Thank you to this session. I'm Vivek Arya from B of A's semiconductor team, and really delighted and honored to have the team from Ambarella join us this morning. And really delighted to have Fermi Wang, the Chief Executive Officer of Ambarella. I'll go through some of my questions, but please feel free to raise your hand if you would like to bring something in between. But very warm welcome to you, Fermi. Really appreciate you joining us.
Thank you, Vivek.
So Fermi, maybe let's, you know, start at the beginning, which is, for a long time, you know, Ambarella has been known as a specialist in kind of vision processing and cameras, but you have made this interesting pivot towards AI. So maybe walk us through, why did you decide to make this transition, you know, and where are you in that transition right now?
Right. So, you know, we started this company 20 years ago purely because we want to—we believe that we want to enable people to, to capture personal video content. That's—20 years ago, that's the intention of the company. So for the first 10 years, we tried to optimize our video capture technology, from the video processing, to video compression, to the storage, to the local storage, so that we can really help our customer build the best camera. We call it human vision camera. Basically, camera is designed for human to watch the videos. But, and then, then luckily with that, we went IPO in 2012. Right at the IPO time, it become very clear that our customer want beyond just a human vision.
They start seeing that computer vision becoming important because neural network enable the brand new trend of AI processing for video, and they want to enable cameras not only can watching, but also can help us to people to understand the environment. That's exactly where we start our computer vision or AI type of video processing. It took us a few years to understand how to do an Algorithm-First approach to provide optimized silicon architecture for AI, for video data on our silicon. Then, the real product we generate, really generate real revenues is our second-generation CVflow engine. We start shipping CVflow engine in 2018, 2017, 2018 time. The CV2 engine, we have our dedicated neural network AI inference engine in there. And the market, the AIoT market as well as automotive market.
And throughout 2017-2018 to 2023, last year, our AI engine-based revenue is roughly 60% of total revenue. So you can see that we transition from a video human viewing camera company to an AI video camera company today. But however, we didn't stop here. In fact, in the last few years, it become very clear to us that the AI performance requirement for our customer, existing customer, and the future customer, going to continue to increase. And the GenAI is just another take it to a different level, but just look at our current silicon that we're producing right now. Our CV2, our first generation of our second generation of AI chip, the average ASP is $20. CV5 is the top end of the second-generation processor.
The ASP is anywhere between $30-$40, and AI performance increase proportionally. Our third-generation AI processor we designed for the automotive, the ASP we talk about is $100-$400. So you can see that the dollar signs also relate to how much AI performance you have. So my feeling is, throughout the year, we start with a human viewing video. Now we are doing AI for video, and we counting on the AI performance requirement from our customer continue to increase in the future. Of course, LLM will be the next generation, and we'll talk about this for fourth generation AI processor engine, but that's where we are at with our current AI for video technology.
Got it. Exactly as you pointed out, I think your 60% AI exposure probably puts you, you know, next to just the one large player, right, in the market.
Yeah.
But if you kinda contrast for me, you know, yourself versus that large incumbent, right? Their, obviously, products are designed for a very different application, right? But they also cost tens of thousands of dollars. Right, you mentioned something that's in the tens of dollars. So what is your AI strategy? How do you pick the specific markets to target? And what kind of differentiates you in those markets versus the large incumbent that's out there?
Right. So in fact, everybody is focusing on training, which we don't do that at all. We focus purely on the inference side, on edge, edge product, edge device or edge server. And the other thing that we don't do is really that we don't do cloud. Data center is not our business, and we purely focus on the areas in AIoT, automotive, basically edge device company. So from a technology point of view as well as the market focus point of view, we are different. However, I really think that the reason you have so much investment on training is because you want to enable more inference in the future. Otherwise, why do you want to do training? The idea is enable different application at the end.
So inference, I think, I truly believe will be have a huge growth potential in the future. In terms of how we differentiate, it's very simple. We believe that starting from day one of the company, we believe that anything related to video or edge AI, from that point of view, you need to have a specialized architecture. If you want to optimize for the power consumption, performance per watt, performance per dollar, you have to optimize for it. And our differentiation is always that we create architecture best for the video and the best for the video AI, and then therefore we can provide much lower power consumption. In fact, when I say much lower, it's not like 20%-30% compared to our key competitors, 2-3x power saving on the same performance level.
