power consumption, become a huge differentiation for us. So that, for example, in CV2, we show that one CV2 family chip running with some of the neural network performance is 4 watts versus our competitor solution running identical workload is at 20-plus watts. That power efficiency is not only duplicating CV3, we make 3x better compared to CV2. I think there are just 2 examples that why CV3 is so important for us. But also, I want to emphasize because for CV3, CV3 is not a chip, it's a family of chips. The idea is to build a family chips from a very high performance level to lower performance level so that our customers can use 1 family of chips to build different performance on their product road map. And not only that, they can leverage the same software among those products they want to build that they can really significantly reduce their software investment and also our support requirement. If you use a same software architecture on the sensitive on the CV3 family chips. I think that's another benefit we're offering to our customers.

You bring up a really interesting point around what I'll call the scalability of that model. And the focus in the past with respect to sensors has really been on the power efficiency side. But as you have an entire suite or an entire family of chips that you can offer your customers grading from the low end to the high end. Can you talk about the customer traction for the CV3 family as you've engaged with them over the past couple of years? And then more formally unveiled that product and how important the scalability is to that discussion?

In fact, the CV3 is not the first family of chips we built. We built -- in fact, the CV2 family that we introduced 3 years ago was a family of chips. And in fact, we have 6 chips in that CV2 family going from the low end to high end. On the high-end side, the highest performance CV2 family is sold around $100 and lower side selling at single digit. And I can tell you that most of our customers, particularly we use CV to sell to the security camera. All our security camera customers take advantage of this road map because they also need a family of products. And with this approach, they use 1 software implementation for all their products in this family definitely help us to not only get better traction and also sticky customer relationship. We -- CV3 we're trying to do the identical things. In fact, when we present our approach, presenting a CV3 family to can go as high as to do that what we announced on CV3 high, which is 1,000 TOPS chip, which can do Level 4 car, and our idea is to scale all the way down to 20 to 40 TOPS chips that can do a very advanced single camera ADAS market, and we can use the same software structure for all of that. And you can imagine that there are multiple chips in between. So this family chip that I think our customers believe if we can deliver on that, they can have 1 software stack and address all their markets. I think that's become a very powerful message and a powerful selling point for us to the customers.

Yes. And maybe segueing that into kind of the competitive landscape and where some of your peers stack up. Obviously, on the semiconductor side, we're all aware of some of the large incumbent players, not specifically for ADAS active safety, but everything else in the electronics of the vehicle. You've got the [ Renesas ], the Infineons, et cetera. But specifically, Specifically, in the automotive ADAS active safety space, there's 1 pretty large dominant player in the world on visioning technology. There's newer players that are coming out, Qualcomm with Veoneer and their River brand. You've also got Qualcomm -- excuse me, NVIDIA that's coming into the space as well. Can you stack up where Ambarella sits versus some of the other major players in the computer vision and the chip industry?

Right. So today, I think the majority -- the -- our biggest competitor is really NVIDIA, Qualcomm and Mobileye. And all of them has -- including us, all of it has pros and cons.

Sorry about that. I think we've got a New York alert coming out.

So I think the when we look at how we compete with those competitors, we mentioned 1 important thing is really power efficiency. Power efficiency is not just against one competitor, it is in fact, against all the competitors and we are significantly below in the power consumption. In fact, when we presented the power consumption to our customer, it's not just because you get a better battery life, it's really about one of the customer told us that every watt of power saving represents $2 or better cost saving. If that's the case, we are representing hundreds of dollars of power battery saving for the customer if they want to reduce the cost on the battery side. So what we are trying to translate is power consumption means dollar, particularly on dollars for our customer. And by using us not only they save on the power, they save on the cooling system, they also well save on the RADAR system because now we integrate all the RADAR processing into our chip. So if you add all that together, that's really our position is really because we have better technology, we design a chip specifically for this particular application. That's why we can achieve this kind of power efficiency. That's how we differentiate. We are not trying to use or repurpose a GPU or a processor to this application. I think in the future, when you continue to try to scale the performance especially, you want to go to Level 4, Level 5 cars, you need to scale the performance by another 4x. You cannot come down just increase the clock rate on your general purpose CPU, GPU. You need to design a specific hardware to address those so you can achieve those kind of performance without increasing the power accordingly.

