All right. Is the mic on? Okay, great. All right. Welcome to the Credo Fireside Chat. My name is Mark Lipacis. I'm the senior semiconductor analyst at Evercore ISI. We're very excited to have Credo join us. This is one of the fastest-growing semiconductor companies with exposure to AI connectivity. Today we have the CEO, Bill Brennan, and the CFO, Dan Fleming. We'll just go through some Q&A, and if we have some time at the end, we'll open it up to the floor for questions. Bill and Dan, welcome. Thanks for joining and welcome.
Thank you.
Thank you.
You just had an earnings report. I think that's 20, 36 hours old or something like this. Why don't you share some thoughts? What are the messages that investors should take away from that earnings report?
Sure. I do want to note, though, that we did it on a Monday so we could be here with you.
Okay, great. Thank you. All right. Appreciate that.
We missed it every year, the way things line up. Thanks for having us. We had our fourth quarter and our full fiscal year call on Monday, and couldn't be happier about the accomplishment of the team really over the last year, even the last several years. Just to put things in perspective, in FY 2024, we finished at about $195 million in revenue. We more than doubled in FY 2025 to about $437 million-ish. This year we just announced greater than $1.3 billion. More than a double and then more than a triple. We've gone through a tremendous transformation as a company, and really driven by AI connectivity. One notable comment that I made on the call was that our revenue for the quarter exiting the year was $437 million. That it was actually slightly higher than the entire fiscal year of 2025.
We're on this incredible growth curve. Right now, mainly driven by the success of the AEC product family. Looking back, we've been working on that and on pioneering that product category really for the last six or seven years. It's really taken off. We view it as a multibillion-dollar, really long-term growth engine for the company. We expect continued growth this year, continued growth in FY 2028. It's a very solid part of the portfolio. I think as we look at Credo going forward, I think we've basically filled in all of the pieces of the puzzle that makes us a full-spectrum connectivity company. We're on the verge of launching what we're calling ZF Optics, which is a transceiver that we've developed. It's a new connectivity category that we're pioneering, just like we did with AECs.
We're developing micro-emitter technology, microLED technology as a different light source, and we can talk about that. We're also going, not just across the data center in length, but also inward towards the die, and we can talk about the die-to-die connectivity that we're pioneering. We couldn't feel better about the accomplishments of the year, and we couldn't feel better about how we look going forward.
Okay, great. All right. I think it's helpful just to level set and to understand how you think about the company and the vision for the company. When we talk about your AEC products, these are systems. Normally when you think about a company like you think of, oh, you're going to sell a chip, but you're really selling a system. Can you just talk about that kind of realization? What led you to the realization that that's the approach that you had to take? Because that was a non-consensus view at the time that you did that.
Sure. Sure. One of the things that I'm proud of around our team is that we don't draw bright lines about the value that we're going to bring to the market. The idea of a semiconductor company building a cable, when you say it that way, it sounds like a pretty bad idea, right? Even when I told my mother, she was concerned. It's not intuitive, I would say to most investors, right?
Right. My board looked at me, and they thought I was completely nuts. What we discovered in pursuing the concept of extending the life of copper by making it an active solution, what we discovered, it's a really tough problem. It's not so easy as selling a retimer to a copper cable company. It's a much more difficult process that you have to go through. Huge amount of innovation that has to happen and did happen. We've built this incredible engine over the last several years where we can quickly bring products to market. Now, you're talking about a vertically integrated stack from the SerDes, which is ultimately the most critical building block from a technology standpoint, to semiconductor design. Component-level expertise. Talking about system-level design, the actual system, the AEC system. Talking about firmware. We're talking about software that sits in the stack.
Qualifying products. When you think about how does that happen? The first time we put a product into a thermal chamber at Quanta, it failed within an hour, and it was a bulletproof design, we thought. Here we sit today with more than 20 thermal chambers in our facility in Taiwan. I call it our little shop of horrors because the goal is to break the link. We get customers' switches. We get customers' NICs. We run traffic at high speed, and the goal is to find weakness in the link, whether it sits on the switch side or the NIC side. The goal of the team is to harden the solution.
