NVIDIA Corporation (NVDA) Earnings Call Transcript & Summary

Good morning, everyone, and welcome to Rosenblatt Securities' Fifth Annual Age of AI Scaling Tech Conference. My name is Kevin Cassidy. I'm one of the semiconductor analysts at Rosenblatt, and it's my pleasure to introduce Gilad Shainer. He's NVIDIA's Senior VP of Networking. Also, we have Stewart Stecker. He is NVIDIA's Senior Director of Investor Relations. So on NVIDIA, we have a buy rating and a $200 12-month target price. And we're bullish on NVIDIA not only because of their leadership in AI but now their ability to expand into full rack-scale deployments, including scale-up and scale-out networks. So we're fortunate to have Gilad speaking with us today. Gilad is a networking expert. Gilad joined Mellanox in 2001 as a design engineer and has served at senior marketing management role since 2005. And of course, NVIDIA acquired Mellanox in 2020 and Gilad serves -- he also serves as Chairman of the HPC-AI Advisory Council Organization, and he's President of the UCF and CCIX consortiums and is a member of IBTA and a contributor to the PCI-SIG, PCI-X and PCI Express specifications. So Gilad also owns or holds multiple patents in the field of high-speed networking. So with that, first, I'll turn it over to Stewart to go over some of NVIDIA's disclosures.

Thanks, Kevin. Thanks, everyone, for having us. As a reminder, the content of this call may contain forward-looking statements, and investors are advised to read our reports filed with the SEC for information related to risks and uncertainties facing our business. So back over to you, Kevin.

Thanks, Stewart. Yes. So I'll kick off the fireside chat with a few questions, and we'll take questions from the audience also. And to ask a question, click on the quote bubble in the graphic on the top right-hand corner of your screen, and I'll read the question to Gilad and Stewart. Keep in mind that this is a fireside chat working towards the understanding of NVIDIA's network strategy. Gilad will not be taking questions around financial guidance. So with that, thank you, Gilad, and great to see you again.

Thank you very much, Kevin.

So maybe we'll start with a very high-level question of, what is the strategic importance of networking in AI data centers?

Well, it's a good start. It's a good start. So first, the data center is the unit of computing today. Previously, it was an element or was a CPU and more GPU. But today, it's not the GPU, it's not the server. It's the data center, right? The data center is the unit of computing that we use. Now networking defines the data center. The way that you connect those computing elements together will define what that data center can do. It could range from just building a server farm all the way to building an AI supercomputer that can run a single workload at large scale and to do amazing stuff. So the networking or it's used to call networking. I'm not referring to networking anymore. It's more like this is the computing infrastructure, okay? It's much more than a switch. It's much more than a NIC. It's a computing infrastructure. And that's why it has become so critical and so important. And that infrastructure will determine what kind of workloads you can do. What will be the efficiency of the data center? What will be your return on investment? How many users? How many workers you can bring in? How many tokens you can support? How many end users you can host on the data center? And this is where the networking or the infrastructure is so critical. Now when you go and design a networking for AI data centers, it's completely a different task than designing networking or infrastructure for the traditional hyperscale clouds. Here, we're not talking about single server workloads. We're talking about distributed computing. We're talking about workloads that need to run on over multiple compute engines, which could be hundreds and thousands and tens of thousands and hundreds of thousands. So you need to make sure that every GPU here gets the right throughput. Every GPU needs to be fully synchronized. So the data that goes over the network needs to hit every GPU at the same time. If you create skews on network, if you create what we call tail latency, then 1 GPU is going to finish later than others. And we all know that when you're running an AI infrastructure, the last element to complete the task will determine the entire performance of the data center. So it's the tail latency, it's the throughput, it's the latency cost. It's making sure there is a congestion control. There is a huge amount of elements that are in that infrastructure. That infrastructure will determine what you can do with the data center that you build. That's why it's so important.

Great. When I talk to investors, they've been hearing terms now that they're just a little confused on, but maybe as you talk about connecting the entire data center and even data center to data center, but the terms of scale-up and scale-out networking, that's new to some investors. So maybe if you could just give that explanation of what's the difference and why are each important?

