How Retimers Built an $80B Company: The Story of Astera Labs

A retimer is a small, unglamorous chip. Astera turned it into a high-margin franchise. Austin and Vik cover why copper gives up at PCIe speeds, what a retimer does that a redriver can't, and how one H100 socket became a moat. Then Scorpio: can Astera take switch share from Broadcom?

Key Takeaways:
- 30cm of PCB trace at Gen 5's 32 GT/s and the signal is gone. Signal integrity is an inside-the-server problem, not just a rack-to-rack one.
- Gen 6 swaps NRZ for PAM4. Four voltage levels instead of two — a third the eye height, ~9.5 dB of SNR gone.
- A redriver amplifies, noise included. A retimer recovers the clock and data, rebuilds the signal, and resets the jitter budget.
- Astera won H100 by shipping PCIe 5 retimers at volume first, then bundling COSMOS on top. The chip became a fleet telemetry sensor.
- Sticky socket: H100 carried straight into an Aries 6 design-in on Blackwell.
- Scorpio moves up the stack. Same signal conditioning IP, higher ASP, pointed at Broadcom's PCIe switch business.
- Astera granted Amazon a warrant — 3.26M shares at $142.82, vesting against up to $6.5B of purchases through 2033. Equity flowing to the customer, and it sizes the Scorpio P-Series commitment behind Trainium 2 and 3.

Chapters:
0:00 The Astera Labs Franchise
3:32 The PCIe Copper Problem
6:53 Signal Smearing and Jitter
11:52 Eye Diagrams Explained
14:59 Equalization: Fixing the Signal
23:52 Redriver vs. Retimer
29:30 How Astera Won Nvidia
35:22 The Blackwell Scare
38:50 Scorpio: Moving into Switches
42:52 The UALink vs. Ethernet Battle
45:58 Taurus and Leo Product Lines

Follow Chipstrat:
Newsletter: https://www.chipstrat.com
X: https://x.com/chipstrat

Follow Vik:
Newsletter: https://www.viksnewsletter.com/
X: https://x.com/vikramskr

Follow Semi Doped:
Get more of Austin and Vik daily, free: https://daily.semidoped.com/

Transcript

SPEAKER_00 0:00

In March 2024, a company almost nobody outside of the industry knew IPO'd at $36 a share. This June it hit $499, joined the NASDAQ 100, and printed 93% year-over-year growth at 76% gross margins. Estera Labs doesn't even make GPUs. They don't make memory. They don't make optics. Astera makes the chips that fix copper, the little devices that catch a dying high-speed signal halfway down the board and relaunch it clean, unglamorous, invisible, and in every single AI server NVIDIA's partnership. Today's episode is the story of how a repair part became a franchise and how that franchise is now going after the biggest socket in the AI rack. Hello everyone, welcome to Semi Doped. I'm Austin Lines of Chipstrat, and with me is Vic Shaker, Vic's newsletter. So, Vic, let's talk Astera Labs, but let's talk re-timers. Like when I first heard about retimers, I was just like, what does that even do? What does that mean? So, what do you what do you say? Should we get into that today?

SPEAKER_01 1:06

Yeah, like let's talk about the problems uh of copper connectivity, which we have spoken about a lot before, but I think re-timers is an interesting concept because we always talk about stuff that is outside, let's say, a server tray, right? We talk about how GPUs are connected together and scale up, we talk about how to connect racks together and scale out and scale across, all of that stuff. But what we've never covered is how does a GPU connect to a CPU within a server tray, or how does a GPU connect to the network interface card or the NIC? We'll keep referring to this as the NIC, NIC, network interface card. So how does that connection happen? You know, it may not seem like it's much of a connectivity problem because come on, how far are these things, right? Like the server tray is maybe this, you know, a big with about two feet wide by four feet deep. But that is still a problem because the way you connect a GPU to a CPU or to a NIC is through the uh PCB that is inside the server rack, right? Like it's it's a host board that is holding all these chips. And so those are not like excellent conducting mediums to carry high speed signals, right? And so that is uh the problem we're going to talk about today and how Astera Labs solved that problem and became pretty much an integral part of what is inside the server rack, a server tray today.

SPEAKER_00 2:36

Awesome, interesting. Let's get into it. Yes, because already you got my wheel spinning because you're talking about copper PCB traces and high-speed signals. And we've been talking about, like with Credo, uh, you know, um, and active electrical cables, like the copper problem. Um, so yeah, let's get into it. When a GPU talks to a Nick and it talks over this uh, you know, PCB trace, like what are we what are we talking about?

