Hi, everybody. I'm Austin Moeller, the Senior Aerospace and Defense Technology Analyst here at Canaccord Genuity. Today, I am speaking to Michael Murray, the CEO of Kopin. Michael, I guess just to start out, can you talk a little bit about Kopin's background as a company and how it's evolved over the past few decades and the core focus today of microdisplays within the business? Sure. Many of the investor meetings that I start off usually start with, "Oh, I know Kopin from 1996, 2006, 2016," because Kopin's a 40-year-old company started by Dr. Fan out of MIT Lincoln Laboratory. I took the company over just over four years ago now. By background and introduction, I'm a turnaround CEO by trade. I spent over a decade at a company called Analog Devices here in Massachusetts. Did a little bit of work with the NSA in Type 1 cryptography thereafter. Worked at a company called Ultra Electronics, which most of you don't know, but it's a 100-year-old defense company out of the U.K. Which we sold to Advent here in Boston for $2.5 billion, which they just actually resold for $10 billion, so they did all right in that. Kopin right now is focused on microdisplay development. We build four different types of microdisplays here in the United States: LCD, OLED, ferroelectric liquid crystal on silicon, and MicroLEDs, which is our big franchise that we're developing. We invented a fifth kind, Austin, which is the bidirectional microdisplay called NeuralDisplay, which actually was the birthplace of our Neural I/o chip, which we partnered with our friends over at Fabric.AI to develop, which is a 1.6 terabit AI GPU transceiver. Can we talk about brightness, energy use, and other advantages of using MicroLED displays in headsets and other applications versus OLED or liquid crystal on silicon? Sure. So brightness as a spec, the displays that you are all probably looking at right now. Roughly speaking, I am going to use a rough term in terms of nits. It is about 1,000 nits of brightness that you are looking at right now. That is a common LCD type of brightness. OLED, very bright. We can get up to roughly 20,000 nits of brightness with the top-end device that we can build. The problem with organic light-emitted diodes is they are organic. They are like your teeth, your bones. If I push a lot of current through them, they will start to degrade over time. It is just physics. The more current and brightness you push through an OLED, the dimmer it gets over time. If you are flying an aircraft, that is a real bad thing. You want to move to a MicroLED. Our current MicroLED that we actually demonstrated here at our tech day yesterday is capable of 1.8 million nits of brightness, which is 3.8 foot-lamberts of brightness. That is basically a laser, for lack of a better term. Because these tiny little pixels act like a laser, that is why we can use them as a data-centric type of medium. Can we talk about the MicroLED production line established for Soldier Borne Mission Command and its current throughput and the capability to scale that capacity with more defense industrial base investment? Sure. So for those that do not know, Kopin received a $15 million Industrial Base Analysis and Sustainment Act award. Which is basically the Department of War reshoring a lot of capability here in the U.S. that necessarily we may not invest in on our own. That production line is now being assembled in Westborough, Massachusetts. We received our custom-made bonder actually last quarter. It is now being implemented. At its peak, with the current structure that it is, we believe it can turn out about 160,000 wafers per year. Depending on the size of the chips that we will produce on those wafers, we could receive up to several million units per year. Can we discuss Kopin's involvement in building displays for FPV headsets for the Drone Dominance Program? How many headsets would be needed for a specific quantity of drones as part of that? Yeah, great question. One that I got pretty much every investor meeting today. Our friends from UMAC aren't here yet, so I'll talk badly about them first. Working with the Drone Dominance Program with The Pentagon, essentially these are all focused one-way drones or OWA drones. Essentially they're bullets, folks. They're not coming back. Those types of drones, the U.S. government wants to build around 3 million drones next year of this class. The Drone Dominance Program is all online. You can actually look it up and see the Drone Dominance Gauntlet winners. Currently, Kopin is supplying headsets to one of the top three. They've already supplied 6,000 drones to The Pentagon, and we've supplied headsets to go along with them. Kopin right now is in 5 of the top 10 in the current gauntlet. What we've learned is the ratio, depending on the size of the drone, if it's a 1 ft and below, the ratio is about 1: 10 drones. So 1 headset for every 10 drones. If it's a foot or bigger, it's about a 1: 20 relationship, so 1 headset for 20 drones. The reason that is the ordinance and the size of the drone itself and the use cases that we see. It's public knowledge that Ondas and UMAC are investors in