Arista Networks, Inc. (ANET) Earnings Call Transcript & Summary

Thanks, everybody, for sticking around that later in the day for this conference. I hope you've enjoyed all this session, and that's my pleasure to welcome Jayshree Ullal, the CEO of Arista Networks. And I think I've decided I'm not going to present Jayshree because everybody knows her. She has been the CEO of Arista Networks for 10 years and you all know as a role, Arista has played in technology infrastructure, in data center networking and much more than that over this tenure. So Jayshree, thank you so much for making the time to close with us this conference.

We spent like the whole day talking about generative AI, infrastructure requires or generative AI, and I just wanted to pick one claim that was made before you joined, which is -- and I won't give you -- I won't tell you who that was, but someone said, well, with a 25 terabyte -- terabit switch, you basically can manage 500 servers. And with the exact same switch, you can manage 7 GPUs. So I thought that was like an encouraging and almost like frightening thoughts to ask you this, maybe this first question about AI. So like, is that true? Is networking as important as that? Is the share of networking as important as that in the infrastructure required for generative AI?

Well, let's see, 8 GPUs and 25 terabits, that means each GPU is generating 3 terabits good grief. I remember a time when we talked about networking in megabits and gigabits, and we were still looking for a use case for anything in the terabits. But I think AI has indeed become that killer application, and it is great for Arista to have an opportunity to participate not only in the front-end cloud network, but now what this new evolved back-end AI network will become. So it's not too far from the truth whoever gave you that source. And I think if you would look back this time last year up here, when you had me on the AI conference, we were still looking for that killer application for AI. And yes, we were talking about GPUs, et cetera. But I think these AI workloads, whether they are collective operations were all reduced or all are dominating the world because these GPUs are data-intensive, compute-intensive and network intensive. Every cycle is a compute, exchange, reduce and then the cycle continues again. And the initiation of this training depends on the suite of GPUs, which even last November until the whole ChatGPT, OpenAI came along, we thought 1 billion parameters was a lot. And today, you can see with ChatGPT, you can record up to 175 billion parameters and in GPT-4 models is expected to be 1 trillion parameters. So clearly, we're pushing the envelope of compute processing as well as the network in a way I never imagined, and we haven't done in the last 3 decades of my networking history.

That's very exciting times. And I think the first question I'd like to -- I mean like after that introduction that I'd like to not to go back to a memory I have of -- I think it was in 2021 at your last CMD, Andy, your founder and Hugh really talked about AI at the time, nobody was talking about it, and they were sharing very, as they always do, very complicated charts explaining how basically it was kind of nice to line up GPUs to manage a very large model, but these GPUs were ending up like waiting for data, 33% of the time. So they explained very well, AI clusters need very, very dense network, very high-capacity network. And so lossless network. You can't basically in the workload, you can't close or you can't afford to loss what you exchanged between GPUs, they explain how they had to avoid collisions and how your technology was very well positioned for that. So my question to you is, well, did you know at the time, if anybody was capable of talking about that, does that mean you have already in 2021, anticipating like the kind of insane scale of deployment that we are seeing ramping today?

I think the truthful answer is no. We didn't imagine the magnitude and scale of AI would take off so fast. But of course, a lot of this growth did start in 2021 back in our labs with Andy Bechtolsheim and Hugh Holbrook as we were starting to talk to customers and reimagining the back-end network. And again, going back to the back-end network, it's not something Arista historically participated in because it was typically a PCIe or some kind of a CXL or NVMe IO, connecting to high-performance compute, servers and storage. But the concept of envisioning a network and going beyond a cluster of servers and building large GPU clusters really started emerging, I would say, in '22 and '23, where we're now finding ourselves in the middle of pilots. And we've had to think of basically 3 things. How do we make our platforms capable of scaling to these number of GPUs with predictable scale, performance, latency and like you rightly pointed out, a lossless fabric? And a lossless fabric doesn't just come from making Ethernet lossless, but it comes from having the right congestion control mechanisms because you've got all these GPUs and traffic coming from the center to the receiver to the uplinks. And as you know, AI and all the AI traffic and performance demands are very different. It comprises of a small number of synchronized high-bandwidth flows that make them very prone to collisions. So we were working a lot on the congestion control algorithms because the last GPU can be the culprit and that determines your entire job completion time. And these GPUs, as you know, are not cheap. They're very expensive. The last thing any of us want is for them to be idling 30% of the time. Imagine a $1 million or $1 billion GPU cluster idling 30% is basically like losing $300 million out of your $1 billion cluster. So a good network can make all your GPUs hum. And this is why we believe the NVIDIA GPUs, which are clearly top of the rung and the only one in the market that's dominating requires a fine network and a scheduled network and a fabric that can deal with that. So I think Andy and Hugh are very much working on it since 2021, but the realization of its scale, its magnitude, really, I think, took off last November with ChatGPT and OpenAI. That's when we really saw the sudden exponential curve.

