Xanadu Quantum Technologies Limited (XNDU) Earnings Call Transcript & Summary

Welcome to our 2026 road map presentation. Thank you very much for coming here kind of in the last minute and also in August. I am just going to flash this in front of you to remind you, there will be forward-looking statements that are accurate only of the date given out. The disclaimers will be provided at this -- the other thing is this slide presentation will be available on our website at 4:15, so no need to frantically take screen shots. And you can read disclaimers at 4:15, if you so desire. And I think with that, I'd like to invite up Christian.

It's good to warm up the audience. Thank you. So thank you, Brett, and welcome, everyone. Really excited today to be here to tell you about our road map particularly and a few other things, both hardware and software road maps. And it should be a lot of fun and particularly looking forward to hearing any questions you may have as well. And we've got a really good turnout, which is awesome to see. So thank you, everyone, for making the trip today. So the speakers today, myself, Raf and Michael. We always like to start off with our mission. It's guided us from very early days from day 1, and it's to build quantum computers that are useful and available to people everywhere. And we often come back to this because a lot goes on during the day, during the weeks and months, and you have to remain focused. And that's one of the characteristics of Xanadu. We're focused on building a large-scale quantum computer. So fundamentally, we're a quantum hardware company, really exemplified by our use of photonics to build our quantum computer. And it's useful through the fact that we're building a large-scale quantum computer. All the stuff that you hear about, fault tolerance, error correction, many, many logical qubits and also the cloud deployed in nature where PennyLane is also wrapping together everything here. Okay. So this is what some of the slides, including this one, is what we presented, I think, 5 months ago now at our first Analyst Day. And a lot of that, obviously, the past stuff has remained the same. I'll just go over it fairly quickly. The hardware is on top, software is on the bottom. First chip fabricated in 2017. And then we had our X-Series, Borealis, and then also the quantum supremacy on that one and then Aurora. And then also, first, on-chip photonic GKPQ. So GKP qubits. Don't worry about the name too much, it just means Xanadu's photonic qubit. Something you may have seen on Friday, along with the CAD 195 million announcement from the Canadian government, is Inception. And so Inception is a nearly 160,000 square feet Toronto manufacturing facility that was announced, I guess, 3 days ago, and we'll talk more about that. And then from the software side of things, PennyLane, obviously, a big market there. It was released 8 years ago, launched the world's first commercial photonic platform, so you could access quantum computers over the cloud. And then you'll see some of the next stuff that we'll talk about in the coming slides, including rearchitecturing PennyLane scale, very important thing. You just can't have PennyLane from day 1 accessing and building out scalable quantum computers. So this was announced this is converted into into USD. It was CAD 195 million from the federal government for our project that we call OPTIMISM. It was executed a few days ago, announced on Friday. We had one of the main ministers, Minister Solomon for AI in Canada. He was there to announce this. Really excited about this, this is what we alluded to 5 months ago. This is part of the money. This is the federal side of things. And stay tuned for any announcements soon related to the provincial side of things. This one was announced first because it was in discussions initially. Use of proceeds, established advanced manufacturing facility. That's why we announced Inception, which I'll talk more about soon. And the cool thing here in terms of finances offsets a substantial portion of the capital for our next phase of development. Okay. We have a track record. It's all about hardware in the quantum industry. It's all about our hardware road map, which we'll talk about. We're a photonic-based company. Pros and cons of that coming up soon, which I'm sure you're all familiar with. But if you check out these 4 Nature papers in the last few years, this is our X-Series, Borealis, which showed quantum supremacy. Aurora, which showed how to network quantum computers together. Remember, the goal is to have smaller well-performance quantum computers, and you have hundreds of them and network them together. This is the only demonstration was published in Nature by any company, which shows how to solve that on a scalable -- in a scalable manner. And then the world's first demonstration on-chip of our GKP qubit, our photonic qubit as well. All of them, as mentioned has made it into Nature. This is more of a reference point. If you were at our Analyst Day, our first one last time, this is the same slide here. We basically compare ourselves to other approaches. So you've got the photonic-based approaches, Xanadu's and other photonics, and you've got the electronic-based ones, which are the other 4 columns there. Pros and cons of each, I'll just quickly go through some highlights. We've demonstrated scalable interconnects, meaning that we had 4 server racks at our network together, could have been 40 or could have been 400. We know to actually how to do that at scale. The clock speed, that's something if you want to keep an eye out over the next year or 2 is going to matter more and more. This is where photonics really demonstrates one of its extreme values where we're orders are magnitude faster than a lot of the other approaches, which is really key. So one way to look at that is we benchmark -- let's say, it took a year to solve one problem for the electronic-based approaches. Whatever that problem is, just abstractly, 365 days to solve it. That's incredible because it would have taken hundreds of years to do it. So you're well ahead. Our computer, because of the clock speed and using photonics, would solve that same problem in 8 hours. So you go from a year, which is already in the promise of quantum computing, but we're better in terms of that, significantly better. The intuition is we're operating at the speed of light. Logical overhead. This is a really key one. One of the things that we decided to do today is go over some of the basics on how to think about quantum computing despite what we may hear in the media and from other companies. And so that's what you'll see. You'll see some basic things about why logical qubits is one important thing, but there are other important things. Hopefully, this helps in the education. The logical overhead is always a challenging one. For us, we believe it's on the order of 100 to 1. So 100 [ crappy ] physical qubits gives you 1 very good or logical qubit. Now everyone, including us, to be fair, we try to be as balanced as possible. We've said we've got 12 logical qubits. We do, but the devil is in the details, and there's many other things to really care about. But it's still an important first step to get some logical qubits. Others are saying -- other companies are maybe saying 2 to 1 or 20 to 1 or something like that. I can categorically tell you that's not true. Now if it is true, then it also applies to us, meaning that maybe 21 is fine, but the logical error rate is disastrous, or maybe it's only correcting a certain subset of areas. Basically, in order to make real money and in order to have significant customer adoption and impact, you need to have these numbers here. And for us, we actually have roughly an order of magnitude better overhead there. Less overhead means less capital means less hardware. So that's a huge benefit for us. 2-qubit gate fidelity, we're up there with the best, I think, tied in this case. Our 2-qubit gate is actually a beam split, or one of the most famous optical elements ever. And so that becomes our 2-qubit gate. Connectivity is all-to-all, along with a few of the others. Room temperature computation. This one is always challenging to hear. We actually have seen other decks by other competitors. I love looking through them and engaging where they're at, particularly the public companies. They will have -- everyone has these crosses and all of that. And we recommend everyone to dive into this and see who's been truthful about it. Often, you'll see across neophotonics, but they're talking about, say, [ PsiQuantum ] or someone else. But ours actually operates. Compute is the right terminology, meaning the logic gates and the qubits and the measurements are all done at room temperature. Now I don't mean cryogenics, and I don't mean laser cooling either. Cryogenics operates around and 10 to the minus 3. These accruing actually operates at 10 to the minus 6, so 1,000x colder than the others using laser cooling. It's kind of a cool technology, literally cool technology. And so the room temperature side of things, we truly operate. Now the others, photonics, not all photonics are created equally. They need -- other photonic-based approaches need to have cryogenics everywhere. That adds cost, iteration times slow down. Superconductors, they have the famous chandeliers. Iron traps and neutral atom use laser cooling, so they don't have cryogenics, but they have laser cooling. They still need to go down, call this, 10 to the minus 6 for that sort of thing. And they sometimes say they're room temperature. I don't know how, maybe it's like saying that some of their part is outside, but the fridge -- yes, it's hard to sort of justify. But the photonics technology and road map, we all have that aspect of things. You're going to need some way of scaling up to do that. Just to be fair, we try to be as balanced as possible. Superconductors, some of them, like, say, IBM are looking to do these kind of tubes. So you got a chandelier here and a chandelier here, and they have like a tube [ coax ] cable where they actually are not sharing photonics, they're sharing still whatever is in the electronics, the [ juries injunctions ] and the qubits credit from that between it. So they're looking to scale up there. Neutral atoms are talking about shuttling things around. That's how they get all-to-all connectivity, but