Quantum Computing: Where We Are and Where We’re Headed | NVIDIA GTC 2025 Fireside Chat

[Music] Heat [Music] Welcome to the stage NVIDIA Founder and CEO Jensen Huang. [Music]

Good morning, welcome to Quantum Day at GTC, the first of its kind. This is going to be a very special event. As you know, I'm a public company CEO, and every so often, someone asks me a question. Most of the time, I'm trying to lower the bar here; some of the time I say something right, and sometimes it comes out wrong.

So, what happened was somebody asked me how long before a quantum computer will be useful. Remember, this is from someone who's built a computing platform, and to me, building NVIDIA and CUDA and turning it into the computing platform it is today has taken us over 20 years. So, the idea that time horizons of 5, 10, 15, 20 years is really nothing to me. Of course, quantum computing has the potential and the hope, all of our hopes, that it will deliver extraordinary impact. But the technology is insanely complicated, so the idea that it would take years to achieve was something I would expect because of the complexity of it and the grand impact it would have.

So, when I said the answer, the next day I discovered that several company stocks, apparently the whole industry stock, went down 60%. And then I started to learn about this, and my first reaction was, "I didn't know they were public. How could a quantum computer company be public?" Anyhow, I discovered they were public companies, and I'm very happy for them. Yes, I'm very happy for them.

So, I said, "Listen, the world's got this wrong. Let's invite all of those companies and more from the quantum computing industry, and to the extent that they don't bring cabbages and apples and stuff to throw at me, this would be an extraordinary moment to learn about the state-of-the-art of quantum computing." There are so many different approaches: trapped ions, neutral atoms, superconducting qubits, topological qubits, quantum annealing, photonics. I thought, wouldn't it be amazing if the CEOs, the technology leaders, the companies leading this pioneering technology, were coming together for the very first time to talk about it? And, of course, in the process, they could explain why I was wrong. This is going to be the first event in history where a company CEO invites all the guests to explain why he was wrong, but I don't know. That's what makes this event so great.

Anyhow, NVIDIA doesn't make quantum computers, but we dedicate ourselves to creating accelerated computing stacks to enable quantum computers. We do the same with self-driving cars. As you know, NVIDIA is probably more integrated into the world of automobiles and autonomous vehicles, and we work with just about everybody in some way to advance autonomous vehicles. Yet, we don't build cars. NVIDIA has a broad range of technologies, product offerings, libraries, and computers. We call it the three-computer solution to help advance robotics in all forms: facility robotics, factory robotics, factories that are going to be robotics to build, orchestrate robots that are going to build products that are robotic. It's an incredibly complicated set of computing, libraries, algorithms, and models, and we approach it as if we are deeply integrated into the ecosystem and industry, and we care deeply about them, yet we don't build robots. We don't build quantum computers, but we are deeply integrated into the quantum computing industry, and we create libraries. CUDAQ is a programming model for hybrid classical accelerated quantum. We have cuQuantum libraries that help you simulate quantum circuits, and DGX Quantum to do error correction of quantum computers. We partner with them, we support them, we help them in any possible way, but we try not to say anything that trips them up every now and then.

So, anyhow, we care deeply about this ecosystem, and I'm really happy to bring many of our partners on stage. I just want to let you know that there were many that we couldn't invite, and I want to thank all of you for your partnership and friendship, and we'll try to invite you next time. But before I do that, we're making an announcement. I'm making an announcement today that NVIDIA is starting a quantum research lab in Boston. It will likely be the most advanced accelerated computing, hybrid quantum computing research lab in the world, and it's going to be located in Boston so that we can partner with Harvard and MIT. Some of our partners will be in there initially, but many others will be working in this quantum research lab in the long term. Quantum Machines and Quantinuum are going to be the inaugural partners with us to build this quantum research lab. So, I'm very happy about that, and we're going to get that going as soon as possible.

Now, what I'm going to do is introduce some of my CEO colleagues. Please join me to welcome them. Alan Baratz. [Music] D-Wave, Ellen, Peter Chapman, IonQ, Peter, thank you. Let's see, I think the next one is Loïc Henriet, I think, right? Isn't Loïc? Okay. [Applause] Welcome, and Rajeeb Hazra, Continuum. Rajeeb and Subodh, come on stage. We're getting nice to see you. So, I think the first thing is, sorry about that. That was funny. That was, you know, for everybody. We do. Hang on a second. Come on, Mikhail. Goodness gracious, I left QuEra behind. Sorry about that. So, it reminds me of a joke. It reminds me of a short story. This is nothing like you. So, anyways, I was left behind. This was in 1995, and we had just started NVIDIA, the 3D graphics ecosystem. There were a whole bunch of 3D graphics companies. It was like a new 3D graphics company per week, and we were the first one to start. But after a couple of years, there were like 50, 60 competitors, and we started. We created an architecture which we chose a technology which was exactly wrong. Okay, and so even though we were the first company to start, because our technology was exactly wrong, we were about to go out of business. And this financial analyst, now an industry analyst that follows the industry, kept a list of all the companies that were building 3D graphics, and it was being published every single week. Then one week it came out, and we were left off that list. So, I was left behind, and I said, "Why'd you leave us off the list?" He said, "Well, I thought you guys went out of business, but you guys are doing great." Sorry about that. Great story. Great story.

All right. So, so, listen, each one of you has chosen different approaches. There are quite a few different approaches to quantum computing. Maybe what we could do is start by having each one of you take a moment and talk about your approach and why you did it. And since Mikhail, since I almost left you behind, why don't you start?

Thank you. First of all, thank you for hosting us. It's amazing to be here. Thank you. And also, thank you for your contribution to this emerging quantum ecosystem. So, we are building quantum computers literally from single atoms, and we assemble and control them using arrays of laser beams, using basically techniques like holography, techniques similar to those used to project, for example, to beam computer images to the big screens. The key advantages are that the atoms are basically god-given qubits. They're all identical, and they are extremely well isolated. We can preserve quantum states for a very long time. But also, we can use light, we can use lasers to control these atoms, position them at will, and move them around, including both during the computation process itself and in particular, it allows us to build a processor where connectivity is basically a living organism. It evolves during computation itself, and this is very special. This allows us to build systems now which have thousands of qubits. It allows us to deploy, for the first time, these techniques of error correction which you mentioned, and execute algorithms with these so-called protected logical qubits. That's great. And this is, yeah, it's a very special approach, and we're in a very special time using this approach. Yeah.

Thank you, Mikhail. Go ahead. Let's take turns. Go ahead.

Uh, so thank you for the opportunity. So, at Rigetti Computing, we develop superconducting gate-based quantum computers. We are based in Berkeley here, and we also have a fab in Fremont. Why superconducting gate-based quantum computing? Gate-based because that's how we know how to do the broad world of computing. That's how classical computers run. Why superconducting? That's an area where, along with us, there are many other companies, including some big companies like IBM, Google, Amazon, as well as the government of China, which is investing heavily in superconducting gate-based computing. The reason we like superconducting is primarily because of its advantages in scalability and gate speeds. We are using fundamentally a silicon chip, so we know how to scale up once we leverage the semiconductor industry and five decades of experience there. And because we are dealing with electrons, our gate speeds are in tens of nanoseconds, and that makes it very easily compatible with the CPU, GPU ecosystem, which is the way we think quantum computing is going to come along. We feel very good about scalability and gate speed. The challenge and Achilles heel, if you will, of superconducting quantum computing has always been fidelity. We get noise because of these intrinsically engineered devices in the chips, just like CMOS technology, and historically, that was in the low 90s and mid-90s. When the qubits entangle with each other, the two-qubit gate fidelity is what we call it. What's been exciting is that in the last few months, Rigetti, along with some other companies in superconducting like IBM and Google, have made some very important strides, and now we are in the 99 to 99.5% two-qubit gate fidelity, which is commensurate with the best out there in other areas. So, we maintain the advantages of scalability and gate speed, but now we feel very good about where we are with fidelity. So, that makes it a lot more attractive within this space. We differentiate ourselves by an open modular approach. We have designed our stack so that if we find a more creative solution out there, whether it's an error correction from a company like Riverlane or CUDA Q from NVIDIA or quantum machines for control systems, we can easily integrate that into our stack as opposed to some other companies like IBM or Google, who have designed it in a mainframe approach. We believe an open modular approach is the right way to build an ecosystem while we are in R&D. So, in the overall journey, our flagship system is an 84-qubit system with 79-nanosecond gate speed. It's available to anyone on AWS and Azure right now. We believe it's one of the best, but frankly, it's not good enough yet for any practical use, as we talked about earlier. And, um, we think we are roughly, well, don't give up yet. Don't give up yet. Let's hear why.

Don't you start, thanks, Jensen, for hosting us. So, at Quantinuum, we build quantum computers using the trapped ion modality, as it's called, and a particular architectural approach called QCCD or quantum charge coupled device. The beauty of this approach is that it produces the industry's highest fidelities, 9s and beyond. There are a lot of challenges in quantum computing, as you know. However, coherence time, scalability, fidelity—these are some of the most challenging properties. And as you're thinking about and listening to these different quantum computing approaches, just listen for those words. I think they consistently help you understand the pros and cons or the challenges and opportunities associated with each one of the technologies. Okay, go ahead, Rajeeb.

