Okay, we'll go ahead and get started. Good afternoon, everybody, and welcome to the first day of Needham's 27th Annual Growth Conference. My name is Quinn Bolton. I am the Semiconductor and Quantum Computing Analyst for Needham. Thanks for joining us. It's my pleasure to host this fireside chat with D-Wave Quantum. D-Wave was the third pure-play quantum computing company to go public back in August of 2022 and remains the only company today that provides quantum annealing systems that solve real-world commercial problems. With 59% of its revenue from commercial customers in the past 12 months, D-Wave is positioned as the market leader in commercial quantum computing systems. Joining me from the company is Dr. Alan Baratz, CEO. Alan, thanks for joining us.
Thanks, Quinn. Pleasure to be here.
Yeah. Alan, just to start off, the company's entering 2025 with some really strong momentum across product development, scientific advancements, and commercial adoption. What excites you most about the year ahead?
Yeah, so you're right. It's been an amazing ride. I will tell you, I remember when we went to our very first investor conference after going public, and we had a fireside chat, and I think we had one person in the room. So things have changed a bit. I certainly wish that the air conditioning was a little better in the room.
Yeah, I agree.
But thanks for the opportunity. Thanks to all of you for showing up. So as we look at 2025, there are quite a few really important deliverables that we're focused on. We've talked a little bit about the fact that we're close to bringing out our Advantage 2 quantum computer, which is our second-generation Advantage system. The current Advantage system is a 5,000-qubit quantum computer. That's what's in production today, supporting customer applications. Advantage 2 will be a significant advance over Advantage. It will be a more connected quantum computer to solve larger problems. It will have higher coherence time, the time within which the qubits stay in the quantum state to be able to solve problems faster, and you'll be able to specify problems with greater precision to get better solutions to problems.
We have a small 1,200-qubit version of that system that's available in our Quantum Cloud Service today for customers to start working with. We've announced the calibration of a 4,400-qubit version of that system, and we're looking forward to getting the full production system out in the market before the end of the year. Second, we're continuing to make good progress on our gate model system. So we are the only company in the world that does annealing quantum computing, and we believe that there are very significant advantages to that in the area of optimization. But we're also working on a gate model system to be able to address the application areas where it excels, for example, solving differential equations for quantum chemistry. We've made good progress on qubit fidelity with our gate model system.
We're hopeful that over the course of this year, we'll be able to demonstrate error correction by putting qubits and control logic on the same chip, similar to what Google did with Willow, although using a slightly different type of qubit than what Google is using. Next, from a software perspective, we have several hybrid solvers that are a part of our Quantum Cloud service that customers can use to solve problems larger than can be solved natively on the quantum computer. And we continue to advance the hybrid solver development work, always bringing out solvers that deliver better performance and better performance in targeted application areas. And finally, and this one I'm really excited about, we started working six months or so ago on the integration between quantum and AI.
And we've done some very interesting work around inserting the quantum processor into the model training process. And we're getting some early results that show that we may be able to train models faster and with lower energy consumption than what's required if you do them purely with GPUs. And so it's early days, but we're actually quite excited about that. And we think 2025 will be an interesting year in all of those areas.
Yes, it certainly sounds like a lot of exciting things. Moving next to my next question, IDC recently called quantum optimization the killer use case for quantum computing. What exactly is an optimization problem? How common is it in today's enterprise? And how is D-Wave uniquely positioned to capitalize on this near-term use case?
Yeah, so first of all, let me answer the last question you asked, which is, how are we uniquely positioned to capitalize on it? Annealing quantum computers are native optimization engines. They are very, very good at solving optimization problems. And in fact, it's now been shown by researchers both in the U.S. and Europe that gate model systems likely will never deliver speed up on optimization problems. So we've got a bifurcation in the application environment for quantum. There are problems that will always require annealing, where annealing will be preferred. That's optimization. And then there are problems that will always require gate differential equations for quantum chemistry. So we are highly differentiated when it comes to solving optimization problems.
