Credo bought Hyperloom, a microLED interconnect startup. Then at OCP this week, Credo's first active light cable turned out to use micro pixels. Why buy one technology and pick the other for your first product? The bigger question is how we connect AI systems. Do we keep pushing each lane of light faster or spread the data across hundreds of slow ones? And if we spread it out, LEDs or pixels? Austin and I are going to explain both and who's building them. Let's get into it.
Hello listeners and welcome to another episode of Semi-Doped. I'm Austin Lyons of Chipstrat. I write a newsletter for institutional investors and tech executives and I also consult with semiconductor companies.
Hey, I'm Vik Sekar and I write Vik's newsletter on Substack and I'm the founder of Semiexponent, a botic research firm on semis. How's it going, Austin?
Hey man, it's going well. Just got back from OCP. Uh, where are you at?
I am still in San Jose in a friend's garage. I actually fly out to Japan today, purely out of, you know, holiday reasons. This is not for semiconductors. I'm not hunting Japanese company alpha. I'm not doing any of that stuff. We have great people already working on that. I'm literally going to go check out Japan for a week with my family on the way back because it's kind of on my way back to India. And so we're like recording this before I hop on the flight. But yeah, this is post-OCP. It was an insane show. There was so many, there were so many people that they had to like close down the registration and people were buying entrance tickets on the black market and the halls were overcrowded and the exhibition halls were full. Very exciting show.
Yes. Yes. It was something. Um, and I had a ton of fun talking to people on the show floor, talking to people outside. A lot of people ran into both you and I since we were there together and said, "Hey, we listen to the pod." So, thank you people who listen. It was fun to see you. Um, I honestly didn't sit in that many talks. I sat in some, but there's just so many people to talk to, so much to see. And also, the talks were being put on YouTube, so I thought, "Yeah, I can watch them later when I get home at 2x." You know, if you go to the Open Compute Project YouTube website, you'll see all the talks there, totally free to watch. I think it's amazing that they did that pretty much on the same day. It avoided a lot of FOMO for the people who were there. And people like us decided let's talk to more people. Let's visit the show floor because they have wheeled in so much equipment and if nobody sees that it's kind of a tragedy. And so we spent a lot of time on the show floor talking to people and that was so much fun. Now I have all the YouTube videos as there is everybody else and so we can go and watch it at our leisure at our correct spacing whatever, right?
Totally, totally. Which I do have some time blocked next week to catch up on them because there were so many good talks. OCP did a good job, couldn't get through them all, but one big talk in the topic for today, we're going to do a micro explainer on micro emitters, so microLEDs or micro pixels or both, lots of micro to talk through. So that was really the talk of the show. Bill Brennan, CEO of Credo, had a keynote very early on. I think it was on the very first day and it was surprising. So let's just dive into, maybe let's dive into the Credo news but then we'll zoom out and talk about micro emitters, microLEDs, micro pixels and everything. But yeah, do you want to give us the news first and then we can zoom out?
The whole stir-up happened on the conference show floor where everybody's like, "Oh, did you catch that talk by Bill Brennan?" And I was like, "Oh, wait. I was watching the stuff over here." But luckily, you know, the YouTube videos helped. I watched it immediately as soon as it was posted. So the whole thing was, the background is that for the longest time we've been thinking that microLEDs are going to be the way that cables are made with light because copper is kind of reaching its limits in what it can do in terms of reach at a given speed. Optics has its like pros and cons. It's expensive. It's power hungry. All of that stuff. So there's always this middle ground where it's a battleground for what kind of technology has to go in and for the longest time a lot of people were working on microLEDs and these LEDs are pretty much like a light bulb in the ceiling but you know it's pretty much a small condensed version of it and there are many companies working on this. Avisa, Credo acquired a company called Hyperloom for 92 million with this technology. Microsoft was working with MediaTek I believe on what they call Project Mosaic and Marvell was working with this company called Mojo Vision to develop microLED. So all of these companies are like artificial reality, VR, AR, VR companies who had these microLED technologies already in production and they were like okay let's pivot to data centers because there's so much money to be made. So there was all this talk but then what happened at the talk at OCP is that everybody thought this is going to be microLEDs or at least I did, maybe everybody didn't, but he got on stage and said yeah we were doing microLEDs but we decided we'll just do micro pixels for the sake of reach and this sparked in everybody's mind the feeling that okay micro pixels as a technology is the future and microLEDs are kind of no more. That is the sentiment on the show floor but I think there's a lot more than that. I don't know how much of it we'll be able to get into. It's really a long discussion. I plan to write an article on that as well. But yeah, let's at least cover the basics of what these things are. Why are we talking about these micro emitters? What are the differences between microLEDs and micro pixels and how to think of this going forward.
