Matthias: good afternoon, Jason. ⁓ Good to see you. ⁓ Welcome to the I wanted to start podcast with a story the origin of Lumai. you invented the technology, the core optical technology in a lab in Oxford, and then you became the CEO of Lumai. How did this come about? How did you realize that you want to move from academia to creating a business? was the real moment when ⁓ got excited about that and wanted to go ahead?
Xianxin: Yeah, so thank you, Matthias, for inviting me to this podcast. ⁓ Yeah, this is a long story. We basically started ⁓ journey like six years ago, When I first started to do the research ⁓ Oxford here, we basically, initially we came up with the idea and then... It was like almost too good to be true. And so we first basically did a very quick experiment at the university in the lab. That was like pre-pandemic. then, the result was actually better than what we imagined. And then one thing led to another. And then we just keep working on the tech, right? And then we saw lots of commercial interest. Once we our paper, then lots of interest out there. And then basically, then we decided, OK, it may be best actually to get started business. get VC funding to basically like commercialize the tech. So that was like how why we decided to Luma from our research.
Matthias: And how long did it take you from doing the scientific experiment, publishing, to creating a company? What's the time span?
Xianxin: So initially we the paper, right? And then we decided to basically ⁓ the idea at a bigger scale before we started the business. So ⁓ we demonstrated the system at a much bigger scale and it still worked really well. And that's why we decided to start Lumai So ⁓ have been working on this tech for like about three years and then we started Lumai, ⁓ also three years ago.
Matthias: Okay, so it took a number of years, got it. Okay. And then when you did this shift, you know, and moved from being a researcher to a CEO, what was the biggest surprise for you on a personal level, but maybe also on a technical level? You know, what, how do you describe that shift from one to...
Xianxin: Yeah. Yeah, I think the transition has been really ⁓ like very rewarding, right? Because when we were working at the university, we only have like a few people, right? A few PhD students, a few colleagues, and then we work on everything. And now actually, we started the business, we actually have a much bigger team. We have much more resources and support, right? And so actually we can do things at a much faster pace. So this is actually really amazing. We didn't really expect that we can achieve so much in such short time. at the university. So I think ⁓ as a researcher, you have your freedom. You can pick whatever you want to do. You don't have any pressure. But then once we started the business, yes, we have lots of resources. We have lots of support. But then we also have a much, much, much more well-defined goal because we have to basically connect the technology to the application to the problem. I think in this sense, we don't really have too much freedom, but You definitely know why you doing this. You definitely know why you want to do it because you want to solve a specific problem So in that sense, this is very different from research But I think this is actually much much better as compared to I just doing random research
Matthias: Okay. I hope you didn't do random research, very focused. Okay, well, maybe maybe for our listeners, let's talk a bit about, you know, the technology and the breakthrough. I mean, if if you look into literature, you know, for many, many years, or even decades, people talked about optical compute, that companies have struggled to
Xianxin: Random but you still got publications, right?
Matthias: move from concept to something that works in prototype or even at scale. is the difference? How Lumai make this What is the technical background behind it?
Xianxin: Yeah, yeah, yeah. So once again, the reason we started LUMi was because we demonstrated our technology at the university lab, and then we saw that, OK, there are other spin-offs, photon spin-offs outside there. and they haven't really achieved what we did. But then they can still like raise lots of funding and then we say, okay, if they can do that for sure, we are far better than yes, we should start a business, right? So that's how we started. back to your question, ⁓ right? ⁓ We are using ⁓ and optical effects to do compute, right? ⁓ in traditional digital domain, you need to use lots of transistors to compute, right? The optical equivalent of those transistors are like optical beams. The more optical beams that we can process, the more compute we can do, right? And then our system, our approach is fundamentally different from other like photonic chip based approach. Photonic chips are basically, okay, you still do compute on a 2D chip, but you change from electrons to photons. Therefore you have some benefits, but doesn't really scale, right? Our system does processing in the 3D space, three dimensional volume instead of a 2D chip. Therefore naturally we can process many, many more beams than what you can do with it. other approach. So that is how our system can scale much, better than any others. And that is actually fundamental to all the benefits, all the compute advantages in our system.
Matthias: sounds intuitive, right? I mean, you move from a chip to three dimensional. what are the challenges if you do that?
Xianxin: Mm-hmm. Mm-hmm. Yeah, today it sounds intuitive, right? But when we started, actually, most people still wanted to do the photon approach because that also has lots of benefits, right? It's tiny, it's still compatible with SCMOS, right? So you will want to continue to compute that way. But nature is very different, right? Our brain, our brain use neurons that are connected in three-dimensional volume, So nature is very fundamental different from digital. So in the past, most compute people tend to tend to think of computed still with the chip based approach. So right now we say, yeah, you can do more compute in the 3D space, but that's not natural for compute people at all. So, in terms of life challenges, there are lots of things there, right? Because, okay, you can, the more, basically you want to process as many laser beams as possible in ⁓ a physical volume, and that requires us to come up with very, very good optical design and with very good optical components. So in the past, people can demonstrate this kind of idea with a few laser beams, but that's just not enough to output from digital. You will need to do millions of laser beams operating at very high clock rate. So, I mean, if you really want to up perform digital computer system, you need extremely good optical design. You need extremely good co-design of both optics and electronic systems. So that has been very challenging, but once again, we have our own unique data encoded methods and some other patterns for us to basically overcome those challenges. That's why we can actually achieve very, good compute performance today.
Matthias: Okay. And how, I me, how do you, I mean, I understand you create millions of laser beams, get this, them in three dimensional space. So that's, that's also, but how do we do, ⁓ mean, I understand the main product that you're doing is matrix multiplication. ⁓ do you do, you know, with laser beams then, ⁓ know, how do you make them interact with each other so you can do matrix ⁓
Xianxin: Mm-hmm. Yeah. Yeah, yeah, yeah. So ⁓ basically, we are not trying to use optics to create a CPU or GPU, right? So what we have is very different. So we basically use nature. We use those natural optical effects, like interference and deflection. We use these optical effects to perform a ⁓ very specific matrix operation, just matrix multiplication. And it turns out that this kind of operation, ⁓ you do it with optics, it's very simple. It's very energy efficient. What's even better is that these matrix multiplication are very fundamental and they are used extensively under all kinds of neural networks. So that's what we do. So back to your question, we use interference, we use this optical effects and lenses and fibers, these components to perform exactly what optics is good at.
Matthias: Got it. it. Okay, so you need you need lenses, you need sensors, etc. ⁓ If I get it correctly, you started actually as a optics, PhD or student or expert. How, you know, how did help? I mean, this is not a quantum computer, I understand. was a traditional computer. So what's the relationship and how did your quantum optical quantum expertise help in?
