Pod Insights Eye On A.I. Oliver Dial of IBM: Quantum Advantage Is Happening This Year About this episode IBM's VP of Quantum Systems, Oliver Dial, has spent his career building quantum computers from the ground up, and he's unusually direct about what they can and can't do. In this conversation with Craig Smith, Oliver Dial walks through where the field actually stands in 2026: quantum utility was achieved in 2023, quantum advantage is the target for this year, and a fully error-corrected machine capable of tackling the hard problems is on IBM's roadmap for 2029. That last milestone, Dial says, now feels both achievable and terrifying.
The episode is worth your time because Dial doesn't hype. He explains why IBM built a 1,000-qubit computer and then took it apart almost immediately, why Google's quantum advantage claims remain scientifically contested, and how a new error-correcting code IBM developed just reduced the qubit overhead required for fault-tolerant quantum computing by an order of magnitude. For anyone trying to understand what quantum computing will actually mean for their industry, and when, this is the clearest map of the road ahead available right now.
If this conversation changed how you think about the future of computing, subscribe to Eye on A. I. for weekly conversations with the researchers and builders shaping what comes next.
Transcript Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.
0:00 Speaker 00 Is there are some problems you can potentially solve on them that you could never ever ever solve on a classic computer? Remember, one end of this wire is up at room temperature where you and I live, and the other end of this wire is at 0.02 degrees above absolute zero. This year, 2026, we're hoping to demonstrate what we call quantum advantage. And I decided this technology was just so cool because...
0:25 Craig S. Smith? Okay, so Oliver, it's great to see you. I met you, as I said before, once you were fiddling around with a quantum computer back in one of the rooms there. And I wanted to talk to you about where IBM is today with quantum. There's a lot of chatter, and it's very hard for a non-expert, which I certainly am, to understand what's happening. At that time, there was a fairly aggressive timeline, and I was with Jay Gambetta, your colleague, and he was pointing out that you guys had met your milestones so far on that timeline.
1:08 Craig S. Smith? I think, if I'm not mistaken, this is the year that you're to get to quantum utility? Is that the term you use?
1:17 Speaker 00 We use the term quantum advantage, actually.
1:19 Craig S. Smith? Quantum advantage, okay. So can you, I guess, first of all, start by introducing yourself? Introduce yourself to listeners, and then a little bit of your background, how you got to IBM Quantum, and then we'll go from there.
1:38 Speaker 00 Yeah, absolutely. So my name is Oliver Dial. I'm a physicist, so I study what's called condensed matter physics. I usually tell people it's physics of rocks, but it's really the physics of sort of how quantum mechanics acts when you get it in really weird circumstances. And I got interested in quantum computing kind of indirectly. I was studying something called quantum dots, which are little boxes you can put only one electron in. And it turns out that's one way people try to build quantum computers. So I sort of transitioned from studying quantum dots as physics object to studying quantum dots as qubits. And I decided this technology was just so cool because it kind of brings together computation, which I'm really interested in, electrical engineering, physics, all into
2:23 Speaker 00 this kind of one package that has the potential to really change the world. And so, once I kind of decided that was an interesting thing, IBM was definitely the place I wanted to go to to do it because IBM has made a really big bet on quantum computing, that it is in our mind, part of the future of computing. And I decided that if I wanted to get more involved in that, that it was just really the place to go. At my heart, I'm a hardware guy. I'm happiest when I'm in the lab, turning a wrench, like you found me, or programming, trying to get one of these machines to do something unusual. But these days, I'm acting as VP of Quantum Systems, so kind of responsible for leading the team that takes all the pieces of these quantum computers, puts them together, and gets it to work as one big piece thing.
3:09 Speaker 00 And that's a complicated job, because these quantum computers have a little tiny quantum core. And then, as you work out from that, there's all this classical infrastructure. There's refrigeration, there's control electronics, there's wiring, there's connectors, and all those things have to work just perfectly for these processors to work as well as we want them to.
3:27 Craig S. Smith? Yeah. And IBM is a systems company at this point, and for quantum as well as classical hardware. And you have a roadmap. Where are you on that roadmap right now?
3:48 Speaker 00 So, yeah, the roadmap you're talking about is something that we publish and update every year, and it shows what we think the future is going to look like reasonably, and what we did in the past. And kind of one of our big statements is we've generally stuck to it. And the reason why that's important is there's a lot of hype in this field. There's a lot of people saying they can do crazy things. There's a lot of people making just very optimistic projections. And so, what we want to be able to say is, look, we've been doing this for years, we've been making these projections for years. And, you know, we changed a little things, but we're basically on the track that we've been saying that we're going to be on. And so, you can trust us in these predictions about the future. A couple of years ago, we hit a point on that roadmap that we call quantum utility, which is what we started this conversation on.
4:32 Speaker 00 And it's a really neat point. It's the place where the quantum computer can do something complex enough that it is challenging for a classical computer to simulate the quantum computer. Right? And so if you, you know, the background to why that's important is quantum computers don't work very well. They're expensive, they have high error rates, they're not very big, they're hard to program. And the reason why they're interesting is there are some problems you can potentially solve on them that you could never, ever, ever solve on a classical computer. And so it's worth all of this rigmarole to get that point of I can solve something I couldn't do otherwise. That's really the allure. And so being able to get to the point where you can't simulate the quantum computer on
5:15 Speaker 00 a classical computer is sort of table stakes. Until you hit that point, anything that you did on your quantum computer, you really should have just done on a classical computer instead. And hitting that point requires kind of two things. One is you have to have enough qubits. That's the unit of computation for a quantum computer. And it's important because the amount of memory it takes to directly simulate a quantum state is exponential of the number of qubits it has. And so if you don't have enough qubits that it couldn't fit in the memory of a classical computer, that opens a set of pretty efficient simulation algorithms. The other important thing is how many operations that you can run on the quantum computer. I said it has a high error rate. And so if you run too many, the machine will make too many mistakes for us to correct.