SPEAKER_01 3:00

Okay, so let's go back from the beginning and like set the stage quickly, but also in a complete fashion so everybody understands what's happening in this uh part of the AI stack, right? Fundamentally, what happens outside, you know, we always talk about NV Link for scale up or you know, Ethernet or you know, uh InfiniBand, all these things. These are all like outside outside the server tray. What happens inside is something that has been around like very long, right? This is then every PC that you build out, it's been in gaming PCs forever. It's called uh PCI Express or PCIe. Okay, so PCIe is how uh the GPU communicates with the CPU or the NIC, for example. It's just a protocol that's been around for a very long time. So the whole idea is that these don't have to run very fast on a single lane basis. However, uh they have many parallel lanes. So a PCIe interface can have uh 16 lanes, and so it's called a PCIe X16. Now the way this is dealt or dealt with or spoken about is that uh you have so many gigatransfers per second, right? So you can have uh, let's say a Gen 5 uh PCIe uh you know has essentially let's say 32 gigatransfers per second. This was actually used in the Hopper era uh anyway for connectivity. So what that means is that why why are we saying transfers and not like bits, right? Transfer we are talking about transfers because we are talking about transferring a symbol. A symbol can be a collection of bits, it can be one bit, is one symbol, which is basically this NRZ modulation, zero or one, very simple. Or it could be a PAM4 modulation, which is like four levels, right? Uh so you could have different kinds like so when you do PAM4, you have two bits to a symbol. So gigatransfers means you're transferring 32 uh billion uh symbols of either one or two bits uh every second. So if you look at that in the time domain, like let's see, like how how so what are we talking about? How many symbols every second? So every every 31.25 picoseconds, you have a symbol going by. So you think of a string with like a lot of beads on it, and each bead is a symbol, and the beads are moving from left to right, right? So that's that's how it looks. Uh if you kind of imagine bits moving on a wire. That's exactly what's happening between the GPU and the CPU.

SPEAKER_00 5:32

Did you say picoseconds, like 10 to the minus 12?

SPEAKER_01 5:35

Yes. Yes, it's very less. Yeah, it's very it's a very small amount of time, and that's what it takes to transfer like a billion uh symbols per second. It takes picoseconds, like uh you know, it's a very, very small period of time. And so the thing is that it's not easy to do this, okay? It's not easy to move bits across from one place to another because it faces several problems. Copper is a problematic medium, and when you try to push these bits from one place to another, things go bad. Okay, so one thing is like loss, right? Like, you know, you put you know a signal of high strength in one end and it loses all its strength by the time it reaches the other end of the wire, right? That's one problem. Then the PCB itself has plot problems and all that. And then these things are like high-speed signals, so you you can think of them like sometimes they ping-pong, so they reflect from that end. So you can think of the beads trying to come back on the wire. You don't want that, right? Those are reflections, right? Uh then crosstalk, you know, if you put two sets of beads uh next to each other. Imagine if one bead on this string is affecting the bead on the next string. That's a that's a terrible outcome. But these things happen in copper. So all of these things, they basically need correction. They need some way to fix this, right? But that's just the copper thing. Actually, I should mention that there are two more problems we should talk about. And that is intersymbol interference is one of them, which means the beads on a string don't actually stay like one bead on the string. Like, imagine that you know they're made of sugar or something and they try to like stick to each other and try to become uh one larger ball of sugar instead of like staying distinct sugars. Like as they travel really fast, you can imagine they like eat up or something and then try to stick and become one and like a blob, right? That actually happens all the time. So when one symbol travels, the next symbol uh kind of starts to overlap with it and it starts to spread out, and that's terrible, right? You don't want inter symbol interference.

SPEAKER_00 7:37

Let me jump let me jump in there for listeners. You know, so you can think of if you're familiar with like square waves or sine waves, like if you have a square wave, you can imagine that as it travels down this medium, it starts to spread out instead of being a square, you know, maybe it's like more of like a hill. And so you can imagine if you're sending these different like pulses in the simplest sense that as they start to spread out, they could start to overlap each other to this point of like the sugar sticking together. So now all of a sudden, instead of like a steep up to one, steep down to zero, some space, and then another steep up to one, you start to get like a ramp up to one and a ramp down, and then maybe it overlaps another one to the to the point where what you receive is like a 0.5 and you're and it's slowly transitioning, and it's a little bit confusing. Is this a one or is it a zero? And then, of course, to your point, as you mentioned earlier, there's this attenuation. So by the time it gets all the way down the copper, it's just like lost, you know, the peak amplitude as well. So um, but but back to the inter symbol um sort of overlap here. You yeah, there's like this smearing where you tried to send certain symbols and they're kind of overlapping each other. And somehow we have to figure out what was said at the end. Yep. Carry back on. What's the yeah, yeah, what's the time the next one that you were gonna mention?

SPEAKER_01 8:53

Yeah, the other one is jitter, which means that the pulse is not where you think it is, right? Uh the pulse, you think you sent it uh some time ago, but by the time it reaches a given location, it's it isn't there. Uh there are random uh variations in the system that cause the pulse to be further than you think it is, or like earlier than you think it is. So when you go in, be like, okay, at this moment I'm going to look and see if this is a one or a zero, and I'm going to like lock it down, right? Like that's what you were saying. Like, let's say that you know the point five question isn't a thing. Like we are constantly monitoring this particular point on the wire, and we're going to be like, okay, when the thing comes, I'm going to call it a one or I'm going to call it a zero. But the thing doesn't come at all, or it comes like you know, late. And you're like, what happened? So now you can't detect the bit because the thing wasn't there when you expected to be there. That thing is that thing is called jitter, and that also causes problems in communication. All of these are big issues.