Xianxin: Mm-hmm. Yep. No. Yeah, yeah. Indeed, I did my PhD in quantum optics. And after five years, I decided to work on something that's easier. And that's why we decided to build optical computers for AI. So if you look at the system that we have, we actually use lots of those similar components and techniques that people will apply in those quantum systems as well. But what's very different here is that we don't really need use like quantum effects, right? Quantum effects are very fragile and they are very sensitive to any environmental perturbation. So we don't need any of those quantum effects. Therefore, our is actually far more robust. It's actually much easier us to scale our performance and for AI. So that actually is very important. So we leverage the technologies, those really advanced optics and electronic technologies that people would ⁓ use in systems. But then we don't have high level of sensitivity and environmental perturbation. ⁓ So we benefit from that. And ⁓ using technologies, we can scale our system. We know how exactly we can ⁓ improve performance. And ⁓ at same time, we don't really need to continue to do very, hard engineering work for another five or 10 years. So ⁓ our computers can be ready basically ⁓ in years'
Matthias: got it but then they don't have the effect that quantum computers have that they scale exponentially they scale more linear
Xianxin: We scale more linearly, but what's more important here is that quantum computer, they can be extremely useful for certain tasks, right? And what we can unlock here is to compute energy efficiency, right? So we can solve today's AI problems and we can use our optical computers in data centers, and we can reduce the power consumption. This is actually very important because normally don't really use quantum computers for these applications, right? They are at their own domains.
Matthias: I I see. Okay. I want to shift a bit and we want to talk about, you know, from lab to company. We touched on that previously a bit. you ⁓ Lumai together with your colleagues, James Bale. What do the two of you bring? How are you complementary? How do you as a partnership? What's your different backgrounds? How does it work as two co-founders?
Xianxin: Yeah, we together with James, we have been working together for six years since the university. And I will say, we basically co-developed the technology, right? We published lots of papers together, we applied for patents, and then we started Lumai from our research. So both of us have technical backgrounds. We understand optics, physics, and AI really well. But the difference here is that Basically, at the university, I was a postdoc, designing the project and thinking ahead. And James was a PhD student. And then he basically implemented all of those He demonstrated the system. He in the lab, like day and night. ⁓ So the that we work is basically, most of the time, I am a thinker. I plan ahead. I plan for the project. And then I look ahead and say, OK, what's the next step? What's the roadmap? And then James execute the plan. And he can basically demonstrate all the technologies, all the... all the inventions really, really well. So he executes the plan, right? And that goes so far ⁓ well. ⁓ now he's CTO in the business and he worked very closely with ⁓ engineering team to basically build our product. ⁓ then I'm a CEO and then I ⁓ ahead on the product market fit, et cetera.
Matthias: Okay, so you're basically first wave and then he's basically making it happen in as a CTO. it. Okay. ⁓ know, your startup that works obviously at the intersection of hardware ⁓ software, I guess. ⁓ hardware is very capital intensive. Many, many startups shy away from hardware. What the ⁓ main you had when you had, you know, to turn from a lab prototype into something that
Xianxin: Yes. Mm-hmm.
Matthias: can be manufactured eventually at scale at this stage still as a prototype.
Xianxin: Yeah, I think in terms of manufacturing, so. the one of the biggest challenge is like on the components and on the supply side. when we decided to use optics to do computing, we initially, well, even now we see a major benefit there because there's a huge optical ecosystem, optical supply chain out there, right? Lots of those optical components ⁓ surprise optical manufacturers that we can work with. ⁓ see that as a huge benefit that we can leverage ⁓ do computing, right? So we work with not just the electronic industry, also optical industry. So that gives us really a very strong advantage there. But in the near term, right now, we have a small team and we are a startup. And naturally, we just need to deal with lots of these optical supplies. And we will need to work on many, many components. So that actually takes lots of time. So we will need to work with them to specify the components and then to test the reliability, to test the matches. So that has been taking time. But I think over time, then we will still benefit a lot from this, because at least we have lots of supplies out there. we can work with. if we don't, then we will have to ⁓ many of these materials or components, and we definitely want to do that. ⁓
Matthias: Okay, so most of the components are off the shelf and you buy them and it's more about the integration.
Xianxin: yeah, most of them out of the shell, but of course we will need to re-engineer them, right? The good thing is we don't need to do it all by ourselves. We can work with a very large supply chain and ecosystem.
Matthias: Okay, cool. I mean, the other thing that happened, probably more recent, I you got a bunch of awards recently, I think you're on the Photonics 100 list and just something like a week or two weeks ago, you won or the company won the Falling Walls Foundation Award 2025. You know, how... How does it help the company? How does it validate the achievements, the technical work? What does it mean for LUMEI? ⁓
Xianxin: Yep. ⁓ Yeah, I think, yeah, indeed over the past like 18 months, we have been receiving lots of recognition, right? And I think this is a very good sign because in the past people have tried different methods to do computing, right? And those attempts haven't succeeded yet, right? And many other companies have pivoted away. we ⁓ we have a very different approach and we can show how our system, how our architecture can scale, how we can improve the performance over time. And over time as we indeed ⁓ more and more demonstrated more and more. And then we can see that, okay, more and more people believing that we can indeed solve the problem. So all of these recognitions basically prove that, okay, as long as we have the right technology, the right roadmap, then actually we can solve the problem. Right. So these are really good. think what's really important here is that we don't really want to like over promise, right? And we say, oh, we can be 1000 times better than a GPU and we can be extremely cheap. Because I mean, there has been lots of promise already, but here, we really want to deliver exactly what we said we will, which is okay, we can still all the challenges we know exactly how to solve the problem. And I think these recognitions are really like showing that yes, more more people believe in our vision right now and they believe in our approach. ⁓
Matthias: And probably, you know, one award feeds to the next award, et cetera, right? You get more known, you're, that's fantastic.
Xianxin: Yeah, yeah, think I think overall, right, we definitely need more and more people to understand to know what optical compute or to know that in the future, right, computing can take many, different forms, right. Optical computing is a very promising approach. So I mean, the more people out there hearing this concept knowing that we are doing this, I think it helps with the adoption a lot.
Matthias: Okay, cool. Let's talk a bit about the market. I your promise when you started, you know, creating a company, promise was, look, we can reduce energy consumption by, I don't know, up to 90 % compared to traditional compute. How does it translate into a business model? What do you want to sell eventually? You want to sell hardware, you want to sell compute time? Do you want to... sell IP and then leave it to other companies. What's the business model that you want to go with your value proposition?
