Hi, I'm Shane Brett, a U.S. semiconductor equipment analyst. I have a quick disclosure to write. For important disclosures, please see the Morgan Stanley Research Disclosure website at www.morganstanley.com/researchdisclosures. If you have any questions, please reach out to your Morgan Stanley sales representative. I'm honored to host Greg Smith, President and CEO of Teradyne. Let's get started. I want to start with a few longer-term questions about Teradyne and the broader test industry, as there's just some remarkable developments underway. Just beginning with kind of the industry environment right now, semiconductor test was about $9 billion in 2025, and it surpassed a 1% buy rate relative to semi industry. Could you walk us through the kind of key drivers behind this increase in test intensity to sort of these levels we haven't seen before?
It's a, it's a really good question. The first thing that I'll say is that, like buy rate as an engineer, I kind of object to the notion of buy rate because it doesn't... Like, it's accidentally right, but it doesn't, it doesn't reflect the actual dynamics of the market, you know. People need to buy new capacity to support revenue growth, not to support the same level of revenue that they had before. What you'll see in markets that are strongly growing is that the buy rate relative to revenue is actually going to go way up. When the revenue ends up being stable at that higher level, then the buy rate drops down. It's not that the capital intensity is reducing.
It's that you need, you need new stuff to build that increased capacity. What we're seeing right now is a massive increase in the capital intensity associated with AI compute, and that's in the memory space and in the SoC space. It's even more remarkable, because if you're talking about a buy rate, that's the ratio of revenue, IC revenue on the bottom and tester revenue on the top. If you look, the things that are driving the revenue are actually the high-margin part of the memory business, you know, so HBM, and the high-margin part of the SoC business, which is the AI accelerator stuff. If you look at the test intensity as like a ratio of the device COGS, it's a much larger increase.
So much of the money that's being made in memory is on ASP increase, and so much of the money that's being made in the AI accelerator space is on the high margin in those devices. It's actually like, it's crazy high increase in test intensity.
I guess to kind of sum up the argument, buy rate is probably something we should look at through cycle, and the kind of buy rate relative to COGS would be the homework that we should do.
I think the thing that I would, you know, for model builders, I would say that the thing you wanna do is you wanna look at. It's a derivative market, you know. You wanna look at revenue growth as the primary thing that drives ATE TAM, not total revenue.
Got it. I'm sorry, I'm gonna dig into this buy rate a little bit. It doesn't appear that this sort of test intensity is set to slow down anytime soon. We all sort of remember what happened in 2020 and 2021. Just how is the current cycle different from the kind of SoC boom we saw five years ago? What differentiates this environment from that period?
Yeah. By the way, like I think of this as what happened in 2022 and 2023.
Right.
You know, in 2020 and 2021, there was a really rapid increase in the size of the ATE TAM, and Teradyne revenue growth was really terrific, and it was really all around the mobile boom. What happened during that period of time was, it was a period of rapid improvement in mobile phones, which drove a higher refresh rate. People bought phones more often. We had a global pandemic, disposable income, you know, like people no longer were spending money on going out to restaurants. All of a sudden, they were spending hours every day staring at their phones, they wanted the best, latest phones. There was a increase in the refresh rate and also an increase in the mix of phones towards the premium tier.
All of that combined to really create a surge of demand around the mobile segment. It was very, very concentrated and very driven by the external factors around the COVID pandemic. What we see here is something that is much broader and driven by factors that have much more longevity. You know, essentially, we are already hearing about NVIDIA talking about demand firming out in 2027. The timeline to build up data centers is such that people are, you know, like buying the land, trying to get the gas turbines, all this stuff out multiple years.
The rate of growth of AI revenue for the for the AI players is such that they're leaning very, very hard into this capital expenditure. For sure there are downside risks against this cycle, but it looks entirely different than what happened in 2020 and 2021. It's far more of a, you know, like the difference between a hype cycle and something that is probably sustainable, at least over the midterm.
If I were to be a bit of a pessimist and say 2026, super high semi test intensity, 26, we may see a digestion. Your kind of rebuttal will be there's a much broadening of the customers, AI revenue seems to have a little bit more durability and longevity. These kind of CapEx plans are kind of spread through years.
