Okay.
Okay. We'll go ahead and get started. Thank you, everybody, for joining us at the 28th Annual Needham Growth Conference. My name's Quinn Bolton. I'm the Semiconductor Analyst for Needham. It's my pleasure to host this fireside chat with Ambarella. Ambarella is an edge AI semiconductor company whose products are used in a wide and increasing variety of edge AI and human viewing applications, including video security, ADAS, electronic mirrors, telematics, driver and cabin monitoring systems, autonomous driving, edge infrastructure, drones, and other robotic applications.
The company's successfully leveraged its heritage with edge perception to enter the edge AI market with, so far, three generations of its proprietary AI accelerator and its complete systems on a chip address the unique AI performance requirements of the edge in contrast to the edge, sorry, to the AI data center. The company's working with approximately 80% of this year's revenue coming from edge AI.
Joining me from the company today is CEO Fermi Wang, and in the front row are John Young and Louis Gerhardy, CFO and VP of Corporate Development, respectively. Fermi, thank you for joining us.
Thank you.
Fermi, maybe just at a high level, the company's undergone a significant transformation over the past decade from a company that used to sell video processors into consumer applications to a company selling AI SoCs into the intelligent edge and edge infrastructure applications for both enterprise CapEx and consumer devices or consumer-driven markets. Can you walk through this transformation? Give us an overview of the company's AI processor families and some of the target markets that you're going after today?
Great. Thank you. You know, as you know, Ambarella started in 2004. For the first 10 years, we built a video processor for human viewing. The most famous customer was GoPro, right, for sports camera. But in 2012, we IPO that year, but also we did some significant change. In the last year that Google published this first neural network paper that we convinced ourselves that's going to be the foundation of models for changing how people are using neural network to change the, we call it video analytics at the time. So with that, we stopped investing heavily on trying to understand what kind of silicon architecture we need to build to cover all of the CNN type of neural network, which we believe that's the foundation of the AI at that time. It took us a few years to build that.
In fact, we take the second generation chip, we call it CV2 family of chips that specifically target CNN type of neural network. We start shipping that chip family in 2018. Today, that family of chips represents 80% of total revenue. That's definitely a significant change. Also, it's fundamentally changed the whole company from a human viewing to a machine perception business. In 2018, when we finished kind of wrapping up a second generation, another thing happened, which is that there's definitely another research paper published on the transformer type of neural network that we believe that's going to be a foundation for many different things. We didn't see GenAI at that time, but we did see that transformer will become important technology for autonomous driving car because Tesla already started talking about it.
So we start our third generation architecture based on the belief that we, in addition to CNN, we need to build on transformer, and that's how we build third generation. Today, we build a family third generation chip. We're just ramping up the first two silicon called CV72 and CV75 in production with the customer last year. Q3 and Q4 start ramping up. So you should say we just start ramping up our transformer-based offering, but we think that total opportunity represents a much bigger market opportunity than CNN. Like I said, when we started the third generation, we didn't know GenAI, but fundamentally that GenAI capability will impact all of our existing customers for security camera.
Today, all the physical AI robotic applications we talk about that in detail, but that's the path we took in the last 20 years: go from human viewing company to perception company to an edge AI company.
No, that's great. And I was going to ask about the sort of transition from CNN to transformer-based, but as you mentioned, the CV2, CV5 sort of target CNN networks, that's effectively 100% of your AI revenue today, which is 80% of total revenue. You've got the CV7X family, the CV3, the N1 that support transformer-based models. When do you see the transformer revenue really starting to kick in? When does that potentially overtake the CNN-based business?
First of all, those two things, CNN and the transformer will coexist. I don't think transformer will completely replace CNN. That's the first thing. CV75, CV72 are two chips that we remember are going to start doing some kind of GenAI, but that's in a smaller large model. However, if you look at our other family, CV3 for autonomous driving and also N1 family for edge infrastructure for bigger language model, I think those markets haven't happened yet. It takes time because GenAI was introduced three years ago. Since then, our customers are trying to figure out how this GenAI is going to impact their business, build up hardware and software to take advantage of that and choosing our platform because we have much better power efficiency for the AI inference.
