Okay. Good afternoon, everybody. Thank you for joining us at the Mizuho Conference today. I'm Vijay Rakesh, senior semi analyst and semi cap equipment and auto technologies analyst for Mizuho. Joining me today is Bill Brennan, President and CEO of Credo. Having joined Credo in 2013, Bill has overseen the scaling of Credo from 15 employees to 400+ today. Prior to Credo, Bill was Vice President of Storage Business Unit of Marvell for 11 years. Prior to Marvell, he was at NEC and Texas Instruments. Dan serves as CFO of Credo. Prior to Credo, Dan served as VP of Finance at Siva Power. Prior to that, he was at SunPower, Marvell, Xilinx, and others. Bill and Dan, thank you very much for joining us. Always a pleasure.
Thank you for having us.
Stocks at 52-week high, can't be more happier than, you know, everything seems to be coming together, so-
Yeah. Thank you.
With that, Bill and Dan, maybe you can get started off and give us a little bit of an overview of what you are seeing. You know, how is the industry evolving, and where do you see it going? And then maybe we'll go into some quick Q&A, sir.
Sure. So maybe I can give a little background on the company for those that aren't familiar with the story.
Yeah.
Credo is a high-speed connectivity company, really pure play. That's what we do every day. All of our solutions are related to connectivity. The first market that we pursued was the Ethernet protocol. So the Ethernet market is really the market within the data center above the server. So all of the switching and routing is done with the Ethernet protocol. We started there first because of the speed. Over the last several years, there's been an exponential increase in the bandwidth, and measured by single lane speeds. It's really gone from 10 gig to 25 gig, to 50 gig, to 100, and now we're working on 200 very actively. That curve represents technology difficulty, so that was the hurdle that attracted us to this market.
We have extremely good designers in the analog space, in particular, that have enabled our solutions to be a lot different from our competition. And so with the business strategy, we decided to target all of the connections in the data center that were Ethernet, including copper as well as optical. And so we've got a full portfolio of solutions now that includes active electrical cables, line card PHYs, optical DSPs, SerDes chiplets, as well as we license our IP to partners that are building larger, more complex chips than we build. One of the things that's happening in the data center is this, you know, this amazing growth of generative AI applications.
And if we think about AI compared to the traditional, the traditional general compute and, and traditional network, the front-end network, as we're calling it now, AI is a, an application that really puts, connectivity, you know, in, in the forefront. You know, if we think about, what NVIDIA has done, they, you know, they, they brought GPU expertise. With the Mellanox acquisition, that was really a connectivity acquisition.
Mm-hmm.
And if we think about, you know, the AI supercomputer application, it's really tying, you know, many, many GPUs together to act as one supercomputer. So it's like a neural network, and so there's really insatiable demand for higher bandwidth. So for a connectivity company like Credo, you know, surely this is wind in our sails because the per lane speeds will go to the fastest available in the market, which today is 100 gig per lane.
Yeah.
And the number of connections has grown exponentially as well. So the way you get more bandwidth and data to these GPUs is really you have to go faster single lane or wider in the number of connections. And so that's probably a good starting point as far as-
Yeah. No, absolutely. So to that point, as you go from compute to AI, you have talked about a much, much bigger TAM, that you work, especially with your core business, which is AEC. I think on the compute side, you're talking 2-4, you know, servers per rack. You know, as you go to AI, you're talking, you know, almost 8-12, but then the number of connections-
Mm-hmm
... go up a lot more, right? I mean, you have almost 50 or 60 connections per rack on an AI server versus, you know, it's probably in the 10-12 in the, in the compute side. How does that? How do you see? So in, in theory, your, your growth starts to accelerate as some of the AI, especially the Ethernet side, not InfiniBand, but Ethernet side starts to come to it. Again, maybe you can talk to that-
Yeah
... how, how you see that progressing.
Sure, sure. So first, let me start with, you know, talking about the network that's dedicated to the AI cluster, and so we talk in terms of that being the back-end network. And so it's a network that is networking all of the GPUs in, say, you know, a cluster of 10 racks or 20 racks of appliances. And if we look at the traditional front-end network connection, that still exists in an AI cluster.
Yeah.