I feel like there's waves in the AI space. In 2025 and into 2026, most of the growth in memory is really revenue growth driven by ASP increases. You know, like more people are chasing after the same number of parts. We're seeing very good, healthy TAMs, but we haven't seen the impact of significant capacity increases in memory. That's more of a 2027 and beyond thing. We are seeing significant growth in AI accelerator and in networking, but none of that incorporates the complexity and the TAM expansion that we would expect to see from the introduction of co-packaged optics for scale out and scale up. Those are things that are likely to like be the next waves falling on the shore.
Beyond that, we are still in a world where edge AI is not yet a market driver. In terms of delivering a compelling user experience, there's a future, you know, where smartphones are running large language models. In order to do that, the complexity of the silicon in those phones, the amount of memory in those phones, is going to need to be significantly increasing. So that would be like the next wave crashing on the shore. Because of Teradyne's exposure beyond just wafer to package, we go all the way from wafer to data center. The changes in data center architecture are a tailwind for our Product Test business.
The need to build and maintain these data centers and to build the equipment is a tailwind around our Physical AI and Robotics business. We think that there's a fundamental transformation of the global economy that is happening because of AI. The thing that drives that whole transformation is the silicon.
Great. I guess let's jump right into compute. 40% of your SoC revenue, just largely driven by networking and VIP. Can you discuss what you're seeing across those kind of two customer cohorts?
Yeah. In networking, Teradyne has a very good market position in the networking part of the business. The most remarkable thing in the networking business has been how quickly NVIDIA has grown in that space. You know, if you look back a few years, the network part of the market it wasn't very fragmented, but there were a lot of players that mattered. Now, essentially, it's kind of three players that matter, and we have a good position with all of them. As data center architectures are becoming increasingly dense from a network connection perspective, and the network switch technology is increasing, we're seeing increased test intensity and increased unit volume for this business.
We think that that, you know, that showed significant growth from 2024 to 2025, and we think that it's positioned for growth out into the future. That's networking. Now, VIP compute, the thing about VIP compute is there are a lot of hyperscaler ASIC programs, you know, different device types, different hyperscalers that have their own ASIC development programs that are going on. At this point in time, there are only really two that are at scale. If you include, you know, edge AI, automotive applications, it's three. It's possible, you know, it's possible and even probable that other hyperscaler ASIC programs are going to scale, but they have a really high bar.
You know, to try to produce more tokens per watt than what you can do with merchant GPU and networking is like, a very difficult challenge. For particular hyperscalers, if a lot of their workload really comes down to something that can be reduced in silicon, you know, video encoding or something like that, I would expect those things to ramp, but at a much lower complexity than the merchant GPU or the ASIC programs associated with training. Right now, that market is very concentrated. There's kind of two major programs that you wanna be a part of. As of the end of 2025, Teradyne has kinda fought our way to about a 50/50 share of that market, which is much higher than our position in merchant compute.
I wanna go back to the networking portion. As customers are sort of pushing the envelope with increasingly complex silicon photonics or co-packaged optics, and you're sort of enabling that, just how high do you think what are you seeing in terms of test intensity for networking and kind of your demand for that end market?
There's a lot of preparation work going on right now for silicon photonics and co-packaged optics. It is not a meaningful part of the end market at this point in time. You know, it's certainly optical networking is a big thing, and optical connections at the sort of data center level and long haul totally in the optical domain. For scale out, you know, rack to rack and scale up, it is not yet really had an impact on the market. That being said, there are many, many development programs that are going on as we speak, those are driving a lot of activity in the market, Teradyne is a very good position. We believe that this is going to begin to have a volume impact beginning in the second half of this year.
In 2027, it will be meaningful but not a huge deal. The best way to put this into context and by the way, there might be upside from this view, but if you look at our long-term model, we have this long-term model that is around like what are the conditions for a $6 billion revenue Teradyne. In order for Teradyne to be $6 billion of revenue, we expect the ATE TAM to be in the $12 billion-$14 billion range. Inside of that $12 billion-$14 billion TAM, we think that the total computing TAM, you know, so that would be merchant computing, VIP computing, and networking, that's probably about $8 billion of that $13 billion, so big.
As you divide that, the merchant computing part of that is probably more than half of that, and then the rest is divided between VIP compute and networking. In the timeframe that we're talking about for this model, where you're talking about this $12 billion-$14 billion TAM, we would expect that the silicon photonics and co-packaged optics would be a, you know, a third, a quarter to a third of that networking TAM. Hundreds of millions of dollars, but not like a game changer for the whole ATE marketplace.
