Hi, Jim.
Hello. Good morning, everybody. Welcome to the Goldman Sachs Communacopia + Technology Conference. I'm Jim Schneider, the semiconductor analyst here at Goldman Sachs, and it's my pleasure to welcome Cadence and CEO Anirudh Devgan to the stage today. Welcome, Anirudh. Thanks for being here.
Thank you. Great to be here.
I've been asked to read a safe harbor to begin. Today's discussion will contain forward-looking statements, including Cadence's outlook on future business and operating results. Due to risks and uncertainties, actual results may differ materially from those projected or implied in today's discussion. With that out of the way, let's get rolling. First question for you, maybe high level. I think Cadence has had a very strong first half of the year with double-digit growth across pretty much every product group, a record backlog, and two increases to full -year guidance. Before we get into individual businesses, how would you characterize what's changed in customer behavior over, say, the last 12 months?
Yeah. Thank you for the question. The customer environment is probably the strongest I have seen it. Because last few years, of course, some companies were doing phenomenally well, the big AI companies or the hyperscalers, but some of them were not. But if you look at in 2026, universally, the industry is doing great. The semi -companies are doing great, and then all the system companies, hyperscalers. The commitment to silicon is the strongest that I have seen. Because sometimes we used to get questions a few years ago, "Well, will all these hyperscalers really do chips or not?" But you can see now the success of. I think that what I would say at the highest level is the environment is good, and we always try to check how long this party going to last, but looks like party's only getting started.
I talk to all the people. They'll be very confident next few years. That's number one. Number two, I think our products are performing great. Of course, we are in the tech business, so best product always wins, and our competitive position is very strong. That's second. Third, we have this new TAM opportunity, new expansion of agentic AI on top of our traditional offerings. So that's all new TAM for us. If you put it all together, these three things are what is driving this growth that you are seeing. Yeah.
Great. Now, you framed Cadence's differentiation as a three-layer cake, especially tempting as we get closer to lunch here. Anyway, a base that's accelerated compute and data, middle layer that's physics-based simulation and optimization, and a top layer of AI agents. Why is that particularly relevant for EDA versus other kinds of software that are in the market today?
Yeah, I have been only saying this for five years now, I think, the cake, and people say, "What?" First of all, things have to have three things. Answer to life is E, the universal constant. This is what my advisor used to say, 2.7. Because if it is less than three, it is too little, and if it is more than three, nobody remembers anything.
Pi is more than three.
Yeah. Pi is slightly more, 3.1. So whether it is 2.7 or 3.1, you can choose your favorite universal constant. I think three, and the reason I call it a cake, you could call it a stack if you want, or the reason I call it a cake is because if you eat a cake, unless you are a two-year-old, you eat all the layers together, and you have to bake all of them together, means they interact with each other. So that is the reason to call it a cake. And the reason I put AI at the top and compute at the bottom, you can put it in because, first of all, the middle layer is super critical, and this is going to happen in all. By the way, it is going to happen in all markets, not just EDA or not just chip design.
You have to ground the AI with physics, especially in these kinds of complicated engineering software or engineering workflows. In some cases, the middle layer may not exist or is maybe simple, but definitely in our business, you need to ground the AI with the physics and then, of course, run it on compute and data. So all three are critical. And the real value will accrue to the vertical application, not the horizontal. Because in the beginning, it is always horizontal. In the end, it is always vertical. Waymo is a vertical application, for example. So they have a AI model. Do you know what model it is? It does not matter. It is can you get from point A to point B? They, of course, have control theory navigation, and then they have the silicon, just to give an example. And same thing will happen in chip design.
The reason I put agents in the top is because agents are very good at directional kind of orchestration. If you want to go from here to Palo Alto, that's directional thing. But actual navigation and detail, they are not as good. But they're great for orchestration, planning, and optimization. So that's why the top layer calls the middle layer that sits on the compute layer.
Yeah. Okay. Now, the bear case that investors often raise with me relative to the EDA industry is that if you have a sufficiently capable frontier model, you could basically automate chip design from prompt and basically bypass the commercial EDA software flows. Why do you believe that's wrong? Specifically, why do you think deterministic physics-based engines and proprietary data are kind of essential, especially for leading-edge designs?
I think they're all going to be important. One thing with AI is people who graduated a few years ago think, "Well, I will make a model of everything. What do I need to know? What is this?" Then people who graduated 30 years ago say, "Ah, it's all curve-fitting. What you need to know is reality, how things work." Whether it is physics or mathematics or economics or whatever it is. The reality is you need both. There's no need to take a side in that. You need both. To have a successful thing, AI itself, it's a nonlinear curve fit, right? That's what these LLMs do: you give it input-output, and it fits a nonlinear model to it. It used to be a transformer architecture. But fundamentally, they cannot do nonlinear differential equation state varying. This is mathematically not possible to do.