they still need to have photonics if they want to go beyond that once they kind of max that out. Just some of the benefits, I think I've mentioned them. Superconducting, significant cooling and networking requirements, lower gate speeds. I mentioned all these sorts of things, but you're welcome to ask me any questions after that. Okay. This is one of those kind of educational slides. And I don't mean to be kind of demeaning or anything like that, but people have different definitions. So we want to kind of get back to the basics on some of these slides. So perhaps you folks already know about it. I'm sure you know about this one. Basically, you can build a qubits using photonics, which is what we're doing. Or not photonics, electronics, roughly speaking. Pros and cons of both that I mentioned before. And so we incurred our quantum states into a physical beam of light, and it's at room temperature. This is one of our earlier chips, and you can see the size against -- I think it was a thumb. And the other based approach is superconducting trapped ion neutral atoms, spin silicon qubits and other photonic-based ones are listed there, and they all face their pros and cons. The other thing we like to get across is we think there's a lot of great approaches and a lot of great people out there. We believe we'll be one of the winners, but we don't think it is going to be a winner-takes-all market. It's hard to really think about too many industries where it's a winner take all. Really, AI is not going to be like that. Automobile, 100 years ago, it wasn't like that. Maybe some examples, like AWS is an example of winner take all to some extent, and maybe Google Search. But for the most part, very, very rare, and we think that we'll be one of the winners and then maybe like 3 to 5 winners, at least this decade, that can kind of scale up to a large-scale quantum computer. Beyond that, it's always hard to -- becomes less about the technology, more about traditional business metrics and strategies. So the benefits of us, I've talked about it already, no cooling required for computation. We do need some cooling, that's the key aspect, but we don't need it for the qubits, the gates and the measurements. They're all at room temperature. We do need some cooling, basically, think of it as turning on the computer or initializing it. They help prepare the qubits. But because they're entangled, the qubit ultimately never sees any cooling whatsoever. Manufacturing, we're using silicon processes. So every quantum company that I've seen, they all say, including us, say, we're CMOS-compatible. That, to my mind, that just means you can use foundries out there. Now I don't think anyone is lying with that comment when you see all the ticks there. But there's a difference between can you in principle use TSMC or GlobalFoundries and actually using it. It's a very different thing. With us, we're actually using GlobalFoundries, Tower, UMC and others because they're very, very, very picky about the tools they use, the processes, contamination, materials and substrates. But when we go to these large foundries that we're already using for high-volume manufacturing, they will take our chip designs, we can e-mail them to them. Now the other approach is, to my knowledge, you can't sort of send an ion trap or superconducting qubit [ Joseph injunction ] to TSMC and they'll make it. They can, in principle. It's just we're not their biggest customers. They want -- they're using -- they're addressing AWS, Apple NVIDIA, everyone. And so you want to make it so it's so easy for them to say yes, and we have that huge benefit. Error-correction flexibility. This is a key one we don't talk about enough. But other approaches -- not all -- most approaches have to already assume a particular architecture for their chips. We're like, we are working on, say, LDPC codes, but we leave that to the end. Our architecture, the designs and the hardware don't really care what choice of error correction you use. And so what that means is we can save time and money by having a more generic system. And that comes back because we're using photonics and we have this all-to-all connectivity. And that's going to become more and more important over the coming years. Compatible with telecom, we didn't have to invent the laser or optical elements, fiber optics and so forth. Faster clock speeds, I mentioned. And the modular network ability, I mentioned as well. I'm going to get to it, but I'd like to be -- like he's imperfect, I'd like to be as balanced as possible. Our biggest challenge is loss. So we talked about how amazing photonics is. Our biggest challenge is loss. So just keep that in mind, and I'm going to come back to that. So let me start down the bottom. So how do we -- every modality is judged on the same fundamentals: fewer errors, lower error rates and more logical qubits. So today, that's what we would like to sort of let people know that that's what we think are the biggest indicators. You often hear originally about things like physical qubit account, logical qubit count and all those. We think these are the 3 biggest indicators. Errors, for us, that's loss. Remember, everyone has errors. It's just ours is loss. So there's nothing unique about us, except we're using photonics, and loss is what's causing the errors. Everyone has this. Everyone needs to have a logical error rate, and everyone needs to have logical qubits. If you just -- you don't need to worry about physical qubits, but you don't need to worry about co-distance and many other things. you just sort of say to a company, including us, how many logical qubits do you have? Well, we're used to that now. You ask them what their error rate is. And then you also asked them about what type of errors that they have. And these are the 3 things to kind of be aware of essentially. So logical error rate, you can think of that as -- remember, there's like 100 to 1. You have 100 worse qubits, and that creates 1 very good qubit, logical qubit. Those worse qubits, they have gates acting on them. If you can get the errors below a certain level, then that's when the logical stuff kicks in. And kind of think of this way, if you add more crap, you're not going to get good stuff until you're below a certain level. You can't take advantage of that. And there's just a lot of crap and a lot of noise everywhere. And so this is the logical error rate where once you below a threshold, you can keep driving it down lower and lower. And you often see metrics like 10 to the minus, blah, blah. And 10 to the minus, getting back to what some folks say is logical qubits, if you have a ratio instead of 100 to 1 but 2 to 1 but your logical cubit rate is 10 to the minus 3, it doesn't matter. It's just not -- the error is not low enough. So this is an important slide, key technological indicators. So for any hardware, including obviously ours, real hardware is always imperfect. It doesn't matter what anyone says, it doesn't matter what we say, no one has perfect hardware, anywhere in the industry, whether they're using photons or electrons. Having said that, if you get the error rate low enough and error correction and fault tolerance, you can tend towards a perfect system. Meaning that the errors you're getting are so low that you're going to get the right answer, which is ultimately what we all want to do. So for us, photon loss is the main source of error. Other approaches have their own sort of error. The error for everyone, you can think of it as characterized as decoherence, or the loss of coherence. And so we all have decoherence. It's when our world interacts with the quantum world, makes it noisy. And photon loss causes ours. Other based -- other approaches that are electronic, they would have, say, laser noise, they have spontaneous emission, they have T1 noise, all the other sort of stuff. The great thing, and also the sucky thing about our approach is we can just sort of say we don't have all those other things. We just have 1 source of errors that we need to overcome, and that's what we suggest everyone here focuses on its loss reduction. And as it says there, in Xanadu's photonic architecture, nearly all these imperfections reduced to 1 primary challenge, photon loss, whereas other approaches has a whole lot of them. Still solvable, to be fair, but it's nice to sort of thing you just have 1. So how do we often -- okay, so we understand that, then it's a case of, well, how are you going to deal with loss? How are you bringing it down? Well, there's 2 ways to reduce loss. It's through hardware improvements and through theoretical architecture improvements. And so the hardware improvements is essentially wherever light goes, it has a chance to lose photons. So when you start the laser and all the way to the end, some of the photons get absorbed, they get scattered. That's a problem because the photons encode information. So that means you're losing and absorbing information, which is not good. So if you can make the hardware smoother and less rough, then the photon has a higher chance of going from the start to finish without anything happening to it. So architecture, we say you need to tolerate abstractly, this amount of loss. And then the architecture team can kind of say, "Hey, guys, we found some other model of the system. You can now tolerate more loss." So you see you're reducing loss by the hardware and improving the loss balance. You're trying to reduce this gap, and then you want to get that gap to 1. And then you have everything you've ever been promised when you get to 1. In terms of the hardware, we have this flywheel happening here where we design our own chips in Toronto or in photonic integrated chips. And we heavily used, particularly last year, more and more AI to really optimize the structure. So you can think of it as if you have too many elements on a beam, on a wafer, on the chip, then you've got more loss. So if you can reduce elements while still maintaining the same functionality, then you can actually have a lower loss system. And then we go to multiple foundries around the world. We have 3 different substrates. So we work with multiple foundries because some are better at lower loss in sort of materials and others. We also won backups as well. So every -- we get -- every few weeks, we're getting back chips and wafers. And so then you go down to the testing. You'll