I guess I want to go back to that $8 billion compute TAM. When we think about it, you said half of it comes from the merchant GPU opportunity where you're starting to make inroads. Half of it comes from networking and VIP. Networking where you have dominant share, VIP where you're splitting it. When I think about the kind of merchant GPU opportunity where you haven't historically had exposure, could you kind of remind us what's the progress on kind of penetrating into that market?
Sure.
What's the kind of next steps in that progress?
Yeah. What we've talked about before and we are on track with what we've said before is that we believe that we will achieve tester qualification in the first half of 2026, and that we will start to see some revenue, some production revenue associated with this in the second half of 2026. It would be like single-digit, low single-digit % of share for merchant GPU in 2026. It's going to take a while to take off from where it is. That program is proceeding on schedule. We are still on track to achieve that. Let me describe a little bit about how this works. When you're trying to qualify a new tester, what you wanna do is you want to make the new tester the only variable.
You pick a stable high-volume device as your vehicle to do that. As you go through it, you can be pretty sure that if there's a difference in the results, that the difference in the results is due to a difference in the tester platforms, not anything related to the device itself. That project is pretty extensive because you have to convert all the libraries, you have to develop all the test algorithms to work across two platforms. Once all of that's done, you've shown that that tester is an equivalent, and you've qualified the tester and the libraries. Now you can go into a mode. The first thing that can happen once you achieve that is you can test that part. That would be the first production usage, is testing the specific part that you've converted.
The second thing that happens is you can go into what I would call a fast follower mode for other devices. You would be focusing on devices that are earlier in their life cycle. You would be trying to develop a test solution for those devices so that the customer would have the choice of loading the incumbent platform or the challenger platform. That's the second phase, is this fast follower phase. We believe that between this initial qualification and the fast follower phase, that that is enough for us to get up towards, you know, like meaningful share, towards 30% share of GPUs. To get higher than that, there are two things that need to happen.
One is that, most customers in this space, including merchant GPU customers, are exploring a DevOps approach called tester agnostic development. What they're trying to do is they're trying to develop the test solution without abstracting the characteristics of the tester that they're going to put it on. We've done this in the VIP compute space. We are applying it in the merchant GPU space. What that means is that the customer can defer the decision about which tester to put it on until later in the process. The other thing that happens is as you do more parts in a fast follower mode, the customer gets more familiarity with your platform, and you have the opportunity to compete to be the sort of the primary platform for a new device.
That's something that's probably 3 years out in the future. You know, it takes a while to get to that point. At the point in time when you have a qualified platform, then it becomes more around the characteristics of the tester, you know. If, for example, we have a better solution for a new technology like PCIe 6, then we might be selected for a part like that as the primary. If we're able to handle higher device power requirements, that would be a way that we could. You know, like, once you're qualified, you can then win those kinds of opportunities. It's like, it's not that I would expect to run the table and own the account, but it gives you more opportunities for share gain once you've achieved qualification.
I guess to kind of sum up what you said there, so I guess the initial step is getting into a high volume manufacturing product.
Yep.
The next step would be getting to a fast follower status and starting to kind of expanding your share from there. If your customer goes towards a tester agnostic platform, then, you know, you're kind of on a level playing field.
Yeah. That like a tester agnostic platform, you could almost think of that as faster follower, you know. It makes the supporting multiple platforms even easier than if you have to do a part-by-part conversion because they would use that tester agnostic development approach even in new part development. They would, you know, like by default, go to their incumbent platform because they know it so well. It reduces the level of effort to support multiple platforms by a lot.
Got it. Got it. Got it. I want to shift over a little bit towards memory. You've broken into compute. DRAM's also an area where you've sort of gone back to that level playing field. Just could you begin by just discussing the changes you're observing in DRAM, particularly as just wafer test evolves from a commoditized step to, I guess, a more performance-oriented one, and just where this DRAM test intensity could get to?
So, um-
Loaded question.
Well, well, no, there's a lot of DRAM wafer sort that is still commoditized, you know. Let's talk about an HBM memory for a minute. To build an HBM memory, you have a stack of DRAM die. Those DRAM die get tested at the wafer level, and that test at the wafer level, before anything gets stacked, that is a relatively low-tech test insertion. You're validating the core of the memory. You're testing that every bit works. You're validating very various timing parameters. That is a tough test step to differentiate on because the test time is largely set by the architecture of the device. Once the DRAM die gets stacked on the base wafer, now you're into the performance test.
