Good morning, everyone. My name is C.J. Muse, Semiconductor equipment analyst, here with Cantor Fitzgerald. Welcome to Canaccord Global Technology and Industrial Growth Conference. Very pleased to have Lam Research and Doug Bettinger, CFO. Welcome.
Thanks for having me, C.J. It's good to be here.
Always great to have you. I know you have something to read, so I'll turn it over to you.
Yeah, let me start. I need to do a quick safe harbor, have a look at this verbal slide. I'll read a little bit, and then we'll get started. Today's discussion may include forward-looking statements that are subject to risks and uncertainties. Actual results may differ materially. Additional information concerning factors that could cause actual results to differ materially from those forward-looking statements can be found on the risk factors disclosed in our public filings with the SEC, including our 10-K and 10-Q. My lawyers are now happy, and now we can talk about what's going on with the company and the business, C.J.
Wonderful. So you've been witnessing a significant upturn, and now this is the second time we're together, and I think the last three weeks, you know, I wanna keep you happy.
Nothing's changed in the last three weeks, C.J., but I cannot.
To keep you in that good mood, I wanna start, you know, with your favorite business that you've highlighted many times, CSBG. Currently-.
Thank you for starting with CSBG.
1/3 of your total revenues and buttressed by an installed base of over 100,000 units. Why is this such an important business for Lam, particularly as you think about, offering critical support and driving your relationships with customers, as well as supporting your capital return plans?
Yep. Let's by the way, thank you again for starting here. I've had three meetings this morning. No one's asked about CSBG at the end of the meeting. It's important. It's a third of the business, as you pointed out, roughly speaking. The way to think about it is there's four components. CSBG is the Customer Support Business Group. It's spare parts, it's service, it's upgrading the installed base, and then it's the older products that we call the Reliant product line. Important component of what we have going on. The important thing, and we give a number at the end of every calendar year, is the number of chambers in the installed base. It was 102,000 at the end of last year. The previous year was 96,000.
Our installed base grows every single year. Why this is important is our equipment honestly runs for decades. It almost never goes away. Over that time period, on average, we generate more revenue after we sell the tool than when we sell the tool itself. As much as we're all super excited about WFE, and I am super excited about WFE, it's very strong, and I know we'll get into that, C.J. I often will describe CSBG as my favorite part of the company's business model because it just goes for a very long time. It's nicely profitable, and it's nicely profitable because it doesn't require a ton of investment. Not anywhere near the level of R&D investment that designing a new etch tool, as an example, requires. It does require investment, but it's not a huge amount.
That's important because it is nicely profitable, more profitable at an operating margin level than selling the new equipment, again, because of the more modest R&D investment. It is also, to your point about capital return, it is very cash generative for the same reasons. It comfortably supports our annually growing dividend and contributes nicely to the capital return and share buybacks. We're also very excited about what we're doing in the advanced service area using Equipment Intelligence, and we've got all these cool videos using cobots to provide service that as we think through how we're trying to innovate here, we're beginning to deliver service in different ways. Service in this industry historically has been show up and do a task.
We are innovating here by, I do call it, results-based contracts, where we know we can do certain things using more predictive type AI algorithms and the cobots to be consistent in delivering the service where we know we can guarantee certain outcomes. That is a really exciting part of CSBG in the advanced service area because it's able to deliver something for the customers that oftentimes they can't get on their own.
You've grown the business at a 13% CAGR over the last five years. You just talked about Dextro, cobots, bringing in artificial intelligence, trying to be predictive. Should we be thinking about a higher content, kind of driver there and that over time we could see potentially, you know, even faster growth?
I don't wanna get out over my skis too much in terms of the faster growth, but yes, what we try to do in the areas like I described, advanced services, is deliver US dollar growth beyond just the growth rate of the chamber count. We've been quite successful at doing that, and we'll continue to drive the innovation there to try to continue to do exactly that.
Just to kind of level set where we are this year, you've got it high single digits, low double digits for calendar 2026.
Yeah.
Are there still kind of headwinds within Reliant or is that kind of done?
Yeah, if you think about the four components I described, spares and service, generally, when utilization in the industry is high, consumption of spares and services is higher. Obviously, utilization is very high in the industry right now, given the strength of everybody's business. That will benefit. The fact that chamber count grew last year will enable that to continue to do well, spares and service. Upgrades, because of what is happening end-side customers, is quite strong this year. It was quite strong last year. And then as I think through what Reliant may or may not do, you have to think about what is happening in the mature node investments, globally. Flattish is probably the right way to be thinking about that, maybe up a little bit.
The fact that everything else is growing so strongly, that component probably is a little bit of a headwind in Reliant.
Gotcha. Makes sense. Maybe moving to the current industry outlook, a lot has changed in the last three months, and you kick-started semi cap earnings with the first kind of WFE official outlook, up 22%, most aggressive of your peers. What has changed and what is driving this robust outlook?