hear about our new facility soon. We can do test and measure there. That's really important. You do test and measure for us to get more information. It's like, well, it's more loss than what we thought. So we'll have to alter it based on this feedback. And we have a dashboard and sort of a repository where we have all the data and information. So basically, you can kind of see this iteration process happening more and more. This is related to the hardware side of things. I can't overemphasize enough. We always get asked about AI and quantum. They do go together, but I would kind of tend to think that quantum is the next big thing after AI rather than trying to force 2 things that are hyped together. Although having said that, there is some overlap. But the biggest thing that we've seen the benefit of AI, significant benefit is using LLMs for the chip design and bringing down the loss there. It's been huge for us. And I think we're ahead of the curve when it comes to that. Okay. So we had this plot, if you remember, 5 months ago. We stopped it here because we couldn't show any forward-looking statements. So we had it from 2023 to 2026. So here, we have -- think of it roughly speaking, as loss. It's more of a loss gap you're trying to close. So when you hit one, you've closed it. And along here, we have the time line. Now if you remember, in our first Analyst Day, we had 2 lines. Just to make it easier and to track things better, we got rid of 1 of them and chose the harder line. Meaning if we have this and we can achieve this line, then the other line automatically follows from it because there's overlapping features there, but there's harder things to do on this one. That's the first thing. The second thing is our optical system has approximately 17 different optical elements associated with the quantum computer. We have loss where we are today and loss where we need to get over the coming years. And we've aggregated them in such a way that instead of having to track all 17, we have an aggregated line for that. So that's where that comes from. The other thing that -- and it's really important, this is what we think everyone should judge us by is this dotted line. Remember, you want to get down to 1. I think roughly, you can see its log scale here. Roughly, this is 2. 1.5 is breakeven on this scale. So breakeven means that you've hit the threshold. Now you need to go below. And as soon as you go below, you start racking up 10 of the minus 3, 10 of the minus 4, 10 to minus 6, minus 10 and so forth. But fairly quickly after that. The key is getting to that. In this case, it's normalized to 1.5. The other thing that I would kind of stress is that this is a midpoint. So if you think of it, it could be above or below, it's like sort of a sleeve that's air bars that encompass it. The key point is making sure it's headed down and making sure it gets to 1. So it's also not something that happens every day. Every quarter, there could be an announcement. But on average, are heading down at the end of the year. And so, so far, if you look at, say, this year, we're on track to hit the target for end of 2026, which is here. We're on track for that. Okay. So crossing the threshold, we need error reduction. In our case, it's loss. Everyone has their version. We need error correction, and that equals a logical error rate. So I've already talked about a lot of these threshold-driven performance. The thing everyone needs to start questioning and asking is what is the error rate? What is logical error rate? That's the key thing. Not really now, but over the next coming years, that's going to get more and more important for folks. And I like this phrase down here. Cut loss enough to get under the loss threshold, remember, heading towards -- starting at 1.5 and going to 1, then the error correction turns pretty good hardware, the physical qubits into almost never wrong logical qubits, the good qubits. So that's that ratio. In our case, 100 to 1 there. The other thing to keep and -- that applies to us too, we've said we've got 12 logical qubits. I think maybe [ Keyera ] says they've got 35. None of them are equal. It's not apples-to-apples. And it applies to us, too. So not all logical qubits are created equal. They only become equal when you also compare the error rates. So if someone says they have x logical qubits, you then ask them what the logical error rate is. And then when those things line up, then you can kind of compare apples-to-apples. And that's where the devil is in the detail. So it doesn't matter how many logical qubits you have. That's not the whole story. But the error rate completes the whole story there. So the error rate, as mentioned, is how long can you compute before the state decoherence before it becomes crap. So 10 perfect qubits can't solve a useful problem. 1,000 that fail after 10 operations solve nothing. So it goes to the extreme of either of them, so you want them both coexisting. So to sum up a lot of stuff, key indicators, loss, key indicator, logical error rate, key indicator, logical qubits. And in between them, you need error correction. We have a whole family of error correction codes, unlike others that we can choose from. And then the networking is that piece that we sold last year with our Aurora quantum computer. Remember, you -- us and others will master a small quantum computer, whether it's in a server rack or a chandelier, and then you need hundreds of these things and you need to network them together. And so we've solved the networking piece. So this is very much a blueprint for anyone doing quantum computing. How is your errors? How is your error correction? How is your -- what's your logical error rate, how you do it with networking and how many logical qubits do you have? It's kind of putting it all together. So Inception. On -- as mentioned on Friday, we announced CAD 195 million to build, among other things, our new facility, which is called Inception. So Inception, we moved in into Inception. It's a huge warehouse, 160,000 square feet. There's nothing in there yet. It looks really cool. It's huge. And we'll be using that money along with our own money to populate it out. So there's a few things we'll be doing in this facility. And it really is state-of-the-art. There will be some of these machines that we buy that no one in America even has some of these tools. So that's state of the art. It will be the best in the world in terms of lower loss for our architecture. So some of the things and the capabilities, test and measure, I mentioned that flywheel. You get the wafers bat. You've got to test and you got to understand like performer or does it not, get some data there, we'll be able to do that ourselves. Heterogeneous integration, that just means 2 different types of chips essentially glued together. We'll be doing that ourselves. Packaging, how do you control, how do you use -- how do you get light in and off the chip. You need to package it. Optically and electronically, we'll be doing that. And then how do you start assembling some of these things together. It doesn't matter if you've got them all. Individually, you've got to start putting them together. And only are we shoring up the supply chain. But if you take packaging, we used to work with a company in the U.S., not far from Lake America. And they were, I think, in Ohio, and it would take up to 3 months to actually package it. And that was fine. But now that we're doing it in our $10 million one as a precursor to this, it was done in 8 hours, 3 months to 8 hours. And so not only are we showing up supply chain, which is really important in quantum, but we're also speeding up our iteration times, which is also pretty amazing. And obviously, all of this helps with loss reduction as well, which is our key challenge. Okay. So I think this is the summing up slide here, yes, before I hand it over to our COO, Raf, to talk about software. So this is kind of a take home. This is our road map. So we have time along here, goes up to 2030 and beyond, 2031. We have key technological indicators. Loss, This is just the loss gap taken from the plot. As mentioned, 1 is where you need to get to, and 1.5 is the breakeven point. So you've got these numbers from the plot. Logical error rate. Now this is the first time we're talking about Xanadu's logical error rate. So this is what's really important. There's a couple of things that I would say as a takeaway from today's presentation. One is our error rate, we anticipate to be far, far lower than anyone else's in the world. The lower the error rate, you can think of that also as being proportional to more applications, higher quality applications you can do compared to anyone else. To our knowledge, looking at publicly available presentations, the most I think I've seen -- someone can correct me, is 10 to minus 12, something like that. So we're far lower than anyone else. That's the beauty of a photonic-based approach. And the other key thing is we've announced today over 1,000-plus logical qubits in 2031. Previously, we announced up to 520. We said 2029-2030 and up to 200 before that. So this is our road map and it comes back to what we've been talking about this morning, how many logical qubits -- you're seeing no physical qubits there. There are physical qubits, but the key thing to judge on is logical qubits and the error rate. And I don't think anyone is claiming that they can get to this. This is the great amazing thing about photonics because the network ability, you keep adding machines. And the error rate goes lower and lower, it's adding extra hardware. So that's really exciting. And then we can break these phrases up into less sort of specifics and the projects. Qubit factory build out, this is the toughest part of what we're doing. Fault tolerance is reached in 2028-2029. That corresponds to between these, so 1.5. And then start building the quantum data center in 2030, 2029-2030. Now while in order to achieve loss reduction and start making sure we achieve our goals, we need to open up a facility. Inception, that was opened 2 days ago. And we'll start building the quantum data center. We anticipate early next year is our road map to start that. A lot to take in, and so I can answer some more questions later. But I guess we'll leave that to the end, and I'm going to hand it over to Raf now to talk about the software. Thank you.

