So, CPX is the socketed CPO where you can put anybody's optical engine and socket it and close it down. Like you know, you put a CPU onto a motherboard, right?
Welcome to another Semi-Doped Podcast. I'm Austin Lyons of Chip Strat, and with me is Vik Sekar from Vik's Newsletter. Hey Vik, what's up today, man?
I don't know. How's it going with you? It's like so much stuff happened to Semi. I don't know what to start about. So, I get like this analysis paralysis and be like, I don't know. Is it like Credo acquisition of Dust Photonics? Is it like Broadcom's deal with Meta to make chips? I don't know. There's so much to talk about. But I suggest like we trim the scope of this thing to keep it like tame. And let's just talk about photonics today because there's like so much photonics stuff going on. But before all that, did you see this piece of news about this company called Nuvia Core?
I mean, I saw something on Twitter, but it's like you said, there's so much going on right now. It just went right past me.
So, let me explain. This is kind of close to me in a sense. So, what Nuvia Core is, it's a new CPU company, but it's founded by the same guys who founded Nuvia, which is the CPU company that Qualcomm bought in 2021. And it turned out to be a really good thing for the company because they kind of made it into the Orion series of chips and CPUs that like run on laptops and such. And it was kind of the core of their whole argument with ARM and the lawsuit about whether it was okay to use Nuvia's architecture and the ARM architecture's licensing agreements that Nuvia had. Was it okay to move over to Qualcomm and make their chips? So, I don't know if you remember the whole lawsuit thing.
Yep. Yep. This was the contention of the whole thing. I think like, was it like last year? Founders actually quit Qualcomm. And everybody was like, wait, what's going on? What's everyone doing? And then I think people watched Gerard Williams, who's one of the founders of Nuvia, painting his room. He's like, yeah, I'm just painting rooms now. Whatever. This is relaxing. And everybody's like, why is Gerard... This is Gerard Williams you're talking about, you know. He's like ex-Apple also. And he was part of the M-series processors. And then like he founded Nuvia, sold it to Qualcomm. And wait, why is he painting his house? Like I guess that's what you do when you win the game. Like apparently not because now it's like Nuvia Core is the new CPU company that is going to, I guess, sweep in and, you know, provide CPUs for the agentic AI era. I'm not exactly sure what it is, but we have a new CPU player in town. And I'm not sure they have a product yet, but it's exciting to see the same guys try this one more time because why not? They sold a CPU company before. They can do it again. Interesting. There you go.
Yeah, where my head goes is like, okay, cool. Respected founders who had done this successfully before at Apple and Qualcomm. Makes sense if they're just kind of doing something similar, but probably different. Obviously, CPU demand is super high right now. It will be interesting to follow and ask the questions like who are their target customers? Presumably hyperscalers. Now we've got Intel, AMD, but then we also have ARM and Nvidia. And then to your point, you know, there's the Qualcomms and there's, you know, everyone making their own ARM-based chips. And so another CPU, it does raise the question of like, what's different? Who's going to buy it? What are others not doing that these guys believe that they will do differently? So, this will be fun and interesting to see what happens.
And in the CPUs for agents and agentic AI era, it's really a question that I have gotten many times ever since I wrote my CPU article. From many people, investors and professionals alike, about what kind of CPU is the best one for agentic AI. Is there like x86 and ARM? I know we spoke about it. And like, wait, ARM is coming up to speed in the whole server era. And people have also argued with me about like what kind of software really doesn't work on x86 and or rather ARM and what kind of software works only on x86. And I've been saying like yes, there's some legacy stuff that happens. But I think it has come down to something very simple right now in the CPU space. The supply for CPUs is so short, the demand is so high that even normal cloud providers are not able to get CPUs anymore. The force of AI has taken over all CPU chips now and like nobody gets anything for anything else. So, the best chip for AI CPUs use, you know, is the CPU you can actually get your hands on. It doesn't matter if it's x86. It doesn't matter if it's ARM. It doesn't matter if it's Intel, AMD, Vera, Graviton. I don't know. Like whoever can make these things fast enough, people will write code software to port it over and get it done. It is so much in demand.
Yes. Yes, that is a really good point. It's not about what is the optimal chip, it's about what is the available chip. There's lots of questions that I have to untangle eventually and hopefully we'll hear it from the CPU designers themselves, which is getting deeper into the workloads of what is the GPU orchestration that has to live on or close to the head node. What is the actual agentic workloads that's doing tool calling? And then what about CPU orchestration like agent orchestration? And where should all of these live like with respect to latency of like talking to GPUs quickly? Cuz even if you're doing agentic stuff as like general purpose as it is like, oh yeah, it's just running code. Every step is like ping-ponging with an LLM. I tell it, hey, I want to write this code and it goes to the LLM and then it comes back. And then the CPU might... the LLM's going to write some code. The CPU might execute it for me or something. But there's this constant ping-pong. And so there's this... to me something I'd love to understand better is just like that... that's obviously like there's some latency there. And so I'd love to get deeper into what are the optimal designs for the different workloads? Is it, you know, Nvidia, it's kind of like a one-size-fits-all. It's like we have a CPU. And it happened to be our head node CPU. And now we're going to have racks of them, you know. And so I would love to pull on that thread more. And I know you've written about it. But getting much deeper and I'd love to hear from people who are actually using these agentic CPUs. But I do think there's lots more to be talked about from all the different players. I think like they were pretty early on in the game of agent CPU, agentic AI.