Okay. And when you think about the competitive advantages and disadvantages in the market more broadly, one of the competitive advantages in the space is obviously being on production model. So some of your peers do have that, that they are with the customers right now. As you think about the technology and how it's going to evolve and specifically moving from Level 2 plus, to ultimately Level 4. Is there a competitive disadvantage in not being on production models right now or is it rather a demonstration that the technology that you're offering is superior, and therefore, there's really no disadvantage for you?

Well, it's definitely a disadvantage, right, because people want to look at track record. However, I would like to argue, first of all, we do have a Level 2+ production with CV2 like Rivian. And it's a production with the EV car. And also, we have multiple ADAS and Level 2+ car in production in China. So we do have a track record. But in addition to that, if you look at our traditional automotive business for the recorders, Toyota, Nissan and some major Japanese companies, in fact, also that BMW is our customer in the [ recorder ]. So we do serve those major Europe, Japanese and U.S. automakers on the recorder side, that only help us to establish our quite ability as auto supplier because those guys are really looking at is whether you understand the quality requirement and you can support them with their business model. And we have proven that we can do both with the different product lines.

And does having that customer in on the recorders and the data loggers, open the door easily to the conversation around your more advanced computer vision chips like CV3 and what you can provide from an active safety and autonomous driving perspective?

Definitely it help us to open the door. But however, at the end to close the deal, we still need to show them that the technology works. So we are going to sample our CV3 chip second quarter this year. But before that, in the last -- you can imagine in the last 6 to 12 months, we have a demo to our customer with our CV2 system that our customer can really evaluate with the real hardware as well as the CV3 simulation system, they can see how to scale the CV2 performance to CV3. So with CV2 system, they confirm our power efficiency. And then on CV3, we tell them how to scale it. And basically, they are waiting to see whether we can deliver that with a real silicon. But I think that in the last few months, we have convinced our customer when we deliver that CV3, we have the best technology in terms of performance and performance efficiency that our customer wants.

Okay. Great. And maybe to ask 1 question before we open up to the audience and see if anyone has questions. But you've mentioned it a couple of times on the commercialization side and specifically the partnership with Rivian on the Level 2+ product. Is it -- that is not the only announcement that you've made recently for some newer entrants in the automotive space. You also have a partnership with Motional, and I believe Arrival as well. Is it fair to look at the partnership with Rivian as obviously a series production contract versus the partnerships of Motional and Arrival maybe more R&D spot buy in nature or R&D projects? And how do you convert those types of projects into eventually serious production contracts?

Well, I do believe that the intention for Motional and Arrival is they're going to become production. It will just take time for them to get there. We started engaging with Rivian 4 years ago. It took the 4 years to go into production. It just takes that long to build a Level 2+ car these days. And I think eventually, Motional and Arrival will get there. It's just taking time for them to really go through all the qualification process. I think that at the same time, all of the current generation Level 2+ cars, it doesn't matter whether you use a NVIDIA or us, it's a multichip solution in there, right? If you opened up with Rivian, you see multiple silicon to try to serve different applications inside the car, even with NVIDIA chip, they still use multiple chips to a Level 2+ car. So the next phase is really when the -- all the Level 2 car companies try to decide the next-generation level to design wins, they are evaluating based on single-chip solution. And that really creates a different requirement and different opportunity for us. It's not a multichip solution anymore. You have to show you can do a single chip with Level 2+ design, that I think really require an integrated chip like CV2 -- sorry, CV3. And -- but of course, that [indiscernible] and Qualcomm is proposing some of the different solutions, that's where we need to compete with them.

Got it. And we'll open up to the audience and see if we have any questions. If not, I'm certainly happy to keep going. All right. So maybe switching gears a bit and kind of getting into some financial questions and specifically the targets that were provided at the Capital Markets Day. And one of them that I really want to focus on is the automotive sales funnel. And specifically, it seems like there's been a lot of progress on that front over the past couple of years. I think the disclosure at the CMD was $1.8 billion over the next 5 years. The last time you disclosed that, it was $600 million for the preceding 5 years. Can you talk a little bit about where most of the incremental business wins are coming from, from a product perspective? And then maybe from a customer perspective, is it a few large major contracts that you've locked down? Or is it progress across numerous fronts?

