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

Welcome back, everyone. I'm Joe Moore from Morgan Stanley. Very happy to have with us today the CEO of Ambarella, Fermi Wang. Fermi, thanks for joining us.

Thank you.

We need to read real quickly this hedge. For important disclosures, please see the Morgan Stanley research disclosure website at morganstanley.com/researchdisclosures. If you have any questions, please reach out to your Morgan Stanley sales representative.

So Fermi, thanks for joining us today. Maybe if we could just start out about the transition that you guys have been on, 45% of your revenue last year coming from computer vision, continuing to move up to 60% this year. More recently, you've sort of talked about this computer vision sort of transitioning to more of an edge-based AI company. Can you just talk about where you are in these transitions?

Right. So the reason we talk about edge AI instead of computer vision because in the past, we focused purely on the video processing and the computer vision is the AI technology for the vision. We are focusing on domain controller for not only automotive but for general other applications, at the same time, we add radar into our product road map. So certainly, we are -- computer vision doesn't cover everything we do. So we believe that we need to give a more accurate representation of what we are focusing on. So AI is the one that will come out and the transition is really not only about technology and the market, it is really about focusing on creating a total solution that we think is important for our customer moving forward. In automotive, you need domain controlling in the radar. In fact, I can say that radar going to go to IoT space also. So in the future, I think that this AI for different sensor modality plus high-level processing is going to be the area we're going to continue to focus on.

You've really transitioned this computer vision analytics market around deep learning and AI. And I guess that seems to me to be a real advantage when you talk about deploying this stuff that some of your competitors either are using programmable silicon or are using custom silicon, this is designed for more of a heuristic approach to solving these problems. You guys have embraced the underlying software transition from day one. So can you maybe talk a little bit about that software first mentality you have?

Right. From day 1 of the company, which was starting 2004, we knew that we have to generally optimize the silicon architecture for the application we're focused on. We call this algorithm, first of all, software-first approach. What we did is we looked at the algorithm and software we need to implement for that particular application. And after we optimize the software, then we come back to look at how to implement those software in a silicon form. And for you do this iteration, not only you can achieve the best performance because you're focused on your algorithm software, but at the same time, when you come back to optimize silicon, you can optimize with die size, you can optimize with power consumption. That's another reason that we can -- really showing up with the best performance -- AI performance today with the lowest power consumption.

And I guess deep learning has obviously been in headlines a lot this year outside of the vision space because of what's happening in natural language, large language models, ChatGPT, those kinds of things. I know you guys have always talked about transformers being an important element of your architecture. Can you just talk to how you apply that same methodology to your markets?

Right. In fact, that transformer is really a new network architecture proposed by Google in 2017, become popular in 2018. And we started paying attention to that because, of course, that's a brand-new architecture then, and we found out that Tesla's going to use that for their -- in the FSD software, so we started looking at how to implement transformer early on. And today, we believe that we are one of the very few suppliers that can really provide a very efficient transformer network implementation on the silicon form and the demo at real time. In fact, at the CES, we demoed the transformer implementation with a customer -- with customer transformer software running on our chip and also we have demoed our own transformer software running on our chip. So that really shows off that our application space hardware has enough programmability and I can change it to a different type of network architectures.

And it seems like what you're really -- this is above my knowledge base a little bit. But when you talk about using transformer inference -- you guys are really an inference company and transformer adds material complexity, multipass to that inference. How widely applicable is that? I mean, can you apply this to other workloads outside of vision?

I think that our current silicon can run any transform-based application. It matters how efficient it is. And you know that all the large language model today and also some of the computer vision software like stable diffusion, all of them using transformer network in their application, although they are not really for autonomous driving or IoT space. We believe that we can do well in terms of performance efficiency, but we need to look at two things. One is we need to verify that we can -- I think we can run, but we need to verify the performance we can get using -- running the [indiscernible] on our silicon. And second thing we look at is whether there is a business opportunity for us in the near future. So we are definitely doing -- starting on that.

I mean, it's really wild to me because if we -- you and I talked about this, we did one of these in December, we talked about some of these things. And it was this kind of esoteric semiconductor question. And now suddenly, I'm getting Google and Microsoft press releases every other day about lowering the cost of inference, lowering the cost of per query on these things. So it seems like there's a real opportunity there to do this and you have technology that might be applicable to it.