Yes. And I'll try to make it maybe a little bit simple because I see that there is terms and people try to define what is scale-up and what is scale-out. First, we can start with examples, okay? When we design an AI supercomputer, our scale-up infrastructure is NVLink, and our scale-out infrastructure could be InfiniBand or Spectrum-X, okay? Those are the examples. Now what's the difference between them? Scale-up is your ability to build a larger compute engine, okay? So in a scale-up infrastructure or connectivity, we're taking those GPU ASICs, let's call it like that, or GPU packages, and we want those GPU packages to behave like one. And in order to build that one, you need to scale up infrastructure. That's what the scale-up network does. It takes those components, making sure that all of the balance between them, kind of the right message rate, the right connectivity, the right elements are there in order to make those engines behave like one. And this is why if you see Jensen keynote, he says that his GPU is the rack, right? It's GPUs, not the ASIC. It's like we have NVLink72. So that rack is the GPU and scale-up network enables that. Okay, so scale-up enables to build larger GPU out of the different ASIC components. Now once you define that larger GPU, now you need to connect those GPUs together. And how many GPUs you connect together, it depends on what kind of workloads you're going to run, right? What is the mission that you want to achieve? Connecting those GPUs together in order to form multiple GPUs that would work together and run those larger missions, this is where the scale-out network is needed, okay? So there is different requirements from a scale-up infrastructure versus the scale-out infrastructure. One's create a larger compute engine and the other one's connect multiple compute engines in order to support the different missions that you want to run on the data center.

So just as an example, it was only last year that NVIDIA was networking scale-up 8 GPUs. And this year, it's 72 GPUs.

Right. And we talked about 576, right, in keynote that Jensen talked about that as moving forward. And it's all been determined according to what is the workloads, what is -- what are the workloads that you need to support. And obviously, as workloads continue to evolve and new workloads continue to emerge and you need to solve new things, then everything in a data center is being added or changed, right, or progress. So one of the example is that, that unit of computing that was maybe a single GPU then becomes 8 GPUs on NVLink. And now it's 72 GPUs and it's going to go to 576. It's all in order to support what kind of workloads you need to run today or you need to provide today.

And maybe you touched on it, just the workloads. What is happening with the AI workloads and applications that are influencing some of the network requirements?

Yes. So what you actually do is you build a data center, right? And that data center is aimed to serve workloads that you define or the workloads that you want to run on the data center. So essentially, everything is -- needs to be connected together, right? And there's different elements that you can look at. First is building or codesigning the network with the compute, for example, right? And this is important because you design a data center, you don't design components. And I'll give you 2 examples of what does it mean, codesign. One example is that in the traditional world, let's say, there was compute engines that doing compute, and then there was networking elements that were tasked to move data, right? That was the separation between them. But when you design a data center to serve the AI workloads and you have the ability to decide where to put to what, and now there is no boundaries. And for example, we took compute engines kind of traditionally run on compute components. We took compute algorithms and running them and we're running them on the network. For example, what we call SHARP in InfiniBand is taking -- is doing data analysis on the data on the network. So the network is not just moving data, it's actually participating in the compute cycles, right? And why are we doing that? Because once you do the reduction operations, for example, on the network, you can save half of the bandwidth that you need to run and you can complete things much faster, okay? So this is an example where you move things from compute to the network. On the other side, traditional network topology was built in the concept of a top-of-rack switch, which means that all of the NICs will go to the switch on top of the rack and then you will connect those switches together. This is the wrong thing to do if you build an AI data center. Because you mentioned, for example, in previous generation with 8 GPUs connected on NVLink. So those 8 GPUs already communicate between themselves on NVLink. They don't want really to continue and talk with themselves again. So why would you put all of those GPUs on the top-of-rack switch? It doesn't make sense. You want to spread that connectivity and have every GPU connect to other GPUs in a fabric, and this is where we created a multi-rail topology, okay? So now the network is designed the way that the compute is running, the way that the compute algorithms are running. And then we're taking some of the compute algorithm actually running on the network because it's much more efficient to do it there, okay? So this is one element of AI workloads required to actually design a full data center, design that unit of computing, and then you want to do that in a full synergy in a full core design, okay? That's one thing. The other thing, of course, is that AI frameworks continue to evolve, right? That's why every year, we have a new compute engine that's coming out. There is new network infrastructure or computing infrastructure coming out. There is a new GPU, there is new NICs. There is new switches in order to serve scale, okay, because we see increase in scale. You're moving from thousands of GPUs to tens of thousands. Year after, you go to hundreds of thousands of GPUs. People are talking about now the million-scale GPUs. You need to actually be able to grow that element. You have so many routes. Just think about it, with all those GPUs, every GPU communicate on the network. There is so many routes that you need to make sure that you send them in the right direction and no one's going to collision with another one so there is no congestion on the network. So there is so much complexity in that network. And that's why you see that there is, every year, there is new generation, new capabilities, new elements that are being brought into the compute infrastructure to support the full data center design and to support the different kind of workloads that we see.