Right. So this is usually my time to refer to my CFO, but he's only 2-week new. So I would -- if you don't mind, I will take over this answer. We knew that the reason we designed this sales funnel and make it public to the investor is because we knew that a lot of things we're doing, the design win we have done, we cannot talk about it. We couldn't talk about. For example, Rivian, we work with them for 4 years and only when they went into production then we start talking about the design wins. So the ideas of a sales funnel is really provide a continued update about our progress that we made in the automotive space. When we disclosed $600 million in the first time that was 2 years, I would say, the end of 2020. It was $600 million, basically, in fact, in there, we talk about this is accumulation with 6 years sales funnel that we built. And a portion of that is the design we already secured. Another portion is the project we are building on. And for the project we're building on, we take multiple haircuts on that. We put -- our judgment call or our confidence how much whether we can win this project in terms of percentage. We're also putting a haircut in terms of who they're going to move -- when they're going to [ SOP ] and also sometimes even what's the total volume of the customer will take at the end. So there you can say that there's a lot of judgment call in there. But however, we try to represent that number in a realistic way, it's not just adding all the possible opportunity together, but we will try to tell you based on the multiplier that we show you that a realistic number. So the major difference from the $600 million to $1.8 billion last year was a few things. First of all, the portion of that is definitely a continuation of the promotion of CV2 for data loggers for the ADAS market and also the DMS market. However, there is another portion which majority is still on the design win level is CV3. We start engaging CV3 design into many different OEMs and Tier 1s. That also represent a portion of the funnel of the $1.8 billion funnel that we talked about last year. If you look at the $1.8 billion, it really represent the -- from last year for the next following 6 years, how the revenue we can generate from automotive. But it's not a straight line. It's more like exponential curve because the biggest variable, of course, that the total unit number is important, but I will ask you to focus on more on ASP side. Today, Ambarella's ASP is roughly in the ballpark $10. But when we go to Level 2+ car, CV3, we're quoting at anywhere between $150 to $300 depending on performance. So you can see that the scale of ASP is the reason that we get a bigger dollar size in our sales funnel. That's definitely one of the reasons.

Historically, in the automotive industry, there's actually an inverse relationship between penetration and units versus ASP. As you get penetration in units going up, ASP, you're democratizing the technology or standardizing it, that comes down. But in this particular case, you kind of have both going up at the same time, which is the applicability of those sensors and the capability that they're providing. Is there another angle within that back-end loaded or let's call it, hockey stick type trajectory for the $1.8 billion in forward sales funnel revenue. How much of it is the units? How much of it is the ASP perspective versus what may also be Ambarella gaining share in the industry maybe in some of the outer years?

I think it's all of the above, right? I definitely think ASP is one significant because when you scale your ASP from $10 to $150, that just help you to accelerate your revenue much faster. And also, we believe we model that at $600 million, we represent 2% of total TAM; at $1.8 billion is represent a total of 7% of total set [ TAM ]. So that also shows you that we are assuming that our market share improved over years. Then the unit number is really about associated with the market share also. So I really think that the increase in market share as well as ASP play the major roles of our sales funnel.

And maybe to get into that market share estimation of 7% in the outer years, is that a target for Ambarella? Is it based on current rates of winning? Or is that kind of, let's say, starting point for where you could grow market share in the future?

The 7% is really representing $1.8 billion divided by total TAM. So basically, we've gone through this exercise to look at the sales funnel and then that's $1.8 billion and look at all the potential TAM in the next 6 years, the ratio is roughly 7% at the end of last year. So that's where we get to the 7% number.

And do you have any estimation for market share based on programs that you're going out and bidding for and what you're winning?

That -- in fact, that all the things you just mentioned is part of our sales funnel. When we build the sales funnel we look at how many projects we can win and at what time, at what cost and what's the probability. So everything -- all the variable you are talking about is embedded in our sales funnel calculation.

Okay. And then, again, to kind of stick on the sales funnel because it's important for us as financial analysts. You -- last time you disclosed it, I think it was $700 million of that sales funnel was actually booked, signed and awarded contracts. And there's another $1.1 billion that is in some form of contract stage. Can you disclose to us where it sits, whether it's an RFQ process, RFI or maybe even pre-RFI?