AI. The demand for AI compute is what's driving it. Obviously I think everybody in the room and everybody listening to the webcast knows there's just real strength in hyperscale CapEx. That's what's driving it, C.J. Yeah, we described a view that WFE this year, $135 billion, up from $110 billion last year. The incremental growth vector largely all about AI compute. It's accelerators, it's high bandwidth memory, it's KV cache driving the need for DRAM. When we look at all of that's kind of the incremental uptick. Now, also what we've been describing, and I think pretty much everybody in the industry has been describing it this way, is WFE is gonna be constrained by the amount of clean room that's available in the industry.
Demand right now is stronger than that 135. There's only so much clean room out there, and it takes time to expand that. That I think is what's gonna modulate the amount of investment this year.
I think you've talked about already 2027 being a pretty good year.
Well, just because of what I described.
Yeah, no, exactly. Curious, as you think about kind of this pent-up demand, how are you kind of working your supply chain to be ready for this? You know, you know, our house view is that lead times are likely gonna be worse this cycle than COVID cycle. Curious how you're thinking about that.
Yeah, listen, there's a large part of the organization at Lam whose job it is to manage the supply chain, to manage our own internal manufacturing capability, and those guys and gals are very busy right now. We are communicating with all of our supply chain partners about our expectations for volume based on what we're hearing from our customers and the timing of that, and what upside might be to make sure everybody's gonna be prepared and ready for it. Our objective at the end of the day is not to be the constraining item in the industry, and I think we will succeed at that. It doesn't mean that it's easy. It doesn't mean that it just happens automatically. It doesn't. We know how to do this.
I've got really capable supply chain organization that is working on this right now as we speak, and we'll manage it reasonably well, C.J.
Sounds good. Let's focus in on Lam. You know, 2025 was a great outperformance year for you. I think you grew revenues 40%, the market was up 10%. I go back to, you know, your intense focus on R&D during the COVID period that I think has really set yourself up for market share gains. You've already talked about being the best kind of house in the neighborhood for 2026.
I like to use that analogy.
You know, what are the key drivers of outperformance in 2026? Maybe, you know, compare and contrast versus what, you know, drove the outperformance in 2025.
I think, I'll just make one observation around our performance in 2025, which is almost 60% of our systems business was foundry logic. People often think of Lam as the memory company, and we love our memory customers. We do very, very well there. It's not that we're not extremely focused on them, we absolutely are. But we made conscious efforts to expand the footprint of the company quite successfully, as you point out during COVID. We grew R&D in a meaningful way because we saw some of the changes the industry was going to experience in gate-all-around and backside power and advanced packaging, that do have footprints in memory, but have very strong footprint growth in foundry logic, right? 3D structures evolving, which are etch and deposition intensive.
We were quite successful, I think, relative to the investments we made, the strength of the product portfolio, you saw it from us last year, and our expectation is you're gonna continue to see it. My comment on we live in a good neighborhood and maybe have one of the nicest houses in a good neighborhood is because architectures are inflecting in the third dimension, which grows our SAM, which grows our footprint, which grows our opportunity. We're very excited about that. I think Tim and I were quite happy with how the company performed last year, and we believe we're gonna continue to outperform this year, too.
Sticking with the foundry logic side of your business, we're now seeing signs of a second and maybe even a third kind of player emerge outside of the lead player. How is that kind of impacting your view of the world and WFE in 2026 and 2027?
Yeah, it's absolutely. When we communicate a view of $135 billion, it's all in for all of our customers. There's not any exclusion anywhere, that's for sure. Yes, it's good to have things broadening out. What is most important though is just the end demand, right? End demand is very strong. Yeah, it's good to have multiple customers supplying that demand, and that certainly is what we see.
I guess maybe sticking with gate-all-around, I think you talked about well over $3 billion in revenues there in calendar 2025. Where are we in that cycle?
For gate-all-around and advanced packaging put together was the disclosure. Yeah.
Where are we in that kind of adoption curve? And then maybe to add to that storyline, you know, it certainly sounds like backside power will be adopted by high-performance compute potentially first for products in 2028. I would think investments at least by 2027. How does that kind of inform your vision for growth and relative growth?
Yeah. Both are important. Both are 3D structures. Just to give you a rough order of magnitude on it, we have said as gate-all-around ramps because of the 3D architectures, our SAM expands by $1 billion for every 100,000 wafer starts at capacity that the industry puts in place for the gate-all-around sheets themselves. Obviously, we're excited about that. We've invested to do well there. We are doing well there, and so you're seeing that show up. By the way, curiously, we've said for backside power, it's also an incremental $1 billion SAM opportunity for us for every 100,000 wafer starts. Your observation of timing is correct, C.J. That's still kind of in the future on the come line, so to speak.
We're very excited about both, and both play to the strength of what we're good at doing, depositing material and removing material, etching material.
Final question on the foundry logic side, but really focusing on advanced packaging. I think you've guided that business up 40% this year. Curious, how do you partition kind of your participation across foundry, you know, versus HBM?