Thanks, Jayshree. And so if we fast forward until today, the least we can say is that the AI cluster's deployment accelerated sharply. And if we try to put numbers on the size of the AI networking market just a couple of years from now, so NVIDIA compute's revenues are at nearly triple this year and nearly double next year, so that will exceed $60 billion. That would correspond to say 3 million, 4 million A100 and H100 GPU unit. So probably around $100 billion in AI servers. And so that would correspond to about maybe $20 billion in AI switches. Is that the kind of scale that you see out there just in 2024?

Oh, wow. Well, I think that's a bit high for us anyway. Maybe that's a good number for NVIDIA because obviously, they are the early participant and winner in the GPU clusters. So maybe a good way to look at this is the networking component is typically 10% to 15% of a GPU cluster, not quite 20, although it may be when you add up all the optics and power requirements. But there's 2 halves to the networking requirements. One is the NIC or the DPU as some folks call it, that connects in all of your service and it's a network interface to the outside world and the other is the switch, which could be InfiniBand or Ethernet switch. So if you look at the market today, we're still a little bit Arista is anyway outside looking in because a lot of these GPUs are bundled in with InfiniBand and sort of selling as a system solution, whether it's a DGX, et cetera. But if I fast forward, and these are 650 market reports, I'd say today, it's probably 2/3 InfiniBand and only 1/3 Ethernet. But 3 years down, it's going to reverse. And particularly with the Ultra Ethernet Consortium coming in with a large ecosystem of familiar tools, troubleshooting, Ethernet, as you know, is deployed not in the millions, but billions of ports. I really believe then we're going to see a market that favors Arista's TAM more than it currently does where it's much more bundled in a vertical stack approach with the GPUs.

Let's go maybe a bit deeper into this InfiniBand versus Ethernet. So if I understand you correctly, when you're very close to the GPU inside the cluster, you tend to have InfiniBand today, and that's about 2/3 of the market. And you guys view Ethernet are one layer behind that. Can you explain us why InfiniBand is a better fit closer to the clusters and why Ethernet becomes a better fit more outside? And what are the evolution in Ethernet that are going to make that switch?

There you go. Absolutely. First, I want to go on record and say neither technology is perfect right now for AI. If you look at InfiniBand, it was conceived in an era 20, 25 years ago, where it was the best technology for high-performance compute, often known as HPC, right, where the #1 metric was nothing to do with AI, but really ultra-low latency and you really look at your megaflop utilization of your back then CPUs and now GPUs. Now I think you can recast InfiniBand for some use cases of AI. But I think there are a number of things InfiniBand is deficient in and a number of things Ethernet needs to improve to become AI. So I think they both have some challenges. Just staying on the InfiniBand, it's a layer 2 technology, it's subnet based. It's technically and theoretically a vendor of one defined by the InfiniBand Trade Association, but really Mellanox NVIDIA is the main provider. So when a customer is looking for a diverse ecosystem, they're looking for certainly more than one option. It's a layer 2 technology. And those of us who have been through networking know the whole route when you mask switch when you can is so important because if you're just based on layer 2 and you're limited to a subnet topology of say, 10,000 or 48,000 subnets, by the way, every subnet is a port. So if I have 2 connections to my NIC, 2 to my GPU, 2 to storage, they all count, that means you've really limited your cluster ratings for InfiniBand in a significant way. High availability is a little ill-defined there. It's, again, not a routable protocol. So you're based on a controller or server architecture. So HA means if something fails, it has to recover. Of course, if it recovers in 30 seconds, it's okay, but if it recovers in a few minutes, that can be very mission-critical. So for all these reasons, I think this is why the Ultra Ethernet Consortium came to being because in Ethernet, and I never will in my career bet on an either not technology as Bob Metcalfe would say. I've done my share of fiber channel, token ring, InfiniBand, FDDI. And ultimately, what Ethernet does and does so well is marry its familiarity with all the additional features and capabilities. And that's what the UEC is working on right now. We're working on a number of congestion algorithms to deal with the per-priority flow control. We're combining that Arista with the right virtual output queuing fabric so that as the centers send a lot of traffic from the GPUs, you can appropriately prioritize them, schedule them in a fabric and send them out with the right buffer, right memory, et cetera. Proper packet spring and load balancing becomes very, very critical because, again, AI traffic is very slow intensive. And what you're dealing with a lot of times is small chunks of large flows. But again, if the flows don't get there in time, you're hung up. So I think Ethernet is the winner because we're going to be working on all these load balancing and visibility and monitoring and congestion control mechanisms that naturally sit on top of today's Ethernet specification, that's standard space. The other reason I'm a big fan is we can talk about building all the back-end clusters, but there shouldn't be silos. You want the back-end to work with the front-end because ultimately, this AI traffic will flow on to the rest of the network. And the last thing you want is an InfiniBand to Ethernet gateway that will slow you down so much because then you lose all your latency advantages of either technology. So the predictability of the performance, the seamlessness of connecting your machines for compute and storage to your general purpose front-end cloud network, I think, is all the reasons why Ethernet is poised to win. But today, InfiniBand absolutely connects most of it with GPUs from the same vendor.