But together, they can provide a good combo. I do believe AI, like we have seen, can provide more scenarios to optimize, then that can be optimized in the physics-based layer. So mathematically, it's not possible to do the middle layer. But we want to innovate in all three layers. We just don't want to innovate in the middle layer, which is classical physics-based. It's the combination of the three layers that will win.
I got it.
Yeah. You'll see that more and more in all industries, yeah.
Yeah. Then why, I guess the other question follows on is like—
Nobody's trying to do that, by the way. All the LLM companies, all the hyperscalers, they're all using our tools to design chips.
Yeah.
Just to be clear.
Yeah.
There are no chips being designed without using our tools, yeah.
Just push that back for a second. Why do you need all three layers together? Why can't we have somebody else's solution for the top or bottom layer and yours for the middle?
Yeah, that could happen. Yeah, you could have an agent, and some customers are writing some agents that call our tools and not use our agents. That could happen. What you have to remember is the top layer is a brand new TAM opportunity for us because what the top layer used to happen, these AI agents, was basically done by humans in the past. Okay? What agents are doing is they're not replacing the middle layer, they're replacing what humans used to do. Okay. Now, in some scenarios, agent could call our tools, but it is not that efficient. Because we wrote the middle layer, we wrote the top layer, so a lot of times we have access to the internal that is not exposed to the user.
Yeah.
But it will be natural for some users, especially in the beginning, to write their own agent. But in the end, they realize, okay, it is more efficient for Cadence to do it. And we have these four super agents, which are more aligned with functions. So like tools, the middle layer will have 30 or 40 products. The top layer we have four super agents, like front-end design, physical design, analog design, and PCB, and packaging. So they are integrated closely with our middle layer, and we have unique advantages. We have 10,000 people in R&D, so they are writing both the top and middle. But even in the top layer, we do not need to get 100% of that market.
Even if some of it is written by our users, or they could have 10 agents, but the four big ones are by ours, and six could be their more domain-specific, that is all fine. Even in the traditional flows, a lot of customers do customization on top of our tools.
Yeah.
Yeah.
On agentic, how do you think about the monetization of agentic? Specifically, where do you expect to drive incremental revenue above and beyond what you are already doing? Is that the new agentic workflow products themselves? And how do you think about the opportunity for a higher consumption of your existing tools?
Yeah. It will be combination of, like we have new business model for the top layer, which is consumption plus subscription. And then, of course, our existing business model for the middle layer. Okay. And a good example of that is, because one worry always is if something is like, let us say, 5x more efficient, then you will use one-fifth of the middle layer. This is also some perception in the market.
And this is not new. Even actually in 2006, I launched a simulator, and it was 10 x faster, and then my marketing team was worried that, oh, people will buy it 10 x less. But that never happens. That is the history of EDA. And the reason for that, there is a fundamental reason, which is different than almost all other software markets. See, that is why EDA is so exciting. And sometimes we get lumped in general software.
Those guys never thought we were software, and we never thought they were software. Our software is so mathematically complex that accessing a website or database, we don't consider that. That's just one small part of what we do. And they thought, oh, we are semiconductors or something like that. It doesn't matter. I think what happens in this kind of application, EDA or chip design, the workload is exponential. Workload is exponential. If you look at TSMC roadmap, next five years, they said that chips complexity or size will go up by 48x. This is not happening in any other software market. I talked to some customers, our big hyperscalers, they're saying every year, if they continue like this, they need to hire 2x more engineers. It's not sustainable. If the workload is exponential, the requirements of headcount is exponential.
You need this 5x, 10x automation. If the chip size is going to be 50 times bigger, there's no way they're going to hire 50 xs more engineers. You need this 5x to 10x improvement to even sustain the growth. I think the customers' headcount will grow, but with automation, with AI, will be less than. This is the history. If you look at the late 1990s and early 2000s, our customers would design a CPU, it would take them five years and 500 people. This is not uncommon in all these IBM, Intel, DEC companies. Now you can design a CPU with 30 or 40 people within six months. That's 100 x faster than 20 years ago. And the amount of silicon is only going up and amount of design activity only going up because exponentially, the size is exponential, also the applications are.
This is going to continue. If you look at the roadmap from IMEC and all that, this kind of exponential is still projected to go till 2042, which is still how many? 16 years at least. And by then they will have some other technology. This is not going to slow down, which is very unique to any other software market. We are always looking at improving the efficiency of our solution, and it gets absorbed even faster. You look at all the roadmaps from NVIDIA or Google or Apple, and they are doing even more and more with that. This is something not to be afraid of. It's something to embrace.