We improve bit error rates by two to three orders of magnitude, meaning that when we go into qualification, we don't fail. When we talk about qualification acceleration as being part of the stack, and we talk about owning the supply chain. I've got a silicon operations team, of course, but I've also got a systems solutions operations team. We're the ones driving the relationships with the supply chain. We're the ones making commitments to our supply chain partners on volume. We're the ones investing. That's why we've been able to really deliver flawlessly as we've ramped, incredibly. I couldn't be prouder of both the silicon operations and the systems operations team. Turns out, the first and the second, and it's typical now that we were targeting an IEEE spec, let's say 400 gig or 800 gig.
When customers ask you, "Hey, can you add telemetry? Can you add special features?" You own the entire vertical stack, you can be open to that. At first I was thinking standard solutions. That quickly changed when I realized we could actually innovate where there has been no innovation previously. That has taken us to the point where I think we're a trusted partner with five of six hyperscalers and a growing list of Neoclouds. We look at from an application space, of course, there's different networks in the data center where AECs make a lot of sense. Our first success was in the front-end network, connecting general compute to ToR. As AI emerged, the back-end network is now the large majority of our business, the scale-out network. There's also disaggregated switch racks.
We see our products fitting all over the data center. It is the best solution, the preferred solution for connections that are seven meters or less. The reason is because it's got bulletproof reliability at a core technology level. It's got a lower power profile. It's got a better cost profile as well.
Can you take that example just one layer deeper when you talk about you're in a thermal chamber and there's an error. I appreciate SerDes as an incredibly important part of the solution. What else? What knobs are you turning in order to take that part that failed, and make it a part that is bulletproof for your customers? Another way to ask the question, aside from the SerDes, what else is a part of the moat for you guys?
Having the ability to have our SerDes engineers right next to our silicon engineers, right next to our system-level engineers, right next to our firmware team. Diagnosing, capturing a failure and diagnosing it is the first step. Then it could be any number of ways that you can overcome that one weakness that you found. It might be, as you cycle temperature from very low to very high quickly, as you cycle voltage from the low part of the spec to max, as you power cycle. You might find a different condition, and maybe they're all corners, but they all matter when you're in the data center, and failures can shut down an entire cluster.
There's not any one thing I can point to, but it's the combination of all of these people working together to diagnose and then figure out how to make the link more robust so it doesn't fail under that specific condition. As we qualify product, we can go through this cycle iteratively for a couple of months.
until we can't make the cable fail anymore. I'm not aware of any other team in the industry that is going to that length. You've got to do it.
If you're not one company doing it, I'm not sure how we would do it.
If we were trying to take a group of companies and act as one.
Did you convince your mother that this was a good idea?
Yeah. She texts me when we hit an all-time high.
Yeah. Great. In AEC, this is $1.3 billion in revenues in fiscal 2026 just ended. That is up 20x over five years. I would agree it's an impressive feat. How do we think about the duration and magnitude from your peak AEC revenues in your estimation? Where does this business unit go from here? What were the key drivers of the expansion?
Well, it's hard for me to call a peak because I don't see a peak.
I see the market that we're growing into is really, it's all about larger numbers of GPUs, larger clusters. It's all about faster speed. There's emerging applications. We talk about training obviously drove the first wave. All of us think inference is going to be much bigger. Now we're talking about agentic. All of these applications drive more connections and drive high-speed connections. From our perspective, we see this. When I think about the market generally, I think about maybe let's call it the pluggable market. When we talk about pluggable optics in a minute, I think everybody is well aware that that market is just growing by leaps and bounds from a volume standpoint. Short connections, AECs are pluggables. They're short connections, a meter to seven meters. They coexist in the same market.
The same drivers that will drive growth in optical transceivers drive the growth in AECs. It's very well known that I think every CEO that talks about copper and optical, they're going to coexist. We're going to live in a heterogeneous world.
We're talking about Jensen, we're talking about Hock, we're talking about Pat Gelsinger. Everybody's in agreement that this whole conversation about optical and copper is one of coexistence. If you believe the pluggable market is going to grow, we're seeing clearly that it's growing rapidly. AEC is a part of that growth.
We can talk about NPO and CPO.
In a moment as well.
Yeah.
Generally speaking, we don't see a peak.
When you think about your four largest AEC customers, to what extent is the growth you're seeing new capacity installations versus further penetration into those existing footprints?