And also, with that, we provide a much better bank saving cost for our customers. That's basically major differentiation for us. Of course, there are other things, but this is really one important thing for us.
Right. Let me just a technical question that, you know, investors have started to kind of distinguish between there has been the AI that has been there for several years, and now there is this new large language models, and, you know, just the word, emphasizing the word large, right? And the generative, right, application aspect of it, right? That's a very different ballgame. So when you use the word AI at Ambarella, which AI are you talking about?
Right.
Is it both? I think just it would help us to understand the-
Right
... the distinction.
So maybe we should look at the how the AI evolution in the last 10 years. When we start doing our AI for HD device, video HD device, we use this technological CNN neural network, right? Convolutional neural network. This was the main architecture for the CV2 our second generation architecture. Then when we come to the third generation, one major difference is, in addition to the CNN neural network, we add transformer technology support, which is another major architecture required for not only for automotive, but transformer network is really basis for large language model recently. So the third generation really increase the support on different architecture level. And I think next generation, we need to find a way to really optimize for large language model architecture.
In general purpose, will help you to really go to market quickly, but for the people who really focus on power, power consumption or the performance per watt improvement, this will become an important differentiation for us.
Do you think competitors from the mobile industry can go after the same thing? Because they are used to making products, right, in sort of that price envelope, and there's a lot of ecosystem that is now going after developing language models, right? I don't know whether I should call them large or small or medium that are transformer-based, but that are going after that specific kind of device manufacturer.
Right.
Do you see that as competition for you?
You know, there is one—if you follow the automotive software development, people start talking about end-to-end AI, which is Tesla's word. But in my mind, that's just one example of large language, large model.
Right.
From that point of view, I think eventually, when you really look at, in the future, when you talk about Level 2+, Level 3 car, I think you're going to see large language model in there. But you might not see 1.4 trillion parameters.
Right.
You might see a fine-tuned 30 billion or 24 billion parameter running a car and provide an end-to-end AI. I, you know, I don't know exactly how many parameters you need because we're still developing, but I don't believe we need 1.3 trillion parameters. But, so for large language model, I also believe that on the data center side, people are pushing the ChatGPT through 1 trillion parameters. But from an edge device point of view or some vertical applications, you're going to see smaller, large language model-
Right
... but fine-tuned for each vertical. That's where I think automotive will be in the long run.
You described several generations of your AI inferencing technology, right? You said you added transformers to the third one. What do you plan to add to the fourth one, and do you think that'll entail working on new geometries, as well?
Right.
Like, what, what are the building blocks of making Ambarella move to the fourth generation of technology?
Right. So when we support Gen 3 for transformer, we kinda defined it for automotive. But if you really want to define the next generation, like, transformer support for larger language model, obviously, if we apply this, we call algorithm-first approach, and understand exactly how large language model is transformer, there's tons of things you can optimize for. So I expect that our next generation CVflow engine will be focused on optimize, providing much better performance per watt, therefore, performance dollar, architecture for LLM. But that's just one direction. To reach that direction, I think we need to invest on the next generation process node, right? And we talk about in our earnings call, we talk about we are kicking off a 2nm process implementation, and we are finalizing our evaluation of 2nm.
We want to make sure that we get a commitment from our foundry partners to support us for this really advanced processing development. This is a Gate-All-Around . It's definitely, I think, difficult transition for foundries as well as for design house, and we want to make sure that we did it right.
Do you have a choice of foundries, or is there choice of one?
No, I think, you know, obviously, Samsung continue to be, you know, our partners on the foundry side. We're still also talking to other foundry can support 2-3 nano, which is TSMC only. So there's no, no, no, suspicion there. There's only two foundry can do this at this point.
Are there any aspirations of taking the fourth generation and moving into the data center as well, or do you think just by application, right? So you have been more, you know, IoT, industrial, and consumer, and then you move to automotive. Do you have aspirations of going into the data center as well?
There is aspiration, but also there is a reality, right? So as a small company, even we think we can build technology, whether we can penetrate the market, that's two different questions. So I think that we will try to take approach that the technology will be suitable for the current market that we are targeting at. Also, there's a chance to upgrade to the data center. But I think, although we haven't finalized this, but I really think that we need to find a way to enable our customers and partners that are willing to work with us on their current market segment as a first priority before we jump to a totally different new market and trying to compete with the, you know, the giant competitors.