Kopin, as well as a company called Theon International, who most of you probably do not know. They are a Greece-based company, and a major supplier to NATO. We're working with all three companies on headsets to drive essentially and pilot their drones. To give you a sense, next year we're guiding to about $10 million of revenue, at least in our first-person viewer drone goggles. If more of the drone dominance competitors win that use our headset, that could go to $50 million of revenue next year. We don't know who's going to win the next gauntlet. That's up to them. All we can do is supply them our headsets and wish them well. Did I answer your question? Yeah, I think so. All right. What does the competitive landscape look like for Kopin today? Are there manufacturers that have similar technology in the U.S. or is it primarily in Asia-Pacific? Primarily in Asia-Pacific. Are we talking Drone Dominance or MicroLED? Mostly MicroLED. MicroLED, really, there's four competitors. When we think about supplying the U.S. or AI infrastructure and AI factories, there's basically four competitors that we consider. The first one actually is a company out of the U.K. called Plessey. Plessey is a very viable competitor in MicroLED specifically. This crappy company out of Cupertino called Apple invested a lot of money into them to build a MicroLED for their watch. Unfortunately, Plessey and the U.K. government did something that I think is really silly, is they actually sold that company to the Chinese government. The point is, Plessey, from our standpoint in the U.S., isn't really a competitor because they can't support the U.S. government since they're owned by the Chinese government. Moreover, last week, this is very important to our friends at Fabric.AI as well as Kopin. The U.S. government basically initiated an enactment that says that even data center interconnects cannot come from China. That includes MicroLEDs and fiber optics and copper. That's a big deal for us. That's opportunity one or competitor one. Competitor two, a company called ams OSRAM out of Germany. We know them very well. They do have a MicroLED. I think the competitive nature with that, with us is their micron pixel structure is about 25 microns. Ours is about 7: 9. So we're a little bit smaller, actually a lot smaller. Because we're smaller, we can switch faster and actually receive better brightness, which is an advantage for an AI chipset as an example. Competitor number 3, Mojo Vision, a company in California. You probably have also heard of this company called NVIDIA. They invested a timely amount of $2 billion into Matt Murphy's company, a company called Marvell. Marvell invested $150 million into Mojo Vision out of California. Mojo Vision is an MIT startup. They still use the MIT fab, which, by the way, used to be my fab when I was at Analog Devices here. We know the structure that they're working with in MicroLEDs very well. We think we have about an 18-month to 24-month advantage on Mojo Vision. The fourth manufacturer of MicroLEDs is Kopin. Well, I guess we should talk about the AI infrastructure opportunity. Can you just go into detail on the AI infrastructure partnership with Fabric.AI and the need for MicroLED-based optical interconnects to replace copper wire and laser inside of data center racks and chips? Sure. Josh is here. James is here. Great partners of ours. They see the vision in what I see as well, which is copper has had its day in data centers. Unfortunately, the reason why you need copper in data centers is there's no other alternative other than a fiber optic laser. Now, there is an alternative, and it's a MicroLED. Copper at its best can probably hit 800 GHz of frequency range. Unfortunately, when you listen to Jensen Huang in NVIDIA, he's already talking about 1.6 terabits to 3.2 terabits of data transfer. That's clearly not fast enough. Therefore, you have to have multiple lines of copper to achieve that rate. When you have multiple lines of copper, it's expensive, the density is higher, and it takes a tremendous amount of energy to cool that system because you're pushing a lot of current through it. That's why you need liquid-cooled data centers. What to do? You can actually use a fiber optic laser. Unfortunately, fiber optic lasers have an MTBF that's really, really bad. Secondly, they need to be calibrated, and thirdly, they're physically bigger, and you need a fiber optic loom to go outside of it. MicroLEDs and our partnership with Fabric.AI is a great solution to provide a transceiver at 1.6 terabits in one chip. Fabric.AI Has funded Kopin to build that chip with them and for them. They have exclusive right to the hyperscaler part of this market. Kopin is leading with a defense-first posture where we'll work on the same chipset but with NIST, DISA, NSA, CIA, et cetera. Of course, the Department of War, who needs these chips as well. We'll be utilizing the same production line that the Department of War has already invested in in Kopin. That Department of War production line will also produce the chipset for our friends over at Fabric.AI. I would assume that The Pentagon likes that