Transcription
Great to be in Paris, right? It's not too bad. Um, I, uh, we, well, we can, you can do a quick intro, but, uh, I head the, uh, tech equity business at, at BlackRock. As, uh, um, full disclosure, I'm, I'm a major shareholder of Lumentum. But, uh, Michael, why don't you, uh, intro?
One of the most influential shareholders in the world, Tony. You get interviewed yourself, right? That's how, that's how important you are. Uh, I'm the president and CEO of Lumentum, an optical company. We make optical components that power a lot of the communications now that are happening in data centers. So, uh, all right.
Thanks for having me. Yeah, I guess the topic of the is light and stuff. But maybe, you know, I was recently at your, um, your headquarters and, you know, with your, uh, chief product and chief technology people, and, um, I thought the framing of kind of like the future of the data center, right? So, um, I thought this four-stage framework was, uh, one of the best framings of kind of the next generation of where these AI data centers are going. So this, this idea of scale across, scale out, scale up, and then scale in. Okay. And, and at the end of the day, I don't know if you want to describe the journey of the, the data center in the era of the internet to the data center of the cloud, and now the data center of the AI, and the kind of the future AI data center. But the convening, the convening theme is, you know, we, we're going to go from copper to optics, and then we're going to go, the, the data center is going to fundamentally change because of this scale across, scale out, scale up, scale in dynamic. And I thought maybe just explain what, why Lumentum and what, why the, how the data center is changing from copper to optics.
That's a lot of scaling to me, a lot of scale, right? So if you, if you think about it, what makes a data center work? You have speed and you have bandwidth. And as bandwidths go up, speeds increase inside the data center, you have something called the copper wall. Copper can only move signals, move data so far. As speeds go up, as speeds go up, the reach of copper decreases. And so if you look inside a data center, there's miles and miles of copper. There's copper that connects one server rack to another. There's miles of copper in the back of the rack. There's copper everywhere. And copper isn't going to go away tomorrow, right? I think Tony knows that. He invests in a lot of different companies that are Ethernet companies that that power the copper. But as you're getting into these increasing speeds, these increasing bandwidths, what's ultimately happening is optics is necessary to transmit and communicate between things.
So if we can take a second on scale across, scale across is really the historic optical business. I mean, Tony's been around it for a long time. Scale across is what the internet backbone is built on. And in this particular instance, scale across means I'm now communicating signals from one data center to another, from one building to another. These inferencing models are so big, so large that normally they can't be contained within one data center, and you have to have communication that goes from one to another. That's so-called scale across.
And that's in kilometers.
That's in kilometers. Yeah, that's a decent amount of distance. What's new about scale across? You, you've talked about it a lot. There's been data center interconnect, which is just low latency connectivity between data centers. Now, with these inferencing models, you have to run at full bandwidth, full rate, and have a wide bandwidth connection over kilometers.
To kind of keep the inferencing models running. So when we go from scale across, and then for a long time, then it was scale, um, scale out.
Scale out, right?
So, why don't you describe that?
Yeah. So scale out, if you, again, you think about a data center architecture, you have these racks. There's, you know, rows and rows of them. We have a picture here behind us. But then there are switches, right? Companies like Broadcom, Nvidia make these switches. And to communicate from these server racks, these racks and racks of compute out to the switch banks, so-called scale out, you are usually traveling over a relatively large distance. And as such, you need optics. And these are these optical transceivers, these little pluggable devices that look like almost like a USB thumb drive that you plug into the back of these servers. And that creates this scale out. I'm trying to move data from these compute banks out to switch banks.
And those switch banks are meters, not kilometers. And those banks, switch banks are usually tens.
And, and, and when you go from scale across to scale out,
This distance is meters, not kilometers. But the bandwidth and the power and the market size is, is that like one, 10 times, one order of magnitude bigger?
Yeah, that's probably one order of magnitude bigger than scale across. Scale across, right? So, you know, there's, there's lots of these switch banks, as it turns out, in a data center. And, you know, that very well. And so that connectivity now is probably 10x more optical lanes than
10x more lanes for scale out
Than we have.
And then now these LLMs and the AI need scale up. So scale up is, how far distance are we talking about? So scale up now is really inside the rack. Inside these compute racks, you have a huge, if you ever looked at an Nvidia rack, there's miles and miles of copper, right? It goes from one, one top to bottom inside these racks. At, at 1.6, 6 terabits per second, which is sort of the newest speed, the newest data rate, you are not able to get copper to work over distances of more than a meter. So now you have a big move inside the rack itself to replace these copper links with optical. And that's what we call scale up.
And where are we in the scale up transition? And how much bigger is scale up versus scale out?
So we're just beginning. I mean, no one has really deployed scale up. Uh, we would expect the first deployments of scale up optics to be in the second half of 2027. We'd expect to see racks inside data centers with scale up optics probably mid-2027. And that's actually an interesting use case. Again, you understand it well. It's now connecting from rack to rack in a cluster. It's not so much in the backplane. We'd expect to see the backplane connectivity start to happen in 2028. Uh, but from a magnitude perspective, Tony, when we said 110, this is probably another decade, 100x
Deployment of optical lanes relative to the first scale across.
And, and why you, maybe, you know, so this, this is now 100x the market, okay? It's starting next year. And because it's necessity, because these models are getting so big and the data movement. Um, why can't copper do it?
Well, copper has resistance, right? The metal has a resistance. And that metal obviously gets hot. But when you're talking about these kinds of speeds, you've covered retimer companies like Credo, Astera Labs, some of the best companies in the world make these retimers. But you'd have to have a retimer almost every few centimeters to overcome the resistance that copper has. Light doesn't have resistance. We, we are able to transmit light through free space, through fiber, through Corning fiber and others. And that's the big advantage of, of optics is simply there is no resistance. So it can go infinitely far. Of course, light has loss. You have repeaters, the scale across has a lot of the repeaters, but it has far less loss than does electrical signaling.
