Every stock or company on this list that we will talk about today has gone up over 100% in just one year.
Hello listeners and welcome to another Semi Doped podcast. I'm Austin Lyons of Chipstrat and with me is Vik Sekar from Vik's Newsletter. So Vik, quick, let's start with a plug for your newsletter because I think you wrote something really interesting this week. He wrote about CPUs for AI and I thought it was super timely because on Nvidia's earnings call this week there was some Q&A talk about the Vera CPU and Jensen talked about how it was great for post-training and had really great, I think he called out the single-threaded performance. And so this reminded me of your article because I think the underappreciated thing, everyone's been talking about the demand for CPUs for AI, but people haven't taken it a level deeper and said, well, what are those AI workloads in which CPUs fit the different types of AI workloads better? So for example, like having many threads and cores for agentic tool use is something that you called out and I thought was really interesting. So like, give me your quick thoughts on that article.
Yeah. So in that article I have like basically a sliding scale of whether any given CPU from any company, whether it's x86 or ARM, can be put into more of a reasoning-heavy workload where its only job is to have fast single-thread performance and make sure the GPUs are fed really quickly and that utilization of the GPU is held high. And on the other end of the sliding scale, I basically have categorized what are called action-heavy workloads where the CPU does a bulk of the lifting. Maybe it goes and scrapes a web page, does some API calls and looks up a database, that kind of stuff. And the GPU is being constantly used, but the critical point is that you need a lot of CPU cores so that you can handle multiple agents and all these agents are going to keep the GPU completely fed. So after that, with that framework, I have like nine different metrics you can rate a CPU on. And so to follow up to that concept, I have like this yellow pages of all like popular CPUs that are ranked on these nine metrics and then sorted into somewhere on this scale of reasoning versus action. Some of them fall in between. Some of them don't do well either way. So it's a very gray area. I'm trying to sort out exactly what agentic AI needs in terms of CPU. So that's the whole article.
Nice. I love it. And if you're a CPU designer, you should read it and be framing your portfolio in terms of AI workloads. Like tell us, I want to hear from companies. Tell us, this is our skew that we think is going to be awesome for agents, you know, like would love to dig into that more. So anyway, everyone go check out that article and I'll give a quick plug for Chipstrat. I wrote, I did a deep dive on Arista this week, earlier in the week, talking about how the Ethernet TAM is expanding not only for scale-out but for scale-up, for example with UALink that Broadcom is championing and how that can benefit Arista. And then I also had a fun look at a startup tackling the economics of lithography. This is a US-based company called Xite. Super interesting business. They're using particle accelerators to generate light and then you would feed that light into ASML EUV scanners. So instead of doing the tin droplet LPP light source is what it's called, you'd actually use a particle accelerator. You can power many scanners with one particle accelerator. And then the business model is super interesting too, which I dig into. You can think of it as like photons as a service, which is really just like a utility. Like what is our electric utility? It's nothing but electrons as a service, right? So this would be a similar idea.
And finally, Pat Gelsinger is the chairman of the board on this company. So always interesting to see industry luminaries helping out startups like that.
All right. So Vik, let's get into the podcast. What should we talk about this week?
So last week we spoke about all these optical technologies because optics is all the rage right now. But in the episode we actually said we'll talk about the technologies and then the companies, but then we kind of ran out of time and one of our keen listeners commented and said, hey, where are the companies you promised? So this is like the follow-up to that podcast episode. So we can dig into some names. Obviously there are a lot of names in the optical space. All of which are very interesting and all of which are doing like well right now. But for the sake of this episode, we're going to like limit ourselves to some of them. And if we left out any, let us know in the comments. We'll talk through them in the future because this is how we know what to create next, right?
Yeah, totally. Always love to hear from our listeners.
Yeah, let's do that. Yeah, let us know. So we will follow up with whatever information we could dig up on all the interesting stuff that's going on. But instead of just talking about companies, I propose that in this one we play a little game. So the way we'll do it is, Austin, I'll tell you what the company does and why it matters, what the moat is and basically why it wins and also what its risks are or what the bear case might be. And then you can tell me if you feel like you will buy the stock or you will not buy the stock, why or why not. And I'll kind of pitch in and say what I think of it and we'll move on to the next one. And be warned that a lot of every stock or company on this list that we will talk about today has gone up over 100% in just one year. Okay? Like the first one we will talk about has gone up 1,000%. So that's the way it is. So we're not going to look at it from the point of like price-to-earnings ratio and all that kind of stuff. It's just like let's see the core thesis, see if it makes sense, you know, and give it a thumbs up or a thumbs down or even a maybe or maybe not right now or something. Doesn't matter. It doesn't have to be binary. And then we'll take it from there. So it kind of keeps it interesting and not like drab as we go through these companies. What do you say?
