Nasdaq London Investor Conference

Good morning, everybody. Good afternoon, everybody. So happy to introduce today Fermi Wong, the CEO of Ambarella, for maybe the sixth time at this conference, several times.

Several times already, yes.

Yeah, so I always appreciate you being here and a really interesting time to see you. You know, you've had a year. This has obviously been a big year around the Edge AI theme. You guys have sort of pivoted and refocused a little bit. Edge AI is now 80% of your revenue. Can you just talk about where you are big picture where you've come from and where you're going right?

So first of all, I want to clarify one thing which is that you know in my opinion We only building one technology which is really a G.I. Technology for hardware and software and this platform hardware software combination can serve many different application including automotive including we call IOT space in on the IOT enterprise security any new, a lot of new applications that we talk about, drone, enterprise, edge infrastructure. So, in my mind, there are many market segment opportunities because of the edge AI technology we provide. So, for me, we're going to continue to invest on the enable more and more edge AI applications, particularly today is video plus AI plus low power consumption. That's the focus of the company and that will be the core of our revenue growth. Of course, we are talking about moving to the edge infrastructure, maybe a non-video data will become also a play in the future. But definitely today, we are focusing on any applications that can take advantage of our hardware and software platform.

That's a really important point. I mean, we wrote, I think in 2019, our big debates report was on, you know, who will win the battle for edge AI, you know, inference. And MRL was featured prominently. So this is not something that's new. This has always been a focus. What's new is maybe the broadening out of application set beyond cars into these other markets. I guess, can you maybe talk a little bit about automotive? Obviously, this technology has really critical capabilities that you would use in the automotive market, but it's just been slow to see these kinds of features adopted. Can you just talk about where we are in that?

So I think you are talking about really the time of driving level two to level four. Maybe just before I answer that question, let me answer a little bit different question. Our automotive business is still 21% of our revenue and growing. And the area that grows is really another interesting AGI market that we developed starting two years ago. We call it AI telematics. The biggest customer is Sensara. And in that application, the first product Sensara uses us to do a camera facing outside the camera, facing inside of the driver to providing more and more AI function for ADAS and the driver monitor system. But now if you look at their promotion, they are talking about more and more HEAI functions they want to integrate to the solution. Now, not only include more camera, but start putting large language model into that space. So that is definitely another example. Two years ago, we didn't know even this application is part of our roadmap, but because Sensara has using our solution penetrated, make us realize that this technology can use, this market can use our technology. Back to your level 2, level 4 question. This is definitely a tough year for atomic driving. If you look at the, it's not only us talking about this message, but a lot of people in this industry, both OEM, Tier 1, and also Semiconductor Company. The reason for that, I think there are two reasons. One is with all the pressure from Chinese OEMs and also Tesla's FSD, people come to conclusion that, most Western OEMs come to conclusion that their product line needs to be reshaped to a wide source to become more competitive. And second thing is, the software stack, the atomic-driving software stack become the obvious weakness for the auto OEMs, and they are trying to figure out what's the best solution for With these two reasons, we saw fewer RFQ available to the bidding. And also, even with existing RFQ, they all push out trying to understand what's the right spec and what's the right timing. So with that, Umbrella's approach is trying to solve the obvious problem. We are trying to offer a software stack, a potassium-worthy software stack. But however, we are not bundling as a black box solution like our competitors. but trying to sell the software stack, enable the feature functions that our customer might find that they can use, they can license part of our software, or maybe even whole software stack, that we are opening up as a licensing model, as a white box solution. And we believe that by providing a scalable software solution that can scale easily from level two and level four, in fact, we'll improve that scalability to some of the OEM out there. And with that, I hope we can speed up trying to solve one of the difficult problems in atomic driving. But nonetheless, this year is definitely a very difficult year for any atomic driving suppliers.

Thank you for that. And I know you've always kind of led with the software-first mentality on these types of products. I know a lot of your engineering workforce is software-based. Can you talk about the importance of that and now as you sort of take this stack of software plus hardware and you can apply it to a lot of different markets, can you talk about the role of software?