Both are very important. Honestly, and you know this, but I'll remind the audience. If you think about HBM growth, and you think about advanced compute growing in AI, they all need to go together at a system level, system architecture level, right? You can't do the compute without feeding the compute engine with data. That drives the need for HBM capacity. It drives the need for advanced packaging and foundry and logic. We do the TSV extremely well, extremely strong, in the Through-Silicon Via process. I call it the drill and fill, right? We etch the interconnect space, and then we deposit the copper conductive material using an electroplating process. That's a tool we call Syndion on the etch side and SABRE 3D on the electroplating side.
As capacity needs grow, we do well because we pretty much own those applications.
Maybe moving to DRAM. How are you thinking about contributions there, both from new capacity greenfield, 1c and beyond kind of transitions as well as the transition to HBM4?
Yeah. All of that is happening, and our position there is doing really well. Again, I think of the need for high-bandwidth memory and DRAM specifically as being driven at least incrementally right now by AI compute. Our share there is strong. Our footprint there is strong. The fact that HBM is growing, there's a trade ratio that requires incremental capacity. You know, it's meaningfully showing up. When I think about the growth drivers of WFE this year, DRAM is probably top of the list.
When I think about kind of the tremendous demand for DRAM bits, and really the lack of available clean room space, and the undersupply, you know, I'm hearing new tools being installed, ripping out rec rooms to bring in more clean room space within the facility, you know, upgrading tools because they have better kind of throughput and can deliver incrementally more output. Are you seeing benefits from that, and is there a way to quantify that?
Only that it's the biggest contributor to WFE growth this year, C.J. I don't know that we've given specific magnitude relative to anything beyond that. Yeah, it's about investment capacity in HBM. It's about the move to more advanced nodes that enable bit growth with the installed base. It's about upgrading the installed base, which, you know, we do quite well when that happens, and it is happening in DRAM. All of that together is driving the growth in WFE and the growth in our business.
Maybe the last question on advanced packaging. How are you competitively positioned with the eventual adoption of hybrid bonding?
Hybrid bonding is enabling all of the advanced packaging stuff. We don't have a specific hybrid bonding play per se, but we do things around it. As that grows, obviously the through-silicon via stuff that we do extraordinarily well grows quite nicely.
Would there be anything disrupted by that or no?
Nothing that I would point to, no.
Gotcha. I skipped one. I wanna go back to DRAM.
Sure.
You secured your first win for dry resist, so that's a low NA EUV with a DRAM player.
Yep. Ramping in production.
Yeah.
Right now as we speak, generating real revenue this year.
How's the customer reaction, and what more importantly is the reaction perhaps of the other two, you know, multinational DRAM players?
Listen, when everybody knows you're ramping something in production, if they're not the customer ramping it, they're looking at it and trying to understand, "Okay, am I missing something? Do I need to move more quickly?" Listen, I still stand by the numbers that we've described for dry resist, which is as we look at this in total, we see an incremental opportunity cumulatively of $1.5 billion over a five year time frame. The fact that it is ramping in production as we speak is a clear demonstration on the value it delivers, a clear demonstration on the fact that actually there's real value here, and it's gonna get pulled through from everybody else.
What does it say that it's one layer? Does that mean it's really comfortable with the technology and we're gonna bring it into HBM? Or is it where they found dry resist particularly great for one, you know, one area?
I think the fact that you're seeing it ramp into production one layer, that was the highest value application, obviously. Everybody that uses EUV has our hardware in the lab and is evaluating the capability here. That doesn't happen if there's not value here. Again, when we look at it, this is gonna broaden itself out. We've announced another tool of record decision from a second customer as well. The future is good for this, C.J., and this is all incremental business for us. This is business we've never done before, and it's hard to find new addressable market in this industry. It's rare that you find like a brand new thing outside of your existing SAM, and this is an example where we did.
I know you're working very closely with ASML on this run, and you know, they've kind of talked about five to six layers of EUV today in Korea going to 10 by 2030. I'm not asking you to corroborate those numbers, but more, you know, do you think that this is a tool where you could secure all layers or would there be-
Probably not all layers, but if those numbers are even directionally correct, the incremental opportunity grows.
Maybe to go back to your prior comment, you know, what are you seeing from a dry resist perspective on the foundry logic side?
Yeah. Like I said, everybody that uses EUV has got our hardware in their lab and they're looking at the capability.
Gotcha. Okay. Well, maybe moving to NAND. We're 20- minutes in. First question on NAND, Doug.
Yeah. Thanks, C.J., for like getting to it, but also waiting.
You called for continued strength here. I think, you know, the lion's share of spending is much more layer count increases. You know, where are we kind of in the upgrade cycle and are you getting sort of any inkling of more meaningful greenfield investments?
Yeah. The way I've described, I'll take you back, I don't know, roughly a year ago. We were describing a point of view that the industry would, over the next several years, invest $40 billion, largely focused on upgrades to get the bit growth that was required. I'll also remind you, a year ago, I think the industry generally was coalescing around a view that bit demand was in the mid-high teens. The world is different now. I think everybody believes bit demand is higher than that. When we look at, well, what happened last year? A lot of upgrades happened last year, C.J., right? Our upgrade business last year grew 90%. A lot of that focused on the NAND set of customers.