they can probably get a better price point on the MicroLEDs for their helmet-mounted displays by building it on the same production line as Soldier Borne Mission Command with the data center optical interconnect. Look, everything in semiconductors is about utilization rate and throughput. The more wafer we ship through our fabrication plant in Westborough, the lower our costs become because our absorption rate of that factory goes up. The more wafer we push, the lower the cost of the Fabric.AI chip, the lower the cost of our color MicroLED for Soldier Borne Mission Command. If you recall, the only competitor we have on Soldier Borne Mission Command is a company called Mojo Vision on the West Coast. Kopin's producing Fabric.AI chips. Kopin's producing our own chips that support the government and our production line, and our production line for other headsets. We're still building color MicroLEDs for other customers, of course, and Soldier Borne Mission Command. We feel like we'll have about a 25%-35% cost structure, advancement, and edge over our competition. It will all be built here in the United States, which for the government and a lot of data center users now is very important. Yes, the government's very happy with our choice to partner with Fabric.AI, build this technology in the United States, and support the Department of War. I guess, can you just talk about how important it is that you can achieve significant power reductions? Both in the power that has to be pushed through the optical interconnect and the cooling that's needed in the data center? Just given there's a lot of public acceptance issues around data centers and their energy use? Oh, absolutely. One of the theses that we worked on actually with James and myself and Bill Maffucci and a couple of the whiz-bang scientist kids that we have working for us is how much money can you save by switching to a MicroLED? The numbers that we came up with was 27 kilowatts per rack per month, which turns into around $25,000 of cooling savings per rack per month. That is millions of dollars of savings for the average data center. The amount of data centers, you would know better than me, but there are many data centers out there and many, many racks. This is a significant overall system savings of power. That is one piece of the puzzle. I know our friends at Fabric.AI are working on some other pieces. We also have another company that Kopin's working with called Heat to Power that takes power from heat generation from the data center and turns it into electricity. We're working with that company on a few new technologies as well. But this is one of the big advancements for MicroLEDs. I believe Lip-Bu Tan, the CEO of Intel, on Friday came out and said they're putting $1.6 billion to work for U.S.-based fiber optic and MicroLED technologies for U.S. companies to build that to save the cost of the data center. All of these companies are now putting money to work. NVIDIA put $6 billion to work, and they're going to put another $2 billion to $3 billion, I'm told, into co-packaged optics and near-package optics before the end of the year. If we just think about the defense budget opportunity for helmet-mounted displays and FPV headsets, how could an additional reconciliation package, whether you want to call it reconciliation 3.0, 4.0, increase the available budget capacity either for night vision goggles, helmet-mounted displays, or for the FPV drone headsets? Well, I think the budget itself is structured and it is adapting to what we're learning in Ukraine. I think that's number one. Number two, the amount of spending in first-person-viewer drones, I think the government's actually been very explicit that they're going to build 3 million drones next year and increase the amount of drones every year, especially in that one-way direction type of drone technology. For us, we follow the budget lines very carefully. We see the available budget for things like aircraft HUDs, so that could be F-35, F-47, and certainly all the rotary wing, so that would be Apaches, et cetera. Those technologies are being upgraded from old LCD technologies to MicroLEDs. We have a number of programs of record that we're working on right now. Too early to tell if we'll win them, but they're significant. So that's number one. Number two, we are seeing thermal weapon sights also be upgraded from LCD to MicroLED as well. Of course, the big one, Soldier Borne Mission Command, and that is a combination of daytime use and nighttime use. That is a $350 million- $500 million contract to me, to Kopin, that we are vying for right now and will have a good bead on in November. The budget process and all those things I just talked about, we have direct line of sight to lines in the budget that are going up, not down. The big iron programs that we were on, like Abrams tank as an example, still do not know what is going to go next with that. We do not know whether or not tank warfare and the Abrams program, what it is going to be in the future and who is going to pay for it. That is a lot of data to throw at you, but otherwise said, MicroLED is going up