# Optica Executive Forum 2026 Complete CTO Panel, Recorded March 16th 2026

https://www.youtube.com/watch?v=5Tslub-Kmxo

[00:00] Highlights from the Optica Executive Forum at OFC.
[00:02] Thanks to these sponsors.
[00:08] But the data center's ging deep inside.
[00:11] Copper running hot.
[00:11] Hear the signals getting f and cold package light is how we win.
[00:20] We're releasing seven sessions from Optica's executive forum because the world is watching.
[00:25] Feedback encouraged.
[00:29] Now enjoy.
[00:33] Well, we continue with the program and the next panel is about CTOs.
[00:35] We brought the CTO to the some of the top companies in the ecosystem.
[00:39] We're going to have a fireside chat.
[00:42] The lights are going to come up and all of you are going to ask questions.
[00:45] We try to find the best possible moderator for this.
[00:47] Ladies and gentlemen, Anthony Yu.
[00:59] Okay, so full disclosure, I'm not Lisa
[01:01] Huff.
[01:05] Lisa Lisa had a travel delay um back on the east coast and she'll be here in time for networking, but when she asked me yesterday to step in and moderate the panel that she carefully curated and organized, I leapt at the opportunity because this is a five-star panel and it's a privilege to be to share them with you because you'll see that much of the direction of the industry that's going to be headed is actually guided by the people that we're going to bring on stage.
[01:30] So, what I'd like to do is have each one of them come up first, talk a little bit about themselves and what they see.
[01:36] We'll then follow through with a bit of moderation on my part with some questions and then I'd like to make room for audience questions also.
[01:45] So, to begin with, we'll just go in order if Greg, it's hard to see in the dark.
[01:48] If Greg could come up, please.
[01:54] So, uh, good afternoon everybody.
[01:56] Uh, I'm not going to talk about scale across.
[01:57] If I wanted to describe, you know, what's new at GF, uh, it's really
[02:02] comes down to one word and its momentum.
[02:05] Uh, this time last year, we had just announced about a $500 million investment to establish an advanced packaging line for silicon photonics in our upstate uh, New York facility.
[02:18] The intent, of course, was securing the supply chain so that literally a bare silicon wafer comes in one end and a known good optical module comes out the other end.
[02:26] That was a national priority for some federal funding that we got and a core uh priority for us to make sure we took complete control of the supply chain by last year.
[02:35] Actually, in the last 12 months, we've announced four M&As.
[02:40] To pick up on the uh comment from the last one, two of them were in the silicon photonic space.
[02:44] Uh in the late summer, we announced the acquisition of AMF, a 200 millimeter foundry that is uh in Singapore.
[02:55] For us that was an important geographic uh step from our supply chain for silicon photonix.
[03:00] Uh the fact that it also picked up a number of new customers uh important in the
[03:04] coherent space uh was really building out our uh our capabilities as a company.
[03:09] And then by December we had announced the acquisition of Infinilink.
[03:14] which may not be the most obvious uh acquisition for a manufacturing services company like Global Foundaries.
[03:19] But what we had recognized is in addition to securing our advanced packaging supply chain on the front end there were certainly customers that were looking for uh qualified IPs, reference designs and even design services.
[03:34] So uh a big part of that acquisition was meeting our customers where they needed us to meet them based on the capabilities that they had.
[03:44] Even our customers that had experienced photonix design teams didn't have enough arms and legs to do the number of products they weren't.
[03:50] So for us that was an important acquisition as well.
[03:54] And I think you'll cons continue to see us doubling down our investments.
[03:59] To date, we're well in uh north of a billion dollars of R&D investment in silicon photonics as a platform over the
[04:06] last 10 years.
[04:08] I think you'll continue to see announcements in that space.
[04:11] And then the the last comment I would make because there were quite a few discussions on MSAs today that it's great to see both the CPX and the OCI announcements.
[04:20] This is uh you know I think I've been in panels in optica forums for the last several years talking about the role of of standards and we know that whether it's the fiber attached process or the architecture of the solutions for advanced packaging it's been the wild west so I you know we embrace the fact that we are seeing that uh convergence come the more important message and you just heard it on the previous panel multissource supply says that I think the ecosystem of making sure that our customers also are going to likely be working with our competitors and by having more and more standardization I think we're going to be in a very good position uh to be able to uh to meet the expectations that our customers have.
[05:02] So um for us we're an open ecosystem.
[05:04] If that means having an EIC from another foundry that we
[05:06] integrate into a package or our pick goes into somebody else's solution uh we're completely open to that.
[05:11] We believe that the industry for the ramps that are coming measured in millions of units per month as CPO takes off in the next several years.
[05:18] It's going to be incumbent upon those supply chains to have an open ecosystem.
[05:22] With that, hand it back to Anthony.
[05:24] Over to Julie.
[05:27] Hi, good afternoon.
[05:29] I for any of you I don't know, I'm Julie Yang, CTO of Coherent.
[05:31] And first and foremost, I just want to say I'm excited to be here at OFC 2026.
[05:35] I've been in this industry over 30 years, and I kind of can't think of a time that I've been more uh excited about the the importance and also the impact of the of the work that we're doing.
[05:48] And um if I look looking you know at what Richard Hoe shared this morning I mean when you think about it the AI data center is now a huge supercomputer and as optics we are inside that machine and that supercomputer is actually gone parallel and so the parallelism makes
[06:07] that computer depend more on interconnect and we are the interconnect.
[06:10] So I'm very excited.
[06:12] Um and what am I excited about?
[06:14] I'm excited about I'm still excited about pluggable transceivers and uh we see new form factors coming out supporting higher bandwidth density.
[06:21] Uh in order to support that higher bandwidth density, we need new lasers and modulators, gallium arsenide, indium phosphide, silicon photonics.
[06:29] So we're working to raise the data rates of all of those technologies in order to support continued bandwidth density um improvements to scale the AI data center infrastructure.
[06:40] And I'm also excited about CPO and NPO.
[06:43] I'm very excited about those actually and and I think those you'll see also might take many different forms.
[06:50] There's a lot of focus around one particular form of CPO and NPO, but I think just like in pluggable transceivers, you may see that take many different architectural implementations.
[07:00] And uh I'm excited about optical circuit switches uh which add a lot of flexibility to the AI data center.
[07:08] I'm also, we talked a lot this morning about scale across, but I'm also excited about the scale across, the coherent transceivers um and uh the transport infrastructure, including multirail.
[07:19] So overall, I think it's a it's uh as as an optical community, I think I've never declared it before myself, but I think it finally might be our time.
[07:28] And with that, I'll hand it off to Hong.
[07:31] Good afternoon.
[07:33] I'm afraid I'm on the wrong panel.
[07:36] I'm not a CTO.
[07:38] So if I'm a CTO, I can talk about all the good things about CPO, all the stuff.
[07:41] But um uh actually ro is more like VP of engineering.
[07:43] So we need to be realistic on what things we need to build, what things we can deliver.
[07:48] So if you look at AI um data center, it has three type of network.
[07:50] We have a scale up, we have scale out and we have scale across.
[07:55] And today we use a pluggable because give us a flexibility to provision different type of optics for different stage of network.
[08:03] For the scale up we use uh the
[08:09] requirement is different than scale out.
[08:11] because the optics there need to be highly available and uh also it's easy to repair because you have expensive uh compute resource behind it.
[08:21] And given that said uh there's a lot of good things coming out of the CPO.
[08:23] First of all, a lot of VC money pulls into that.
[08:25] A lot of innovation coming out of it.
[08:29] So, we can use CPO type of integration to make make pluggable more energy efficient and more reliable.
[08:36] So, I think that's a lot of good thing we see coming out of the CPU.