6:00 Speaker 00 And the reason why that's important is there are also simulation techniques that you can use for circuits that are short, that are efficient. So 2023, we hit the barrier of the point where the quantum computer becomes challenging or impossible to simulate on a classical computer. This year, 2026, we're hoping to demonstrate, or we're hoping for someone to demonstrate, ideally not us, what we call quantum advantage, which is the point where you're using a quantum computer to do something better, faster, cheaper than would have been possible on a classical computer. And it's a little bit different from being able to run something that you couldn't simulate because now you actually have to get an answer out of it. It's not just enough to, you know, show some speckle pattern or something like that. And really, quantum advantage can be phrased two ways.
6:44 Speaker 00 One of them is a very strict scientific question of can I prove that the quantum computer gave the right answer? Can I prove that I got it faster than I possibly could have on a classical computer? You know, can I make a rigorous proof of this claim? And that's considered sort of an important scientific issue because that settles the question of whether there's some computational regime that we can practically achieve that was unachievable on a classical computer. There's also a very nitty gritty, call it heuristic advantage. I ran this on a quantum computer and I got a better answer than I could have gotten on a classical computer and I'm happy with that. The reason why we think these things are going to happen this year, well, first of
7:29 Speaker 00 all, people have already claimed that nitty gritty heuristic advantage over the last year or so. But the main thing is that we think that the computers have advanced to the point, the algorithms have advanced to the point, and the tricks for correcting the noise have advanced to the point where this is going to be possible for a wide variety of users on our systems today. And so, you know, we're really hoping for this collaboration where we bring the hardware, really smart scientists bring the problems, and then the flip side of the coin, really smart scientists also try to solve these problems on classical computers. Because of course, what we're going to find is when somebody does something for our quantum computer, a whole bunch of computer scientists are going to go to work with their AI coding agents or whatever and try to replicate the event on a classical computer. And so we need this conflict, we need this back and forth between the quantum computers
8:14 Speaker 00 and the classical computers to come to a bit of a close before we can really claim that.
8:18 Craig S. Smith? Yeah, and quantum advantage is one of the confusing things because...
8:24 Craig S. Smith? Google claimed the quantum advantage, I think, two or three years ago. And, you know, different companies, I think, mean different things by quantum advantage. So, from IBM's point of view, and also you talked about the number of qubits. Explain why the number of qubits is important and how many qubits are you working with at this point? And how do you create a qubit in your system? Because there are different kinds.
9:00 Speaker 00 Great questions. I'm sure you've heard the phrase, there are lies, there are lies, and then there are benchmarks.
9:05 Craig S. Smith? Yeah.
9:06 Speaker 00 It's a little bit like that.
9:07 Craig S. Smith? Yeah.
9:08 Speaker 00 That one of the big questions on quantum advantage is verifiability.
9:12 Craig S. Smith? Yeah.
9:13 Speaker 00 Can you show that the quantum computer actually produced the correct output? Because it doesn't really do you any good to say you solved a problem if you can't say for sure that that happened. And so, a lot of the back and forth around Google's claims on that are around verifiability. That they solved a random circuit sampling problem, where there's supposed to be some distribution of outputs that's hard to compute classically. And they were able to show that sort of, for smaller circuits, the distribution had some measurable fidelity. Sometimes it gave the correct answer, the number was kind of one time in a million or something. And then they ran a little bit bigger problems and said, by extrapolation, we should still be getting the correct answer some fraction of the time. And honestly, it's probably true.
9:56 Speaker 00 But if you want, again, if you approach this from a scientific question, it gets a little bit dicey making that claim, because you do have that extrapolation step in there. There are some replicas of that that have kind of bigger differences between the quantum and the classical runtime at smaller circuit sizes where you really can't simulate it. So maybe that claim has gotten stronger. But again, it's, you know, fuzzy because of that verifiability aspect. What we've tried to do is help set up something called the Quantum Advantage Tracker. Which is like hugging face for quantum in some sense. It's a place where people can post specific problems, performance on a quantum computer,
10:39 Speaker 00 and performance on a classical computer, and really let this debate happen out in the open. And it's actually started to get quite a few events put on there.
10:48 Craig S. Smith? Is this a leaderboard that you guys publish?
10:51 Speaker 00 It's a leaderboard. We initially set it up, but you just you make a issue on GitHub to add something to it. It's not like we're not like we're editing what goes into it that it comes from the community.
11:07 Craig S. Smith? Yeah. And then on the number of qubits and the kind of qubits and...
11:13 Speaker 00 Oh, yeah, so many questions there. So quantum computers are a little bit different from classical computers, or at least they are today, that you have memory and you have a CPU and you have storage, and these are all very distinct things. Think of a quantum computer as an accelerator. It sits off to the side of the classical computer. It's controlled by the classical computer, and so we send in a classical description of a problem. We manipulate the quantum state inside of the quantum computer, and then we measure the qubits, which is a way of converting it back into classical data that we can pull out to our classical computer again. And that's really important that we have to begin and end with classical data because that's the only thing we have access to as human beings. So the number of qubits describes how complicated of a quantum state you can manipulate in the center there.
