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

Great. Welcome back, everybody. I'm Joe Moore. Very happy to be here with the CEO of Ambarella, Fermi Wang. So Fermi, maybe we can just start out. You have -- in a period where there's more enthusiasm for AI than we've ever seen, you guys have 60% of revenues coming from edge AI. So maybe put that into -- I mean you didn't really call yourself an AI company in the beginning, although that's what it clearly was even from then. So maybe you can just talk a little bit about what you guys do in the edge AI world and how that fits into this overall AI ecosystem.

Thank you, Joe. I think going back 6, 7 years ago when we started our, what we call, CV2 family chip, in fact, at that time AI is very simple. It's really about object detection, classification, traditional computer vision, neural network to enable some basic AI functions. But things changed quite dramatically. In fact, based on CV2 family, last year, we generated 60% of total revenue based on AI silicon, mainly CV2 family. Then we come out with our third generation of our computer vision chip called CV3, which is also AI-based. But this time, we targeted much higher AI performance for transformer-based network and targeted at automotive. When we started building CV3, in our wildest dream, we never thought that gen AI will happen. So we really defined CV3 family for -- based on transformer for the autonomous driving, Level 2, Level 3 and Level 4. Unfortunately, that transformer, because of Tesla, become a major requirement in automotive, and we can prove that with CV3, we can do much better AI performance than our competitors. In fact, if you look at the software stack, we have a software stack -- a Level 3 software that we demoed at CES. 95% of total performance around our AI processor, only 5% processing from traditional CPU like [ ours ]. That's how much AI that we moved to -- for the autonomous driving. Of course, then with latest gen AI, which is also transformer network-based, we found out that our CV3 chip can perform a lot of gen AI performance. So basically, you can see that, in the last few years, we started with CV2 family for just very simple computer vision, traditional functions, moved to transformer-based for autonomous driving, domain controller. Now we're addressing even bigger opportunity of gen AI.

And this isn't like a trend-following thing. I mean this is a really true thing that the people doing image resolution in automotive, the incumbents are still mostly using heuristic algorithms to do that. And you guys very early on adopted the fact that AI software was the way to resolve this. Is that right? And can you just talk a little bit about that software approach that you have?

Exactly. So the software stack we developed is really focused on people, call it, entering AI. Let's call it almost all the modules in there, including perception to the sensor fusion to the path planning even for the controlling car, all of them are controlled by AI. For example, we also use heuristic at the beginning. For example, when we acquired VisLab, which is really the company that generates software stack for us at the beginning, in 2015, their software stack mainly is running on CPUs, they will gradually move it. But what's the -- why moving to AI is so important, I'll just give you an example. In Europe, when we do the road testing, one of the most difficult problem is round about. When car approaching round about, in the past, with heuristic algorithm, we specify exactly how the car should perform when it reached every roundabout.

As a human, I have that problem too.

And then after we switched to AI, when the first time I see the car, I was shocked how it approached. Because when the car approached the roundabout, it pushed -- it didn't stop at the line; it pushed ahead over the line, just like a human driver because you get better visibility that way. Then I go back to answer the question how did it exactly learn from the human behavior. That's one simple example why learning with -- although that I don't think they understand why they want to do it, but because they achieve better performance and accuracy, that's the model they continue to use. So we believe that a lot of heuristic codes should be replaced by AI in the future, particularly for the autonomous driving.

So I mean I think we define a lot of the AI opportunity today by training in the data center, which is not a market that you have pursued. And it sounds like after some thinking about it, you're not pursuing inference in the data center either. So can you talk about the merits of what you are doing in the markets that you are focused on?

Right. The reason that we decided not to go to the data center is just still we're a much smaller company trying to go to a very large market and it's best interest for us to focus on the area that we have a differentiation as well as have some expertise in there. That's why we focus on edge AI for the inference at the gen AI market, particularly for the market that we are familiar with: security camera, automotive, robotics. Those markets that we are very familiar with, we have customer base. And that's why we want to focus on those markets with gen AI first, but that doesn't mean we'll limit ourselves to those markets. We believe that the market we want to focus on will continue to be edge AI, but any market which would require very low latency, very low power consumption and also require data privacy for gen AI. Any market satisfying one of several of those requirements should be our target market.