SPEAKER_00 9:50

Gotcha. Okay, yeah. So you're saying like with jitter, it's like if I'm thinking about like I'm on a clock, I'm at the end of the wire, I'm checking, and it's like, what am I reading? What am I reading? What am I reading? It but but nothing comes, but then it comes a tiny bit later, like in between readings or something, you know, it's like, what's going on here? But let me ask you, Vic, um, between the inner symbol interference and the jitter, like how much of that is deterministic? Like we can figure out, oh, this is copper and we know how a high frequency symbol behaves, and how much of that is just like random noise or random something, randomness?

SPEAKER_01 10:26

I think that um intersymbol interference is somewhat deterministic and you can kind of figure it out. And there is a way to do this, and I'll get to that. Jitter is more random, you really can't do much about it. And all this affects uh how we will ultimately detect if it's a zero or a one. See that the whole problem here is to tell whether at any particular instant you have a zero or a one. If it's a two-level modulation, if it's a four-level modulation like PAM 4, you have to tell which of the four levels it is. And typically they they talk about it in PAM four as uh zero, one-third, two-third, and one. Instead of trying to use something else, you know, they just uh uh break up the one into four four levels, right? So sure it's up to you to tell what the level is. That's the whole problem.

SPEAKER_00 11:22

Uh yes, okay, yeah, I can see how that's challenging, especially for PAM 4. Like if you're trying to read zero volts is zero zero, a third volts is zero one, two-thirds volts is one zero, and a volt is uh one one or something like that. Yeah. How um now all of a sudden when there's that smearing and stuff, it's not like, oh, is it a zero or one and we're at 0.5 and we don't know. Now it's like really crazy. It's like, oh, it's supposed to be zero or zero point three three, but I got zero point two eight volts or zero point one six volts, or you know, it's tighter. There's like less, there's less margin.

SPEAKER_01 11:55

Yeah, yeah. There is a nice way to visualize this, and you will see this mentioned a lot in this interconnect world. And those are what are called eye diagrams. Okay, and I'm just gonna explain quickly what that is. So imagine that this one particular time called a unit interval, that is basically the 31 picoseconds that I mentioned. Because you know that that is the symbol uh period, right? That's that's the period. So that thing is called a unit interval, and is usually on these charts you'll see it as UI. And so you're looking in that unit interval and you're seeing all the bits that go by. Like you're looking at just one bead on the string, but this that bead is moving, right? Like you're continuously looking at the string. Think of it that way. And what you're looking at is an overlapping bunch of zeros and ones, for example. And always you want to have some kind of a separation between zero and one, right? So that you can clearly say anything above 0.75 is a one, or anything below 0.25 is a zero. Like we just round up that round down that way. And so, but nothing should be there between 0.75 and 0.25 because you don't know where it goes. The uncertainty is too much, right? Yeah, so that that gap is usually referred to as the eye opening. The eye of this signal must be open because if the eye closes, which means there is no space between the voltage levels, you can't tell what it is. So the eye opening determines everything.

SPEAKER_00 13:21

Okay, okay. So you you basically take a bunch of snapshots at this unit interval and then plot them all like on top of each other, right? So you take like a thousand of them. And in the ideal sense, you would have like a like a rectangle. You have a bunch of ones and a bunch of zeros, and that's it. But of course, it there it like takes time to get up to one and time to get back down. And so that's why you kind of end up getting this I shape because it like ramps up and ramps back down, and and same. And but to your point is when you throw thousands of these measurements on top of each other, there should be some sort of gap where it's very clear, hey, there's nothing in here, so it's clearly a zero or one. But if they start to like, if that gap closes, that's bad news because hey, some of these measurements that we took are in the middle and we won't know what to call it. Exactly.

SPEAKER_01 14:06

So now the whole problem of why Astera Labs uh solved this problem is that they needed to find a way to keep this eye open when the communication is going between a GPU and a NIC or a GPU and a CPU. And the distance isn't in the range of like about 30 centimeters. It's not that much, but it's still required because at the speeds that we are talking about, because Gen 5 is 32 gigatransfers per second, Gen 6 is 64 gigatran transfers per second, and it goes to PAM 4. So that makes it harder because you're adding more levels and you're going faster. So it really requires retiming. And then in the future, um, we will go to Gen 7, which is at 128 gigatransfers per second with PAM 4. So Gen 6 is already in, I think, in the Blackwell era of chips, it's being used. So yeah, we'll get to all this. But so Astera Labs is basically their job is to keep the eye open within a compute tray. That's their whole purpose with re-timers. So, what are re-timers, right? Let's get to that portion because we need to talk about, okay, now we have intersymbol interference and we have jitter, whose main effect is to close the eye. Now we have to find a way to fix it, right? And there are a few tools to do this, okay? I think we should first introduce the idea of uh equalization. Uh it's a fancy way of saying, look, I'm going to lose half the signal like down the line. So what if I just boost the signal twice now so that by the time it reaches that point, it'll still be like what it was here. Because I amplified it beforehand, and so that when it degrades later, I can I can uh keep the voltage level constant, right? So that is equalization. You're basically applying the inverse of what will happen to the actual signal, right?