For every appliance, there's a single, front-end connection. But if I can talk about the bandwidth today, you know, that we're seeing for the front-end and back-end networks, one of the programs that we've worked on, the front-end network is connected with a single 100 gig connection.
Yeah.
So 100 gig for the front end. The back-end network on a single appliance has got 4.8 Tbps bandwidth, and that's represented by 12 ports of 400 gig that are all running at 100 gig per lane.
Got it.
So you think about the sheer bandwidth from the front end compared to the back end, and you think about the number of connections.
Yeah, the back end is opportunity there.
Right. So those 12 ports are served by 6 AECs.
Got it.
So they're 800 gig switch ports that connect to two 400 gig ports of the appliance. And so you're talking about 1 versus 6, that's the ratio. And if you think at about a rack scale level, you're right, you know, 48, 48 connections for the back-end network and 8 for the front end, so 56 per rack in application today. And that's moving to a design that's got close to about or maybe a little bit more than 200 per rack. And the reason there is we're not going faster single lane, so more and more bandwidth is required-
Yeah
especially as next-generation GPUs come to market. And so moving from 100 gig to 200 gig, the market is not ready.
Yeah.
And so what we're seeing is going wider with the number of connections. And so the application itself is absolutely wind in the sails for connectivity companies like Credo.
Yeah. You are starting to see 800 G pick up, you know, by the end of this year into next year. And that, at 100 G, it starts to get into where you need an AEC, you know, optical DSP start to-
Yeah
... optical, you know, cables start to hit a ceiling there in terms of power consumption and signal noise, et cetera. Are you starting to see that, you know, especially some of the big guys, NVIDIA, is now talking about 800 G going to 1.6 and all that, but are you starting to see that pull-through happening?
Sure. So it's, it's important to think about what, the way things were traditionally done for rack scale connections-
Yeah
... meaning connections that are 2.5 meters or less. Traditionally, they've been done with passive copper cables. You know, the world that I walked into 11 years ago here at Credo, there was an assumption by the market that when passive copper cables were no longer something that could be used because of the speed of the connection, that short connection would become optical-
Yeah
... laser-based optics. And so we saw the opportunity because, of course, we're building optical DSPs. We thought if we put a DSP in a and made an active copper cable or an active electrical cable, that would be, we felt, a very huge market. I have slides going back to 2014 that we're theorizing-
Yeah
... you know, the size and possibility of this market. And so we started working on it, you know, about probably seven years ago or eight years ago, and we've really pioneered the market. And I’ve said many times that 100 gigabits per second per lane, DAC is dead, so the passive copper is no longer usable.
Yeah.
You know, the product, compared to a laser-based optics solution, it's just so much better from the standpoint... Yes, you can start with half the power, half the cost, probably 100 times as reliable, and the form factor is,
Right
... very rugged and thin, so the routing density can be achieved. And so I can say today that this is the de facto standard for-
Yeah
... for rack scale connections at 100 Gbps-
Right
... per lane. And so, yes, we're seeing adoption. We're engaged with all five hyperscalers-
Right
... at, at different stages. We're in production with two now, going this fiscal year with a third-
Yeah
... and the next fiscal year with the fourth. But also we're seeing, you know, I've probably got, you know, 6 to 8 other significant customers-
Yeah
... that, the decision-making is, you know, just very, very easy.
Yeah.
Of course, they're going to use AEC-
Right
... and not laser-based optics.
And talking about the next generation, I think, you know, obviously, when all these pictures floating around of the Microsoft next Maia, you know, console, it all shows these purple cables. So probably all the Credo cables in there. But are you seeing a similar pull-through on, like Amazon Graviton and, you know, Meta is trying to build their custom silicon as well? And when do you actually see that start to gain momentum? Because a lot of these guys have still been ordering a lot of NVIDIA stuff, right?
Sure.
I mean, even this year, there's a pretty big install of NVIDIA DGX, you know, B100 and all, or whatever, H200 going on. When do you see the custom silicon side start to get pull-through? And that's where you will get a pull-through on the Ethernet AEC cable side, I guess.
Yeah. So I think, you know, first of all, the real question is InfiniBand or Ethernet?
Right.
And the reference to NVIDIA, yes, they are shipping-
We'll hear that tonight on Broadcom, so.