because we know the Chinese have that capability, which means they can see us before we see them, and that is a bad thing. Can you discuss the NDAA requirement that all defense components need to be sourced from outside of China, preferably in the U.S. by 2030, and how that plays into how you are thinking strategically about the data center market, which I assume the focus is on government data centers in the D.C. beltway? Yeah. I have a couple of use cases that are kind of fun to talk about. I will start with the last one first, if that is okay. Let us think about large language models for a moment and think Austin is a GPU. I am an input. You are all the output, okay? If I were to start to integrate data into the input of a GPU that states that people with A+ blood do not need a vaccine for SARS or any sort of disease that could be out there. I am training Austin to not worry about people with A+ blood. He is doing his large data model set, and he comes up with a vaccine that states that we do not have to worry about people with A+ blood. Johnson & Johnson and Roche, they start producing all of their vaccines, and they're not worrying about people with A+ blood. Guess what? The people with A+ blood are now vulnerable to that disease. If you want to target regime change, you can target a human genome. You can target a person based off of a pacemaker. You can do the same thing using large language models and actually induct regime change through learning models, and that is a very scary thing. The reason we think about this is we've done it as a country. Stuxnet as a program is exactly that. It's a large language model that looked for certain aspects of Iran's nuclear power plants and how it worked, and when it found those specific things, it turned them off. Large data models are no different than a virus. If you inject it in the beginning on the input, the downstream output is yours, and that is catastrophic. That is why NIST, NSA are all working now to provide data and technologies that are built in the U.S. that do not have backdoors in them. That's very important technology to get right. If you're not going to allow GPUs to be built in China, you certainly shouldn't be allowed to build interconnects in China because, A, they have a processor, B, they have memory, and three, they control your entire system on the input of your large learning model. That's catastrophic. I can't stress that enough. Okay? That's the scary one first. What was the other one? Drones. Okay. Drones, same problem, but that's more of a source of supply, right? Having a source of supply that isn't coming from China is critically important so that they don't turn it off if there were some sort of activity, one. Two, I can tell you if you come to my shop, I will show you the active processors that are in every single microdisplay in the world. These are active components. They're not dumb. The TVs that you have at home have a microprocessor in them. They have a Wi-Fi connection, a Bluetooth connection, and a speaker. That's all you need. A microdisplay has all of those things except for the speaker. So it's a system that you can hack through if it's built in China and you put in the right backdoors. That's also bad. For those two reasons, the U.S. government has mandated that AI transceivers, co-packaged optics, and microdisplays need to be built here in the U.S. That will be the source of supply for drones, AI infrastructure, and certainly AI interconnects. I think that to your point, we could just talk about the recent news that, I guess the U.K. Ministry of Defence had some unmanned surface vessels that I guess had Chinese firmware on them, and they were transmitting all of their data back to China? Yeah, a perfect example. This isn't new, folks. Of the things that we learn, and my background from doing work with the NSA in Type 1 cryptography, the things that you read in the paper, there's 10 other ones that you don't read and won't know about, right? These things, unfortunately, are invented by us and our adversaries learn them because we use them against them, right? That's where Stuxnet came from. We actually used it against the Iranians. The Iranians used it against us. That's fair. The point is, when you think about where the MDA is now and NATO is now, we've learned time and time again that chips coming from China have a backdoor, and they call home. We've got to get this right. I think we made a major mistake offshoring a lot of our production and manufacturing capability, not only as an economy but certainly from a national security perspective. Allowing access into our data centers is just horrific, and catastrophic. I'm glad to see that the U.S. government is mandating that we build these chips in the United States. Do we have any audience questions for Michael? Just one maybe on the AI sort of optional interconnect. It seems like it has good affinity for the GPU, but then it strikes me that it will need to be packaged in close proximity to the GPUs. It seems like those design and architectural decisions are made with pretty long lead times. So even if it works and it is something