SPEAKER_00 16:04

So, okay, let me let me interject really quick. So back to the eye diagram. So you're saying the problem is the eye starts to collapse, especially if we just don't do anything, but it we're increasing the speeds. Of course, it's copper, and so kind of the the distance at which you can effectively like keep the eye open is shrinking. Um and so that so the question is, you know, how do we keep the eye open? And you're saying like one of the simplest ways is sort of like, hey, if if if we if we know that it's gonna like uh attenuate or degrade as it goes through, can we just amplify it? And it's kind of like um if you and I were in a hallway and I was talking to you, and then you went and went a lot farther away, I would just shout so that by the time it got to you, it would sound normal.

SPEAKER_01 16:47

Yeah, exactly. That's a good analogy because you know you can kind of see how far away that person is, and there's no point talking softly because you know they're not gonna hear it. So you would like to adjust your volume to be good enough so that you have a good chance of them hearing it at the other end. Yeah, exactly. That's a good analogy. I like it. That's on the transmitter side though. That's that's what you can do by shouting in louder. But what can the receiver do? What can the person on the other end do, right? Uh that's that's what uh like what you will often uh see as CTLE is it's called continuous time linear equalizer. Okay, so this is um like a long acronym, but it's in concept it's relatively simple. Okay, um it is just that you want to amplify the high voltages more compared to the low voltages so that you can kind of stretch the eye open. Imagine that you you have, I remember I told you that you know 0.75 or above is a one. Now, what if you can just amplify the 0.75 to make it 0.9 so that the eye is even more open? And what if you can make the 0.25 you know 0.1? You know, so now your eye is like in the much nicer and much wider open. So you can amplify the highs while keeping the lows low. Okay, so that's that's one way the receiver can deal with this um uh problem of all these channel imperfections that happen. Uh this is still analog stuff, okay? There's no like digital work going on. This is just a simple amplifier. But the problem with this is that it amplifies all the high frequency noise stuff, like if there's any noise and all that, everything gets amplified because it's it's just a dumb amplifier, like it amplifies high signals and keeping the low signals low. That's it. That's not a good thing. And if there is any interference from the neighboring uh bits moving up and down, it'll amplify that as well. So it's not it's not entirely very clever, uh, but it's it's a necessary part of the system.

SPEAKER_00 18:46

Yeah, it reminds me sort of of like naively, like if I'm the receiver, if you're at the other end and we got far away and you shouted, and then uh I can kind of hear you, but instead I like put headphones on and I just hold a microphone, and then the microphone sort of like does that. It like hears you and I turn up the gain or something, and it's like, oh, okay. You what you said at one volt was down at like 0.75 by the time, but then this microphone just amplified it back up to one volt, and I can hear it fine. Yeah, yeah, yeah, yeah, yeah.

SPEAKER_01 19:13

Yeah, exactly. So now there are other other ways. There's one other way we can do, one other thing we can do with the receiver. And I think it's a little bit important. And this is uh this is the clever one. So what this does is remember that it's a very good question you asked me before as to which of these errors uh intersymbol interference or jitter is like deterministic. And the fact that intersymbol interference is deterministic actually helps because you can predict you know how this spreading of the square pulse into hills happens beforehand by running some tests on the channel, and you know how it spreads, and then you can like backcalculate and uh subtract it out of the received signal because you know how much spreading has happened, and so you add it back or subtract it back and you reconstruct the waveform. This is called a decision feedback equalizer. So it just knows what's going to happen in intersymple interference and it cancels it out.

SPEAKER_00 20:12

Gotcha. Nice. That's that is pretty clever. So does it look at like it depends on like what I just received? Like, oh, I know that I just received one, one, zero, one, and therefore I can calculate sort of in real time what kind of like noise or smearing or interference I would get from that particular pattern based on this particular channel.

SPEAKER_01 20:36

Yeah, so you it has some kind of a characterization, uh, and then it uh reconstructs it based on how much it knows the spreading or the smearing has happened. Very clever stuff, you know, very complicated stuff. And it is because of this that even GenFi PCIe works. It's a very essential component. Uh, these DFEs are very essential components, right?

SPEAKER_00 20:57

Does that really quick, does that have to get tuned in manufacturing like per board or something?

SPEAKER_01 21:02

I think it does, yeah. I think it does. Um interesting. So it's not it's not all that straightforward to do, but uh it is a system that's been around a while, so it's not like ultra new, but these are all like um very clever uh chip design techniques. You know, there are these. I have some friends who actually do this for a living and they're really good at it. Uh they're like these analog guys, they're awesome. Okay, so and when you go to these conferences like ISSCC and all that, you'll see all these designs of how they do this. It's uh quite complicated, let's just say. Uh just we're just trying to cover it at like an extremely superficial level, but the reality is very, very hard. Uh finally, I think it also extracts what is called a clock and data recovery. So uh, what that means is whenever, let's say I'll be like, Austin, whenever you see a symbol go by, just clap. Okay, just clap for me. And I'm going to sit here and kind of write down all the times that you clapped. And I'll be like, okay, I know the period of the symbol. I know, I know the clock now. So now I can set like a metronome. You know what a metronome is? It's like that music thing that goes click, click, click, click. So I can set the speed on my metronome. And uh that should work out well because now I know when the bits are coming just by hearing the sound. So I have recovered the clock this way, right?