Okay. And then we've talked a lot about CV2 and CV3 over the last 2 or 3 years. But if you could put this kind of into a perspective that now we're looking at transformer inference, in particular, some of the complexity of that. Can you talk about the role of CV2, which is your existing revenue stream, and then how CV3 evolves from there?

Right. So in fact, it's very easy to just -- using our ASP change to simplify the discussion. Before AI chip, our corporate ASP was $6, $7. After we introduced the CV2 family today, our ASP is $12, $13, basically doubling it. When we look at CV3, our ASP is going to be from the low end of $40 to high end north of $400. So we're going to continue to see the ASP improvement. And we believe when we hit gen AI, our ASP going to be even higher than that. So I think that's why a few years ago, we decided that instead of trying to identify ourselves as a video company, we want to identify us as a company that continue to help to focus on AI performance at the edge. That's how we define ourselves at this point.

So maybe you could educate me a little bit on this like in the sense of -- it seems like for a long time, AI was focused on convolutional neural nets. And the inference task was not simple, but more manageable. You didn't need -- even in cloud, you didn't need $25,000 cards to do inference. You migrate to transformer and the inference task becomes several orders of magnitude more complex and suddenly we're using this very expensive hardware. And I think the same thing is happening at the edge. Can you just maybe explain a little bit that transition?

Okay. So for example, our CV2 family, the problem it's solving is really face detection, license plate detection. There's an object. And this object is a car, is a dog. It's really in the object classification, detection, maybe some of the open space detection, really low-level object categories. When you take down to the transformer level, you are talking about different level of integration. For example, the transformer is famous for bird's-eye view based on multiple camera, it basically integrates multiple camera input and build a bird's-eye view of the car, so you know how to navigate through traffic. You can see that the performance requirement to integrate multiple camera compared to just identify few faces in cameras is different level of performance requirement. And let's take one notch up is, for example, at CES, we demoed a neural network -- a gen AI called Lava. The idea is Lava taking live video input to N1 chip and N1 chip is running this neural network called Lava. The idea is the Lava output is describing impacts what I see on the video. So basically, you can consider it as a video to test conversion. And this is something that cannot be done in the past because the description is very detailed. It's not just, say, they're having people there. There's people wearing black suit, wearing black pants, drinking coffee and they also identify how many people are there on the space, what kind of environment they are. It's in detail what kind of -- the camera things that they are describing. So basically, you can imagine that it's as good as a security guy sitting in front of a monitor describing what he sees. And suddenly, this becomes very useful tools for any company security -- say let's -- in airport, they have 1,000 cameras. If you can see that 1,000 camera feed into the software, every second, you can generate a text report about what happening in the airport. And with all of the text search technology we have today, you can have a database of what's happening in an airport every second. Think about how powerful that is and that's going to change the security requirements in different environments. So let's just give you an example from very simple face detection to all the way to software can monitor all the security camera. That's the transition we've gone through in the last 8 years.

Great. That's very helpful. And so I guess the process of turning all of this into revenue, which has been a little lumpier along the way. CV2, can you talk about the growth profile that you still have? I mean, CV3, a lot of exciting opportunity, but still pretty far from revenue. But can you look at the IoT market and the automotive market as kind of the CV2 opportunity?