So, thank you for filling it in. Basically, we have the industry's highest fidelities. You asked about what's our approach to it. Our approach is to extend the QCCD architecture into higher levels of scale. So, we will have 50 logical qubits, highly reliable qubits, this year, 100 logical qubits in about 18 months, and we see a clear path to millions of qubits early in the 2031-32 time frame. Another part of our approach is not just building the science of it. Our approach is we work on really, really challenging problems. It's like living on the edge. We work on them with hardware and software we build and with customers. So, we aren't trying to look for a solution looking for a problem but start with the notion of what big, hairy problem are we going to lunatically go attack, and then build our capability across hardware and software to be able to do that. And I hope we get a chance to talk a little bit about what else is out in the industry today. That is a beautiful combination. Well, welcome to GTC. [Music] We have a lot of hairy problems here.

All right. So, thank you, Rob. Yeah, Loïc, go ahead.

Thanks a lot for hosting us. At Pasqal, we build quantum processors that leverage neutral atom technology. This is quite close to what Mikhail and QuEra are doing. This technology has several key advantages, like scalability. We can trap and control many of those qubits right now—thousands of them. Also, it's easy to control them with laser light, and it's a relatively recent technology compared to trapped ions and superconducting qubits for developing quantum computers. Actually, it's more recent, but there is a lot of progress and momentum in terms of gate fidelities and scaling of this approach. At Pasqal, what we are also committed to doing is working very strongly on the engineering of those devices to turn them from lab experiments into real industrial products. I think we all agree here that usage and adoption are very key for the entire quantum community right now, and at Pasqal, we really want to deliver on that promise. Over the past 18 months, we've delivered, and we will deliver, four machines worldwide, including one in France, one in CANI, and another one in Germany at the Jülich Supercomputing Center. That's about it for Pasqal. We work with neutral atoms and focus very strongly on engineering.

That's terrific. Thank you, Loïc. Go ahead, Peter. Please tell us about your company.

So, I'm the chairman for IonQ. We're a trapped ion company, like Quantinuum. Trapped ions were actually used back in 1995 when they were looking at atomic clocks. Atomic clocks and our technology have a huge overlap. Back in 1995, a team at NIST, and one of the co-founders of IonQ, did the first-ever quantum logic gates. All of this craziness that you see coming all started in 1995 from that experiment. So, we're now 30 years into this investment in trapped ions. Very similar to other modalities, we use individual atoms and we use lasers to do computation. We're down at 0.02 nanometers. When you look at the silicon industry, they're way up there compared to where we play. We play with individual atoms. So, the advantages of our technology are that you can have a room-temperature quantum computer. It can fit in a rack, and to be honest, it looks kind of boring compared to what you probably imagine because it'll be rack-based and room-temperature. The other advantage is that, because we use optics and lasers, you can network them together to do distributed quantum computing to get to larger numbers of qubits, and you can reuse the existing infrastructure of the internet using fiber optics. And as has been mentioned, trapped ions have the best average two-qubit gate fidelities. In that sense, they lead, and that's a fairly large advantage because it means the amount of error correction you need to do will be less than maybe some other modalities. But each one of us, and I just say since 1995, it's amazing how many different modalities have shown up, from free qubits to the amount of progress that's been made. So, it's really quite exciting from that point of view. It's great to see the leading companies on stage here today with us. So, yeah, thank you, Peter.

Alan, go ahead.

Thanks, Jensen. So, we are a superconducting company, similar to Rigetti. We believe that superconducting provides the best balance between qubit fidelity or the quality of the qubits and gate speeds, time to compute. But we're actually quite unique from everybody else on this panel and pretty much everybody else in the industry because we use annealing technology as opposed to gate model technology. Without going into the details, annealing is a much easier technology to work with. It's easier to scale, and it's much less sensitive to noise and errors. And probably the best proof point for that was in a paper that we published in Science last week, where we performed a useful computation of properties of magnetic materials that would take nearly a million years to compute classically. And then this week, we actually put a paper on the arXiv where we showed how to use that computation to create quantum proof of work in a blockchain. The idea would be that you would use the quantum computer to create the hashing function and use the quantum computer to validate the hashing function. We now have this running on four of our quantum computers as the first distributed quantum application, where we're generating hashes, validating hashes, and we think this could be a very interesting, much lower energy consumption approach to blockchain. You know, every time someone in the quantum computing industry achieves a milestone, it stirs up a fair amount of controversy among the others. Did you stir up any controversy with your achievement, given that the achievement-to-controversy ratio is literally one to one? So, the answer to the question is that I've received a lot of positive feedback from my colleagues in the industry. Just to be clear, and only since you asked and I don't like to name names, but I will. There was a paper that came out of some researchers at the Flat Iron Institute in New York. This is a really solid research team. What they were able to do was advance the state-of-the-art on classical computation in tensor networks, and what they've been able to show is that for the smallest instances that we computed, which we also computed classically on GPU clusters, they could do it a little bit faster. Now, they made some claims about how that undermines the results, but not at all. We computed multiple sizes, multiple lattice sizes, multiple evolution time frames, multiple properties on the lattice, and so this is a very strong result that's actually been in the public domain for over a year now.

Yeah, that's terrific. So, I guess the question that stirred up a fair amount of excitement is really about what is the definition of usefulness in quantum computing, and when do we expect that? Before we answer when we expect that, maybe we'll build up to it. What are some of the early applications you think that would be worthy of the endeavor of a quantum computer? Number one, and number two, how do you define usefulness?

Maybe I will start kind of at a high level. Quantum computers are really a fundamentally new scientific and engineering tool. If you look at the history of science and technology, whenever you come up with a new tool, the first thing you use it for is to really advance the science and actually make scientific discoveries. In fact, quantum computers literally allow us to go into corners of the universe where we have never been. If you go to these corners responsibly, you always find something interesting. So, I believe there is a huge potential to use the machines which are either existing already now or which are being developed in the near term to really advance this scientific frontier and actually make new discoveries. There have already been discoveries made using quantum computers, but to be maximally honest, there were very few. The way many of us are thinking about it is that now we are in the era of quantum discovery, where we can use these machines to actually explore the physics of complex systems, maybe related to things like chemistry and material science. The field is now really ripe for using these machines to push into these scientific directions and start making discoveries. Often, they are things which may not be directly commercially relevant. For example, understanding properties of systems away from equilibrium—much of the world around us is not in equilibrium. These are the kinds of things where I expect a lot of progress will be made in the next few years, and often these are the things that then translate to applications and start new industries in a way that is not possible to predict. That's why this field is in a special point right now. That's the trapped ion perspective. What about a superconducting or neutral atom approach?

Neutral atom. It's a small differentiation, but I'll follow on what you were saying. We agree with the fundamental premise that scientific discovery is going to be taken to a new frontier. But we are seeing applications today. As I said, we focus on what is the big problem for a customer or a partner that we want to solve. We are seeing applications in the area of chemistry, like how do you get to new refrigerants that have certain sustainable properties? How do you generate hydrogen from water more efficiently without needing platinum as a catalyst for the reaction? In biology, we're looking at how peptides bind. These are very specific instances, and doing that gives us a good way to understand two things: what algorithms do you need to attack it with, and what capability do you need in the machines at a certain point in time? You ask the question of what the performance standard is. Coming from a classical background, you had performance per watt, then performance per watt per dollar. We're getting to a point where, if you look at it through the lens of big problems you want to fundamentally solve, with a figure of merit of either solving it more accurately or solving it more accurately and with less energy and cost, we are getting to the point of what is your scale of computation, usually the

Number of qubits, but also what fidelities and error rates you can sustain to make those qubits useful. I'm not saying there's a perfect ratio of those things, but they are generally leading us to say how useful is your powerful quantum computer, and that can only be done through the lens of looking at big problems and saying how do you solve them with the help of a quantum computer, not necessarily replacing a classical computer with a quantum computer.

Rajeeb, one of the areas I do wonder about is whether quantum computing is just poorly positioned. Let me take a swing at it. There are so many things in industry built on fundamental sciences, and quantum computers, in their broadest sense, can be the ultimate instrument for understanding the basic sciences that affect that industry. However, because it was described as a quantum computer instead of a quantum instrument, people have a notion about what a computer is—you should be able to run Excel super fast, and every respectable computer should be able to run a game like Crisis. There's a common sense about what a computer is, and it's attached to memory, network, storage, and a programming model. I wonder if it's just a wrong mental model, and as a scientific instrument, it is extraordinary. As you say, the opportunity to understand science deeper along the way is extraordinary, but to position it as a quantum computer per se and hold it to the standards of a computer per se that we all understand, I wonder if that could be a reframing that allows this entire industry to be much further along, frankly, in that reframing as a scientific instrument for very important industries.

Go ahead, yeah. I totally agree with what you said. In some sense, the term "quantum computer" is misleading because people expect that you can replace a classical computer with a quantum one. It's not like that. It's more like a very complementary tool. We like to call our machines quantum processors—very specialized machines that you can use in a complex workflow alongside CPUs and GPUs, but really for specialized tasks. Once everyone agrees on that particular way of using quantum computers or quantum processors, it will be easier to work alongside classical computing and not work towards replacing all the compute capacities that are in place.