Yeah. Yeah, so, when they ship an entire cluster-
Yeah
... you know, that's where the InfiniBand is used.
Yeah.
But also, they're selling their GPUs without the cluster, and it's being applied in the Ethernet with an Ethernet protocol.
Right.
I think the five hyperscalers in the U.S. have all spoken and, you know, the long-term plan for all of them is Ethernet.
Yeah.
That's kind of publicly available information.
Yeah.
And, you know-
I think Arista also talked about that every, out of every 5 installs, like, only 1 is InfiniBand, all the 4 of them are asking for Ethernet or something like that.
Not sure if I have that-
Yeah
... that kind of specificity.
Yeah.
I can just say long-term, at the hyperscalers-
Right. Got it.
... Ethernet is preferred. And so, you know, what we see at the cable connection, if there's a NIC, we see that NIC.
Yeah.
We don't really know what's behind the NIC. It could be something internally developed, could be, another GPU from any of the other companies.
Right.
What we interface with is the NIC.
Yeah.
The custom silicon that's being developed, they're going native Ethernet, right? So that's very clear that-
Right. Yeah
... if it's internally developed with native Ethernet, that's the protocol that-
Yeah
... will be used. But I think, I think that whole discussion is, you know, it's been popular this year, but I think as we think about our business long term, I think at the hyperscalers, it's very clear that,
Right
... that we'll address the wide opportunity of Ethernet.
The long-term roadmap is towards kind of the whole Ethernet and.
Sure
... Ethernet and all the-
Yeah
... possibilities beyond that. I want to go to the next big segment that you guys have talked about is optical DSP.
Sure.
You have a product, the LRO in Linear Receive Optics, and that seems to be getting a lot of traction. You talked about trials with three customers, you know, three hyperscalers. Maybe talk a little bit about that. What's the value proposition? How does it compare to a full DSP? What's the cost, you know, efficiency that you deliver to the hyperscalers and, you know, how that adds to their TCO, et cetera, et cetera. Yeah.
Yeah, sure. So the way that things have been done for the past decade, as far as optical modules go, is that there is a DSP done in each end of the connection on both lane, or on both directions of traffic. So always there's a bidirectional connection that's being made. And doing DSP, what is that? That's basically equalizing the signal. It's basically recovering the clock and data information and transmitting a clean-
Yeah
... clean signal. So the way things have been done is that there's a DSP on both ends of the connection, on both directions of traffic. A year ago at OFC, now this is going back to 2023, there was really a call to action to address the power challenge for optical, specifically for the 1.6T generation, 1.6T port generation. These are eight lanes of 200 gig.
Because there's a whole new architecture that's coming on for that.
You know, it's a whole new architecture. It's a very steep climb.
Yeah.
The call to action was, how do we, how do we take the power out?
Right.
Because we're, you know, heading to much higher power levels-
Yeah
... as we go from 100 gig to 200 gig lanes.
Right.
So there were companies for 800 gig ports that were showing optical modules without any DSP.
Yeah.
This is called, this is what we refer to as LPO. Now, that approach has many, many pitfalls.
Yeah.
Of course, you go to a non-standard-
Right
... space interrupt is impossible, basically. There's no interrupt. There is, you know, fine-tuning, hand-tuning of every lane because,
Yeah
... because the signal integrity is very difficult. The promise there was to reduce the power of the DSP, which is the highest power component-
Yeah
... with the module. So we thought this is, you know, of course, being an optical DSP supplier, we thought, you know, there's a lot of pitfalls.
Yeah.
But we didn't just say it's always going to be a full DSP world. So we went to work one month later thinking maybe we take out half the DSP.
Yeah.
We have a chip in each end of the connection, but we'll do a DSP on just one direction-
Right
... of the traffic, and basically equalising that copper signal coming from the switch. And the data that we got was quite good, right? And so we decided to bring a product to market. We brought a product to market by November. By March, we were demonstrating with, with, with three partners at OFC in 2024. And what we showed is that, yeah, we can, we can build a, a, you know, a, a, a product that's a, an 800 gig module, whether it's an AOC or a transceiver-
Yeah
... with solid, you know, solid bit error rates.