[08:39] And uh for 200 gig, I think we're pretty sad with a lot of knobs we can tune to optimize the power efficiency for pluggable.
[08:50] There's a TRO.
[08:53] If you really want to be aggressive, you can go with LPO and the 400 gig is right out there.
[08:55] The 400 gig is 30.
[08:58] Um, if we can deliver the uh electrical channel to meet the same requirement as like 15 dB, we can do LPO and 25 dB we can do TRO.
[09:12] But the real question, what was going to happen after 400 gig?
[09:17] There are three possible solutions out there.
[09:19] There's a micro ed micro vixel and also there's a micro ring and they have a different uh technology advantage and also challenges.
[09:28] Micro ED I think if you compare that has a it's the most power efficient solution and the v micro vixel is most mature technology.
[09:39] Uh the biggest challenging I see from micro ring you need a provision a lot of laser because uh silicon photon cannot generate light.
[09:46] So need a lot of light come in and out.
[09:48] I think the jury is still out there to be decided what will be the ultimate best solution after 400 gig.
[09:55] So we'll see what's going to happen after M plus2 and plus five years.
[10:05] Good afternoon everybody.
[10:07] Um my name is Matt Cesack.
[10:10] I'm the I move my water over here.
[10:10] I'm the CTO of Lmentum.
[10:13] First of all, it's just absolutely fantastic to be up here today.
[10:16] I mean you have such an amazing panel of distinguished technologists.
[10:20] So great to see so many colleagues I think out in the uh audience.
[10:22] Um hopefully you know I can offer a little bit of uh perspective um in terms of the way Lum thinks about the market and the way that you know I think about technology at the company.
[10:34] Um you know I think we all know and it's become a little bit cliche but you know AI is changing everything.
[10:38] I don't think that's really a surprise.
[10:40] It's not that novel but it's more than just technology right?
[10:44] It changes the way we engage with our customers.
[10:46] It changes the way we think about our competitors.
[10:48] It changes the way we think about standards.
[10:51] It changes the way we develop our products.
[10:54] A lot of people talk about one of the impacts that AI has had is that design cycles are a lot shorter.
[10:59] Uh I talk about design cycles that are 18 months, 24 months sometimes.
[11:04] This is fundamentally significantly faster than anything we've had to do uh do in the past.
[11:08] And so, you know, in the face of
[11:14] all of this change and in the face of all of this kind of uncertainty, you know, one of the ways I like to think about uh momentum is, okay, well, what are you good at?
[11:22] What do we know? What's our core expertise?
[11:27] And how do we use this core expertise to build differentiated products in the marketplace?
[11:30] Uh, and there are two examples just want to highlight.
[11:33] You know, the first example is our optical circuit switch.
[11:38] Um, optical circuit switches have gotten a lot of attention over the last couple years.
[11:42] Um, but if you really think about it from Lmentum standpoint, the circuit switch that we built is based on MEMS.
[11:50] Um, and we've had a MEMS team in Lmentum for more than 20 years.
[11:55] So, when we were looking at, okay, this circuit switch technology is coming up and you know, like it looks really interesting.
[12:01] It looks like it could have a big impact on the data center, you know, well, what should we build it with?
[12:05] Well, let's build it with something that we know.
[12:06] Um and we've built many products out of our you know MEMS technology over the years.
[12:10] The other example uh that I bring up is
[12:15] actually our ultra high power laser.
[12:18] Um that technology is actually based on our pump lasers.
[12:20] And if you know anything about pump lasers, pump lasers need to have absolutely ridiculous and amazing reliability.
[12:27] So when it came to designing something for CPO given a lot of the perceived challenges associated with uh reliability especially with things like lasers it felt very natural to turn to uh pumps as a core technology uh for that particular uh application.
[12:44] So you know again the perspective here is you know really know what you're good at and once you know what you're good at you can really uh build differentiated products in the market.
[12:54] Okay.
[12:54] So, just to put things in perspective, the power of this group that I'm I'm on the stage with here, we have the the two leading um 35 laser companies, Lumenum and Coherent, that basically were in the news a few weeks ago where they're lock that's the that's the source of our 800 gig, 1.6, and 3.2 gig lasers.
[13:14] We have Greg Bartlett from
[13:17] from Global Foundaries investing a billion dollars.
[13:22] Global Foundaries now controls 65% of the silicon photonics um foundry market.
[13:25] And we have Hong who is typically modest but she's the chief researcher on on optical networking at Google and she's a data architect and was one of the early drivers for optical circuit switches at Google.
[13:39] So it is really a privilege to be with you all today.
[13:43] Julie Hong, Matt and Greg, thank you very much.
[13:46] So we've had so many different um wonderful panel discussions today.
[13:51] I'm going to kind of move it a little bit forward here because we have such um savvy minds here looking both at what we're doing today and into the future.
[14:00] So, we've heard comments today that said that we're at an inflection point.
[14:04] We've heard about this is the year of CPO or NO.
[14:09] If we come back here three to five years, what hurdles are in front of us, CTO's and fellows and SVPs, what is in front of us that you think we need to kind of look ahead to and start
[14:18] preparing for as we as we look at the future of optics?
[14:23] Uh why don't we start with you Matt?
[14:27] Um so you know I think it's kind of interesting if you think about CPO.
[14:30] actually CPO has been around for a while.
[14:32] Um I remember coming it's an overnight success right?
[14:34] Yeah, a bit of an overnight success, right?
[14:35] I remember coming to OFC, I think, all the way back in like 2015 and seeing, you know, things like CPO switches around and everything else, right?
[14:43] So, you know, CPO as a concept, I think it's it's been around.
[14:48] Um, I think the question is more about kind of like why now, you know, like why all of a sudden is it like over the last like 18 months or so has CPO achieved such kind of relevance?
[14:59] And I think, you know, really if you look at what's going on, it makes complete sense, right?
[15:02] I mean, you see all kind of uh announcements about how people are chasing power because they're trying to build bigger and bigger data centers, right?
[15:09] And how actually the access to power limits the amount of compute you can put inside some of these data facilities.
[15:12] So, okay, where can you save power?
[15:15] Well, networking is becoming a significant contribution to the total
[15:19] power inside of a data center.
[15:23] So, CPO, um, I think we've all looked at the, you know, the amount of power that can save.
[15:26] It's, you know, pretty significant.
[15:28] Um so that's you know one of the main drivers uh in terms of like why CPO is happening today.
[15:33] Now if you look at ecosystem readiness in terms of like are we actually ready for CPO I think a couple things.
[15:37] I think at the fetonics level and at like the laser level the component level the answer is pretty much yes.
[15:44] You know I think we know how to make lasers.
[15:46] We know how to make fetonics.
[15:48] I don't think that's so much the question.
[15:49] I think the question is more about all the other things you know.
[15:53] It's like hey to make CPO a success right?
[15:55] It's not just about making lasers.
[15:57] It's not just about making picks, right?
[15:59] It's about making connectors.
[16:01] It's about making shuffles.
[16:03] It's about, you know, how you going to do your edge coupler and is that process really manufacturable?
[16:07] So, you know, really, I think is it is it ready?
[16:09] I think it's almost there, but I do think that there's a few more things that have to cross the finish line in order for it to become a real reality.
[16:17] on uh so um internally we did um multiple
[16:24] round of CPU type of products every time.
[16:27] when you have a product you have to justify ROI.
[16:31] so one hard lesson we learned if the electronics can do 30 can do it's very hard to for optics to compete in terms of cost efficiency and also the reliability and uh for if you look at the power for CPO.
[16:45] look at just module level might be big saving but you have to look at the system level and look at the 100 megawatts class level I can give you ballpark so the power distribution for TPU uh in TPU most of power is in the computer and the memory optics and third is tiny portion if you design the architecture correctly.