With that, it takes time to build up, but we talk about a few important things. First of all, that we have the first edge inference, edge infrastructure, sorry, it's a long day. Edge infrastructure business that design win last two quarters ago, and we definitely believe and focus on GenAI type of application for aggregate different video sources. And that definitely is happening. And we believe that when the customer shows up with that product, it will trigger a lot more discussion. In fact, if you go to, have you gone to our CES demo just happened last week? Majority of customers is demoing GenAI type application on our silicon. So I think it will just take time, but I think that in the next year or two, we're going to see tremendous upside, tremendous new design wins with the CV3 families.
Okay. So the CNN wave, if I call it waves, I know you guys have historically liked to talk about waves of revenue growth. The CNN wave is probably today over $300 million of annual revenue. A few years out, how big could the transformer wave be?
I think that there are two ways to look at the problem. First of all, that ASP growth is significant. CV2 family, the average ASP is from $15-$75. We look at third generation for CV3, CV7, and N1 family, the average ASP from $20-$400. So the ASP growth is going to be significant when we enable transformer, which is very easy to understand because to run a multiple billion parameter model will require much bigger AI performance and much bigger die. Therefore, the cost of the transformer-based AI solution is a lot more expensive than CNN base. The ASP growth per chip will give us a significant boost.
But at the same time, we believe that when we have this new GenAI type of application coming up, not only will it enable our current customer to upgrade their product cycle, but we're going to see new applications. So I think the unique growth will be there also. So I'm quite excited for both of the potential growths on that.
On both units and ASPs. You mentioned transformers enabling some newer applications. Can you give us some tangible ideas or applications where you're seeing transformers being applied? And is it opening up new markets for you? I mean, you talked about edge infrastructure, I think being one of those applications.
So, I think, let me give you two examples. One is on the robot side. In fact, autonomous driving car is a special robot. We've been talking about that. In fact, drones also, flying drone. I believe all the robots, including autonomous driving car and the drone, will use transformer type of end-to-end network as the final solution for the mass production software. Autonomous driving cars are there already. If you talk to Chinese guys, even Western OEMs, they already understand that the only way to go into high-volume production, give you the best performance, is go to the transformer-based BEV model. And that's happening on the car side. We haven't seen on the robot side because today, most of the robots still buy in components, buy camera here, buy radar there, buy central computer and pull the thing together.
The software architecture hasn't been as advanced as autonomous driving car, but it will be there. I think when most advanced robots go into production, the software will be end-to-end AI model based on transformer. That's just a first example. At the same time, GenAI based on transformer creates a totally different market segment. At the CES, we were talking about our go-to-market strategy, but we invited two of our software partners to demo their software running our N1 family chip. They brought in. In fact, we engaged with them in November last year. In six weeks, they put all their software and the model, GenAI model, onto our platform.
And what they demo is really something I never thought of before because one customer demoing that in a drive-thru for the fast food, that the camera, existing camera, picked up the cars and the large language GenAI model not only identifies the usual things that the model, the car, the color, the license plate, but also they determine how long it stayed there, what kind of food they ordered, do they come back again? Log all the time, and all the data being marked based on the GenAI type of models. This just creates so security cameras suddenly become your marketing collecting data, right? And that's definitely opened up opportunity. Same thing you can talk about Starbucks. People come in and out of Starbucks. You log all the customers come in. Or you talk about hospitality, about hotel.
There are so many different applications that we haven't even talked about. We didn't even know that our system can be capable of. Suddenly our software guys show us in six weeks of working time. That just shows you how powerful when you have a portable and programmable AI architecture for infrastructure, a lot of things can happen.
There's a lot of talk about agentic AI. How are you seeing agentic AI? Is that another wave after the transformer wave? Is it part of the transformer wave?
In my opinion, those two things go in parallel, and agentic AI can leverage all the GenAI things and also make it easier to programmable and also more powerful, so I don't think they are mutually exclusive, but for example, a lot of applications were showing this agentic AI already and also using GenAI model in there, so I think from that point of view, I don't think they will coexist for a while.
Okay. And so that agentic AI is part of the transformer wave that ramps over the next couple of years?
That's right.
Okay. If you separate your chips into endpoint and edge infrastructure, what model sizes can you support with your third generation chips?