Now, um, you know, in terms of still focused on scale up, which is like the next big market, there are like, in the semiconductor world and the, and the compute world and the memory world, there's all these shortages. Just not enough memory, not enough compute, not enough wafers, not enough anything. Can you explain kind of the, you know, within the components of an optical system, the lasers? Can you describe the challenges you're facing? Obviously, you know, Nvidia is making investments in, in this space. Where does the supply chain sit in, in, in the optical supply chain, and especially lasers?
Yeah, look, one of the key components in these optical systems are lasers. They're basically semiconductor devices that throw off light. That's how the whole thing starts. Uh, this is an industry, and we started talking about at the beginning of the conversation, that's used to telecom customers, AT&T, Verizon. And the numbers of lasers that those kind of customers would deploy are in the, you know, hundreds. Right now, we're talking about hundreds of millions. So for us to get our fabrication facilities up the curve, to go from thousands of of of wafers to millions of wafers, is no small feat. You know, you've, you've covered us for a long time. We've now put online five different wafer fabs. There's a different material. This is not
CMOS.
What's the material?
It's called indium phosphide, right? A very complicated name, but it's basically something that can emit light. It's a property that, uh, a material that can emit light. And we have five indium phosphide fabs that we're trying to ramp to scale. We see a huge shortage. And that's one of the reasons I think Nvidia invested in Lumentum and invested in one of our largest competitors, who also has incredible indium phosphide manufacturing capability. Between the two of us, I don't think we can service the demand that Nvidia and others are now putting on us to solve this resistance problem in the data center. And so the, the shortage of indium phosphide, I think, will become even more acute than what we see from the memory guys.
Just maybe, just to dimensionalize it, uh, stay on the lasers for a second. If, if you have a Grace Blackwell rack and a bunch of Nvidia chips, how many lasers do you need if we go to scale up networks with full co-acked optics and things? What kind of the, how much density of, how many lasers do we need?
Well, a good example, Tony, is, is right now for scale up optics, the idea from most of the co companies is to put this light source, an external sort of pluggable light source. In that light source are eight distinct lasers.
Eight lasers per light source.
Eight lasers per light source. And per, um, per rack, you probably have a thousand.
Thousand.
Of these in each different rack. So you're talking about 8,000 lasers per rack. And there's so many different racks. I mean, you, you know the numbers better than I do. The opportunity is immense.
This is why we need hundreds of millions of lasers.
There literally hundreds of millions of lasers. Yeah.
And, and then, um, before we get to scale in, these lasers are low power, small lasers. But if we go to data centers in space, okay? You need a different kind of laser, but it's also a laser because they're connecting the, the data centers in free space optics. What kind of laser is that?
So, I mean, interestingly enough, if you think about scale out, scale out, which is, as you correctly said, the, the first place that lasers were adopted inside the data center, those lasers are actually relatively modestly powered. And figure of merit is 100 milliwatt. When we talk about transmitting over kilometers on scale across, those lasers are hundreds of milliwatts each. And it turns out scale up is built on those same lasers. So it's 100 milliwatt, 400 milliwatt lasers are quite typical. We think that 400 milliwatt lasers that we use for scale up can also be used in space. Right? You need some amplification and things after the laser, but generally speaking, because of the dispersion properties of light and things like that, you can actually get these same 400 milliwatt lasers to work in space. I mean, it's the same lasers that we use right now. If you think about fiber optics, you have to run fiber optics across the ocean.
Yeah.
Which is a big problem, right? And so, running fiber optics in the Mariana Trench is no easy feat. We can do that with these same 400 milliwatt lasers with some repeating, right? With some reamplification, but we can basically accomplish that. And so the same holds true for space.
Are you part of, uh, SpaceX's constellation?
Well, I, I, I, I can't tell you, Tony. Okay. But,
Uh, you know, we, we definitely have lasers deployed in space. Yeah.
All right. Last one. The last new market, scale in. Just, uh, if you could, again, that's now millimeters on chip. So kilometers, meters, centimeters, millimeters. How much more bandwidth? How much market size to move data millimeters with fiber?
Yeah, I mean, this is, you know, literally a a thousandx. So we go 1, 10, 100, thousand. So it's a massive, massive market. And this is exactly what you're describing. So on a printed circuit board, on a PCB itself, you have the need now for the same reasons, the resistance of copper on the printed circuit board, the distance over which high bandwidth data needs to travel. You have a need to replace the copper traces that sit on a printed circuit board with optics. And so what we've seen from our customers, again, it's not next year, it's not the year after. It's probably 2029 before we start to see this. But even in the high bandwidth memory, uh, where you've covered one of the memory bottlenecks right now, one of the problems is you have memory stacked in stack die configuration with the compute with an Nvidia GPU or a TPU, what have you. And people are looking to disaggregate that for cost reasons. And when you do that, now the bandwidth right between the memory and the compute block is so high that you want to use optical interconnect rather than copper.
So you think that's by end of the decade that we'll start to see that?
I think it's the end of the decade. Yeah, we'll start to see that. And it's, it's massive numbers. I mean, it's a different kind of laser. You don't need all the power. So, we've been talking about scale up being 400 milliwatt lasers, really high power to throw, yeah.
Up and down the rack, data up and down the rack. This is very short distances. As you correctly said in your question, it's centimeters, right? And in centimeters, you can use lower power lasers. You just need a lot of them. You need an array of lasers to be able to handle that data rate. And if you can put together the array in the right way, you should be able to to really drive a lot of, uh, data over an optical link.