[17:08] so for us it's about the um used to be two 3% now it's five to six%.
[17:12] same thing for the IPU like Nikos's compute is even smaller portion and for the networking switch uh optics and 30 is a big portion.
[17:22] It's a 30 30%.
[17:25] optics 30%.
[17:28] But the networking switch if you look at with the switch everything plug in uh we compare the power CPU and the pluggable.
[17:34] the CPU gave you power saving of 20%.
[17:36] But you really have to look at the class level 100 megawatt cluster networking small tiny portion about four or five%.
[17:48] most of power in that building 100 megawatt is consumed by uh GPUs or TPU.
[17:50] So you have this four and 5% it's worth optimizing but you have to balance uh how you going uh are you going to cause any straining in terms of computer resources and how you can really optimize end to end.
[18:10] Julie.
[18:11] yeah I think uh I think uh well it probably won't surprise any of you who know me to hear me say this but I actually think you're going to see the coexistence of pluggable transceivers and CPO NPO and that's because they solve different architectural problems.
[18:24] I mean, when you
[18:26] think about a pluggable transceiver,
[18:28] it's like very it's adds a lot of flexibility.
[18:29] I mean, Han can speak better because she's the one who runs the data centers, but adds a lot of flexibility.
[18:34] There's the multi- vendor ecosystem.
[18:36] There's the fact that it plugs through the front face plate.
[18:39] You know, you can support many distances out of the single form factor.
[18:42] So, there's lots of value.
[18:43] But then the CPO, it allows a an improvement in power consumption, albe even if it's modest on the system level.
[18:50] and um also increased bandwidth density.
[18:53] And so I think you you'll see in cases where flexibility is more important than um bandwidth density and uh power consumption, you'll see pluggable transceivers.
[19:04] And where bandwidth density and power consumption are more important than flexibility, then you'll see CPO and NPO.
[19:09] So I think you'll see them coexist.
[19:12] And I think you'll um Yeah,
[19:16] thank you, Julie.
[19:18] Greg, I'll probably agree with a lot of what's already been said.
[19:22] Um, starting maybe with Julie on pluggables and CPO.
[19:25] Pluggables has been proven to be a
[19:27] volume business, right?
[19:28] There's hundreds of thousands, millions of those going into the market.
[19:29] CPO has not yet done that.
[19:32] And, you know, to bridge on Matt's comments, whether that is the turning mirrors or microoptics or uh the harnessing for the unique form factors that many of the customers have today.
[19:46] that is a business yet to be proven can take to volume and the volumes will be substantial in the next three to five years right we're talking millions of units per month for things that uh there is no infrastructure in the industry so certainly as we look at securing uh an ecosystem and you know for example on just on fiber attach you know we have demos this week with teramont senko and corning and continue to look at all different forms of how to get light in light out uh of these platforms.
[20:18] The key is not the demonstrations of one and a half dB insertion loss.
[20:20] It's being able to do that on 40 fibers on hundreds of die millions of units per month.
[20:22] Uh you
[20:27] know, the good news that's what
[20:28] semiconductor foundaries do for a
[20:30] living. The bad news is there's still a
[20:32] lot of work to be done to uh to get the
[20:34] infrastructure in the industry uh there.
[20:36] So that's that's I think one challenge.
[20:38] The other one that I would highlight and
[20:39] it's a little bit of the same story. We
[20:42] know that 400 gig you know we've got eye
[20:44] diagrams on silicon in 400 gig we'll see
[20:48] where that goes but as you start then
[20:50] thinking about introduction of new
[20:52] materials berium titanate thin film
[20:53] lithium nabate it's the same thing again
[20:56] what is the maturity of the supply chain
[20:58] on there so I think much of this is the
[21:00] hard engineering of preparing to scale
[21:02] the business that's where a massive
[21:04] amount of our investment is going
[21:06] actually is to be ready first and
[21:07] foremost to support the wafer scale ramp
[21:10] secondly the the manufacturing
[21:12] especially in the packaging front where
[21:13] a significant amount uh of the
[21:16] complexity exists.
[21:18] >> Very good. I'm going to turn back to you
[21:20] Greg. Um just I haven't properly
[21:22] introduced myself. Jose, did you have a
[21:24] question?
[21:27] >> Hello Anna Gonal from Ironics here. I
[21:30] have a question. So after the
[21:33] announcement this afternoon of Nvidia
[21:36] saying they are going to include optics
[21:38] in the scaleup networks uh do you think
[21:42] the hyperscalers are going to follow
[21:44] this lead for the scaleup?
[21:48] >> Uh I think Nvidia announced uh no CPU
[21:53] scale out switch at the last GTC and we
[21:57] can check with them how many CPU
[21:59] switches they have deployed.
[22:07] Maybe I could add to what Hog said by
[22:10] just saying that um you know if you deal
[22:12] with all the different hyperscalers on a
[22:14] daily or weekly basis like I do then you
[22:16] realize that they all have very very
[22:18] different architectures and they make
[22:20] decisions based on different um
[22:24] constraints and different uh almost
[22:26] business models in many cases. And so I
[22:28] think each of them is almost a market
[22:30] unto themselves and just because one of
[22:32] them does something doesn't mean that
[22:33] everybody else will. That's my view.
[22:36] >> That's good. I I'm going to to turn back
[22:39] to Greg a little bit. Just a proper
[22:40] introduction for myself. I realize I'm a
[22:43] little rude. I retired from Global
[22:45] Foundaries about a year and a half ago.
[22:46] I ran the uh Silicon Potics product line
[22:48] at Global Foundaries, retiring in
[22:50] October 2024. And Greg here, he leads GF
[22:54] Labs, which has got G Phetonics at the
[22:57] center point of university and industry
[23:00] collaboration. So Greg, as you as you
[23:02] and the rest of your esteemed peers
[23:04] here, look at how to operate in this
[23:07] environment where things happen so
[23:09] quickly. I mean, you blink your eyes and
[23:11] new standards are introduced or the
[23:13] transitions from 200 to 400 to 800 are
[23:16] being compressed. How do you actually
[23:18] manage long-term research, quote
[23:20] unquote, long-term research in an
[23:22] environment where things are are
[23:23] shifting and changing so rapidly given
[23:26] by driven by AI? Have fun with that
[23:29] question.
[23:31] >> I I thought you agreed to not ask any
[23:32] inside questions.
[23:35] >> Um yeah, no, it is actually a very
[23:37] significant challenge uh for the
[23:39] corporation. The answer for us and you
[23:42] know just the release of the uh the OCI
[23:45] MC MSA was such an important step. We
[23:48] have believed in slow and wide for a
[23:50] very long time. You know DWDM designs is
[23:53] is something we demonstrated five years
[23:55] ago with 16 and 32 lambda. So we we felt
[23:58] like those were important as we're
[24:00] pleased to see that it's coming back uh
[24:02] after a tremendous amount of uh uh focus
[24:05] on single lambda on there. But the
[24:08] answer is have your wings spread pretty
[24:10] broadly. I mentioned double downing on
[24:12] the investments. We don't presume and
[24:15] just modulator work indium phosphide,
[24:17] berium titanate, thinfilm, lithium
[24:18] nabate, the polymer solutions on there.