people want to do, how long, like what generation of GPU could that technology from you or Fabric.AI actually be commercialized? Yeah, great question. It is a very intimate design cycle. There is no question. So I am going to answer the question in three parts if that is okay. The first part is we have learned a lot between the two companies recently that these transceivers, a standalone transceiver, does need to sit very close to a GPU, and the intermediary between those two are three ways. Over the air, through a fiber optic connection, or through what is called a waveguide, basically a piece of glass between the two. But they actually do sit today very close to one another. If you look at a very dense H100 as an example or multiple H100s, they are actually very close in proximity. So the distance between that we need to go is not that far. I do not want to speak for Fabric.AI. They are here if you want to talk to them. But we have a number of demonstrations that we want to do over the air through a micro optic, like a waveguide, and certainly using fiber. We are capable of all three of those capabilities and modalities. So that is one. Two, when we think about some of the technologies that we are focused on, chip-to-chip is very short run. Board-to-board, which in a rack, it looks like a pizza rack, so you have a 19-inch rack, one GPU here, one GPU here. You have maybe 2 inches between those two boards. That is not a long run for a MicroLED, not at all. We think we can actually do that today. So, rack-to-rack or anything above about 300 mm, it gets really, really tough for us to transmit at that level. You can receive at that level, but transmit is really tough. I would say short runs, yes, we can have glass or fiber optic or over-the-air. That is all doable with our chipset that we are developing right now, and we hope to demonstrate that, and will demonstrate that at CES in January. It is about an 18-month design cycle from there to get into the next generation of boards. That is about the right level, in my experience, about 18 months to get into production. Secondly, one of the very interesting data points that we learned from our white paper that was just announced, we actually had over 200 downloads in a week. But they were from really interesting companies that I did not expect. They are more of the belts and suspenders type of companies that provide racks and services like Celestica, Foxconn, Penguin Solutions, companies that we have never really dealt with before, nor did I think would be interested in this technology. They are, because they are the ones that are structuring the communications of the racks themselves. It has been a really interesting learning experience thus far. It sounds like mid 2028. Yep, I think that is fair. Josh and I have to have a conversation with James to see do we want to speed that up, because one of the big things that we were concerned about was the receive side of our chip. Kopin had never done that before, but we figured that out and we are actually going to buy it. We are not going to build it. So we have a great supplier on that now, so that risk has been retired. Now it is building LEDs, and we do that better, I think, than anybody in the world. So our risk just went down. Our speed just went up. Maybe we should go monolithic right now, and still do the demonstration, but get to that tapeout as quickly as possible and shorten that 18 months. We have an 18-month lead right now, and I do not want to give that up. I hope that is helpful. Yeah, thank you. Fabric.AI guys made a very interesting statement, which was that with MicroLEDs, you could conceivably have each photon be a transmission channel, basically. Correct. In my very positive mind says, "How would you even go? You've got to transmit and you've got to receive something. So maybe you can send it out and transmit. How would you even go about identifying and isolating any individual photon to get those channels? It's called interpolation, or combination. That's a very tricky thing. Can I give that to you? Yeah, sure. I'm going to do hand signals. I'm sorry. If you have four LEDs, the ability for you to have four LEDs and transmit data depends on how thick your LEDs are, the spacing between them, and how good you are at shooting against the receiver. Thinner is better. If you have a 10-micron pixel, you need about a 20-micron spacing between them, so that when you shoot the data, it doesn't hit the next pixel. Does that make sense? Besides that, you've got a receiver. It looks like an inverse checkerboard. One checkerboard is black, white. The next checkerboard is white, black. You're always transmitting and receiving data. As you space those things out over time, the ability to hit that pixel becomes harder because light tends to look like a flashlight. It comes out collimated and then does this. As it does this, it hits a different pixel. When it hits a different pixel, things go bad. That's the test that we're doing between distance and pixel size, and that's the experimentation that we're doing right now. They have to be close enough so the light doesn't disperse