Right. So CV2, we talk about in fiscal year 2014, which is last year, our total ramp -- 60% of total revenue comes from CV2 family. So 60% of revenue comes from AI. And we expect that this year it's going to continue to grow, although we haven't disclosed the percentage, but it's -- obviously that's going to continue to grow. In fact, the growth -- the percentage revenue go from 20% to 45% to 60% in 3 years. That just shows you how fast our CV2 AI revenue growth. We expect the revenue to continue to grow. For example, we talk about CV5, which is a high-end CV2 family chip. It's our first 5-nanometer chip, ASP is roughly $40. And last year, we shipped 0.5 million units and this year, we're going to double it. And that is a high-end chip, the volume is going to continue to grow. So that just shows you how fast and how important this AI growth for our revenue. CV3 will take this to a different level. Of course, automotive is a much bigger market, but at the same time, the ASP. We talk about our low-end CV3 chip is at $40 and our high-end CV3 chip is going to be $400 or above. That ASP jump by itself going to provide our next step function of revenue growth. So it's not only just that we add more value to our customer, we also collect more value from our products.

And I apologize if this seems like a negative-leaning question. I'm actually -- I'm so enamored with the technology, and it's been hard to get to the revenue level that I thought you'd be at. So that's why I'm exploring it this way. But if you look at surveillance, I mean, there was a very clear growth trajectory where you had video processing and a low ASP, you're replacing it with AI at a higher ASP. And you did do that, you did continue to ramp that. But you had more headwinds in the legacy than I expected there to be. So maybe if you could just talk a little bit about that transition within surveillance. And you mentioned how you can see that continuing to move forward. You've gotten past some tactical inventory correction, which we don't need to rehash. But can we just understand the bigger picture opportunity evolving?

From the IoT space, I think the biggest problem is when we identify AI, we decide that's where we need to focus 100% of our attention and we decide not to take out any video processor chip, which basically become a downside of the problem. And so that you see -- well, in the IoT space, you see the AI chip continue to grow and our video processor revenue drop. That's probably the biggest problem. But I think the really biggest problem for us last year is inventory correction. I won't say the 100% of problem last year was inventory correction, but it played a very big role. And that's why I think our investment on AI will pay off in the next few years because our video processing revenue has become much smaller and our AI revenue should continue to drive our growth. I think I want to make a point on the CV3, going to automotive business definitely taking longer than I thought. And that has been a problem for us. The biggest pushback from the OEM was not our technology, it was our scale, and that's become very clear. And even like I admitted to our investors that had we working a much bigger-scale company, we could have collected a lot more revenue. But the reality is, we are a smaller company. So the way we deal this problem is, first of all, we work closely with large Tier 1s and hopefully -- and by showing up with a country partner or Quanta in front of our OEM customer, their size help us to address the scale problem that OEM has. That's definitely one thing we spent a lot of time last 12 months to resolve.

I mean if you do the work here, your technology is really good. Like everybody that you talk to talks about the breakthrough capability that you have, particularly with the right power envelope and things like that. So how much of that automotive problem that you talk about is the evolution of ADAS, in general, being a little slower? It seems like the focus has been on battery power, EV, a lot of the innovation kind of got siphoned away. And even now, I mean you have these wins that you talked about with Continental and Bosch, which for anybody who hasn't seen them, you should see them at CES. Both over the last 2 years, these are really intricate kind of 12-camera demos and things like that. But the time frame of that being implemented in cars seems to have been push backed. Not all that's in your control, but how do you think about those dynamics?

I think what happened in the last 3 years become very clear to everybody is the focus on Level 4 is dying.

Yes.

Even Level 3 got pushed out because there are so many regulation problems. We have been focusing on purely on Level 2+. Level 2+ is not one product, it's a very wide range of performance. In fact, in China, our low-end Level 2+ solution sell for $40, our high-end Level 2+, we can sell for $200. That's actually how wide the range is and how many products can be in there. And I think after the COVID, everybody reset their product planning from Level 3, Level 4 to Level 2+ because they need -- everybody need revenue. All automotive guys focused on getting product out, get revenue. And the fastest way is focusing on low-end Level 2+ or maybe try to compete with Tesla with FSD time performance with higher end Level 2+ +. Either way, that's a focus right now. That's one of the reasons of delay, but I think everybody focused on that. And I expect that will become the mainstream product in the next few years. But I think back there, CV3 chip is not available until last year, that definitely is one of the reasons that we didn't see enough revenue design wins early enough.