[24:21] We can't pretend that. We won't know. Uh
[24:24] you know heterogeneous integration,
[24:25] thermmo compression bonding, uh
[24:27] dialectric bonding, hybrid bonding. You
[24:30] have to have all of those capabilities
[24:31] and a lot of the uh activity and the
[24:34] investments we're making are capability
[24:35] tooling uh to establish that. So uh it
[24:38] it is something where you have to have a
[24:41] lot of activities and a lot of tentacles
[24:43] out in the industry whether that's with
[24:44] research consortia you know we work with
[24:46] the institute of micro electronics in
[24:48] Singapore and of course we're part of
[24:49] the IMAC uh program that has a massive
[24:52] photonics uh program it is having the
[24:55] bandwidth to stay engaged that you're
[24:56] not disrupted when an emerging
[24:58] innovation uh comes to the forefront
[25:01] >> and then Julie as the SVP of of opto
[25:04] electronic devices and modules how do
[25:06] you als also take what's happening out
[25:08] there and convert it into rapid
[25:10] introduction to the market with
[25:12] reliability and cost in mind. How do you
[25:14] manage in this environment? I would say
[25:16] that the as I couldn't agree with Greg
[25:18] more in the sense that when things are
[25:19] changing so rapidly so what we've uh
[25:22] definitely our time to volume has
[25:24] compressed dramatically probably like 3
[25:27] to 5x over the last decade and so um and
[25:31] when things are changing and the new
[25:33] data rates are coming faster and faster
[25:35] and when that happens you don't have any
[25:37] choice but to basically take more
[25:39] parallel paths than we used to like at
[25:41] 10 gig we would do all our perfect
[25:43] little simulations and we'd have one
[25:44] path and that would be the path we take.
[25:46] >> Those were the good old days.
[25:48] >> Exactly. But now we have to do many
[25:50] things in parallel because many of the
[25:51] components we're using it might be the
[25:53] first thing that has come out of the fab
[25:55] and you're putting all of that together
[25:56] but you're doing that to hit a timeline
[25:58] that if you don't hit you miss the
[26:00] business opportunity. So I actually for
[26:02] about five years ran our consumer
[26:03] business actually and so that operates
[26:06] with a certain cadence and and you know
[26:08] if you miss the train by one day you're
[26:10] out and I feel like the um our business
[26:13] is becoming more and more like that with
[26:15] the the the application being AI
[26:19] actually.
[26:19] >> I see. And Matt, you lead technical
[26:22] strategy at Momentum. You've probably
[26:23] seen compressions between the 100 to
[26:25] 200, 200 to 400, 400 to 800. How do you
[26:28] adapt again along the same lines of
[26:30] these questions?
[26:33] >> Um, well, I think there's a couple
[26:35] things. I think anytime there's an
[26:36] industrywide move in terms of speeds,
[26:39] right, it's not just components that
[26:40] have to change, right? I mean, it's a
[26:42] whole ecosystem. It's components, it's
[26:44] electronics, it's not just electronics,
[26:46] it's also connectors, form factors,
[26:49] everything. So, when there's a major
[26:50] change like this, everyone kind of has
[26:53] to move uh at once. And really the way
[26:55] that we typically handle something like
[26:57] this is we rely on our partners. We
[26:59] build strong partnerships across the
[27:00] ecosystem uh to go and make sure that
[27:03] when those transition happens uh we're
[27:05] ready for it.
[27:06] >> Very good.
[27:08] >> Okay. Hong, I'm going to shift gears
[27:10] again and don't be so modest because you
[27:12] are the chief architect here. So, we've
[27:14] heard a lot of themes this morning about
[27:16] the scaling of these data centers, the
[27:18] scaling of AI, and how we need to have
[27:21] cost and performance be optimized and a
[27:24] lot of focus on cost. As an architect,
[27:27] where do you what do you what do you see
[27:29] data centers needing to do to be able to
[27:31] accommodate, you know, the the rate of
[27:34] innovation in the components relative to
[27:36] cost and and performance?
[27:38] >> First of all, the pace of doing the
[27:41] products is much much faster. We used to
[27:43] do like a 10 gig optic for like seven
[27:46] years
[27:47] >> and so one product can last seven years.
[27:50] So we only have like two optical
[27:51] engineer working on that. Now we have
[27:54] how many I don't know uh five products
[27:57] every year and then we have different
[28:00] network we have a GPU network TPU
[28:02] network we have scale scale up scale out
[28:05] and we have a training network and we
[28:07] have sampling they all have different
[28:09] requirements in terms of bandwidth
[28:11] latency all those things. So we do have
[28:14] to uh design different optics for
[28:16] different type of networks. We're trying
[28:18] to leverage the same critical components
[28:21] to make sure we have near-term we have
[28:22] something to deliver but in the meantime
[28:25] you can optimize for you know for
[28:28] efficiency or the but one thing for sure
[28:31] you have to make sure the things we
[28:33] deliver can not zero to one we can scale
[28:36] to millions rapidly in in a short period
[28:39] of time.
[28:41] And are there any any comments or
[28:43] developments that you're thinking about
[28:45] relative to optical circuit switches
[28:47] from a Google perspective? and the
[28:49] optical circuit switch I think is a it's
[28:52] a is enable enabling technology for TPU
[28:56] scale up network with the optical
[28:59] circuits we can scale to close to uh
[29:01] 10,000 TPUs in one super part when you
[29:05] have this larger super part you can put
[29:08] a lot of amount of parallelism in that
[29:10] in that part you don't have to go to the
[29:12] scale up uh scale you don't require very
[29:15] high bandwidth for the scale out and I
[29:18] don't need a CPC to a CPU to do the
[29:21] scale up because I have optical circuit
[29:23] switch. I don't need a external
[29:25] electrical switch and I don't need a
[29:27] CPU. I can opt out directly go to
[29:30] optical circuit.
[29:32] >> So Julie and Matt both of your companies
[29:34] Coherent and Lmentum have talked
[29:37] publicly about consideration of optical
[29:39] circuit switch wedges in into data
[29:41] centers. Um where does that stand? Is
[29:43] there anything you can share with the
[29:44] audience here in terms of your thinking
[29:46] and over the last couple years has
[29:48] really moved to the forefront? Um where
[29:50] does that stand with with you first
[29:53] Matt?
[29:55] >> Um well I think uh optical circuit
[29:58] switches are going to be a big thing. Um
[30:00] I think you know we publicly mentioned
[30:02] in our last earnings call that we have a
[30:04] $400 million backlog uh for optical
[30:07] circuit switches at Lmentum. So they're
[30:09] really seeing a lot of traction. Um and
[30:12] I think you know one of the more
[30:13] exciting things is uh the use cases are
[30:16] also expanding quite a bit as well right
[30:17] you know if you look at circuit switches
[30:19] it started kind of with elephant flows
[30:21] like many many years ago right and uh
[30:24] has since moved on from things like
[30:26] elephant flows to spine switch
[30:28] replacement from spine switch
[30:29] replacement to some of the use cases
[30:31] that Hong talked about um which cover
[30:34] things like uh scale up um and now
[30:37] they're even being used in things like
[30:38] network failover uh where you you know,
[30:41] if you have a rack of GPUs that
[30:43] potentially is going down, you can
[30:44] connect them up through uh a different
[30:46] set of um circuit switch ports and as
[30:49] the rack goes down, you can bring up a
[30:50] spare rack, uh to manage reliability.
[30:53] So, you know, I think we're seeing the
[30:54] very beginning of optical circuit
[30:56] switching and I would even expect as
[30:58] time goes on, uh there'll be even more
[31:00] use cases, uh which will come up, uh as
[31:03] people get more familiar with the
[31:04] technology.