Right.
But our approach, you maintain industry standards, you have interoperability.
Yeah.
You've got, you know, robust signal integrity. You've got the same test points that the industry is familiar with. And so it's basically what we showed was lowering the power from 15 watts on an 800 gig module-
Yeah
... to 10 watts or below. So, since OFC, we've now engaged with, you know, additional partners, so now we've got more than a handful of partners that are developing solutions.
So going from 2 DSP to 1 DSP goes from 15 watts to 10 watts on a-
The overall module power.
Got it. Okay.
So it drops the DSP power in half.
Got it. Okay.
Yeah. So it's I think it's promising technology to the market. Power is discussed more and more and more-
Yeah
... and now we're making our decisions, you know, say, on process geometry-
Right
... really as power is the primary decision-making point.
Sure. Yeah.
So, for example, as we go towards our DSP for 1.6 T, we're moving straight to 3 nanometers, so we're kind of leapfrogging-
Right
... the competition that's in 5 nanometer, and that's because of power. In 5 nanometer, the solutions in the market are, you know, 20 watts or higher.
Yeah.
which, you know-
Right
... I mean, the ceiling on some of these standard connectors is 20 watts.
Got it.
Our target is to stay in the same kind of power range as an 800 gig DSP, so 10 watts or below. And the LRO version-
Yeah
... will be 5 watts. So that's, you know, that's what we're trying to deliver to the market-
Got it
... is power that kind of fits within the existing model, the existing ecosystem.
Got it. Okay. I want to get Dan maybe talking now. So when you look at the optical DSP side, obviously, you know, it's a silicon chip, much higher margins versus doing an AEC cable. How do you see that ramping? And, obviously, this year is a year where you're going to see a pretty significant ramp on the optical DSP side for Credo. How much of the contribution will be LRO, or do you see LRO more of a fiscal 2026 kind of contribution? And is that, again, back half fiscal 2026, or where how do you see that coming through, I guess?
Yeah. Specifically with optical, we've stated that 10% of revenue is our target for fiscal 2025. If I step back a minute and talk about our kind of long-term margin model that we've laid out for everyone going back almost three years now, our long-term margin is 62, 63-65%, and that's with IP as a percentage of revenue being 10%-15%.
Right.
So if you back it-
That's pretty big this quarter. Yeah.
Yeah. So if you back into that math, that means the product alone revenue is about 60%. And then we further laid out kind of a broad margin profile across product lines or product categories that we have. Optical being on the higher end-
Yeah
... you could think of Inphi was kind of a pure play, so you'd have comparable margins to that, you could presume. Our chiplet business, our Line Card PHY business, is kind of right within that, target long-term model, and our AECs are a little bit below.
Got it.
That's broadly speaking, you know, there's going to be some variability across different products, different customers, et cetera. So, what you'll expect from a gross margin perspective this year, fiscal 25, you know, we ended fiscal 24 at 62.5% gross margin. So, quite good execution, great performance. We were helped by a large contribution from IP revenue at the higher end of that range-
Yeah.
- along with, kind of an outsized engineering services revenue, particularly in Q3-
Yeah
... to help develop the next generation chiplet with our lead chiplet customer.
Right.
So what you'll see this year, we laid out a kind of margin target of 61%-63%, so kind of similar to last year.
Yeah.
Although services revenue won't be as,
Got it
... large of a percentage as it was in fiscal 2024.
Less offset, yeah.
Then by fiscal 2026, you would expect we'd kind of be right within that range.
Right.
There's a lot of levers involved, and optical, of course, as that contributes more and more, that that's very accretive to the margin that we have.
Would you expect a lot of revenues in fiscal 2026, kind of as some of these-
Fiscal 2026, yes.
Yeah. Yeah, so-
So as we go into fiscal 2025 and we're engaged in development-
Right
... it takes time to get through qualification and qualification with hyperscalers, so it's more of a 2026 type of-
Got it
... revenue impact.
So, post your earnings call, I mean, we got a lot of questions on some of the comments that you guys made. One was 75% of revenues are AI. So maybe you can talk to how you classify that. And second was you obviously see doubling this year. So maybe you can, you know, talk to how you classify and what's the confidence interval around that, and how you see that.