No, I think it's all RFQ. This is after RFQ and the [indiscernible]. That's where you can have information about the unit number. Without a unit number, you cannot build the model. So -- and this you have to really talk to your customer to a level that you understand the range of the ASP and also which model will be done. What's the annual rate per year for the multiple years. Unless you have that information, then you can start building a model. So all of that is after the RFQ level. And some of the -- of course, there are some is closer than the others. So you can say -- we can say that we're only 6 months after the last projection, we're going to give another update in November this year. So you can see -- I definitely believe we'll continue making progress on those RFQ we talk about. And those $700 million design win that we already have is that we received a letter from our OEM customer, "You're reward -- awarded for this project for how many years, what models and times." And also, we know there's active engineering on both sides to work on that then with the category that kind of design win as a win in our sales funnels.

Okay. And we have a few more minutes. I've got a few more questions, but I want to make sure the audience has been satiated as well. So are there any questions in the audience? All right. So one in the back.

Go ahead.

Sorry, I don't know where the mics are? Yes. Go ahead.

[indiscernible] What could we expect going forward? Has something like this happened in the past in any way? And how are you assessing the situation?

Right. First of all, Samsung is a very close partner to us. In fact, I don't know whether this is -- you heard about this. We probably are the only semiconductor company exclusively using Samsung today. I don't think there's any other semiconductor company can say that. And also, we started working with Samsung many years ago. In fact, I think for people who follow Samsung, their first mass production process was 45-nanometer, and that's the process is only designed for Apple. In fact, for 45-nanometer process sensor process, they had 2 major customers, one is Apple, one is Ambarella. That just show you that our track record with Samsung since 45-nano, we have been using that exclusively through many different generations. And with that, we really built a strong partnership with Samsung. And -- but we did talk about we have a shortage probably from Samsung just last quarter. That was the first time we deal with Samsung with this kind of problem. Of course, that comes a surprise, and we are trying to work out -- find a way to talk to the top management at Samsung to make sure that this is a one-off event and making sure that we understand why it happens and how we can resolve this problem quickly. And we talk about -- when we announced this problem, we talked about Q2, Q3 will be impacted. Hopefully, Q4 will be recovery. We still believe that's the case. And at the same time, we are trying to continue to work with Samsung to make sure that a similar situation will happen again. Again, I really think that today, if you look at the potential partner on the foundry side, it's either TSMC or Samsung, maybe you can throw Intel in there as a possibility. But -- so having a very close relationship with the foundry is an important strategy for the company.

We got 1 more up here.

[indiscernible].

We would definitely consider that in fact, we talked to them 6 years ago when they are trying to become a foundry supplier. We evaluate their technology at that time extensively and just because we are in this business, it's our job to make sure we understand all of the foundry situation and making sure we make the best decision for the company. I think Intel's new offering, although we haven't really spent time to evaluate. But I think that in my view and based on whether I think that TSMC and Samsung is still ahead of Intel. And I'm definitely looking forward to have another opportunity to engage with them when they can share more information on the new foundries they are building. And I think that's probably where hopefully, that can give us another option to consider.

Okay. And maybe 1 last one to close it up here. Another big announcement that you made earlier this year, I think, was the resegmentation of the business previously at security, automotive and other. Now you're just going to have automotive and nonautomotive. Really what does that say about the growth strategy at the company and where you're focusing your efforts? And as you look out at the business 5, 10-plus years from now, really where do you think automotive will occupy?

Right. So thank you for that question because a lot of our investors still remember us as a GoPro supplier and try to make sure that's not the case anymore. In fact, in the -- after IPO, we understand that for us to become a very a semiconductor company with scale, we need to go to a market that give us the potential growth. Security camera is big enough for that. Automotive definitely is a market that we need to conquer. But all the other consumer market that we used to have is so small and the dominant buyer is just 1 or 2 customers, and we try to work away from the market as much as we can. But of course, the opportunity will still -- if there's an opportunity comes, we still serve them, but we don't design chips, particularly for those markets anymore. So we -- our focus moving forward is really providing computer vision chip targeting today, security cameras and also the derivative of security camera plus automotive. But in the future, I think the bigger opportunity for us is really let's put all the robotic application into one package. I think that's going to be a huge [ opportunity ] I also believe any robot which does require to move can use a CV3 family chip to the bank controller because anything need to move, need to be able to have the same problem in autonomous driving. You need to see around, you need to make a decision how to drive just like autonomous driving cars. So in my view that all our investment for the automotive will become a viable solution for robotic applications.

Fantastic. All right. I think with that, we're out of time. So in the interest of getting everyone off to their next meeting, we're going to close it here. Fermi and Brian, thank you so much for joining us today and supporting us in the summit again.

Thank you.

Thank you.