SPEAKER_00 22:23

That's pretty clever. And that kind of helps with that problem that we talked about where it's like if I'm on a particular clock and stuff isn't flying by right when I expect, I'm like, man, this isn't lining up. But but you're saying, no, no, do the opposite. Just watch what's happening and then figure out the clock from that.

SPEAKER_01 22:36

Yeah, figure out the clock. So this is all the these are all the techniques that people use to correct it. And pretty much this is all there is to it. So you can do the the just to quickly summarize, you can do the equalization at the transmitter thing. Remember, you said like just shout louder, you know? Or you can do at the receiver side, you can do the equalizer on the receiver side, which is like just put a little amplifier or some microphone and turn it up. Uh or you can do the decision feedback equalizer, which is you know what the smearing is happening. I don't think we have a very good analogy for this, but you know, you can reconstruct based on the inter symbol interference. And finally, you have to recover the clock too, by you know, basically trapping out and fitting a metronome to it or something, and you figure out the clock. With all this information, you can kind of tell to a high degree of accuracy uh what the received signal is going to be doing. And you can correct for errors with these methods, and that is essentially with all these, build up a full you know, uh PCIe channel in the transmitter and receiver, and you with this chain, you can build what is called a re-timer. So, and I I wanted to also mention the clock and data clock recovery process because it re-times the signal, right? And I'll tell you what that does in just a minute. So now I'm gonna just explain what is a readriver and what is a re-timer, because this is a very uh there's there is a difference here. Okay, so re driver has it does everything that I just said, but it does not do the clock and data recovery part. Um and it it also does not do the decision feedback equalizing because I think it needs some amount of uh like uh sophistication to do that. So imagine you're just shouting louder and you're using an amplifier on the other end or a microphone and picking up uh the signals better. That is called a re driver, which is just an analog megaphone.

SPEAKER_00 24:34

That's it. Yeah, you're just driving the signal on one end to the other.

SPEAKER_01 24:39

Yeah. So what are the benefits of this? Yeah, it uses lesser power and energy. It's less complicated, you know. A re driver is simpler and uh things are good. That's it, you know, and you don't have much latency, you can do this quickly, low energy, all that stuff. But the downside is that when you go to faster and faster signals, you got have more and more crazy things happening to the signal. You can't recover it. So now what do you do? You have to go for the whole enchilada and go for full retiming where you get the decision feedback equalizing to remove intersymbol interference, and you recover the clock. So it's not just that uh you uh amplified it, but it's the retiming process basically accumulates all these bits and sorts them nicely into the windows they are supposed to be in. Because you remember you have the metronome clock, so you know where the bits should lie, because your your the clock that you recover tells you where they should lie. So you can take these bits that arrived late or that got messed up somehow and slot them into the correct time slots. And now you have a beautifully recovered signal because you took the bits you received and you put it on where the clock should be. So you re-time the received signal based on your recovered clock.

SPEAKER_00 25:57

Yes. Okay, okay. So to sort of simplify it, like back in the analogy of like when you and I spread out, the further out we go, I shout louder, you have a little microphone and headphones, but eventually it just doesn't work. Um, and so because whatever, in this analogy, we're just too far apart. And so this essentially with the re-timer, are we ultimately just putting someone in between us and like I'm shouting and they're listening to what it what I was trying to say, and then they're sort of cleanly repeating what I was saying down to you.

SPEAKER_01 26:33

Um I think it's a little bit different. That sounds like it's more like a repeater to me, a repeater amplifier to me. This is not really that. This is uh like somebody who knows what is going to be said on the other end to some extent. Like, I'm gonna talk to you about you know re-timers, and you are an expert on re-timers, and so even if you hear it faintly, you kind of have an idea of what it is that is being said. When you hear a garbled word, you're gonna be like, oh, he's talking about retimers. So he probably meant clock and data recovery. Because a general person who's not familiar with the domain would not understand that word, but because you heard it halfway and you know the context, you can kind of reconstruct it better than the normal person would.

SPEAKER_00 27:27

Nice, nice, very good. Okay, okay, I'm tracking.

SPEAKER_01 27:30

Okay. I hope everyone else is too. But yeah, so this is very important. So these retimers, now that we under got to the point of retiming, these re-timers are what make sure that bits from a GPU make it to the NIC so that you can then go do scale out. Or the bits from the GPU make it to the CPU, and then you can do all this agentic fancy stuff, right? The bits have to make it from point A to point B, otherwise you're not gonna have any communication. So this was this is a problem that essentially Astera Labs solved uh for Nvidia. And in their HGX baseboard, I think the H100, which is just it had like eight GPUs in it, they needed to make sure that things can communicate uh with each other on the board. On the on the board. So that's what this re-timer chip did, and it really got off Astera Labs to a running start for sure, because you need about a one is to one, I think, uh a re-timer chip for every GPU. You might even need two on either end. No, so that's like a fair bit of content that goes into uh NVIDIA uh H100. And at the time that HV1 you know came out, like that was when like AI exploded pretty much. Totally. And yes, winning this slot was like fantastic for the company, right? Like because every hopper tray that was made had Astera chips in it. Everybody, and they had like a large, even now, they have a pretty large market share. So imagine how much money can come out of that business alone. And this was called their 80s series of re-timer chips. This was like their first product line.