It's a combination. We're not fully penetrated at any one of our large customers today. I will say at the same time, we're adding new customers. Our potential to add significant customers is growing. I mentioned on the call that we're very deeply engaged with five of six hyperscalers. Also there's an emerging group of Neoclouds that we'll do very well with over time. Especially for the Neoclouds, reliability is going to drive their cluster performance, it's going to drive their compute yield. If you talk to any one of those CEOs, all of them that I've talked to say they got to put the best possible cluster out there for their customers.
Highest performance, highest compute yield, most reliable network. This is a very easy discussion where we could say we can help you from a time to stability. You buy $ billions of gear, and if it takes you a week to get them up to the point where they're revenue generating, or if it takes you two months, clearly AECs are preferred. ZF Optics, by the way, offer the same thing. Generally speaking, one of the takeaways from today that I hope we can establish is that two or three years ago, reliability, network reliability became our North Star. I think about us as a full spectrum connectivity company with a huge emphasis on delivering reliable connections.
How many $100 million scale AEC customers would you see on the horizon? How big can that number be?
Greater than six.
Okay.
That's for sure.
The NeoClouds are deploying a large number of GPUs. They're building very large clusters. I expect definitely more than six.
You created this category. It's very hard to sustain 100% share, right?
Sure.
How do you suggest investors think about what is a sustainable number? I know you don't want to give anything up, but what's reasonable?
Sure. We never had the goal of having 100% market share. That's highly unrealistic. It's a large market. Clearly large markets attract lots of competition. When I think about how Credo competes, I think about how we engage with our customers. If we do a very good job with our customers, and I'm talking about understanding their challenges, understanding the things they are trying to deliver to their customers, and understanding the things that are problematic for them, we can have a closer relationship and ultimately design a product that meets their needs. The way we're organized, our goal is to deliver it first. I want to be first to deliver the next generation. I want to be first to be qualified in their test clusters, in their full qualification. I want to be first to ramp.
Once you become first to ramp, it becomes a question about what drives a need for a second source. Is it risk reduction? Because if you're in high volume and you've got a highly predictable supply chain that never misses, you might be adding a little more risk into the scenario if you're trying to, say, take a part of the cluster, take a part of the volume, and rely on somebody new. Generally speaking, my team, the way we compete is to try to do a better job, to try to deliver the right solutions first, get qualified first, ramp first, and ultimately deliver in a way that's flawless. That's the way you maintain market share in a competitive environment.
I'm not worried about somebody offering a lower cost or something. These are not the dynamics.
My customers are looking for trusted partners that can deliver on bringing clusters up quickly and keeping them at maximum uptime.
How do we think about, as you go from 200 gig to 400- 800 solutions, how do we think about the ASPs?
The ASPs are going to vary based on a lot of things. The length of connection, the number of connectors. We've got solutions with eight connectors.
Eight modules. Five, four, three, many a number. We've got lengths up to seven meters, from one meter to seven meter. Speed helps. Any time you've got an increased lane rate, along with that comes cost increase as well. Building a three nanometer chip for a 1.6T cable is more expensive than building a 12 nanometer chip for an 800 gig cable.
You have to charge more.
Sure. We're going to charge based on a combination of things. To get to the point I think you're making is that there will be a content increase as we go from 800 to 1.6T.
Mm-hmm. Is that like the speed doubles, so the ASP doubles? Is that a good framework or is that not the right framework?
That's probably a little too aggressive.
I would say that as this 1.6T market takes off, I can look at two different ways to deliver the 1.6T bandwidth. When you think about a new GPU that requires that bandwidth, don't think in terms of that you need one transceiver or one AEC that does eight lanes of 200. Today in production, many of the sled designs have two physical ports. You can have, say for an 800 gig port, two physical ports. You could have four lanes of 100 and four lanes of 100, or you could have eight lanes of 50 and eight lanes of 50 in each one of the ports. A lot of the sled designs that we're working on with partners, they have two physical ports. You can either do eight lanes of 100 in each port, or you can do four lanes of 200 in each port.
