Synopsys, Inc. (SNPS) Earnings Call Transcript & Summary

Good afternoon, everyone. My name is Blair Abernethy. I'm one of the software analysts here at Rosenblatt Securities. And I'm thrilled to have Synopsys joining us today. I first need to read a quick safe harbor agreement language, and then we'll get into our Q&A session here. Today's discussion may contain forward-looking statements related to Synopsys' current outlook, expectations and beliefs, which are subject to risks and uncertainties that could cause actual results to differ. Please refer to Synopsys' most recent SEC filings for a discussion of risk factors that may materially affect these statements. So with that, joining us today from Synopsys is Stelios Diamantidis, who is the Executive Director of the Center for Generative AI at Synopsys. Stelios has been with the company for about 7 years, is a fellow electrical engineer as I am. And looking forward to our discussion around AI. Also, Director of Investor Relations, Phil Lee, is with us to provide additional support. So gentlemen, thank you, and welcome.

Thanks for having us, Blair. Appreciate it.

Just another footnote. [Operator Instructions] And I'll be happy to enter that into the conversation. So just to kick it off here, Stelios, I'm sure you're really busy with a title like center for generative AI at a company the size of Synopsys. Tell us a little bit about what you're doing these days at Synopsys.

No, I'd be happy to. So as many of you may know, I'll just very quickly go over what Synopsys does first because I think it'll give us the right context. So we deliver silicon systems design solutions, and those solutions maximize our customers' R&D capability and productivity. Our motto here is it's our technology, but it's your innovation. And so we want to make sure that innovation happens as quickly as possible. So as part of doing this, we've pioneered a lot of solutions, a lot of technologies over the last almost 40 years that Synopsys has been in existence. And we've been part of many, many, many design journeys, as we like to call them, that our customers have taken through those innovations. So the center of generative AI is actually something very natural to us. It is us recognizing that there's a new technology that has huge potential, huge potential for us to deliver new solutions, improve the access to our existing solutions and even benefit internally when it comes to the way that we design software or even the way that we drive our operations. And that is really the scope of this center here that we've been running for the past 18 months or so.

Great. Great. If we look back and maybe start with a little bit before the onslaught of generative AI into the mainstream. Talk about a little bit about what Synopsys has been doing in the last 4 or 5 years. I think your first product, really AI-enhanced product came out about 2 years ago, but just maybe give us a sense of what Synopsys has been delivering to customers utilizing AI.

So absolutely. So when you talk about AI, it's always good to define what is it, right? And so for us to call something AI, it's more of a qualitative definition. It means that you've achieved something, some software capability that has reached sort of human level of complexity when it comes to decision-making or effectiveness. So it really is a sort of soft definition, but it's a hard definition. It's very hard to get to that level of accomplishment with a piece of software. So when it comes to the broader context of machine learning, clearly, Synopsys has been working with machine learning-based solutions for many, many years, mostly models embedded in our tools that can help us sort of help tools become more acquainted with the design environment, learn from interactions and become faster, more convergent really as an under-the-hood capability, but that's not what we call AI. I would say that for us, our AI journey officially probably started in 2017 when we first started having line of sight in new technologies, new technologies that have the potential to allow us to solve problems that simply we haven't been able to solve before in chip design. And these are the problems that generally engineers with decades of experience tend to tackle. And what started as a very ambitious research program even surprised us that it became an actual product in 2020, so 2.5 years or so since we started dreaming. And that product, of course, is design space optimization AI or DSO.ai. It's a product that helps solve significantly difficult problems in the area of digital design, primarily physical design, soft floor planning, placement, clocking and routing of chips, one of the biggest problem spaces in chip design. And so in 2020, we went to market with our first solution, and it was really quite amazing to see this journey of AI-enabled solutions sort of take shape. Not only did we grow the DSO.ai in the digital space to the tune of, I believe now it's north of 400 commercial tape-outs that have been achieved. But we've also been able to take the technology branch and apply it to many other complex problems and chip design like, for example, the implementation and selection of tests for verification; the semiconductor test process, which is a very expensive step in the chip design, overall chip design manufacturing; and even analog-level problems that have been deemed to be intractable on before. And we're still seeing this great applicability of the space optimization technology, which is reinforcement learning-based technology way before ChatGPT came into play.