I'm actually struggling with the concept. I don't know how to think of a quantum computer as an instrument when it's being used for materials discovery, blockchain, or improving cell tower resource utilization. It's true that there are many applications I would never try to run on a quantum computer, but for applications that require extensive processing power, these machines are very powerful, and I think they go well beyond just instrumentation or measurement.

Sorry, I'll jump in. It's okay. Yeah, I'll jump in on that. I was actually just trying to help. We saw your help. You know, let me tell you, this whole session is going to be like a therapy session for me. A long time ago, someone asked me what accelerated computing is good for. I said, a long time ago, because I was wrong, that this is going to replace computers, this is going to be the way computing is done, and everything is going to be better. It turned out I was wrong and unnecessarily wrong. It's better to be narrowly focused on something and be extraordinarily good at it. But the moment you cross that line and start talking about the traveling salesman problem, it became unnecessary because that problem is obviously being solved as we know it today. Uber cars are showing up, maybe they're three seconds late or 30 seconds late or whatever, but they show up. I do think that holding ourselves to a bar to solve a problem that is unnecessary for quantum computers to solve, quite frankly, to change the world, takes focus away from something you uniquely do, and quite frankly, sooner than later. Anyways, that was just my swing at it. Go ahead, Peter.

Well, here at the show, we are actually showcasing several applications that are showing quantum is now one of them. One is with ANSYS, and you might know one of their products, which is LS-DYNA, and it runs obviously with GPUs today. We announced that we've integrated our quantum computers with LS-DYNA, and we saw a 12% increase in performance in modeling a blood pump. So, this is the first time, I think, that quantum has been integrated with production software. We also announced, with NVIDIA, AWS, and Atos, a 20x improvement in a chemistry application. What's amazing about it is that we did that on 36 qubits with our existing system today. By the end of the year, we will have 64 qubits. Every time you add a qubit, you double the computational power. So, that's a 2 to the 28th increase in a single generation of chip, which is roughly 260 million times more powerful. By the end of the year, one would expect that for things like LS-DYNA and for chemistry applications, there will suddenly be huge performance increases. So, we're working right now on these applications, which, to be honest, many of you use, to be able to have a significant impact using a quantum computer. I do think that your statement about 10 years is interesting. We think of ourselves as being where you were 10 years ago. We hope that, obviously, 10 years from now, we'll be up there with NVIDIA and all the other giants in the rarefied air. But it does take a long time to go from a startup to where you are today. It's completely fine to sit down and say that for the quantum industry, it's going to be another 10 to 15 years to get to where NVIDIA and all the other giants are. It's just not going to start then; it's starting today. You're going to be much larger than NVIDIA. There's no way we're going to be a relic of the past.

All right, so if it seems like there are so many different approaches to quantum computing, and it is so diverse in its approaches, why is it that this industry doesn't quickly discover a more promising approach, as you see each other's work, and naturally, through evolution, people select the best approach and then everyone starts to advance the whole industry in a unified way much more quickly? As I observe this industry, it's surprisingly diverse, and there are thousands of flowers blooming. When does it become a garden? I'll just say that if you look across us today, there's a number of people you heard are using individual atoms, lasers, and all the rest. So, we actually have more commonality than most people expect. Often, and so I would think that, in the future, there is more sharing, and maybe even the ability to work together because the promise of quantum computing and what it can do for mankind is so significant that it's actually larger than any one of the companies that are here sitting on stage today. So, mankind has a whole range of significant problems to be worked out, and we need quantum computing to be able to solve them. We're still obviously new, and ways to build qubits are being found every day, but I think that over the next couple of years, we will start to coalesce to probably two or three different approaches. Some of us will probably come together because we do share the underlying technology. And so, it makes sense.

Many of the problems you've described, the precise answer is not exactly known because, as you know, fluids are quite chaotic, and it's hard to know exactly what the right answer is. In those kinds of examples, using AI for emulation can give you tens of thousands of times speedup, orders of magnitude speedup from where we are today with principal solvers. How do you guys think through that? What is the point of solving that problem when classical still has orders of magnitude of progress in the next couple of years ahead of it?

Well, there are orders of magnitude of progress, but at the same time, there are problems that are just impossible to solve classically today. There are problems in the area of drug discovery, global weather modeling, and even the application I shared, which is the basis of our paper a week ago, computing properties of materials. We use Frontier, which is basically one of your systems, a massively parallel supercomputer at Oak Ridge National Lab, and it would take millions of years to perform the computation. So, the point is, there are still hard computational problems that are out of the reach of classical, and AI isn't going to address those problems either. They're just out of the reach of classical, right? Exactly. Go ahead, Rajeeb.

So, I'd like to make a point. It's interesting that it took us 30 minutes to get to AI, but from our point of view, where we see it quite interestingly is as an extension to your question on whether we should call it a computer or not. These quantum devices or tools or instruments, as you call them, are expanding the ability for us to access data to train these AI engines that previously was not possible. So, if you're going to solve a chemistry problem today, humankind's max ability is defined by things like density functional theory or other approximations to the quantum space. The world of that information that we haven't had is like trying to train autonomous vehicles by giving them city grids of 500 feet squared and not having the detail of lanes or other things. So, what we see is this concept of a computer brings in the idea that I have a computer A versus a computer B, so A must run the thing faster than B for it to be better, at least technically, until you tell them what the price of A is. We don't see it that way. We see it as what can A and B do if A is the established classical, well-honed frontier models, how are we training those models, and are we enabling those models with the data so it can now continue to be agentic and continue to reason and continue to do things that otherwise we'd be pulled back to do? We call that gen, not AI, gen Q AI, and that kind of breaks the paradigm of a computer competing with another computer. It's two computers now working together, two completely different ways, but they are input and output for each other. You're right. The output of the quantum computer is the input into these large language models (LLMs) and the training methodologies. So, you can have LLMs that actually understand things like ground state energy and ground state configurations of molecules. You can then use them to start doing perturbation theory on whether a molecule will last inside the body and deliver the drug at the right kinetic paradigm or not. That is how we see quantum as an addition of a tool or an instrument into what is already a developing and rapidly maturing and improving compute paradigm.

In the few minutes we have here, what are the things that we could do in the world of accelerated classical computing to be helpful to all of you so that we could advance your work much more rapidly? Can we start with the middle, come on, Mr. Frenchman, let's go.

Yeah, thank you. I guess it's very important, as was said earlier, to couple as best as possible the various compute modalities, like classical and quantum. For the moment, it's not really a matter of bandwidth or being able to collocate to do things fast because we're not there yet. It's not the pressing paradigm. At some point, it will be a problem, but not now. Now, it's really about identifying the key problems and domains where we can collaborate and leverage the best of both worlds. I fully agree with what you said about using a quantum computer or quantum processor to process and create data on a problem that is inherently quantum, which classical struggles with naturally, and then use that natural advantage in a larger workflow with CPUs and GPUs, and couple that quite well at the software level. I was just going to say that we use your GPUs to design our chips and often do co-simulation to make sure that the quantum computers are working. When we look to the future for quantum computing, it's going to be a set of classical systems sitting right next to a quantum computer, and the two of them are going back and forth. It isn't something where one is replacing the other; they're working together. If you look at the same things we're doing today, we're applying machine learning to be able to figure out how to build optimization not only for the quantum computers themselves but also how they run. So, it is already a synergistic relationship between classical computing and the strange thing is that our quantum computers are almost entirely classical. The only quantum part happens to be a little chip and a couple of atoms at the center. The rest of it is entirely classical. So, I wouldn't short any NVIDIA stock at the end of this. I think you have a strong position going forward. But I would expect that in the future, it will be a QPU, a GPU, and a CPU all working together to solve problems.

In fact, if I could just add a little something, you probably observe that NVIDIA accelerated computing is the largest volume parallel computer the world has ever seen. Yet, we don't call it a parallel computer. It was opposed to sequential computing, and the mistake of that approach, the mistake of that positioning, is that Amdahl's law doesn't work that way. There's no reason to replace something that does an incredibly good job. You should add to it and ride the wave of the momentum that's been created. That's why we decided to call it accelerated computing. It's still a computer, and that really revolutionized how people thought about us and how we thought about ourselves and our work. I think the idea that this is a quantum computing industry or a quantum computer is less good than a quantum processor that's going to make every computer better is a better way to think about it. Rejeeb, go ahead.

Some of the paradigms have to change in the way we are thinking about quantum processing or quantum computing. Some people have observed that thinking of a human brain and how it works is closer to a quantum computer than our conventional thinking of HPC and how HPC should be integrating with quantum computers. We are dealing with analog inputs, analog outputs, and simultaneous multiple variables at the same time, exactly like the way the human brain and our neurons work. Fundamentally, we may be limiting ourselves by thinking of quantum computing in the context of classical computing. We may have to start thinking broader and say, what are the kinds of things we could potentially envision when a quantum computer is brought in conjunction with HPC, and how can quantum computing help genAI get to AGI? Those are some of the trickier things that we could use a quantum computer for.