Xianxin: Yeah, I think at the heart, we have this really truly unique technology that can significantly reduce power consumption of compute. And this technology itself is extremely valuable. Because ⁓ today, we knew, like all the hyperscalers and people are building gigabit data centers. ⁓ And then we talking about hundreds of billion investment into this super expensive data centers and digital compute. And everyone knows that this is not sustainable. We are developing this energy efficient solution for AI, for data centers and many other applications. So over time, then we will improve our compute performance and the energy efficiency linearly. And we see this to be really unique here because we don't see any other alternative approach that can improve the energy efficiency. So as long as we can deliver this energy efficiency to be at least 10 times or even 100 times better than what you can do with digital.
Matthias: you
Xianxin: then actually there are many different business models that we can think about. main focus ⁓ to the hardware, to provide a data center compatible hardware solution. But of course we also have software integrated in our system. I think time we can also choose to provide service. We can also ⁓ basically like sell provide inference service, but then our cost will be much much lower because our system is far more energy efficient. ⁓ So I think we sell the hardware ⁓
Matthias: Thank
Xianxin: ⁓ or we sell the service, because our technology is far more energy efficient, it means that we can still deliver the cost saving, the cost benefits to the users, no matter which business model we choose.
Matthias: Got it. And the main use case is AI. If you talk about selling tokens, it's... Yeah, I am concerned so... Yeah, okay. Okay. You know, many people are excited apparently about the company. You know, if I look at all the awards that you got, you got finance now. So what is the timeline? You know, when can I buy a box from Lumai for my data center?
Xianxin: Yeah, it's AI inference, AI application, and also AI as well. Yeah, so today we already have a testbed the lab. We have been doing lots of testing and this is actually already the world's most advanced optical processor, right? But we are not planning to ship out 1,000 of these internal testbeds yet. So we will keep it internally and we will work with a few early customers and design partners on this internal testbed. we expect to start to our first commercial products in ⁓ three years time ⁓ the end of 2027 or early 2028. By that time, we will be able to ship our ⁓ processors to data centers.
Matthias: OK, OK, probably more in a pilot mode. But end of 27, beginning of 28, you want to have a box that I can put in my data center and then run.
Xianxin: Yeah, standards of a boss, everything is compatible with existing infrastructure, but just like.
Matthias: No water cooling, air-cooled data centers, everything's in the... You don't need a lot of cooling, I guess, for...
Xianxin: No, no, who will want that? No, because our system is energy efficient, right?
Matthias: Yeah, got it. I mean, you mentioned you said, look, you are the most sophisticated optical ⁓ computer today. I mean, talk a bit about competition. mean, there's traditional players like NVIDIA, there's NOAA Photonics, there's LumiCore. do you differentiate yourself against those?
Xianxin: I think because our technology can achieve really ⁓ impressive energy So initially our initial market will be those power-sensitive, power-limited applications. So for some, you're building 100 megawatt or 1 gigawatt data center, you can buy tons of GPUs from Nvidia, right? That's fine. As long as you can afford to do that. But there are also many, many other small data centers and many other users that don't really have that kind of gigawatt power capacity, right? They probably, can't really install like 100 kilowatts SilverRex. They only have like 10 kilowatts. In that case, they can't really benefit much from Nvidia GPU, right? when you at our system, we actually consume much lower power. So for those users, for those applications that are power limited, power constrained, they can actually extract much more performance from our system, from optical ⁓ systems. So we target those power-sensitive, power-constrained applications, and we ⁓ deliver much higher ⁓ capacity ⁓ digital for those users. So that's how we do it.
Matthias: ⁓ OK, but compared to other optical compute platforms, I mean, you mentioned before, there many other players. So what's difference? How do you?
Xianxin: Yeah. Yeah, to be honest, we don't really see many optical competitors out there, right? Because right now, even with our current prototype, we are far ahead of any other optics already. So with our architecture, we know that we can uniquely achieve very high energy efficiency and very high compute performance. So we just don't really see any other approach coming close to our level of performance. So there might be some other markets for those optical processors, right? It could be like edge processing. It could be like optical sensing, but that's not our focus. We focus on those large scale AI inference data center kind of compute.
Matthias: Okay, cool. That's very good. you're, I mean, obviously you doubled the team in the last, let's say 12 months since, I guess, since we invested, right? And you got sufficient capital. What's your role in, you know, building the team? What are you hiring for? What kind of, you know, profiles do you want to have? And what's the challenge, you know, in scaling? I don't know what scale you are, yeah.
Xianxin: Yeah, so right now we are still a small team, small but really focused. Right now the team size is about 20, 21. Most of them are engineers, as you can imagine. We have lots of hardware and also some software engineers at this point. We also expect to double or triple our team size next year. So yeah, most of the current team members are just engineers working on the product. Speaking of challenge, mean the biggest challenge clearly is always like talents. We just need talents, right? Yeah, I think this is the same for most startups.
Matthias: you Yeah. And how do you attract talent, right? Why would they join you and not a corporate or different startup? What do you promise them? Is it the technology that gets them excited? Is it the Is it the stock options that you're giving them? How?
Xianxin: Yeah, I think... Yeah, I think, I think... Yeah, so I think we are quite different from many other startups, right? Because we have something that's truly, truly very interesting, very revolutionary, right? So most of our optical electronic engineers, for example, they join us not because, ⁓ this is a very high salary and very good options. No, because they are just really, really interested in this technology. They really want to develop the world's first optical computers, right? That is very, very attractive. So there's no secret, really. If there is, then our security is, okay, we just offer this truly unique and really rare opportunity for those engineers, for the team members to really revolutionize compute and like change AI, right? And this is like extremely rare opportunity and that's what attracts the talent.
Matthias: Okay, so you attract them with your vision, the technological vision that you have, that they can change the world. Let's talk a bit about our partnership. So we, Constructor Capital, invested in your Series A. We were the lead investor there. Then you had a bunch of followers. Why?
Xianxin: Yeah. Yeah. Mm-hmm.
Matthias: How did you make a choice? How did you choose Constructor Capital and why did you choose us as a lead investor? You can give me the brutal truth, right?
Xianxin: Yeah. Yeah, I think working with Constructor so far has been really good, right? Because there lots of investors out there, and what we are doing is truly deep tech, right? And that's hardware. As you mentioned, this is actually notoriously difficult, right? Hardware means hard, right? But Constructor really understands deep tech, right? You guys really understand hardware. For example, you also invested in QoEra, right? Quantum computing as well. Basically, ⁓ working with Constructor really helped us a lot in ⁓ not just working on the tech itself, but actually prioritizing the technology that we have. So that's been very useful and very helpful. And Serge, for example, he himself is a very successful entrepreneur. And so he actually also shared lots of his own experience as well. And then we can also get plugged into the huge network of Constructor. So we can talk to lots of different AI companies and hardware companies as well.