SPEAKER_00 29:18

Yes, yes. Okay, let's get into Estera. So, what I'm hearing you say is that when NVIDIA exploded and things took off with the Hopper era, for every single GPU that was ever sold, there needed to be a retimer so that the GPU could talk to the Nick, for example. And my question is, how did a startup own that socket? Where did Estera Labs come from? And how did they get into presumably the H100 reference design so that they would get designed in? Like that's amazing. That feels hard to do, but it's incredible.

SPEAKER_01 29:51

That's a good question. But these guys, uh, I was looking at the company earlier. Uh, all the founders are actually ex-Texas Instruments guys, and they were basically high-speed interface folks. And their ultimate bet was that in 2017 uh they left TI, I believe. And their bet was that that PCIe 5 would be required on every server board. That bet was turned out to be right.

SPEAKER_00 30:20

Gotcha. So they saw they saw this coming. They they said, oh, every server board, even presumably CPUs, um, would need PCIe, Gen 5. And at that speed, we think Copper's gonna fundamentally have a problem, and that these read timers will fundamentally be required.

SPEAKER_01 30:38

Yeah, they they they they foresaw that coming, and that like when as soon as servers go like faster, that uh they're going to have to use these chips. And so they they made this whole company. Uh, and then it was it was great when um the H100 and the Chat GPT revolution happened. It was great for Estera Labs. Uh but the question was when I was researching this, I was like, why not why not other companies? Like PCIe has been around a long time, right? Like, but what's the what's the big deal? Like, why Estera Labs, why not Broadcom, why not even TI? Um, I think one of the smart things they did was uh they have that they I think they were first to PCIe 5 market. So that is one thing. And the other thing I believe is that they have this monitoring platform they built pretty early on called Cosmos, which not only tells you when the link is going down or something is going bad, but it, you know, all this channel impairments, you know, all this uh data that we all this technical stuff we spoke about, they could capture everything on that platform. And you know, hyperscalers uh love that stuff too. So I think their software platform had something to do with it as well. And then when they get qualified, it is a major win because now you're locked in. It's a very sticky slot. Nobody wants to change this stuff if it works, right?

SPEAKER_00 32:03

Totally, totally interesting. So you're saying, like when they were seeing that this socket was needed and re-timers would be needed, presumably others saw it as well. People who work on the Gen 5 spec, for example, would have known well in advance, hey, we're trying to qualify at 32 gigatransfers per second, this is gonna be a problem. Um, but Texas Instruments has a huge catalog, for example, and so do other companies. And so they might have just thought just about like the re-timer as a component in their catalog. But what you're saying is maybe Estera thought, like, no, it's more than just a re-timer, it's also um a sensor that can give you health monitoring information about your entire fleet. So we'll capture telemetry, we'll build software around it, give you a user interface, and it's not just a little um black box component on your baseboard, but it's actually something that can give you all sorts of insights. So that, for example, if you imagined that the world of tens of thousands, hundreds of thousands, millions of GPUs in a training cluster, imagine if they can all be telling you about the health of their link, for example. I I can definitely see how obviously that's supremely valuable for the end customer, but also that it might take a company oriented around that mindset of this is more than just a little component on a board, but obviously it's very customer-centric. And then to your point, of course, sometimes when there's I should say a good time for someone to try to take some market share is to obviously be first to that next generation thing or create a market. And uh, of course, if they were there first and they got designed in, then all of a sudden now it's like their socket to lose, even though they're a startup. And again, this is the beauty of fabless um companies, which is you know, your fab partner can manufacture at scale. So if you can get designed in, you can figure out how you know how to build these things at scale. And now it's sort of your game to lose.