11:57 Speaker 00 And basically, you can think of it as a computer for doing linear algebra, if that makes any sense to you, matrix multiplications, in an enormously large state space, as long as you're kind of manipulating inside of this quantum this this quantum device. And specifically, it's a vector space that has 2 to the number of qubits dimensions in it. So complex vector space. So that's why the number of qubits is really important, right? If that vector space is too small, you can't represent the problems you're interested in. You could, like I said, just do it on a classical computer anyway. If it's small, like, you know, four qubits, I could do on my phone quite trivially.
12:38 Speaker 00 So to get this to work, we need to have a way of storing information that the universe can't see. Because if you know anything about quantum mechanics, you probably heard something about collapse. That if you have something in a superposition, and in the case of our qubits, that might be a superposition of 0 and 1. If you look at it, it gets to decide, I'm either 0 or I'm 1, and I'm back to classical information again. And so I have to have a way of doing something that looks a little bit like digital computing, because I have the 0, 1 state, but that's very, very well-isolated. And there are a few dozen ways people do it. One way I mentioned at the beginning with my background, that you can take a single electron, and you can store 1 and 0 as two directions of what's called the electron spin. Basically, electrons have a little magnetic moment, just like a bar magnet. They can point up or down.
13:23 Speaker 00 And because that's a really weak magnetic field, it's hard to see which direction the electron's pointing in, and if you're careful enough about it, you can actually make it impossible to see for a little while. And so you can use that to store quantum information. At IBM, we do something a little bit different. We use what are called superconducting qubits. And in superconducting qubits, what we do is use metal, superconductors, to make little resonators, L C oscillators. Think of it like a pendulum, but the electromagnetic version. And we can use those to store 0 microwave photons or 1 microwave photons, and those are our two states. And so, keeping that hidden from the universe is a matter of making sure the photon can't leak out of the box that we put it in, this little resonator, and making sure that it
14:08 Speaker 00 doesn't get absorbed by anything. The reason why we do that is, superconducting qubits, it turns out, you can make using the same tooling and processing and approaches that you use to make classical digital logic. We use the same nanofabrication tools, we use the same clean rooms, we use all the same techniques for packaging. It's just at the end, we make a device that we cool down to near absolute 0, as opposed to a device that you put into your laptop over there. So it's a really good match for the types of skills that we have here at IBM. And they're also sort of the leading modality as far as the number of qubits that we can build, and leading or near-leading as far as what are called the fidelities, the error rates that we can get. You asked about device sizes.
14:51 Speaker 00 Currently, our flagship devices that we have in our fleet are 156 qubits. To put that in perspective, the Google supremacy experiment you mentioned was done with 53 qubits. Around 100 qubits is where it gets basically impossible to do classical simulation someplace between 50 or 100. So we're significantly past that threshold. We have done experimental devices that are quite a bit larger. A couple of years ago, we made a device called Condor that was 1,000 qubits, just to prove that we could, pretty much. It's not still online. We actually took it apart almost immediately because the air rates in it were not low enough to productively use all 1,000 qubits.
15:32 Speaker 00 And so it was sort of an expensive technology demo, not something that we thought would be useful to our clients. We do have a new device that's coming online this year that's called Nighthawk. We name all of our devices after birds. And that's actually 120 qubits. So it's a little bit smaller than Heron, but it has much higher, what we call, connectivity. If you want to run calculations on these quantum computers, you need to get the qubits to interact with each other. Think of it as if you're doing classical logic, you can do anything that you want if you have, say, and, not, and exclusive or. Actually, flip or is enough. Sorry, I don't usually do classical logic. It's the same thing for quantum logic. There's a gate set that you can use to construct any arbitrary computation.
16:16 Speaker 00 And one thing that you need is a way to entangle two qubits, basically. We do that with literally wires on a chip with what we call tunable couplers. So we can turn on and off the interactions between the qubits. All of our previous devices had at most three connections from each qubit. And so if you wanted to do it, gates between qubits that were far apart on the chip, you have swap, swap, swap, move the data around. Nighthawk has four wires instead of three, which doesn't sound like a lot, but it's a big improvement. And so you can do computations more efficiently on it, even though it's got a slightly smaller number of qubits.
16:48 Craig S. Smith? Yeah. And I don't want to turn this into a course on quantum computer, but your qubits are photons?
16:58 Speaker 00 So.
17:01 Speaker 00 So I'm a physicist. I call a lot of things photons.
17:03 Craig S. Smith? Okay.
17:05 Speaker 00 So our qubits are superconducting resonators, so an LC circuit, if that means anything to you, that basically a little patch on the chip that has a capacitor, two pieces of metal that are close together, and something called a Joseph's injunction in between that acts like an inductor. And so the energy alternates between electric field between those capacitors and phase, basically current inside of the Joseph's injunction, back forth, back forth at a frequency that's set by the design of it. We designed that frequency to be about 5 gigahertz. So it's a microwave frequency. And just like optical light, if you go cold enough and look carefully enough, microwave light has photons. And so we can store either zero or one photon
17:47 Speaker 00 inside of that little teeny tiny electromagnetic resonator.
17:51 Craig S. Smith? Yeah.
17:51 Speaker 00 So the data is a photon, or no photon, but it's stored in this little superconducting box.
17:58 Craig S. Smith? And that superconducting box is one qubit.