Yeah.
That the productivity will actually help us sell more, right?
Yeah. Can you say something about, so if you think about the monetization of it in terms of revenue terms
what is different about your agentic flow that's actually driving, accelerating recurring revenue growth today versus the past, things like Cerebras and other AI features, which were maybe in your core offering, but where we didn't see that kind of acceleration revenue?
Yeah, that's a very good question. Of course we always did a lot of good work. But what is new with this agentic AI, and we always wanted to do it, this is going back decades, is we wanted to automate more the running of our tools. Our tools are fairly complex, and typically what happens is they run for a few days. This is not like it doesn't run for 5 minutes, right? If you're doing some blog, it will run for a few days, do all kinds of optimization. But what the customers are doing is they run it one time and design is naturally iterative. So they have an RTL, they would change it, and then they would run it again, and they change it, and they run it again. Okay.
Typically a user would do three or four experiments at a time because that's what typically humans will do. But if an agent is running it, first of all, agentic is much more meaningful to us than GenAI. Because some people said, "Well, GenAI has been around for four years. Why did it not have a big impact on chip design?" Because GenAI helps improve the IO of the tool. You can look up documentation or whatever. Okay, that's useful, but that's not earth-shattering. Okay. What is interesting in agentic AI is that you can define a workflow or a graph, or you do A, you do B, you do C. If you get stuck, you do and this is all relatively new with Claude Code and all about a year or a little more than a year ago.
This kind of workflow combined with our base tools can give a lot more productivity. When the agent runs it runs like 100 experiments. It is not running three or four experiments. This kind of workflow is the new thing. That is why I am so confident that our agentic solutions will have a real impact versus GenAI a few years ago. Cerebras and all were good, but now with agentic and the base, it calls more of the base.
than less of the base. This kind of productivity, this 5x, 10x productivity, or at least several xs, is possible and will help meet the exponential demand of our customers.
Yeah.
The demand for all these—we are engaged with all the top companies with all our agentic solutions, and of course the usage of the base tools is also going up like you see in our results.
Where do you think Cadence is getting most competitive traction today? What are the product areas represent the most remaining market share opportunity for the company over the next few years?
I mean right now we are doing well in almost all of our products, which is great. Normally you always want to see that, but it does not happen that often. Right now I think we are hitting in all cylinders. We are not dependent on one critical area, but right now all of them are firing. EDA, anyway, we have the broadest portfolio for chip design. I do not know how familiar you are. We not only do digital design, we do analog, memory, mixed signal, packaging, PCB. Cadence has always had the most complete portfolio, and then we work closely with TSMC for a long time, with Arm for a long time, and now with Intel and Samsung. Core EDA is as strong as it has ever been. Then we put all the agentic on top of that, right?
I think we are definitely leading in agentic. Then hardware, which is like hardware acceleration, which can run things like 1,000 times faster. We are the only company that designs our own chip actually at TSMC. If you look at our hardware system, these are as complex as the latest GPU or XPU system. These are liquid -cooled, fully optically connected racks. Then we have a 10- or 15-year lead in designing our own. So that is hardware. The demand for hardware is going up because first of all, more people are designing chips, but hardware is used in proportion to the size of the chip. If the size is going to go up by 48x in the next five years, so that is a systematic improvement.
Then IP was the weak point of Cadence historically, and I did not invest as much in IP because it is not as profitable as EDA. Now I think especially with AI and 3D IC, there is more opportunities in IP. If you look at IP, our business is up 30% this year. Was up I think, 30% last year, probably. So last three years it has grown much, much higher than the market, and I feel that IP can still continue to grow well with all this Intel and Samsung and of course TSMC. So I feel all these three major areas, three or four, and system business is growing pretty well. So we are in a good position. The main thing is our customers are growing. If the customers are growing, they want to do more and more innovation. Yeah.
Yeah. I want to get back to IP, but first to just close a loop on hardware for a second. You have talked about demand being supply-constrained, I think. What is structurally driving that demand for hardware? Is it the scale, the designs, which you mentioned, or is it also your customers shifting towards emulation as more of a strategic capability rather than sort of a project-level thing?
Yeah. One thing, I do not know how familiar with this is, give me a few minutes to explain what these hardware systems do. We call it hardware, but it is hardware plus software. People would call it full stack, basically. Basically, what happens is, at this point, you cannot design any complicated chip without these systems. It is not possible. There are multiple reasons for it. What these systems will do is even before, let us say, you are designing a chip for nine months or 12 months, whatever it is, 6 to 12 months typically is the design time. You want to verify the chip in your environment, whether it is a software environment, is it Windows or CUDA or iOS or whatever it is.