Well, this is going to be the beginning of a great conversation for the industry. It's a great pleasure for me to host all of you, and this is just the first of many in our series. I'm looking forward to it. Muel, do you want to finish?

Yes, yeah. So, I think these are all great examples. I want to go back to the point I made before. In a sense, a quantum computer is not a hammer; it's a precision instrument. What you want to do with it is solve the hard quantum part of a problem. This is kind of our vision: if you have a problem you want to solve, you want to solve as much as possible with classical computers and identify the hard quantum part. Then, you find an algorithm, a good error-correcting code, the right compiler, and write the right decoder, all optimized with the specific quantum architecture in mind. In all of this process, what you want to do is outsource as much as possible, at least at this time, to the classical part. This could be CPUs, GPUs, depending on what you want to do. And, of course, at the end, you want to use the output of the quantum data to train your models and improve them. That's how we see the real value of quantum computers emerging in the next couple of years. Surely, one very productive use case. Thank you, guys.

Okay, how about we just have to be very quick. Next year at this time, what are we going to be talking about? So, let's just quickly go through. Go ahead, Alan.

Next year at this time, I hope that we're talking about how quantum is helping you to do better model training and inference with lower power consumption. Okay, go ahead, Peter.

First, quantum applications in production, helping customers take workloads and I hope that we'll see, along the same lines, the first prototypes of a new kind of AGI based on quantum, talking about the learning things that we got from all the usage of all the computers and processors that are being deployed today in the field.

Mhm, yeah, Rajeeb, I agree with the previous speaker's theme. We will see, in the next year, the first real tangible use cases of an AI agent working in conjunction with a quantum computer, doing things it couldn't have otherwise done before, and done with a tremendous amount of trial and error.

Mhm, mhm. Okay, Sub, I hope a year from now we are at a point where there's a little less skepticism about quantum computing, and we start talking about how exactly it will be valuable in a data center, and we can show some real-life cases.

Mhm, will be different. So, I want to see 10 new scientific discoveries in physics, chemistry, and biology, and maybe other areas, which would be delivered by the quantum machine.

Well, guys, let's go make it happen. All right, guys, thank you. Thank you, thank you. All right, our second panel. Our second panel. Thank you, thank you, guys. Either way, either way, don't worry. Yep, don't worry. Don't worry. Okay, our second group. Thank you, thank you very much, Rejeeb. Our second group: Ben Bloom from Atom Computing, neutral atom qubits. Go ahead, Ben. Come on in. Hey, Ben. I'll shake all your hands in a moment. Uh, Matthew Canella, CEO of Inflection, neutral atom qubits. Hey, Ben. Thank you. Thanks for coming. John Levy from Seek, superconducting qubits. Hey, man, nice to see you. Perin, CEO of Alice and Bob, superconducting qubits. Ellis and Bob, okay, QCI, Rob Shov, superconducting qubits. Nice to see you, and then Side Quantum, Pete Schulbot, single-photon qubits. Hey, how's it going, Pete? Yeah, sit down, sit down.

So, very quickly, how about let's go through again. Let's start from this side. What is your approach, and why did you choose it?

Yeah, so my name is Ben Bloom. I'm one of the founders of Atom Computing. We build quantum computers with neutral atoms. You heard a little bit about neutral atoms earlier, so I'll reiterate some of the good points and hide some of the bad ones. Generally, we can make very large numbers of qubits. We're one of the first companies to breach a thousand qubits, and we can do this with very high fidelity. We can also do operations with these qubits that are just very coherent. It also allows us to do things like all-to-all connectivity, which allows for a variety of quantum error-correcting codes and applications to be run on the system.

Yeah, go ahead. First of all, thanks for having us, Jensen. Yeah, it's great to see guys. It's great to be up here with you. This is going to be fun. At Inflection, we are also building our quantum computers using neutral atoms. And I think Ben did a great job explaining that, and as did folks on the last panel. So, I'll just say that neutral atoms are a highly flexible technology, and that's because they take place entirely at room temperature. So, because we don't need a freezer, you can actually shrink, cost down, and field-deploy this technology. And so, what we do, I actually brought a prop here. We can trap our qubits in these ultra-high vacuum cells, and then they're atoms, these qubits are atoms, and then we can arrange them and do interesting things with them with lasers. And we think, as I think Misha said in the first panel, atoms are nature's perfect qubits, but they're also nature's perfect clocks and nature's perfect sensors. And so, we actually point this core atom technology at those three areas: clocks, sensors, and computers. You can think of them as a sort of continuum of complexity on what you can do with neutral atoms, with computing being the most complex and clocks being the least complex. We're following a tried-and-true monetization and market development strategy of monetizing those areas where we actually have true quantum advantage today, like clocks and sensors, and using those learnings because there's a lot of leverage. All the underlying components are the same. Those learnings and those gross profit dollars help us push the limits and get to ultimately quantum advantage in the computing world. And so, that's what we do, and we are doing interesting things in the computer alongside your fantastic quantum team, Jensen, Sam, Alisa, and others. So, yeah, appreciate that. Thank you very much, thanks, Matt.

Yeah, go ahead. So, I'm John Levy, the CEO of SEEK. SEEK stands for Scalable, Energy-Efficient Quantum Computing. And what we've heard today is that there are multiple ways of building quantum computers with different kinds of qubits. But what we also know is that qubits alone don't build a full-stack computer. You need to be able to do readout, control, multiplexing, reset, error correction, GPU integration, the full stack. And so, at Seek, that's what we do. We have actually built digitally controlled computers. This is an example. This is Seek Orange, and it's the world's first digitally controlled and digitally multiplexed quantum

Computer, and we're putting all the core functionality of a quantum computer on a chip. Now, the only way that we can do that is if we're incredibly energy efficient. And you talked in your keynote about the importance of energy-efficient systems. And so, if you think about building a regular quantum computer, say, doing superconducting, you might use two to five watts of power to run a quantum computer just to control a single qubit. We've gotten that down to three nanowatts of power. So, we're energy efficient. And the last part is that we're all digital. And so, that enables us to avoid one of the major sources of noise, cross-talk, in quantum computers. But it also enables us to connect to other digital chips like GPUs and CPUs. So, our notion is to create a platform for heterogeneous compute, where we basically take this idea you were saying in the previous panel about computing and accelerated computing, and we think that the way to accelerate computing is to seamlessly integrate the way you've done it with NVLink and a GPU and a CPU, a QPU, and that's the infrastructure we're building.

Yeah, that's terrific. Thank you.

Yeah, to you. Thank you.

Yeah, thanks, Sean.

So, at Alice and Bob, we're superconducting chip designers. Now, we design superior superconducting qubits for error correction. And you know, in quantum, error correction is all you need. So, our technology, the cat qubit, has a first layer of error correction directly built within the qubit. And it's so powerful, so hardware-efficient, it slashes the number of required qubits for impact by up to 200-fold. Think about it for a minute. That's not only reducing the cost and complexity of the system; it's shortening the timeline significantly. If you think of it in terms of Moore's law, it's nearly a decade of head start we're getting. And so, Alice and Bob, this is how we're turning decades into years.

Yeah, it's terrific.

Hi, yeah, thanks for having us, Jensen. This is really a fun event. So, I'm Rob Schulov. I'm one of the founders and chief scientist of Quantum Circuits, which is in New Haven, Connecticut. We're a spin-out from Yale University, where a lot of the superconducting folks were trained, and some of the main ideas came from. I'm glad you brought up the issue of error correction; I think that's a good thing to talk about. At Quantum Circuits, we believe error correction is the key to obtaining useful quantum computations. We actually have a bit of a different take. It's somewhat similar to what Alice and Bob is pursuing, but our mantra is "correct first, then scale." We don't want to make machines with millions of very noisy qubits and then try to figure out how to program those or how to build the error correction as a software layer on top. What we're doing is we have a new paradigm within superconducting circuits. It's a thing called the dual-rail qubit, and that's got essentially error detection built in at the hardware level. So, that's got a couple of advantages. We get all the speed and scalability of superconducting devices, but now we're starting to see performance that rivals the ions and the atoms, trying to square the circle and have the advantages of both of these different types of technologies. So, we think that the enhanced fidelity we can get by detecting the errors is going to get us to use cases in the near term that are interesting, especially for this kind of scientific discovery that was mentioned in the previous panel. It's also a way for us to scale more efficiently to fault-tolerant machines. So, we want to scale, but we want to not just do that in a profligate way. We want to scale in a way that's really giving value and suppressing the errors dramatically. I think the main challenge for the field, whatever the technology, is to show that error correction really works and we can suppress things down to levels that have never been seen before with physical qubits.

That's terrific. Thanks, Rob. Go ahead, Pete.

Yeah, thanks a lot for having us, Jensen. I appreciate it. I really hope they're paying you well to make sense of all of this complexity. So, I think it's fair to describe Side Quantum as sitting on the extreme end of the spectrum of quantum computing companies. From the very beginning, we've been singularly pigheadedly interested in building very large-scale, universal, fault-tolerant machines—on the order of a million qubit scale. The approach we use to build that is we use single photons, particles of light. We made the first demonstration of two-qubit gates by our CEO, Jeremy, over 20 years ago in Brisbane using those photons. Now, we put those on a chip, repurposing the silicon photonics technology that was originally developed for data center applications, which I was really excited to hear you speaking about in the keynote. We think that gives us profound advantages in overcoming the scaling challenges that face our field: manufacturability, cooling, power, connectivity, and control electronics. That leverage has put us in this position where we're now breaking ground in the next few months on very large-scale, data center-like quantum computers in Australia and in Chicago.