Yes. Yes. Sounds good. I agree. Also, everyone was basically like, "What? Are microLEDs dead?" You know, that's like when we saw people that came right from that like I guess it's micro pixels now. But, um, you know, maybe also focus on Credo really quick and then we'll get into that specific technologies. You know, anyone who's been following Credo, obviously there's this whole conversation over the past year of, you know, is it electrical, is it optics? And Credo was the first to come out with the AEC, the active electrical cable, which was as we continue to talk at higher and higher speeds, the reach of copper shrinks and Credo said, "Hey, if we essentially put some digital processing on both ends and throw a little bit of power in there, we can extend the reach of copper." And but still there's, as you look forward on the roadmap, the question is, you know, well, how long will AECs last? And there's long range communication over optics using very powerful lasers for example but this kind of like medium reach there's a question in like the portfolio gap of passive copper, AECs, maybe something else, ALCs, active light cables, and then you know your traditional longer reach optics and in this sort of intermediate area is where people were expecting Credo to come out and have these ALCs driven by microLEDs, but then yes, the talk was about like, hey, we learned a lot from microLEDs and actually micro pixels are a good fit here. So that's again the framing, but now let's get into it. So maybe start us with microLEDs. And then of course I want to go into micro pixels because people have probably heard a little bit of microLEDs over the past, but not as much about micro pixels, although you know we have had a conversation with Al Yun of Pico, which was helpful and people can listen to that as well. But yeah, let's dive into microLEDs.
Sounds good. I want to set the basis for what these are called micro emitters. And essentially, like you mentioned, there's a reach gap, but I just wanted to mention one other thing about why this reach gap is important. Now, because obviously lasers, indium phosphide based lasers can go a long way. And I want to mention why this material matters because indium phosphide is in short supply and everybody like you know is scrambling to make this happen because optics is becoming a big thing in AI. But yeah, indium phosphide optics can go you know thousands of kilometers if you have to, you know, but we don't need all of that for the data center and the point is that why do you use a technology that is in short supply, capable of a lot more when you don't need it? Can we use the right technology for the right application? Copper beyond 7 m at you know 224 gig is about you know like already pushing it. PAM6, which is a new modulation scheme, is being considered to extend the reach of copper so the copper work is going on on one side but it's becoming very difficult. Optics is like an overkill technology with indium phosphide so there is a reach of between say 5 m and 50 m where there's a lot of applications you could do other things and this is important because now it's becoming way more important to expand the number of GPUs in the scale-up domain and it is becoming too hard to put too many GPUs in a rack. So even if you know we were at the keynote from Nvidia at OCP and Ian Buck showed the picture of like eight NVL racks connected together in an NVL 576 and of course that picture showed that it was co-packaged optics hooking them all up together. So that's Nvidia's solution but everybody need not have to use co-packaged optics or near-package optics. You know instead there is a way to connect these things with other forms of light right and so 50 m is enough to connect a lot of this stuff and the question is what is that technology that brings us to micro emitters so this is what the industry has been working on.