Okay. Great. You're connecting these hundreds of thousands now of GPUs. But maybe if I even roll it back a little bit with the Mellanox acquisition. At Mellanox, you had both Ethernet and InfiniBand. I guess what's the difference in the 2 offerings? Is there scale-out? And if we go into depth and talking about the scale-out networks, how do you decide whether InfiniBand the right solution or Ethernet is the right solution?

Yes. Actually, we let our customers decide what makes sense for them. And maybe I'll start a little bit in the history. Yes, Mellanox did start with InfiniBand. And InfiniBand was built for distributed computing workloads, okay? It was built in a sense that, first, it is lossless. Unlike traditional network that lossy was fine, meaning traditional network works in a way that if there is collision on the network, you don't try to solve that collision, you just drop packets because it's okay. I can retransmit the data. But when you deal with distributed computing applications, if you drop data, you're going to retransmit the data but you don't just retransmit the data to a single GPU set, for example. The fact that you retransmitted data to a single GPU and that GPU become late in the whole scheme of the workload, now everyone else is waiting, okay? So you cannot -- you don't want to retransmit data. You don't want to drop data. So InfiniBand started as a lossless network. You don't want to drop data. You don't want to create the latency. And then InfiniBand, in order to do that, brought congestion control and adaptive routing elements later on and so forth. And it was great for scientific computing, great for HPC and essentially, it's great for AI because AI is distributed computing. And today, InfiniBand is still the gold standard for AI. Everyone that builds a network always compare its network to InfiniBand. Even when we did Spectrum-X, kind of creating Ethernet for AI, we compared it to InfiniBand. That's the gold star. It is the gold star. It brings element that no other network exists and it's a great solution. So if you build an AI factory, single job, running large-scale in InfiniBand, there is nothing better than InfiniBand, okay? It's the gold standard. Now NVIDIA also brought Ethernet, right? We designed Ethernet for you. And you can ask -- if you have InfiniBand and InfiniBand is so great, why did you guys bought Spectrum-X? And the reason for that is that we believe that AI is going to go everywhere, okay? Every data center will run AI. And therefore, there will be AI clouds, multi-tenancy, multi-workload and multi-users. There will be AI in enterprise, right? I'm talking about enterprise AI. We see a lot of enterprise now adopting AI. Those areas are being built by people that are familiar with Ethernet for many, many years. They build their software stacks. They build their management tools all on Ethernet, okay? And if they continue running with Ethernet and keep their management and keep how they support their enterprise company and so forth, that would be much better for them. AI is evolving so fast and, therefore, start learning how to handle InfiniBand, for example, manage InfiniBand, meaning they're going to lose the train, right? So we wanted to help them. We knew it's going to go to everywhere. And everywhere means that we want to bring Ethernet to AI, okay? We want to enable Ethernet as an option for AI. And for people that build AI data centers and they are familiar with Ethernet, their software depends on the software ecosystem of Ethernet, okay, all the tools that were created. Their own management infrastructure that is there and was built over the years and progressed over the years around on Ethernet, we don't want them to create -- recreate it again, okay? So for them, Ethernet is a great thing. And this is where we built Spectrum-X. Now one important thing to kind of know what we did in Spectrum-X. Well, Spectrum-X is the first generation of Ethernet for AI because nothing in Ethernet fits AI before Spectrum-X came. Spectrum-X is actually not the first generation. And the reason is that what we did is that we brought things from InfiniBand, from the multi-generations of InfiniBand that continue to evolve over the years. We brought those elements to Ethernet, okay? So that's why Spectrum-X on one side, it's kind of the first Ethernet for AI, but what we brought inside has years of development on the InfiniBand side. So that's why it came in very mature, very quickly and actually completely aimed to solve the problems of AI on Ethernet. And for example, we brought lossless to Ethernet because we don't want to drop packets, right? So we brought lossless. We brought adaptive routing capabilities. So you have lots of flows between GPUs. You want to make sure that every flow will go in the best available path, right? It's like solving the routes in a sense. We brought the congestion control to Ethernet. Okay, no collisions. You want to make sure no collisions. That application, one application cannot impact another application by creating collisions on the network. We brought many things from InfiniBand into Spectrum-X and actually created Ethernet for AI. And now you have InfiniBand, which is a gold standard. And if you're running, building supercomputer in a single job, if you know how to manage InfiniBand, use InfiniBand in the past, there is nothing better than that. And if you're running Ethernet, you can keep running Ethernet. We brought the best Ethernet for AI on Spectrum-X. And a good example for that is that Spectrum-X is running more than hundreds of thousands of GPUs, more than 100,000 GPUs in a single data center for a single workload. There is no other Ethernet technology that they managed to achieve what Spectrum-X did. And the reason is Spectrum-X is built for AI. So we have a great Spectrum-X. There is great InfiniBand with Quantum. And now people can choose what makes sense for them based on their workloads, what they need to serve, what they're building, what's their familiarity, what is their software ecosystem and so forth.