That performance test is quite differentiated between tester suppliers that, you know, you have differences in capital lifetime, in signal integrity, yield, throughput, all of that. There's a lot more differentiation potential at that test step. After the wafer gets diced up, each of those singulated HBM stacks, in most cases, that stack is not tested again before it's put onto a substrate with an accelerator. There is an increasing trend towards testing each one of those stacks by itself. Okay? Singulated stack test. Our gains have been primarily in the area of wafer stack test. In that space, we've been able to, you know, better than 50% share against the commercial part of the market.
There is one major memory manufacturer that has their own test solution, neither us nor Advantest is getting that part. For the part that's up for grabs, we're kind of splitting that. Advantest's share is higher in the non-differentiated wafer sort space. Our share is very good in the stack test and also in, like traditional DRAM, DDR5, LPDDR6. We have good share in that final test as well. The way to think about this is, for HBM, as you look at HBM, there are kind of two parameters around what describing the HBM stack. One is how high, and the other is the HBM standard. If you go from 8 high to 12 high, that is a driver for the TAM of wafer sort.
If you go from HBM3 to HBM4 to HBM4E, that is primarily a driver for the post-stack wafer test. 'Cause you have new data rates, you have new test coverage that you need. That's the, you know, that's the part of the market that gets driven by where we are strongest.
Across all three steps of DRAM test, we're sort of increasing intensity. You guys are all kind of, it's like a high tide lifts all boats situation with DRAM test.
For DRAM test, yeah.
Okay.
I mean, there is, you know, it is unlikely that there is going to be something that takes the wind out of those sails. You know, if you look way back in time to, like, 2008, there was a fundamental change in, like, flash memory...
Yeah
... where flash memory controllers could begin to work with perfect flash memories. More and more built-in self-test was being incorporated into devices, and the site counts went way high. We're now at a point for, like, memory wafer sort, it's one touchdown, and getting to higher throughput than one touchdown per wafer is very high. The rate of technological change for the final test is also very high. The kinds of things that depress the TAM are absent, and the kinds of things that increase the TAM, like unit volume and complexity, are going up.
Got it. I wanna move over to a market where you've historically had really high share, which is SSDs and HDDs, just storage. What do you see in that market just now that we're seeing kind of the end markets for that market get stronger, are you starting to see signs of that tester purchases pick up?
from a HDD rotating storage, we are definitely seeing signs of life.
Okay.
You know, that. This is a part of the market that works against pretty long lead times, and they had, you know, even when demand in that market was quite low, we were still shipping at low rates and increasing the amount of capacity in the market, because essentially the players there wanted to keep us in the market so that we, like, when they needed more capacity, that we would be there to do it. We went through that whole period of time. We have now gotten to the point where all of that installed capacity has been filled up, and the market for that is growing significantly from 2025 to 2026, and that's HDD.
In HDD, we also are winning new customers, so we've historically had 2 big customers in HDD, we've now won a 3rd customer in that space, and so we expect to grow associated with that from a share gain perspective.
That'll be in systems test.
That, yeah, this is all part of the Integrated System Test group, and those results, we aggregate those results financially into our Semi Test group.
Yeah.
We don't count that as part of the Semi revenue or the Semi TAM.
Okay.
Okay? Because what we discovered is that as the Integrated System Test group was doing more and more system level test of semiconductor devices, it really made sense for us to aggregate the go-to-market and the technology between the Semi Test group and that group. We combined that, we combined those results, but we do break it out in terms of the SoC and memory revenue and the TAMs that we serve.
Yeah.
On the flash side, there's a disconnect that we haven't figured out yet. We see that from a data center perspective, you know, data center architecture, server architecture, that the amount of flash storage, like the SSDs that are required going from generation to generation is significantly, like, it's like a triple in flash capacity. At the same time, we don't see the investment in capacity from the flash manufacturers.
On the front-end side.
Well, from the manufacturing side, you know.
Yeah
... in terms of the foundry investments that they're making or the test forecast that they have. I don't think that that mismatch is going to last forever. It's either gonna turn into additional upside in HDD or we're gonna see sort of more investment allocation from the memory manufacturers in the flash space, and that's probably gonna drive, TAM increases for us there.