18:01 Speaker 00 One qubit.
18:02 Craig S. Smith? And so you have to build an array of now 156 qubits. Why is it, it is it, uh, sounds like an engineering problem, but why is it so difficult to expand the number of qubits? Is it the noise problem, the error correction problem?
18:21 Speaker 00 There are a lot of problems. That's what makes it so much fun.
18:24 Craig S. Smith? Yeah.
18:24 Speaker 00 So one thing to realize is although we're using semiconductor technology, these qubits are actually pretty big. Ours are about three quarters of a millimeter on a side. So if I handed you one of these chips, you could actually see our qubits with your naked eye. I mean, to me, that's amazing. Like, we're talking about entanglement and superposition of things so big you could see them, although I'd have to freeze you to death before you got to see it happen. So some of the things that makes it really challenging is that we have a lot of I/O. If you think about your microprocessor or a bit of memory, it has, you know, a few hundred pins on it. But then it has billions of transistors and millions of bytes of memory. For our devices, for each qubit that we have, we actually need a wire to send microwave signals into it to control it.
19:09 Speaker 00 And then we need another wire for each coupler that it's connected to to control it. And then we need another wire for each group of qubits to measure them. And so our 156 qubit device, I'm going to get the number wrong if I try to do off the top of my head. It's easier for Nighthawk. Our 120 qubit device has on the order of 480 control signals going into and out of it. And those are not digital signals, those are analog microwave signals. And so figuring out how to pack that many signals into a small area in a way that's low cross talk and accurate is extremely challenging. But then we also need 480 signal generators to make those. And each one of those, it looks a little bit like the transmitter in your cell phone.
19:53 Speaker 00 But it needs to be much more precise, much lower noise. So it's a little bit of a specialty object. That's what we call the control electronics. And then we need the wiring to connect these things together, which sounds like, okay, wiring, big deal, Oliver. Except remember, one end of this wire is up at room temperature where you and I live. And the other end of this wire is at 0.02 degrees above absolute zero. So a little bit colder than deep space. And so the thermal gradient across this wire, it's a little bit like trying to air condition your house with all the windows open. If you make that wire wrong, it carries a lot of heat into our system. And so we need to be very careful about how we dissipate that heat down the length of the wire. So those are just a few of the engineering problems that you need to solve as you scale.
20:36 Speaker 00 All of these, you know, they add complexity to the system. But as you said, we're sort of a systems engineering company at our heart that this is the kind of problem that we're used to dealing with.
20:47 Craig S. Smith? So I understand tunneling diodes and when you bring it down close to absolute zero, what you're saying is?
21:00 Craig S. Smith? The photon in that junction stops moving.
21:07 Speaker 00 So the junction works as long as you're colder than TC, as long as the superconductors are superconducting.
21:16 Speaker 00 For the materials we use, colder than about 1 degree is enough. The reason why we need to be so much closer to absolute zero is actually thermal photons. If I had an IR camera and I looked at you right now, I would see you were glowing.
21:30 Craig S. Smith? Yep.
21:31 Speaker 00 And if something is glowing at the frequency of our qubits, it's adding photons to it. And so again, it messes up our quantum computer. So we need the chip to be so cold that it's not glowing at 5 gigahertz. And that's where that absolutely way colder than deep space numbers come from, is not so much from getting the Josephson junction to work or even the qubit to work, but just protecting it from infrared radiation. I said infrared radiation, I should have said heat radiation at microwave frequencies.
21:58 Craig S. Smith? Right.
21:59 Speaker 00 As for the photon, so there's a lot of different types of superconducting qubits as well, I should mention, that store the information in slightly different ways. The description I gave you is appropriate for a type of qubit that's called a transmon, that looks like a capacitor shunted by a Josephson junction. And in those transmons, what you're doing is basically keeping the Josephson junction in a place... Oh gosh, I'm trying to figure out how to explain this in a simple way. Josephson junctions have what people call a phase current relationship, that there's this relationship between how much phase is dropped across it, think of it as a little
22:42 Speaker 00 bit like magnetic field, and how much current is flowing through it. And that change phases whenever there's a voltage across it. And so if you think about it, what I'm saying is that voltages drive currents inside of Josephson junctions, but it's really voltages drive derivatives of currents. If I put a voltage across it, the current slowly goes up. If I reverse the voltage, the current slowly goes down. So in electrical engineering terms, it looks like an inductor. In this limit, as you know, I'm sure you can make these things do all kinds of weird tricks, this is just a particular way we use them. The cool thing about it is it's a ridiculously nonlinear inductor. And the reason why that's important is, you know, I talked about zero photon and one photon. You never said, well, what about two or three or four or five or six?
23:27 Speaker 00 And the answer is qubits shouldn't have a two state or a three state or a four state or a five state or a six state. That we need, really need a way to isolate that zero photon and one photon manifold. And the way that we do that is the Josephian junction, because it is so ridiculously nonlinear, we can make our oscillator so that if we're putting one photon into it, that takes five gigahertz photons, but the second photon has to come in at 4.8. And so we can make those frequencies so different that you can really isolate those bottom two states. I'm not sure I really answered your question very well.
23:59 Craig S. Smith? Yeah, but that's, that gets me there.
24:03 Speaker 00 Yeah, I've never tried to explain this without a chalkboard and a lot of diagrams.
24:08 Craig S. Smith? The issue to get, so you have those 156 qubits. And you say that they're particularly useful for matrix multiplication. Is that?