[31:05] >> Yeah. Yeah, I agree completely with
[31:06] Matt. Yeah. So, uh, in production, you
[31:09] know, we're shipping, um, 64x 64
[31:12] systems. There's 30 300 300 by 300 or
[31:15] 320 by 320. We're seeing people
[31:17] interested in going to 512 x 512, so
[31:19] higher radex, and going the other
[31:21] direction, compressing more into smaller
[31:24] boxes. And so what surprised me actually
[31:27] about the optical circuit switch was
[31:29] exactly as Matt said, how the use cases
[31:32] broadened first, but also the depth of
[31:34] the c the number the sheer number of
[31:37] customers and the speed with which they
[31:39] started to adopt the technology. I think
[31:42] because they saw the value in it. So we
[31:44] announced at our own call that we've
[31:45] we've shipped to over 10 customers
[31:47] actually optical circuit switches. So I
[31:49] do think it adds it it in some cases
[31:52] again Hong can say better because she
[31:54] architects the network but what I see
[31:56] from our customers is that it's it
[31:58] becomes a new architectural tool that
[32:01] actually makes people think differently
[32:03] and come up with different solutions
[32:04] that they didn't hadn't thought of
[32:06] before they had that flexibility of the
[32:08] optical circuit switch and that's why
[32:10] we're seeing those use cases expand. So
[32:12] yeah very exciting.
[32:13] >> Very good. Before we go to the audience
[32:15] questions I did actually have a question
[32:17] submitted to me. Did you
[32:20] >> I did have a question submitted to me
[32:22] that that I guess I'll address first to
[32:24] Greg here. So we've seen that um in the
[32:26] 800 gig transceivers is a shortage of
[32:29] supply relative to demand. I think
[32:30] estimates anywhere between 40 and 60%.
[32:33] So as transceiver market just takes off
[32:37] the plug transceiver market, how are
[32:39] foundaries like global foundaries and
[32:40] packaging houses trying to keep up with
[32:42] this kind of demand and how will they
[32:44] actually close that gap both for foundry
[32:47] and also Julie and Matt for lasers.
[32:51] >> So it's not yet at fab bursting levels
[32:54] of capacity. You know in most of our
[32:56] facilities we run around a million
[32:58] wafers per year and the business isn't
[33:01] there now. We are reconfiguring our our
[33:03] Malta facility, our upstate New York
[33:05] facility, roughly doubling the capacity
[33:07] year on year, and we're going to
[33:09] continue to do that. It still does not
[33:11] represent a significant fraction of the
[33:14] uh of the facto's capacity on there. I'm
[33:16] much more worried about the downstream
[33:19] uh capabilities, especially in the
[33:21] packaging front. That's a few years away
[33:22] still on uh on CPO. Uh but for us uh it
[33:26] is you know we're making material
[33:29] capital investments uh in the wafer
[33:31] capacity uh there but uh it's manageable
[33:34] within the normal remix we do within our
[33:36] facilities.
[33:37] >> Julian man I'm going to phrase it I'm
[33:39] going to phrase the way the question was
[33:40] put in here. What about everybody other
[33:43] than Nvidia in terms of lasers?
[33:45] >> Well just generally speaking I think
[33:47] from a big picture perspective as Greg
[33:49] said I mean we announced our last earn
[33:51] our booked to bill is four. We're booked
[33:52] out for the end of this year and you
[33:54] know probably by June we expect to be
[33:56] booked out till June of next year. So it
[33:59] is uh there is a lot of business in the
[34:01] optical industry right now and and many
[34:03] of our product lines were not demand
[34:05] limited we're supply limited. So what
[34:06] are we doing about that? So we can add
[34:08] more assembly and test capability for
[34:10] things like transceivers for optical
[34:12] circuit switch but in in um in the
[34:15] transceivers we're also seeing a lot of
[34:16] constraint on the indium phosphide um
[34:18] capacity that you heard on the panel
[34:20] this morning also. And so what we're
[34:22] doing there is we're actually uh
[34:24] expanding our indium phosphide capacity
[34:26] and in particular uh we brought up 6-in
[34:29] indium phosphide in our Sherman Texas
[34:31] and Yarfala Sweden fabs uh to bring
[34:34] extra capacity on for the industry. Um
[34:40] >> uh so you know we're both a transceiver
[34:42] as well as a laser uh supplier. So maybe
[34:44] I can talk just a little bit about both
[34:46] of those for a second. Um on the
[34:48] transceiver side uh I think this
[34:50] technology is you know if you look
[34:52] historically at it it's a mix of kind of
[34:54] like hightouch and automation you know
[34:57] meaning that there are still you know a
[34:59] number of transceivers that are made by
[35:00] hands people are still kind of you know
[35:03] doing a lot of manual assembly and
[35:04] things like that and now that gets
[35:06] combined with some automation capability
[35:08] but the volumes that people are talking
[35:10] about from a transceiver standpoint you
[35:12] know really are so large right that it's
[35:14] very difficult I think to continue you
[35:16] know some of these
[35:17] uh sort of hightouch kind of processes,
[35:20] right? So at Lmentum um one of the
[35:22] things we're doing to address the
[35:24] scalability of transceivers, right, is
[35:26] expanding our manufacturing footprint
[35:27] and continuing to look at ways to build
[35:29] automation into the transceiver line. Um
[35:32] now from a laser standpoint, uh you
[35:34] know, I very much echo what Julie said.
[35:36] I think many of us are uh sold out from
[35:39] a laser standpoint. And I think the
[35:41] demand on lasers has been um very high
[35:44] uh over the last uh I would say uh many
[35:47] many months. Um and you know us as a
[35:50] ecosystem we're all working very very
[35:52] hard to scale our capacity to meet the
[35:54] demands uh of our customers. Um and for
[35:57] us that includes you know significant
[35:59] capacity expansion on the Indian
[36:01] fossified side.
[36:03] >> Very good Jose I think we're ready for
[36:05] audience questions. Uh so going back to
[36:08] CPO topic, we've been talking about CPO
[36:11] for a long time. Um advantages in power
[36:15] consumption, bandwidth density, but from
[36:18] my standpoint, the main surprise of last
[36:20] year was incredibly high reliability of
[36:23] CPO. I think Meta reported 3x
[36:27] improvement. Nvidia reported 10x
[36:30] improvement in resiliency which
[36:32] translated into I think doubling of GPU
[36:36] utilization efficiency in case of meta
[36:39] 5x improvement GPU utilization
[36:42] efficiency at Nvidia. So I'm just
[36:45] wondering Hong, you've been u an
[36:48] opponent of CPO. Were you surprised by
[36:50] this results? And
[36:52] >> are you warming up to
[36:54] >> are you warming up to CPO? No,
[36:56] >> never.
[36:58] >> First of all, I'm not surprised because
[37:00] what you compare the data reliability is
[37:02] uh uh the traditional optical pluggable
[37:06] using all the failure is coming not from
[37:08] the components itself, it's from the
[37:10] epoxy wound. But if you use CPU type of
[37:14] integration for pluggable, you will have
[37:16] same level of reliability. And the
[37:19] biggest problem of CPO is not about
[37:21] reliability. You can have 3x improvement
[37:24] and reliability. But the mtr meanantime
[37:26] to repair time is much much longer than
[37:29] pluggable. Let's say I have this quarter
[37:32] million very expensive computer resource
[37:34] behind my CPO and in case there's a
[37:37] failure, I have to swap the whole tree.
[37:41] So no failures were observed over 40 40
[37:46] million hours I think is the latest data
[37:47] from meta. So no unrepaiable.
[37:51] >> I think you need to check meta. What is
[37:52] that the plan of record for the next
[37:55] generation networking fabric? Is a CPU
[37:57] or pluggable?
[38:01] >> If if something is so good, why is a
[38:04] yeah next generation fabric? It's not
[38:07] it's not a CPU.
[38:10] >> I have a feeling I will never win this
[38:12] argument.
[38:17] >> Hi, my name is Arand. I lead the
[38:19] Stanford Fidonics Research Center and I
[38:21] also consult with with consortia like
[38:24] Foton Delta. So I live at the
[38:25] intersection between university
[38:26] innovation, startups and industry. And
[38:29] my question for you is when I look in
[38:30] this room I see a huge um concentration
[38:34] of innovators across the whole value
[38:36] chain, if you had a magic wand to make
[38:39] us all collaborate better, what would
[38:41] you fix and why?