So that's just the mix. You can put those features into Ethernet, and that's more of a physical layer features that you're doing. And so it doesn't affect the customer's Ethernet that they have been running for years?

So it's not in the physical. It's a combination. It's the physical, it's the link layer, the transport level. It's the way that the NIC runs with the switches, okay? One of the things that made InfiniBand so great is that it's a platform, okay? It's not a separate NIC, a separate switch. It's like it works together. The NIC get information from the switch network in order to determine the flow of data. The switch element knows how everything in the data center behaves, okay? It's like you need to know not just your own status when you do routing on the network. You want to know the status of your neighbors, and the neighbors could be the NICs on the switches. Because if my neighbor switch has some issues, for example, I don't want to continue to send data to the same area. So there is a global load balancing that happens. And the NICs work in conjunction with the switch, it's a full end to end, okay? So this covers everything from PHY to link to transport. Now on top of that, you have all the management stack and you have all the cloud management tools, for example, and hosting multiple tenants and so forth. That runs on top of that infrastructure, not on the network, okay? So what we brought into Spectrum-X cover all the infrastructure limit. Everything that runs on top of that could be the same. And this is where it goes easily into people that build Ethernet for data centers and build their software ecosystem. Now it actually goes directly there and you bring them the elements of the infrastructure that are needed for running AI training or AI inferencing.

Okay. So your Spectrum-X could be mixed in with standard Ethernet? Some other racks could be running standard Ethernet?

So Spectrum-X is Ethernet, so it's interoperable with any other Ethernet devices. So if you build, for example, an AI data center, you built a unit of computing, right? So that means that Spectrum-X will be the scale-up infrastructure, for example, and it covers the full stuff. Now that data center can be connected to other parts of your infrastructure, right? It can connect to storage, it can connect to another data center, it can connect to users, their desktops and stuff like that. And this is where you might see other kinds of Ethernet, right, connecting to desktop. Traditional Ethernet is great. And of course, you can connect that traditional Ethernet into that data center that has Spectrum-X for the scale-out infrastructure.

Okay, great. And maybe if we could touch on too, you had mentioned the NIC cards and your DPU or the BlueField. Can you talk about the importance of that, of having control over the DPU within that same network?