SPEAKER_01 33:54

Yeah, this is the benefit of startups, right? They can pick a problem and just go after it. Bigger companies uh are not as nimble, and that's that's one of the nice things about this industry that somebody can come in and disrupt something amazingly different, and which is why we would love to you know also talk to startups on the podcast and stuff. This is a very interesting uh industry because Broadcom did come in later uh as a second source. They did come up with a uh a similar re-timer chip uh for doing Gen 5 timers in 2024. But by the time when you're designed in, you're designed in, right? Like it's hard to move because the hyper the NVIDIA will have to re-qualify everything and they have to make sure that the channel link budgets are correct. It's like a big problem. Once you're designed in and it works, you're done. You know? So the question is did Astera keep the slot when Hopper went to Blackwell? Uh yeah, they did actually. They did uh keep it in the B200 boards as well. And uh I believe it was the Ari6, which is the Gen 6 part uh of the same family, which was doing 64 gigatransfers with PAM 4. So they had like a fancier chip, of course. This this chip has more ASP, good for money, right? Yeah, yeah. So uh all through, you know, the money every time they sell these chips, the revenue keeps going up and up and up, you know. So you can see all these charts probably elsewhere, like that people are charting all these revenue tramps, right? Like for Astera Labs. So it's all public information. Uh but yeah, but one interesting thing that happened in the Blackwell era was people got a little spooked because Nvidia said, like, okay, this is not happening, guys. Like the distance between the NIC and the GPU is too far. So we're gonna move it close by. And now everybody like freaked out because what does that mean for Astera Labs, right? Now, if you put them close up close together, why do you need like Gen 6? Uh I don't know, some fancy re-timer chip, or maybe they you don't even need re-timing. If it's close enough, maybe you just go with a re-driver, cheaper chip, you know, lower power, all well and good. Like, you know, so that was like a bit of a scare. But as uh reality would have it, it uh turned out differently. Do you want to know how? Yeah, yeah. So what happened was uh it turns out that not everybody deployed uh racks this way. The the the gray the Grace platform, the Grace Blackwell platform, uh had variations. Like there were a lot of customized deployments going on. For whatever reason, I don't fully understand why there were customizations, but everybody didn't use Nvidia's so-called reference design where things were placed in a certain way. There were customizations. As soon as customization started to happen, it wasn't it the reach went up again. And so the re-timers were required. And then the second thing happened, which is the birth of like custom accelerators, like XPU, the training thing, you know, all of that came in. And once all that came in, like you're not no longer tied only to NVIDIA, like the ASIC world is wide open for you. You can go and do uh put re-timer chips in all of those accelerators.

SPEAKER_00 37:05

Yeah, absolutely, totally. There's a new TAM for you. And I could see, you know, once people got used to having the re-timers and the telemetry information, it it might also be a little scary to take that information away. Even if NVIDIA said, oh, if you use if you put everything close together in our NVL72 um configuration, our reference design, you won't need it. It'll be great. But I could I could see being like, no, no, I think we do want to pay to have that telemetry and to make sure those signals are clean.

SPEAKER_01 37:36

Yeah, yeah, yeah. Exactly. So I have some numbers actually. Uh so in 2023 their revenue was about 115 uh million. Uh after their uh Hopper thing took off in 2024, they were at nearly 400 million. And in 2025, they moved to other product lines, which we will also briefly mention uh right after this. They were at 850 million for 2025. Uh, and in Q1 2026, my notes says that uh it is they're already at about 300 million, and it's just the first quarter. So you're like looking at a company that's going to do over a billion in revenue now. So all these are revenue numbers, by the way. So yeah, so you're looking at a company that you know uh revenue just three years ago was a hundred million and now they're at a billion. It's an interesting story, and that's this general story of you know how re-timers and how getting the signal from just a GPU to the other end of the line, whether it's a CPU or a NIC, could make a company a billion dollars, right? It's it's quite a nice story.

SPEAKER_00 38:43

That's amazing. So, of course, one must ask is it just re-timers or is there life beyond retimers?

SPEAKER_01 38:50

There is life beyond retimers because this is where their next uh you know big big move is really, you know, the big socket that you mentioned in the cold open basically is that they're like, why why stay within why stay within the tray, right? Like we know how to condition signals. Why don't we uh go up and make a so full switch? We'll make a switch, you know. And this is competing with basically the NV switch, or it's competing with uh the Broadcom Tomahawk switches. So they have their Scorpio, I believe, P and the Scorpio X switches. Uh the P is basically for uh scale out. So this is going to be pretty much a Tomahawk replacement switch. And they have uh Scorpio P series switch with up to like 320 lanes, it's a very high red radix switch. And then the X series that they have is basically uh for scale up, it's like the NV switch replacement that uh like anybody else could do with like merchant silicon. So NV switch is basically NV Link and Nvidia's switch, right? But if you wanted a similar performing thing, you could do uh Scorpio X. So there's that.

SPEAKER_00 40:11

Yeah, totally. So you're saying Asteris said, hey, what are our core competencies? It's it's networking, but it's it's signal conditioning. And we are doing that really well with re-timers, but ultimately let's move into switches. Is that because conceptually a switch is like signal conditioning of every lane plus routing or something? So like is it is it pretty conceptually similar?

SPEAKER_01 40:35

Yes, uh the signal conditioning part is, but then the actual switching process in packets and stuff isn't very simple that straightforward. Because so far we're just talking about signal conditioning, but how the silic the packet is actually switched within uh a Scorpio switch isn't that straightforward. It's like it literally like how to build switches, right? Um and this this complexity, because it's not incremental, is exactly what makes this a very high ASP part. Like they expect a lot of revenue uh from the Scorpio switches uh because I believe the a uh the training three uses this Scorpio X as well in scale up. So yeah, it's it's probably going to be uh quite a bit of revenue from the Scorpio uh parts uh going forward.