Got it. You started to talk about the different products, so kind of the 60% related to AI, so CV2 and onwards, but video processor is still, I think, almost 40%. So I was hoping you could kind of double-click on these two and help us define how do you look at, you know, sort of the gradual kind of-
Right
... deceleration of one and the acceleration of the other?
Right. So first of all, last year, roughly, our total revenue from our AI video chips is 60%.
Right.
The percent will continue to go up. In fact, it's driven by two things. One is our CV5 family. In fact, we talk about that this is really the ASP at north of $30-$40 range.
Right.
That will help to drive our ASP growth as well as our revenue growth. The other thing is that the point, when we look at the recovery from this inventory correction, we see the strength of a CVflow, our AI engine chip, more than on the traditional one. Our video processor, we have not tape out any chip without our CVflow engine in there. So meaning we haven't really invest on the video processor-only market in the last three years. With that, I expect there's continued to have a long, long tails of a reduction on the video processing. We see huge deduction, reduction last year, but we're gonna see much smaller reduction on the video processor this year. I believe our growth on the computer vision will cover and outpace our reduction on the video processor moving forward.
Got it. You mentioned the third generation of AI inferencing. Could you help us understand the difference between the CV3, CV7, and N1? You know, sometimes I get lost in these different products.
Right.
So how, how do you kinda distinguish? Are they according to different power specs? Are they according to different applications?
Right.
What's the differentiation?
Yeah, so like I said, CV3, from technology point of view, there are two differentiation versus second generation. One is the transformer support, the other one is the power efficiency. Running the same workload on second generation and third generation, the power efficiency is 3x. So with the same workload, you, you can reduce the, the power consumption for one to one-third. But however, from the silicon point of view, when you see a CV7 family, which is really for our AIoT, when you see CV3, is really for our auto-grade, Level 2, Level 2+ cars. Right? So those are two very distinguished, just like we disclose our revenue. There's AIoT revenue, there's auto revenue. So CV7 is for this side, and CV3 is for the auto side.
The N1 is really created for the LLM, which, you know, we start this process just last year because when large language model pop up 17 months ago. N1 is going to be a LLM processor that we designed for LLM, and that's definitely a focus area we're going to continue to invest. We can continue to see AIoT chip, auto-grade chip, and the LLM chip from us.
All right. Now let's kinda move to redefining your you know by product rather than you know these technologies. So right now, it's about two-thirds IoT, you know, one-third automotive. What is that target mix you are working towards? I think just before our discussion, you were talking about the auto shows. So I imagine autos has got to be a much bigger part of the mix-
Right
... going forward.
In fact, when we start our kickoff for CV3, our intention is that, you know, automotive has to become a major portion of our revenue in the future. It has to be in the near future. In fact, with the current design wins, the, with our current Chinese OEMs, we are talking about the end of 2025, early 2026.
Right.
So even with that one design win, I expect that the ratio of our AIoT versus auto will gradually move to 1- 1, and with more design win, that will move to even bigger. So automotive will continue to... We invest so much money on the third generation AI engine, in particular for automotive business. We need to quickly generate revenue on that. That's the number one priority for the company right now.
Got it. How do you, you know, from a competitive position, kind of stack Ambarella versus, say, a Mobileye first, and then against, you know, Qualcomm and NVIDIA and some other names later?
Right. So for Mobileye, I think really, you know, for people who really enjoy a black box approach, Mobileye is always their first approach. But however, for the OEMs who want to control their own software and their own differentiation, like the design win we had-
Right
... we talk about, you know, that's not the strength of a Mobileye. In fact, in those kind of design win, we face NVIDIA and Qualcomm or Horizon in China a lot more often. So I think that's a two-sided competition. We continue to believe for the people who are OEM say, "Okay, I need to have a product shipping in the near term, and the black box is okay-
Right.
Mobileye usually have the edge. But however, for people who really said, "You know, I have my own software, I want to differentiate," then, we face other competition.
... Right. So how do you make the distinction for me between, you know, where does ADAS stop and where does Level 2 start, right? Because, right, Mobileye will obviously, you know, suggest they can do both, right. But today, a lot more of the focus seems to be in kind of the front, right?
Front ADAS.
Front side ADAS.
My, my definition of front ADAS is just a single camera facing front, with maybe 2 or 3 radars on it. So that's one V, we call it one video, 2, 3, three radar, 1V3R , 1V5R type of solution. That's called-
Right
traditional ADAS. When I go to Level 2+, I'm talking about multiple camera, multiple radar, with central domain controller.