Yeah, I think that's definitely important. You know, one thing, one statistic is important that although we are a semiconductor company, our engineering source, the highway ratio, the highway to software engineering ratio is 1 to 6. That just shows you how important the software size is important for us. But it's not only just software by itself. For any silicon, building a software SDK is important. And it's even important for us because every time we try to convince one of our customer switching their software platform from the NVIDIA GPU engine to our platform, the biggest resistance is how you help them convert from the CUDA to our Cooper SDK. And that is not difficult. It's really the mindset that, you know, I spent so much time getting CUDA already. Why do I need to spend time to convert to a different SDK, so the only way we can solve that problem is from the business side to convince people that there's a power advantage for them to move. But at the end, that software structure that is not only mature but also flexible enough for people to move any Kuda software to our platform is fundamentally important for us to have a successful business because almost every customer we have today, they all use NVIDIA in the previous generation one way or another. So, from that point of view, that mature software is not only just an SDK but a compiler to move help people to port any trained model to runs on our ship and also even in the application level that we show people example how to run application on our software platform. All of those software are important for us to win design wins.

Great. Thank you. And so then as you talk about those new verticals, can you talk about what they are? You've talked about edge infrastructure. Can you define kind of what that means and some of the examples?

So in fact, I would just say 12 months ago, our largest market is still enterprise security, and it's not true anymore. It's not that the enterprise security slowed down. In fact, we still see very strong growth on the enterprise security. is really that other areas of AGI start growing and going faster. We talked about two new markets in an earnings call two quarters ago. One is drone, one is infrastructure. And drone, you know, these two are totally different applications, in my opinion. But funny thing is they can use the same hardware and software structure that we are providing to our customers. And for the drone, we are offering two types of solutions. One is, if you only want a drone to capture video, in fact, one of the products that our customer introduced, they put a 360-degree camera under the drone, so when you fly, you're not only seeing one direction, you'll see everything surrounding you. And that really helps people to navigate the drone if you're using manual navigation. That's one product. The other product is really, a drone is just another type of robot. And all the, everything was developed for a time driving car applied to a time driving drones. In fact, you can view that most of the drones today, we call a level two plus drone because you still need people to manually control it, but there's a lot of already a time of function in there. And the drone will move to level three, level four, probably faster than the cars. From that point of view, you need a very powerful domain controller on drone to perform those functions, to avoid objects, to navigate, to understand the performance. So from that point of view, I think our CV3 family product that we define for the atomic driving will eventually go to the atomic drones. So that is a market that we think is important. Although the market is still relatively small, it's 10 million units per sumo drone today, mainly dominated by the DJI. But the window opportunities opened up because DJI got banned by the United States government so that 1.5 million consumer drone market in the United States opened up for fight. And we are seeing multiple of our customers fighting on that. So that's just one new opportunity from zero to meaningful to us very quickly. The other one you asked about, edge infrastructure, which is even more important for me. Edge infrastructure means, in the past, we sell our solution to what we call edge endpoints, cameras, or any form of cameras into a different device. But edge infrastructure is really the aggregate different type of cameras and performing higher level functions in a box that we never do before. So in fact, we announced our first product two quarters ago. And the applications for that particular application is very simple, is try to aggregate multiple camera feed and for example, in this hotel floor, say there are 20 cameras. Most of them probably is not even AI enabled, let alone the chat GPT. So if you want to upgrade those cameras to be running chat GPT type models, the easiest way is in your engineering room in this floor, plug into appliance box with one of our N1 chip, and the fit, those 20 camera fit into that box, and then you run the larger language model on that box, and so that the older fit can suddenly be upgraded by the chat GPT-ready. So from that point of view, that becomes the easiest way to upgrade install base camera. that, you know, for security camera alone is 2 billion install bays worldwide. So that's, we're talking about a huge opportunity, not alone for the hotel, but retail, any retail store probably have four to eight cameras. You can easily upgrade in the same way. So we are viewing that as opportunity. But we're still talking about video-related edge infrastructure. They're definitely, they're non-video-related edge infrastructure. I think all the corporations start talking about how to upgrade using, training their own LLM, but they want to run the LLM on the on-prime servers, not instead trying to run at the AWS or other cloud services. From that point of view, you need on-prime edge servers or edge infrastructure boxes that can provide similar performance. And why we have an advantage? Because all of the applications we talk about power efficiencies continue to be important. The engineering room here, I bet you it's not well air-conditioned. The power consumption is definitely a problem. Even the power supply to the box is sometimes limited by the configuration. So from that point of view, I think that our power-efficient solution for N1655 is suitable for that.