Yeah, that's terrific. Thanks, Pete.

You know, one of the things that's really quite challenging for people working around the industry and certainly observers of the industry is that the science is very different in many of the different approaches, and there are quite a few different approaches. If there were just two approaches, you could wrap your heads around it, but there are quite a few different approaches. The science is new, of course, every aspect of the engineering, the manufacturing, all of it is new. Even the programming model, how you think about programming these things, is new. Comparing them is difficult. For example, on the one hand, the last audience was already talking about the usefulness of their computers in running industrial software. On the other hand, there's some common sense about the number of qubits necessary to have a productive and functional system. We're at 36 or x number of qubits at the moment, and now you, Pete, are talking about a million qubits for a fault-tolerant machine and a productive machine. So, how do you bridge that gap from where we currently are, what's the current state-of-the-art, versus where do you think we reach a phase shift? It's likely not to be a very specific point, but the usefulness of these quantum computers will become more and more useful over time. When do you see that transition happening? When do you guys all see that? Where are we today, and where do we kind of likely say, "Yeah, that's a really good quantum computer"? When is that phase shift happening? How do you guys see it?

Well, I think it's really important to just keep scaling. I think Pete's probably right that some of the biggest problems you have to work on that are actually going to change the world are going to require millions of qubits. So, you have to make sure that you are scaling your quantum computer really, really fast. We don't want Moore's law scaling of factors of two or root two. We want factors of 10, and we want it every few years, and that's what we do at Atom Computing. I think that, in the end, there are people who are using quantum computers now who are making progress and finding useful problems, but when you talk about utility-scale things that are going to change the world, you have to get to a million. I think it's important also, Jensen, to define terms because there's not to confuse things—there are physical qubits, and then there are error-corrected logical qubits. Error-corrected logical qubits are really the key to the kingdom here. The ratio of physical qubits to error-corrected logical qubits was once thought to be 10,000 to 1, but it's probably closer to 100 to 1 or now. So, you're going to need multiples of the number of physical qubits and then run error correction software on those to get the logical qubits you need. I think to answer your question, the consensus is around 100 logical qubits, you can start to do interesting things with quantum computers that classical computers can't yet do.

It's interesting to think about how to scale these quantum computers because, for example, there was a really wonderful paper that Google did in late November around error correction. If you looked at the Willow chip and the setup, it was a great demonstration of doing error correction. Each qubit, if I'm not mistaken, required five separate cables. If you think about trying to scale a system with a million qubits, are you really going to have five million cables? So, I'm using this as an example because there are so many things like that in quantum computing that we need to solve. We're solving that by doing multiplexing and by doing chip-to-chip integration so that we can solve that problem. But it's one of a thousand engineering problems. So, it's really a pick-and-shovel approach to taking each one of those and trying to solve them. We can't just solve one of those problems; we have to take a comprehensive view and solve them all. I think there's probably broad agreement that unless we can figure out how to build quantum computers on a chip, we're never going to get there. So, that's the major goal: to scale on a chip.

Yeah, the academic community has it figured out pretty straightforward. It's a hundred logical qubits, as you mentioned, with their error rates remaining at one per million at most. Logical qubits are not completely error-free. I mean, your classical transistors still, from time to time, happen to make an error. Now, the thing is, not every physical qubit is born equal. The size and complexity of your system to get to 100 logical qubits might vary a lot from one platform to the other—from hundreds of thousands of physical qubits on some modalities to just a couple of thousand on others. To answer your timeline question, when does this shift happen? I think it's by the end of this decade for sure. So, 2030, and that's where you see the inflection of the exponential power. You know, an exponential curve, when you zoom out, it's dead flat, and then it's a hard wall. So, where this inflection or just the beginning happens, by 2030, you'll feel the wall climb.

That's terrific.

Yeah, it's an interesting question. I think it's a bit of a fallacy that quantum computing is going to be in development and then there'll be a flip of the switch, and it's everywhere, solving all the world's problems. It's really going to be more like a knob. We're going to be turning up the volume steadily, and you know, we can start to hear the music now, and eventually, everyone will be able to hear the music. But I think also, you know, I liked what you said in the earlier panel about how these are different from regular computers. That's a completely different paradigm. So, a thing that we're doing now is we're really learning how to program these machines, and we also have to learn how to deal with the errors that are always going to be in quantum computers. We're never going to have perfectly fault-tolerant digital computers that work as reliably as a GPU or a CPU does. Because they're going to be used for special-purpose things, and you get the answer right one time, and then you know the answer to a question you never had. I think the right analogy is to think of the early days of electronic computing with vacuum tubes. You know, it was imagined that we'd only use them for cryptography or maybe modeling bombs. And von Neumann had to come up with the von Neumann architecture, and the idea of a compiler was new. So, I think we're in the era now where the machines are powerful enough that we can do that kind of discovery of what it is to program things. We're going to see applications of these things that are not what we anticipate today, is my guess.

Yeah, well, it's really great that you said that. One of the things that we hope for and we'd like to be able to contribute is to help discover those programming models and help invent that programming paradigm. There is an unnecessary expectation, and it actually sets the industry back, frankly, that the quantum computer is going to be better at spreadsheets. That's an unfortunate and unnecessary expectation. The reason why we're involved in all this is because we have such incredible, great grand hopes for you, that we're going to go discover new ways to solve very, very challenging problems, but not so that we can go solve food delivery. I really wish my burger would show up an extra three seconds earlier, but I could live with it. But there are some things that simply won't get solved without quantum computing, and I do think that the framing of our collective understanding of quantum computing is going to be really helpful for the industry. So, can I just follow up on Rob's comment? If you go back to 1946, and somebody dragged you to the basement of the University of Pennsylvania to see the ENIAC, do you think that anybody at that point said, "Oh, I'm going to use this to remotely call a car that's going to optimize the route, be able to pay for it remotely, and communicate that to my friends in more or less real time"? That wasn't what people were thinking about. They were tracking missile trajectories on application-specific devices. And I think the idea of creating an open space to explore and discover is exactly where we need to be, and that's what we need everybody to be working on.

Yeah, that's really terrific. Thank you.

[Applause]

You know, I do want to say that I think one of the big changes between a classical computer and a quantum computer, at least how we understand it now, is this idea that a quantum computer is a big compute resource. A lot of the classical computers and the ways we use classical computers are kind of big data resources. I think it's going to take the combination of classical computers and GPUs and everything to actually understand how you even use a big compute resource. If we succeed, we're going to build a bunch of supercomputers that are really, really good at understanding the physical world, and we have to figure out how to use those efficiently and actually bring them into normal everyday processes.

Yeah, when you spoke earlier about scaling, I'm excited by the fact that you're not limited by Moore's law. Moore's law, as you know, is not based on a fundamental law. There were some principles involved in it, but one of the things that's really great to see from the industry here is the rate of scaling is not a factor of two every couple of years. Because at that rate, it will take 30 years, and we do know we need to scale. If you look at the past 10 years of scaling, it's not an indicator of how fast you guys are actually scaling now because of the new science and the new methods you guys are using for quantum computing and these quantum processors. You're scaling a lot faster, frankly. So, can you guys talk about scaling and where you see scaling in the next 5 to 10 years? What's enabled you to scale faster, and what's the technology that allows you to scale better today?

I mean, I think the answer is that we can use classical computers. We're learning how to build control devices, we're learning how to use light efficiently, and we can just trap more atoms, control more atoms. We start off with clouds of millions of cold atoms, and we have the ability to load 10^7 to 10^8 atoms per second that are just ready to be qubits. It's up to us to go and figure out how to build the control infrastructure around that. That's classical computing, it's photonics, it's RFSoCs, it's all these pieces of equipment that are now just able to be bought by us.

The interface between your processor and our processor is that interface sufficiently well designed at this point? No, I mean, I think that every step of the way, we're trying to make our systems faster. GPUs and CPUs will have to get closer and closer to those actuators because, at the end of the day, everything is just governed by the speed of light, and you want your computation to go really, really fast. Any physical distance between GPUs or CPUs that are understanding the errors that are occurring in the system is just going to slow down the quantum computation. And just rounding out the neutral atoms, and then we'll let the other guys talk about scaling their modalities. But like Ben said, the number of physical qubits for neutral atoms isn't really the bottleneck because we can put millions of qubits into this little device here. It's really just our ability to control those precisely with lasers and then, basically, error-correct those codes, error-correct those qubits. And so, it's interesting in that when you think of scaling, it's not just about putting more physical hardware in there. It's really about more precisely controlling these god-given, nature's qubits with lasers.

What's the latency that we have to achieve? It's probably associated with the coherence time of these qubits and the time it takes for us to do error correction or whatever control algorithm and send some signal back. How much time do we have in that end-to-end round trip loop?