And I'd say that although micro pixels, VCSELs, I don't want to differentiate too much but for the sake of the micro aspect of it we'll just call it micro pixels, have been around a very long time, right? Like VCSELs have been used in data centers, that's one thing to remember about it and we'll cover more about this, but microLEDs, and we should enter into talking about microLEDs now, is that it is a far newer technology for interconnects. It has not been deployed in data communication applications like long, you know, LEDs have been in displays and all of that stuff, but for communications it hasn't been a mainstay whereas VCSELs have been a mainstay of communication systems for a long time before we went to single mode optics, right? But you know, VCSELs have been around. They have been deployed. They do work at scale with limitations, but they work.
This episode of Semi-Doped is presented by Crusoe. They build the power, the data centers, and the cloud behind AI. Did you know Crusoe can help you tune an open-source model on your own data for just a few bucks? Frontier models are very good, but they don't know your company's terminology, your part numbers, or how you like your reports formatted. So you keep stuffing instructions and context into the prompt of an expensive frontier model to get the job done. But for simple repeatable work like reports, an open-source model trained on your data can actually match a much bigger model. Think of a new grad with three months of on-the-job training. You don't always need the PhD. And the open model is cheaper and faster to run. Fine-tuning used to mean renting GPUs, working on orchestration and Kubernetes, writing training code, and babysitting the entire setup process. But Crusoe's serverless fine-tuning removes all of that. Upload your data sets, pick an open model like GLM, Llama, or DeepSeek, and they handle the training. You pay per million tokens used, not per GPU hour. And on the smaller models like Qwen 3 8B, $2 can take you a long way with millions of training tokens available. Once the training is done, you can then deploy them on Crusoe Cloud in just one click or download the weights and run them anywhere. Your data only fine-tunes your model. And it doesn't have to be work data either. You can try it on how you write emails. Pick a task, compare the tuned model to your current prompt, and see how it does. Crusoe is giving all new accounts $5 in credit to get started today. That's enough to start tinkering with fine-tuning right now. Links in the description.
So, microLED. So, let's talk about this new incumbent technology, the new technology that wants to, you know, displace optics, like you know.
So, essentially what it is is that it's literally a diode. Okay, it's if anybody who's done like a basic electronics course, it's just like a P-type semiconductor with an N-type semiconductor. It is made out of gallium nitride and it emits light approximately in the 420 nanometer range which is actually blue light and you can see it in these interconnects that like you know if you look at the place where the micro LEDs are you can see blue light being emitted. So it is blue LEDs that are doing this. You have red LEDs, you have green LEDs. Anyway, so a lot of the work around microLEDs has been at 420 nanometers. And the thing with microLED light itself is that it emits light based on what is called spontaneous emission which means the moment you apply a little bit of voltage to this diode, it turns on. I mean like it doesn't have like a threshold voltage like the minimum voltage after which it turns on. It's like a continuously varying thing. It's like the dimmer switch in your wall, right? The moment you start raising the dimmer switch, you know, you don't have to like push it 75% of the way up for it to even turn on. Okay, this is an important analogy to remember because we need to use the same framing when we get to talking about pixels. But LEDs just turn on as you increase the light. You know this just like you're used to at home. So what ends up happening is that this is a different form of light emission called spontaneous emission where it's just like sending out photons randomly and they're not very correlated to each other. This photon decides to leave at this moment in time. The next photon decides to leave whenever it wants in time. They're not like coordinated. They don't march together. They don't do anything like that. Just leave. And then the other thing is this just like an LED in your home, right? This light, this very very broad angle light, you know it's not like a pencil beam of light that you can use as a laser pointer, you know that is a pencil beam light but LEDs are like dispersive, they give a wide angle, okay. And so the third feature or bug, I don't know, of microLEDs is that they don't have a very good line width so if you look at it on a spectrum around 420 nanometers they're a pretty broad spectrum and so this is important because as this broad spectrum light goes along an optical fiber it faces what is called chromatic dispersion. And I know this is a lot of terms but we have to cover these things.