SPEAKER_00 41:27

Speaking of training, does Amazon have warrants in Estera Labs? I feel like a lot of times I'll have to look it up. I think they do. I think a lot of times when uh a company like Amazon works with a very early stage company like Astera Labs, um, they also to make sure that all their incentives are aligned, they end up getting some warrants or some ownership of the company. And that that obviously can help um Estera get the opportunity to build out their port silicon portfolio and have the get the right to play in train in future training. And then of course, um it maybe incentivizes Amazon to put Estera Silicon in their silicon when they can, instead of, for example, Broadcom.

SPEAKER_01 42:17

Yeah, yeah, yeah. Oh, I mean, that was that there's another big big story there, which I'll be briefly mention because it's really interesting. Uh because, you know, so AMD actually wanted to put Aster, because they are all about UA link, right? Like the open standard of hooking stuff up. Uh AMD actually wanted to put a UA link switch, but there was no UA link switch in the market. So what Broadcom did was they were like, okay, this is the only thing that was available was a Broadcom Tomahawk switch, outside of like Nvidia's NV switch or whatever. So there's no option. AMD had to go with Broadcom. And so what Broadcom did, like this is what, this is ruthless business, right? What Broadcom did was they immediately exited the UA Link consortium and they're like, Ethernet is it, you know, they had like scale up Ethernet, SUE, and then they merged it with like OCP and called it eSun. And they were like, that's it, you know, we're not doing UA Link. We are in. If you are designed in with us, we're going to go Ethernet all the way. It's basically puts you know the nail in the coffee for all these other people who are trying to do UA Link. And they're like, no, we have the only switch in the market, and we're not doing UA Link. Screw that. Like we're going, you know, Ethernet. That's it. No more. And remember, the training thing with the Scorpio switch right now is uh is a PCIe switch. It's not Ethernet.

unknown 43:39

Right?

SPEAKER_01 43:40

PCIe could be used for all this as well, by the way. And uh Amazon Trainum is running on PCIe. But the future of Scorpio switches actually has UA Link in there. And it's probably a 2027, I think, story. Um, along with Marvel, who's also building something like this for UA Link. Now the the risk here is like, will AMD and the others uh go uh go away from using Ethernet, eSAN, and go jump on UA Link and buy all these people's uh switches? Or are they going to say, no, forget it, Ethernet is it? We have been designed in. Remember Astera Labs' own story, right? When they got the Aries slot on the Hopper series, Broadcom couldn't nudge them out. Now what has happened? Broadcom took the AMD slot uh because there was nothing else in the market, and now they've cornered the Ethernet way and exited the UA Link Consortium. And now can they get back in? That's very interesting to see.

SPEAKER_00 44:40

Yeah, well, we yeah, we should do a deep dive on this because I know that with the Helios platform, AMD is doing UALOE. UAL, UA Link over Ethernet, and so they're using like the UA link protocol, but they're doing it over an Ethernet switch. And so I do think they're planning to take a step in that next gen of Helios to moving to UA Link. Yeah. But to your point, Astera Labs in the market didn't have the UA link switch ready in time, so this is like an intermediate step. So I think that you know Helios 500 and beyond will be interesting to help tell us what direction things go.

SPEAKER_01 45:18

Yeah, for sure. UA Link is supposed to be the more purer networking uh you know, you can, you know, GPUs again, you can access each other's memory and all that nicely with the UA Link protocol. It's built for this kind of stuff. It's very you know low overhead, very lightweight protocol. Yes. Uh Ethernet uh and eSun and all these are like kind of uh derived from the previous era uh of Ethernet and is not optimized for what AI needs today. UA Link is more so. So we'll we'll have to talk about these standards another day because that's like a whole different subject. But I just wanted to like mention like the two other things that it's only a mention. I think Astera Labs has um uh their Taurus product line, which is essentially uh basically the same signal conditioning product, but you put it inside cable and you make an active electrical cable out of it, like AECs. So they have like these chips that they can put in AEC cables. This is not like a high margin business, and they are very clear that look, our gross margins may go down because this is not a high margin business, but we want to do it anyway. Like if there's a way we can use our existing technology, we will, right? So they're like, we'll do it. And then their newest line, I think, um, is the Leo uh product line, which is basically a CXL controller, which is used to maybe pool memory and do that kind of stuff, right? So we it's still that's still new. Um, so but there is some promise. There are apparently some design wins already. So we'd have to keep a tab on that. But uh CXL is a new topic, a new episode for another day.

SPEAKER_00 46:59

Totally, totally. CXL is old, but it's new, and it's coming back maybe. We'll talk about it another time. But with that, listeners, we hope that you learned a lot about read timers. We hope that you learned a lot about Aster Labs and you had fun with it like we did. So thanks for listening. Check us out on YouTube, leave us comments, share this with your friends, check us out on Spotify, Apple Podcasts, wherever you listen. And of course, check us out on X. We're trying to post this, of course, on X as well as our daily takes. If you like our daily takes, check out our newsletter. Just go to semidope.com and you'll find it under the daily tab. And then, yes, also follow us on X. We're trying to share some clips there as well so that if you missed some old stuff, you'll get to see it again. Um, so thank you for your support and until next time.