24:23 Speaker 00 Well, it's the operations that you go on, them all have representatives in linear algebra. That most of our gates you can write as unitary transformations, which are basically matrix multiplications that don't destroy any information, that they have an inverse. A little bit more complicated than that, but good enough for our purposes. It's just, again, they're in this sort of ridiculously complex or ridiculously large space that you're performing them.
24:52 Craig S. Smith? Yeah. And in that I'm interested, and my podcast is primarily about AI. How does that matrix multiplication relate to AI? I mean, is there an application? Will you be able to run artificial intelligence?
25:15 Craig S. Smith? Algorithms that depend on matrix multiplication on a quantum computer.
25:22 Speaker 00 So a lot of the time, we talk about kind of near horizon applications versus kind of long horizon applications, that there are things that you can do or almost do in our quantum computers today. And then there are things where you would need a very much larger error corrected quantum computer. We really should talk about error correction at some point to execute. So for things that you can do today, we need problems that map really efficiently onto the quantum computer. And so you're looking mostly at things in chemistry and material science, basically using the quantum computer to solve problems in quantum mechanics. And in fact, there are some chemistry demos now where we're really approaching the accuracy of what you can do with the kind of most up to date classical chemistry techniques. So we're really getting close there.
26:09 Speaker 00 In this near term period, a lot of the time, we're dealing with heuristic algorithms. Because at the end of the day, we're sort of typically competing against classical heuristic algorithms anyway. It sort of makes sense that you're pushed into that limit. And so the horrible thing is the only way to really test some of these things is to try them and see what comes out. And AI is very much in the category of like the king of heuristic algorithms. And so it's just really hard to say what it looks like to do it on a quantum computer when we don't have one large enough yet to actually represent any of these models or to do a circuit deep enough to run them. There have been some proof of concept demonstrations of something called kernel estimation, trying to find a good basis to represent information
26:52 Speaker 00 in so that it's sparse, which is a part of a lot of AI models, that if we deliberately design a problem so that there is a feature that we know quantum computers can recognize more efficiently than classical computers, and one of those is period estimation, like if I have some signal figuring out what the periodicity of it is, we can show that the quantum computer can run. And again, totally just toy small problems, not anything anyone would be interested in. But the problem is, nobody really knows how often do problems like that actually come up in real life, because we wouldn't be able to tell.
27:29 Speaker 00 So right now, sort of AI on quantum computers is very much in the field of small toy problems and wondering what it would really look like. There are a lot of heuristics that people are playing with in optimization, which is sort of one step removed from AI. You can kind of think of it as almost a training phase that look really encouraging. But again, they're heuristic. And again, not surprising, all the best classical optimizers are heuristic. So we're ending up in the same space in quantum.
27:58 Craig S. Smith? And I'm just going to stop for listeners to find heuristic.
28:02 Speaker 00 Oh, heuristic.
28:03 Craig S. Smith? There are a lot of.
28:05 Speaker 00 Yeah, sorry, I'm a scientist. I sometimes talk that way. There are some algorithms where I could sit down with a chalkboard and I could prove it to you.
28:12 Craig S. Smith? Right.
28:13 Speaker 00 That matrix multiplication, I can sit down with a chalkboard. I can write down the algorithm. I can say this does it. Fast Fourier transform is a great example. It's a way of going from time series data to what frequencies are in it. I can prove to you that that algorithm works just by doing math. There are other algorithms, quicksort, another great example. There are other algorithms like beam search, genetic algorithm, gradient descent on neural networks, where I can't prove to you that it's going to give you the right answer. I could even maybe prove to you that it's going to give you the wrong answer. But in practice, they work very well. And so we refer to those as heuristic algorithms,
28:54 Speaker 00 meaning you can kind of look at them and say, it makes sense that this should work pretty well. Maybe you can make a bit of an argument why it could work pretty well. But I can't prove to you that it's going to give you the answer you want. You just sort of have to try and see. Everyone's solutions to traveling sample split problems solve and fall into this category as well. So an example of, you know, you mentioned there are very few quantum algorithms. An example of one of those blackboard algorithms is called Shor's algorithm.
29:22 Craig S. Smith? Yep.
29:23 Speaker 00 And it's a pretty famous one because it's an algorithm for factoring numbers. That's a problem that's believed to be hard for classical computers. And actually, you mentioned.
29:37 Speaker 00 Oh my gosh, I forgot his name, Scott Erickson earlier. You know, he has this sort of famous description of Shor's algorithm of, you know, three things are possible. Uh, one is that it's easy to factor numbers classically and somebody's gonna get a Fields Medal. Another is that, uh, it's impossible to build a quantum computer or gonna figure out why on the way and somebody's gonna get a Nobel Prize. Or the third is that really quantum computers can do things that classical computers can't. And that's kind of an amazing outcome. So Shor's algorithm is one that I can prove on a chalkboard. Um, but it is incredibly hard to run. Nobody can build a quantum computer anywhere near close to be able to run Shor's algorithm on a usefully sized problem today. And so it's one of those kind of distant future algorithms.
30:20 Speaker 00 And so the other place where we could talk about AI and quantum, just to circle back to where it came from, is this distant future world, that there are large chunks of linear algebra making some assumption that you're able to prepare the input state sufficiently that you get, um, speedups with quantum computers on. Um, but the prefactor, the, the cost of running it on a quantum computer, because they're so much slower to start with, is so large that even if you had a computer that could do it, it would only be a winning situation for really absolutely enormously large problems. Um, we're a long ways off from that being possible.