Right. So let's go from the low-end side first. In fact, our low-end chip, we call CV75, which is our lowest-end third generation AI architecture, is a two-watt chip that we demoed, that we can run 2 billion parameters of a DeepSeek model, a reasoning model on that two-watt chip in real time. I think that's fundamentally a breakthrough because before that, nobody can even think that a two-watt chip can run large language models, let alone a reasoning model. That's definitely something I think not only surprised us, but also surprised all the customers. Now people are thinking how to use this kind of capability. So I think that for this kind of low-end chip, anywhere from 500 million parameters to 2-3 billion parameters is doable. And the cost, I can say that, is roughly $20 in that plus-minus.
So for $20 with a 2 W chip, you can do this kind of capability. I think that opened up definitely new opportunity for us. But if you look at our high-end side, today our N1 family chip can easily run 34 billion because we didn't push the envelope. We didn't design the chip for the higher model, but we can easily go to 70 billion parameter, even higher if we have enough different capacity. So from that point of view, I think that anywhere from the half a billion to 100 billion in the future, we're going to have a roadmap, a scope of different silicon to cover this wide range of application. That's where I think we spot.
How do you think about techniques like distillation, where you're taking the learning of much larger models and distilling it onto some of the smaller models? I mean, I know you're never going to need to have the same number of parameters supported at the edge as you do in the data center. But with distillation and your ability to support, say, up to 100 billion parameters, it feels like you'll have some pretty capable edge infrastructure, edge AI devices.
Yeah. Those kinds of new distillation model and also MoE type of applications definitely are going to make the edge application even more powerful because you can run, really can select for a particular vertical, you use a large language model, but you only use a portion of that to enable the application. And the DeepSeek model is a distilled model that we talk about. We can run a two billion reasoning model. I think we haven't figured out exactly what kind of application that can enable at edge, but sooner or later, somebody is going to figure that out because we are not doing application. We are helping our customer build application, but sooner or later, some of our customers are going to figure that out. And that we have that capability is important. We're going to continue to maintain that. That's a 5 nm chip.
When we go to 2 nm, we might be able to drop this to a watt, maybe even lower than that to run billion parameters. Then maybe you can put something here and having really nice applications to run through. So I think this is where we feel excited that when we continue to drive our performance per watt with our architecture and most advanced node, definitely there are better things will happen.
Okay. So I mean, to summarize, it sounds like today you think your hardware capability may be ahead of the types of models that are being run on the edge and you're sort of waiting for the models or the applications to catch up. But from your perspective, you're there now ready to enable these applications as they're developed.
Right. So in fact, when we demoed this two billion parameter DeepSeek to people in April last year, the first reaction is very consistent. I don't think this is possible before they saw this, right? Now people suddenly realize this kind of thing is capable, is possible. Then people start thinking how to apply for it. And I think that's where our customer is, and I definitely believe some of them will figure out.
They were looking behind the curtain for that link to the data center. Switching gears, maybe just coming back to sort of the markets and the business. You had a very strong fiscal 2026 entering the year. I think you'd guide it for mid-teens to high-teens. You're on your growth and looks like you'll exit closer to 37%-38%. I know you haven't reported the January quarter yet, but just talk about some of the drivers of the outperformance last or in fiscal 2026.
In fact, five years ago, I think you'll remember the biggest market for us is enterprise security camera, but it changed dramatically in the last few years. Last year, enterprise security continued to be a huge, very healthy business and continued to drive business and growth for us. But there are two other markets really show up and ramping up in a way that surprised us. One is telematics, right? Samsara type of a market. They basically put in edge AI without camera into the fleet management and add tremendous value. If you please go to Samsara's website and look at how they use AI to help to drive revenue growth on the service business, it's amazing, right? That business alone is driving our automotive growth in a range that we didn't expect. The other one is portable video, right? You remember many years ago we had GoPro.
This is much better in terms of, in my opinion, right? Because, for example, there are multiple different types of cameras available. It's not just a sports camera. You have 360-degree camera, a wearable camera, you have drone camera. So the whole slew of different applications and continue to drive ramping up all of that different time, but most of them ramping up last year. That both of them give us growth that more than what we expected at the beginning of the year.
Okay. Perfect. And looking into fiscal 2027, I know you haven't given guidance yet, but as you start to contemplate sort of your guidance or thoughts for fiscal 2027, what are some of the puts and takes that you're considering that could drive that 2027 outlook?