Got it. DRAM, pretty strong. HDD and SSD sort of waiting for demand to pick up.
No, no. DRAM, really strong.
Yeah.
HDD has inflected.
Okay
and will grow. SSD, we're waiting for a shoe to drop.
Okay.
Like, we haven't seen that inflect yet.
Got it. Got it. Got it. I wanna move on to mobile. Historically, really strong end market for you. We're seeing packaging transitions towards more chiplet architecture, but unit demand probably feels a little bit lackluster this year. Just how are you thinking about this mobility end market? What do you kind of outlook for 2026 there?
If you look at the there's sort of three things that affect the size of the mobile market. One is device complexity. The next is unit volume, you know, so handset volume. The third is tester fleet utilization. We think that 2026 is gonna be like one of the best years we've had in a long time in terms of device complexity. That's both transition, node transition and also, for the premier tier, a transition in memory technology. That is a tailwind. The unit volume right now, it looks like that is likely to be a headwind, and that is kind of price elasticity with demand, that if phones cost more because memory costs more, then people are gonna buy fewer phones is the equation that people are working through in their minds right now.
We'll see. We'll see how much margin phone manufacturers are willing to eat in terms of higher memory costs in order to maintain unit volumes. The third thing is tester utilization. We have a huge fleet of testers to test mobile devices, and that gives our customers a lot of ability to optimize the way that tester fleet is used.
They've been able to increase utilization in a pretty significant way over the past few years. We always think that they're just about out of cards to play there, but they continue to surprise us. We, we believe that we're looking at a mobile market that is, you know, as strong or a little bit stronger than it was in 2025, in 2026. There's some upside potential around complexity, and there's some headwind associated with unit volume.
Got it. Got it. I guess kinda wrapping up this talk on semiconductor tests, so you have great things happening in compute, DRAM's great, mobile, we'll see. How are you getting your supply ready to meet this tidal wave of demand?
Well, I think the same thing, the same reason that we're seeing opportunity in the customer space, the main reason we're seeing opportunity in, like, merchant GPU is really around supply chain resilience. That, in essence, dual, competitive dual sourcing is a way to provide significant resilience in your supply chain. We know that from the customers that we serve, we apply that same approach to our procurement. You know, we are primarily outsourced manufacturing. We have enabled our highest volume platform to be produced by multiple CMs, and we are working with both of those CMs to increase their capacity. Behind them, all of the component suppliers, we've established an envelope that is very aggressive against our 2026 demand fixture, because we are on a platform that has legs.
We're gonna be. This platform is gonna last a very long time. We wanna make sure that we have enough material to be able to respond to upside demand. The key difference between us and our competitor is this idea of using competitive dual sourcing in our supply chain.
If your customers come over to you and tell you, "Greg, we need more testers," you will be ready?
Yes.
Okay.
We'll be ready with really competitive lead times.
Great. Great. I guess last question. We've spoken about a really strong demand environment for 2026, but at your latest earnings, you did talk about a first half-weighted demand environment. Can you just kind of talk about that? Just what do you see between the first half and the second half? Is there anything different?
Sure. I mean, one of the things, you know, we talked about a first half weighting. Part of that is limited visibility into the second half of the year. You know, economic conditions, programs that have not yet ramped, share gain that we're not sure of. Part of it is uncertainty. Part of it is also that many of our most important customers are leaning hard into the, into capacity adds in the first half of the year. If that's the case, it's like you have a certain number of bullets in the gun. If you shoot a bunch of bullets in the first half, you don't have that many bullets to shoot in the second half.
That's part of it, is that we don't see the same kind of, like untapped potential for the second half that we do in the first half. That doesn't mean that, like, we think that the market is calming down. It's more that the market doesn't follow pure calendar year cycles anymore. You know, that, Like one way to look at it is that 2025 was back half-weighted, 2026 is front half-weighted. The other way to look at it is we're in the middle of a 4-quarter boom. Right? That started in Q3 of last year, continues through Q2 of this year, then there'll be digestion, then we'll go into the next. I don't know whether we're working in a five-quarter, six-quarter, seven-quarter new world based on what we're seeing right now.
Got it. Well, I think that takes us to time. Thank you very much, Greg. That was amazing.