and get- Exactly right. Over-the-air, exactly right. If you use an intermediary or a medium, it is pixel to pixel, receiver to receiver, no big deal. But for us, we see a tremendous value in being over-the-air, as well as a low-cost optic being a waveguide or something along those lines. For those types of short-run applications, we think we have a very viable solution. I was in California last week meeting several of the largest companies in the world, and they said, "You have got the best transmitter in the world, by far. Do not worry about the receiver. Get it right for Fabric.AI. It is going to be a great chip." We think it is going to take a lot of that market, which is $63 billion worth of an optical interconnect market. There is lots of room to play. Can you talk about, are there technical specs yet for what you are looking at developing? We are talking about onboard transceivers, pluggable transceivers. We are talking about how many lines in parallel, what data speeds at each level, what line code you are using even. Sure. Has that been defined yet? Or is it way too early? No, we have a spec with Fabric.AI. I am not going, that is their spec. They can inform you on it. What I can say, and Josh will beat me up if I go too far, but 1.6 terabit bandwidth, must have. Certainly the keying, we have figured that out, will be NRZ. We also know that from a structure perspective, it is going to be a transceiver, so transmit and receive. Pluggable or onboard? That is all onboard. Pluggable, when you say pluggable, we could have a solution for that because that is basically connecting to a cable or a fiber optic. Is that what you mean? Connecting to the electrical to a SerDes. Oh, to a SerDes? In some cases, we may not need a SerDes anymore. Yeah, you still do because it comes out into the processor one way. It depends on if you're in the processor. Whatever, 32 parallel or 64 parallel traces. If you are in the processor, you do not need to. You can take that data straight from the wire bond, and you are not even passing a SerDes. That is a co-packaged optic in that case, to be fair. When you talked about comparison with electricity and the like, you are comparing that with active electrical. Your comparison was against a short reach active electrical cable? It was a comparison against active copper SerDes. Are you looking at maximum 3-ft reach, generally? Too early to tell. 1-ft reach. Right now, what we're focused on in development is spacing that product as far as we can until that interpolation becomes a problem. So too early to tell how far we can go, but I think the original spec was 300 mm. That's what I thought. About a foot. Yep. Uh-huh, and most active copper is in the 3 ft-7.5 ft range, isn't it? In most use cases, like for rack. Yeah, you can run copper as long as you want. No, you can't. No. Not at those line rates speeds. Oh, no, not at- 1 to 800. No, but if you have a temp sensor on a board, yeah. If I go look at a spec, I don't know, pick somebody, pick the chip guy's spec is under 7 ft, right? But it's definitely well above 1 ft, so that means how much of the market is sub-10 that maybe these guys want to have? How much are we talking about? How much of the inside the rack market is available for only 1 ft? Could have tell you, but it's all racks, so you're- I know, but normally the rack could be 7 ft, the full rack. You can only do within the single rack or an adjacent rack, that's a very tiny portion of the market. Yeah, but we're doing GPU to GPU. We don't have to have every, the GPUs will hand off messages to each other that'll still be faster and more energy efficient than- To do GPU to GPU, NVIDIA has to design you in, right? No. Depends how the data center door is designed. Depends whether we're doing over-the-air or putting some fiber in there. Depends how, or on the data bus itself versus you're saying with SerDes. There's a lot of factors. Well, I understand, but generally, to have widespread adoption, would one of the GPU or TPU vendors, not just NVIDIA, have to design you in? No. The data center, the hyperscalers have to design us in, at NVIDIA or- Would they have to be designing anybody in? They generally have been going with, unless it is a normal rack, not a GPU rack. Yeah. Again, it all, it depends on the hyperscaler. It depends on so many design factors. Well, I think you are also conflating two different things. There is transceivers that can sit in a cable, like a 3M cable or a short cable. Microsoft has proven this, by the way. They have a cable with a MicroLED at the end, and it plugs into a rack. The end of that cable is another GPU, right? So that is a use case that you can actually look up, but that is a near-package optic. A co-packaged optic is what you are referring to, is a transceiver that is inside the actual package of a GPU. Two different things. All right. Well, very engaging conversation, but thanks so much, Michael, for taking the time to speak to us about Kopin today and give us the latest updates. Thank you very much for having us at the conference. It's been wonderful. We have a full day today and a full day tomorrow.