30:56 Craig S. Smith? And, and that's because of the, the number of stable qubits.
31:00 Speaker 00 The number of qubits and the noise keeping us from running very deep circuits today. Um, there's a final algorithm people talk about called Grover Search. I don't know if you've heard of this. Um, it's kind of, it's a fun one because it's one that you can actually explain to somebody that knows a bit of linear algebra in a couple of hours, how to do it and prove that it works. Uh, and it does a weird thing. It searches an unsorted list in a time that's the square root of the size of the list.
31:28 Speaker 00 It's kind of, okay, that's a little bit weird. Like normally you have to search on about, or about half the list. So if I gave you a list of 10 items, you have to search about five of them. With this algorithm, you only have to search three. And it sounds odd, but if you have a problem where you, it also works on functions of the list. So if you have this list of a billion possibilities and you want to say, give me the one that this function returns 7 for. I can do that and only search a square root of a billion as opposed to half of a billion. And so if I had a huge, infinitely large computer, this is maybe really cool. But it's even further out there because again, the prefactor is so large that you're really just better off building a massively parallel classical computer and doing it that way.
32:13 Speaker 00 And, and this is another one of these places where I think people get really confused because it's an easy algorithm to explore. And so they think it's going to be important technologically, but really it's just so off, so far off there. It's unlikely to matter in either of our lifetimes.
32:26 Craig S. Smith? The, the, the challenges that, that everybody's working with is noise isolating from noise. And noise is, can be many things. It's not sound waves, right?
32:43 Craig S. Smith? And, and, and the other is error correction. The, the noise issue, I think everybody can conceptually understand.
32:54 Craig S. Smith? Error correction, can you talk about that?
32:56 Speaker 00 Absolutely. So to circle back to our roadmap, this is something we think is going to kind of come into the world as far as things that are useful for other people to use as opposed to research projects, kind of in the 2029 timeframe. So it's not that far away. And the game is pretty simple that it shows, maybe it's best to start with kind of classical error correction, which maybe not everyone's familiar with. It gets used a lot in communication. Imagine I'm trying to send you some data, but every now and again, my radio receiver or your, my transmitter or your receiver makes a mistake. And so I want to send you a one. I can do something really simple. I could send you three copies of that one. And then you'd look at them and if you saw 111, you know, it's a one. If you saw 110, you'd say, well, probably it was a 1, right?
33:41 Speaker 00 And so you can correct some errors that way. There are more complicated error correction schemes set on, based on, for example, parity checks. So imagine I was trying to send you eight bits. I could send you those eight bits, and then I could also send you whether there was an even or an odd number of ones in them. And that would be enough information that when you got those eight bits, provided there was only one mistake, you would be able to say this is correct or this is incorrect. Yep. If I added two parity bits, you would actually be able to correct the mistake, or you'd be able to detect larger numbers of errors, right? So I can build different codes that protect different amounts of data with parity checks. The more parity checks I add, the more mistakes I can correct, and, you know, the better the code works, but I had to send some extra data.
34:25 Speaker 00 And there is a huge theory of how to do classical error correcting codes, because communications is so valuable. It's very advanced, like, error correcting codes are actually built into the, I go back to the cellphone, the transmitter and receivers in your cellphones. We can also use error correcting codes on quantum computers. It's just, it's a little bit trickier, because if you look at the qubits, you destroy the quantum state. And so if I do that simple repetition code we started with, where I send you three qubits, and you've got to look at them and say, well, are they all, do they all match? That doesn't work. But those codes based on parity checks can actually work. And the reason is, I can design circuits that will tell me the parity of a set of qubits without actually telling me what any of the qubits were.
35:08 Speaker 00 And so I can kind of do a blind error correction that way. The tricky thing about this is, we have two types of errors we need to correct in quantum computers, what we call bit flip errors, one goes to zero. Yep. And what we call a phase error, which is where superpositions come out at the wrong angle. Think of it as a timing error. And so we need to, to correct one qubit, we need to use two error correcting codes. And I need to use enough of these parity checks that I can get to error correction as opposed to error detection. And our error rates apply to all the operations, including the ones that we use to prepare and measure the parity checks. So if the error rates are too high to start with, the code actually does more harm than good. Like, it actually damages the information as opposed to correcting it.
35:51 Speaker 00 So you put all these things together and air correction is actually pretty challenging. When people talk about air correcting codes that have enough protection for me to run some of those long distance, those far future algorithms. They get answers that we need air rates that are better than what we have today. If you look at our best processors, kind of three to 10 times better. So not, you know, miles off, but better. But the really big problem is with the air correcting codes, people understood until recently, the overheads in terms of number of qubits were huge. It would be pretty typical to talk about using 100 physical qubits to make one air corrected logical qubit. So we're saying we could take our best processor today and we could make one logical qubit out of it that was good enough for a far future system.
36:35 Craig S. Smith? And that's why people talk about logical qubits now.