Matthias: Yeah, okay. Okay. I thought actually, for me, it was quite interesting. You know, once we joined, suddenly you had so many people coming in, right? You had, you were completely oversubscribed. It shows a bit it shows a bit that you need a lead investor, because many people don't understand deep tech that well. And then there's a lot of capital once once you have the lead investor. that was what I learned from it. But okay, so we invested 10 million, ⁓ I it's a year ago. ⁓
Xianxin: Mm-hmm. Yes. Is that it? Mm-hmm.
Matthias: What will happen next? How much of that did you spend already? long does it last? When will you do series B? What do you have in mind there? ⁓
Xianxin: Yeah, just to clarify, our next round is Series 8. But yes, we are halfway there, right? So yeah, we got two years runway with our 10 million round. So our catch runway is until December, December 2026. So we are halfway there. We are planning to open our next round Series 8 in Q1, 2026, in a moments' time.
Matthias: ⁓ OK. All right.
Xianxin: far, with that funding, we doubled the team size and we ⁓ spent a lot on the tech development, on the prototype, on the test bed. ⁓ so we are now very, close to our very important milestone of ⁓ this most advanced optical processor and running actual LAMA type of transformer inference in the system. So these are very important milestones. And as I mentioned, these are far ahead of any other ⁓ in terms of the TRL, Technology Readiness Level, in terms of the performance. So with this kind of milestones achieved, then we are very confident that it will prove the powerfulness of our architecture. And then we will open the next round and then we will build a product.
Matthias: Okay, so the next round is for building a product. It's obviously much more than 10 million that you need to raise. Okay, shall we talk a bit you as a strategist? mean, you obviously ⁓ already or mentioned that optical compute is... maybe the most sustainable way to deploy AI infrastructure. But how do you expect the AI hardware landscape, the compute landscape, maybe even also software landscape to look like in five to 10 years? So are these approaches in parallel for special use cases for edge versus data centers? What is the role of quantum? Are there any other compute solutions, neuromorphic, et cetera?
Xianxin: Mm-hmm.
Matthias: So how do you see this landscape evolving?
Xianxin: Yeah, I think the first thing is that now, the cost on the infrastructure is just too high. And everyone needs to build gigabit data centers and ⁓ fill it the latest generation of NVIDIA GPU, then the cost on the infrastructure itself is just so high. It's not sustainable. Then how can you actually benefit from AI applications? ⁓ So the will have to come down significantly. And that's why energy efficient compute is extremely important because that directly reduces the cost of the infrastructure. That's one thing. Whether you use Optis to do that or other methods, sure, but you have to significantly drive down the cost of compute. And we do see OptiCompute to be basically the most promising way to achieve that goal. ⁓
Matthias: Do you think they replace traditional computer? Will it be for special use cases? How do different models play together?
Xianxin: So compute in general will become very heterogeneous, right? Because today basically everyone just use GPU for everything, for convenience basically. But when you think about productizing AI and benefiting from AI, you definitely need to think about the cost. and ROI, And therefore for certain applications, then you would consider using ASIC. You will consider using Optus because that gives you the best performance over like more general-purpose GPU. So in the future, no, people will continue to use GPU for certain case. People still probably use GPU to train their models, But when it comes to the deployment to the application, then you can use far more cost-effective solutions like optical computing that we have.
Matthias: Okay, so it's more for inference, it's maybe more for edge use cases, etc. So the hardcore training, okay, got it. Look, ⁓ last few minutes on ⁓ road ahead. You ⁓ what you up at night? What is like the biggest technical or business risk? You know, how? ⁓ What are you scared of, right? It sounds all very good. You you have a roadmap, you won awards, you got money, you're to raise series A. So what is the nightmare that comes to you when you sleep and...
Xianxin: Yeah, yeah, yeah. I don't know. mean, there are many, things that came to my mind. But a few things there. One thing is, say, in two years time, three years time, we built the world's most amazing optical computer. And you can see that, OK, the performance is absolutely amazing. It's much better than GPU. a very important question that is, How do we actually sell this to the market? Because today not too many people know of optical computers, right? They only think of GPU. They don't know any other alternatives, right? It's similar with quantum. When you think about quantum, most people, don't know what quantum computing is. Therefore, the market adoption, that market education is quite important. and it also means that for us, we are not just like working on the metrics itself, just maximizing the technical performance. No, we actually, we need to do lots of work on the user interface. or making our optical computers compatible with today's infrastructure because we don't want people to change how they program. We don't want to change how people compute today. We want people to basically treat our optical computers as a black box, just a job in replacement for electric use. Therefore, we are doing lots of work on that to ensure that people don't really need to learn physics, learn to manipulate laser beams to compute. So that's what we are doing. And I think this is very important because by the end of the day, don't really
Matthias: Yeah.
Xianxin: care about whether you use office or GPU. don't really care. They just want it to be as easy to use as possible, but also much cheaper, right? And also doesn't consume too much power. we ⁓ to make sure that we can deliver those benefits and we are hiding those benefits in the system. And they still use Python. They still use PyTorch. They ⁓ use the tools that they use today ⁓ program an optical computer. So that is something we think a lot about internally. ⁓
Matthias: Okay. Okay, so it's not the compute itself. It's the interfaces. It's the embedding in the system.
Xianxin: It's the interface. It's how... Yeah, yeah. And also making sure that ⁓ our optical system, our optical computers is compatible with today's infrastructure. We don't want people to build a new optical data center just to host optical processes, right?
Matthias: So that's your single biggest worry that you can make this work as well and that you find the right people.
Xianxin: Well, yeah, I mean, there are many, many other challenges, of course, right? Like we need to deliver the performance at a much lower cost and we will need to ensure that our software is compatible. We can support many, many models. But far as we see, there are no fundamental challenges there, right? And those are ⁓ ⁓ And as long as we can deliver performance and we see a path to achieve those performance, then we can solve ⁓ those So those are not real concerns.
Matthias: Okay, cool. So we are in 2025. said next year will be series A, 2026. How, if moving now to 2030, let's say, how does a successful Lumai look like? What would be your dream? How? Not a nightmare, but a dream. How ⁓ does it...
Xianxin: Yeah, I think first, in five years time, will people still use Nvidia GPU as ⁓ are using it today, everywhere? I that people will be to use some alternatives, some new solutions out there, otherwise. the power consumption and the cost will be like insane at that time in five years time. So that's one thing. And second, we definitely hope that by in another five years time, right, we will be able to ship our optical processes, not just like a few, but actually thousands or 10,000, 100,000 of these optical processes to real data centers. And people can start to use optical computers. They can start to program optical computers for different workloads. And they will see that, right, it's so natural to use optics, right? happened long before. So it's like today people think of Optis as like really new, very different in five years time. think people should look at Optis and say, yeah, of course you use Optis right now. So I think that will be successful.
Matthias: Look, I mean, we obviously hope as a partner and investor that this vision will come true. I'd to thank you. I I thought it was very insightful and all the best and I guess I see you soon in some other occasion. Thank you. Thank you for listening. Cheers. All the best.