I think fast-forward to where we sit today, that $40 billion likely happened sooner than we previously expected because bit demand is stronger. At some juncture, what we've said is, yeah, you're gonna need some wafer capacity put in place. When we look at the industry this year, a lot of upgrades happening. When upgrades occur, actually wafer capacity goes down. You lose wafer capacity because throughput extends. That is what we see happening this year, that you likely have a reduction in wafer capacity, but at some point, the industry is gonna get around to adding capacity, C.J., is our point of view based on the strength of demand and especially related to KV cache driving the need for storage here.
That was the next question, KV cache Jensen announced GTC Washington, D.C. What are your thoughts here and any work in terms of the incremental bit growth for the industry?
Yeah, we've looked at it. I haven't quite articulated it. Bit demand is now this much higher, but it's clearly gonna be that much higher. We've done some analytics around for every 2 million accelerators, a percentage of incremental demand for NAND based on this. Again, it's an estimate. Everybody's got slightly different numbers. I think you've got numbers that are also out there, but it's incremental and we're excited about it, and I'll let my customers describe to what magnitude it ends up showing up.
Wow. five plus points to industry growth potentially.
I guess we'll see.
Yeah.
Right? You, you'll do your own math. Everybody will do their own math, but it's clearly incremental.
Maybe moving to margins. You've had multiple quarters now above your calendar 2028 target model of 50%. I guess, you know, within that, can you discuss A, pricing environment, B, what inning are we in in terms of the benefit from the Leipzig facility and C, are there other tailwinds that we should be thinking about?
This is a way I wouldn't want people to, like, jump too far ahead of the model. Yeah, we've delivered on it. There's some puts and takes relative to gross margin. But I think we are quite pleased with past year, we printed two quarters with gross margin above 50%. We just guided to 49%. We're in that range of the financial model a couple of years ahead of time. Your observations on the contributors to it are exactly right. We developed a couple of years ago a strategy from a manufacturing supply chain standpoint that we wanted to be close to where the customers were. That's enabled a more efficient cost structure for us. That is showing up in the P&L. It's pretty much in the P&L as we speak, C.J.
There's not much left there 'cause we've executed on this already. It's starting to asymptote in terms of that, but there's still incremental opportunity. Customer mix is probably a little bit of a headwind this year. If you think about the growth this year, it's the biggest customers which, you know, tend to get the best pricing. That's a little bit of a headwind, but we're managing it pretty well right now, I think. The other thing with the close to customer strategy, you know, the dollars we need to spend on freight and logistics, inbound, outbound benefit from being closer. That shows up as well. We should still be thinking about 50% as the objective of the company.
If you look at kind of your downstream customers, their margins are now extraordinary, particularly for the memory players. You know, in this kind of environment and given the value that you add, you know, is there potential willingness to deliver better margins on new products or is that challenging?
Listen, you're always working best you can to get paid for the value we're delivering. This year is no exception. Yeah, that's part of the calculus certainly. Yeah, we're in a good spot.
You know, we started off the conversation talking about CSBG and so clearly there's a bit of a razor-blade model for Lam. So how does that kind of play a role in, you know, maybe putting a cap on where gross margins can go on the tool side?
Well, it's certainly part of the thought process, which is when you're winning a position, it's not just, "Okay, we're selling the tool." There's the spares and the service and the upgrade opportunity that's gonna continue into the future. That's certainly part of how we think through how to put a package for the customer together and make sure they're getting the value, that we're getting the value, and it all generally works out in the portfolio at the end of the day.
Makes sense. You talked earlier about putting development closer to customers. Curious how that might translate into actually knowing your customers' challenges and problems better and feeding back to kind of corporate Lam and being able to bring to market maybe, you know, higher productive, more valuable and therefore higher margin kind of tool set.
Yeah, listen, I think we have for several years had a lab strategy to also not just have the factory close to the customer, but to have our lab footprint close to the customer. What that has meant is we've expanded, certainly we've expanded labs in the United States. That's still the focal point of what we do in both California and Oregon. But we've also expanded our lab presence in Korea. We're working in Taiwan. We've got lab footprint in India. The benefit of being closer to where the customers are is the customer can show up in the lab every single day. You can move wafers between their facility and your facility very quickly, as opposed to needing to fly it across the Pacific Ocean and back. You can drive it back and forth. That has been very beneficial for the customer.
The time to solution is much quicker. We've suggested that we can actually deliver time to the customer at twice the rate by being in closer proximity to where they are. I think it is a very differentiated strategy that we've adopted versus some of our peers in the industry.
Maybe a very high level question, looking out maybe kind of three plus years. If we continue on this kind of sustainable path for WFE and semiconductor revenues, you know, I think McKinsey talked about $1 trillion by 2030. They've now raised it I think to $1.6 or $1.7.
I think everybody generally thinks $1 trillion is gonna show up this year.
Exactly. If we are on that 1.6, 1.7, maybe even two, and even if you account for kind of lower WFE intensity, given the higher margin stack in high-performance compute, you're still talking, you know, a $200 billion plus kind of number. What do you worry about either for Lam-specific or for the industry to support that kind of growth?