Well, the good news is that it's really about the ratio of how much you can get done to the actual coherence time of the underlying modality. Neutral atoms have quite a long coherence time relative to other modalities, but we're talking microseconds.

[Music]

Yeah, so we actually think that, at least with our technology, which is applicable across all quantum computing modalities in whole or in part, when we focus on superconducting, we think that you need to be able to have less than one microsecond latency in order to do error correction before the next error can show up. So, our goal in connecting to GPUs and to your GPUs is to get to the 500 to 800 nanosecond range so that we can actually take advantage of using a GPU for global error correction. We think we can parse it by doing some on-chip, some with a pre-decoder operating maybe at 100 millikelvin or a Kelvin, and then do the rest with a GPU doing global error correction. But again, that latency question is really, really critical to being able to do that on a timely basis before the next error shows up.

Yeah, milliseconds is easy, microseconds is challenging, nanoseconds is hard. So, we're already in the hard space.

Yeah, exactly. That's okay. We like hard. So, the goal again is, you know, 500 to 800 nanoseconds is really our goal. It's not just about latency, by the way. It's also about throughput because if you have a large error-corrected computer, you have a pretty big stream of results coming back, telling you where the qubits went wrong, how you have to adapt your algorithm to steer it back to the correct answer. You know, that's a multi-scale problem. So, you probably have to do some amount of classical compute with very tight latency, and then there'll be more complicated computations that you have a little bit more time for because you're doing multiple rounds of error correction. This is a very interesting thing to explore, and I'm sure that very hard error correction is a very hard problem. Throughput doesn't scare me, but latency and throughput requirements together scare me. It's exciting, though. An essential thing for any of these modalities to really work and to scale is going to be to build the special-purpose classical computer that is the control system and drives it, doing all this magic of error correction. This control loop, error correction, control loop right here is a really exciting computer science problem, and I hope all of us together can make some real breakthroughs in the next several years. Throughput independent of latency is not an extraordinarily hard problem. Latency independent of throughput is not an extraordinarily hard problem. The two of them combined is an extremely hard problem. In fact, I was talking about it in my keynote. Large-scale inference of AI is a problem kind of like that, especially when you want to interact with the AI.

Pete, go ahead.

Yeah, Jensen, so I think more than one of us on this stage has gotten ourselves in trouble by making comments on timelines, of course, and found ourselves in difficult positions and so on. But I really appreciated the other piece of your quote that went missing, which was that we have to scale up by 100,000x, five orders of magnitude. And I think in our field, that is really hard. In our field, unfortunately, there is a lot of wishful thinking from people desperately hoping that we can make the machine useful before it's useful, but you have to scale up by about 100,000x. It's natural to say that that's a multi-decade exercise for human beings. It feels that way in the dead of night. But you, of course, have extraordinary demonstrations that it doesn't have to be a multi-decade exercise. XAI Colossus built a 100,000 GPU cluster famously in 112 days. How did they do that? What miracle did they deploy to get that done? The answer is the trillion dollars in 50 years that has gone into the semiconductor industry. I came back a couple of days ago from GlobalFoundries, and I've been raging about the insane leverage that we can access there for eight years without ever having actually stepped inside the fab. And I was extremely lucky to get in there a couple of days ago, and it's a religious experience because you see just how insane the capability of the fabs, the contract manufacturers, the OSATs, and so on, is. That has always been our thesis, and to your question on when, there are basic questions

That you should ask, and you know, this is a very complex field. Everyone is arguing for their particular technology, and I'm no different. But there's a whole separate set of questions, in addition to fidelity and so on, which is, can you tell me who your fab is? Can you tell me who your OSAT is? Can you tell me who's doing contract manufacturing, and where is the site where you're building the machine? These are necessary but not sufficient conditions to actually being ready to build a genuinely useful machine.

It's very hard, but that is what we've really spent our resources towards. And I think it's very exciting that you now see other players in the field taking steps into that same kind of regime.

Yeah, in fact, Pete, as you were talking, all of the things you're mentioning are very sympathetic with us. Recently, as you know, we introduced this idea called silicon photonics co-packaged optics, and the semiconductor physics had to be invented, the packaging technology, how you stack it, how you manufacture the entire supply chain. I introduced a whole bunch of new supply chain manufacturers. We were able to leverage many existing industries. If not for that, and we had to invent everything from the ground up, it would have been impossible for us to do that. In a lot of ways, you guys are doing what we did with silicon photonics CPO, a multi-scale challenge; the sciences have been being invented in your case. So, this is a really hard problem.

However, exactly as you guys are saying, we have to find a way to carve a route, carve a road for you to be successful as soon as possible. Almost every day, as you scale up to this future of quantum computing, if you look at the reason why we're here, we found a way to explain our story in such a way that we didn't set a bar so high that we couldn't reach it. On the other hand, we selected the problems that we could solve in a quite unique way early on.

In a lot of ways, Nvidia democratized parallel computing, if you will, but we did that on the backs of computer games. Some of the 3D graphics we rendered in the beginning were not exactly right, and there were a few missing pixels, and there were some gaps in the Z-buffer, and people accepted it because it was a game, for crying out loud. It gave us the opportunity to scale our economics, scale our technology, scale our footprint, and then we selected the right science, the right industries, field after field after field. I'm hoping that for all of you, we don't set a bar so high that we can't reach it. We shouldn't be held to a standard of computing, which, as you know, is quite high. The robustness of computing, the repeatability of it, the whole industry, everything is a very high bar. We had to go find an industry where the bar is quite low. Because there's no alternative to quantum computing, the bar is very low. You know, a computer just simply can't solve it. Like, for example, you couldn't buy a personal computer in 1995 and play Quake without Nvidia in it. You can't reasonably do that. And so that bar was, in fact, incredibly low. Now, of course, you guys are doing something much harder than that today, but I do hope that together we find a path for you guys to be successful sooner than later.

And Matt, you were saying earlier about your approach, and I love it.

Yeah, well, we honestly took a lot of inspiration from you. Find the markets where you can provide true commercial advantage and address those, get feedback from the market, learn how to fulfill customer expectations, and then find the next one. With neutral atoms, we're very lucky in that they are flexible, where we can and we do have three orders of magnitude improvement over current standards in timekeeping and sensors, which are very large markets in and of themselves. But using that, if we say timekeeping was your gaming market, maybe.

That's right. You're using all that to actually get real feedback from the market and ultimately develop that commercial muscle that we need as an industry to actually sell these things for real use cases that people are going to earn a return on their investment. And I also heard something really, really clever. Of course, I knew about this in the work that you guys are doing, that you would stand on the shoulders of classical computers, extend it, and make it do something extraordinary. We didn't replace the computer; we added to it. And what I would explain to people in the beginning is that we never made anything worse. Just turn me off, you know, but I don't make things worse. Parallel computing, as you know, violated Amdahl's law and in many applications made it worse. But accelerated computing, we did no harm. One of the things I really like about adding these to GPUs and CPUs is that you keep adding something, and you make that computer more and more special, more and more capable. So, yeah, it's interesting that there are people now talking about quantum AI and the idea of tightly integrating a QPU into a GPU and a CPU creates that opportunity. So, let's explore the space.

Yeah, right. Go ahead, Pete.

Yeah, I very much agree with that sentiment. There is a caveat that I think is important to add, which is that we love to talk about the idea that conventional computers are integrated with quantum computers, and we absolutely believe in that. There will need to be a large conventional supercomputer, GPU cluster, or whatever, preparing Hamiltonians, preparing input to the machine. But when we talk about this, there's a very seductive idea that the whole will be greater than the sum of the parts. That idea leads to an interpretation that we don't need a good quantum computer; we can take a not very good quantum computer, plug it into an incredibly high-performance conventional computer, and the whole will somehow be greater than the sum of the parts. I think you have to be quite careful about that. In some cases, you can see a rationale for that, but in some cases, there is no reason to believe that taking a small, low-performance quantum computer and plugging it into an incredibly high-performance conventional computer is going to make things any better. It is going to make things worse.

That's exactly right, and in fact, that was one of the challenges of accelerated computing. We sat next to a processor that was getting better by a factor of two every single year, and the R&D budget of that industry, the CPU industry at the time, completely dwarfed the GPU. It was a million times larger R&D budget per year. So, what are the odds that adding a GPU, built out of such a low R&D budget, could add value to a system that sustains such an enormous R&D budget? The answer for me was to keep narrowing and narrowing the application space. I kept lowering the bar for myself so that the problems we solve are so specific. Now, the challenge, of course, and this is going to be a challenge for your industry as well, is that if you end up finding an application space and it's very specific, then the size of the market is not so large that it can sustain your growth. You have to ultimately find that flywheel in anything that's in computing, which is what we do for a living. In anything that has a very high computation requirement, you need a flywheel to get you there. That flywheel starts with solving a problem better than anybody, eventually getting to high volume, which generates more R&D budget, which allows you to build something better, which allows you to get more high volume. That flywheel is insanely hard to get going.

You guys understand, and so, but I have every confidence that this problem-solving approach you guys are trying to solve—there are problems that are simply impossible to scale for classical computing. To the extent that we can find a way to lower people's expectations of us and narrow our own aperture of problems we want to ambitiously go solve in the beginning, so that we can catch that flywheel ourselves, I am absolutely certain this is going to happen. I have every hope and expectation that this team is going to do it, and I really love hanging out with you guys.