Xianxin: Yeah, thank you. Yeah, thank you, Marius.
Xianxin: Yeah, so thank you, Matthias, for inviting me to this podcast. ⁓ Yeah, this is a long story. We basically started ⁓ journey like six years ago, When I first started to do the research ⁓ Oxford here, we basically, initially we came up with the idea and then... It was like almost too good to be true. And so we first basically did a very quick experiment at the university in the lab. That was like pre-pandemic. then, the result was actually better than what we imagined. And then one thing led to another. And then we just keep working on the tech, right? And then we saw lots of commercial interest. Once we our paper, then lots of interest out there. And then basically, then we decided, OK, it may be best actually to get started business. get VC funding to basically like commercialize the tech. So that was like how why we decided to Luma from our research.
Matthias: And how long did it take you from doing the scientific experiment, publishing, to creating a company? What's the time span?
Xianxin: So initially we the paper, right? And then we decided to basically ⁓ the idea at a bigger scale before we started the business. So ⁓ we demonstrated the system at a much bigger scale and it still worked really well. And that's why we decided to start Lumai So ⁓ have been working on this tech for like about three years and then we started Lumai, ⁓ also three years ago.
Matthias: Okay, so it took a number of years, got it. Okay. And then when you did this shift, you know, and moved from being a researcher to a CEO, what was the biggest surprise for you on a personal level, but maybe also on a technical level? You know, what, how do you describe that shift from one to...
Xianxin: Yeah. Yeah, I think the transition has been really ⁓ like very rewarding, right? Because when we were working at the university, we only have like a few people, right? A few PhD students, a few colleagues, and then we work on everything. And now actually, we started the business, we actually have a much bigger team. We have much more resources and support, right? And so actually we can do things at a much faster pace. So this is actually really amazing. We didn't really expect that we can achieve so much in such short time. at the university. So I think ⁓ as a researcher, you have your freedom. You can pick whatever you want to do. You don't have any pressure. But then once we started the business, yes, we have lots of resources. We have lots of support. But then we also have a much, much, much more well-defined goal because we have to basically connect the technology to the application to the problem. I think in this sense, we don't really have too much freedom, but You definitely know why you doing this. You definitely know why you want to do it because you want to solve a specific problem So in that sense, this is very different from research But I think this is actually much much better as compared to I just doing random research
Matthias: Okay. I hope you didn't do random research, very focused. Okay, well, maybe maybe for our listeners, let's talk a bit about, you know, the technology and the breakthrough. I mean, if if you look into literature, you know, for many, many years, or even decades, people talked about optical compute, that companies have struggled to
Xianxin: Random but you still got publications, right?
Matthias: move from concept to something that works in prototype or even at scale. is the difference? How Lumai make this What is the technical background behind it?
Xianxin: Yeah, yeah, yeah. So once again, the reason we started LUMi was because we demonstrated our technology at the university lab, and then we saw that, OK, there are other spin-offs, photon spin-offs outside there. and they haven't really achieved what we did. But then they can still like raise lots of funding and then we say, okay, if they can do that for sure, we are far better than yes, we should start a business, right? So that's how we started. back to your question, ⁓ right? ⁓ We are using ⁓ and optical effects to do compute, right? ⁓ in traditional digital domain, you need to use lots of transistors to compute, right? The optical equivalent of those transistors are like optical beams. The more optical beams that we can process, the more compute we can do, right? And then our system, our approach is fundamentally different from other like photonic chip based approach. Photonic chips are basically, okay, you still do compute on a 2D chip, but you change from electrons to photons. Therefore you have some benefits, but doesn't really scale, right? Our system does processing in the 3D space, three dimensional volume instead of a 2D chip. Therefore naturally we can process many, many more beams than what you can do with it. other approach. So that is how our system can scale much, better than any others. And that is actually fundamental to all the benefits, all the compute advantages in our system.
Matthias: sounds intuitive, right? I mean, you move from a chip to three dimensional. what are the challenges if you do that?
Xianxin: Mm-hmm. Mm-hmm. Yeah, today it sounds intuitive, right? But when we started, actually, most people still wanted to do the photon approach because that also has lots of benefits, right? It's tiny, it's still compatible with SCMOS, right? So you will want to continue to compute that way. But nature is very different, right? Our brain, our brain use neurons that are connected in three-dimensional volume, So nature is very fundamental different from digital. So in the past, most compute people tend to tend to think of computed still with the chip based approach. So right now we say, yeah, you can do more compute in the 3D space, but that's not natural for compute people at all. So, in terms of life challenges, there are lots of things there, right? Because, okay, you can, the more, basically you want to process as many laser beams as possible in ⁓ a physical volume, and that requires us to come up with very, very good optical design and with very good optical components. So in the past, people can demonstrate this kind of idea with a few laser beams, but that's just not enough to output from digital. You will need to do millions of laser beams operating at very high clock rate. So, I mean, if you really want to up perform digital computer system, you need extremely good optical design. You need extremely good co-design of both optics and electronic systems. So that has been very challenging, but once again, we have our own unique data encoded methods and some other patterns for us to basically overcome those challenges. That's why we can actually achieve very, good compute performance today.
Matthias: Okay. And how, I me, how do you, I mean, I understand you create millions of laser beams, get this, them in three dimensional space. So that's, that's also, but how do we do, ⁓ mean, I understand the main product that you're doing is matrix multiplication. ⁓ do you do, you know, with laser beams then, ⁓ know, how do you make them interact with each other so you can do matrix ⁓
Xianxin: Mm-hmm. Yeah. Yeah, yeah, yeah. So ⁓ basically, we are not trying to use optics to create a CPU or GPU, right? So what we have is very different. So we basically use nature. We use those natural optical effects, like interference and deflection. We use these optical effects to perform a ⁓ very specific matrix operation, just matrix multiplication. And it turns out that this kind of operation, ⁓ you do it with optics, it's very simple. It's very energy efficient. What's even better is that these matrix multiplication are very fundamental and they are used extensively under all kinds of neural networks. So that's what we do. So back to your question, we use interference, we use this optical effects and lenses and fibers, these components to perform exactly what optics is good at.
Matthias: Got it. it. Okay, so you need you need lenses, you need sensors, etc. ⁓ If I get it correctly, you started actually as a optics, PhD or student or expert. How, you know, how did help? I mean, this is not a quantum computer, I understand. was a traditional computer. So what's the relationship and how did your quantum optical quantum expertise help in?