Yes. So the DPU actually bring another element of that infrastructure. And so when you build a data center, there is no 1 network. Traditional world, there was 1 network. If you go to hyperscale clouds, there is 1 network. You build an AI supercomputer, different story. And you mentioned that already because you asked the scale-up and scale-out. So here are 2 networks, right? There's at least 2 networks, scale-up and scale-out. Now there is also access network, meaning users need to access the data center. That's the third network. Now there is also a storage access that might be even a fourth network, okay? So there is multiple elements in that AI data center, and there is different components to each. So if you look on NVIDIA AI data center, we use NVLink for scale-out. We use Spectrum-X for scale-out or InfiniBand. And that scale-out includes the switch and includes what we call SuperNIC, okay? And that SuperNIC has compute element inside of it in order to determine that injection rate and process telemetry from data from the network and so forth. And then we have the DPU on the access network because what the DPU enables to do is to move the data center operating system from the server compute engine into something else. And that greatly help with security, okay? So if you build a data center and your hypervisor, for example, your hypervisor is going to run on the same CPU that hosts the user applications, you have a security threat because the user can get access to the hypervisor and now can control the entire data center, okay? So in order to make it much better, you want to separate the infrastructure domain from the application domain. The CPU will host the users on the system, but you're going to run the hypervisor, for example, or other element of the infrastructure operating system on a different element, let's say, completely separate from those applications where the application is running. This is where the DPU plays a role. So we're running -- the DPU is being used in order to run the data center operating system to provision the servers, to do the secure access to the user that's coming into the data center and so forth. So DPU is the north -- what we call north-south kind of the access network. And then SuperNIC and the switches or Spectrum-X and InfiniBand are part of the scale-out infrastructure, kind of the -- some people call it back-end network or some of them call it compute network or compute infrastructure. And then you have also the scale-up where there is another element of NVLink.

Okay, great. So the DPU also it's kind of freeing up the CPU for doing cycles of what it's good at and the DPU. So it's -- yes, that's -- maybe if we switch over and talk a little bit about the scale-up network, the NVLink. You have NVLink and NVSwitch. There's other topologies out there. I guess what's the advantage of NVLink over there's UALink and even Broadcom on their earnings call, they've been talking about just using Ethernet as the scale-up. Can you kind of give the gives and takes of each one?

Well, I can definitely talk about what we do. So first, scale-up is not easy to do. It's very not easy to do. It needs to take those GPU ASICs and make them one, okay? It needs to form like 1 unit out of a lot of ASICs together. And therefore, it's not just the huge amount of bandwidth that need to run between them. It's you need to have a very high message rate that everyone will -- all ASICs will connect and communicate together as like 1 unit, okay? You need to have a very low latency between them. It's a very tight network. And because of that, we are trying to put everything in a rack, okay? So we can use, for example, copper for that connectivity because copper, first, it consumes zero power. And because of the huge amount of bandwidth, if you would do it on something else, it's going to be a good amount of power being consumed there. You want to make it very resilient and so forth. So we want to maximize copper. That's why we want to put everything in a rack, in closed rack. And this is where density becomes an interesting element to deal with, and this is where we bring liquid computing into the game because we want to pack everything to increase the density so you can maximize copper and build like 1 unit, okay? So there is a good amount of complexity in actually building an NVLink element. Now one thing that isn't as obvious that NVLink brings is it's working, okay? NVLink, it's in the fifth generation. And essentially what made InfiniBand so great is because it had many generations in it, right? It continue to evolve and it continue to be better and better and better. And that's what makes Spectrum-X so great because we took all the 25 years of InfiniBand and put it on Ethernet, okay? So putting an idea of a network and says, okay, one, the first shot, my first shot is going to make it so great. In reality, it's not the case, okay? So this is a complicated element. There is a huge amount -- just think about NVLink72 is like 130 terabytes per second, okay, in a single rack. It's like the entire peak Ethernet traffic is just running in a single rack, okay? This is what you need to support in that set. So it's fifth generation. It continued to evolve over the years, connecting more and more GPUs. We brought SHARP into NVLink. Actually, there is compute engines. There's compute algorithms running on that NVLink when you're running everything together. So this is kind of NVLink. I tried to give you a little bit on the complexity of it. And obviously, having the fifth generation, it just show that you evolve from GPU to GPU and you bring more elements, more capability. You need to adjust to the workloads, okay? I'm not sure that I mentioned it, but the reason that we're annual cadence on the infrastructure, not just on the compute is because the element that you need to bring into the infrastructure, including those data algorithm that are being added from generation to generation because the workloads are different, because the workloads are being modified. And as the workload is being modified, the compute algorithms need to be modified, and that's impact what you put on the infrastructure, which include NVLink and the rest. So this is where the cadence and being robust, it's working, it's amazing technology. It's liquid cooled, it's the dense, fully copper, and that's what make NVLink, NVLink.