First of all, I still think that our growth in those markets is going to continue. There's no reason for them to stop, so I expect that our enterprise security continues to have a healthy growth, and I expect that the market will talk about. It will grow two ways. One is ASP, one is unit number. We talk about both. Obviously, we haven't given the guidance because we still have six months, six weeks to work out end of Chinese New Year. There's industry, all of some of the uncertainty on the market space. We'll take the time, but definitely in our February, when we announce our year-end announcement in February, we'll give official guidance for next year.
Okay. Perfect. You referenced CES show last week. I know this is always a major event for you to display new products and capabilities to customers as well as investors. Can you talk about sort of the three or four highlights coming out of CES, either from customer engagements and demos? You also had your product and technology briefing where you updated investors. So what are the three or four key takeaways from CES from your perspective?
This is not the order of priority, but it just comes to my mind. The first thing is we announced new product. It continues to expand our product portfolio. We already have a family of 15 AI chips. Think about it. From, like I said, $15-$400, any performance price point, we give you a chance, but we continue to expand on that, and we talk about CV5 is a major growth driver for us last year, and we announced our CV7 chip, which is 2x, 2.5x more AI performance and lower power consumption, believe it or not, and we figured out how to do that. You can imagine that everybody that designed us for CV5, they will move to CV7, so that, although it's really a progression on the product development cycle, but I think it's significant for us from the near-term revenue growth.
The second thing is really our long-term growth. I think we announced this new go-to-market strategy. By the way, it's not replacing our current business model that we go direct. In fact, in addition to the current business model that we do direct sales, we created a new go-to-market strategy that we try to engage partners, GSIs or ISVs, that they can come in here to help us to address those applications segmented with a small customer at the beginning. For example, physical AI, robotics. Most of robotic applications today, we have to deal with small customers, but tremendous opportunity. How to make sure that we engage as many as we can without using direct support business model, we can't. We cannot scale. So that's where we realized that we have a shortage on this. So that's the first thing. Also, edge infrastructure. This is a brand new business.
There are many different applications. We need a software vendor to help us to bring us to McDonald's, hotels, those things we just mentioned. We realized that we need to enable that, and to enable that, we need to open up a software platform, building a hardware platform people can buy, and we give them the complete platform. Any customer, without asking our permission, they can build software and pull software and evaluate performance on that. That's critical, so that's our new go-to-market strategy. We believe they are important for two reasons. One is to enable new customers in those more segmented market space, which we talk about that. The other thing is when we do this right, in the long run, it will build definitely a lot more sustainable tail revenue on our current existing business, and I think that's another reason we want to do it.
So that's the other thing that we feel really excited about. So I think the third thing is really about another thing. Again, it's not trying to replace our current business model. With our ability doing edge AI, IPs, as well as 2 nm, we start seeing a lot of opportunities that our large customers want to work with us to do custom chip or semi-custom chip design. We engage the first design win for our first 2 nm chip, and the large customers not only pay for the NRE, but also commit certain value. And this is a great business model. We kind of share the cost, but they get some benefit by putting some of their own differentiated IP into the chip so they can protect themselves and become more competitive.
At the same time, not only can we lock in a one big customer, but the chip can go to sell to other adjacent markets. This kind of business model, we look at it and say, wow, if there's more customers who want to do this, we'll definitely open it. But guess what? In the last 12 months, we have multiple conversations on this topic. We only engage with customers on two conditions. One is the market leader in a large market opportunity, right? Has to be large. Second, the chip they want us to build leverages our IP. So for example, they can use our perception IP or our AI inference engine or both. The ideal situation is more than 90% of the IP of that chip coming from Ambarella. Our customers are putting some of their IP in there so they can protect themselves, differentiate themselves.
But at the end, we can sell the chip to others without enabling their proprietary data. So that business model makes sense, and we are definitely engaging multiple conversations. I think that this business model makes sense, particularly for the customer that is large enough, but they don't want to spend hundreds of millions of dollars to build their own silicon. That's where we are right now.
You mentioned targeting very large customers. I think at the investor briefing, you were sort of saying, hey, these types of ASICs have to have $100 million plus lifetime value before you would start to engage. How many, with your resources, how many ASIC projects do you think you could support?