And so when IDC says that optimization could be the killer app for quantum, that's a very important statement for us because that is our sweet spot. So what is an optimization problem? An optimization problem comes up over and over again in businesses. Basically, workforce scheduling is about optimization. Last-mile routing for delivery is an optimization problem. Cargo loading is an optimization problem. In fact, most of the important hard problems that businesses need to solve are optimization problems. And we're unique in the industry in that the approach to quantum that we've taken, annealing, is the approach that can address that class of problems.
And so annealing is uniquely positioned to solve optimization. And again, you're the only annealing quantum computing company.
We absolutely are. Now, on the planet, exactly. Now, you will, I believe, in 2025, start to hear some of the other quantum computing companies talk about not necessarily an annealing quantum computer, but annealing mechanisms as being important for solving optimization problems. Because I think this point that annealing is a native optimization engine and really well-suited to optimization is starting to become more and more well understood. And we've actually seen some gate model companies try to simulate annealing on their quantum computers, on their gate model systems. They don't do very well. And that's kind of like back in the day, if you wanted a graphics processing card for video games, you could run it on a CPU, or you could go buy an NVIDIA card. Which do you want to do?
Do you want to simulate it on a gate model system, or do you want to get the annealing quantum computer?
Yeah. I'm sure lots of folks in the room want to know the answer to this question. Last week, quantum stocks took hit when Jensen at CES made a comment that it may take 15-30 years to get a very useful quantum computer. I have a feeling you may disagree with that statement. So how do you, what's your reaction to his remarks?
Yeah, not only do I strongly disagree with the statement, I've strongly disagreed with his statement in many interviews over the course of the last week. So the problem is that Jensen has assumed that all quantum is the same, but it's not. There are different approaches to quantum. We've just talked about annealing versus gate, two very different approaches to quantum. So Jensen doesn't understand annealing. I mean, his company, to the extent they're working on quantum, they're working on simulators for quantum computers.
They're working on software development tools for quantum computers, but that's all gate. So he thinks quantum is gate. Well, it's not. There are other approaches. Annealing is commercial today. We have companies that today are using our quantum computers as a part of their business operations. NTT DOCOMO is using us today for cell tower resource optimization.
Pattison Food Group is using us today for workforce scheduling. We're close with Ford Otosan, a joint venture of Ford and Koç Holding on autobody assembly. Interpublic Group for promotional tour planning. We are commercial today. So not 30 years from now, not 20 years from now, not 15 years from now, but today. So how does that jive with 15- 30 years? It just doesn't. And by the way, the other thing that I think it's important for Jensen to understand is we've demonstrated quantum supremacy on our quantum systems. We've been able to solve an important, useful problem in the area of material simulation on our quantum computer in minutes that would take well over millions of years to solve classically. I'm not talking about what Google did, where they've got some contrived problem.
It's called Random Circuit Sampling, which is basically trying to simulate what the quantum computer is doing itself, of no value at all. This is a real-world problem that we can solve millions of times faster than the fastest supercomputers in the world, which happen to be massively parallel GPU systems. So think about that, Jensen. We are solving important problems today that your systems just can't solve. You might want to think about that.
Yeah, it certainly seemed to me sitting in the audience when he made those comments that he might have been referring more to a truly fault-tolerant, error-corrected gate model quantum computer.
Exactly. And to be fair, we are years away. I mean, we need error correction. I mean, we're building a gate model system as well. You need error correction. You need scale. I mean, you're going to need to be able to interconnect either chips if it's superconducting or traps if it's trapped ion. There's still work that needs to be done. So yeah, it's going to take some time on the gate models.
Maybe just for folks in the room that may not know the difference between quantum annealers and gate model quantum computing, can you talk about the difference? And we kind of touched on the readiness and the maturity where gate model is certainly behind annealing, but maybe what's the big difference between annealing systems and gate model systems?
Annealing quantum systems do only one thing. They solve a very specific mathematical problem, which is finding the lowest energy point in a multidimensional landscape. That's all they do. Now, they use quantum mechanical effects to do that. They use superposition. They use entanglement. They use tunneling. This has all been proven multiple times by us and academics around the world. Why does solving that problem matter? The reason is that pretty much every optimization problem can be easily converted into that problem. The way you program our system is you take your problem, you transform it into the problem of finding the low energy point in a multidimensional landscape, you submit it to the quantum computer, and then you get the result back. That's what annealing quantum computers do.