Great. And you also announced NVLink Fusion. So you opened it up that it's not a closed network. Can you talk about the advantages of NVLink Fusion?

Yes. So once you get a sense of how complicated scale-up is, and people might say, "No, it's easy." No, it's not easy. There is a huge amount of complexity in it. Then why not to help our customers that wants to build their own custom accelerators, for example, leverage what we invested for years building the best scale-up infrastructure with the liquid cooling, with the density, with all the aspects of that, the performance of that, why wouldn't we let our customers leverage that huge amount of investment and making it easier for them to take those accelerators that they build, those custom XPU that they build, the custom accelerators actually leverage our infrastructure to build a solution for them, okay? We design a data center. We design it as a whole. And then you can take pieces off it. And you can take the GPU, you can take the CPU, you can take them both together. You can also take the infrastructure if you want to, okay? So this is where we build or working with ecosystem, which includes MediaTek and Marvell and Alchip Technologies and Astera Labs, for example, and CPU suppliers like Fujitsu and Qualcomm and working with them so they can leverage what we do. But infrastructure, we started this talk with saying the infrastructure becomes a key element. And essentially by having NVLink Fusion, we enabled that key element to be used by people that needs or wants or require to build their own accelerators and now they can leverage what we did, what we designed and actually get great data center for their own custom elements.

And maybe just to understand that, is it Fusion link like you mentioned Qualcomm, just use them as an example, if their CPU wants to connect, do they pay a license for the NVLink? Or do they just start using your NVSwitch?

I think that there is element of NVLink that they will need to connect to. So essentially, they need to get the interfaces and they need to get, for example, an NVLink [indiscernible] that the CPU can connect to it. And once they have that, they connect to their NVLink Switch so they can acquire the NVLink Switch, and they can acquire the entire elements that also come there with the liquid cooling and all the stuff. So they are taking elements from us. They're taking the API from us. And obviously, we work with them, and they can build their own system.

Okay. And we are getting lots of questions from the audience. But the people, you've introduced the silicon photonic switches at GTC. People are asking NVLink, when does that go fiber? And also when we talk about scale-out, is that already fiber? Or what happens with silicon photonics tied into all of this?

Yes. So different elements here. So first on the scale-up, let's put it like that. Copper is the best connectivity. Copper is the best connectivity, zero power. It doesn't consume power. It's very reliable, okay, and it's very cost effective. So you would like to use copper as much as you can, as much as you can, for any connectivity that you can. And therefore, we are trying to put as much as compute density in a rack because within that rack, we can use copper. And that's why we're investing a lot, right, to increase the amount of compute in the rack. So we can use copper and run that because there is nothing better than copper. Now when you go to the scale-out, this is where you're talking about distances, right? Because now we have racks that needs to be connected and you're out of the reach of copper, and you need to go and use optics and you need to use optical connections. Now in traditional data centers, the amount of connectivity between rack was very, very -- between racks was very, very small, okay? So there is no much optic transceivers or optical connections that were there. When we look on an AI factory, every GPU has a NIC out, right? So if we look on Blackwell, every Blackwell has an 800-gig NIC that goes out. So the scale-out infrastructure, actually, there is a good amount of optic connectivity. We need to use around 6 transceivers for every GPU. So if you build 100,000 GPU data center, it's like 600,000 transceivers. And now the power that's associated with the optical network becomes something that can consume up to like 10% of compute. So if I'm building 100,000 GPUs and I can add another 10,000, that's not a small number, okay? So now the power becomes something that you want to look how to improve it. And we all know that the limiting element in building data center is power, right? It's not really space, it's actually power. So as much as you can save power and you can redirect it to compute engines, that's a great thing to do. The second thing is that data center increases in size, and it go fast, right? It's like 2 weeks ago, we talked about like 16,000 GPUs there, large data centers. Now you're talking about hundreds of thousands of GPUs. So 100,000 GPUs, 600,000 transceivers, and it takes time to install that and it takes time to manage that and you might need to replace elements. There is so many components that you need to deal with, okay? So this is kind of -- this is the right time for improving optical network for the scale-up. And the way to improve that is to introduce co-packaged silicon photonics, right? And co-packaged silicon photonics, what that means? It means that instead of having the optical engines in every transceiver, I will take those optical engines and put that next to the switch and package it together with the switch. Now what did they do here? First, I reduce distances, okay? So if the optical engines in a transceiver, it needs to go the distance through the transceiver, the cage, the PCB, the substrate to go to switch, reduce the distance and with that distance, I reduce the power. So now on the same ISO power, I can put 3x more GPUs. On the same ISO power of the network, I can connect 3x more GPUs. That's huge. Now I'm reducing transceivers, okay? Now I have 1 transceiver per GPU, not 6. Think about how many elements you reduce from the data center, which means it's not just I increase the resiliency of the data center because now there is less elements, I also reduce the time to operation. I can build the data center much faster. So CPU brings such a greatness element, and we started with the scale-out because, again, it's like 10% of compute power. I can increase that number, it's huge, to reduce number of components. There is a huge amount of benefits bringing co-packaged optics into the scale-up infrastructure. Now on the scale-up, as I can use copper, I'm going to continue to use copper, okay? So we increase the density with copper because there is nothing better than copper. As long as you can use copper, we use copper. Okay. So this is where we continue to use copper. We announced that we are having 576 GPUs on copper, NVLink, scale-out, it's multi-racks, distance optics, this is where co-packaged optics will be a great thing.