Xianxin: Mm-hmm. Yep. No. Yeah, yeah. Indeed, I did my PhD in quantum optics. And after five years, I decided to work on something that's easier. And that's why we decided to build optical computers for AI. So if you look at the system that we have, we actually use lots of those similar components and techniques that people will apply in those quantum systems as well. But what's very different here is that we don't really need use like quantum effects, right? Quantum effects are very fragile and they are very sensitive to any environmental perturbation. So we don't need any of those quantum effects. Therefore, our is actually far more robust. It's actually much easier us to scale our performance and for AI. So that actually is very important. So we leverage the technologies, those really advanced optics and electronic technologies that people would ⁓ use in systems. But then we don't have high level of sensitivity and environmental perturbation. ⁓ So we benefit from that. And ⁓ using technologies, we can scale our system. We know how exactly we can ⁓ improve performance. And ⁓ at same time, we don't really need to continue to do very, hard engineering work for another five or 10 years. So ⁓ our computers can be ready basically ⁓ in years'
Matthias: got it but then they don't have the effect that quantum computers have that they scale exponentially they scale more linear
Xianxin: We scale more linearly, but what's more important here is that quantum computer, they can be extremely useful for certain tasks, right? And what we can unlock here is to compute energy efficiency, right? So we can solve today's AI problems and we can use our optical computers in data centers, and we can reduce the power consumption. This is actually very important because normally don't really use quantum computers for these applications, right? They are at their own domains.
Matthias: I I see. Okay. I want to shift a bit and we want to talk about, you know, from lab to company. We touched on that previously a bit. you ⁓ Lumai together with your colleagues, James Bale. What do the two of you bring? How are you complementary? How do you as a partnership? What's your different backgrounds? How does it work as two co-founders?
Xianxin: Yeah, we together with James, we have been working together for six years since the university. And I will say, we basically co-developed the technology, right? We published lots of papers together, we applied for patents, and then we started Lumai from our research. So both of us have technical backgrounds. We understand optics, physics, and AI really well. But the difference here is that Basically, at the university, I was a postdoc, designing the project and thinking ahead. And James was a PhD student. And then he basically implemented all of those He demonstrated the system. He in the lab, like day and night. ⁓ So the that we work is basically, most of the time, I am a thinker. I plan ahead. I plan for the project. And then I look ahead and say, OK, what's the next step? What's the roadmap? And then James execute the plan. And he can basically demonstrate all the technologies, all the... all the inventions really, really well. So he executes the plan, right? And that goes so far ⁓ well. ⁓ now he's CTO in the business and he worked very closely with ⁓ engineering team to basically build our product. ⁓ then I'm a CEO and then I ⁓ ahead on the product market fit, et cetera.
Matthias: Okay, so you're basically first wave and then he's basically making it happen in as a CTO. it. Okay. ⁓ know, your startup that works obviously at the intersection of hardware ⁓ software, I guess. ⁓ hardware is very capital intensive. Many, many startups shy away from hardware. What the ⁓ main you had when you had, you know, to turn from a lab prototype into something that
Xianxin: Yes. Mm-hmm.
Matthias: can be manufactured eventually at scale at this stage still as a prototype.
Xianxin: Yeah, I think in terms of manufacturing, so. the one of the biggest challenge is like on the components and on the supply side. when we decided to use optics to do computing, we initially, well, even now we see a major benefit there because there's a huge optical ecosystem, optical supply chain out there, right? Lots of those optical components ⁓ surprise optical manufacturers that we can work with. ⁓ see that as a huge benefit that we can leverage ⁓ do computing, right? So we work with not just the electronic industry, also optical industry. So that gives us really a very strong advantage there. But in the near term, right now, we have a small team and we are a startup. And naturally, we just need to deal with lots of these optical supplies. And we will need to work on many, many components. So that actually takes lots of time. So we will need to work with them to specify the components and then to test the reliability, to test the matches. So that has been taking time. But I think over time, then we will still benefit a lot from this, because at least we have lots of supplies out there. we can work with. if we don't, then we will have to ⁓ many of these materials or components, and we definitely want to do that. ⁓
Matthias: Okay, so most of the components are off the shelf and you buy them and it's more about the integration.
Xianxin: yeah, most of them out of the shell, but of course we will need to re-engineer them, right? The good thing is we don't need to do it all by ourselves. We can work with a very large supply chain and ecosystem.
Matthias: Okay, cool. I mean, the other thing that happened, probably more recent, I you got a bunch of awards recently, I think you're on the Photonics 100 list and just something like a week or two weeks ago, you won or the company won the Falling Walls Foundation Award 2025. You know, how... How does it help the company? How does it validate the achievements, the technical work? What does it mean for LUMEI? ⁓
Xianxin: Yep. ⁓ Yeah, I think, yeah, indeed over the past like 18 months, we have been receiving lots of recognition, right? And I think this is a very good sign because in the past people have tried different methods to do computing, right? And those attempts haven't succeeded yet, right? And many other companies have pivoted away. we ⁓ we have a very different approach and we can show how our system, how our architecture can scale, how we can improve the performance over time. And over time as we indeed ⁓ more and more demonstrated more and more. And then we can see that, okay, more and more people believing that we can indeed solve the problem. So all of these recognitions basically prove that, okay, as long as we have the right technology, the right roadmap, then actually we can solve the problem. Right. So these are really good. think what's really important here is that we don't really want to like over promise, right? And we say, oh, we can be 1000 times better than a GPU and we can be extremely cheap. Because I mean, there has been lots of promise already, but here, we really want to deliver exactly what we said we will, which is okay, we can still all the challenges we know exactly how to solve the problem. And I think these recognitions are really like showing that yes, more more people believe in our vision right now and they believe in our approach. ⁓
Matthias: And probably, you know, one award feeds to the next award, et cetera, right? You get more known, you're, that's fantastic.
Xianxin: Yeah, yeah, think I think overall, right, we definitely need more and more people to understand to know what optical compute or to know that in the future, right, computing can take many, different forms, right. Optical computing is a very promising approach. So I mean, the more people out there hearing this concept knowing that we are doing this, I think it helps with the adoption a lot.
Matthias: Okay, cool. Let's talk a bit about the market. I your promise when you started, you know, creating a company, promise was, look, we can reduce energy consumption by, I don't know, up to 90 % compared to traditional compute. How does it translate into a business model? What do you want to sell eventually? You want to sell hardware, you want to sell compute time? Do you want to... sell IP and then leave it to other companies. What's the business model that you want to go with your value proposition?
Xianxin: Yeah, I think at the heart, we have this really truly unique technology that can significantly reduce power consumption of compute. And this technology itself is extremely valuable. Because ⁓ today, we knew, like all the hyperscalers and people are building gigabit data centers. ⁓ And then we talking about hundreds of billion investment into this super expensive data centers and digital compute. And everyone knows that this is not sustainable. We are developing this energy efficient solution for AI, for data centers and many other applications. So over time, then we will improve our compute performance and the energy efficiency linearly. And we see this to be really unique here because we don't see any other alternative approach that can improve the energy efficiency. So as long as we can deliver this energy efficiency to be at least 10 times or even 100 times better than what you can do with digital.