Maybe we could talk a little bit about that. In the past, I feel like we've sort of moved a lot of the intelligence onto the camera, where you're doing a lot of the edge AI kind of resident in the camera it's very clear you know you guys have been pretty dominant in that business um the value proposition's pretty clear of you know moving that intelligence into the camera when you talk about moving into an edge-based kind of box do you still get the same benefit of performance per watt are you more putting yourself in competition with gpus and things like that just just what's the value

proposition right so i still think that the performance per watt is important particularly for the first application I talk about. The engineering sitting here, a lot of box is supplied by power over Ethernet. So basically, your AI performance for the box is defined by how much the power efficiency that you can get out of that chip. So yes, power efficiency continues to be important. But there's another driver is really the to the The box itself, you know, most of a GPU box require heavy air condition, you know, water cooling system. I don't think that's widely available in a, you know, server room. Even in my company, I don't have a water cooling system in there. So from that point of view, if you really want to have a powerful, you know, on-prime servers, I think that power efficiency continues to be an important factor.

Great. I guess, you know, maybe if we could talk a little bit about the surveillance market, more, you know, home surveillance and things like that. I know that used to be a bigger category for you. There's a lot of price sensitivity. The cameras have to meet really low price points. But it also seems like the value proposition is really strong. And, you know, as a consumer of video cameras where you see all you can do is turn the sensitivity up and down, there's really limitations to that. When you talk about doorbells and things like that. Is there going to be an application for you guys as the sort of intelligence in those devices grows again? Or, you know, to what degree have you had to walk away from those opportunities?

Well, if you asked me this question 12 months ago, I would hesitate. But today, I am convinced there is definitely an opportunity. In fact, if you look at all of the home security suppliers that are working on Ring, Nest, they all are enabling a new service by user-only clip type of a vision language model on the server side. So the video stream, streaming from your home to the cloud. At the cloud, they store the video and apply this clip, vision language model on that so that you can provide more services. They are charging $9.99 per month for that service. But we all know that Ring and Amazon and Google can do that because they control the cloud. But all the other major consumer security camera customers, They, when they try to use the cloud to pull out the service, they are limited by the cost and also the transmission bandwidth, the storage cost, and the processing cost on the cloud. In fact, I will, in fact, when I talk to them, they are convinced that if this kind of similar service can be offered using an edge device, that some, if that the clip model can run on the camera, and although you pay a little higher price on the processor and the memory, but it can easily be compensated by the lower cost on the cloud as well as the transmission cost. From that point of view, I think that new service enabled by the vision language model is a clear way to upgrade that service, and I believe that, you know, our new chip can and run two billion parameters chat GP model for two-watt chip. That will definitely enable this kind of service in the future.

Yeah, I mean, the value of these applications really seems to be growing. Can you talk about robotics a little bit? And I guess it seems like drones is on the path there. You know, you see, you go around Los Angeles, you see little refrigerators driving around delivering stuff. Like, it seems like there's a lot of, you know, before we get to the humanoid robot upstairs, there's a lot of applications for vision in these robots. Can you talk about your view on that, Mark? Yes.

It's become clear. In fact, I have been saying this before. I view that autonomous driving car is just one special type of robot. That applies to drone, too. So I think today, if you look at the biggest robotic application is autonomous driving car and the drones. And new applications were popping up. And I think so when I look at this robotic application, I focus on what we call mobile robots. Any robot that need to move, and that can take advantage of all our investment on our CV3 technology designed for the autonomous driving. So, AMR or any other human role in the future, any drone needs to move and they need to understand the environment, need to find a way to maneuver over different objects and design the path they will need to move, and then finally decide what kind of function you need to do. This really sounds like a time driving car for me. So from that point of view, we believe we will continue to focus our robotic development on this mainstream revenue generation models first, meaning cars and drones, and use that to continue to fund our investment in this direction. That's why, in fact, we definitely continue to invest on autonomous driving car because everything we invest in that direction will be heavily reused in the robotic application. But the biggest problem for me in all the new robotic applications is it's very segmented. There are a lot of developers and they're all trying to demo and showcase their products in a prototype form. How to enable those guys is important for me because we're not talking about one, two larger customers anymore. We're talking about hundreds of different robotic applications, and we need to engage with them. So we do have a plan. In fact, at the CES, we're going to have a technology conference. We're going to highlight our new product and new technology, and we definitely will highlight how we want to develop a new go-to-market system to address these robotic applications.