The -- some of the early results from DSO in terms of PPA improvements are quite impressive like material improvements on customers' designs. What are you seeing in some of the other newer areas? Do you have enough feedback yet from the verification test or analog kind of AI-enhanced products?

Yes. And I think that's probably one of the most exciting aspects of this technology. While the problem space has changed dramatically, even value is captured differently. So as you said, in digital design, value is PPA. Well, in verification, value is time to regress or time to coverage on coverage completion, right? And then in test, it is the size of the -- really the time that you spent on the tester and in analog is the time that it takes you to, for example, migrate an analog circuit from one process to another process. So wildly different definitions of the problem and yet the value, both in terms of the applicability of the underlying technology and the magnitude are very similar. So for example, in verification, we have seen 5 to 10x faster closure in many, many design cases, which changes the way that you think about verification overall, right? Generally, verification is a problem where one -- an engineer generally with a lot of experience tries to figure out exactly what parts of a -- of an overall functionality of a chip to test in the time they have available. So now they have a lot more time available. So they can do a lot of different things, and they can change the way that they design actually fundamental. Same thing for, let's say, for example, test. If you are -- if you have the ability to reduce the patent counts that directly correlate to the amount of time you're going to be spending on a tester, we have seen easily 20%, 30% patent count reductions for our test space optimization technology. And that has led to amazing, of course, changes in the way that people plan products and so on and so forth. I think the impact of the solution, the magnitude of the impact has been very substantial, and I would say at the same level of impressive results as we originally saw with our digital design solution.

And these products are all -- they're all add-on products, if you will, to an existing tool chain that a customer might be using, right? So your -- Synopsys is actually getting additional revenue for these products. You're not charging on a seat basis, though, right?

The products are licensed independently, but they are used in conjunction with the underlying tools. And that is actually a very important aspect. When you look at problems outside of EDA where, for example, you had reinforcement learning-based systems learn how play chess or go or even other -- even within the chip design space, you've seen different kinds of solutions trying to create circuits from first principles. Very exciting technologies, and we really love to see that kind of innovation, but they all run into some very specific problems when it comes to scaling because getting one great result is actually impressive. But being able to consistently get great results for all chips and all verticals and all target processes is actually an extremely difficult thing to do. So our AI doesn't learn how to build circuits, if you'd like, with the transistor and a whole bunch of wire. It learns how to build circuits using Synopsys tools. So it leverages our 40 years of experience solving known problems. We don't recreate the wheel here. And we focus on our energy and our investment in learning how to navigate these tools into the more intractable problem spaces where no known mathematical solutions actually exist. And so they're companion to our product suites, and they grow together. In fact, it's practically impossible to talk about one without the other. And hence, we see them as addressing one space, the digital design space, the verification space and [indiscernible].

Okay. Okay. That makes sense. And how has the take-up been of these new products by your customers? Are they -- do they take a long time to pilot with them? Or are they put into production fairly quickly? What's kind of in your experience in the last couple of years?

Even since 2020 when we first went to market with a brand-new solution, we've seen tremendously quick uptake in some of these products. Now tremendously fast in our world doesn't mean hours or days, still means weeks or months because at the end of the day, you are putting a new capability into production. And you are putting a lot of money behind the chip program for which should be applying this capability. So it does take time. But in my experience, my years in EDA, I haven't seen a technology like DSO, for example, going from 0 to 400 tape out to about 3 years or so. And that means that people put those into test environments, then put them into pilot environments, then we test chips and finally deploy them through production. So to reach 400-plus production tape-outs, it means that you've actually applied this to a lot of circuits. And I'll share one more thought from -- even from our early days when we started talking to customers about DSO.ai, as you can imagine, when you walk into an engineer's den, I will call it, basically a leading semiconductor design company where you're introducing a new capability, and you use the term AI, immediately, people, of course, want to see it working. And they don't want to see it working on some easy examples that some of the more junior design that are taking on. They want to see it working in some of the hard-to-close design, some of the fastest, most energy-efficient designs they have available. And so that for us was really the sort of like measuring stick, if you like, for the solution. And being able to walk in and demonstrate results at that level of capability was quite incredible. So yes, fast, but certainly in our space, even fast takes a bit of time.