If we very quickly, what do you think we'll end up talking about next year because I want everybody to come back. This is such a great show, and remember, this is our first time, so if we're a little clumsy, lower your expectations, but next year it's going to be incredible.

Yeah, let's go around quickly. What do you guys want to talk about next year?

I think there will be some amazing demonstrations with quantum error correction in the next year. I think the industry is expanding so quickly that we'll be trying to tamp down expectations, like you said. I think there's going to be such amazing velocity occurring with quantum error correction. I think we'll hear a lot of great progress on continued increases in logical qubits. And then, from our perspective, with our commercialization strategy, we did about $30 million in revenue last year. So, we really hope to be telling you about a heck of a lot more, selling some of these early use cases.

Yeah, man, that's real money. Go ahead.

Yeah, first off, I like the premise of your question that there is going to be a quantum day next year. So, absolutely. From our perspective, it's the integration of all core functionality of a quantum computer on a chip, full stack.

Go ahead. To you, Pete.

For us, it would be error correction, better architecture, but most importantly for the audience, it's novel algorithms. There is so much room for innovation in quantum algorithms. We're seeing completely new subroutines emerging every couple of years. Think about it as someone inventing the fast Fourier transform all over again. It's paradigm-shifting for whole industries each time. So, when we're looking for those niche areas you've been mentioning, those core algorithms, those routines can completely change the reach and the impact of quantum computers and get this flywheel going. In fact, there's a misunderstanding that quantum computers are going to take classical algorithms and just make them go faster. The whole point is to invent new algorithms that are ideal for this new form of computing.

Go ahead.

Yeah, I'm going to talk about error correction as well because, again, that's the key thing. We've really entered an exciting phase where you can build machines on which you can run error correction routines. For a couple of decades, Shor discovered error correction the same year he discovered his factoring algorithm, so we know it's possible in principle, we know what the math is, but now this is becoming a practical discipline. We can build and test things, and now you can say, "Oh, wait, this is the flaw my hardware has. So, here's a code that's much more optimal." And we can try and adapt our hardware to go in that direction. So, I'm super excited about that whole area. Just from the downloads, the accelerating downloads of Qiskit, the number of people who are using it to discover new quantum circuits to simulate and discover new algorithms is growing. So, I'm super excited about that.

And so, Pete, what are we going to talk about next year?

Yeah, so now we're making thousands of wafers of quantum chips at GlobalFoundries at a pretty high level of maturity. We're building large cryostats with no chandelier. We're stringing together heaps of optical fiber. And I really hope that you'll have us back next year. Next year, I think I'll have to wipe the mud off of my boots before I come up on stage here because, as I said, we're breaking ground on these two very large sites—half a million square foot kind of sites in Australia and Chicago.

Pete, you look like a builder from Australia.

Well, I'm trying my best, but thanks very much for having us.

All right, guys. Thank you. Thank you. Next. Just to introduce the next crew, thank you guys. Really enjoyed it. Thank you guys very much. Really appreciate it. So, this is our last panel. It sounds like next year we're going to have demos. What do you guys think? Next year, we're going to do demos. Okay, all right. Hey, nice to see you. Nice to see you. Hi, Krista, nice to see you. Welcome.

And so, we had all these scientists here, and the thing about quantum computing is that most of the CEOs I meet and talk to can understand the basic technology because we're in the computer industry, and of course, there's always computer science. But there's not a lot of basic science, and basic science is hard, as you know. Quantum basic science is quite hard. Most of the time, you're talking to CEOs who are maybe refactoring the way a computer is going to be architected or designed, but the basic technology is understandable and maybe it's applied in a different way. In quantum computing, the science is new, the science and engineering are new, the manufacturing is new, the software programming model is new, and the way you think about algorithms is new. And, of course, one of the areas that I, you might have noticed, I want the industries to succeed, and so I can't help but try to advise it—not that I'm giving good advice necessarily, but to narrow its focus on applications so that it's not held accountable to the expectations of other forms of computing.

Now, here, the three of us work in large companies, but yet you have quantum computing initiatives in your company. How do you guys think about quantum computing in the context of your overall computing and industrial computing business? What do you hope to achieve, and what are some of the challenges that you see in ultimately making quantum computing successful?

Yeah, so I'll start. Thank you, Jensen. Great to see you. It's wonderful to be here and to be with all of you as well. When we think about quantum computing at Microsoft, obviously, we have a large cloud, and we have many customers, whether they're enterprises or scientists, practitioners. It's really about empowering our customers with the most powerful quantum computing at every moment in time, today and tomorrow. We are a platform company looking to bring forward a quantum computing platform to enable new scenarios, new applications, emerging capabilities, and disruptive capabilities.

What technology did Microsoft choose, and why?

So, we have a couple of approaches. We both partner with other quantum processing unit providers, quantum hardware providers, and we also have a long-term investment in an approach called topological qubits, topological quantum computing. In fact, we've just had a breakthrough announcement in the last month, and this week at the APS March meeting for physics, we shared more data on our approach around the topological qubit. The idea here is that you encode the information non-locally, meaning the qubit isn't just a single point. The information of the qubit is spread across a device design where it promises to protect the qubit more, but it also gives us a very nice control profile. So, we can use digital control instead of analog control, and it can simplify the amount of control requirements in the quantum computer itself. This is called the Myana 1 chip that we shared last month.

Okay, we're going to come back to you in just a second about applications and how you think about computing platforms and that kind of stuff. Mon, if you could help us understand what approach AWS selected, why you guys did that, and how you see quantum computing in the overall context of your computing strategy.

Sure, thanks for having me. It's a great opportunity, and this is my first GTC; it's been a wonderful week. We built quantum computers based on superconducting technology. So, we give strong emphasis to error correction. We believe that error correction is really going to be important for quantum computers to deliver their long-term promise. We announced recently a superconducting chip called Ocelot. You know, Ocelot is a kind of wild cat. The name is between Schrödinger's cat and oscillator. So, scientists came up with this name.

That's clever. That's almost as clever as Nvidia.

[Laughter]

So, Ocelot demonstrates error correction in a scalable architecture. And you've heard this term "scalable" a few times today. Scalability is an important term. Now, why superconducting devices? My mental model hinges on three terms: knowledge, speed, and experience. Knowledge: there's a lot of experiments out there done with superconducting devices in industry and academia, proof of concept for error correction, so good basic knowledge and background. Second is speed: they're fast. They use microwaves, so it's easier to implement error correction. Finally, experience: AWS has a good amount of experience with custom silicon and semiconductors, and some of this experience can be translated to superconducting technology, at least from the operational point of view. Of course, we are open-minded about all other ways of building quantum computers and are very excited about the progress happening across the board.

Yeah, and Krysta, back to you.

The technology is still developing, making great strides. Frankly, if you just timed the milestones of the industry, the rate of milestone achievements is accelerating, and that's really terrific to see. But you still want to find early applications or ways to think about quantum computers as you're developing them to make them useful, to find use for them as you go. How are you doing that?

Yeah, so I mentioned our work on our Myana, our topological qubits, where we're really focused on utility scale. It has the right size, speed, and reliability and controllability to reach millions just in the palm of your hand. We're also working on types of qubits today that we are building to have 50 logical qubits this calendar year. For example, with Atom Computing, Ben Bloom was on the stage earlier in this session. We are working together to co-design the architecture so that we can enable the most and best logical qubits. When you look at upwards of a thousand physical qubits, a thousand neutral atoms in this platform, we can work on how we arrange them, how we move them to best enable, say, 50 logical qubits that have better performance characteristics than the underlying physical qubits themselves. Now, with those 50 logical qubits, we can look towards showing early applications. Around 50 qubits is where you can start to outperform classical computing, and then around a hundred logical qubits, you can start to outperform applications in the space of science, where you might be looking at different materials models, quantum magnets, and so on. Our intention in the next few years is to work with 50 and then a hundred logical qubits, and then build it up with several hundred where we're really pushing the limits on the applications.

At the 50 qubit area, what application space are you mostly focused on, and how are you selecting your applications?

Yeah, so definitely, the most promising set of applications is in chemistry and material science, biochemistry as well. As we look at the rise of AI, the AI capabilities that have emerged are just immense and tremendous. We don't want to replace AI with a quantum computer. I really view a quantum computer as an accelerator, something to accelerate the other compute we already have. We need to integrate it with AI and high-performance compute in the cloud, not replace it. It's really about bringing those together. In the next few years, I think it's all about using the quantum computer to produce highly accurate data. I used to think of a quantum computer as a standalone solution provider—run a problem instance, get a solution out. But that's not the right way of thinking about it. The other way to think about it is, I'm getting classical bits out. What are classical bits? Data. What do I do with classical bits that are highly accurate for what they represent? I use it for training data. We know that we can use lots of training data to train an AI model, and we can augment that model with small amounts of high-quality data. We've seen this in different spaces, like machine translation and other tasks, where small amounts of high-quality data can make a huge difference in the task you want to use that model for. And so, in the space of chemistry and material science, we're focusing on using the quantum computer to produce highly accurate data that can be used to train AI models and improve their performance. I think this is an incredibly promising direction for quantum computers with 50, 100, 150 logical qubits, where those logical qubits have to be better than the physical qubits they're built from.