Right
... central domain controller. That's my different, my definition change, right? But from function point of view, ADAS is really just emergency brake. For Level 2+ , there are two levels of autonomous driving, in fact, defined by Tesla. One is really called Autopilot, right? It's really highway-level autonomous driving with multiple camera. The next level is really FSD. Basically means you can do autonomous driving-
Right
... on the local road or city roads. Those two require totally different performance level. In fact, we believe that although both of them call Level 2+, and the AI performance requirement between these two can be 2x or multiple x difference in terms of AI performance requirement to reach these two levels of a Level 2+. So I think Level 2 is very well-defined, but it's more by dominant that. Level 2+, there are a huge range of performance and requirement. It can go as low as a 5V, 5 video, 5 camera, 5 radar, all the way to 12 radar, 12 camera, 5 radar. This wide range, it can be defined by software and also a combination with hardware accelerator.
I think that's where we're focusing on, really making sure that our CV3 family of chip, from the low-end 635 to the high-end 685, we have a family chip that can address this Level 2+ applications.
Got it. One other thing I found fascinating is your participation in the domain controller market, right, with CV3. Because, you know, there you have a different set of incumbents, right, with NXP and others, right, who have a lot of scale and incumbency in that market. So first, what made you go after that specific market? And I know you have a very important win there. When does that start to become tangible for you?
Right. So first of all, I think in any car, there are many different domains, right? The domain that we're interesting is really just safety domains, which is controlling the car to drive. And NXP and other traditional automotive companies, they really focus on other domains for providing controllers. So for the safety domain, the biggest requirement is really video and radar perception, and you do fusion on top of that. That require really high level of the video processing and the AI performance. That's where we think we are better than NXP or other traditional controller or the microcontroller companies. So I really think that our key competitors in this safety domain is NVIDIA and Qualcomm. Where will it make a difference?
I think, I hope that we're gonna continue to build out our momentum on the 2026 revenue, from the first design into other revenue, even getting radar inside of this revenue ramping up. So hopefully that we can continue to build momentum on that for our 2023, 2026 revenue.
Got it. Just, near term, how is automotive market demand? Because that is the one market where, you know, it has not quite had the flush that we have seen in the industrial-
Mm-hmm
... market. So what are you hearing from your automotive customers?
Well, I think right now, I think a lot of our automotive customer are dealing with the inventory correction that other market has been dealing with the last few years. You see that, automotive can start digesting the inventory. So that's definitely impacting the market size and momentum. But however, if you want to look at, say, even though people are saying, "Oh, even EV growth is slowing down," but I still believe that EV is going to continue to grow in terms of the market share, although the, it's not ramping as fast as in the past. But eventually, I think EV car is going to continue to become the mainstream.
On top of that, we start seeing, you know, Level 2+, really the real autonomous driving, or I should say, I should say that assistant driving, to a point that is better than just an emergency brake. It can help your car to navigate through the highway or local streets. I also believe that's a huge growth in penetration area. Today, Level 2+ penetration rate is probably 4%, 5%, less than that. I expect that the penetration will go much higher, and that's where I think will help us to get revenue growth.
Got it. Conceptually, why should inventory matter to a next-generation technology, you know, company? Because, you know, you're still in front of customers adopting a lot. Is it that you think that they have a lot of inventory of, you know, pre-built cars from prior generation technology? Like, why, why should inventory matter to you if, if you are developing a lot of new technology?
Not for me, but for our customer, when they start dealing, building with inventories, when their revenue starts to reduce, and they kind of start cutting back on the investment, right? In fact, the first thing they cut back is Level 4 cars, which-
Right
... become very obvious. But however, I truly believe that the EV car and people's investment on autonomous driving continues. That should not impact us a lot, right? So although I made some comment about general market situation, but I think the investment for our customers, and for us, for, and from Ambarella-... to this, auto driving is, it going to continue to be, a mainstream.
Got it. Just quickly on the near term, I think you, you know, you kind of described interestingly the profile in terms of there is a cyclical part of your business, and then, right, there is a secular part of the business. When do you think Ambarella kind of transitions to something that is a lot more secular, right, than cyclical?