Do you have an idea if we can get to 576 with copper? When do we have to cut over to optics?

It's a good question. Over the years, there was always people saying, "Oh, this is going to be the last generation of that." It will be the last generation of that, right? It's like every time when people say it's going to be the last generation, apparently, there is another one. So as long as we can pack, we'll pack.

Okay, good. Great. I'll see if I can -- let's see if there's a question we haven't covered yet. If you can answer this. If you're winning in the market, are you displacing what Marvell and Broadcom solutions are or solution providers like Coherent, are you replacing their designs?

The first answer is not really, okay? The reason is the following. First, there is many infrastructure in the data center, and there is many areas that requires a need to use transceivers, okay? So on the scale-out infrastructure, we're going to introduce co-packaged optics. North-south network, for example, require transceivers. We put transceivers on NIC and so forth. And since the data centers are growing and the market is growing, there is enough for everyone. And therefore, we're not replacing anything, but there is different infrastructure and there's infrastructure areas that require transceivers. That's one thing. The second thing, we are working with that ecosystem of partners, and they are part of our CPO infrastructure. So they are contributing into what we're doing on CPO and they're bringing our elements, and we're working with the ecosystem. For example, we announced working with TSMC on packaging, but we're working with a lot of vendors that you mentioned on lasers and optical arrays and the different elements that we need for connectivity. So they are contributing to our CPO infrastructure as well, they have more or a good amount of transceivers to continue and support. And data center is growing. AI is going everywhere. There is enough for everyone.

Great. I think we're out of time here now. So I would say in summary, you've got the scale-up and scale-out networks and you're tying together hundreds of thousands of GPUs to act as one big GPU. And if people want to come into the NVIDIA network, they're open to do that. You think you've got the right solutions. So maybe if you want to give a closing remark also.

Yes. I think you had short questions and I had long answers, and I'm sorry for that. In the past, people data center budget was focused on, let's buy as many servers as we can. And then we have something left, we may connect them to get there. If something left after that, we may do some storage and stuff like that. I think now folks realize that the infrastructure is key, okay? It's not just network elements, buying a NIC and buying a switch, no. You're buying a spaceship, okay? You're buying a supercomputer. You're buying something that requires the kind of to be fully synchronized with the data center, and that infrastructure will determine what data center will do, okay? That infrastructure will determine if those compute engines are just a server farm or that's an AI supercomputer for training or inferencing, okay? So it's a key element. Its importance will continue to increase, and we'll see innovative technologies coming into the infrastructure. So it's something that keep us exciting, keep us exciting. Yes, so this is where the infrastructure is. I think more people are interested in learning about it, and I'm happy that we were able to talk today. And I hope that we provided people with more or better understanding about the infrastructure that we built.

Yes. That's great. Thank you. Thanks, Stewart. Thank you, Gilad. Thank you very much.

Thanks, Kevin. Thanks, everyone.