We can have a chip behave like a chip, RTL, we can make it behave like a chip even before it comes back from TSMC or any foundry. That is used to not only develop software, but also verify the functionality of the chip. Because if you can boot some OS on top of your chip and run your application correctly, then, of course, you know the chip is correct.
Yeah.
That is only possible with these kind of Palladium systems. Then they become irreplaceable. Otherwise, what will happen is you would do the design, and then you would check, and then you would redo the design and take a few iterations, which is the old way of doing it. Only a few companies are doing it. Most of them have moved to hardware-assisted design process. The second reason they are popular is not only can you verify the chip, you can write your software. Because if you emulate the chip, and these are custom chips that emulate the chip like 1,000 times faster than CPUs. They are still slower than real life, but much, much faster than anything else. So you can develop all your software. All these system companies, hyperscalers, are developing chips. Of course, they have software to develop.
For those two reasons, it became irreplaceable, and then the amount of hardware you buy is proportional to the size of the chip, which is going up. So one, it became irreplaceable. Two, there are more chip designs. Three, their size of the chip going up. So it has been record year for, I do not know, last six years. I do not think that is going to slow down.
Yeah. Very good. IP, let's come back to that one for a second. In terms of your market position there, you've got a very wide product breadth across a bunch of areas, including DDR, SerDes, PCIe, even processor cores to some extent. Maybe talk about the diversity of the IP offerings and what are the specific areas where you feel like you have most competitive advantage?
I think IP, the interesting part is, of course, we focus on lower nodes and HPC IP, which is exactly what is, of course, growing the most. Because we didn't want to do all parts of IP because it's not as profitable, and also we want to do, of course, where the buck is going, and we focus on five or six critical pieces of IP. Some of it we developed, some of them we acquired. This is the SerDes IP, the PCIe, UCIe, which is chip-to-chip, HBM connection to memory, DDR. These are, in terms of design IP, t he critical IPs that a lot of customers want. And then the other key thing that happened is our team is much better than before. In the end, these are standard-based IPs. The customer will buy if the PPA is good.
In the end, it's not just having the IP, just like in anything-
Yeah.
it's how good your IP is. So our team is—we, anyway, I personally believe all the leaders should be highly technical and engineering background. So that's true for all my GMs, and I think we have a—This is one thing that has changed in the last few years. Our EDA teams were always world-class, okay? Hardware teams world-class. Now our IP team is world-class in terms of design capability. And they can also use AI to further accelerate their own. The output of the IPs are very competitive at TSMC and other foundries. And then the third thing that happened is these other foundries also want to get in, so we need to develop IPs for them. Whether it's Samsung, Intel, Rapidus, along with TSMC. I think these three things, our focus is correct in terms of the market segment.
Our team is much better, and the PPA is much better. PPA is power performance area of IPs. The market is naturally growing with newer foundries.
Got it. Okay. I want to move on to your last segment, system design analysis.
Yeah.
My personal interest is I think this is the most interesting segment you have in terms of the evolution.
You've talked about SDA enabling companies like aerospace and defense OEMs to simulate a whole system. How different is a product strategy when you're selling to somebody like Boeing relative to somebody like NVIDIA or AMD? Do they want the same physics models, or do you have to take a fundamentally different approach to R&D for that?
No, it's similar. That's why I did it. I don't know if you know this. I'm the one who started in 2017, and people thought this was preposterous, like why would EDA and SDA be together? Because they were not together. There were multiple reasons for it. I don't know how much time I have to explain the reasons. At the highest level, first of all, the math is very simple. R&D is very similar, and SDA is easier than EDA. Of course, the SDA guys don't like it when I say that. EDA algorithms are much more complex than SDA algorithms. Electromagnetics is much simpler than circuit simulation, but they're in the same direction. They're also mathematical software. All companies want to expand, but you want to expand in your core strength.
Because they ask me, "Okay, Anirudh, you're going to be CEO. You became president, so you're going to be CEO, so what is your strategy?" The strategy is go amplify your core strength. What is our core strength in Cadence, or my background, or all the EDA is numerical analysis, computational software. Like I said in the beginning, this is not like some database software or look up a website. This is mathematically deep, as deep as you can get. Of course, everybody thinks what they do is hard, but you can look at what we do. It's the most difficult CS plus math plus physics. So that could be applied to systems. Then the question is, why do you apply it to systems? Because if you look at the market, so that's our core strength, mathematical software.