And Krysta, just to be sure, you guys all got it, but one of the areas that's super exciting is in the area of materials and biology, where we would like to train a model. We would like to train a representation of biology, but where does that training data come from? It's not like we have sensors and instruments collecting data about biology, cells, and proteins. Now, we could simulate that using our quantum computer and use that as ground truth data to then go train an AI model. And once we have an AI model, it's a lot more malleable. It's a lot easier to apply. We could use it to do all kinds of experiments with.

That's exactly right. The idea is to get a faster, more predictive, more accurate AI model. This is a classical AI model that deploys in your current infrastructure, and it's fast. So, that's really the promise—you're using the quantum computer to produce data that you cannot otherwise efficiently get on this planet.

That's powerful. Perfect. Simone, go ahead.

Same. I like this perspective. I attended your keynote, and you showed us a slide with four phases: perception, generation, agency, and physical AI. Of course, physical AI is about robotics, autonomous vehicles, but

It's about our physical world, a larger container for that. As Krysta said, what is the role of quantum computers in that phase, physical AI?

Quantum computers are the only instrument we have today that we know so far for accessing that layer of physical reality, which is quantum physics. That layer of physical reality is governed by certain laws that do not apply to the physics around us that we experience with our senses. So, quantum computers are going to be catalysts for scientific innovation that will allow us to discover certain things that are very hard to predict. In a way, we must build quantum computers because otherwise, that layer of reality will not be accessible to us.

Quantum computers will work in partnership with machine learning and AI. In my opinion, in the fullness of time, we will do science together with machines. We will ask questions to machines, machines will ask questions to us. There will be formal verification, formal reasoning, different types of compute, and quantum computers fit very well into this picture. Indeed, we have not been able to compute like nature computes in many cases. Nature is incredibly efficient, and I think of quantum computers as enabling us to take a step closer to computing like nature does—being able to see and understand electrons in a new way. We can't do that in all cases today efficiently. So, it really takes us a step closer, and combining that with AI will, I think, be revolutionary. And this paradigm of using the quantum computer to get the ground truth to train a classical computer's AI model, which is much easier to use than the entire software stack and all the applications, and quite frankly, very cost-effective. The training data that's going to come from that quantum computer will not be easy to get. Obviously, it required the endeavor of humanity to get there, but now that you have that simulator, you could, within 50 logical qubits, be able to solve what would otherwise be intractable equations. Now, you can go train a model that we can easily apply. We've now taken this incredible valuable asset and extracted it, making it simple to use for all of the computing world.

Indeed. You told us about AI factories. Maybe we'll have quantum AI factories.

That's right. That's really exciting. So, we're currently at how many logical qubits?

Yeah, so interestingly, one year ago, had I been on this stage a year ago, I would have said we had zero logical qubits that were better than the physical qubits they were built from. And within the last year, we went from, in April last year, showing with Continuum four logical qubits. Then five months later, we tripled that number to 12 logical qubits. And then two months later, we over-doubled that and showed 28 logical qubits with Atom Computing. So, now that shows that progression. It goes back to the acceleration you mentioned, Jensen, that we're seeing in the field. Now, this calendar year, we're working on the 50 logical qubits with Atom Computing, and the next generation of that system will be 100 logical qubits on a 10,000 physical qubit machine.

Yeah, that's really incredible. That's really incredible. So, Simone, one of the great capabilities of neutral atoms and approaches is the scalability capability of it. All the things you said about superconducting approaches are absolutely true. How do you see these approaches at some point emerging, or do you see a grand solution emerging from the industry where everybody says, "Yeah, that's it"?

Back in the old days, in my generation, we had this thing called TTL to ECL, and we eventually all settled on CMOS. During my generation, there was a lot of arguments and debates about ECL versus CMOS. The reason for that was because ECL is essentially static current; it's always high but it's constant. In the case of CMOS, it gets higher and higher, and there's dynamic power instead of constant current power. So, there was a lot of debate about which one was going to be better long-term, but ultimately, CMOS won because it was scalable, and you could solve all the other annoying issues with just more transistors. Is your industry also discovering something similar to that, where maybe all of these challenging, annoying issues ultimately go away because of scalability and just having a lot of logical qubits?

The short answer to your question is that I don't know. The longer answer is that there is some transition happening in the industry. Historically, people ask, "How many qubits?" So, everybody is obsessed with how many qubits they have. You're at the supermarket in the morning in front of tomatoes, and a random guy comes and asks, "How many qubits do you have?" [Laughter] People ask this question less now, which is great because people are starting to understand that it's not just about the number of qubits. It's about the quality of the qubits, the error rates, and the coherence times. It's driving me crazy too, but it doesn't matter. So, everybody has this obsession, but it's shifting. There is an interesting signal that is emerging at the moment, which is error correction. In the last 12 months, a number of different experiments have been done with different modalities—ions, neutral atoms, superconducting devices, and more exotic devices. This is probably a theme that is going to become stronger and stronger and will likely determine which of these modalities is going to emerge in the fullness of time.

Yeah, well, go ahead. Maybe I'll just say that you mentioned flops. Here, it's not just more is better. We need better, also. As we talk about logical qubits, not all qubits are created equal. We need qubits that are able to extend how much computation we are doing. We are not building a storage device; our intent is to do computation we cannot do with all the other compute we have on the planet. So, that means we need to improve the qubits. Physical qubits fail once every thousand operations, and a thousand operations in your computation is not enough. We need to do computation that ultimately has more like a quadrillion operations. So, there's a large gap to close there. We use error correction to close that gap, but when you use error correction, you need more physical qubits as well. You're using hardware and software to close the gap. The goal is, as we look towards 50, 100,000 logical qubits, it's not just the number of qubits that changes. When we say 50, 100,000, it's also the error rate. It's how good those qubits are, those logical qubits, that also changes and needs to change by orders of magnitude at each increase. When we talk about a thousand logical qubits, we want those to be as good as one in a billion, only one fault in a billion operations, or much better. At 100, you really want only one fault in a million, 10 to the minus 6, is what we say. So, we really have to increase that, and that makes it a little more challenging.

Yeah, it's really true. One of the things I'm really excited about today, and I was quite excited to do this event today, but I was also quite concerned about how we would take a conversation that is deeply scientific and very technical and ultimately connected to developers, which is what GTC is all about—people who are trying to do hard science, people who are trying to create impossible applications—and to start the journey of connecting the dots between the science and usefulness for them. I thought we did a fantastic job today. There were several things that I think were properly and successfully conveyed. The idea of a quantum computer is not to build a computer that replaces computing but is a QPU that is added to a GPU to a CPU to extend classical computing to do things that otherwise cannot be done. There are some domains, some useful domains that we can imagine, for example, in biology and chemistry applications and material sciences applications, where we could use a quantum computer to make a classical computer way better to solve problems that otherwise cannot be solved. For example, to create the ground truth for biology, to create the ground truth for atomic physics, and that enables us to use AI as we understand it today, quite reasonably well, and improving itself at a million times every couple of years, to now amplify the capability of AI to be able to use AI to solve drug discovery, material sciences, and biology applications. If quantum computers singularly addressed that, that's all it did, but that's the reason why we need to build these machines.

Thinking about the timeline, certain questions that you are familiar with: how long is it going to take? Well, first of all, this is like the space program. The goal is to go to the moon, but we're going to discover a lot of things on the way. We discover firemen suits trying to go to the moon. This is going to happen with quantum computers as well. It's a grand adventure, but we need to get there. It's not a binary thing; it's not "oh, we have a quantum computer, we don't have a quantum computer." It's a journey, and lots of things are going to be discovered in science and technology as we get there. In terms of timeline, I mean, I myself grew up in Tuscany, in a village in Italy. There's a city called Pisa, and there's a tower which is not straight; it's leaning. Well, that's by the way a result of quantum computing because a classical computer would make it be perfectly straight. [Laughter] It took them 200 years to build this tower. Now, if they did a good job or not is debatable, but 200 years, right? So, only a quantum computer can make a structure like that stay up for so long. The point I want to make is that it takes time. Quantum computers are going to be so impactful that it's going to be a great party in the end. I think it's fantastic that Microsoft Azure and AWS are integrating quantum computing into their platforms so that developers around the world can gain access sooner rather than later. If everybody had to go build their own quantum computer to be able to enjoy this and the instruments you build, it would take so much longer. Now, you've put it in the hands of browsers, and anybody could use it. I'm looking forward to 50 logical qubits very, very soon and 100 shortly after that. I think the progress of the industry is incredible. We had a great panel today. I know they're all friends of ours and friends of the industry, and in a lot of ways, although there are many competitive approaches, the industry is working so closely together to want everybody to succeed together. So, I think it's really great to see. And if I had to be wrong to show everybody in the world that quantum computing is worthwhile to do and that the industry is built of amazing people and the work that the industry does is going to make a great impact, and if I had to make a mistake in order for us to demonstrate to the world, mission accomplished. Thank you all for coming.