Yes, yes. Interesting. Let's maybe shift the conversation over more onto the last 2 years or so with generative AI and maybe explain a little bit about what you guys are looking at. And I think both -- I'll let you define it the way you want, but it's certainly I look at it and I go, this is opportunities to improve productivity at the customer site, change the tools, but also these internal things you can do at Synopsys that can drive productivity and margin improvement using this technology. So maybe lay out for me how you guys are approaching it.

Oh, yes. Absolutely. And you're absolutely right, Blair. That's the way to think about it. There are opportunities, both internally and externally. And maybe for the first time, we can serve them both with very similar underlying technology, right? So generative AI gives us a new tool belt, if you like, a new set of capabilities that we can apply to difficult engineering problems. And so it's very much ideal for some of the challenges that our customers face with designing chips. And I group them in 2 categories very broadly, right? One category is making it easier, faster to become more efficient with the existing tool set. So the existing products, which means that we can now come up with capabilities like our recently announced in Synopsys-built AI copilot, which is a conversational front end to tools, right? So now engineers with maybe a few years of experience can right there within the tool environment, ask questions, identify next steps, compare those with the experiences of literally hundreds or thousands of other designers that have been part of our design journeys over the last 35 years. And that is extremely powerful, as I said, particularly as we are really gasping for talent worldwide when it comes to chip designers. In addition, we can also improve the way that we support our customers by helping them get answer to questions faster and most importantly, within context, so we can answer things. Not generally as in, what would you do for a design that exhibits a behavior like this? But more specifically, what would you do for this design that you're working on in the current context, which is extremely more powerful. On the other hand, now there are new things that we could do with our tools. So on one hand, we had accelerating the way that people interact with our existing products. Now we have opportunity to solve new problems that we were unable to solve before. And we leverage the power of generative AI, which is generally what content generation and the conversational interface. So not only can you find out how to do things, but you can also create agents that actually do things for you. Can you read the spec and, for example, create some interfaces for me so I can review? Now that task that I just described in a few words and makes a lot of sense to everybody is weeks of work of an expert designer, right? And we could just describe it very quickly, succinctly and get going. And then we can come back and revise. So great opportunities for external users, but equally great opportunities internally when you think about the things, again, that generative AI is very good at. So the data is in. There are -- it is clearly a huge productivity boost for software engineering. So a lot of folks outside of EDA, outside of design automation or IP, I've seen great value using things like GitHub Copilot to create software, where Synopsys has a lot of people who create software, software for our tools and also designs for our IP products. Accelerating those teams means that we can benefit internally from a lot of the benefits of generative AI when it comes to our engineers on productivity. But equally important, again, internally, our execution machine, right? And I'm talking about our sales force. I'm talking about our marketing organization, our finance department. Everybody who in their regular course of business create content, perform research and can now upgrade their everyday operations through the part of conversation through generative AI. And so that's a fourth area overall, if you'd like, of applicability of generative AI that we have explored successfully, and we're now in the process of leveraging internally as well.

How long is that road for starting to roll these tools out internally, and you start to see sort of impact from it?

We are already in pilot programs for this technology. So it is -- there's a pipeline of capabilities. There is an engineering road map. There is a deployment road map. We're not quite at the point yet that we can broadly deploy them. But we are at the point where we have exposed them to several hundred users already across all these domains. And we're in the process of doing all our good hygiene, right? Turning great technology and 10 great examples into full on, all the time, great productivity, 0 hallucination in case of journey, then we're well along this journey. And I think we are actually comparing to what we see in the industry in a pretty good spot for leveraging these technologies. [Technical Difficulty]

Sorry about that, everyone. We're just back. I'm sorry, Stelios, your last couple of comments I missed. We had an issue on our end here.