Matthias: you
Xianxin: then actually there are many different business models that we can think about. main focus ⁓ to the hardware, to provide a data center compatible hardware solution. But of course we also have software integrated in our system. I think time we can also choose to provide service. We can also ⁓ basically like sell provide inference service, but then our cost will be much much lower because our system is far more energy efficient. ⁓ So I think we sell the hardware ⁓
Matthias: Thank
Xianxin: ⁓ or we sell the service, because our technology is far more energy efficient, it means that we can still deliver the cost saving, the cost benefits to the users, no matter which business model we choose.
Matthias: Got it. And the main use case is AI. If you talk about selling tokens, it's... Yeah, I am concerned so... Yeah, okay. Okay. You know, many people are excited apparently about the company. You know, if I look at all the awards that you got, you got finance now. So what is the timeline? You know, when can I buy a box from Lumai for my data center?
Xianxin: Yeah, it's AI inference, AI application, and also AI as well. Yeah, so today we already have a testbed the lab. We have been doing lots of testing and this is actually already the world's most advanced optical processor, right? But we are not planning to ship out 1,000 of these internal testbeds yet. So we will keep it internally and we will work with a few early customers and design partners on this internal testbed. we expect to start to our first commercial products in ⁓ three years time ⁓ the end of 2027 or early 2028. By that time, we will be able to ship our ⁓ processors to data centers.
Matthias: OK, OK, probably more in a pilot mode. But end of 27, beginning of 28, you want to have a box that I can put in my data center and then run.
Xianxin: Yeah, standards of a boss, everything is compatible with existing infrastructure, but just like.
Matthias: No water cooling, air-cooled data centers, everything's in the... You don't need a lot of cooling, I guess, for...
Xianxin: No, no, who will want that? No, because our system is energy efficient, right?
Matthias: Yeah, got it. I mean, you mentioned you said, look, you are the most sophisticated optical ⁓ computer today. I mean, talk a bit about competition. mean, there's traditional players like NVIDIA, there's NOAA Photonics, there's LumiCore. do you differentiate yourself against those?
Xianxin: I think because our technology can achieve really ⁓ impressive energy So initially our initial market will be those power-sensitive, power-limited applications. So for some, you're building 100 megawatt or 1 gigawatt data center, you can buy tons of GPUs from Nvidia, right? That's fine. As long as you can afford to do that. But there are also many, many other small data centers and many other users that don't really have that kind of gigawatt power capacity, right? They probably, can't really install like 100 kilowatts SilverRex. They only have like 10 kilowatts. In that case, they can't really benefit much from Nvidia GPU, right? when you at our system, we actually consume much lower power. So for those users, for those applications that are power limited, power constrained, they can actually extract much more performance from our system, from optical ⁓ systems. So we target those power-sensitive, power-constrained applications, and we ⁓ deliver much higher ⁓ capacity ⁓ digital for those users. So that's how we do it.
Matthias: ⁓ OK, but compared to other optical compute platforms, I mean, you mentioned before, there many other players. So what's difference? How do you?
Xianxin: Yeah. Yeah, to be honest, we don't really see many optical competitors out there, right? Because right now, even with our current prototype, we are far ahead of any other optics already. So with our architecture, we know that we can uniquely achieve very high energy efficiency and very high compute performance. So we just don't really see any other approach coming close to our level of performance. So there might be some other markets for those optical processors, right? It could be like edge processing. It could be like optical sensing, but that's not our focus. We focus on those large scale AI inference data center kind of compute.
Matthias: Okay, cool. That's very good. you're, I mean, obviously you doubled the team in the last, let's say 12 months since, I guess, since we invested, right? And you got sufficient capital. What's your role in, you know, building the team? What are you hiring for? What kind of, you know, profiles do you want to have? And what's the challenge, you know, in scaling? I don't know what scale you are, yeah.
Xianxin: Yeah, so right now we are still a small team, small but really focused. Right now the team size is about 20, 21. Most of them are engineers, as you can imagine. We have lots of hardware and also some software engineers at this point. We also expect to double or triple our team size next year. So yeah, most of the current team members are just engineers working on the product. Speaking of challenge, mean the biggest challenge clearly is always like talents. We just need talents, right? Yeah, I think this is the same for most startups.
Matthias: you Yeah. And how do you attract talent, right? Why would they join you and not a corporate or different startup? What do you promise them? Is it the technology that gets them excited? Is it the Is it the stock options that you're giving them? How?
Xianxin: Yeah, I think... Yeah, I think, I think... Yeah, so I think we are quite different from many other startups, right? Because we have something that's truly, truly very interesting, very revolutionary, right? So most of our optical electronic engineers, for example, they join us not because, ⁓ this is a very high salary and very good options. No, because they are just really, really interested in this technology. They really want to develop the world's first optical computers, right? That is very, very attractive. So there's no secret, really. If there is, then our security is, okay, we just offer this truly unique and really rare opportunity for those engineers, for the team members to really revolutionize compute and like change AI, right? And this is like extremely rare opportunity and that's what attracts the talent.
Matthias: Okay, so you attract them with your vision, the technological vision that you have, that they can change the world. Let's talk a bit about our partnership. So we, Constructor Capital, invested in your Series A. We were the lead investor there. Then you had a bunch of followers. Why?
Xianxin: Yeah. Yeah. Mm-hmm.
Matthias: How did you make a choice? How did you choose Constructor Capital and why did you choose us as a lead investor? You can give me the brutal truth, right?
Xianxin: Yeah. Yeah, I think working with Constructor so far has been really good, right? Because there lots of investors out there, and what we are doing is truly deep tech, right? And that's hardware. As you mentioned, this is actually notoriously difficult, right? Hardware means hard, right? But Constructor really understands deep tech, right? You guys really understand hardware. For example, you also invested in QoEra, right? Quantum computing as well. Basically, ⁓ working with Constructor really helped us a lot in ⁓ not just working on the tech itself, but actually prioritizing the technology that we have. So that's been very useful and very helpful. And Serge, for example, he himself is a very successful entrepreneur. And so he actually also shared lots of his own experience as well. And then we can also get plugged into the huge network of Constructor. So we can talk to lots of different AI companies and hardware companies as well.
Matthias: Yeah, okay. Okay. I thought actually, for me, it was quite interesting. You know, once we joined, suddenly you had so many people coming in, right? You had, you were completely oversubscribed. It shows a bit it shows a bit that you need a lead investor, because many people don't understand deep tech that well. And then there's a lot of capital once once you have the lead investor. that was what I learned from it. But okay, so we invested 10 million, ⁓ I it's a year ago. ⁓
Xianxin: Mm-hmm. Yes. Is that it? Mm-hmm.