In either case, we're going to see a content increase, and so either a doubling of the 800 gig. Why would somebody do that? There's a tremendous amount of work that has to go into maturing all of the links in the ecosystem. Some customers will say, "Let me take the GPU first. I'll de-risk the connections, and then I'll move to the 1.6T when it makes sense from a maturity perspective and a cost perspective and power as well.
Fair enough. All right. Let's shift to optics. You recently updated your FY 2027 optical revenue target from $500 million to $600 million. You highlighted optical DSPs, silicon photonics PICs, ZeroFlap optics, all could represent $100 million. That's three segments at 100. That's about half of that $600 million bogey. How should we think about the rest of this? What should investors know about what you're doing with ZF Optics?
Sure. All three of those markets that we talked about represent multi-billion dollar opportunities long-term.
If we look out to 2030, I think the silicon photonics PIC market by itself is forecasted to be a $5 billion market. DSP, we all know that's a multi-billion dollar market as well. ZF Optics could be the largest of the three markets if we look at the customer traction that we're getting and we expect over the next several years. I mentioned that they're all growing. All three segments are growing faster than what Dan guided for the year, which was greater than 80%. He didn't guide 80%, he guided greater than 80%.
I can tell you that all three of those segments that we talked about for optical are going to grow more than 80% and even I think more than the company's growth. I think maybe to get a little bit close to the answer that you're looking for, ZF Optics has the ability to ramp very quickly and become the largest segment for optical portfolio.
I think you mentioned potentially doubling or tripling of that business in fiscal 2028. How do you get there? Is this like where are you in the qualification process and design win process?
Right. Go back to what I said about AECs. Go back to about the capability that we've already built as a company.
to bring system-level solutions to market. Our ZF Optics products have already gone through our little shop of horrors in Taiwan. We've already hardened the solutions. We've got interoperability with many switches in the market, many NICs in the market as well. We're well along the way to being able to ramp that. I mentioned on the call 90 days ago that we were engaged deeply with two hyperscalers and two Neoclouds. What I'll say today is that that has broadened in the last 90 days. I feel very good about the customer discussions that we're having. Understand what we're delivering here is not a commodity solution. What we're delivering with ZF Optics, it follows directly in the path that we went down with AECs.
We had a customer come to us and say, "Look, link flaps are killing us in a sense that it's taking us two months to get our clusters up to the point where they can be revenue producing." The term they used was time to stability. "We're struggling. We can't rearchitect. We can't go seven meters or less like xAI did. Long term, we're losing a significant percentage of uptime because link flap, this intermittent connection and disconnection, this can actually shut down an entire cluster." We looked at this problem and said, "How would we attack it?" We had done a huge amount of telemetry work with our AEC family. First product we ramped, we were able to detect when a ToR port with a signal integrity was declining, and before it failed, switch the data to a second ToR.
This is a pretty smart cable. The challenge there was to say, how do we do the same across a large number of GPUs? The connection we're really targeting is the NIC to the ToR. There's no redundancy. You lose those links in a link flap manner, you could shut the whole cluster down. The idea was, okay, during bring-up, could we design a system that was smart enough to be able to determine is there ESD damage, slight ESD damage? Okay. When you fire up a rack, you have to be very careful as you're installing them. It's very easy to damage, but if you don't damage it enough and it actually works for a while, you've got a latent failure that's sitting in your network. We designed a system that we can self-diagnose ESD damage.
We can also diagnose if there's dust in the signal path, whether it's on the fiber plant or on the transceiver side. Multipath interference is killer. As they're turning on their clusters, this enables them to get to a point of stability within a week. As far as in production or in long-term running the cluster, being able to sense in a real-time basis continuously when any given link is crossing a threshold that's set typically by the customer on link health. A link that's declining, being able to say, "Let's proactively take that link down so it doesn't crash the cluster. We'll take that GPU out, put it on a list for maintenance." We can capture exactly what the failure mechanism was that caused the link health to decline. We're delivering a product that's never been designed before.
Right? We're delivering a solution that's got so many features that go beyond commodity, that it's a very easy discussion for people to become interested. These are things that the entire industry are fighting. When it comes to manufacturing these solutions, yes, we're through the qualification process. We've been working on the supply chain for more than a year now, knowing that we're going to be able to generate a huge amount of demand, if we're not able to deliver, those two pieces come together. They have to come together for you to be able to build a business, but much more so for ZF Optics than it was for AEC. There's so many different components that we've got to go lock in supply on. We've got three partners that we've got to lock in volumes on for assembly.