It's execution, right? You have to deliver. You got to be ready for growth. I don't know if those numbers are precisely right, but like I described, we live in a good neighborhood. You got to make sure your house looks good and the customers are coming to visit you and that you're delivering for what they need. It's all right now about head down execution. Make sure you're taking care of the customer, managing your lead time, managing your supply chain, managing your quality, hiring people to install the incremental volume, getting them trained, getting them ramped up and being ready. That's absolutely what we are laser focused on right now and what historically at Lam we've been very, very good at doing.
We've got 30- seconds left. Low capital intensive industry, low capital intensity for Lam, delivering great free cash flow. How should we think about kind of ongoing capital returns?
Really no difference. You know, we've described a capital allocation strategy of returning 85% of free cash flow to shareholders. That's still what we intend to do. Growing the dividend on an annual basis, that's still what we will continue to do. We've done it every year, I think, since 2014 when we first put the dividend in place. Should be a good free cash flow year this year, CJ, and we'll continue executing on what we've told everybody we're gonna do.
Perfect. Well, thank you for the time.
Of course. Thank you, C.J.
My name is C.J. Muse, semiconductor equipment analyst with Cantor Fitzgerald. I'm very pleased to have Ambarella. We have Louis Gerhardy, VP, Corporate Development, reports to Fermi Wang, CEO, and responsibility for planning and capital market activity, including investor relations. Welcome.
Thank you, C.J.
Great to have you here. Maybe just to start, with the current environment, while you just reported earnings a couple weeks back with a solid beat raise, could you give us a brief recap of what you think were the most important takeaways?
Sure. Yeah, maybe before doing that, we have a presentation online, and please feel free to download. Take a look at Page 2, the risk and forward-looking statements. We will be making forward-looking statements, so please refer to those disclosures there. We reported about two weeks ago and a solid quarter, in-line revenue and EPS $0.02 or $0.03 better. For this full year, we're in our fiscal 2027. Our revenue estimates on the street went up about 3%. The story is like last year, we grew 37%. The year before that, we grew 26%.
This year, at the current time, we're expecting to grow 10%-15% year-over-year, really on the back of some very strong new product cycles we see occurring this year, and then again next year. That is the summary of the quarter. Unfortunately, as the call was winding down, there was an ITC decision that was made against one of our customers, and for about four hours, wasn't clear what the impact would be. We woke up our sales team in China, and they talked to the customer, and turns out there's no impact from this ITC ruling, but it did create some fear, uncertainty and doubt in our stock price, which I think it's led to an interesting opportunity today.
Just to follow on that, I think, obviously, semiconductor stocks have declined given what's going on geopolitically, but I think yours is down, you know, more than the stocks. I'm getting kind of what am I missing question from investors.
Yeah.
I'll ask you the same.
Yeah. I think there's two ways to think about that. The most important that you want to hear from me is kind of what we can control, and I'll cover that. I think in terms of stock price performance, the move from 70 - 50 or so where it is today, much of that is this FUD that was created by this ITC decision. Again, our customer Ambarella, we see no impact from that ITC decision. Just the short story, and I'm happy to go into it more, is that the product that was found to have infringed was already redesigned, and it's on the store shelves as of July of last year, 2025. That seems to be a non-issue, but it did create some FUD.
Then, of course, with the war and the geopolitical situation recently, that's created some other uncertainty. Just my view structurally of the market and how our stock trades is that, if you get a little bit of selling from the passive side, whether it's an index fund or an ETF or something thematic, then there's some FUD on the active side that just creates an opportunity like we have in front of us today. What are we missing? You know, it's a very good question in terms of our call. If you listen to our earnings call, I think you've got the full story there, you know, at this point in time.
The question we get fed back to us is, hey, you're talking about new product cycles, but your revenue growth is going from 37% last year to your guidance at the midpoint is 12.5% growth. That's quite a bit of deceleration. What am I missing? The first point I'd make is that Ambarella, you know, has a reputation of being quite conservative in how we give guidance. In fact, last year, when I was here, we were talking that our revenue growth would be mid- to high-teens. We ended up doing 37% growth as we could ascertain how our customers', you know, orders would play out throughout the year. We kept raising the numbers.
Not promising that's gonna happen this year for sure, but where we sit at this time, we feel good about 10%-15% growth range. The reason that we're not guiding higher at this time is very similar to last year. While we know we have, you know, some really attractive new product cycles, we wanna understand how will our customers take those to market? When will it start? What is the slope of their growth? How will our customers' customers adopt these new products? As we learn that throughout the year, we'll update our guidance appropriately. You know, when investors say, "What are we missing?" You know, that's how we respond. We've got rich new product cycles, and we're just taking a conservative stance with how our customers bring these to market.
Makes sense. Maybe taking a step back and taking bigger picture, you're building edge AI platforms. Ambarella is essentially transforming from an edge perception to an edge AI kind of company. How, I guess, should we be thinking about the required kind of ingredients to be successful here? You know, in terms of what you're targeting, you know, where does edge AI kind of begin and end?