Absolutely. I can summarize again. We are in a good stage of pilots with these technologies internally across multiple areas that I mentioned before. And that would include the software engineering productivity and also our execution machine productivity with several apps. We're seeing very, very good results. We're just going through the maturation process. And I'm really looking forward to exposing those to more of Synopsys.

Fantastic, fantastic. One of the other areas that's really important for Synopsys is on the IP side of the business. Maybe you could just describe a little bit for us what you're seeing there or what's happening vis-a-vis AI technologies in and around impacting your IP business.

Oh, absolutely. Our IP business has really grown tremendously over recent years, I believe. And Phil can keep me honest here. I believe it's now north of 25% of our revenue as a company, if I remember correctly. And by volume, we ship a lot of IP. In fact, I think we're probably the highest by volume provider of IP capabilities out there in the industry across really all verticals. So it's a very exciting space. We also have a lot of great designers that work in IP. And particularly, when it comes to analog IP. Analog IP is really the very hard stuff that makes your PCI Express work when you come down to electrical-level interactions between transmitters and receivers. So we have tremendous brain trust here, and we also have a wealth of technology. So there are many opportunities to amplify that with AI. First and foremost, the concept that we can now, particularly with generative AI, take this technology and amplify the throughput of our own teams. At the end of the day, we're not different from our customers when it comes to EDA productivity. So engineers here can leverage our design space optimization technologies. Designers here can leverage our upcoming generative AI capabilities to accelerate the ramp-up time, make their teams more efficient. So there's a tremendous opportunity for us to become more both effective and efficient in the way that we design IP. Now at the same time, we also -- since we are providers of the building blocks of the AI revolution in our IP business, and I'm talking about memory interfaces. I don't think it's news to anyone that AI likes a lot of memory and also data movers. AI also like data movers like PCI Express, which we announced just this week. And so there's opportunity, I think, for growth in the AI vertical market overall here with our IP products, both in data centers where, again, we're looking at memory and data movers. But also in the edge where you're looking at capability to process deep neural networks and small transformers in situ within smart devices themselves. So really across the board, all the go-to-market axis of IP gets served here, right? We have higher -- more workloads that are specific to environments that require specialized computation and we can supply that. We need more memory, we need more data movers, we can supply that. And then AI software and AI-driven design tools allow us to supercharge the design machine. So we're certainly looking to take advantage of that as well.

And maybe on the customer side for the IP is who is it that's pulling this down from you? Is it the more established players, the larger companies? Or is it the start-ups? Who's drawing down this AI-related IP?

I think I'm going to defer to Phil for that one. If you have more recent information, Phil, please feel free to chime in.

Yes, Blair, it's really all of the above as -- just for the reasons that Stelios mentioned. I think in addition to some of the AI capabilities that we're enabling within the team, there's also several drivers that's driving that business. And overall, we expect our design IP revenue to grow in the mid-teens. A few of those drivers are in addition to the standards accelerating. I think if you look at the standards a few years ago, a new one would come out every 3 or 4 years. That time has really compressed, and a new standard is coming out every 18 and 24 months now. The second thing that our IP business really benefits from is around the idea of supply chain diversification. A lot of our customers are looking to manufacture not only at one foundry but at multiple foundries just to ensure they have a supply of the product that they commit. And the way we sell our IP is our IP titles are unique to each process node and each foundry. So a different title -- you'll need a different title at TSMC than you would at Intel than you would at Samsung. And then the third factor that's really benefiting our IP business is you still have a lot of our traditional semiconductor customers who have in-house interfacing that are designing some of the standard blocks. And in a lot of those cases, is that the best use of their resources in terms of where they allocate their engineering talent or does it make more sense to outsource that to a third party like a Synopsys? And a lot of times it is. So we're benefiting from that trend of outsource -- increased outsourcing as well.

