Quantum San Diego Convening | Alan Ho, Co-founder and CEO of Qolab

0:00 So, I'm going to talk about commercialization of quantum computers and actually I put to this slide together about an hour ago um because um there's a very big uncomfortable question I'd like to start us to ask and and try to answer. Which is how much will a useful quantum computer cost? All right, and this is very relevant right now because there is an RFI out there which is asking um about say scientifically relevant fault-tolerant quantum computing systems. I know there's a few people from the various labs here. And in the RFI it specifically asks provide a cost estimate for the proposed system.

0:44 So, this is something I know that nobody wants to talk about because when they actually run the numbers it looks really scary. Okay? But I'm here today to see if anyone wants to collaborate to solve this problem. And I'm going to give you some numbers that I've come up with. They're a back-of-the-envelope numbers. The second thing is that we know in order to build a quantum computer we need a supply chain. And if you look at the supply chain of let's say the semiconductor industry, it looks something like this. So, at the very bottom you have the foundry tools. Then you have the foundries that make the chips. You have EDA manufacturers like Synopsys, Cadence that build the design tools. You have fabless quantum computing companies like Collab, right?

1:27 Qualcomm is actually another fabless company, right? Then you have components on top of that, dilution refrigerators, um uh uh um quantum uh control systems. You have uh ODMs. These are like the Foxcon's and the Wistron's of the world. And then you have the computer manufacturers like HP Enterprise, Dell, or some of the hyperscalers. And they're asking the questions like this. At the very bottom, you know, the tool manufacturers are asking how many tools I want to make. How many wafers am I going to be selling? How many engineering seats will I have? These are all numbers that are really important to make the supply chain work. And if we don't have a supply chain, we can't bring down the costs. Hey, we can't even manufacture it, let alone bring down the costs. So, we have to kind of tackle this very, very uncomfortable problem.

2:21 So, I'm going to talk about one problem, which is the holy grail of quantum computing, I guess, which is quantum enhanced foundation models. And the reason why this is important is because there's simply so much money flowing in the into AI at the moment. On on the order of 4 to 8 trillion dollars is just committed by the hyperscalers alone in quantum in in these data centers. The largest data center right now is being built in Ohio, I believe, by by Google. It's about 40 billion dollars.

2:56 And today, the training the training versus inference split is around 50/50, but over time, there's going to be a lot more inference than training. I think the really important thing is, especially for the people in the audience, is that the hardware spending right now for AI is speculative. And to give you a little context here, about 2 years ago, the largest training cluster was about 100 million dollars. Last year is between 4 and 7 billion dollars. This is the Colossus. This is the open XAI's big data center.

3:34 This year, next year, probably in the tens of tens of billions, if not hundreds of billions of dollars. And they don't know whether or not if you throw on more GPUs, whether that's definitively going to give you better performance. People think there is, there's some good extrapolations because of doing that, but they really don't know. And this is very important for us as a community because this means that the AI people who are already allocated dollars for buying hardware are willing to spend money on speculative things, which is very important.

4:06 Also added to that is that the next wave of AI is probably around physical AI, of which quantum computing is an important part of quantum is a important part of physics. So you can take a look at our We have a paper called how to build a quantum supercomputer and in it we talk a lot about different We did resource estimates based on realistic systems. And so I'm going to talk about the alpha-conotoxin as a good proxy for simulation, all right? So it takes about 26 days to simulate on a decently fast quantum computer. Before any of you ion trap And and by the way, superconducting qubits are about a a thousand times faster than neutral atoms and and ion traps. So this is actually for dollar for This this speed over here is actually very fast.

5:00 However, when you put that in a real-world system, let's take an example inference workflow. An inference workflow with generative AI today that people are doing this in drug discoveries that they would use generative AI to generate a 10,000 candidates, use screening to get down from 10,000 to 1,000, use HPC simulation down to get to 50, and then and then the fifth they would have a lab that would do some tests to come up with 10, and 10 molecules would go to some sort of human trials.

5:30 And so one could ask what how's quantum going to come in? Well, probably come in here, right? To cut down the number of actual lab lab molecules you need to to verify. So you need to simulate on the order of 50 molecules, right? And if you actually go do the math, and you do the work and the workload, um you know, people actually Sorry, there's a mistake here. Um but at the end of the day, if you want to have a workload of simulating one molecule per week, you need roughly about 530 quantum computers for inference.