Because they're solving only that one problem, and because the annealing process runs in typically microseconds, you don't need long coherence times. You don't need the kinds of coherence times that you need on the scale of gate model systems, which are quite different. Gate model systems are more like classical computers where you program the problem by specifying the sequence of instructions needed to solve the problem. These are quantum mechanical gates. The problem is it could take hundreds of thousands or millions of gates to solve a problem, and you need coherence through the processing of all of those gates. How do you do that? Either you have infinitely long coherence times, which will never be the case, or you're able to do error correction, which gives you the benefit of that.
But error correction itself is a very challenging technology where there's still some hard problems that need to be solved.
Each gate, as you go through that sequence, introduces errors.
Potentially introduces more error.
Yeah. I know you touched on the difference between your quantum supremacy work and Google's. I don't know if there was anything else you wanted to add on that. So I had it here on my list of questions. I know you sort of touched on it, but.
Yeah, sorry for jumping the gun. Sometimes I get excited. Look, I think the important thing to keep in mind here is that there are three companies that have demonstrated quantum supremacy. Google has demonstrated it on random circuit sampling, a contrived problem. Quantinuum has demonstrated it on random circuit sampling, a contrived problem.
D-Wave has demonstrated it on material simulation, basically simulating the properties of materials as they go through a phase transition. None of the other quantum computing companies have demonstrated it. In fact, none of the other quantum computing companies have actually even shown a speed up over classical on anything. I'm not talking about supremacy where classical can't do it, but even a small performance improvement on a real-world problem. So yeah, I mean, there's a ways to go in the gate space for us as well as everybody else, but annealing's here today.
Got it. And that quantum supremacy work, you've submitted it to a technical league of technical computers.
Yeah, we submitted it to one of the top technical journals. We're just about at the end of the peer review process. We're actually waiting for a publication date. We better get it soon.
Okay. Looking to 2025, will this be the year where we see more and more customers with quantum and hybrid quantum applications built with your technology going into production and being used on a daily basis?
Yeah. I mean, look, we already have the first of those customers today that are using it on a daily basis. We have others that are in development with us. We have a professional services team that works with customers to help them build out their applications. And so we have a number of others that are in development. And then we have customers that do it on their own. The NTT DOCOMO example, they did it by themselves.
They didn't require professional services from us. In fact, it's an amazing application. And we didn't even know about it until they wrote a paper on it. I mean, we knew they were a customer and they were buying time from us, but we didn't know exactly what they were doing. So we help many of our customers. Some of them do it themselves.
Perfect. You'd mentioned quantum and AI and the intersection thereof. Can you expand on those opportunities? You talked about the model training. I think there's also, and I'll be honest, I don't know what it means, restricted Boltzmann machines. I think that where you've done some work there as well.
So I think about the opportunities for synergy between quantum and AI as falling into three areas. The first is kind of obvious and the most near term. That's simply the two working together to solve a problem. So imagine the following. I use AI to predict product demand over the next year, and then I use quantum to optimize the supply chain to meet that demand. So here we've got each of the technologies solving the piece of the problem that they're best at solving, working together to solve the complete problem. Second, and second is the one that I referred to, which is quantum as a part of the AI model training or inference process.
And there's some recent work that was actually done at NVIDIA that was published that looked at the two primary approaches to model training right now, what's called diffusion and transformer, and how if what you do is you insert in the middle of the model training process a, it's called a small latent space, a small subset of the variables that are representative, good representations of the problem, and you train the model with that inserted, then you can get better results, significantly better results. And it turns out that our quantum computer is really good at doing exactly that, at generating samples from small latent space models, which allows you to tune up those small latent space models very quickly and with very low energy consumption. And that's the kind of path we're working on.