Do you think that the productivity improvements that can come from applying more -- some of the more advanced tools, can that help with your -- with Synopsys' margins in the IP business?

I think definitely, and I think it speaks to the reason that Stelios mentioned, right? So our customers are benefiting from this. And given that our IP team is a team of chip designers are -- really capable chip designers, they can use a similar type of improvement in their design process as they work through the block sets and titles that we build out for the -- for our customers.

Fantastic.

Yes. Absolutely.

I guess if I look at maybe take the conversation up a little bit and just talk about the competitive landscape a little bit. We've been talking about your technology and what you guys have been doing in the last couple of years. How is that positioning you competitively? And how do you see that landscape unfolding over the next couple of years?

No, absolutely. I can speak to that as well. So I was very nervous for the first 1.5 years when we announced DSO.ai, and we watched the months go by and we didn't see any competitive announcements because I was beginning to get worried that maybe this is not going to go down the path that I was envisioning. And I was almost relieved when I saw one of our competitors announced a solution that sounded very similar. And so I think this is a natural course for our business. What's very, very important and then -- and that's always the leader's advantage is to continue to be first to coming up with the technology solutions, right? That's your advantage to know where you're going to go next. And I really am excited that we have a very strong road map as to where we want to go next. And now we have actually built also an incredible track record as to where we've been already, right? If you look in just 3 short years from the first AI application to now what, 5 or 6 AI applications really up and down the technology stack to the first introduction of a generative AI capability in November, less than a year since ChatGPT came out and set sort of a standard for this stuff. The technology announcements and partnerships we have established the collaboration with the likes of Microsoft and NVIDIA that we've talked about already, we're going 1 million miles an hour, and I very much expect competition. We'll also take the steps, and we'll identify those paths. And we'll come up with our competitive offerings and that's natural course of business. But I think we've established some good daylight, and we will do everything we need to do to keep that daylight, make sure that we bring this technology solutions to our customers first. And that would come in all flavors of incrementally improving capabilities and disrupting the market as well.

Interesting. Interesting. And maybe shift a little bit over to the extent that you can talk about Ansys, the acquisition that's pending for you guys. Stelios or Phil, how are you thinking about their capabilities in this area? And how do you work your products together with their products?

Absolutely. So I'm sure a lot of the details will take some time to figure out. However, I can tell you how I see it as an opportunity, right? So we talked about 2 significant businesses that Synopsys runs. Our Design Automation business, we call ocular tools, that help engineers create new circuits; and then our IP business of existing design circuits that can be dropped into any new chip. Now there's a third business, and that is our systems business. And that is really thinking of the hardware/software interaction, thinking about designing entire systems of chips, potentially multiple chips, really thinking about the software bridge. And that's where you'll start getting a lot of the real-world interactions that are very, very interesting. For example, if you're designing electronics for an autonomous vehicle, you are getting a lot of test traffic, no pun intended, that comes actually from real traffic conditions, right? So applying that to the electronics world creates opportunity to create digital twins. And digital twins for electronics are extremely powerful because now our customers can now test the systems and the software they intend to run in the systems to the tune of 1 million or 2 million, 3 million or 5 million [ lens ] code before they commit to an architecture, which is a huge advantage. For me, I get really excited thinking about taking that principle and applying it to every physical interaction out there. And I think with Ansys, we're going to have this expanded portfolio that allows us to think about electronics, mechanical, environmental, like all at the same time and make this applicability of digital twin technology really global, really apply to everything. And then if I think about the things that we've done for electronics with our design space optimization technology, verification optimization, now generative AI, well, they sound like a perfect application for this world of future digital twins that can model the world, right? Because now customers can explore, predict, optimize and even generate content for extremely complex systems, right? So I think of it at that level of capability, and really it's a limitless set of opportunity. I think I also draw inspiration from Jensen Huang's recent sort of view of the space of we're going to go out and we're going to look at the world, and we're going to model it, and those models are going to come from Synopsys. And so we see some really great opportunities there for the company as the ecosystem is taking shape.