Yeah, for sure. So, we really have to keep a tab on this stuff and see how it evolves because it's a new space for a lot of people. And I was watching the interview that Jensen had on Dwarkesh podcast. And he was explaining things about so many things. A good podcast. You should really watch it if you haven't. And I think the EDA tools are going to get very sophisticated. There are going to be agents running all kinds of EDA workloads. And Cadence and Synopsys are going to sell more licenses than you can ever imagine because agents need to use licenses to do stuff. And he says engineers are going to deploy tons of agents. And the reason this hasn't happened yet is that we haven't really figured out how this tool calling with CPUs really works yet. We're still working on that part. Once we figure out how CPUs will really orchestrate and work with a wide variety of tools, there is so much untapped potential in what agents can do when they start calling tools, really calling tools, that the best is yet to come. So, we have to wait and see how this evolves. And I'm sure we'll keep talking about this. But like we promised, maybe we should like go and talk about photonics. Because CPUs are cool and all, but you know what? Everybody does photonics now. Talk photonics or you're not cool.
That's right. Let's talk photonics, man. Let's see. Should we start with the Credo acquisition?
Yes, we should. I did so many Substack articles on this. And everybody has been writing about this. So, definitely this is something we should at least talk about because it's very important in the frame of what Credo has been typically viewed as versus what they are viewed as now.
Yes. Totally. Impact to the stock, impact to the narrative. And lots here. Let's start at the top. So, Credo acquires Dust Photonics for $750 million cash upfront and almost a million shares and an earnout as well on top of that. So, the total potential is just over a billion dollars. I had never heard of Dust Photonics. What did you learn about them when you looked into them?
Yeah, I had not either. But apparently this is a company that had investors like Actuaries Management and Gavin Baker behind it. Seems like they are founded in 2017. They're an Israeli company that is fabless. Which is interesting because the moment you say they're fabless, you have to ask, so who is the fab? And in all likelihood, it's Tower. So, there's like this common beneficiary of Tower Semiconductor becoming the photonics TSMC. Still that narrative keeps happening. Anyway, they have about 70 employees. And they've been doing photonic chips. They have this whole photonic IC portfolio that does all the way 400 gig, 800 gig, 1.6 T. They have a roadmap up to even 3.2 T. So, they're a very interesting company. But to me it seems like, yeah, they have a couple of core IP. We'll talk about it for sure. But essentially they're a photonic IC company that seemingly has some hyperscaler design wins already. I don't know what it is exactly. But seems like it's a successful company that was basically up for grabs by somebody like Credo. It's a match made in heaven because it brings all the abilities that Credo doesn't have right there and becomes like a perfect synergy between what Credo offers and what Dust Photonics does as photonic ICs. So, it's a perfect match.
Yes, so let's go into that. So, everyone thinks AECs, copper cables. Credo has strong serdes. I think they actually started just licensing serdes IP, but then they moved into the AECs. They kind of pioneered that and made the end-to-end cable. They also have DSPs, optical transceivers, the ZF optics. So, tell us, as Credo started as more of like they're an interconnect company, but they started on the cable side. They've been moving into optics. So, how does Dust Photonics fit into who Credo is becoming and where they're headed?
Credo is one of those copper cable companies that is also vertically integrated within the cable section. It's not like what Marvell does, for example, on this front because Marvell makes these like DSPs and they make the active electrical portion, you know, the things that do the forward error correction and the amplification and the equalization and all that at either end of the cable, but they don't actually own the cable. Credo does the cable, too. So, they're an end-to-end cable provider, which is why they were so popular in all those purple cables we've seen in the past. And they basically are useful for scale out. Like scale out is basically Credo's bread and butter, really. And even their optical interconnect portfolio, the ZF optics, is for scale out. And that is their, you know, their hottest area that which they provide the best solutions for. And then they have this ZF method where they have this thing called Pilot, I think, the software platform. It monitors all the link health, telemetry, and all the metrics that comes from whether the link is on or not. And then it reports back via software screen, you know, an interface where you can see which link is going to go down and like replace it preemptively. This is becoming more and more... it's always useful to contrast this with Marvell, also because Marvell has a similar thing called Reliant, which is a telemetry system where they monitor link health and all that. So, this is where things are. So, they have excellent serdes team. They have been doing serdes design for copper cables, of course, and they have optical DSPs, so they do it on that side as well. Now, the acquisition of Dust Photonics brings in the ability to make photonic ICs in house. That is something they did not have before. So, this fits right into the missing portion of Credo's optical portfolio, which means now they can compete on anything that you want, near package optics, linear pluggable optics, co-packaged optics, whatever that you want. If you can make silicon photonics chips and you have the ability to drive them with serdes circuits and you know how to now monitor telemetry and you understand how all these interconnects fit into the context of a data center, you have yourself a powerfully integrated vertical stack of technology that goes all the way. So, that's why people love this a lot and the stock has been up like I don't know, 50% over the last week. And typically Credo has been synonymous with copper so much so that when the CPO news came out, the stock has been dropping. On this podcast we have maintained that copper is never going to go away and we are always pro copper and pro optics because they both have their uses and needs in the data center. So, to me this is great that it is vertically integrated now, Credo as a company, but I always was never very happy with the whole sentiment of Credo equal to bad because copper going away.
Totally. So, okay, let me pause you here and ask a question. So, why does Credo need to own the silicon photonics chip as opposed to source that from some other vendor? What's the benefit to being fully vertically integrated there?
Well, there are a couple of things I can think about, actually. The first thing is that when you own the serdes, you might as well own the design of the chip itself because you can optimize all of this in one go. And then the second thing I can think of is the supply chain issue. You're not here waiting for somebody to make you photonics chips. Because everybody is trying to service their own needs. It's not entirely prudent to try to keep sourcing chips from other places. So, if you own the entire design flow end-to-end, then you can provide a complete solution to somebody who says, you know, you don't have this. Optics is the way ahead, but, you know, CPO is coming, but Credo is not ready for it. How are you going to address this CPO world? So, if you have your photonics solution in house, then now you know that question is answered.