Yeah. Yeah. Let me jump in here. So if it's supposed to be centered around 420 nanometers when you say there's like a broad dispersion, you're saying some come out at 416 nanometer wavelength and some come out at 422 nanometer wavelength. So that's what you're talking about here. So the light coming out has different wavelengths.
Yeah, that's what line width is. So ideally you want, if you draw the wavelengths of a light source, you know, across wavelengths, you want it to be like a single line. You don't want like half a light bulb shaped dispersion of like wavelengths. You can't have everything from 400 to 440 or something like that which is approximately the bandwidth of this light source. It's about 10% I would say. So you have anything from 400 to 440. So this is what results in chromatic dispersion. And let's get into quickly what that means. Chromatic means you know the color, right? Dispersion means kind of how they scatter or something, right? Dispersed. So what happens is if you send 400 nanometers down an optical fiber, it takes I don't know let's say 1 second. Okay. Now 440 nanometers should also take 1 second so that they all arrive together, right? All the wavelengths should arrive together. Only then will you know it look like the same thing has arrived at the other end. But it doesn't. You know different wavelengths travel different speeds. So what happens is like all these different wavelengths arrive at different times. And when that happens, you know, when you send a pulse, that pulse has basically spread out in time, which means that you can't send pulses fast enough because now they start to overlap, right? So when you can't send pulses fast enough because they start to overlap at the other end, that means you're data rate limited, right? You can't send data fast enough. You can't keep sending bits because they're going to start overlapping. So you can only send bits slow enough so that you know you can still account for overlap at the end of the cable.
Gotcha. Gotcha. So if you have a very precise line width, it's a lot easier to make sense of the signal once it propagates down the waveguide or whatever. But if your wavelengths are spread out, then they kind of smear and maybe overlap and by the time it gets to the end of the waveguide, it's harder to tell was this a one or a zero. And to your point, then you can't send signals quite as close together, which ultimately means you can't send information as fast enough, right? So I'm assuming that you're leading us to say that the problem with LEDs is that they can't be used fast enough. Maybe that's like the problem for data centers.
Uh, yes. So it's a combination of speed and reach. So you see, chromatic dispersion limits how much spreading happens. The further you send the signal, the more the spreading. So if you want to limit spreading, one way to limit it is that you use it only for like I don't know 5 m for example, right? And then the spreading is limited. So you can send a reasonably high data rate. Now if you're going to keep it like for 100 meters, it's going to spread so much that your data rate, you can't send data fast enough. They're going to overlap. Right?
Is that spreading? Let me ask you really quick. Is the spreading a function of distance? Because the photons are just bumping into like the waveguide or something like that? Why does it... Yeah.
It'll spread more the longer you go because think about, you know, like if two cars are racing, right? Like a drag race for example. Two cars like arrive, they go a short distance in a drag race but the cars like are literally like I don't know microseconds or milliseconds of each other in a drag race, right? But think about like a cross country race that happens for 24-hour cross country or something. Yeah, yeah. You have the cars coming at much more different time intervals, you know, the leader might come in first, the next car might come in like 10 seconds later, 1 minute later or things like that. Got it. Got it. It's the same thing. Or you could think of this as a person running on a track like the 100 meter dash has like extremely close timings, but you know the 400 meter will have different spacings of people and the marathon will have even more spacing.
Yes. Yep. Got it. Thank you. Carry on.
The other thing is that microLEDs fundamentally you can't turn them on and off fast enough to send data fast enough. Okay, this is a fundamental problem of the device and right now we can only send about I don't know maybe 3 to 5 Gbit per second on a cable of like I don't know 5 m reach, right? I mean the LEDs can only like turn on and off so quickly that you can send like 5 Gbit per second. So it's limited in both directions, right? The LEDs can't go faster inherently and they don't have reach. So you know both ways.