6:05 It's very big, something that nobody really wants to hear again. Okay. Uh what about synthetic generation of data? So, AI, a lot of it is garbage in, garbage out. So, a lot of people are thinking about using quantum computers to do simulations to train machine learning models. There's also another really important part of quantum uh of AI called reinforcement learning. And the area of reinforcement learning, what one would do is they would create an agent that's powered by LM uh that would generate data. That data could be then simulated by the quantum computer, let's say binding energy of a molecule to some uh some protein, and then uh it would update to the it would update the the large language model. So, again, this is this is very similar to the inference use case.

6:56 You actually note that because of this whole agentic workflow, uh CPUs are now back in vogue. Right? And the And the that's because there's a lot of these coding systems where they would use a generative AI system to generate code, but that code has to be run on an actual CPU to see if it actually executes. That's why Intel stock is up like crazy. Similar thing can happen with QPUs. You can use a QPU to simulate molecules generated from generative AI systems, and you can be in this kind of in reinforcement loop.

7:27 But assuming, let's say, you need on the order of 10,000 molecules at minimum, you're looking at the order and you do all the math, you're order of 821 uh quantum computers, which is again a lot of quantum computers. So, if you actually go back of the envelope and try to figure out, okay, what does that mean from a manufacturing cost perspective? Again, very daunting. Let's say this rounded up 1,000 quantum computers per data center, right? Uh in our module we'll talk about a little bit more. We're we're trying to put it together 20,000 qubit modules.

8:02 Assume three wafers per module. You work out all the math, it's about 500 and uh uh for a data center is about 535,000 wafers. Assuming that there's going to be a 3-year uh hardware refresh, you're talking about 145,000 uh wafers per year, which is about the which about same order of magnitude a fab would run uh building uh wafers 24/7. So, this way you can actually start back of the enveloping all these costs.

8:34 Now, what does that mean? Let's say the biggest data center is $40 billion, right? And let's hypothetically say that the budget for quantum computing is 10%, which puts that $4 billion. It's still very big number, right? That means that the quantum computer each would have to cost about $4 million. Quantum computers, unfortunately, do not cost $4 million. So, let's take a look at the costs and where do they come from? This is 150 cubic quantum computer, not a not a 3 million cubic quantum computer. And if you look at the cost, majority of it is in this wiring. We talk about the rat's nest of wires. We got to get rid of all of this. There's also a lot of other costs like the filters, the amplifiers.

9:14 Actually, the fabrication of the QPU, I put this here as just a rough number, is actually the lowest part of it. It's all these other components. So, how we're going to get that how we're going to reduce all these costs? Well, the way we can do it at least at least with Colab's plan is that we want to have cryogenic integrated circuits. Put all that rat's net rat's nest of wires, filters, amplifiers, etc. all onto one system.

9:37 Uh and so, the first part is signal processing. You can put the wafer you can put the um, the signals instead of coaxial cables. You can use flex if you go into 4K. Uh, you can use cryogenic multiplexing to reduce the number of wires actually going down into your system. And then one can use uh, filtering solutions. This is a coplanar waveguide filtering solution to integrate the filtering right into your interposer. Likewise, instead of using very expensive TUPAs and and HEMT amplifiers, one can use uh, Josephson photon multipliers and slug amplifiers to get the signal out. So this way we believe that we can significantly reduce the cost. And if you look at this chip, it yeah, it is artistic rendition and it looks like a H100 GPU and that's on purpose cuz what we want to do at Co-lab is leverage as much of the semiconductor industry as possible to reduce the costs.

10:30 Okay, reality check. The current DOE budget is about 1.3 billion for both AI and quantum. Right? So how how are we going to build a useful quantum computer with that budget? You probably can't. You really have to figure out how to leverage the much bigger budget of the four to eight trillion dollars that are being allocated in next five years. So my opinion is that the best strategy is that the DOE provides a proving ground for the commercially relevant fault-tolerant quantum computer and leverage the existing commitment towards hardware from the industry to build these quantum computers at scale.