I agree with that because in the past, majority of the people doing new network was focusing on training because that's a difficult problem and you need a very powerful platform and inference was not there because the people using inference on the service side is pretty little compared to training. But suddenly, because of ChatGPT has 100 million people using it on a regular basis, suddenly the pinpoint of providing performance is not on the training anymore. It's on the inference side. Inference has dominant the server performance requirement. From that point of view, it really created a question whether there is a requirement for optimized dedicate hardware on the server side around inference. And that's something we need to figure out.

So we have leadership technology in a really important new area. You have struggled a bit with numbers recently. So maybe we could talk through some of that. You've had a couple of supply chain issues in the last couple of quarters, and you've had inventory dynamics in the -- particularly in the IoT markets. Can you just talk about where you are in terms of resolving some of those issues?

Right. I think that we believe the current weakness that we're seeing in the Q1 that we told you is because of inventory control, inventory correction on our customer side. We gave an example in our script saying that one of our customers saying their revenue growth based on our product is growing 10% to 15%. But at the same time, when we look at their appeal to us and the forecast to us, it's down 20%. So there's a gap in there, which we think is because of the inventory correction. And we are seeing similar patterns of different scales among all of our IoT customers. And that's the problem we're dealing with. That's why we talk about our Q1 guidance is weak, but we also believe that, that's a low point from an inventory correction point of view. And we hope that in the future we will start recovering, but we just don't know how fast the recovery is going to look like.

If you look at the surveillance markets, you have a very obvious growth story there because you're moving from your traditional video processing socket to a computer vision socket that's more than 2x the value. So it's surprising, I guess, to me to see the magnitude of the drawdown that you've seen. Are you sure that you're not like losing market share in some of those legacy areas? How do I think about framing that longer-term opportunity?

Well, there definitely is we -- our CV products replacing some of the video processing product, that's definitely is one factor we are considering. But there is another concern. Three years ago, our Chinese security camera business was 30% of total revenue. This year, it could be 0. And that transition is also something we're experiencing. And on top of that, we're talking about inventory. So we think that our design win pipeline is still very healthy, and we're confident that what we are dealing with today is inventory problem, not the design win problem.

Okay. Great. And then you talked about CV being 60% of your revenue is up from 45% this year. Now we don't -- there's no revenue guidance associated with that. But on my number, that would be kind of a 10%, 15% growth in computer vision, which would be a pretty big deceleration from, I guess, 70% plus last year. Is that -- how do I think about that? I mean, I'm not asking for guidance, but I guess I would think that's relatively conservative relative to the opportunities that we know you have.

Well, I think that we need to -- in fact, when we give the guidance, the biggest worry for me is do we consider all the potential inventory. So that's where -- I won't call it conservative, but definitely, when we give the guidance, we want to make sure we consider all the potential problems we're going to deal with on the inventory side when we give the guidance. Computer vision, I continue to see not only that give us twice ASP growth, but also we're going to go to new applications, right? We talked about [indiscernible] access control. We talked about potential Kodiak Robotics. I think the auto is our biggest opportunity in the next few years. So I think I'm confident that computer vision, I should say, AI application, going to continue to drive our growth and after we correct -- finish this inventory correction.

Great. Maybe pivot and talk a little bit about your automotive opportunities. And I want to get to the new stuff to CV3, but I guess if we look at where you are today, you have a very strong position in areas like surround view, video digital review mirrors, driver monitoring. Some of those categories seem kind of slow to take off. And it seems like driver monitoring in particular, we know there's regulatory support, we know as part of the European luxury car standards. And I know from talking to your customers how strong your position is. So when do you start to see those markets contributing more to revenues?

We've definitely seen that those markets are going to continue to grow. In fact, on top of that, even when you talk about regulation, you talk about in China, they finally approve that they can use electronic mirrors without using optical mirror associated with together. So suddenly, there is a regulation to support better electronic mirrors design wins. So I think from that point of view, we believe that the growth of the DMS and the e-mirror will be important for us. But however, sometimes ASP is on the low end side compared to our ADAS or CV3 revenue. So the impact to our total revenue growth is limited by the ASP, not because the unit number grow.

But it still seems like a pretty relevant number as those things start to ramp. And then I should mention your automotive funnel for the next 6 years, you've talked about a kind of $2.3 billion opportunity for a business doing about $100 million a year. So obviously, a lot of growth potential in those numbers. Maybe before we get to CV3, if you talk about the opportunities in ADAS before you get to that level of domain controller, what are the prospects for forward-facing single camera type of wins for you guys?