Fantastic. And I guess, I mean, the other thing that's been happening partly what we're talking about here is moving from just individual chips into systems of chips, right? And that's introducing whole new levels of complexity. And I guess that's partially what you're trying to address here with these AI tools.

Yes, absolutely. The fundamentals of how we build chips have been changing. The drivers have been there for a while, complexity and the ability to also address supply chain issues. And all of those are coming together with more 3D structures. We're still in the sort of, I would say, early innings when it comes to 3D, and we're seeing a lot of commitment into that space. Early innings as in we're actually doing chiplet-based design, but we haven't really gone into deep 3D engineering. I think that's yet to come. But it's opening up at the same time, whereas the system giving us big problems that we can address computationally, the physics now through these 3D structures allow us to produce ever more complex chips in systems of chips that can help us go out and deliver the actual computation of horsepower in packages that were simply unfathomable just a few years ago. And in an environment that can be very noisy, can get hot or have high temperature gradients where environments where you're operating from unreliable power sources because you're mobile or because you're going through environments, that put stress on your power supplies. We cannot go out and do all this and deliver the systems. And that's creating a tremendous, I think, vector for growth.

And obviously, these physics issues are something you've -- that you've been addressing with Ansys for a number of years with their radar product, right?

Correct, absolutely. But things like electromigration, IR drop, we've been able to have a very successful partnership over the years and show that we can deliver incredible solutions together.

If we just switch back to generative AI for a moment because it's part of what you're looking at and researching at the center, how are you looking at -- are you looking into more design automation using AI that is more generative opportunities actually -- well, we're generating code for programmers now. And they're taking it, and it's giving them a 15% to 20%, 25% lift sometimes in productivity. Are there -- are you looking at similar opportunities in electronic circuit design?

Yes. I think the opportunities are there. Now the technology today, really when it comes to generative AI has been around the use of language. And so we have these machines, these large transformer machines that can essentially help us predict the next thing in a series of things, I'll just put it that way. So there are opportunities for us to be able to, as I said before, help designers become more effective by learning methodology or interpreting outputs or debugging problems, right? That would be Stage 1. Stage 2, you can help them with generating content. Now this content can be RPL for a digital circuit, can be new tests, for example, for a test environment. It could be all kinds of things that you can generate that today are solely generated by end users, by designers. Now does that mean the liner is out of the loop? No. But then again, how many designers do you know that really take pleasure in repeating the same kind of circuit 100x over the next few years? Generally, the -- solving tough problems is what designers get excited about. So if we can help them do that faster and take away some of the repetitive nature of doing things, I think we'll actually help them become more productive and happier with their work. And then from there, once you've delivered sort of the fundamental building blocks, you can start imagining sort of a new plateau of design solutions where you're operating in a more abstract level. And that's really where I think things like agents will come into play where I now know how to design FIFOs or fundamental building blocks. I can have agents and do that for me. And I can tell you in simple instructions instead of being coding assistants. And a simple instruction like just design a FIFO with these characteristics for me. I'm not going to use it for something else, and then the FIFO agent can go out and do that for you or build a model that follows this section of such and such standard. I'm just getting started, things like that. So some of these things are here today, certainly in the lab, the future, I think the sort of the dots, connecting the dots to what's possible, the things that I mentioned, I think, are quite possible. And then we, of course, we take it from there. There are many applications as you start thinking about future potential models that are not bound by language that could completely reverse the license space as well.

Still, it's interesting because as you're seeing in -- on the language side of LLMs, enterprises are looking at tuning or making their own models based upon their own internal enterprise data. Let's say all the HR departments are all -- all of the contracts the company may have with its customers and so forth to train models, is that something that you would look at? And I guess given Synopsys' IP, it seems like you would have maybe a bit of an advantage in this area because you have a lot to train on.