Yes. And you talked about China, when you say the focus has shifted to L2+. Is that also a comment on the Chinese market? And then, just in general, it seems like the regulatory environment in China is much more pragmatic. Is it safer than a human driver versus in the U.S.? It needs to be completely safe.

I think what you said is correct. In China, they are even more -- even faster to move toward to Level 2+. In fact, it's clear majority of design win activity today in China is Level 2+. And again, in China, most of the EV company target Tesla performance and say, "I want to be Tesla." Right? So that's the goal. So I think in China, I would say, even majority of the design win activity we are seeing is focused in that area. Either it's very low end, go-to-market quickly, make money, or a Level 2+ complete with FSD so they can claim they are the performance leader. That's the range we're seeing in China. In China, I think, although EV is slowed down -- EV market is slowed down, but I think if you look at a Chinese vendor, they now only think that -- they are not looking at the Chinese domestic market anymore, they think the exporting business is a market they can tap into. And they -- that's one of the reasons they need to use foreign components. And that's where we think -- that help us to get some tailwind help on the Chinese market.

And it's really interesting. I was looking -- I went to the Qualcomm Automotive Day a year or so ago, and they talked about all this progress knocking out Mobileye from some key accounts, which they've done publicly. And they did that without really showing any technology breakthrough. And I actually asked them how are you knocking out an incumbent without a technology breakthrough and the answer is kind of people want seamless integration with the infotainment system and they want it to look like an Android phone. So people don't just turn this over to CarPlay and things like that, which was kind of a depressing answer to me a little bit because it sort of says that the market -- the evolution has really slowed down. So like do you see that changing? I mean the promise of L3, L4 is still really exciting to me. And is it really a regulatory environment enough just to sort of keep that from ever happening?

Right. So I think L3, L4, will happen, but you will push out for different reasons. I think technology is not ready just yet, it will play a major role, particularly in U.S. where regulation is tough. But let's talk about the integration. I think everybody agree that more integration drive down the cost, so that it's easier to sell. But for automotive, I think the best integration is not to combine the infotainment system with safety domain. In fact, even in the safety domain like autonomous driving, there are still plenty of things you need to integrate. For example, your camera and radar system should be integrated. Today, most of the cars still putting radar systems, camera systems, totally independent, you use -- putting out outside your domain control and try integrate later. I think that's wrong. The first integration need to happen is in the domain controller for the safety domain, everything needs to be integrated into the domain controller. That's the first step of integration. There is many argument to support infotainment system and the system to integrate. The main argument is cost. But there is some other argument, for example, for the cybersecurity. When your infotainment system downloads software every day or every week and how do you protect your safety domain from virus to go into your safety domain? That's a real problem, right? Cellphone is not the best example for cybersecurity. So I think there are many other reasons that we should keep a lot of domain separately. And on each domain, you should integrate as much as you can. For example, I definitely think infotainment system should integrate everything possible like a cellphone does. But with safety domain, just integrate enough sensor, enough processing, enough AI performance to Level 2, Level 2+ and Level 3, we still have a long way to go.

Great. And then you mentioned radar, Maybe you could talk a little bit about Oculii, the acquisition that you guys did, which seems like it's potentially a breakthrough in terms of implementation of radar. Can you talk about what traction you're seeing with that?

Yes. So in fact, after we acquired Oculii, we quickly decided that the best way to apply this technology is to use it in a centralized radar, meaning integrated radar processing into domain controller instead trying to use the edge processing like what's being used today in the last 30 years. Today, every radar, all the single processing radars happen on the edge where you have a single radar module processing down there. And at the end, each radar had generated object list and pass through the central processing and integrate that way. We think if you work for like 30 years, but with the current technology, the best way to do this is you take all the raw data of radar signal and the camera radar signal and integrate everything in the domain control like CV3. With that, you do a low level of sensor fusion, you can achieve the best performance, and we prove that with our own software at CES. In CES, if you're driving our car, you do see that very dense point cloud that we take in from the radar head into domain controller. We're taking 6 radar, plus all the camera. We have enough performance to process everything at a point cloud level to achieve the best object detection. So that, I think, is the future and we're already start seeing OEMs ask for these features. For example, in Europe and in China, would be the RFQ, we already start seeing OEMs specify saying they want centralized radar requirement. I think that's the progress. I hope that we can ride on this progress.