Right. So for ADAS, we have been winning -- we have been talking about design, but our investors continue as well is a smaller OEMs that is not the major OEMs. But I think one major thing that we think can help us to change that is when we announced a partner with Conti & Bosch, they will not only help us on the design win with the CV3 pipeline, but also gives us enough credibility on the CV2 side or OEM who consider us in the past, but worry about whether we can have scale. Now we don't have the scale problem anymore because we are working with major Tier 1s, all those things. So I really think that ADAS will continue to be an area we need to focus on and try to get -- continue to get design wins on the big OEMs. We have announced a lot of ADAS design. In fact, we announced quite a few just this time also. But like I said, we need to have a bigger OEM design wins to drive a bigger growth on the ADAS side.

Yes, it seems like the barriers to entry that are coming down faster than I might have thought in terms of doing those types of L2 functions?

Right. I think because -- they are more and more software partners available in there and they are ready to take things into production. In ADAS, we don't have an Ambarella software bundle to it. In CV3, for the Level 2+, we have our own software that we bundle with to do demo. But also, we announced Conti and us going to have a joint software development for -- which will be available for OEM who want to take that software into production. So things changed quite dramatically in the last few years in terms of people understand our benefit as well as the mature software solutions.

You should still have a pretty strong vector for growth in automotive even before you...

I believe so.

And then CV3, maybe you could just for people who aren't aware of it, can you talk about the -- obviously, that's been a lot of your investment in the last several years and a lot of your investment now from the design win pipeline perspective. Can you just talk to that product?

Right. CV3 is really, we call it, domain controller. So the idea is this is not only just processing video, it is really taking all the sensor modality that is required to do a Level 2 or Level 2+ even Level 3, Level 4 cars. And we take all the sensor modality, we do all the single processing, including both camera and the radar maybe even though it's ultrasonic. And to fuse them together at a very low level to a low level sensor fusion and to try to provide the best perception solution. And then on top of that, we're going to provide enough solution based on new network or other algorithms to past planning to drive policy or safety modules, all the software stack will be available on our CV3. So domain -- CV3 is a complete domain controller on both hardware and software side. We expect one single chip CV3 will provide all the function and to drive -- for the driving solutions.

And you've had 2 really important wins here with both Continental & Bosch. I think you had some -- there was some press today where Bosch was highlighting the importance of the Ambarella relationship. Can you talk about the importance of those wins? And obviously, those aren't things that drive imminent revenue because they have to turn around and sell it to OEMs. But can you -- the demos, by the way, thought really impressive, both that you guys did at CES, but also the ones that Continental did in its demo. Can you just talk to [Technical Difficulty] might be getting with OEMs?

Right. So first of all, we definitely talk to OEMs with our Tier 1 components today, but more importantly, I think we are those guys -- basically announcing County & Bosch really give us credibility of size, right? People have been complaining about Ambarella is too small compared to NVIDIA and Qualcomm. But when we partnered with Conti, that kind of solved the problem because now we are not standing there alone, we're standing with the big guys that OEM is familiar with and that adds credibility. But on top of that, I want to emphasize the software announcement with Conti again, we are not only -- Conti is now just providing a box hardware platform to the OEM anymore. The software that County and Ambarella are going to jointly develop is going to be either a backup solution for OEMs who are doing software development itself or the solution for OEM who doesn't have a software solution itself. So with the Conti collaboration, we think that we will have a complete hardware -- software solution that offer to OEMs who need it.

I want to ask about scale because since you mentioned it. You're competing with NVIDIA primarily in kind of L4-L5. You're competing with Mobileye which is part of Intel on L2, you're competing with Qualcomm mostly on L2. Those are the 3 biggest semiconductor companies on earth. And the -- in an environment where everyone is very focused on supply chain and the automotive customers are thinking -- making sure they get it. Obviously, you've proven yourself in terms of technology capability. And I would argue versus any of those competitors, you have real leadership, particularly in kind of the sense of like EV battery powered -- power-sensitive stuff. So you have really good technology. Do you feel like your scale is something that holds you guys back? And what can you do to sort of bridge that gap?