Ambarella's heritage, you know, founded in 2004, IPO 2012. Heritage's first 15- years was video processing, or call it a perception module, where we'd collect information through the lens of a camera, and then data would be provided for human viewing purposes, frequently in a battery-powered device, so we had to be really efficient and provide high resolution. Around 2015, began to look at video analytics, which became AI. By 2019, we're out with our first proprietary AI accelerator that we didn't sell standalone, but we took the AI accelerator, and we integrated it in with a perception engine to single-chip AI SoC for the edge, where that single chip could ingest data from a variety of different sensors or even digital media.
The AI would do the processing and decision-making that would enable a machine to, you know, perceive the world and make decisions either partially or fully autonomously. Here we are, fast-forward to 2026, 80% of our revenue now is coming from edge AI. That business for us in edge AI grew about 50% last year. It is the story at Ambarella now. If you kind of break down the key functional blocks in our chip, I'd argue the AI accelerator part of our story, which is indifferent to how the data is ingested. Again, we can take in digital media, we can take in physical AI, many different types. That is our strongest value proposition. You asked also, you know, how do we define edge AI?
I think there's a couple of ways, maybe two ways I can quantify it. Before I even do that, Ambarella has historically in a network provided the human viewing and AI to this point at the edge endpoint of the market. Edge endpoint being the terminal device in a network. Again, it's often operating off of the battery. That has been our first market, and in all of our revenue, today in AI, that 80% of the $390 million we did was all edge endpoint. We have now said in addition to that, we're moving into the next ring of opportunity, which would be the edge infrastructure, which would be that first point of aggregation. That's a new business area.
Of course, we'll have a small amount of revenue this year, but we're starting to talk about more design and activity and articulate the use case and taking some other strategic actions to better position ourselves for that market. In terms of quantifying how we define edge AI, if we think about it in terms of like the AI TOPS performance, if you will, for us, we're in that range of 1 TOPS to 500, you know, effective TOPS type performance range. There's a lot of guys below 1 TOPS. Of course, there's a couple of guys, you know, above that range. That would be one way to look at it. Another way to kind of quantify what we mean by edge AI is how many parameters can we support.
We can support anywhere from tens of thousands of parameters for like a CNN type network up to 34 billion parameter models, in chips that we have today with line of sight to support a 100 billion parameter models, in the future. Those would be some of the ways how we define edge AI, what it means and what our opportunity is.
Maybe to level set kind of where we are today at that 80% of your revenues, you've got 12 different AI SoCs, and I believe you've kind of highlighted 370 unique AI customer projects. You know, is there kind of an easy way to define, you know, the end market you're targeting, you know, i.e. robotics, auto, edge infrastructure within kind of where we are today and how you see that kind of progressing over time?
The answers are very different for those areas. Our business today of that $390 million of revenue last year, 78% was from IoT, many different areas I'll cover, and then auto was 22%. Of the 78% IoT, think of, you know, three buckets. The first two, roughly 45%, be the similar size, 45% or so each, and that would be security, and the second one would be portable video. The third bucket has a lot of green shoots and in many different markets, some of which can be very significant. We'll cover that when we talk about the opportunity in the future. You've got things like access control in there.
You have wearables, you have robotics, many different form factors, and you have the edge infrastructure I was talking about before. You have enterprise video conferencing. All these areas are happening for us to some degree, but not generating much revenue today. On the auto side, and again, just covering our revenue from last year, it's 95% of the revenue, again, 22% of the whole company. 95% of that's coming from a category that's growing very nicely called telematics. It's safety, it's ADAS, like L1 - L2 type stuff. About 5% would be what we call auto autonomy, which would be L2 + to L4 type applications. That's how the business is set today. Turning to the opportunity question, what do we see as our biggest markets at this time long term?
Clearly, robotics, where we've started to generate revenue in drones. We've announced some wins in different form factor systems we can talk about if you want. The second market would be auto, and I think most investors are pretty well aware of our position in auto autonomy. We're still looking for our first material passenger vehicle win, but we have commercial trucks. The third area I've already mentioned is the edge infrastructure. That first point of aggregation in a network where multiple endpoints get aggregated in a single, you know, box and some AI is performed on all that data.
Perfect. At CES, you announced several additional go-to-market initiatives. I guess how incremental is that to what we just discussed?
Ambarella's invested cumulatively, like in that 10-year period when we first started looking at video analytics and then AI, about $1.3 billion into edge AI R&D specifically focused on that market. We've only had revenue really the last six years from the market. Our cumulative revenue is about $1 billion. That's changing because as I said at the beginning, you know, edge AI grew 50% for us last year, so that we had $310 million out of $320 million of edge AI revenue last year. Our AI R&D last year might have been $170 million, right? We've turned the corner, and that AI—edge AI business is growing very quickly.
What these new initiatives you referred to, the new go-to markets mean is that all of our business today has been direct to customer. We're saying, how can we accelerate the revenue growth even more with the existing technology we've developed? We've introduced two incremental go-to markets. One would be an indirect sales channel, which means we'd be able to support smaller to mid-size customers, and that is new to Ambarella. 100% of our revenue is direct today. We have no channel presence at all. That's one new initiative that's well underway. In fact, we had some announcement just yesterday at Embedded World, where we have a large presence in German exhibition that's happening right now.