Yeah, it's interesting because you've historically had fairly concentrated customers in automotive, enterprise security, markets like that. Okay, makes a lot of sense. You know, one of the questions we get a lot, particularly when you start thinking about these more consumer-centric applications, is gross margin. You know, you have a model of 59 to 62. You've had a tendency to walk away from markets where you don't see the value a little bit. You know, is that going to be the right margin structure as you think about your future business mix?

Well, in fact, all of the consumer occasion that you're talking about, look at our drone, We're talking about a $25 chip. In fact, the customer drone, the Pusumo drone they're selling is $1,000, so it's not cheap. So definitely there is value, and people want to buy high quality, particularly if you want to compete with DJI. The quality has to be one of the major concerns. So from that point of view, definitely price is important, but gross margin, I think, is important. The most important thing for me in the last few years is we gradually start to realize that, that while we try to maintain the 59% to 62% gross margin target, we are willing to trade off a little lower gross margin to higher revenue, therefore higher leverage on the operating margin side. That's the thing we're trying to talk about. I think we're only willing to do with large customers, and today, in the past, we talked about automotive customers can be one of them, but today, our largest customer is on the consumer side. So it's not the consumer side market driving us to lower price. It's really that they have the volume, they have the potential higher revenue growth for us, and that's where we're willing to trade off our growth margin.

And drones in particular, I mean, DJI was once a big customer for you guys, and I know geopolitics was part of the issue there. But is it also that there's just a lot more value going into these drones now? When you were doing more kind of image sort of capture, now you're doing more image analytics.

Right. So if you look at how DJI drone has been used in, although it's a consumer drone for the consumer video capture, but they've been reusing many different applications, right? And I've seen people using a DJI drone for inspection, for the, you know, many different type of other application that's not possible to use in any other technology. So drone, you know, to my surprise, when we were working on drone 10 years ago with DJI, the whole market was like 1.5 million units, and people think that would be saturated maybe two. Today we are talking about 10 million units of consumer drone, and out of that drone market, 9.2 is a consumer or a consumer, and 800,000 is a commercial. So I do believe that this drone market will continue to grow because people start identifying more and more commercial application. But I think the right approach for me is we need to focus on the customer who has an ambition to be the player in the consumer side or consumer side so that they can drive to the scale to get the best commercial scale so that they can compete in that 10 million units market. And with that, I think they will have a capacity to develop a solution for commercial drones. That commercial drone is a lot more profitable, but however, if you don't have the scale, you won't be able to compete with a company like DJI, who's already in the market and down in the market. So I think that the business model approaching this drone market is very important. I think that technology matters, quality matters, but more importantly, there is already a dominant supplier. You need to find a way to coexist with that.

And just to double-click on that, the military drone market seems like a very obvious application where you really need good computer vision, but it's also one that's specialized, people that are optimized around military applications. Could that be an application for you guys as well?

Well, we don't design chips for the military grade. However, I do believe some of our customers or design houses building a camera with our commercial-grade chip and selling to that market. But we don't have any customer that's really in a military level of customers. Okay, great.

Maybe if we go back to the automotive opportunity, I mean, the technology that you've delivered is really a breakthrough, and we've seen that years ago, and you've gotten wins with some of the biggest tier ones. that specialize in autonomy, and we just haven't seen adoption yet. You know, I guess where do you think that stands if you look over the next three to five years? Can people look at the advances of Tesla's FSD and do nothing? Do you think that, you know, that there's a call to action there that we need to start implementing some of these features?