I'll assume that applause was for me, not for Christian. Thank you, Christian. And we talked a lot about hardware so far. But in order to satisfy that mission of building quantum computers that are useful and available to people everywhere, you really do need a software stack to be able to enable all those applications, to make it useful and of course, provide that availability. PennyLane is that leading software framework with over 30% of developers last year having used it. And this is across large multinational enterprises, government labs, academic institutions. There's kind of an issue in quantum computing right now in that there's a large amount of investment happening into the discovery of various quantum algorithms and applications across many industries. But fundamentally, we're bottlenecked by the availability of hardware. So this utility scale quantum computer coming online, but also the availability of proper tooling to be able to build these applications out. If you ought to think back, this is like sitting in the 1940s, 50s of digital compute, thinking about breaking codes and solving chemistry and physics simulations. But nobody is thinking about what the Internet and e-commerce and all these things are going to bring because that tooling is not yet available. That tooling also provides some issues right now in terms of what's available out there in terms of being more broadly adopted. One, a lot of it requires pretty high expertise in order to be able to leverage. You do need to be a quantum expert for a lot of the tools to be able to program. One of our investors in the past did a survey and identified that there's less than 2,000 people worldwide that have that deep expertise in quantum application development. There is substantial vendor lock-in. If you go start developing on something as simple as IBM Qiskit, you very quickly are tied to that stack. Maybe you're looking at NVIDIA CUDA-Q, you get tied to NVIDIA GPU accelerators for your workflows. There's also fundamental issues that Christian already alluded to that limit existing software and being able to actually scale to the level that it needs to, to provide applications for utility-scale quantum computers. To give you an idea, today, you're dealing with handfuls of qubits deployed for tens or hundreds of gates if you're running on [ Neskara ] hardware that's publicly available. If you're going to be running on a utility-scale quantum computer, you're going to be working with hundreds to thousands of logical qubits operating billions of gates over thousands of repetition. So trillions of steps. And these programs today are not able to be captured by the tooling that actually exists. And then finally, adoption does require effort. Nobody wants to pick kind of the wrong stack. But in the world of multiple winners, a lot of the value is going to push into software. So on various other compute platforms, whether it be digital CPUs, whether it be GPU accelerators, whether it be AI, we see that more and more of the value actually does get pushed into the software stacks. NVIDIA did this incredibly well. It's predominantly a hardware company using a software layer to actually lock the entire ecosystem into their hardware. AWS built an entire platform in commodity compute that was able to provide huge amounts through a software layer of value. And this is exactly what we're looking to do with the entire PennyLane family of programs, not just PennyLane on its own, but also Catalyst and Lightning, which I'll touch on a little bit. We're definitely focused on developing those high-performance compute and AI applications for quantum computers. So these range from existing ideas that we have around materials and chemistry, out to pharma and drug discovery to crypto analysis and broader optimization in places like finance systems. But this is really built on top of PennyLane as that software development kit for our universal [ photonic ] quantum computers. Over this year, we've embarked on an enormous effort to actually raise PennyLane from something that was previously targeted. [ Nesara ], and, I would say, early midterm quantum computers to now being fully future-proof for that utility scale level. In addition to that, we also continue developing Catalyst, which is our compilation layer, actually being able to take these programs and compile them down to things machine code that could actually run on individual quantum computers. This is an incredibly important step as it can actually provide huge optimizations and how much you can reduce the gate count required for a given program. And some of our examples, we're able to take a program and reduce the computational resources by a factor of 10 just by [ acquiring ] the proper compilation passes. This means that you could take an application and run it effectively a year earlier than you would expect on hardware based on currently publicly available road maps. And then as these utility-scale quantum computers are not available here today, you do need a suite of simulators to actually be able to validate, prove out and understand the full barriers to running these in production environments. And we've deployed those with Lightning simulators that are now available as part of PennyLane as well. Really important part too that I'm sure everyone here knows is that PennyLane is very much hardware-agnostic, and this goes both for the quantum hardware. So there's no reason why you could not run on any of the systems or modalities listed below. And in fact, are ready for superconducting trapped ion neutral atom and photonic. We have a number of publicly available devices that have built-in plug-ins within PennyLane, so you can access those, but also on the classical compute accelerators. So we have a long-standing partnership with NVIDIA, but also a recent one with AMD, where we're targeting both of their devices, along with other less traditional compute accelerators in order to let people leverage the top computational tools today to be able to rapidly develop quantum applications. This has actually attracted a huge amount of adoption for PennyLane. Like I mentioned, already 30% of quantum developers out there have adopted PennyLane within the last year. Over 1,800 publicly available software projects referenced PennyLane as a dependency. So we're actually building on top of it. That includes 1.1 million package installs monthly on average across the entire family, so PennyLane, Catalyst and Lightning simulators. And like I mentioned, there's a large number of hardware and simulators that are available. So I believe 51 kind of across PennyLane. This has really been a massive adoption for us. It's been great. It's also been great in that the reason that people are adopting this is the true fundamental technical capability of it. I mentioned, they're available in all modalities, Catalyst and just-in-time compilation having the huge performance boost. Christian identified the importance of error correction. And this is something that is fundamentally built into PennyLane at its core level. The other big part was originally when PennyLane was being built out. It was considered to be a quantum machine learning software stack. This was -- I'm going to say some of the researchers at Xanadu were the founding figures within that field, and it still holds true throughout. And that's differentiable end-to-end, meaning that it directly integrates with today's advanced AI platforms like [ TensorFlow ] and [ PyTorch ] and allows for easy transition of developers within those spaces to be able to actually be incredibly familiar with PennyLane as a programming language. One of the major reasons why this massive adoption has happened, aside from the technical piece, is the fact that learning has really been the final that's allowed developers to come into our software stack. According to the Unitary Foundation, over 50% of developers within the quantum space cite poor documentation as the reason that they kind of drop off from various other quantum packages. We have been leaders in terms of providing material to not only get new people into quantum computing, but also transition existing developers from traditional industries like high-performance compute and AI to become quantum developers. This means we have over -- close to 600,000 codebook submissions. This is our coder size type of exercise platform where it allows people to self-assess their skill level. We have over 200 demos. So these are the latest research papers that are coming out, but are then fully implemented by the community within our ecosystem and made publicly available. We also have over 4,500 distinct learners within our code book. So people that are active progressing. And it has become so popular that in fact, a number of our enterprise partners have requested an enterprise version of this to be built to upskill their workforce in quantum development. One publicly that we released recently and we're able to talk about is our partnership with Lockheed Martin, where they're looking at about 200 people internally every year being interested in transitioning from AI and classical developer roles to quantum developer roles but realistically only have resources to take a handful of those folks. So we were able to work with them to build a self-serve platform where the remainder of those people are able to continue on that learning path. Lockheed Martin is not our only partner. And education is not the only place where we partner with folks, definitely partner on a number of application development. I think this was disclosed previously. But in chemistry, finance, machine learning and fluid dynamics, we work with folks like Volkswagen, Toyota, BMW on some battery and sensor use cases, with Rolls-Royce, the aerospace company on that computational fluid dynamic applications, with folks like Mitsubishi Chemical Group on next-generation photoresists for semiconductor fabrications. But then also with folks like AWS, AMD and NVIDIA. We're working to build out the next generation of tooling. If you're familiar with AWS Braket, PennyLane is closely integrated and in fact, supports many of the features on that platform. If you're familiar with AMD and NVIDIA and what they're doing in the accelerated compute space, you will see quickly that those devices are natively supported within PennyLane, allowing it to be able to deploy from your laptop to the largest supercomputer in the world without having to change the code in any sort of way. And then there's a wide range of companies like [ Agnostique ], [ Unoses ] and [ MetanAI ] that are quantum start-ups that are actually fundamentally building their businesses on top of PennyLane as well. We're also doing this as part of the broader open source community. So developing this on your own is, of course, quite challenging, but we're able to now have an incredibly good reputation across various research institutions and the broader community, allowing them to contribute back into our code base. PennyLane itself has over 2,500 scientific citations, over 75 outside contributors that are pushing code and improving in PennyLane every single day. We're actively involved with the MLIR ecosystem. This is the middleware stack for AI compute that is now being leveraged within quantum. And this just drives broader adoption across enterprise, government lab and, of course, academic partners as well. And really, it allows us to be able to start looking at early commercial adoption of PennyLane. The long-term vision is definitely 1 of serving as that funnel for hardware and providing an enterprise-grade platform where you're able to make the adoption and integration with quantum computers very straightforward. But we don't have to wait for the utility scale quantum computer to be able to do that. Already today, like I mentioned, workforce training and enablement is something that we're aggressively pushing on. We have a dedicated platform that we've built out that we're continuing to provide to various partners. We're continuing on our algorithmic engagements. So this is actually identifying $1 billion-plus value problems that our partners have through commercial engagements and being able to capture the IP to be able to deploy those problems on a utility scale quantum computer. It also gives us a fantastic ecosystem as we're getting developers onto PennyLane to be able to provide future products. There's many ideas that we're currently following, it's just getting an understanding of when is the right time in the journey, especially as hardware develops to be able to deploy those. And then ultimately, all of this is building out a very large funnel of not only individual developers and companies that are fully integrated and understanding quantum computers and how to develop applications for them, but also a wide range of ready-to-go applications that have been developed on PennyLane that could be deployed on our quantum computers. That's, of course, will be a first-class citizen within the PennyLane ecosystem. With that, I'll pause there, and I'll hand it back to Christian for a quick summary.

