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

Welcome to Day 2 of Citi TMT Conference. My name is Kelsey, one of the analyst here at Citi covering U.S. semis. We are very excited here today to have Fermi Wang, CEO of Ambarella, here with us to discuss more about the company and to share the outlook.

Perhaps to level set, Fermi, could you give us an overview of the company's heritage portfolio and also what opportunities you are the most excited here today?

Right. Thank you, Kelsey. And first of all, thank you for joining us today. Ambarella's fabric semiconductor company started in 2004. And throughout the years, we have been starting focusing on digital video application now in the last 10 years, really focused on building solutions for edge AI processing. Basically, what we are saying is that we're building a SoC that integrates not only just an influence engine, but also image processing pipeline, CPUs and peripheral so that we can become the SoC for the whole system. And there are so many different verticals we are targeting at. Anything related to the edge device, which require AI and influence. We are basically in there. The biggest markets like enterprise security to the portable video to the telematics for automotive. And now we are talking about robotic application. In the latest earnings call, we told about how to penetrate edge infrastructure with new partners that CapGemini and MagicHub. So basically, we are really focusing on edge AI for the last 10 years. And hopefully, the market will start taking off as people expected. But definitely, that's really what we are focusing on.

Right. On that front, what are you seeing today that gives you that confidence that we are sort of reaching inflection point for edge AI? I mean, we have seen a lot of news flow on acquisitions that big analog companies acquiring smaller edge-related companies. So could you talk about that a little as well?

Right. So first of all, I think our confidence is based on our engagement with our customers. In fact, all of the customers we talk to, they all want to put more and more AI performance into the edge device. It's now -- in 10 years ago, we start with CN type of network. Now they moved to transformer, then move to large language model. All of that are available in our silicon hardware software platform. So -- and one thing that becomes very obvious in the last 24 months is everybody want to continue to increase the AI performance on their device, mainly because the larger language model. So -- however, edge AI or physical AI, people have been really talked about a lot about, but the biggest problem for this market today is there is no one vertical that like coding for data centers, right? Today, majority data center revenues is generated by one big application, but there's no such application on at that edge. So that's where our strategy always focused on multiple different possible verticals and hopefully, the one that will grow. You talk about the acquisition. This ADI just announced a deal, which we acquired a start-up company for $1.5 billion, which has little revenue and their AI performance -- by the way, the AI performance we offer at Ambarella is in the family between 1 top to 1,000 top. And based on the public information I can find for that company is really about sub 1 type to 1 type of performance. So yes, they are edge AI company by this a little bit overlap. But -- and we are generating $400 million revenue versus [ Sun ]. So I really think that's indication that edge AI is happening and a lot of the companies said, I need a road map and they try to find a way to buy acquisition by using acquisitions to fill the gap.

Right. On that front, could you characterize Ambarella's differentiation?

Obviously. So every time we try to bid on a design win, the biggest differentiation for us is always performance per watt. That performance can be AI inference performance per watt or video processing profit per watt or other related functions per watt, right? So at the end, it's power efficiency. We understand that, in fact, most of our customers when they build their first-generation product, they use a GPU to do it because it's easy to program, widely available to do that. But when some of the applications they are using is really run into the barriers to power consumption because some of the bookings just require battery or using a very limited power dissipation solution. So from that point of view, power efficiency is the most important differentiation for us. But the second one is really that because of heritage is on digital video application, our digital video processing technology is better than I think most of our competitors -- maybe all the competitors out there that -- that's another huge differentiation. But the third, I think a lot of people ignore is -- if you look -- we look at our product portfolio, we have 15 AI SoCs that we offer to our customer. But on top of that, there's a unified SDK called Cooper that cover the whole product line. Basically, if you design one product based on our chip using Cooper, you can easily move the cell application to different silicon without too much modification. And that software architecture, I think, is critically important for our customer who has multiple pipelines to build and they don't want to reinvent the software from one product to another.

Got it. How should we think about as you compare in NVIDIA, they have their related Qualcomm, NVIDIA, Synaptics related offerings. How should we think about you relative to...

So you're talking about 3 types of competitors. NVIDIA, obviously, they can deliver really high performance. That's their strength. But like I said, GPUs power consumption has been well-documented problem. And we think that if you our customers who really want to have efficient, power-efficient solution, PU is not a bad solution. So from that point of view, kind of we carve out that based on politicians. Qualcomm is a little different. Qualcomm is definitely more efficient than the GPU. But however, when they compare to our approach, we still have a significant power efficiency. But on top of that, we continue -- but because they are closer, so we definitely have a lot of overlap with Qualcomm. So we always view Qualcomm as the biggest competitor to us. Synaptics is different. Synaptics is just -- they came in to play this AGI 2 years ago with -- by acquiring product line. And their coverage is really from 1 top to maybe 10 top type of location. So yes, there is some overlap, but it's not a lot. So from my point of view, we don't run into Synaptics that much. But when we go out to compete, we're going to NVIDIA and QualcoMatt Smith all the time.