Yeah, let me level set, let me talk about our Q4 of fiscal 2024, where we stated that 75% of our revenue, and our overall revenue was just over $60 million, so call it $45 million, was AI related. Now, the easiest way to think of this is, if you think of the hyperscalers, what they're deploying today, a lot of the dollars being spent are being deployed on AI-related deployments. So, and that impacts everything that we sell, even if we're selling our line card PHY chips into an Arista or a Cisco, those are being deployed in backend networks. We talked a lot about our chiplet business in fiscal 2024, which was very strong. That's our lead chiplet customer, automotive company-
Yeah
... they developed the Dojo supercomputer.
That's right.
Our chiplets are being sold into that. So there's a chiplet revenue that's AI, plus there's the engineering services piece-
Got it
... for their second generation. So, long story short, the easiest way to figure out what the AI contribution is, is to identify what is not AI specifically.
Got it. Okay.
One piece that was non-AI related was our, a portion of our IP revenue, about $10 million of it, was the last tranche of a USB-C4V2 license contract that we have with a large consumer company.
Got it.
So that we did not classify as AI. Longer term-
Sure, yeah
... who knows what USB-C-
Yeah
... comes classifying as. But and then the other piece is of the Microsoft front-end network cables that we've been selling since our fiscal 2023, we estimate that probably about a third of those are for general compute, not for connecting back-end networks.
Got it.
That estimate is based on what we've heard, you know, what they tell us. So it's not a precise measure-
Got it
... but that's,
Fair, yeah
... that's how we categorize it.
Got it. And then on the visibility on the 100% ramp, I guess, the chiplet guy has been spending like crazy, but he's been spending on DGX too, so.
Yeah. I mean, generally speaking, we see strength across all product lines, all product categories coming up this year. However, there's really two main drivers in that step up in revenue-
Yeah
... Q4 to Q4, and what we expect them to be is our second AEC hyperscaler. You know, they've been contributing to revenue for us for a while, but in our Q3 of fiscal 2024, kind of the start of a ramp.
Yeah.
You can-
Where you have, like, 3 cables or something with the second customer?
So, uh-
Yeah
... there's probably 12 different programs.
Right.
But, there's a few key programs that we expect to ramp throughout the year.
Got it.
And that will really. That's the biggest factor in that.
In the fiscal 2025
... in that.
Got it.
The other piece of that is-
You see that the H2, fiscal 2025 or starting into fiscal 2025?
Well, we've described the H2 as kind of an inflection point-
Got it
... really driven by that, along with Microsoft, the front-end network cable volume returning to-
Right
... what it had been, you know, pre-
Okay
... 18 months ago.
Yeah.
So there's a number of factors.
Got it. Okay. Maybe I'll take some questions from the audience.
Sure.
Repeat it if you can hear me, but you talked about... Thank you. Thank you. My name is Janine Prezioso. I work for Mizuho. You mentioned the word 10 watt, 20 watt, and I'm picking up on electricity generation, and that is-
Sure
... a big factor in all of the compute power that we need, or that we have, and to kind of power it. I just wonder what your outlook is for that, and how are these hyperscalers... I understand some of them are building their own, you know, power generation on-site, but how do you view that in terms of, I guess, a headwind going forward?
I would say it's definitely not a headwind for us. You know, it may be a headwind at a certain very, very, you know, macro level, you know, to identify space and you know, power source and that whole bit. But, you know, that, I don't think that will be a headwind for us. You know, we're really kind of ground floor with new deployments, and new deployments are accelerating, if anything, right now. So I think it'll be quite some time before we would say, you know, our business is impacted by some macro-level concept like that. So I think the next three years are pretty locked in, you know, if you talk to those companies that focus on this business.
Thank you.
Any other questions?... Jesha, go ahead.
If we look at the top five customers, how much percentage of revenue those top five account for now and maybe a year from now?
So, I guess when you say the top five customers, that you mean in the market, potential customers?
No, for your, for your customers.
Oh, I see. Okay. So when we think of our top five today we've got great diversity, and two hyperscalers only in the top five. And so I guess maybe restate your question.
So he's asking your top five customers are 50% or 70% of revenue?
Yeah.
How do you see that?