36:40 Speaker 00 Yeah, and so that logical qubit would be 100 physical qubits plus the air correction apparatus around it. So if you think about that, it just really pushes the system size and cost, if you're using a code like that, to something that we felt was kind of beyond the reach of engineering. Last year, that changed. And the reason why it changed is a little bit subtle. People have also known about, you know, as I said, this is a really well-studied problem from communication, classes of codes called low-density parity check codes that in principle could have much lower overhead, many fewer additional physical qubits to protect the same number of logical qubits. It's just there was a dearth of kind of practical implementations that people knew that they
37:24 Speaker 00 can map to hardware, where people knew exactly how to do the center extraction, where people that measure the parity checks, where people knew how to determine from the parity checks what the errors were to start with. And we were able to fill out the details of one of those codes that we're calling the gross code and show an air correcting code that's in order of magnitude more efficient than anyone ever knew was possible before. Now, that didn't come for free, but the reason it didn't come for free, I actually think is really fun, because I told you I'm a hardware guy, I just love to make, you know, make things that work. The earlier air correcting code, air correcting codes, most popular was called the surface code. Only required nearest neighbor connections between the qubits. Remember I said there were wires that connect qubits we used to run operations?
38:08 Speaker 00 And so you could use qubits laid out like on a checkerboard and implement that code. To get these high rate codes to work, you need non-local connections. You can't do them in two dimensions anymore. You need a way to pop out of the plane and kind of do a highway overpass. And so the really great thing about the gross code is we were able to work with our air correction team here and kind of sell them, well, we think we could make a chip. Maybe we could have six connections per qubit. We think we could get the wires to go about 10 qubits away and still work. And they were able to design an air correcting code specifically around the constraints of what we thought we could build on the chip. And so that kind of that co-design is where this really came from. And so the other really exciting thing that happened this year is, well, end of 2025,
38:49 Speaker 00 we're now testing in 2026, is showing all the features that we need to implement that air correcting code on a single chip. And it's a really experimental one. It's not one that we're going to make available to clients over the cloud because at the end of the day, it implements two logical qubits, not very useful. But it shows all of these neat features. And it's really the prototype for these systems that we're going to be trying to build in 2029, 2030, 2031, that let us tackle these big chalkboard problems.
39:16 Craig S. Smith? And when you talk about 156 qubits, you're talking about physical qubits, not logic, logical qubits, I mean, not arrays of, yeah. The, how much of the work is on software, in software? I mean, working out solutions using code? How much of it is, as you just described, designing new configurations of hardware?
39:56 Craig S. Smith? And the new configurations of hardware sounds like a very expensive process. And is that done here? I mean, for example, the quantum computers I've seen here, are you?
40:16 Craig S. Smith? How fixed is that hardware? Are you constantly changing it?
40:22 Speaker 00 We have a huge research effort changing, improving that hardware, both from the design viewpoint, the materials that we put into them, how we fabricate them. It's a big chunk of what we do.
40:34 Speaker 00 Our team is probably about one-third focused on the hardware and two-thirds focused on the software, which kind of gets a little bit to your first question.
40:44 Speaker 00 That tackling the hardware part first, because that's just the guy that I am, until a couple of years ago, every quantum processor we built was built in this building. We have a small research fabrication facility here. It's small by comparison to like a Samsung or a TSMC or an Intel, that we're really able to get this entire thing off the ground. Recently, we've moved all of our fabrication, up to the, most of our fabrication, I should say, up to the Albany Nano Technology Center, which is a state-of-the-art 300-millimeter, again, research pilot production facility, not, you know, full-on production line that's co-run by IBM and NY creates. And the reason why we did that is our processors are getting more complicated with these extra layers.
41:29 Speaker 00 They run 24/7. They have modern, more automated tools. And so it was able to build these more complicated devices in the same amount of time, which is great. The finishing packaging of the devices is still done here. For everything else, a lot of it comes from elsewhere. The, you know, the control electronics, it's designed by the same people that designed System Z. But at the end of the day, it's built by subcontractors. And then it all comes together here or at our site up in Poughkeepsie, New York for assembly and test.
42:00 Craig S. Smith? And those quantum computers that I've seen, you have them in various places around the world for various research institutes to work with.
42:14 Craig S. Smith? When you change things here and you like what it does, do you then go out and change all those other
42:24 Speaker 00 Yeah, so we We have sort of two classes of client accessible quantum computers. One is we have our own data centers, which is where people can have cloud access to our machines. You actually get 10 minutes for free every month. So if you want to go try it, you can. And for those, we kind of upgrade them when it makes sense to us. Typically, they'll be running for about two years before we upgrade them. For the ones that you're talking about in other people's buildings, I think there are 13 of those right now around the world. And four more getting installed towards the end of this year. It's actually a pretty big fleet. Those are actually managed service agreements. So we still own the quantum computers that we install them at the customer site.
43:09 Speaker 00 We manage it. We keep it calibrated. We keep it running. They get to use it to run whatever they want to on it. Those typically include a midterm upgrade after about two years with the expectation that this is a technology that's moving so quickly that by two years out, you're really going to want the latest and greatest anyway. And that's something we coordinate with when we actually have something that's appreciably better.
43:33 Speaker 00 We turn these devices around pretty quickly. You heard me spew out about a half a dozen bird names. But we're working on three different versions of that Nighthawk processor right now. And whichever ones look good, that'll be the next one. So it changes incredibly quickly from a hardware perspective.
43:52 Craig S. Smith? So on the timeline, quantum advantage? A verifiable quantum advantage this year? Or I should say, repeatable quantum advantage, right?
44:06 Speaker 00 Verifiable. And if you go look at the tracker, arguably, we could have a discussion. Maybe it's already there. It's close.
44:14 Craig S. Smith? Yeah. And what's the next milestone? You know, without going out 10 years.