Great. Thank you, Raf. So we're talking about 4 different ways of commercializing. This is also from our Analyst Day, it's a repeat of that. So I just wanted to make sure folks understand that looking to sell through, make money through the hardware access to the cloud, selling individual computer systems is part of it. Advanced photonic devices and IP licensing, that's a key aspect of what we do as well, really the world's leaders when it comes to that ultra-low loss. And then as Raf was just mentioning, the other revenue market that we're after is commercializing PennyLane as well. Michael, is this your slide or -- okay. So I think it's quite self-explanatory. $686 million of capital sources. We have a synthetic ATM, which we've mentioned before. I think everything is Canadian government funding, that's $195 million from the federal government. That's a nondilutive loan. And then we have the synthetic ATM as well that's ongoing. Okay. So some of the key takeaways. The first pure play publicly listed for photonic quantum computing company, deeply technical leadership and experienced board, category leading technical position, particularly with the Aurora quantum computer, we solve the networking side of things. That's key to how you actually scale up quantum computers, not just ours, but nearly all else, all the others. Raf went over the PennyLane. It's really 1 of the leading software stacks out there and targeting meaningful end customer commercialization by 2029-2030. That corresponds to our large-scale quantum computer and quantum data center. As we've mentioned in the past, 2, the top 2 priorities for us doesn't include revenue, not at this stage. It really is hardware and technical milestones, achieving them. Because ultimately, the only thing that matters in this industry is that you've built a large-scale quantum computer. And secondly, some partnerships with some great companies. We announced AMD and Lockheed Martin at the start of the year. And hopefully, partners like that will become our first and earliest adopters of our technology as well. And as mentioned, available funding that we have is USD 686 million, and we also have a synthetic ATM. Okay. So we'll leave it there. Thank you very much for everyone's attention.