Absolutely. In fact, one of the things we talk about, this is really a bad year for the autonomous driving, but people are still trying to figure out how to compete with AFSD. And now I start, I think, hearing people that are talking about end-to-end models in the Western world, which is a good thing, because without that, I don't think you can compete with AFSD. But however, to run the end-to-end model, both on the hardware side and software side, is a huge commitment. We know that because if you look at the software model that we work with, that the company will acquire, and we take a few years to get to a point that our software stack is too large a model. But to make that combine the two large models become one end-to-end model, it takes effort. But we know how to do it, but it takes years to get there. So I really think we talked about this just a few minutes ago. I think one of the biggest bottlenecks for us to get penetration into the market is we need to start selling our software in a way that adds value to our customer. How to get a better perception with our procession module that we can do sensor fusion between a camera and a 4D image radar and also running everything in a large end-to-end model that runs on our 685, we can demo it. When we demo this and take that software ready to be in production, I think that's where one of the solutions we think we can help that to resolve the current situation that people are looking for software stack and they haven't found one. But more importantly, we believe our approach is scalable. When I say scalable, it means I think our approach can scale from level two to level four. Of course, you need to reduce the number of hardware, number of sensors, but if you are training that model properly, you should be able to scale your performance down in a way that you can easily using a certain end-to-end model to address level two plus to level four applications.

I want to follow up on that. Let me see first if we have any questions from the audience.

Just wondering, what do you think the market is missing? umbrella?

I think, 99% of the AI investment is still on the cloud, although I think a lot of people here to listen to this presentation because you appreciate HAI, but I think the majority of the industry still think HAI is on the niche. if you don't think they can become then that's probably one of the reasons that they don't pay attention to umbrella. But I really think personally, I think, give another 10 years, I think HCI can be as big as the cloud. Because there are so many applications that you're looking at today. It has to be, you know, implemented on the edge. Robots. It's an obvious one. There are many other applications. If that latency matters, if the privacy matters, if the, you know, the private data matters, it has to be on the edge side. So from my point of view, I truly believe that when people realize there are new applications that will require running the AI on the edge that we should get our fair chance to be competing in the space.

to follow up on Otto, I mean, how much of these advances are tied to EV? Because it feels like with internal combustion implementing a higher degree of autonomy, there's just a lot of technology challenges that need to be solved with, you know, physical actuators and things like that. It's just easier if you're redesigning the whole vehicle around EV to start implementing these features, as Tesla has, as Rivian has. I guess, do you agree with that? And it seems like that's a really strong positioning for you guys, because a lot of the stuff that we're seeing in internal combustion, you know, is not going to translate into an They just can't meet the power budget that you can meet.

Well, if you asked me this question 12 months ago, I would agree with that. EV and the, and the atomic driving really come hand-to-hand, but now with the new people start delaying the EV distribution and slowing down the timeline for the EVs, I, we start hearing a lot of OEM customers start saying how we can implement atomic driving on the ICE cars. And in fact, we start seeing RFQ bidding on that. Because those cars need to have the atomic driving to be, stay competitive. So with the EV schedule got delayed, it really brings more attention to the ICE car and the automatic driving. So I think, although we're just hearing it, I won't be surprised to start seeing the automatic driving function being enabled on the ICE.

I mean, it's incredible to me that we've had the breakthroughs that we've had on reasoning models at the edge, and we've actually moved backwards in autonomy. It seems like we can only move forward at some point.

You know, I don't want to comment on the political environment, but that's a reality you deal with. But reasoning model, let's give you another example. You know, we can run a reasoning model on our 2-watt chip today. You know, we talk about this, that our CV75 is a 2-watt chip. We can run a 2 billion parameter deep-seq model on that. But the problem is, what's the re-application with the reasoning model for each device. I think whoever figured that out is going to be one of the biggest potential customer for me. We are not the one to drive application for each AI, but we are enabling all the functions that are not possible in the past, but now we are definitely thinking that with our silicon, we enable something that's impossible, and hopefully our customer can take advantage of that.

Great. Well, congratulations on all the progress, and we'll wrap it up there. Thank you very much.

Thank you. Thank you, guys.