I mean it seems like it's a no-brainer to implement radar that way when you describe it.

It is. However, the problem is without the optimized software, the amount of information you bring to the domain controller is huge. You really need to find a way how to bring in less information, but still processing huge amounts of data. That's, I think, the beauty of Oculii algorithm. And that's IP. That's the reason we believe because of the IP, and that's the reason we acquired them is because the IP, we believe that centralized radar can be done efficiently. Without that IP, it's a difficult problem.

Okay. So I guess in that context, how do you guys think about LiDAR? Obviously, slower L3, L4 means less LiDAR anyway. But are you still approaching -- is there still some sensor integration that has to take in LiDAR? Or do you think the radar plus optical is adequate?

I think for people who can afford LiDAR, you should use LiDAR for Level 4. You can use as much sensor as possible to improve the safety. However, when it comes to Level 2+, you might not have a luxury to use expensive LiDAR, particularly, we believe that the centralized radar can be a good replacement of LiDAR signals. And hopefully, the demo we showed to our customer at the CES that within a one CV3 chip, we can take in multiple camera, multiple radar, all through sensor fusion in that chip and convince people that's where we need to go.

Great. Maybe you could talk a little bit about the progress that you're seeing in automotive today. We talked about L3, L4, but you obviously have driver monitoring. You have digital rearview mirrors, things you've been talking about for a while. Where are you at monetizing those?

In fact, we continue to announce design wins and the revenue from those areas. But however, the percentage, the penetration of those markets still continue to be small. And that's why you don't see a huge jump on our revenue and ASP is pretty low. It's single digit or to low teens of ASP for DMS, OMS or even ADAS today. My belief is that when Level 2+ start taking over the market, most functions will gradually be integrated. So I think that that's another reason that although we continue to try to win designs for OMS, DMS and ADAS with the CV2 family of chips, we know that when CV3 comes in, that will take -- will replace those silicon solution with a much better integrated solution. So I think that's how we view the market right now.

Okay. And then in terms of your traditional IoT markets, can you characterize -- we've gone through an inventory correction that you talked about. It feels like we're through it. It feels like you guys saw it before others and kind of have emerged from it before others. But you're still kind of along the bottom. So can you just talk -- can you characterize that for us? And do you think you have enough visibility to sort of feel really good about the growth from here?

So first of all, we definitely believe that the worst of inventory correction is behind us. We said that with some data points. First data point is through our last 12 months, we continue to talk to our customer. A big portion of the customer is on their way out of this problem already. Some companies still kind of managing the inventory, but I think a big portion is not. Two, we've been watching our booking and ordering for the last few months, the recovery is significant and also continue to stabilize our business. So from that point of view, we feel comfortable with our Q1, Q2 booking right now, and Q3, Q4 remains to be seen. But at least the momentum we are seeing gives us confidence that we have bottomed out already.

And I would think -- I mean the fact that some of the peers have really struggled would actually help the sentiment for you guys a little bit.

I hope.

We saw -- I mean companies like Silicon Laboratories saw a lot of the same issues that you saw later on. Okay, that makes sense. And then where are you with consumer surveillance? That's a market that it seems like a lot of these AI features would really be relevant, but it's also really expensive for cameras at those price points. So can you talk about that?