Can you guys hear me? All right. That works. All right. We have a lot of questions. Excellent. So this is the Q&A section. We have mics moving around. So if you could -- let's start with John.

Thanks for doing this, guys. 10 to the negative 16, super impressive. How are you going to get to that kind of error rate? Could you elaborate a little bit on your confidence there too, Christian?

[indiscernible] confidence. We're in the deep, deep tech sector, everyone in quantum computing. So there's going to be small error bars, that should be taken into account for us and anyone in the industry. We're very, very confident in that. I would say that the challenge really is getting the loss to the threshold. That's where the biggest challenge is. After that, when you go below, you pass through -- so I didn't mention there, but 10 of the minus 3 is where our threshold sits is for our -- the way it's been normalized for our system. So we're going to hit 10 to the minus 3, and 10 to minus 3 is related to where we are with this loss reduction or this closing of the gap. So once you pass that, the way anyone gets lower is you keep adding more stuff to it. Because once you're below a threshold, the really cool thing is now that you add more stuff instead of the errors going up, which you would think with error correction, it actually goes down lower. The reason I bring that up is that for us -- and I believe it's easiest for us in terms of this is that we can just keep adding these server racks, these modules to our system because now we know how to solve the interconnect problem. And the assumption there is the interconnect problem has been solved for a certain loss as well once you hit that -- the loss targets. And then the more you keep adding, the more the lower that it actually goes in the error rate. So very, very confident of that. That's not the challenging part. The challenging part is getting to that threshold. And everything else there, as you saw, it's an aggregate of the top 17 optical elements. And we needed to, in this case, normalize to 1.