You know, Macnica made some announcements about, you know, our indirect ecosystem. It's not gonna generate revenue this year, but it should start to contribute next year and potentially long term be very material because the revenue from that channel can be long tail. In terms of the incremental new go-to markets, indirect is one, and then the second one that we also kind of formalized during our CES briefing would be a semi-custom chip strategy. In that case, we'll work with either existing or new customers. We have a lot of interest in this, and they have come to us given our established AI position in edge. We ship 42 million edge AI processors to date, and they might wanna use our edge AI accelerator. They might wanna use our perception module.
Hopefully, they wanna use both, put some of their own IP in there. Maybe we give them field of use exclusivity in a certain market, but we can sell the standard product to many customers. That's a new initiative that actually will start to generate revenue in the first half of our fiscal 2028. We haven't said too much about it because there's a customer behind this with some field of use exclusivity, and they kind of dictate what we can and can't say. We have said this first semi-custom project is 2-nanometer gate-all-around, and we've taped it out during CES. As the wafers move through the fabs, you know, we'll give more updates and refine our thoughts.
Currently we're saying that first semi-custom project will generate, you know, revenue in the first half of next year. That's part of our new product cycle story I was talking about before.
Is that on the infrastructure side or?
We've said it's edge, endpoint related on the IoT side. It's not an automotive product.
How have discussions kind of proceeded since then for potentially other kind of design wins?
There's a lot of interest. I think it's because you had mentioned some of the data points before, but you know, we've really built a platform for edge AI with these 12 chips, having shipped 42 million edge AI chips cumulatively, having ported successfully 200 different model architectures. Not models, model architectures. We haven't found a model that we can't do well on our proprietary AI accelerator. The cumulative R&D investment I mentioned, we have 370, you know, customer AI projects that have gone to production. You know, this is getting out that we do have a strong platform in that portfolio of chips. Just as importantly, we have our Cooper Developer Platform , which is just critical in order to enable customers to get their software and their application stack programmed efficiently into our chips and into the market.
With that said, when I talk about platform, it's both the SoC as well as our Cooper Developer Platform .
Excellent. Maybe sticking with platform, you're starting to ramp your transformer business with your third generation offerings, CV75, CV72, CV 7, as well as N1, versus your just CNN second generation offerings. Curious, you know, what kind of customer traction are you seeing there?
Well, transformer, you know, synonymous with GenAI discussion, is something that we think is going to happen in all of our businesses. In fact, some markets, maybe robotics, couldn't really happen without it because the incremental utility of GenAI and transformer-based neural networks is so powerful. That, by the way, doesn't mean, you know, CNNs go away. CNNs are so efficient at doing processing for things like detection and classification of objects. They'll always be used. This new generation of chips, starting with the CV72, CV75 , CV 7, N 1 family, can do both CNN networks as well as transformer networks. It'll coexist, and the chips can do, you know, both type of processing.
Probably the first market you'll see it with Ambarella will be in the security camera market, which was, by the way, given AI at the edge use case is so strong in that market, five years ago, that was the first edge AI market that took off for us. It'll probably be the first to take advantage of these new capabilities of the GenAI.
I guess I think agentic AI has really kind of come to the forefront in the last kind of few months. Just curious if you take a step back and you look at that, how does that kind of inform the go-to-market strategy, the potential opportunities for you? You know, as we investors think about you building kind of your edge AI business, are there particular like end market subsegments that we should be watching closely? Is there a very large kind of pipeline opportunity for you, or is it really more breadth of smaller opportunities?
Yeah. Yeah, just kind of agentic, you know, brings you back to the kind of GenAI and what incremental utility it does offer. Of course, you know, reasoning and diagnosing problems and taking corrective actions and, you know, navigation is an example of that which is used in robots or cars, as an example where GenAI can be used. Creating content, whether it's using text to create video or creating text from video and summarizing the scene, which is another very powerful thing that some of our customers are doing. Then there's, you know, information processing that GenAI can do, which is, you know, summarizing a complex document or solving complex problems.
The one you asked about is really about operational automation, which is where agentic AI comes in and, you know, with a defined purpose, it can execute like multiple tasks. So it's a very powerful catalyst for the type of AI processing we sell because there's this agent, you know, might exist in a microcontroller, but then the processing probably gonna be different types of hybrid processing, where some is done locally in the edge, some might be sent into the data center. We think these different value propositions I mentioned will play out in all of our markets. It's really up to our customers. We show them what our processing can do.
We might use open source models, like if you see it at CES, we'll give examples to customers, "Hey, this is what we can do." It's really up to the customer to take those new capabilities and to create their own secret sauce, which is the application layer that sits in our chip. Our value proposition is really about providing the most efficient processing and development platform to enable a lot of different markets. The current state, you know, the answer to the last part of your question is right now we see just incredible breadth of edge AI opportunities. Some of them I mentioned are very significant today. Some are out in the future, significant ones, again, security market, portable video, and then a variety of different things in automotive.