So it's basically using the quantum computer as a part of the AI process to reduce time and energy consumption. The third one is much more speculative, but interesting. And that is there's some evidence that if you train a model on a quantum distribution rather than on a classical distribution, that you can get a more accurate model. Now, the intuition behind this is that quantum distributions have more degrees of freedom than classical distributions. So you might be able to get a better fitting model.
This is kind of like if you want to fit data points to a line, it's not going to fit as well as if you fit it to a quadratic curve versus a cubic curve. So quantum distributions have those additional degrees of freedom. And so there's some evidence that you may get more accurate models. I'm not entirely convinced of that, but we're exploring that as well. So those are the three areas where I think there's a lot of synergy between AI and quantum, and we're doing work.
Perfect. Excuse me. The company is not historically focused on complete system sales, but last week you announced such a sale, and this is sort of a little bit of a shift. Can you talk about how selling complete systems fits into your business model going forward?
Yeah. So you're absolutely right. Our business model has been focused on quantum compute as a service. We have our quantum computers. We have our quantum cloud service. The computers are called Advantage. The cloud service is called Leap. We currently have four quantum computers in production in our cloud service, two up in Vancouver, one in Southern California, one in Germany. And we decided to pursue that model because we are much more focused on commercial than everybody else. Everybody else is more focused on government research grants, but we're focused on commercial and helping businesses actually use quantum. And so we wanted to lower the barrier to entry for them. And so it wasn't about shelling out $5 million or $10 million or $20 million to buy a quantum computer.
It was about spending $300,000-$400,000 to get started, maybe with a proof of concept, maybe get some early applications into production and grow from there. However, we've started to see increasing interest in owning our systems, mostly because when you own the system, there are parameter changes that are accessible to you that may be useful in certain environments, and you have more flexibility in how you integrate the quantum computer with classical systems. So for example, you could imagine a quantum computer in a supercomputing center focused on AI workloads where you need to really get the latency down between the massively parallel GPU systems and the quantum computer.
Got it, and so you can kind of on-prem tune sort of the capabilities of the annealing system.
You get tighter, faster interconnect and integration, and you can better tune the parameters.
Okay. And what are you seeing in terms of just the pipeline of opportunities for on-prem annealing systems? Is that something that you've seen building over the last year? How big could that opportunity be? I'm not asking for financial guidance, but are you excited about this becoming a substantial part?
I mean, look, I think both the quantum compute as a service and the system sale models will be important for us. Quantum compute as a service takes a bit longer to build the revenue because it's ratable revenue. So it's nice because you get the backlog, but it takes time to build that up. System sales, it's a quick hit of revenue, cash and revenue. And there's nothing wrong with that. So I think a good mix in the model can be quite important and quite valuable. And we have, over the course of the last year, seen more and more interest in having an on-prem system.
We announced that Davidson Technologies, a government contractor in Alabama, has just finished a new headquarters building, and they're going to be housing one of our quantum computers there. And then we just announced that another company has purchased a quantum computer. We haven't announced two, but that announcement should be coming out within the next week.
But Davidson is not.
Davidson is not.
This is a separate system.
Yeah, this is different from Davidson.
So Davidson will be part of the Leap service that the computer has to.
The system that we install at Davidson will initially be a part of the Leap service. The work that we're doing, it's just going to take a little bit longer, is to lock down that facility. It's a secure facility that can run classified applications. That will be the first of our systems and maybe even the first of any quantum computer that would live in a secure facility.
Got it. So more dedicated to the government. So still a quantum computing as a service model, but not open commercial.
Yeah, and when we do that, the front end of our quantum cloud service actually runs on AWS servers. It's not the Amazon Braket quantum cloud service. It's our own Leap quantum cloud service, but we run the front end of that software on CPUs and GPUs at AWS in Portland, actually. In order for classified applications to run, we'll have to duplicate that environment in the secure AWS environment, so that's work to be done.
You'd mentioned Davidson. You mentioned Pattison. You mentioned NTT DOCOMO. But maybe can you give us some other recent examples of D-Wave customer use cases, what folks are using the annealing systems for?