Matthias: What will happen next? How much of that did you spend already? long does it last? When will you do series B? What do you have in mind there? ⁓
Xianxin: Yeah, just to clarify, our next round is Series 8. But yes, we are halfway there, right? So yeah, we got two years runway with our 10 million round. So our catch runway is until December, December 2026. So we are halfway there. We are planning to open our next round Series 8 in Q1, 2026, in a moments' time.
Matthias: ⁓ OK. All right.
Xianxin: far, with that funding, we doubled the team size and we ⁓ spent a lot on the tech development, on the prototype, on the test bed. ⁓ so we are now very, close to our very important milestone of ⁓ this most advanced optical processor and running actual LAMA type of transformer inference in the system. So these are very important milestones. And as I mentioned, these are far ahead of any other ⁓ in terms of the TRL, Technology Readiness Level, in terms of the performance. So with this kind of milestones achieved, then we are very confident that it will prove the powerfulness of our architecture. And then we will open the next round and then we will build a product.
Matthias: Okay, so the next round is for building a product. It's obviously much more than 10 million that you need to raise. Okay, shall we talk a bit you as a strategist? mean, you obviously ⁓ already or mentioned that optical compute is... maybe the most sustainable way to deploy AI infrastructure. But how do you expect the AI hardware landscape, the compute landscape, maybe even also software landscape to look like in five to 10 years? So are these approaches in parallel for special use cases for edge versus data centers? What is the role of quantum? Are there any other compute solutions, neuromorphic, et cetera?
Xianxin: Mm-hmm.
Matthias: So how do you see this landscape evolving?
Xianxin: Yeah, I think the first thing is that now, the cost on the infrastructure is just too high. And everyone needs to build gigabit data centers and ⁓ fill it the latest generation of NVIDIA GPU, then the cost on the infrastructure itself is just so high. It's not sustainable. Then how can you actually benefit from AI applications? ⁓ So the will have to come down significantly. And that's why energy efficient compute is extremely important because that directly reduces the cost of the infrastructure. That's one thing. Whether you use Optis to do that or other methods, sure, but you have to significantly drive down the cost of compute. And we do see OptiCompute to be basically the most promising way to achieve that goal. ⁓
Matthias: Do you think they replace traditional computer? Will it be for special use cases? How do different models play together?
Xianxin: So compute in general will become very heterogeneous, right? Because today basically everyone just use GPU for everything, for convenience basically. But when you think about productizing AI and benefiting from AI, you definitely need to think about the cost. and ROI, And therefore for certain applications, then you would consider using ASIC. You will consider using Optus because that gives you the best performance over like more general-purpose GPU. So in the future, no, people will continue to use GPU for certain case. People still probably use GPU to train their models, But when it comes to the deployment to the application, then you can use far more cost-effective solutions like optical computing that we have.
Matthias: Okay, so it's more for inference, it's maybe more for edge use cases, etc. So the hardcore training, okay, got it. Look, ⁓ last few minutes on ⁓ road ahead. You ⁓ what you up at night? What is like the biggest technical or business risk? You know, how? ⁓ What are you scared of, right? It sounds all very good. You you have a roadmap, you won awards, you got money, you're to raise series A. So what is the nightmare that comes to you when you sleep and...
Xianxin: Yeah, yeah, yeah. I don't know. mean, there are many, things that came to my mind. But a few things there. One thing is, say, in two years time, three years time, we built the world's most amazing optical computer. And you can see that, OK, the performance is absolutely amazing. It's much better than GPU. a very important question that is, How do we actually sell this to the market? Because today not too many people know of optical computers, right? They only think of GPU. They don't know any other alternatives, right? It's similar with quantum. When you think about quantum, most people, don't know what quantum computing is. Therefore, the market adoption, that market education is quite important. and it also means that for us, we are not just like working on the metrics itself, just maximizing the technical performance. No, we actually, we need to do lots of work on the user interface. or making our optical computers compatible with today's infrastructure because we don't want people to change how they program. We don't want to change how people compute today. We want people to basically treat our optical computers as a black box, just a job in replacement for electric use. Therefore, we are doing lots of work on that to ensure that people don't really need to learn physics, learn to manipulate laser beams to compute. So that's what we are doing. And I think this is very important because by the end of the day, don't really
Matthias: Yeah.
Xianxin: care about whether you use office or GPU. don't really care. They just want it to be as easy to use as possible, but also much cheaper, right? And also doesn't consume too much power. we ⁓ to make sure that we can deliver those benefits and we are hiding those benefits in the system. And they still use Python. They still use PyTorch. They ⁓ use the tools that they use today ⁓ program an optical computer. So that is something we think a lot about internally. ⁓
Matthias: Okay. Okay, so it's not the compute itself. It's the interfaces. It's the embedding in the system.
Xianxin: It's the interface. It's how... Yeah, yeah. And also making sure that ⁓ our optical system, our optical computers is compatible with today's infrastructure. We don't want people to build a new optical data center just to host optical processes, right?
Matthias: So that's your single biggest worry that you can make this work as well and that you find the right people.
Xianxin: Well, yeah, I mean, there are many, many other challenges, of course, right? Like we need to deliver the performance at a much lower cost and we will need to ensure that our software is compatible. We can support many, many models. But far as we see, there are no fundamental challenges there, right? And those are ⁓ ⁓ And as long as we can deliver performance and we see a path to achieve those performance, then we can solve ⁓ those So those are not real concerns.
Matthias: Okay, cool. So we are in 2025. said next year will be series A, 2026. How, if moving now to 2030, let's say, how does a successful Lumai look like? What would be your dream? How? Not a nightmare, but a dream. How ⁓ does it...
Xianxin: Yeah, I think first, in five years time, will people still use Nvidia GPU as ⁓ are using it today, everywhere? I that people will be to use some alternatives, some new solutions out there, otherwise. the power consumption and the cost will be like insane at that time in five years time. So that's one thing. And second, we definitely hope that by in another five years time, right, we will be able to ship our optical processes, not just like a few, but actually thousands or 10,000, 100,000 of these optical processes to real data centers. And people can start to use optical computers. They can start to program optical computers for different workloads. And they will see that, right, it's so natural to use optics, right? happened long before. So it's like today people think of Optis as like really new, very different in five years time. think people should look at Optis and say, yeah, of course you use Optis right now. So I think that will be successful.
Matthias: Look, I mean, we obviously hope as a partner and investor that this vision will come true. I'd to thank you. I I thought it was very insightful and all the best and I guess I see you soon in some other occasion. Thank you. Thank you for listening. Cheers. All the best.
Xianxin: Yeah, thank you. Yeah, thank you, Marius.