[38:43] >> You know, I actually think we
[38:44] collaborate pretty well already. So you
[38:46] know our company we do a lot of work
[38:48] with universities and I'm engaged with
[38:51] we are engaged with almost all the
[38:53] startups and we collaborate very well
[38:55] with our customers
[38:57] um you know definitely I do think that
[39:00] um you know our R&D is very expensive
[39:03] actually um and in general our margins
[39:07] are not very high so it's a little hard
[39:09] to or not as high as perhaps we they
[39:11] should be and so you uh so in in just
[39:15] the sheer your ability to invest in the
[39:17] roadmap that you know is needed. It can
[39:20] be a challenge. You're always I mean
[39:22] it's true of any business but I think
[39:23] exceptionally true of our business where
[39:25] we're always making these decisions
[39:27] about what we can afford to work on and
[39:29] not work on. And I do think that that's
[39:31] why you see I mean some publicly but
[39:33] maybe some things that are happening
[39:35] that are not public where you actually
[39:37] see um you know capital money coming
[39:40] into the optical companies because I
[39:42] think that that people who rely on the
[39:44] optical technologies understand that. So
[39:47] I think that is a it has been a
[39:49] challenge but I actually think that you
[39:51] know it's starting to change actually.
[39:53] >> Yeah it's kind of a good question. Um I
[39:56] mean I think maybe just come back to the
[39:58] comment that I made kind of at the very
[40:00] beginning you know given the pace of
[40:01] innovation you know like it's really
[40:03] important like you you basically don't
[40:05] have time to develop a new technology
[40:06] right I mean if you're trying to do
[40:08] something completely from scratch that
[40:09] you're unfamiliar with it's very hard to
[40:11] meet the timelines you know so I think
[40:14] you know there's probably opportunity to
[40:16] go out and you know if you need a new
[40:18] technology in your suite you know like
[40:20] okay you know we wanted to go develop I
[40:23] don't know some other technology that we
[40:25] didn't have, right? I think that's a
[40:27] perfect opportunity to go and
[40:28] collaborate, you know, a perfect
[40:29] opportunity to help us build out our
[40:30] core expertise. So, I think that there
[40:32] are opportunities like that that exist
[40:34] and I think again that's really driven
[40:35] by the pace, you know, the pace of
[40:37] innovation that's being asked of us.
[40:39] >> Do you think that um the relationship
[40:41] between the the cloud service providers
[40:44] and the component providers? Do you
[40:46] think that's as good as it can get or
[40:48] what area let's push the question a
[40:50] little bit more. What area of
[40:51] improvement could happen when you work
[40:52] with the ultimate
[40:54] user which would be the cloud service
[40:56] providers.
[40:58] >> Maybe I could give an example that
[41:00] involves me and Hong actually.
[41:03] So in the old days when we made
[41:05] transceivers, we actually sold them
[41:07] through an OEM who would then sell into
[41:09] Hong. When you sell through an OEM and
[41:12] sell into Hong, you never really
[41:13] understand her requirements because
[41:15] they're translated through someone whose
[41:16] business needs and motivations are not
[41:18] the same as yours. So by the time it
[41:20] gets back to you. So it's like
[41:21] impossible for us to collaborate
[41:23] together to take the best of my
[41:25] technology to get to um what what Google
[41:29] or or people like uh you know the CSPs
[41:31] need. But then whenever we can
[41:33] collaborate directly then Hong can say
[41:35] this is what I actually need and then
[41:36] like oh here's all my technologies I can
[41:38] try to address. I'm thinking of the
[41:40] 10GSR light actually that we developed
[41:43] together. And so I do think that it's
[41:45] very important for the CSPs or any big
[41:48] customer to be working with the person
[41:51] who actually owns the technology that
[41:53] you care about. That if there's many
[41:55] layers between you and the person who
[41:57] owns the technology you actually care
[41:59] about, then your requirements may not be
[42:02] getting translated properly to that
[42:04] person and that person's innovation may
[42:05] not be 100% properly focused on solving
[42:08] your problems. And also I want to
[42:10] emphasize important to work with um the
[42:14] one advantage uh for for me for us we
[42:18] work closely with Google deep mind team
[42:21] on the modeling software and hardware
[42:23] code design. So we come come up with the
[42:25] most efficient hardware and then we also
[42:27] work with the system architect to see
[42:30] what is the most efficient thing to
[42:32] scale the instead of scale the component
[42:34] just component power efficiency come up
[42:37] with new architecture to scale the the
[42:40] know to scale the system level
[42:42] efficiency one example is OCS when we
[42:45] use OCS to replace the spine layer and
[42:48] then that savings are 30 or 40% and when
[42:52] we use OCS to scale the scale up network
[42:55] the power saving there is is much much
[42:57] bigger than the power saving from the
[42:59] CPU and the more uh and more importantly
[43:03] I think it's uh you have to work with
[43:06] the deployment team see what is the SEO
[43:08] target you see I give this a cool
[43:09] technology but will take days to repair
[43:12] if things break so it's you really have
[43:14] to work with all layers of the the
[43:17] infrastructure to come up with a
[43:19] solution that is uh not only uh
[43:22] wonderful but also deployable.
[43:25] >> John,
[43:25] >> thank thank you very much for the
[43:27] question. Uh Ryan Carverton, Sarah
[43:28] Capital, you know, you guys are all in
[43:30] this scale business. You know, you're
[43:32] three of you are scale suppliers, one of
[43:33] you is a scale consumer. And so when you
[43:36] think about you guys are in a unique
[43:38] position where you know kind of what you
[43:40] do dictates the market or what your
[43:42] customers tell you that you guys
[43:44] implement like you guys are driving the
[43:45] market because of your scale. And so
[43:47] it's kind of a two-part question, but
[43:48] but on one side, how do you think about
[43:51] um internal investment in your own R&D
[43:55] for things that you know how to
[43:57] commercialize versus the potential to
[43:59] acquire new technology that seems to be
[44:01] very pervasive in the space and maybe
[44:04] the inability to a question or comment
[44:06] earlier in an earlier panel about the
[44:07] commercial viability of that and and how
[44:10] do you guys kind of think about um just
[44:13] your kind of role in in defining the
[44:15] market versus the the startup community
[44:17] that's out there that's looking to sort
[44:19] of penetrate their way into the market.
[44:21] >> Greg, why don't you take that first
[44:22] because you talked about the M&A that GF
[44:24] has had
[44:25] >> and I'll I'll actually come back to the
[44:27] last question as well. uh Ashkan and I
[44:29] were on a panel a few years ago and it
[44:31] was a debate about light in light out uh
[44:34] and it's like why can't the foundaries
[44:36] just align on a standard on said well
[44:38] why can't Nvidia AMD and Broadcom go
[44:40] align and it was one of those taste
[44:42] great less filling kind of things and
[44:45] you can't under appreciate the role that
[44:48] our customers play in uh defining what
[44:52] it does first and foremost it is a scale
[44:55] game what does it take to scale a driver
[44:57] customer that has a market and needs you
[44:59] to go do something. We don't we don't
[45:01] develop technologies just to do it. We
[45:03] do research on those things so that
[45:04] you've got a broad pipeline, but
[45:06] ultimately it comes down to locking in
[45:08] with a customer that says, you know, in
[45:10] 24 months I want a million units a month
[45:12] in this configuration on there. And that
[45:14] flattens a lot of the necessary
[45:16] discussions up and down the the supply
[45:18] chain on there. So I never underestimate
[45:20] the role of these MSAs
[45:22] among uh the drivers that really do sit
[45:25] between the data centers uh and the uh
[45:28] the folks that are manufacturing
[45:29] components for it. So um from a from a
[45:33] research standpoint you know we
[45:36] participate in as many consortia as we
[45:39] can that give us as broad a perspective
[45:41] as possible. uh I think our most
[45:44] successful participation has been at IMC
[45:46] not just in photonics but we have
[45:48] translated technology out of a research
[45:51] uh uh environment there across multiple
[45:54] uh fronts. Um we certainly we're a me
[45:58] limited partner in three or four
[46:00] different venture capital funds that
[46:02] give us a tremendous amount of insight
[46:04] into the startups that come. Uh and we
[46:07] either want to be their supplier or
[46:08] their acquirer and we're uh we're open
[46:10] to either one of those considerations.