Not a lot. Right now we have one, but I think that one a year, maybe it's the right trade-off. But if we think this is a big opportunity for us, we can scale our ASIC to be that. But we are not there yet. With our current committed R&D dollars, we think we can easily do one a year. But keep in mind, this chip is not just for that customer. It also can be used for other markets.
So you can then take it, maybe disable the customer IP, but then take the design and sell it into it?
That's right.
Okay. And then lastly, as you go down this path of semi-custom, custom ASICs, how should investors be thinking about gross margin or perhaps more importantly, operating margin as it compares to the base business?
First of all, I think for the first model, first design win we had, that already factored into our current business model. So the gross margin, the total corporate gross margin is still 59%-62%, and this collaboration is not going to change that. So moving forward, I would think that would be a model. We want to try to repeat that. Definitely that I think is doable, particularly for those chips that we leverage and sell to other customers easily to maintain that 59%-62%. So from that point of view, I think that's going to be continuing a healthy business for us.
Perfect. I wanted to come back to the IoT business. I think historically you talked about security cameras. Professional security cameras in particular have been the biggest sort of portion of that business, but you've seen a nice diversification, 360-degree cameras, drones. Just talk about the diversification opportunity going forward. What excites you about IoT as we look into next year?
Right. So last two years, just so many new applications popping up. We talked about two of them last year. Now we talk about drone. If you haven't seen the drone, in fact, we started working with drone 10 years ago with DJI. So I'm kind of familiar with the drone market. I'm particularly excited about this opportunity, not because it's a drone, it's because they bring a brand new feature into a drone market that did not exist before. I suggest that you go to that website to look at. They basically provide you an immersive experience. You put on goggles, it's a 360-degree camera. So what you see is not decided by the direction of the camera. It's decided by how you move your head. You see different directions of the... So basically, this capability did not exist before, right?
You have to really play with it and understand the potential impact to the drone business. In the future, the guy who flies the drone doesn't need to know where he is, he didn't know where the camera is, he didn't need to move the needle to move the controller to decide where he wants to see. By moving his head, he sees everything around him. It's hard for me to describe it. You have to really experience. I have been doing drone for 10 years. This, I think, is a fundamental technology change for how people operate the drone in the future, and you can imagine, I think the majority of the drone market is going to move to this direction very soon.
Okay. And as you diversify away from security cameras, can you give us a sense within the IoT business today, how much of it is security-driven versus now these other applications?
First of all, I want to emphasize one more time, enterprise security continues to grow for us.
Yeah, sorry, I don't want to...
So I think if you look at the IoT business, if you look at security as one portion and the other portable video plus other possibility as another batch, security is a little bit less than 50% right now. In the past, it was more than that. You can see that the other side definitely grows very well for us. In this market, there are so many things. Let me talk about another potential market, wearable market. We started talking about the wearable camera market 10 years ago, right? Now it's become really significant because it's not only policemen wearing it. I know that our customers are selling wearable cameras to a 7-Eleven clerk, people doing retails, putting a camera there so we can record all of the interactions with customers.
In the future, you can imagine that this has become AI-based, so we can do a lot more with that camera. So with that, I think just that one simple application, we do see new applications on that. But there are many, many more applications in the IoT space.
Okay. I know we've mentioned edge infrastructure a few times. Your first edge infrastructure design, I think, goes into production in the second quarter of this year. You haven't sort of named a specific application, but can you give us some sense of the types of edge infrastructure applications you may be targeting with N1 and future generation products?
In general, I think the N1 AI box is trying to aggregate the edge endpoints. Let's say we do cameras. Let's say the edge endpoints are cameras. Let's say this hotel floor has maybe, I don't know, 10 cameras. You can imagine that you plug in an N1-based AI box, connect to a networking infrastructure, and because the networking infrastructure can stream all the video from the camera to that box, and suddenly you upgrade all the 10 cameras to be GenAI-based enabled, right, without changing the camera. So basically, that kind of aggregation makes a lot of sense for different types of applications. In fact, what we just talked about, you know, Starbucks, it's the same thing. The camera that exists for the security camera, you feed those video feeds to an AI box, and suddenly it becomes a marketing data collection box for you.
So, I think this kind of aggregates multiple edge endpoints and adds a new functional application to it is the first application we see. I imagine we're going to see a lot more in the future.