But you're still at that 100 to 1 physical to logical at the 10 to the negative 16?

That's correct. That doesn't change. Those numbers will be pretty fixed. It's proportional to that, but yes.

You referred to this as copy paste, I think, to me in the...

Yes.

That's the evolution there.

That's right. That's right.

Harsh Kumar, BMO. Just -- and feel free to get technical or detail about this. You're basically dealing with photonic qubits, and you said that your error consists of the photons getting lost or the information getting lost. So how do you -- this is kind of the mechanism of error correction. I'm trying to understand how do you know what's lost? And what's the false information that's being absorbed, how do you separate between the two to get to the proper amount of information that you're desiring to get to?

Well, what you can do, just as a very simple example, let's say you have -- it's going to be quite small, but it's a mantle's, let's say the chip is this big. So you couple in fiber optics goes in a fiber optics goes out. And so you need a pump or put laser light through and it propagates along this distance here. And so what we wanted to -- and let's say there's no optical elements just propagation loss. So how much of the light actually makes. And like you said, some of the light gets absorbed or scattered as it's going from A to B, and then it gets coupled out. There's loss there as well. What you do is this is a known problem, and you forget about quantum, and you just shine classical light through it. So like a traditional laser. And then you know how much -- you can calculate how much light you're expected to go in, let's say, it was a perfect system and then you detect how much light you actually collected at the end. And then that tells you how much the loss, is the difference between the 2. And it's more detailed, but that's essentially the way it is. You measure the light that comes out and compare it to the ideal case where you assume, say, 0 propagation loss for instance.

Craig Ellis, B. Riley Securities. So -- no offense, Christian, but I'll direct 1 to Rafal. Although it may be a question for Michael. So I really appreciate all the information on PennyLane and how we're taking that capability and trying to develop an ecosystem as many others have in the past to attract developer interest. But what I'm not clear on is how the pricing model is going to work and how we actually monetize it? Is it as a service, onetime seat license? How is that going to work?

Yes. So I think in the long term, it really depends on how the industry does develop. What I can tell you right now is that -- and for example, the educational offering, the appetite from our partners is very much as part of broader engagements and kind of on a per seat basis. But what I will say is that all of this is always with an eye towards that utility-scale quantum computer. The driver here, like Michael and Christian often say, is not the revenue. It's the adoption signal. It's the large number of developers that we're able to get. It's the fact that we're getting Lockheed Martin and our other partners hands on with more people with PennyLane, that is the push. Because what we want is not I'm not sure what the TAM of quantum compute is today, but it's small. We don't want a big chunk of that. We want a huge chunk of ultimately, the TAM that will be here in 2030 when the utility scale quantum computer comes online. So I think great question. It's something that we're continuously testing, discussing with our partners internally and externally. But not a major driver until that utility-scale quantum computer lands.

[indiscernible]?

No, I think that the third of the last slide that Christian presented, which was also from our first Analyst Day, covers it. That's the way that we're thinking about it. How those models flush out really depends, I think, how many people are there at the starting line. I'd like to think of that as the starting line. But I think that slide does a good job of laying out our thought process.

Great. Todd Coupland from CIBC. How does your position now with your technical spot in '26 stack up wit DARPA's Stage C?

Not there now, we're along that line would you need to be to satisfy Stage C?

Yes. We're -- the easiest way to sort of say is we're on track the time lines here and the time lines that I mentioned there. As you know, Stage B is roughly a 12-month program. It could be longer, it could be shorter, honestly, depending on DARPA's resources, how quickly they can get for the analysts analyzing the hardware. But we can't mention more until we find out how we're doing, but so far so good, and I can say we're on track. The key thing really is the on track really relates to the loss. That's really the be all and end all, where we are with loss reduction. As mentioned, this has to make sure it keeps going down. There will be quarters where we may, like I think last quarter, we mentioned a really good fiber-to-chip coupling result. We'll announce them every now and then, but it really has started the end of the year seeing that go down and making sure, ultimately, we're hitting that value of 1.

[ Joe ], Morgan Stanley. Just a question on the software side. I want to understand something you said about Q-to-Q. I would understand reservations with IBM and Google that you're tied to their hardware. I guess, NVIDIA would argue that they're sort of agnostic to the quantum side of it, and you're locked into the coprocessor and simulation side. But what I might not understanding about that, why would people be afraid of NVIDIA locking in?

Yes. So I would make the argument that we're actually pretty locked into still the hardware side as well. And it's because that advanced compilation requires hybrid compilation between all your computational resources. And if you're going to limit yourself in building a quantum computer to just NVIDIA GPU accelerators, for example, you will never be able to achieve above megahertz clock rates. And on the [ street is ], NVIDIA is great at wide pipes, but not low latency compute. If you want low latency, you're probably going to be going with AMD with [ RSOC ] or FPGA products through [ Xilinx ] or others. And now if your software stack can't compile down to those types of resources, you disqualified a whole host of quantum computers. So it does look like people often think only on the simulation side, but there's the entire run time and compilation side as well that you want to make sure you are relatively vendor-agnostic because right now within the broader quantum ecosystem that is very actively playing out. And at Xanadu, we're not even limiting to only existing hardware, we're even thinking things about like what happens when there are custom ASICs that people start building out specifically for their purposes for their quantum computer and how do you have a broad plug-in architecture that can support a rapid development of those types of developments that you're able to compile down into.

Yes. Nehal Chokshi from Northland. Thank you for showing that road map in terms of DB improvement. It is somewhat nonlinear as we go through time. So it looks like it's carefully thought out with respect to particular improvements in components, particular components. So given that context, what is that 1 photonic component or material parameter that's going to give that big boost, especially going from '28 to '29?