Let me say, we don't track that number specifically, but it's, there's five U.S. hyperscalers that dominate the CapEx spend-
Mm
... that buy from companies like us. So longer term, if you take five customers, and we're at all of those five hyperscalers longer term, you would expect it to be a very, very high number. So maybe talk about our 10% customer. Yeah. So for fiscal 2024, which we just ended, we had three 10% customers, the largest being 26%, second largest being 20%, and then fifteen or sixteen percent. Right... for the, for the third largest. But, but it's a, it's a high number today even, right? Based on that.
I have one more-
Yeah
... quick, unrelated question.
Sure.
There's a lot of discussions around InfiniBand and Ethernet.
Sure.
May I? I would like to know your thoughts on it.
Sure. So our thoughts are based on the feedback that we're getting from the five U.S. hyperscalers. It's very clear to us that long term, all of them plan to use the Ethernet protocol. So if you look at their roadmaps, how they're, you know, building the architectures, a lot of them are building internal solutions, and across the board, those are all Ethernet. So I think that's, that's only a function of time.
All right, any other questions? Okay, so Bill and Dan, I mean, you guys talked about USB-C into consumer, and obviously, you have the MacBook Pro and the iPad Pro and all those ramping with that. Do you see a possibility to get into the handset side, and why not Android?
Yeah, so generally I'll maybe give a little bit of background.
Yeah.
We initially targeted the Ethernet market because that was going at the fastest speeds. We've recently talked about PCIe. We're targeting Gen 6, which is 64 gig PAM4. The degree of difficulty for that technology is kind of, you know, right down our alley. We're going to bring the same kind of advantages to that market that we brought to Ethernet. We had a customer approach us more than five years ago, talking about USB.
Yeah.
It was a real head scratcher at first for me because-
Yeah, sure
... you know, we're talking about high, high speed connectivity-
Yeah
... and you don't think of USB that way. But they were talking about a, a new protocol, the next generation USB protocol, being two lanes of 40 gig PAM3. And they were talking about power numbers that were so low, probably because everything is battery powered.
Yeah.
So low that, you know, one of the, you know, the deep competitive differentiations that we've got is our architecture at a SerDes level, kind of a core technology level. I knew from the first conversation that if they were serious about that spec, we would be the supplier of that.
Nice.
Here we are today. We've delivered an IP to them that they're going to integrate into their products, and we expect that, you know, most likely towards the end of this year, they'll introduce products with that next generation USB protocol. You know, the conversation about where it goes, of course, we have no insight as to where they're gonna deploy it. But I think generally, when you think about why not Android, for sure, there's an opportunity there.
Yeah.
And so when I think about our team, I don't put boundaries on where we're going to play. I don't put boundaries on the protocol. I don't put boundaries on the types of products and solutions that we develop. And I think naturally that as the market goes towards this high speed, there's going to be great opportunity for Credo. And I think that when you think about, you know, plugging into a USB port, of course, it's plugging into SerDes, right?
Yeah.
Although that's not a market that people have really talked about too much as being, you know, interesting from a chip perspective-
Yeah
... a lot of people are calling it edge AI now.
Yeah.
The announcement this week, earlier this week was, you know, or last week, I can't remember when it was, but that was interesting.
Sure.
You know, the idea that you'll be holding, you know, edge AI devices in your hand. And if you think about, you know, the ultimate application, maybe a combination of what's being done at the handset level or the device level, you know, that combined with the cloud, it could bring applications that we haven't even thought of yet. So I'm actually bullish on AI, not just for the cloud, but also for the consumer.
Right. Yeah, got it. But when do you see that picking up? Do you expect like a royalty? If it gets into the handset, let's say, would you classify that as a separate line item, edge AI, or does it, does that get clubbed into something else? When do you see that number picking up, I guess?
It's part of the IP revenue.
Right. Okay.
We essentially disclose IP and product.
And that's. So the IP revenue would go by units, I guess. Is that right?
That's the way the royalty work, yeah.
Okay, got it.
Yes.
Okay. How much time do we have? We are almost out of time.
Second.
Great. Fantastic.
I appreciate all the good questions.
Exactly.
Thank you for the interest.
No, thank you very much. Thanks, Bill.