44:21 Speaker 00 So we have Two roadmaps. We have a development roadmap, which is about devices that we think are computationally interesting that will go out to our clients. And for the next three years, 26, 27, 28, what's on that development roadmap are improved versions of Nighthawk, which is our current state-of-the-art processor, improved in the sense that the gate fidelities will get better, so fewer errors. And that's really that materials and physics research and larger versions of them. And that's really that engineering and system scaling. And so these are going to get continuously more and more capable. On the background of that, we have a lot of work on how to use them better, using techniques like air mitigation or even
45:03 Speaker 00 smaller versions of air correction that don't give you large-scale fault tolerance, but can fix some small errors, fix some specific errors, getting put into the circuits people run so they can accomplish more and more. And we try to capture that with a single number, which we call the number of gates the device can run, the number of operations it can run.
45:24 Speaker 00 And the other roadmap is the innovation roadmap. And that's the one where we capture the research demos, the things like these air corrected devices that we don't think are computationally useful yet. And so over the next three years, this year, we're going to be demonstrating logical qubits. Next year, we're going to be demonstrating connecting two modules of logical qubits together, because to get the system size that we want to build, we can't do it as one huge monolithic chip. And then the year after, we're going to be demonstrating universal computation on these guys. And at that point in 2029, we are going to be hopefully bringing those air corrected devices to clients, and these two roadmaps kind of merge again. So I have to say, you know, I've been at IBM for over a decade now. I've been doing quantum computing for much longer than that.
46:09 Speaker 00 I used to think it was really a question of, am I going to see an air corrected quantum computer in my lifetime? And now to be saying, actually, we think we're going to be building one that's big enough to be computationally useful in the next four years is both amazing and terrifying.
46:24 Speaker 00 But it's really these new, more efficient error correction codes have rewritten what's possible in the next decade.
46:32 Craig S. Smith? The, you know, I was asked to do some writing for different clients. I was asked to write a piece about now is the time that enterprises should start training a cohort of people so that they understand quantum algorithms and they understand the timeline so that when it arrives, they'll have some workforce or subset of the workforce that's ready. Do you think that's too early?
47:08 Speaker 00 I don't think so from a couple of perspectives. One is that right now is the time when you really should be figuring out what algorithms that you want to use because everyone has their own problems that they want to solve. And how to map that, even in the abstract sense to a quantum computer, whether you're using one of these heuristic algorithms or a blackboard thing that you can run in a decade, four years, it's a big challenge. And finding people that understand both your problem and quantum computing well enough to do that mapping, it takes time. The other thing is, I kind of like to say, if I could have told you four years in advance, System 360 is coming out and you could have been,
47:51 Speaker 00 you know, ready and raring to go on the day that that thing was available, that's a huge advantage. And so kind of from both of these perspectives, if you're a large enterprise that's investing on, you know, this four and five year horizon, I think it is absolutely the right time to be thinking about it. The other thing to keep in mind is we do have these near-term heuristic algorithms where we're getting really close on optimization, we're really getting close on chemistry, where it seems like we may well cross some additional thresholds in the next couple of years. I can't promise it, they're heuristic, but it's very encouraging.
48:28 Craig S. Smith? And final question, a minute ago, you said, you know, in four years, that's kind of terrifying. Is it terrifying because you now have this goal and, you know, are we going to get there? Or make the deadline, so to speak? Or is it terrifying because of the power of quantum computing and how it's going to impact the world?
48:54 Speaker 00 I think the impacts are going to be overwhelmingly positive, actually. That, you know, we named the ages of man after materials. And so to say that I have a computer that's going to help material science is, you know, how can that not be a wonderful thing? For me, it's terrifying. And we talk a lot here about cycles of learning. You know, I'm going to build this device and from it, I'm going to learn the things I need to do the next one. And from that, I'm going to do. Four years is short enough that I can now count the numbers of cycles of learning that we have left before we have to deliver that device. And although it's possible, it's honestly, it's going to require a lot of very hard work. And we know things are going to go wrong in the meantime.
49:35 Speaker 00 And so it's, the concern is more about timeline, about are we going to hit that 2029 year? Or too many things going to go wrong, we're going to slip out a year. As I said, we put a lot of pressure on ourselves to stick to that roadmap. And so the sense in which it's most terrifying to me is, yeah, keeping on that schedule. Like doing science on a schedule is a tricky thing.
49:58 Craig S. Smith? That's right. Okay, Oliver, well, thanks very much. That was fascinating and I'm sure will be fascinating to listeners. There's a lot of information there. And I encourage people to run a transcript and run it through ChatGPT. If there are things that you don't understand. So, great, I hope we talk in four years and you can tell us how things have advanced.
50:25 Speaker 00 Well, I hope I see you back in the building again sometime before that. And maybe we can take you around some of the labs.
50:30 Craig S. Smith? Yeah, I would absolutely love to, yeah.
50:33 Speaker 00 So, and yeah, it's been an absolute pleasure. And, um, yeah, maybe next time you could also tell me how we can use AI to make this all just a little bit faster.
50:43 Craig S. Smith? Yeah, okay.
50:44 Speaker 00 Take some of the tear out.
50:45 Craig S. Smith? Okay, great, thanks.
50:47 Speaker 00 Thanks.
Previous Why Agentic-First Startups Won't Disrupt Enterprises as Fast as Everyone Thinks | Kris Lovejoy Next The Single Biggest Barrier to AI Adoption Isn't the Technology — It's This | Errol Gardner of EY Page updated October 30, 2026 · transcript and summary are machine-generated