[46:12] So for us, you know, it's having the uh
[46:15] the scanning the horizon, looking at
[46:17] what innovative uh things are there and
[46:19] to not be bashful, right? If it's not a
[46:21] core skill that we've got, but we
[46:22] believe it's going to be important, then
[46:24] uh certainly look to look to acquire it.
[46:26] >> Well, from Lmentum perspective, I mean,
[46:28] if you look at the history of the
[46:29] company, it's essentially kind of a
[46:31] series of acquisitions, you know, I
[46:33] mean, everything from all the way back
[46:34] to SDL where we acquired pumps to
[46:37] Neopetonics and Claro. I mean the
[46:40] company uh is very familiar with what it
[46:42] takes to bring in kind of new technology
[46:44] I think from a from an M&A perspective
[46:46] but you know thinking about it more from
[46:48] a technology perspective I mean I think
[46:50] kind of similar to what Greg was talking
[46:51] about you know we also spread our bets
[46:53] we have things that we're good at but
[46:55] you know we're also looking at a bunch
[46:56] of new technologies which may not
[46:58] necessarily be in our pipeline but the
[47:00] key thing there is we need time to
[47:02] develop it you know like I said I mean
[47:04] some of these new technologies it's very
[47:05] difficult to turn around a product in 18
[47:08] months so you we'll do some technology
[47:10] investigation at the very beginning and
[47:12] kind of see how it goes um just to
[47:14] explore things and
[47:16] >> I think I could add I mean we are as you
[47:18] said we are one of the biggest in the
[47:20] industry and we take that responsibility
[47:21] very seriously actually um and so you
[47:24] know we've driven a lot of the
[47:25] multissource agreements for this
[47:27] industry and a lot of the e standards um
[47:30] for the communication link um and I
[47:33] think that you know we try to be a
[47:35] thought leader and a good uh technology
[47:37] partner to to the eosystem system,
[47:39] right? But I do think and I do think we
[47:41] have a ton of capabilities inside of the
[47:43] company. Obviously, um but I do think
[47:45] that there's always a healthy ecosystem
[47:47] when there are startups in the
[47:48] ecosystem. You know, the startups um
[47:50] they they uh they push all of us to be
[47:53] better and a lot of times they have the
[47:55] opportunity to focus on one technology
[47:57] as opposed to us where we have to do all
[47:59] of the technologies at the same time.
[48:01] And maybe they can focus on one before
[48:04] it's kind of on our radar screen. So in
[48:06] our case, you know, I I just have a we
[48:08] have a long pipeline of things we do
[48:10] internally, but then we have a healthy
[48:12] respect and interaction with all the
[48:14] startups and just um you know, in the
[48:16] end, I have to have access to all the
[48:18] optical technologies and and as Hong was
[48:20] saying about things coming faster and
[48:21] faster and faster. When things come
[48:23] faster and faster and faster, you may
[48:25] not be able to do everything in house,
[48:26] right? And so then you're best off
[48:28] having all the partners. So I think it's
[48:29] a good to just overall have a healthy
[48:31] ecosystem. Yeah. Hello, Bushnesar with
[48:34] Aluma. With the supply chain constraints
[48:37] you talked about earlier, can you
[48:38] comment on the impact on the
[48:41] hyperscalers and how they're adapting?
[48:43] Thank you.
[48:46] >> It's a question for me, right? So I
[48:48] think uh with all the constraint we have
[48:50] especially on Indian for laser all those
[48:52] things there are way to get around it if
[48:55] you uh design architecture correctly. I
[48:57] can give you example. uh if you do IMDD
[49:00] you need the four laser but if you do
[49:02] coherent you only need one laser you
[49:04] can't do four and you can't do both you
[49:06] get extra uh but you can get actual uh
[49:10] freedom to to scale because if you you
[49:12] feel constrained by the I by the laser
[49:14] right
[49:15] >> my name is uh Frank Chong from source I
[49:19] have a similar questions um um similar
[49:22] to like we just for the argument there
[49:25] so I thought this is a pure plugable
[49:27] panel but we still have to talk a lot
[49:29] for the CPO look like if you cannot give
[49:32] a talk this year without talking about
[49:33] CPO but my question that so we talk
[49:36] about the CPO will be the end game right
[49:40] somebody keep telling me CPO will be the
[49:42] end game there so will be happening
[49:44] someday and the PL we enjoy the market
[49:46] for pluggable but look like the industry
[49:50] keep investing CPO and someday CPO will
[49:53] resolve all the issues that's pretty
[49:55] clear right if you keep investing the
[49:57] issues will will be gone. So my question
[50:00] probably is uh when do you think we
[50:03] starting worry about our jobs if we have
[50:06] to yeah we have to to do plugables.
[50:09] >> Could you repeat the last part question?
[50:12] >> Okay. Okay. Sorry. Sorry.
[50:13] >> Just the last part.
[50:14] >> I I think I got your point. I I think uh
[50:17] CPU has nice property. uh there's but
[50:21] there there's a good use cases
[50:23] especially for like I give you example
[50:24] network interface car where you have we
[50:27] still going to run out of density
[50:29] problem because for the GPU network it
[50:32] requires know very high bandwidth per uh
[50:35] network interface car if you pl put the
[50:38] plug there it's very hard to scale but
[50:40] there is endpoint where the
[50:42] serviceability is not issue you can put
[50:44] CPU there
[50:46] >> you my question mainly is maybe
[50:48] explain more clearly my question that
[50:51] everybody told me right the CPU will be
[50:53] the end game someday that we always
[50:55] there right some with a very bright
[50:57] future someday the plug will stop right
[51:00] run out of gas
[51:01] >> I don't really agree totally agree with
[51:04] you on that one Frank but anyway let me
[51:05] just say that um so we put out a market
[51:08] model last year and which was actually
[51:10] based on Vlad's data like counting but
[51:12] also our own internal input and we have
[51:14] pluggable transceivers growing through
[51:15] 2030 and at least the the the most
[51:18] recent market model we put out. So, it's
[51:20] 25 billion by 2030. And then CPO is
[51:23] smaller than pluggable transceivers. So,
[51:25] as much as we talk about C, I mean,
[51:26] we're all talking about CPO because it's
[51:28] a new technology. It's something new.
[51:30] It's bringing a lot of volume to our
[51:32] industry, which is very valuable. But at
[51:34] least what we've published is we think
[51:35] pluggable transceivers are bigger than
[51:37] CPO through the end of the decade.
[51:40] >> Yeah. If I can maybe add one more thing,
[51:41] I think CPO and transceivers, they're
[51:43] different spots in the network, you
[51:45] know, and think this is the other thing
[51:46] to think about, right? like um uh
[51:49] transceivers are very good for the front
[51:51] end, right? There's a lot of diversity.
[51:53] There's uh very high volumes. That's the
[51:55] side of the network which faces the kind
[51:57] of outside world. Um that's the
[52:00] hyperscaler network that was developed
[52:02] back in the mid you know 2010s or so.