Sure. Sure. Yeah. Sounds fun. I will say not financial advice. Do your own due diligence. Don't invest in this or any of these companies because we say so. This is just a fun way for us to think through the company's thesis and their moats. Totally.
All right. Where should we start?
Okay. So I think this one I will let you explain to me and we'll see what comes out of it. Okay, first one on the list, AXT, Inc. So this is a $1 billion market cap company. You would call them a micro cap really. It's a pretty small company. They make indium phosphide substrates. So this is the base wafer that every EML laser and silicon photonics component starts on. And AXT, which I think they're headquartered in California, they have 60 to 70% of global share. And I guess why we would be talking about them right now is they would sort of be at the very bottom of the optical supply chain in that everyone downstream depends, or above them if you will, depends on indium phosphide substrates. So if you don't have the substrates, you can't make lasers. Can't make lasers, you can't make transceivers. Can't make transceivers, no AI optical networking. So obviously that would be a key bottleneck even if they aren't huge, they are very important. So I think when you study AXT, the moat there would be, so then you might ask, oh wow, if they have all this market share and they're pretty small, couldn't other people just jump in and also start making indium phosphide wafers? I think that the moat here that they can ride on is that indium phosphide crystal growth, it's not trivial. It's sort of part art and part science, part material science, decades of process know-how. And obviously just because you make something that's hard to make, you still have to take it to market. And I think AXT, and listening to their most recent earnings call, they have a lot of the biggest names that are buying from them. They've got a massively growing backlog. You can tell the demand is huge and obviously they should be tied with the inflection, like the optical inflection. Again, if you start with the AI racks, like the more optical fibers and lasers that a rack has as these data centers get bigger and bigger and the more maybe scale-up CPO that happens, like this all sort of ties to needing more indium phosphide. Let's see if I have to think through risks. I think maybe one interesting almost irony here right now with all of our geopolitics is that this is a very important optical supply chain component and it's an American company but the substrate is made exclusively in Beijing and therefore subject to Chinese export controls. So we're not talking about American export controls into China but actually Chinese export controls back to the United States. So that's obviously, on their earnings call they actually had for the first time some permits were delayed from the Chinese government and it actually impacted their revenue. So that would be the big risk. One big risk for this company is just how much of their company is dependent on geopolitics and maybe things that are a lot harder to directly control. So how about I pause there. How's that for like a high-level overview?
Sounds good. That's a good overview. I think it has all the pieces in there. I just wanted to add like they have a significant market share like 50 to 60% which is why they're dominant. But the competitors would be like Sumitomo. I think they have something like a 25 to 30% share. But like the fundamental problem, like in what you explained here was not what you said but with the company is that their manufacturing is subject to Chinese export controls. It is not in the hands of the United States even though it's a US company. And if you look at the stock price of this company, it was about a dollar or two but which has gone up to like 40 bucks in the last year. So and if you look at the earnings and their numbers like the next quarter guidance is minus 4 cents to plus 2 cents earnings per share which is like almost, I mean they're making the, what's funny about this is that I think people who ended up buying this stock at like a dollar and you know sold it at like 40 dollars are making more money than the company itself. So it's kind of funny. So if you ask me, would I invest in this? I'd say no. Even though the entire AI industry hinges on this, but the fact that it is subject to a foreign government's whims and fancies and that the earnings per share is almost zero and the stock is at like extremely elevated levels growing over 1,000% puts it out of the range of risk for my personal taste. What do you think?
Sure. Totally. Yeah. Yeah. I mean, it feels like that's going to be the common theme with any up-and-coming component supplier in the optics supply chain, which is probably as they're growing, their earnings per share. Maybe they're even not profitable yet because they've got to invest into more manufacturing capacity, for example. And yet people who have like smart investors have already said, oh, let's look up and down the supply chain and invest in these companies. So if they've sort of inflated their price ahead of the growth of the company, the argument is probably like it's not that it's not investable, it's just not now. Like it's always a timing game as far as like is it priced to perfection? Now, as for the not investable part, you know, you make a good point that people have to be comfortable with the risk of the Chinese export controls. So if you're investing that would be a risk you'd have to be comfortable with.