If you look at the market, I always thought the market will evolve into these three concentric circles. Again, this is obvious now, but the silicon is in the middle, then system, and then data. A perfect example is a car, right? Or self-driving. You have all the navigation data. Then you have the car, which is mechanical plus electrical, hardware plus software, and silicon that drives the car. This is going to happen in all markets. So if you take those three concentric circles and overlay the strength of ours, which is computational software. Of course, computational software applied to silicon is EDA, chip design, EDA and IP, and that was always our core. Always wanted to make sure that we are number one in EDA because the other mistake people move is they expand into other markets but lose focus on the core market.
Our focus from the beginning is EDA should be number one. That's why OM invested in EDA versus IP, even though IP is interesting now. But in EDA, we have the broadest portfolio. We are clearly the company to work with. Then if you apply computational software to systems, that's SDA, and we want to do things which are synergistic to chip design. So which is like thermal analysis, electromagnetic analysis, things which are 3D IC, which are closer. Then computational software applied to data is, of course, AI.
Yeah.
By the way, the AI is even simpler than SDA. The AI people don't like that either. That's just linear algebra. I took seven courses in linear algebra in undergrad. Don't forget even grad school.
I think I forgot about that in kindergarten now.
Yeah.
Yeah.
The algorithms of AI are even simpler than. But it's good. I mean, it has a lot of application. But it is same kind of computational software applied to chip design, which is the most complex, and of course, growing exponentially. Then systems, and then data.
Great. Just a minute or two left, but I wanted to quickly ask you about physical AI.
Yeah.
How should we think about physical AI being a long-term opportunity for Cadence, and where are companies seeing practical value in those applications today, your capabilities today, and how should we think about physical AI in terms of magnitude of revenue contribution over time for Cadence?
Yeah. I am super excited about physical AI, and have been for some time. We are, of course, excited about the current trends of data center. I mean, those are huge. Physical AI will be a very big application. If you talk about the cake in the beginning, the three-layer cake. Also for five years, talk about three slices of the cake. These are vertical slices. Because in the end, of course, the value will be vertical, not horizontal. The big slice right now is data center and infrastructure, and I think we are very well-positioned. We are working with all the Mag 7, like we discussed. You can see it in our results. The other thing in strategic direction is you want to make sure you don't miss any of the other big things.
One thing is you have to grow in your core strength, number one. So I explained computational software. Numbertwo , you have to grow with the market. Then chip companies are becoming system companies and AI companies, which is obvious now. Look what NVIDIA is doing, or Broadcom and Google and Apple. Number three, you don't want to miss any big trends. We always over-invest ahead of it. Not too much, but always ahead of the big trends. What are the big trends? If the three layers are horizontal, the three vertical slices are data center first. We are very well-positioned. Then I believe physical AI will be huge because these are all trillion-dollar markets.
If AI is good enough to reason and talk and see, imagine what could happen in cars and robots and drones, and these are trillions of dollars of market. The third slice I always believed is sciences AI, which is life sciences and other deep sciences. I think what happens is people confuse that all these three are happening at the same time, and to some extent they are, but they have a peak of each cycle. So I think data center is in peak. I think physical AI may peak in next 3 to 7 years. Then life sciences and all will maybe 5 to 10 years from now because that's another important thing, it's very difficult. So we want to invest in all these three slices. So we of course do life sciences, as you know.
Physical AI, we did acquisition in Hexagon to get the best middle layer for that. It is the best robotic simulator. The opportunity for the physical AI is not just. The AI model will be different. It will be a world model, right? If you go back to the three layers of the cake and put the physical AI slice, the top of the slice is different because it is a world model, not LLM. There is no data for the world model. You have to do a lot more simulation. Therefore, we invested in Hexagon D&E business for simulation. It will also drive a lot of silicon. Our traditional business. The silicon in physical AI will be more mixed signal silicon, for cars and drones, which is anyway Cadence's traditional strength. You can see that in Tesla or Rivian or BYD or Xiaomi.
I mean, I just came back from China. It is amazing what is happening in Xiaomi and BYD and NIO. They are all designing chips. They are all our customers. Same thing with some of these U.S. companies like Tesla. What they are doing is remarkable. Rivian, and some other even traditional companies are. Because the criticism has been, oh, this is a very slow-moving market, but I think this self-driving is completely going to change that.
Yeah.
Drones and all. It does not mean that we do not love data center. Of course, we love data center. We just want to make sure we are ready for physical AI, ready for science AI.
Great. It is a great place to end it. Unfortunately, we are out of time. Anirudh, thanks for being here with us here.
Thank you.