This GenAI transformer capabilities enabling so many other markets to practically just become more intelligent and create an entirely new industry. We see more of the breadth now than that one pipe cleaner, but there will be pipe cleaners in the future, just not, you know, calling out exactly what it will be, but could be something in robotics, it could be edge infrastructure, it could be both. It's just hard to be precise with the timing on that now.
Makes sense. Maybe a couple of quick hits on other parts of your business. Within enterprise security, you know, how would you characterize the demand environment across both enterprise surveillance and consumer lending?
Yeah. Enterprise security, again, is how Ambarella first came into AI because the value proposition there is so strong for it. That market, think of globally ex China, maybe there's 600 million -700 million installed base, and it's growing at a nice rate with a very strong replacement rate, new product driven cycle caused by AI, which is all CNN today. For us, it's all for everybody. CNN in the endpoint is what the market is today, and maybe only 20% of that installed base is converted over to these CNN based cameras. We see that CNN penetration continuing to rise, creating new product cycles, and we see the GenAI wave behind it. Maybe with products announced this year starting to generate revenue towards the end of this year.
Behind that first CNN wave, then there's the GenAI wave, which comes with a higher ASP for us 'cause transformer processing, you know, is more complex. There's one other thing I'd like to say about the security market, which is our kind of foundational market for edge AI. Now it's diversified into so many things. That security market is called, you know, physical security, and that continues to be a good growth driver. However, these cameras, they all have their own IP address. They're plugged into the network frequently with Power over Ethernet, and they're starting to be used as a tool by businesses to run their operations more efficiently, whether it's in a retail store or it's a kiosk or it's on a factory floor.
You're starting to see what used to be called physical security or security cameras morph into an entirely new domain, which should cause the growth rate of that installed base to increase so that our business isn't mostly driven by replacement rates, but it's also going to be driven by an acceleration of the installed base as an entirely new category for these products begins to materialize. It's a very exciting space. Even though the market's been around a long time, it continues to offer very good growth for us this year as well as long term.
Maybe a quick hit on auto, 22% of your revenues. I think you're heavily tilted towards kind of safety telematics. Curious, as you think about kind of AI, generative AI, how is that kind of impacting your design wins or your kind of outlook, mix wise across safety, telematics, and autonomy?
Same story. It'll be used in certain markets faster than others. You know, I would say in areas like telematics is on a very rapid progression where there's more sources of data coming into our chips. Lots of interest in using new types of networks today. Again, the business is all CNN, but chips like CV72, CV72 give the customer an option to do CNN networks in that design today, but then they can do a software upgrade and just use the transformer capability in those chips. Telematics would be one example, but then you start talking about autonomy. GenAI is critical for that market in terms of being an enabling capability to help autonomous products just, you know, do reasoning. For example, I mentioned navigation and reasoning, but just helping that market, you know, really take off.
Where are you, I guess today within that market?
In auto overall, most of our revenue is in the first category, safety, ADAS, telematics. 5% of our revenue is coming from more on the autonomy side. We have commercial truck wins, so very complex systems like an L4 truck like Kodiak, you know, Aurora are some examples of trucks that we're in, but that's a lower volume, you know, high value type of market. What we don't have on the autonomy side of our automotive business would be passenger vehicles. We have a lot of passenger vehicles in that first bucket I mentioned, but on the autonomy side, we don't have the first big passenger vehicle win.
Gotcha. Maybe moving to financial model. I think, you know, one of the key aspects to whether it's the investment presentation you have on your website or, you know, your discussions at CES, I think that one of the keys is kind of your content story, as CV2 pricing $15-$75 going to CV3 $20-$400. How should we think about kind of the timeframe of that adoption curve of moving up the stack, and how should we think about kind of the impact to gross margins and what that kind of fall through will look like to earnings and free cash flow?
Yeah. A big part of Ambarella's story has been ASP growth, and we expect that to continue. All of our new products in design, things we haven't announced and things that we're beginning to define all command an ASP way above the $15 that we had for fiscal 2026. Lots of confidence. ASP timing of how that comes into the mix, you know, is the more difficult challenge. Just like this year, you know, we're taking a conservative stance and tell our customers we see more from them. In the ASP side, you know, lots of confidence.
Our customers at the edge are demanding just more and more AI in applications that used to be single-digit ASPs. They're now buying double-digit ASP parts and asking for more AI because they're using it for so many different features in their systems. Moving to other parts of the financial model, our long-term gross margin guidance is still 59%-62%. There's no change in that. We are at the low end of the range because in the last year, remember the upside that we delivered. We said, "Look, we'll take more revenue, slightly lower gross margins to drive that positive leverage over this huge, you know, investment that we've been making in terms of AI." That's resulted in, you know, our operating margins and earnings coming in better than expected.
That's where we're at with the model.
Fantastic. Well, I think we've run out of time.
Thank you.
Thank you very much.
Yeah. I appreciate it.
Appreciate your time.