Yes. I think absolutely, the scale was a problem when deal with OEMs. But however, like I said, we work on that. In fact, there are a few arguments. First of all, for our OEM customers who really want to compete with Tesla, technology matters. I really wanted to consider to show them to compare our solution to our competitors, make sure they understand we can offer better solutions to our OEM. That's the first. Second thing is that's another reason we continue to focus on the importance of Conti & Bosch because the scale they add to and their relationship with the Tier 1 -- sorry, relationship with OEMs, it's a major, major factor for us in the last few months after we announced that. So those 2 things definitely. But however, I would always agree that if we are in a much bigger platform, we can do even better with our current technology. So I never say no to that. But however, we are not going to sell this company, just putting ourselves the block to say, "Hey, please buy us." That's not our intention. Our intention is that we need to continue to run this company as an independent company when the opportunity presents itself, finally is to consider whether that's the right thing for us to do.

Great. And these wins, it seems like obviously, Continental and Bosch are both leaders in these markets, they've made a large investment to work with you guys. I guess when you think about your funnel, how are you thinking about those things translating to revenue? And should we think of that being upside potential to the funnel over time as you start to convert some of those?

Right. So last year, we talked about $2 billion to $3 billion total funnel size and with Conti & Bosch, you definitely expect this year going to be better. And in fact, I think just by working with the next thing, we'll definitely improve our percentage -- the possibility of winning new design wins. So I think that will be positively impact our funnel numbers that we will talk about in November this year.

Okay. Great. And then can you talk about a couple of developments. First radar, I want to ask about software as well. But you acquired Oculii, you have some really groundbreaking technology in the radar space. How do you look at that being integrated into your solutions?

So we announced -- Oculii radar is fully image radar, is fundamental technology. But on top of that, we announced a brand-new architecture, we call centralized radar. We think that's -- it's enabled by CV3 as the domain controller, plus Oculii's unique algorithm, radar algorithm. Together, that's the combination of this 2 enabled this centralized radar. I think very few people can do this because the problem we solve is most of 4D image radar companies trying to use a huge antenna array, 48x32 or 24x16, huge antenna array to get the distance accuracy. We are not. We are using 4x3, 8x6 antenna array, but we pass that information into a CV3 or the centralized -- raw data into a centralized domain controller and we use our algorithm -- Oculii algorithm to extrapolate the point cloud that we need. And we can achieve similar performance in terms of accuracy and distance just like any other 4D imaging radar, but the benefit is -- because, first of all, you can do sensor fusion with a video data point as well as radar, and you can do a low-level sensor fusion which nobody else can do. Two, because we use much smaller antenna, so the cost and the power on your radar module is much, much lower, but we add not only the better performance, better accuracy, but also lower bounds because of our approach. So this centralized radar is being well received by the customer. And that's definitely something we think that will be another way to help us to win OEM design wins.

Yes, it's a really unique capability. What's the path to the revenue generation? Like how long does that take?

Well, it's because it's associated with CV3, so it definitely is related to CV3 design wins. That will take -- probably, we're talking about 2026 right now. However, in between, we continue to sell the Oculii radar at edge, that we announced with [indiscernible] that they're using -- basically the solution is running Oculii software on a TI chip, TIDSP at edge, and that they pass the Oculus to the processor in the middle. So we -- for each car, we have a front radar and the back radar and our total content is probably in the $30, $40, in that range. So from that point of view, we are also -- we already have a road map for 4D edge radar. Now we have a centralized radar. So in the next 3 years, we're going to monetize on the edge radar side, but also focus on design wins for the centralized radar.

You had automotive OEMs that were investors in Oculii before the acquisition. So I know there's a lot of traction there potentially. And then can you talk about software. You announced Continental hardware win aAnd then a few months later, you talked about incrementally software wins. Can you talk about the modernization of that? And I think when you look at like the NVIDIA Mercedes relationship, which is really good economics for NVIDIA, they're attributing a lot of that to the DRIVE software platform. So how big do you think that software can be for you over time?

If you look at the Goldman Sachs report on the -- sorry mentioned their name. So the software stack worth is roughly $700-plus in 2025. And we believe that's realistic number for software content in Level 2 plus autonomous-driving car, but of course, the value will continue to go down. So the way that Conti and us are working together is we want to split, whatever the price we want to get on the software side. So that's in addition to the ASP you can get on the CV3 side. And that's the first thing. But also because that -- all the software can be upgraded in the future, we believe any time you upgrade software, you probably get also your fair shares of software content on the upgrade. So that's 2 ways we're going to get -- monetize our software.