Yeah. Well I mean there are people who are going to invest in it but maybe this is why I don't make any money. I'm very risk-averse. So you know be noted that whatever I say is going to come from a background of being very risk-averse but yeah maybe we go to the next one?
Sure. Sure. Yeah. Let's see, you take, so we talked about substrates. You should take it up a level. Let's go to manufacturing. Why don't you talk to semiconductor?
Okay let's do Tower. So Tower Semiconductor is today the undisputed leader in the silicon photonics foundry. So what Tower actually does is they are like the TSMC to all the chips that the fabless companies design. So Nvidia makes like the GPU and they give it to TSMC. They make the chip and they give it back and then they sell it right. Nvidia sells it. So that is the Tower equivalent for the photonics industry. So companies can make photonic designs which means they make these waveguiding structures and Mach-Zehnder interferometers or ring resonators for modulation. All of this stuff happens on the silicon photonics chip and then it's sent off to Tower Semiconductor who manufactures it and gives it back. The thing is they also make other things. They call themselves the specialty analog foundry. So they don't really focus on digital technologies. They more focus on silicon germanium which is very important for driver circuits for lasers and they also do RF and power chips, power ICs and things like that. They are actually an Israeli company but they have a fab in Newport Beach also. So they have such a strong hold over the silicon photonics wafer fabrication market that you know they, I think seven out of the 11 top datacom transceivers all use Tower and they have a very big market share in what they do. So for the optical world and when it comes to this new generation of 1.6T which is really ramping heavily in the industry, they stand to be the dominant player in this game. I have a quote from the last earnings call from their CEO Russell Ellwanger. He said, Tower's exceptional ability to scale the capacity flawlessly in partnership with our customer has made 1.6T the fastest growing silicon photonics node in the industry to date with Tower being by far the majority supplier of 1.6T silicon photonic integrated circuits. So that's their position on this thing and from their last earnings call right, their revenue has grown significantly in this area of silicon photonics and silicon germanium revenue, it's grown like 75%. And if you look at silicon germanium alone it's grown like 115%. So it is like a doubling every year kind of thing and it is expected to even continue to double through 2026. So that is their stronghold over this thing and they also announced that they are going to 5x their silicon photonics wafer capacity by investing, I don't know, something like $950 million in additional silicon photonics capacity. So think about that like they're 5xing their silicon photonics wafer capacity and this is not just like because they feel like it because they actually said that 70% of all the capacity has already been reserved through 2028 backed by customer prepayments which means they've already paid for it and they're using that and expanding up to 5x the silicon photonics.
Amazing. Yeah, you can get a good rate on the debt you take out to build plants when it's backed by customers.
Yeah, absolutely. It's significantly de-risking. Yeah, totally.
And I believe I heard on the earnings call that they went so far as to say like, hey, we're not even pressuring customers to prepay. They are offering it up because they care so much to kind of get in line and like reserve that capacity.
Yeah, that's the thing. And they're all over the silicon photonics thing because they announced some partnership recently within the last 10 days I'd say with Cintel Photonics to make these DWDM, we spoke about this for long-haul optics which is dense wavelength division multiplexing, but that's going to be like used for CPO apparently. And this is like the slow and wide approach like HBM is to a GPU, like DWDM laser is like a way to transmit optical data where each lane isn't really fast but you got many parallel lanes and you can send a lot of information just like HBM is like each lane isn't that fast but you have so many parallel lanes of electrical connections you can send memory like high bandwidth, same idea. So they're doing it with Cintel Photonics. They also recently, I think two days ago, announced some partnerships with Salience Labs for optical circuit switches, you know, so really this is a foundry that is like really well positioned to hang on to the upcoming optics wave. And because all the wafers, if they're actually available from China, from AXT, you know, this is the company that would actually process them into meaningful circuits, right? So this is the next layer of the cake.
All right, what's your take?