So consumer surveillance is really designed for home. Like I said in the past, there are obviously 2 segments right now. One segment is extremely cost-sensitive. On Amazon, you can buy $20, $30 home security camera today, and they try to bundle services, try to make money on service side. And also, there's a different approach, which is using -- putting a lot of AI into the camera. And so they can provide better service that way, but service fees are higher. And the trend we're seeing is more going to this cost-effective solution because that's where it's easier to gain market share in terms of unit number. So that's the direction. And that's where we have stopped investment for a long time, right? We don't -- we haven't invested on a video-only solution. We focus on the AI solution. So for the consumer IP cam, that's definitely an area we don't -- we're losing market share. We talked about this before. But hopefully, if they come back to -- if there is any customer who wants to have a meaningful AI performance, I think we still can be a supplier to them.

And when you say you've lost market share, I mean, is that -- you still have some of the premium design, some of the top door bells, things like that, you guys are still in that?

Absolutely. In fact, that not only we had still some critical design wins, we're still getting design wins. It's just that there's a lot of -- even for the customer who has high-end solutions, they also want to do low end. Those -- for those low-end market, we don't have a solution for them.

Okay. So you have consumer surveillance that has some potential; a professional surveillance here past the bottom of the inventory correction with a good ASP lift; and then in cars, you're seeing a lift from some of the CV2 designs. And then CV3 is more 2026. Is that the right time frame to think about?

Yes. We have been saying that CV3 will start 2026 ramping up in China, and outside China will be 2027 story.

Okay. Okay. And your visibility into that at this point is good. I mean when you -- I know you have a funnel. We can talk about where the funnel is derived from. But I know you have some probability weighted. What's your probability weighting on some -- do you have any of those that are moving into the high probability at this point for '26?

Since we haven't updated the model for November, but we definitely think we move -- we are -- some of the design moving to the right direction. There's nothing moving to 100% yet, but is definitely moving to the right direction. But however, even just for the one business, we talked about [ $800 million ], which is guaranteed, which is a design win we got. The only thing will change is whether the volume -- unit volume can vary when they go to the production. So that's where we're looking at it right now.

So I mean you are talking about less than $100 million of trailing automotive revenue; a 6-year funnel of $2.4 billion, which includes a fair amount of probability-weighted stuff in the back half of that 6-year window.

Correct. So we have -- out of $2.4 billion is $1.6 billion of that is weighted. Probably $800 million is...

Okay. Great. And then can we talk about other opportunities for -- outside of vision? You're not going to participate in cloud-based inference, but you've talked about inference at the edge. What does that mean? And what types of LLM, types of opportunities might you see that Ambarella could participate in?

Well, like I said, initially, we're going to focus on existing customers, which means it's video-centric application, but I do believe that the few things that will help us to go beyond video application. One is the delay issue, right? It's really the -- how -- the real-time response time. One is power consumption. But there is another very important thing is more and more edge customer when they retrain their LLM model, they want to keep the data private to the company. So which means is people are going to try to fine-tune their data at the edge, also inference at the edge. So from that point of view, I do expect for people who are paying attention to latency, power consumption and the data privacy, well, we have a chance to penetrate and that's where we focus on beyond just video applications.

Okay. Great. And then what other video applications do you see? I mean you talked about access markets, you talked about robotics market, which of those are you most excited about?

Well, first of all, do you talk about LLM or just...

For vision as well.

For vision, I think robotics continue to be an important market for us. Particularly now I think with the latest LLM development, I think a lot of people that are familiar with this market I talked to, all of them agree that eventually LLM will be the model being used on robotics. That one single model controls the whole robots. It doesn't matter it's industrial robots or family robots or whatever robots. Even if you come to Level 3, Level 4 robot, I believe LLM will become eventually the best solution for this kind of thing. So I think...

It cannot be cloud-based?

It cannot be cloud based. Think about your car controlled by a cloud. So I think -- well, [ Cruise ] already proved that doesn't work. So I think the key is making sure that we continue to develop a platform, lower power enough with an outperformance to run a large language model. But however, I want to point out, not all the language model need to be trillion parameters. In fact, if you focus on controlling a robot, you can remove anything related to the cooking, New York Times, movie theater, all of the unnecessary information, cut it off, retrain purely for the robots. I think the model, a small model like 30 billion, 70 billion can be very useful for -- if you focus on one particular vertical.