11:11 All right, so call to action. Uh, there is an RFI a DOE RFI right now. Uh, if you're interested in collaborating on the RFI, please uh, contact me allen@colab.ai. And I also want to note that we have a consortium that we're working with. You know, we can't uh, do it all alone. It's called the Quantum Scaling Alliance. Many of those companies are actually in the audience today. Quantum Machines, Hewlett Packard, Applied Materials, they're actually all here. And we know that we have to work together to figure out the supply chain, the costs, and as well as building the technology. So, let contact me if you're interested in participating in the RFI and collaborating on this project. Thank you.

11:55 >> [applause] >> Uh yeah. So, um uh com- recycling uh so, I think we're actually one of the really important things if you look at this dilution refrigerator is that you want to uh this is the the the dilution refrigerator is something that one we can reuse. Right? So, I think as we are building these systems and we're thinking about scaling up, we actually want to build systems that are upgradeable. That's the best way that we can drive reuse. So, for example, if one can uh come up with the base system uh that we can put bigger and bigger chips with more and more flex wires, that is something that I think is really important. And it's also able to scale with the cost cuz right now, you may the com- cost of the components, okay, maybe $7 million is your your budget right now. So, at most you can get 500 150 cubits. But then, as technology improves, let's say we just throw on flex wiring with the filters, maybe from uh from 150 cubits, you can go to, let's say, 300 or 1,000 cubits, you want to make sure your system you want to be able to reuse as much of the dilution refrigerators and other systems as possible.

13:08 >> I guess there's a question? >> Yeah, Scott actually Oak Ridge National Lab. Thank you for your talk. This is very helpful. Uh the RFI is out. That $1.3 billion, I think, is just for AI. I don't think there was any money budgeted yet for quantum. >> Oh, that's great. Thanks for the great news. >> [laughter] >> And on your I was not aware of the Google data center going in at you what you had 30, 40 billion?

13:31 >> Yeah, it's a 40 billion-dollar data center. I think it's in Columbus, Ohio. >> So, Microsoft at a recent DOE conference said that their quarterly spending right now is 37 billion. >> Correct. >> One out of four electricians in the US are working on their projects. >> Yeah. >> Yeah. >> Yeah. >> So, those numbers check out. >> Yeah, yeah. >> But thanks. I'll check I'll get with you later. Thank you. >> Okay, so um I just want to make a comment about the calculation for the um using quantum computer to generate token for uh AI generative AI. The one thing you should consider is the actual value not only the cost is like if you use the data for pre-training the the value could be much exceeding you know the kind of particular application like in inference for like drug discovery.

14:28 Because we don't know what would foundational uh the next generation foundational models that are equipped with quantum data would be able to do. There is some emerging phenomenal cross-domain type of property uh like you know once you feed in the data they might just have different ideas about the bunch of other things that not might not be obvious to us to estimate now. So, that might you know the value could be exceeding what we are estimating.

14:57 I'm saying that you know quantum enhanced foundational models could be much more powerful if if we you know include it in the pre-training. >> I don't know. >> Absolutely. And so, that's why I think like when you um let's say for a cluster right now 40 billion dollars uh the question is what percentage of that cluster would they spend on quantum? And I think the interesting thing is that the the the stakes are so big for AI that they're willing to spend a lot of money on it. So So I think that's this is kind of like why I think this is a particular application that we should all be paying attention to. And you're right, it could be people might be willing to spend more spend more than 10% of their cluster. It could be 30, it could be 50, it could be expand the cost of each data of every training cluster. So something that's something to think about.

15:49 So I I think that's hard to tell because like you know the way that you want to use a quantum computer is to use it for like applications that would be impossible to do on a classical computer. So for example, when we did this quantum supremacy experiment, we worked with Oak Ridge quite a bit to simulate what was the speed up or or improvement. And basically it was something like a hundred a 300 second at the time it was around 300 it was on the order of about a billion billion times speed up and a trillion times improvement in energy efficiency. Those numbers just don't make any sense when you say oh what percent how what percentage can you do more? So it's really what you want to do is you want to find applications that are you know so cost prohibitive to to to to work on.

16:46 Now there's there is one thing about the quantum data that right quantum data is garbage in garbage out. So you're going really going to be competing against things like DFT and there's still an open question especially for synthetic data generation which is what is the amplification of the error in the synthetic trend the synthetic data when you actually create a generative model. That's still a very open question in science. So it's it's very hard for the qualify like what is the improvement what is the improvement of um what is the value of just reducing the error in the quantum simulation?