It really is the 17. There are maybe 1 or 2 that were pretty much already there in terms of the number of 9s and say, fidelity we need. But you can make, say, 1 quarter, we make progress on the second out of the 17. That's good, and then you need to go to the next 1 or the 1 after that to improve it. So it really isn't -- the way I would look at it is the 1 thing that's already been chosen, that's loss. But then you really do break it down into many other things. The other thing too is you can trade off numbers as well. So you can kind of say I made more progress than I thought here, and that will allow the other 1 above it, for instance, to tolerate more. So it's really hard to sort of narrow it down to 1. Another way to look at it is, let's say, you had 16 of them amazing and 1 is really, really bad. You can't do anything. So they're all really interconnected, and you really need to make sure that you're lowering the loss on all 17.

One other question, maybe quick, maybe not. I talked to you about this before, but it's hard to have right my head around it. So why can Xanadu tolerate any photon loss at all? Because there are some other photonic quantum computing players out there. They're thinking about, hey, we got to be able to do single photon detection, no loss of photons at all. So can you talk to that?

The single photon ones, yes, they will have loss. In that case, they're losing the whole carrier. So when you have a single photon you're losing the whole information that's been encoded. Technically, for us, we're using many photons, where it's -- theoretically, it's looked at as having a mixture of vacuum mixed in, basically, so you never really lose it. So it's just a quirk of ours where you don't lose it, but you actually get more noise to the system. That's how you theoretically work it out. So -- but to answer your question, ultimately, it's just the type of encoding that 1 approach compared to another approach was taken. We're taking a many photon approach, where you can incur information using the GKP states. Which has many photons and a super position of photons, whereas the other one, roughly speaking, is a single proton. So it has -- we have more resistance to our GK. We have the first layer of resistance is with the GKP coding. That already tolerates a certain amount of loss because you're actually not worried about a single photon in that case.

[indiscernible]?

Photonics? That's correct. There is another 1 in Japan that's been getting some funding. It seems like they're doing GKP, but it's unclear and they're using different sort of rate or repetition rate for theirs. But for the most part, yes, there are some in superconducting. [ Nord Quantique ] has it.

[ David Williams ] with Needham. So just -- you talked about the 17 components earlier, and you've got an Inception that you're building that's for a couple of different purposes, really for packaging, bringing that in. How do you think about these components? Are there opportunities for you to bring those in-house or maybe to accelerate this path forward on some of these components that aren't quite performing as you would think? Or how is that?

100%. Yes, for sure. So the packaging, like you mentioned, is one. I gave that example where we it was taking 3 months. And with last year, we announced a $10 million kind of a baby version of Inception. We spread that up to 8 hours. But another example is the heterogeneous side of things that I mentioned. It's for test and measure, which you can measure the loss and then iterate to bring it down. Packaging, you mentioned that's what it's doing to reduce the loss. Heterogeneous integration, this means that -- 1 way to look at it, perhaps instead of using fiber optics to connect 2 different chips, you can actually place them on top of each other. And now you're reducing the opportunity for that to be lost. Because when you take light on and off chip, there's a coupling efficiency, which can -- you'll lose light as well. And then it propagates through fiber optics, which is less of one, but it's still losses there. Now if you put them on the top, there's no propagation compared to that, there's less propagation risk there. So we'll be doing that at an Inception as well. And there's also the assembly side of things. The assembly side, you have to do a lot of making sure that you put it together in the lowest way possible and you're optimizing it from architectural point of view as well. So yes, the answer is yes. It's a huge yes there.

Kingsley Crane, Canaccord. Thanks for doing this. It's really helpful update and encouraging advancements over the past couple of days. So I think 1 of the things that was talked about in the advancements was commercializing photonics IP outside of quantum computing. And I just want to check if there's been a change in stance within the company, like maybe that there's more potential to do that and if that could be like a revenue driver in advance of 2030?

Yes. It's definitely something that we're continuing to explore. What I'll say is maybe the bottleneck or the barrier to being able to fully push on that is just understanding how much of a gap there is between our IP and a dedicated product. Packaging, Christian mentioned, we have world-leading packaging capabilities. So the ability to move light on and off chip. This is something I'm going to say every single hyperscaler out there is super interested in as picojoules per bit transferred is now 1 of the major metrics that they worry about. So ultra-low loss coupling is important. But the process we're going through is understanding and is there a gap, if any, between what we have and what they need. If there is, do we want to build that product on our own? And what would be the cost and deviation from our path? Do we want to partner with somebody in order to be able to do that? So this is just what's happening internally. I can tell you we've been progressing on that thinking. So it's something we're going to definitely continue exploring. But still a priority and still a topic internally.

It's [ Ruben Roy ] from Stifel. I know you guys are calling this the technology road map. But Michael, maybe just high level on the funding, congrats on the CAD 195 million. How much of this is fully funded? Any changes to the spending kind of forecasts? Or how you're thinking about spending against the new road map?

Yes. I know I'd reiterate some of what we shared. So as we talked about on the last earnings call, you're certainly going to see R&D and CapEx increase in the second half of the year. It's -- also, as we move forward, that will start to accelerate, but we've said that in the past. At this point, we're not ready to update our full outlook. We're in the process of continuing to evaluate that. I would say there's exciting work as we do that. You hear some of the things that we're working through in some of those. And so we're in the process of optimizing around those decisions. And when we're ready, happy to update on that. Yes.

Thanks for putting this together. [ Jesse Sobelson ] with BTIG. Just curious, why the range in logical error rate guidance for 2028? Is it all just loss reduction? Or is there any dependency on the error codes being used? Or any other factors?

The range was more about -- you've got to pass through all those 10 to the minus numbers as you're reducing loss. So the range is -- once you hit that, fairly quickly start reducing it. So in that particular year, I think 2029, we expect to see orders of magnitude in loss in the range of the error rate being reduced. So it's just in that year, over that 12-month period, we'll be passing through all those error rates.

Paul Treiber from RBC. Just on -- you mentioned building the qubit factory is probably more challenging than reducing losses. Can you elaborate on that? What's the challenge there? And then what's the path to reduce or to create the qubit factory?