We're well down that path, and we'll be at a point this year, exiting the year, where we're talking about monthly volumes measured in hundreds of thousands of units, and then doubling and tripling that the following year. Even doing it the following year as well. I feel very confident. This is not the first rodeo for my systems operations team.
I feel very confident about delivering the volumes to the market and delivering on the promise of more reliability.
You also mentioned near package optics. What are the most unique elements of Dust Photonics and near package optics PIC?
Reduction in lasers. That applies to pluggables, that applies to NPO, that applies to CPO. They've got a design that we think is best in class in the industry, but the key takeaway is 75% reduction in lasers required. That's an obvious one when the whole world is talking about how are you going to get the lasers if you need fewer lasers? Not only does it give you an easier path to supplying, but it gives you power advantage, it gives you a cost advantage, it gives you a reliability advantage as well.
Mm-hmm. What kind of visibility do you have for that?
For the ZF Optics?
Yes.
Visibility in what sense? On the supply side or on the demand side?
On the demand side.
Demand side. I'm more worried about selling out than I am about not being able to generate the demand. It's going very well.
I want to shift to microLEDs, unless there's something on optical that we didn't cover that you want to make sure we hit.
There's lots of things we can talk about.
Right.
Let me go and-
We got 11 minutes and 32 seconds.
Let's talk about micro. I think it's an important conversation. Hope we can get to OmniConnect as well.
Yes. On microLEDs, the MicroLumen acquisition, what do investors need to know about the microLED market opportunity?
Right. Following the North Star of reliability. MicroLED, different light source than lasers. At a core technology level, think wide and slow.
We're talking about equivalent reliability to AECs. Think about the length being extended from seven meters to 30, and think about the form factor being much more friendly, a fraction of the diameter of an AEC at any given length. We think that that becomes an important part of our portfolio because it offers customers the same kind of dynamics with AECs, so bulletproof reliability, lower power, and potentially a better cost structure and a friendlier form factor and a longer connection. We think that market is quite large.
I mentioned earlier that we think at a minimum, it's double the AEC market.
Okay. That's a pretty big statement, I would say. We're looking forward to seeing that play out. All right, let's shift to OmniConnect. This is an ultra-efficient, 112G, very short reach SerDes. What are the starting use cases? How do we think about this?
Think about the OmniConnect concept as being able to make late binding decisions on any kind of IO that you're building. Like a composable XPU. First application is memory, but next application could be any kind of connections, either scale out or scale up, or even NPO for that matter. We will build gearboxes that will interface with our customers' composable XPU. Our first customer is a company that is, I think, getting great recognition in the market because of things that they've announced. They're building the highest performance inference engine that we're aware of, and it's really memory. Where other inference designs are limited to, say, 128 gigabytes of LPDDR, they announced a system of 2 TB .
We're talking about frontier models that can all be in DRAM, super high performance, for given inference applications like real-time AI-generated video, which is going to be a huge market. The way that we do that, there's two physical limitations. If you look at other designs in the die photos and the system photos that are out there. If you're looking at an LPDDR SerDes or PHY, and you're trying to connect just 128 gigabytes, the photos you can find online, 75% of the beachfront is occupied by the PHYs to connect just 128 gig. The beachfront density, it is what it is. It's not good compared to what we're doing. Orders of magnitude better with the approach we're taking. From a beachfront throughput or just the terabyte per second that we can move in and off the XPU, orders of magnitude better.
The other thing is with those SerDes or PHY, you can only get about 25 millimeters or one inch away from the die. That's the reach. Our solution, we have 10 inches. You look at these photos and 128 surrounds. There's no more room. Okay. When you can get far away from that XPU die, then you can start talking about, we could design with a partner up to, we think 6 TB. 2 TB is first up. We've completely blown the lid off of both physical limitations, beachfront as well as distance, the reach from the main die. That in of itself is a big opportunity, and also when we think about even other designs, like designs that would be maybe more for training. The way that you do that is you've got to source CoWoS, you've got to buy HBM.