Great. So do I have a question from the audience? We can give them a mic.

Yes, I wanted to get some clarification on your comments about inferencing at the edge and going on beyond existing video-centric. You mentioned power consumption, latency and privacy of data. What exactly -- what use case and deployment systems are you referring to? Are you putting these into servers? Or I mean what are they sitting next to? Can you give us a little bit detail?

Let me give you an example happening in my company. Today, our CV3 programming is pretty unique programming, right? And we try to use a copilot to do C programming. But when you come to CV3, we believe that we need to fine-tune one of the LLM for this, right? So this inference and -- this training and inference data we need to stay in our company, we don't want to -- anybody use CV3 programming data we put into the training being used by the public domain LLM. So you have to fine-tune this LLM based on our own data and inference locally. That's one example.

Is it deployed in -- so I guess 2 questions would mean fine-tune using like RAG type of retrieval augmentation sort of techniques. And then what is the ultimate end product that CV3 will be? And is it going to be in the server? Is it going to be deployed in the end device? Like what's the...

The end device can be a PCIe card sitting on server, sitting on local or AI reference machine. You design only for the local -- one particular application, right? So for example, in one that we try to use our company is really used for to generate code, generate document, particularly for our silicon.

Okay. So it's going to be in the server, but it will sit on something like a PCIe card. And what were you saying? Does it need to be like in the smartNIC format to receive that data? Or is it offloading the data off of the central processing unit?

It will be -- because we don't want this data being communicated outside, so you assume that this is really a [ box ] sitting alone by itself and all the data in and out is -- all proprietary data will stay there, right? So it's a PCIe card, but with our CV3 chip in there and running our own models.

And you're going to market with sort of like the OEMs that are specifically -- or are you going to like -- sorry, are you going to deploy this through, say, Quanta or some of these ODMs? Or do you feel like that these can be deployed in general service, kind of general purpose servers like HP Dell, et cetera?

So we do not plan to sell the model we built for CV3. But definitely, the model will be, we'll sell in PCIe card with Quanta. Quanta can be selling PCIe card for anybody who want to develop this similar kind of model for your corporation and you can run the local model on the PCIe card.

So -- but if you talk it this way, you're going to that model. I mean it's not going to be -- it's going to be very -- it's not going to be a standard product. It's going to be very customized. So time to market, it's got to be really, really long, right? Or is there -- am I missing something? Could you make this sort of like a standardized platform that is quite easier to be built?

Thank you. for example, we use this as one example, but I don't think we're going to do a standard. For the edge server, it has to be specified for certain applications. For example, that for the security camera example. It's going to be some of our customer security camera customer, build their own bots, selling to airport, selling to the schools selling things. So it cannot be just general purpose putting a cloud.

And it seems like it's a very focused effort on things that you can do without, things that you're going to be particularly good at that you don't have to invest hundreds of millions of dollars to compete with?

That's the key, right? I think with us, we try to be focused and reduce the OpEx expense to enable our first few customers.

So that chip is essentially a CV3 with some of the automotive stuff stripped out of it?

I shall say, yes. Yes.

Okay. Okay. Any other questions from the audience? Can I actually ask a financial question with the nonfinancial motivation, but your gross margins have always been really high. And you always have legacy revenue, it seems like that's a little bit of a headwind. Have you thought about sacrificing some gross margins to preserve some of that? Is that a trade-off that you guys could make?

We have been doing that for a long time. In fact, the trade-off is this. If you ask me to serve a customer who only wants to sell a low gross margin, that's not what we want to do. However, for a large customer, they have a mix of high gross margin and low gross margin business. We have to do low gross margin to keep them with us. That's where we're willing to sacrifice gross margin to keep the customer 100% with us.

So you're not hamstrung by a low 60s gross margin like this?

No. No. Well, 60% gross margin is a target, I think, we can achieve. But it's not vice versa.

Yes. Well, with that, we'll wrap it up, Fermi, thank you very much.

Thank you very much. Appreciate it.