So it sounds like a lot of retail-oriented, excuse me, applications where you collect data and then you can do better product placement or you can drive greater sales of your existing products.
Right. And this kind of thing can go to hospitals, hotels, schools, you name it. Anywhere you have an installed-base security camera, you can turn them into totally different applications.
And as that business ramps, I think myself and probably many investors think infrastructure could be a higher gross profit or gross margin business. Does that come in at healthy margins relative to your 59%-62% long-term corporate average?
It has higher ASPs, but similar gross margin.
Similar gross margin. Okay.
Right. So the ASP, in fact, the N1-665 we talked about is a three-digit ASP, which is higher than any other ASP we have in the company.
Excellent. Just shifting over to automotive quickly, can you give us a sense of what you're seeing in the automotive market, demand by geography, and most importantly, sort of the interest in Level 2+ adoption?
First of all, I think 2025 is really a very weak year for automotive business as a whole. In fact, a lot of Western OEMs trying to decide on the design win for the Level 2 or Level 2+, most of them got delayed. The reason for that is they need to figure out their software strategy. Also, they need to respond to the pressure from the Chinese OEM that continues to increase the performance, increase the function and features. So I think Western OEMs are trying to figure out how to respond with their roadmap. So I hope that they kind of settle down and get to, in 2026, they get to a better year for that. But for us, we didn't change our strategy.
We continue to focus on the several OEMs that I think are within our reach that we want to get design wins from there, so that business development activity didn't stop, but more importantly, I want to say one thing. It's become clear for Ambarella that all the investment we did for the autonomous driving in the last five, six years will apply to all the robot applications. If you look at any robots that are moving robots, they will need a domain controller, maybe a lower-end domain controller of a CV3 family to control the whole robot, and drones are going to be the first market to go there. I expect to see a level three drone in a couple of years. That just shows me that in the future for drones, it's two-chip solutions for me.
One is to do the video collection, the other one is to navigation. Maybe at a certain time, they will combine together, but it will take time to get there. So the ASP is going to definitely grow even if the market didn't grow. So from that point of view, I think that our investment into autonomous vehicles or let's talk about robotic applications, we're going to continue to invest and leverage in that already heavily invested autonomous driving software.
For the automotive market, I know obviously you've focused on the CV2 family, CV3 family, but you're also developing the full automotive stack, and there's a software opportunity over time. Can you give us an update just on the software stack and the potential for software sales going forward?
Right. So first of all, I would like to highlight that a lot of people talked about end-to-end AI software for autonomous driving. Today, we don't have one model to do end-to-end, but we have two large models that we can do end-to-end AI today. I think that's really a significant achievement for us to go there. And this investment, like I said, will go to other robotic applications. But for even autonomous driving, we proved to our customers that we are advanced enough that people can enjoy, or if they want to license a portion of this model or maybe even a full set, we are open to license either a black box or source code. Either way, it's within our business model. So that flexibility definitely opened up conversation for certain OEMs that are looking for software solutions.
So we do have some software partnership with, for example, Conti for Aurora type of project, but that kind of thing is already in there. What we're really focusing on with our current software is how to secure design wins for both hardware and software revenue, and we want to sell that. But on top of that, licensing can be another way we generate revenue from the software side. And this licensing does not only apply to autonomous driving, but also applies to any drone or robotic applications that we talk about.
Last question for me. You talked about the delay in the Level 2+ market last year. Hopefully, we start to see more activity. I think the market's been waiting for Ambarella to announce sort of your first big Level 2+ wins. What's your sort of level of confidence, or how can you discuss your engagements for Level 2+ as you're looking to this year? When would you expect to be able to announce some of those potential wins?
You know, this time last year, we thought we won VW, and then last minute we lost it. So there's no guarantee in any form, but I definitely think that throughout the last year, for the company to really spend time to evaluate our technology, they like our hardware and like our software. So our technology is not the problem. The real problem that has been challenging upon us is always that we are too small compared to our large competitors. So what we are really focusing on is to make sure that to engage high enough management team to make sure that we won't be ruled out at the last minute of what happened last time. So from that point of view, I think we know where our weakness is, and we try to address heads-on. Hopefully, that will significantly improve our opportunity this time.
Okay. Perfect. Okay. We'll go ahead and get started. Thank you everyone.