We're providing a path around that. You can do it with LPDDR now, both from a performance standpoint and a memory size standpoint. We avoid the expense of HBM and CoWoS. We're working with several customers now globally on solving these challenges, whether they're for inference or training.
I think, if I got it correct, initially you thought of the use case, you're thinking about $2,000 - $3,000 per accelerator. Is that how we should be thinking about this?
Sure. Yeah. You can run the numbers. Now I know that Positron's got much bigger targets than 100,000 units.
Yeah, right.
If you think about 100,000 unit increments, these are numbers that really move the dial.
Okay. Got you. All right. Dan, I don't want to let Dan off the hook here. Dan, what are your updated views on the gross margin evolution as you ramp these new product categories, and you mix over time? You have a target, I think, of 63%-65%. How do all these new products impact this model?
Yeah, one thing we've said is that all of these different products that Bill has talked about, we expect to fit within a very similar gross margin profile. Now, one important update that we gave this week was, if I go back five years as we were preparing to go public, we came up with a kind of a long-term model, which was very instructive. Gross margin, 63%-65%, operating margin, 30%-35%, net margin, 28%-33%. In FY 2026, this was a story of scale. We were certainly subscale five years ago as we were preparing to go public, FY 2026 was a really transformative year for us, where we exceeded all of those long-term targets. In fact, we haven't talked about a long-term net or operating margin for quite some time. 28%-33% net margin.
We were in excess of 50% in our Q4. We gave guidance that we would be around 50% for fiscal 2027. Similarly, with gross margin, we said that for fiscal 2027, we expect to be broadly consistent with fiscal 2026, which was 68%. Not really talking about a long-term model anymore. We'll give updated guidance as we see things play out on the horizon. The most important takeaway is all of these different products that Bill talked about, we expect to be of a very similar gross margin profile.
What about the OpEx side? Bill just talked about a million products that you guys are. How should we model?
Yeah
R&D expense?
That's another thing where our long-term model, we're below that in terms of percentage of revenue, for sure. The thing that really stands out or to bear in mind, some of these things sound very different, but our core technology, our SerDes technology is what really drives our innovation.
That is used across the spectrum of all of our product portfolio. It's a highly leveraged R&D model that we have. Sure, we have a very strong and big team focusing on AECs and all of these other areas now, but fundamentally it's a very leveraged operating model that we have, and you've seen that leverage play out over the course of the year. The last point I'll mention, while we mentioned that 80% top-line growth or greater from fiscal 2026 to 2027, from an OpEx standpoint, I mentioned about 50%. Top line growing greater than 150% versus the OpEx. There will continue to be operating leverage over the horizon.
All right. Good stuff. We only have a couple of minutes left. As you guys talk to, I know you talk to a lot of investors, what do you think is the biggest disconnect between your kind of perception of the opportunity and your fundamentals and the investment community?
Three years ago, I think we were viewed as the emerging AEC category creator. It is true that last year, the large majority of our revenue was driven by success in bringing that product to a preferred category for short connections. I think a lot of times we get classified as a copper company, I think that perception's going to change pretty quickly. With the announcement of the acquisition of Dust and bringing that silicon photonics PIC capability in-house, we also decided to maybe talk a little bit more about the other parts of the portfolio. I think that the perception over the next three years is going to change. I think it's going to change. There are other waves coming with AECs and OmniConnect.
I think the takeaway, the change in perception should be that we're really a full-spectrum connectivity company that is delivering the highest reliability in the industry, from die-to-die, all the way to any length within the data center. I think the story is unique, and I think that within that three-year period of time, you'll see us be very balanced between copper and optical. I think we're going to be talking about driving revenue at a very fast pace, even as a small company, reaching $1 billion in revenue is a big thing. We flew by the milestones from achieving $500 million run rate to a billion-dollar run rate. Now we're on a more than a billion and a half dollar run rate as we look backwards. We're focused on $5 billion and beyond.
Call it $5 billion, and then $10 billion. I think we're addressing the kind of markets that will take us there and take us there within the decade.
Good stuff. I think we're at time. That'll have to be the last word. Bill, Dan, thanks for joining and sharing your insights.
Yeah, thanks a lot.
Thank you. Yeah.
Great presentation. Yeah.