You got it. It's absolutely true. So early on in the generative AI space, everybody got excited about models. And then I think I'm going to plagiarize a quote that came out really very quickly once people started looking at these models, which was there is no moat around models. And so I think that exactly what you described, organizing enterprise data and using model technology that's evolving by the day or the week to get you the best performance, deployability, responsible AI use, governance is the name of the game. But you just can't overlook the data. Really the data defines all these other downstream decisions. It's no different for design data. So you need to have a strong command of your data. And you need to basically put your data in a position that it can work for you and create economic value almost like some real estate around that data. And then have, of course, all the key technologies to very quickly deploy, operationalize that data, turn it into value, right? That goes without saying. But you cannot overlook design data. And you're absolutely right. One of the things that I think Synopsys has done well and it becomes a tremendous advantage for us in the generative AI space is these kinds of data, whether it's from our own IP or even thinking of product as data-generating machines. This is really a tremendous way of operationalizing generative AI and putting it to work.

It's fascinating. And not to nail you down but is this -- how far into the future are we talking about here? Because we've seen some very simplistic things come out pretty quickly leveraging large language models. But are we 5 years away from a tool that can really help the designer actually generate circuitry they're asking for? Or what are your thoughts?

I think that when it comes to helping the designer, we are here -- where they're today, meaning we already are helping the designer do things faster, better, easier. And then if I look over -- as you said, the horizon of 5 years, I think what you're looking at over the next 5 years is a constant drumbeat of capability that is going to be coming in and really building upon the prior capability set to add more value and more power to that concept. Now in 5 years, where will we be? What level of intelligence and automation? Nobody has a crystal ball, but I think that this virtuous cycle between building great data assets, making them easily deployable and then having the designer come in and add their decision-making and experience on top of it, I think we're going to see some really fast ramps of productivity several times over, over the next 5 years. And I'm also sure that while we are at Synopsys are adding a lot of value on top of these existing technologies and we're even researching new technologies, nobody really saw a lot of the future before November 2022 when ChatGPT came about. And I think we'll experience a few more of those moments in the next few years as well. So being flexible and being able to take sharp returns is going to be key as well.

Yes. I totally agree with you. It's been very interesting to see how rapidly it has evolved with LLMs. They were certainly in research papers 2, 3 years prior to that, but they seem to be too big to deal with. And yes, here we are now with the open source models now approaching the capabilities of some of the best private models that they've spent hundreds of millions of dollars on. So I guess it's early -- too early to say, but are you thinking internally that you -- that Synopsys will build your own branded model? Or are you going to just leverage third party? How are you going to approach it?

Well, I think presently, we see tremendous opportunity to leverage existing capabilities. So whether it's hosted, very large models. And what does that mean? That means that you require large computation clusters that can typically be found in sort of cloud environments, all the way to small models. I mean now we're talking about 7 billion or 10 billion parameters being small models, but that's something in by itself. Even small models that can be deployed in a completely different way in environments with much traditional computation, available cheaper hardware and then also being able to access data that's generally difficult to get to because it's core IP or because it will just generate a millisecond ago, right? So you have to really leverage the entire gamut of capabilities. So I think a successful strategy really mandates this. And I see a ton that can be done with what's there today. In fact, we're always doing a lot of it. But going forward, I really very much see a future where Synopsys and other technology leaders, too, will want to explore more customized models that really fit our data signatures a little better than spoken language, which is what the kind of technology is based on. And those will give us even more opportunity, right? So that's what we're saying, I see a future cycle of several ramps of value as some of these things come into play with quite a fair amount of certainty as well.

Yes. Fantastic, fantastic. But listen, we're up against our time here, Stelios and Phil. I really, really appreciate it. You guys are at the leading edge as always or ahead, which is fantastic to see. So looking forward to more great ideas coming out of Synopsys in the next couple of years.

It was a pleasure being here. Thank you so much.

Thanks for having us.

Thank you.