Yeah. I mean, I think that my favorite use cases really are Docomo, Pattison Food Group, Interpublic Group, Big Ad Agency, Ford Otosan, Mastercard. It's interesting because it's across a variety of different industries and use cases, but with kind of interesting ability to take those use cases elsewhere. For example, workforce scheduling, broadly applicable. So we're working now to take that to other companies in other industries. We talked about SavantX and cargo loading and unloading at Pier 300, Port of LA. So yeah, we've got a good list.
Can you talk about some of the benefits and opportunities integrating high-performance computing and quantum computers that you're working on?
Yeah. I talked a little bit about it a few minutes ago, but this really is all about the ability to kind of have workloads that require some portion of classical compute and some portion of quantum compute. So our hybrid solvers that I mentioned previously, that's exactly what they do. We use CPUs and GPUs on the classical side and then the quantum computer on the quantum side. And what happens is the classical compute takes in the full problem. It essentially looks for the hardcore and sends that to the quantum computer. So we're solving some really interesting problems using that today, but we're solving it on a couple of GPUs that we spin up at AWS when we need them. Now imagine doing that in a supercomputing center.
Imagine, for example, if we did that at Oak Ridge, where they have Frontier, which is one of the most powerful supercomputers in the world, and it was tightly interconnected with the quantum system, we'd be able to solve even larger problems even faster.
You've talked about growing government interest in annealing systems and near-term quantum computing applications. Can you share more about what you're seeing in that area? And then maybe a related question, the National Quantum Initiative Reauthorization Act, if that gets signed when the new administration comes in, how does that benefit your business?
Yes. So a couple of things. First of all, the U.S. government, through the National Defense Authorization Act last year and this year, as well as through some new language in the National Quantum Initiative Reauthorization, is being asked to, A, be inclusive of all quantum technologies, so not just Gate, and, B, put a focus on near-term applications. Now, that's very interesting because, as we just discussed, the only company that's ever been able to show better than classical on a real-world application to date is D-Wave with its annealing quantum computers. So if we're talking about near-term applications that benefit the government, that's a great opportunity for us because we're kind of unique or further ahead than anybody else in the near-term application space. So we think that's going to play well for us.
As far as the National Quantum Initiative is concerned, that is the premier U.S. government funding mechanism for quantum. That was passed under Trump during his first administration. It needed to be reauthorized about a year and a half ago. It still has not been reauthorized under the Biden administration. We believe that when Trump comes back in, it will be authorized. And as a result, the premier funding mechanism from the U.S. government for quantum will be back in place and kind of open for business again. So we're looking forward to that.
Last question for me, and then I'll pause and see if there are any from the audience. The company made two announcements last Friday. The first about 2024 bookings exceeding $23 million, which certainly surprised me. It was a much bigger number than I was expecting. I think that's up 120% year on year. And then you also announced a new $150 million ATM facility. I know you're in the quiet period, so you probably can't say too much, but what can you share about those two announcements?
So first of all, I talked a bit ago about system sales being like a quick hit. So okay, we sold a system in Q4. That was our first system sale. So a portion of that got reflected as bookings in Q4. And that was one of the drivers of the 120% year-over-year bookings increase. But even more interesting, that was $18 million in bookings in Q4, which is a 500% increase over Q4 in 2023. So we felt very good about that. We think that the system sale as an additional component to our business model may allow us to kind of grow the business faster because, as I said, quantum computers as a service takes time to build up, whereas system sales can go a bit faster.
And so a good balance between the two, we think, will be important. Secondly, we also announced that we now have over $175 million in the bank. And that is the largest cash balance that the company has ever had. And so we now also have all the cash that we need to execute. We did another $150 million ATM just in case we need it. I mean, what was nice about the end of last year was that we were able to raise a fair amount of cash toward the end of last year with very little dilution. I mean, when the stock price is at five or ten dollars a share and you're raising $75 million, that's not a lot of shares. So we're at the point now where we think we've got the cash we need.
We're not worried about cash, but we have the ATM to be opportunistic if there's a point in time when we can raise some more cash without a lot of dilution because we do care about our investors and we don't want to drive dilution.
It sounds like it's more flexibility.
It's all opportunistic. Yeah.