So the IoT segment currently is roughly about 75%, 78% of total revenue today. Consumer -- sorry, security and consumer video, roughly about 45% wearables, 10%. Could you walk us through the key demand drivers for each of these end markets?

Enterprise security and the...

Yes. So consumer video...

Well, enterprise security is really our largest application that we've been addressing for is really for the security camera in the past. But I think that market start changing. The way it changes is that a lot of security camera become serving different purpose. For example, when we talk to edge infrastructure companies, one -- I just give you an example that, for example, in the Starbucks, you have many camera serve as a security camera in the past. Now people say, I want to link, connect all the security camera to a box. And in that box, you run ALM so that now that box can tell you how many people come in, how many people stay there, how many online, how many -- and how much time each one of them spend in the store, what they buy and how they pay for it, whether the cafe shops clink. So suddenly, ALM can start providing customer marketing data as well as operation. So that -- but you're using the same security camera hardware. So from that point of view, I think that the driver in the past was purely security. Today, we are start seeing a lot more marketing-driven solution for that. That's one. On the consumer video side, in the past, after iPhone was introduced, cell phone basically replaced all of the steel cameras. But now there's a new trend, right? In fact, people building really high-end video capture device, handheld stabilized handheld device and start selling at a very expensive price, but the volume is at the order of 10 million units a year. So that's -- those high-end video capture device become a new category that we are addressing. That category including all the 360-degree camera and high-end sports camera pocket, there is no handheld stabilized camera, drone camera, they all package into that. That market by itself, we estimate anywhere between 10 million to 20 million. So -- and ASP is pretty high because not only you need high-end video processing, but also you need to be able to run AI, a lot of AI on that. So from that point of view, it becomes a sweet spot for us. So that's another driver for the consumer side.

And wearables is about 10%. Is there any growth driver there?

Well, in fact, in the past, wearable is really for police, law enforcement. Now we are seeing a lot more -- for example, a lot of our customers are talking to us that the retail stores, all the people in that store wearing the wearable camera, start documenting the conversation and providing actual angles of security. So wearable camera go way beyond law enforcement and we are starting seeing that. And right now, that market is really limited by power, how much power you can -- power consumption can be made available in that camera and how long you the life -- the battery life, you can have those. So that's really played to our strength. However, I rethink we are still at the early stage of this wearable because I think when wearable camera become really expand to all of the service category that everybody should wear one so that can start collecting -- became not only for security care, but like collecting the marketing data, the purchasing habit, all customer data can be -- also become quite interesting for the people who use those kind of wearable cameras. So I think wearable camera start from the law enforcement now turning to a marketing tool and will continue to evolve from there.

Is there a way to think about the growth rate for these 3 end markets?

Those markets, obviously, Enterprise Securities grows at a certain rate because it's very mature. And I think that consumer market is really growing fast in the last few years and wearable definitely grow much faster -- so -- but however, the base is small. And there is also a similar market at robots, robotic application. People talk about it all the time. But the base is so small, most of the customer there is a small, midsized customer, although the growth rate is high, but the base is low. But another one that we talked about in this time to disclose our same discussion is really the edge AI infrastructure, which we believe that people will build each box or edge appliance. In the past, most of them just have a powerful CPU in there. In the future, will be CPU plus a lot of AI inference engine. So that will become an interesting market, and we think that's definitely another huge growth area for us.

Could you talk about how that will play out? Is that on our phones? Is that...

I think it's -- for example, today, if you go to any retail store, right, if you look at their back end in the store, they have probably a huge box, huge Intel CPU box sitting there for inventory or for the pricing or the cash exchange. And most of people said, okay, I have that CPU box, but I need to add a lot of AI too. That's where we think opportunity. Almost every industrial PC or kind of edge CPU box in the past, when they try to operate to the more AI performance into it, I become opportunity for it.

Okay. Got it. So I know that you guys read had a relationship with Hanwha and signed an LTA deal of about $800 million. Could you share what prompted the move to the LTA deal? And why now?

Right. So in fact, this process is initiated by Hanwha. I think it's -- Hanwha is a huge Korean group that has many, many different subgroups. The sub group we work closely was called Hanwha Vision, which is the security camera division. In the past, we sell our chip to Hanwha Vision, but Hanwha Vision also developed silicon for the edge video product. And they approach us because as a whole group, Hanwha decide they have to have a unified edge AI strategy. They want to have one hardware, one silicon, one platform for the whole group. And that makes a lot of sense. But however, they try to evaluate who can build become a pioneer on that. The requirement is very simple. There has to be very power efficient. They need vision processing. They need a very powerful AI influence engine and they need to access 2-nano, 4-nanometer process node, right? If you look at the combination of that, Ambarella become one of the very good candidates to talk to. So we talk to them, we quickly come to the conclusion that it's beneficial for both sides to have collaboration. Obviously, there's other business discussions come at the end. But at the end is really that we help Hanwha to build a family of chips that they can use internally and also give them a certain benefit. In return, they're basically helping basically give a spec and making sure that we can become the dominant supplier -- semiconductor supplier to the edge AI strategy.