Great. So I have one more question and then we can open it up to the audience. You put this focus on CV3. It's obviously a pretty large opportunity. Does that take away focus though from some of the other potential wins you have, CV2 and some of the other areas? I mean, it seems like we're -- in 2019, there was a major investment in CV. We're 3 years away from monetization. We're now at the modernization phase and investing in another big wave. And I'm very excited about the investment potential that you have, but how do you make sure that you're kind of monetizing the investments you've already made?

In fact, there's -- you're asking about our trade-off in terms of our investment. CV3 is a huge investment for us, and we haven't monetized it yet. We haven't seen revenue from that year. But because of the focus on that investment, definitely, we have to give up something. What we really give up is we stop investing on video processors, particularly on the low-end side. A lot of video processing today, for example, for high [indiscernible] we are selling $3, $4 video processor. We're kind of saying that's not our business. We don't want to spend another silicon on that business. We still sell into the low-end business by selling our current chip with that lower [indiscernible] trying to maintain this year. But we have not -- a video processor chip targeting at $5, $5 below ASP. That's the trade-off we decided to make. And because we truly believe -- in the next 5 years, video processor market will quickly transition to computer vision. And in 5 years, I don't think that there is a video processing market anymore. So although we can probably lose some short-term opportunity because there's still a lot of low ASP opportunity out there, but I think that our focus will pay off in longer time.

Great. If we have questions from the audience?

I would like to know your talent management strategy. Do you think right now is a good time to do some [indiscernible] savings, layoff employees or maybe lease is a good time to hire more software engineer?

First of all, we need to watch our cost expenses, but that's not at the cost of losing our long-term strategy, right? We will continue to focus on our CV3, our investment, our schedule that we commit to customers will be met. But however, we are also watching the expense on the edge that we -- if there's any [indiscernible] we're going to train it. But we haven't laid off anybody because we think that we -- our engineering resources are very efficient. But at the same time, we are taking other opportunity to cut expenses so that internally that is not visible outside, but we definitely want to watch our expenses without hurting our long-term investment.

In Joe and Fermi, it's Charlie from Cowen and I cover foundry sector, TSMC. So my question is about your foundry strategy because you just announced that you will use Samsung for your CV3 5-nanometer foundry, but you should be aware that 2 years ago, some of your peers, Qualcomm and NVIDIA, they use the Samsung Foundry and separate a little bit, right? So can you kind of go through why you do this very concentrated -- big concentration on the Samsung Foundry.

Right. So because we're a small company, we cannot afford to have a dual source for each process, right? When we commit to 1 foundry for 5-nanometer, for example, we can afford to do another one. But however, at any process note, when we start new process note, for example, we're going to evaluate 4-nanometer and the 3-nanometer very soon. We keep our option open. We definitely look at all the possible options. So the reason that we use Samsung extensively for a long time, I think, first of all, like we talked about a small company that the priority at TSMC probably very low, but also I think we get a better services and the price, particularly on the price side. on Samsung that really help us to get a 60-plus percent gross margin. I think that's definitely one of the reasons we want to continue to use Samsung. But however, I truly understand that when you move to next process, which is really risky because the technology maturity is very important. So we need to evaluate properly when we go to 4-nanometer which we haven't made any decision yet.

I mean, can you talk to -- last minute we have -- there's been a number of OEM announcements that your competitors have made that when we've dug around that there's still opportunity for Ambarella that's pretty material. And so I don't want to put you on the spot by mentioning any of those individual names. But when you look at these high-profile wins, does that still leave opportunity for Ambarella at those customers? And what does that opportunity look like? Is it 2026 CV3 stuff? Or is there stuff that you can do along the way?

For the domain controller thing, we're talking about 2026 because that's how long it takes. And I also believe there are a lot of things open to us because if you look at the first generation domain controller solution people selected, there obviously is a problem, right? The biggest problem is power consuming so high, you need water cooling. And that is a fundamental problem I don't think that NVIDIA can address yet. So without addressing that problem, I think for our customers, it's natural for us to start looking at possible other options. So evaluating other solutions is ongoing, and I think we are definitely thinking that we have a chance to bid on those projects. And for ADAS, it's probably totally different dynamic, right? There are a lot of people using a black box solution, which they want to have more flexibility, which I think we can offer.

Great. Well, we'll wrap it up there. Thank you so much for your time. Thanks.

Thank you.