Yep. Yeah, well so it sounds like a very definitive business that it's already a strong business and I think it's fair to say we can see that it's going to grow. It's not speculative. How about that? You know, they're ramping up capacity 5x. It does make you wonder like, oh, are we going to get to an over-supply situation? Especially as we'll talk about, you know, GlobalFoundries is going to ramp up their capabilities as well. And so, you know, when we look at the logic market, it's not quite a good analogy because TSMC is kind of the monopoly there. And so if you look at the logic market, you'd say like, oh, it's just going to go up and to the right forever like GPUs and TSMC has just done awesome because they're the only one that can make it and it's so important. Maybe when you look at this space, you could argue like, okay, well, it's a little bit different because there actually is going to be a competitor and so there could be some price dynamics and especially as they're both ramping up capacity. On the other hand, it takes a long time to turn on capacity. People need all these silicon photonics chips today, yield matters. So as we see with like, you know, in logic as others, Intel Foundry and Samsung are trying to catch up, you know, obviously it takes time to ramp not only to like be able to do the process node but to be able to ramp it to manufacturing scale where you have really good yield. There's also a lot of risk I think for people when they switch foundries or dual-source is not free, right? And so there's all sorts of incentives to think about and generally even your investing timeline matters too here. But it does feel like you could safely say like, wow, it looks like Tower is really well positioned for the next several years if logic is anything. And then of course if you look, maybe a better analogy would be memory manufacturing where there's three competitors and yet the whole pie is growing. All three competitors are doing well. Obviously HBM wafers were in high demand. DRAM wafers high demand. Even storage and so on. And so if you look at those other manufacturing markets as analogies, it's hard to see how Tower doesn't continue to do well here. So I say all that to say it feels like a lock for investing in them.
Okay, good. So I agree with you. But I wanted to mention two things. The first thing is that actually when it comes to a wafer fab like this, having a lock-in in a process technology is actually a big deal. You just, it's very difficult to, you know, swap out because if you have a legacy of products you just iterate on them and you figure out all the yield issues, all of that stuff, you know, you have a good relationship on both the design side and on the wafer pricing when you can promise large quantities of wafers. And so that tends to become a lock-in and a moat actually, if you can be the dominant player and people, it is hard to come in as an alternative, right? The same thing that happens to TSMC's awesomeness in digital circuits. So it is no competitor really although there are like, you know, a second place is like a far away second. And the other thing is that in the world of HBM when you have three companies then they are like trying to all get a growing piece of the growing pie. You know, it is actually possible because there are things like a JEDEC spec that determines what the connectivity speed should be. And so as long as any supplier can meet those electrical specs, and then you can get enough supply and pricing, companies are willing to swap out or de-risk by having multiple suppliers and putting them against price constraints against each other. It is possible to do that but it's a lot harder when it comes to foundry technology because like I said moving is a difficult process and so sometimes companies do have parallel fabs running but it's a lot of internal work to do that. So that is actually a benefit for Tower.
I think you're calling out the difference too like it's not fair, it's not apples to apples to compare to memory because those memory manufacturers are vertically integrated, they are IDMs, they're making the end product and then selling and competing on that product, the actual memory, which is a bit commoditized in the sense that there is a spec and they can all show that they meet the spec. And so, yeah, it would be more analogous to look toward TSMC where it's just pure foundry, not also making the product.
Yes. Yes. All right. Now, that's, I agree with you. I'm very bullish. Well, let's talk about GlobalFoundries. I'll let you take this one.
Okay. GlobalFoundries. So who is competing with Tower in the fabrication of silicon photonics chips? GlobalFoundries is trying to play there. I think most people probably have heard of GlobalFoundries, has a long and interesting history. You can read about, I actually touched on some of it in a SkyWater post that I wrote. GlobalFoundries is a sort of foundation node foundry, meaning they're not making the leading edge 2 nanometer chips. They are the third largest pure-play foundry and they have more business lines than just photonics. They do traditional analog, RF, CMOS. But they do have GF Photonics spelled with an F, F-O-T-O-N-I-X for listeners, GF Photonics. And that is the only 300 millimeter monolithic SiPh plus RF CMOS platform. So you can have photonics and electronics on a single chip. And so why do they matter? They are the alternative to Tower for silicon photonics trying to catch up. They do have lots of big brand partners, Broadcom, Cisco, Nvidia, Marvell. And let's see what other notes do I have. Oh, interestingly the government is sort of putting a claim on them and championing them there. They got a one and a half billion dollar CHIPS Act money for domestic silicon photonics. They also acquired AMF, Advanced Micro Foundry in Singapore, which was the world's first specialty SiPh foundry. It was a 200 millimeter foundry with plans to scale to 300 millimeters. So that was a bit of a land grab.
I think you mean millimeters.