Got it. One last question around that hardware sale. Obviously, somebody's got to maintain service, perhaps do upgrades. Is there a service element to the hardware sale? Does that get booked at the time of sale? Does that come in over a couple of year periods?
Yes and yes. There is a service and maintenance component to the sale. And so that revenue does get recognized over the life of the system as we continue to maintain it versus the revenue associated with the actual system sale itself, which would be recognized when the system is delivered.
Got it. All right. I want to pause here. I know we got a lot of folks in the room. If we have any questions from the audience?
Talk about Mideast as customers. A lot of interest in quantum and AI and a lot of money flowing into U.S. companies. I'm not talking about the investment side, but as much about use cases, and I am familiar with some people working in those areas, so I'm curious if that's an opportunity for sales or is that going to be shut down by export controls?
That's a really interesting question. Currently, we have no export restrictions. Quantum is a sensitive technology, but the way the export restrictions are written, they are written around gate model qubits and achieving a certain level of fidelity and a certain level of scale. That's the point at which export restrictions come in. There are no restrictions on annealing. Will that continue to be the case? Not clear. We have chosen to operate as if there were restrictions. We do not sell in restricted countries. We do not allow companies in restricted countries to access our cloud service. There was a paper from some Chinese researchers on factoring semiprimes on our quantum computer. That was done about six months ago. They did it while they were in school here in the U.S.
But we're actually kind of making changes to how you access our system so that that also won't happen in the future. So while we have no export restrictions, we also care a lot about doing the right thing and protecting sensitive technology because we know this technology is really important. Now, as far as the Middle East is concerned, I was in Dubai and Abu Dhabi. I was in UAE September of last year. We talked to a number of different potential customers about application opportunities, mostly the Abu Dhabi National Oil Company, ADNOC. And they have quite a few interesting applications that we're looking into.
But that one is likely going to result in a system sale because, at least for government-owned businesses like the national oil company, the data can't leave the country. So we would have to have a system installed there. And so we're looking at all of that. Yeah.
Yes.
Hi. I have a question. So I understand that D-Wave is very good at annealing, so can do optimization. And you can do certain maybe optimizing for AI. But for AI, the mainstream based on the deep learning is all based on the neural nets, which is heavily based on the matrix operations, the matrix multiplication operations. That does apply for all these CNN, RNN, LSTM, transformer all these. So do you think for quantum annealing or other methods can be useful for this kind of mainstream operation instead of just optimization?
Yeah. We can solve linear algebra problems. So we can address optimization. We can address linear algebra. We can address factorization. We cannot address differential equations. That's the one area that annealing can.
How do we solve the matrices, let's say the GEMM, VMM, those like?
Yeah, so let me talk about actually how we do this, and I'll use a little bit of jargon, but I'll try to explain it. So one of the more interesting models or approaches to building models is through what's called a restricted Boltzmann machine. Now, this is just a mathematical construct that has a lot of parameters associated with it, and the way you train the model is by kind of building a restricted Boltzmann machine that has parameters set almost randomly and generating samples, see if those samples match the distribution from your input dataset. And if they don't, tuning the parameters in the model and repeating the process.
Turns out that restricted Boltzmann machines are not used all that much in AI these days because the computational load is way too heavy, so there are computationally simpler approaches, but not necessarily as good. Restricted Boltzmann machines fit natively into the annealing quantum computer. So that's exactly the computation that we solve. I said it's the low energy point in a multidimensional landscape, but the way you get there is through a computation that looks a lot like the computation that a restricted Boltzmann machine would go through to generate samples. So that's essentially the way we do it. It's not like we are doing matrix multiplication natively. We're building the models in a slightly different fashion.
Because from my understanding, the VMM, GEMM, matrix operation accounts like 70%-90% of the operation of the whole AI. I don't see the Boltzmann machine was kind of like a neural operator in the mainstream, like TensorFlow or anything. Do you think that one this can be in the?
Because it puts a much heavier computational load on the processor. That's why restricted Boltzmann machines have not been used. But now we can actually use them and do them even faster than the matrix operations on GPUs.
Thank you.
Any other questions?