I think, you know, next year we should go back to the normal business. This year-
Right
... we continue to deal with a little bit of inventory correction. You look at our first quarter, we did $54, and now we are going to do $62 this quarter. So we kind of still dealing with some hangover from the inventory correction, but I think that will be done by the end of the year.
Got it. A few kind of financial questions on gross margins. I think you have been consistently outperforming your 59%-62% long-term target. But do you think that as the auto mix grows, as you transition to these newer products, right, that we should be expecting gross margins to come back into that long-term model? Or you think there are drivers to keep it above that long-term model?
I think that for automotive, if you look at how big our competitor are, and they are famous being, you know, cutting price to keep their market share, that's where I think that we need to be realistic, that we're going to be ready to defend ourself. So I think that we make the comments about 59, we're going to gradually go back to 59%-60% because we need to be ready to compete in that space when our bigger competitors, they can use this as a weapon. So I think that's just a comment for that. But however, while, you know, continue to develop new technology and that go to different market, you know, we- our ASP continue to go up, right?
You know, from the CV2 family, from the $15-$20 to CV5, $30-$40, for CV3, $100-$400. When your ASP goes up, I think that's where I also would suggest our investor focus on. We'll continue to maintain a healthy gross margin while expanding our ASP range.
Got it. You know, one thing Ambarella is well known for is just kind of the commitment to R&D, right? Regardless of whether it's up cycle or down cycle. That's been very impressive, and that's obviously the reason why you have all these new products coming out.
You know, competing with NVIDIA and Qualcomm, this is the only way we can survive, right? Without technology, we are nothing. So making sure that we are not only investing, in fact, investing in 2nm is like a defensive move. It, it didn't give us any advantage because our competitor are there, too. Our differentiation is from our own technology, building our own CVflow engine for AI inference, building our video processing, much better AI video processing technology. Those are the core technology that will differentiate us.
Got it. So is there a certain way for me in which you manage the company and say, you know, if, let's say, top line grows, I don't know, say, 20% a year for the next three years, this is how much we should plan to have operating expense grow, or, or these two are done independently? Because a lot of times you're making these investments years in advance.
Right. So first of all, for CV3 investment, we are done, right? Because we spent in the last few years, we spent so much money to develop a brand-new architecture, brand-new market, software stream, automotive market, software. This investment is really mature, and we should continue to spin out new chip and new product and giving design win, without adding too much of a brand-new investment into this market. The new investment is really on the Gen AI side, which, you know, from my point of view, is not brand new because the first generation, like I said, 2nm , we've done 5nm before. So we could see incremental investment on the new technology. But however, how to generate a lot more revenue from the our investment on third generation AI is a key.
If we can—when we can enjoy our investment on the automotive market, I don't think we can. We need to another high percentage of our revenue on R&D side. We should continue to have a leverage on the gross margin and operating margin.
Got it. And then, China, what percentage of the mix is that? And do you think that there is any concern about whether it's restriction or, you know, whether it is, let's say, extraordinary exposure to the Chinese EV makers? How should we handicap China as a part of-
Right
... Ambarella's growth strategy?
Today, our total revenue selling into the Chinese market is 15%, right? So it's already reduced quite a lot compared to that in the past. However, I think that the geopolitical discussion is always valid, right? There are two directions. One direction is the U.S. try to limit the high-end AI capable chip to China. And also, China want to make sure that people use Chinese silicon.
Right.
So both sides, we are facing the pressure. But however, I really think that the current U.S.'s target is to make sure that high-end AI for training not going there. Inference engine is not a target yet. If that become a target, that become a really industry-wide problem. So, it's hard for me to comment on that. But on the Chinese side, Chinese always want to use their own silicon, if possible.
Right.
However, if you look at that, a lot of Chinese OEMs also want to do export business. I think that will force them to use, consider, you know, non-Chinese component for that part of the business. So the balance will be there. It's not like they can continue to count on Chinese component only for the complete automotive business.
Got it. And just finally, I know we are a little bit out of our time, but maybe just leave our audience with: what are the next two or three important catalyst milestones that we should be looking forward to for, for Ambarella?
First of all, I think most important thing is the design win for automotive, right? This is something we invest so much money on the automotive side. We need to continue to find a way to create a revenue on that. The second thing is really, like you said, technology. What's how we view we can continue to improve our technology, build our differentiation. Those are two things that we need to balance moving forward.
Terrific. On that optimistic note, thank you for me.
Thank you.
We appreciate your time.
Thank you.