Are there any other opportunities where you see these kind of long-term agreements will happen?

Absolutely. In fact, when you look at the data center market today, almost all the large companies are looking for their customers. Everyone wants to build their own chip. And because they can afford it, we'll go Amazon, they all can build their own chip. But when we come to edge AI market, there are a lot of large corporations. But if they want to do a custom chip, they will find out that they can't -- the economy doesn't work out for them. Because the size is not as big data center, the price is not as big. So they -- the best case is if they want to build a custom chip themselves, they found out that build 2-nanometer chips too expensive, they cannot justify. So this is why the semi-custom chip business model comes in. We believe that we can help those large corporation build a chip they can use. But in exchange, we can sell the same chip to other market that they don't care. So just I believe that just like in the data center, many large corporations want to control their silicon road map, that apply to edge AI market also. So when edge AI market become bigger, I expect more and more large corporations say, how can I control my own silicon road map. But then I think that will add to our opportunity to grow in this business.

Got it. I'll pause there for -- and opening up to the floor for question.

So your SoC platform clearly in the competition, especially on the energy perspective. Have you received any validations from automotive OEMs specifically like your software layer on top of that?

Well, we don't have any design win we can announce. But however, we've gone through many cycles of evaluation. So the -- from the auto OEMs point of view, we have announced level truck deal like Aurora. And also, we announced automotive OEM -- sorry, Tier 1 like Conti and Bosch, they are partnering with us. We obviously continue to talk to potential partners in that. So that's on the time driving side. But our total revenue, 30% of total revenue comes from automotive. That's come from other areas like telematic for commercial vehicle telematics, several large companies in the telematic is using our solution today. And also that electronic mirrors, ADAS, DVRs, we have significant design win. So in fact, Q2, we have a record automotive revenue. But all of that has come from the one side of this auto-driving we still need to get or design wins from big OEMs. Any other questions? Yes.

[indiscernible]

Between the OEMs -- sorry -- right. So first of all, we start serving this edge AI with our, we call CV2 family chips 10 years ago. We start developing that. That family chip is really focused on relatively simple one compared today is CN-type network, base detection, obkect detection, those simple tasks. That's where we start. Then starting 5 years ago, transformer type of network come in and large language model happen. That's where you start triggering fast pace of AI performance requirement. And every time there's a new model coming out, that trigger more AI demand. Just one example, robotics right now talking about using VRM, VLA, now VRX, just look at the performance requirement among those models is tremendously different. And every time the model changes, we customer engage say, okay, are you going to handle this? So from that point of view, different applications definitely have different paths of the AI performance requirement, but all of them is on the path of getting higher and higher. In fact, we talked about X7, which is our accelerator product just in this earnings call. The reason for that is when we talk to our existing customers say, okay, the LLM changing so fast. And in the past, every time we build a product, the product exists for the market for 4 to 5 years. Today, nobody believes that for one product defined for this year, 4 years down the road, the same LLM still works. So I think then become -- how are we going to address that problem because the performance requirement is so big. So the easiest way for today is we design a baseline product with SoC chip, and we leave option to add our accelerated chip X7 into that port when it's required. So you design in a fundamental hardware and software and just -- and also add option that you can increase the performance. And those kind of option was not required in the past. And the only reason for that is really that nobody can predict how fast LLM evolve in the next 5 years. So that's just kind of example how the conversation we have with our customers.

Could you talk a little bit about robotics at humanoids? I think this is a theme that people are just sitting on or thinking that's going to be coming. How close are we? Or what are some of the customer conversations?

Right. So first of all, robotic and humanoid is quite different, right? In fact, here's my personal opinion. I think humanoids from a technology point of view is more difficult than Level 5 autonomous driving car, okay? And Elon Musk promised us in 2015, we will have Level 5 cars in 2016. And in 2026, we still have to have that. I just show you that the best engineering team, the best engineer talent still haven't really resolved the Level 5 autonomous driving car. I'm not saying humanoid is going to take 10 years to do it. But if that -- if I'm right, if I -- if that problem is harder than an autonomous driving car, it will not happen next year. It will take more time to evolve. So from today, to when the humanoid become available, we're going to see multiple generations of different type of robots performing different functions at different AI performance level. And just that's an opportunity which will focus on so trying to really put our eyes on the home run opportunity. So our approach is we engage with humanoid company. We have a couple of design wins in that category. But the volume is not really -- I won't bet that's going to change materially about our financial performance, but that's really to try to prepare everything on the road map we have both on the hardware and software side. So we won't miss the opportunity when it happens.