[52:04] But CPO if you look at it today at least
[52:06] how people are kind of thinking about it
[52:08] is it's very much in the back end you
[52:09] know where there's all this connectivity
[52:11] of racks and everything else. And you
[52:13] know, we can debate whether you know
[52:15] today or 5 years from now, 10 years from
[52:17] now, right? But I think the problems
[52:20] that pluggable solve versus the problems
[52:22] that CPO solves, they're just they're
[52:24] aimed at different things. They're not
[52:25] not aimed at the same thing.
[52:27] >> All right. Hi, uh this is Rafi Islam uh
[52:30] from Cactus Materials. We're the
[52:32] manufacturer of the lasers. Looks like
[52:34] lots of discussion on the indium
[52:37] phosphate lasers like you know this is
[52:39] what the bottleneck is. Have you looked
[52:41] into the gallium arssonite like the
[52:43] pushing it to the 13 10 1550 uh we're
[52:46] looking into some g dilute nitra we
[52:48] already uh provided like you know some
[52:51] some of the samples to some of the
[52:53] customers I'm just wondering that what
[52:54] is the volume you require for indium
[52:58] phosphide that that the laser for
[53:00] example you're looking at that your
[53:02] backlog now and how it possibles to the
[53:05] other way like the architecture you
[53:06] mentioned something that might be four
[53:08] versus one lasers
[53:10] what are the other technology you are
[53:12] investing in as of today just to because
[53:14] it's a huge investment by the way indium
[53:16] phosphate for it cannot be scalable
[53:19] there's a lot of issues in there yield
[53:21] is not that good uh all sort of thing
[53:24] but what is your company is doing to
[53:26] address that problems thank you
[53:28] >> I mean in our case we have gallium
[53:30] arsonite fabs also so we used our
[53:32] gallium arsonite for 980 pumps and so
[53:34] that would be more for the transport and
[53:36] telecom side or scale across now and als
[53:39] also um well actually people are putting
[53:41] them inside of transceivers but I guess
[53:43] that's for the ZR also but also the GI
[53:45] Marsite vixel so gallium Marsite vixels
[53:47] and um and with the GI Marsite vixel
[53:50] what you're seeing is people starting to
[53:51] go to different wavelengths and things
[53:52] like that so but I would say overall
[53:55] that's more like 1060ish
[53:58] than you know 1310 or 1550
[54:01] >> yeah I would agree with Julie I mean
[54:03] actually both of us I think we have
[54:05] expertise both in Indian phosphide and
[54:06] gallium marsa kind of similar to the
[54:08] comment that I was before I think
[54:09] they're used in different areas. Um, now
[54:12] there are some interesting projects. I
[54:13] think Julie and I would actually agree
[54:14] on this that uh, you know, microixels I
[54:17] think are very interesting for some of
[54:19] these highdensity sort of interconnect
[54:20] applications because of the supply chain
[54:22] that it gives you. You know, big wafers,
[54:25] >> something which has been scaled into
[54:27] volume previously through 3D sensing. So
[54:30] yeah, I think gallium arsenide is a very
[54:31] interesting material. It doesn't solve
[54:33] exactly the same problem as indoite
[54:34] solves. It's a little bit different, but
[54:36] it's it's definitely an interesting
[54:37] material system.
[54:39] Yeah, I just want to echo that I I agree
[54:41] I agree with Matt that I'm very excited
[54:43] about pixels and small arrays of vixels
[54:46] for data.
[54:48] >> Okay, Jose, we got time for one more
[54:50] question. Two questions. All right.
[54:52] >> Yeah.
[54:54] >> George Zeras from Oreo Networks. Hong,
[54:56] you mentioned clearly that um uh
[54:59] interconnects is not the key point but
[55:01] the system optimization for AI clusters
[55:04] and TPU with OCS uh become more
[55:08] efficient comparing to other systems. My
[55:11] first question is do you see TPU pods
[55:13] with OCS to grow significantly larger
[55:16] >> and the second do you expect to have
[55:19] benefits if the OCS becomes a lot faster
[55:22] in terms of reconfiguration?
[55:25] uh so optical circuit switching has all
[55:27] the benefits is a data rate wavelength
[55:29] agnostic but does have a drawback is a
[55:32] slower switching so there are two places
[55:34] OCS will be a good fit first is for the
[55:37] spine switch where the data coming is a
[55:40] highly uniform so we use OCS to do slow
[55:42] switching and for ML if you look at a
[55:45] traffic c pattern it's a it's a
[55:48] predictable you know where the packet is
[55:51] go how long's going to be there and so
[55:54] this is different than traditional uh
[55:56] data center network where it's a highly
[55:59] bursty and you have to use electrical
[56:02] packet switch to do statics
[56:03] multiplexing. So I think the if you look
[56:07] at the Gemini training pattern you have
[56:09] this periodic networking traffic. So
[56:12] this is the place where we see uh
[56:14] optical switching uh circuit switching
[56:16] shines. So we talk about the efficient
[56:18] efficiency. You really have to design
[56:20] the architecture to get the best
[56:22] efficiency out of every components.
[56:25] >> Hi uh my name is Alexa and I'm with
[56:27] BlackRock. Um so in one of the earlier
[56:29] panels today we heard commentary on
[56:32] industry constraints not being in the
[56:33] raw capacity for the modules but rather
[56:36] um for some of the higherend components.
[56:39] um is that like a a broad-based shortage
[56:41] or is there any particular kind of
[56:42] component that you're seeing um that
[56:45] being more attributable to? I would say
[56:48] again we said but for our side it's
[56:49] really uh indium bossified is the
[56:52] primary constraint I would say uh I
[56:54] think uh a lot of the constraints coming
[56:56] from the architecture so if we want to
[56:59] do CPO for scale up it will require 10
[57:02] times more laser then we do use optical
[57:04] circuit switching because it's just all
[57:06] the bandwidth need to go through the CPO
[57:09] and also if you have a smaller scale up
[57:11] domain the networking requirement for
[57:13] scale out has to be like 4x or even 10x
[57:16] higher Then we the way we do scale up.
[57:20] So that requires 4x 10x more laser. So
[57:23] you really have to design first
[57:26] efficient network. So you don't have to
[57:28] use that many laser. So we don't we
[57:29] never had the laser problem and because
[57:32] first we go and get all the laser but we
[57:35] also don't use that many laser as like
[57:38] scale up network and scale out.
[57:42] >> Thank you.
[57:43] Yeah, I guess I just I I think exactly
[57:45] Julian and Hong said it extremely well.
[57:47] I think um and maybe the only thing I'd
[57:50] add to this is you know indium phosphide
[57:52] it's more than lasers. You know indium
[57:53] phosphide extends into other devices
[57:55] like MLS.
[57:57] >> So you know I think um yeah addressing
[58:00] addressing the shortage of India
[58:02] phosphide is a major priority I think
[58:03] throughout all of us in the in the
[58:05] ecosystem today.
[58:07] >> Okay. Okay. So, I'd like to uh have you
[58:09] join me in thanking our panel for their
[58:11] comments and participation.
[58:28] Everybody talking about scaling AI,
[58:33] but the data center's choking deep
[58:36] within.
[58:39] Copper running hot. Yeah, the signal's
[58:43] getting thin,
[58:45] so we flip the switch now.
[58:49] Optics is in.
[58:52] Bandwidth climbing fast. Racks are
[58:55] running red.
[58:58] Cloud demand exploding overhead.
[59:04] Pluggables fading as the limits clos in.
[59:10] Cold package light is how we win.
[59:15] It's photo
[59:17] baby 2026.
[59:21] Riding that light way.
[59:24] Doing new tricks.
[59:27] From the fiber in the ground to the chip
[59:29] in my hand. We make that sunshine jump
[59:32] on command.
[59:34] Yeah. Foodx baby
[59:39] 2020.