And data sensor networks rely on pluggable optics technology today, but we hear more and more the supply chain talk about Linear Drive optics and co-package. And we even discussed with a company called [indiscernible], an all optical fabric, which would allow to interconnect thousands of CPUs, GPUs and memory chips so that they would behave as a single logical virtual machine similarly to what is done today in Google's TPU clusters. And I also recall that you explained very well in the conversation with Pierre last year that [indiscernible] had a number of issues for traditional networks building up flexibility and reliability for performance. But I wonder whether AI could be changing the equation. What are the benefits of Linear Drive uptick versus DSP uptick? Do you see co-packaged optics adopted at scale for AI Networks? And would that impact -- how would that impact the optic supply chain?

Right. First of all, we are big fans of all kinds of optics ServiceDesk is not an optical manufacturer, but we believe we have to be an important enabler. I'll kind of parse that question from the last year, I answered it to now where it's not so much that co-packaged optics couldn't work. As you have these dense GPU clusters, the troubleshooting of co-packaged optics, it becomes a real challenge. And this is why Arista launched in OFC earlier this year, the capability of Linear Drive. Just a quick education on Linear Drive. Linear Drive is the ability to have the electrical service drive your optical for longer distances without using DSP technology. Now DSP is definitely required digital signal processing when you're going in long-haul distances of 100 kilometers. But majority of the AI clusters, as you well know, are within a data center or even within a short haul where you don't need optics, you may only need coax cable or shorter distances. So the combination of these non-optical cable options, combined with Linear Drive now can make your capability of taking AI traffic a lot longer, a lot farther in a much more cost-effective fashion and not have to worry about all these troubleshooting techniques. Why did we come up with Linear Drive? If we didn't do Linear Drive, the 800-gig switches, when they connect to the optics, would be -- the optics would be 60% of the power. And this would be a huge problem. As you know, there's GPUs and all these other things that already might start needing liquid cooling and immersion, but we were very concerned with the power envelope. And by bringing in Linear Drive at higher speeds, we're now having the electrical characteristics of the switch drive the optical distances without using the same power, and we're able to reduce the power by at least 50% and make it a smaller envelope of the total switch cost. So I think Linear Drive is very exciting, not only for price performance and reliability, but also for power reduction.

Thanks, Jayshree. Let me maybe come back to what you described with the market shifting like [indiscernible] Ethernet and its role you described at the Ultra Ethernet Consortium. So it's really like you describe an industry effort to get the standout to address the very specific needs of AI clusters. And so my question would be, in that framework, how does Arista fair? You will have competitors and even like the main player in InfiniBand today is actually very active developing Ethernet technology as well. So how do you go in that -- where do you describe so well and that seems to be so almost like I would say, deterministic, how does Arista differentiate and add value -- unique value to AI clusters?

Yes. No, I think you bring up -- first of all, I think you know, Pierre and Antoine, we are very used to competition, and we welcome competition, and we've always had to prove to the world that we are better from an innovation technology and architecture point of view. And I think the role of Arista will be to bring the union of the best-of-breed chips that will have this UEC capability, the best-of-breed platforms that we will bring to bear either with the 7800 AI Spine or a 2-tiered architecture, if you're starting to go beyond 1,000 GPUs to 4,000 to 10,000 GPUs maybe for generative AI and inference applications as well as to, over time, a distributed scheduled architecture where you need to go to over 10,000 GPUs. So we will be building different use cases and platforms. And don't forget our software stack. Anybody can build a commodity Ethernet something. But very few have invested like we have over a decade or more in building that software stack with that published subscribed model and we will be adding AI features to that because, as I said, AI is very, very performance and visibility and traffic demanding. I'll give you a couple of examples on how we'll be differentiating. For example, the UEC form is defining packet spring all the way from the [ net ] to the switch. Arista is already working on this kind of UEC compatible capabilities by developing dynamic load balancing algorithms where you can take all that traffic coming from the ingress ports, host ports, and go through the uplinks and go through the receiver ports and rebalance all of these links, whether it's based on different hashing mechanisms, UDF, QPair IDs, our customers can now pick and choose the packet header fields. And these are now capable of better entropy, better efficient load balancing and efficacy of their AI workloads. So this is something we've been working on in the lab since 2021 that's coming out. Another area is AI network visibility and monitoring. When these packets and flows are traveling at record speed in the training phase of these large LLMs, the accuracy of these LLMs is super important. So in addition to features we've already had in the U.S. like Latency Analyzer, Arista's new AI Analyzer feature monitors and reports traffic counters. And these have to be done at micro second levels because you've got to be able to detect the microbursts, [ not at millis ]. So it's very helpful to detect and address microbursts that are difficult to catch in normal traffic because they're going at AI speeds. So congestion control, load balancing, increased visibility and an array of platforms are examples of how we'll be leveraging Ethernet and don't forget IP to make this possible.

This is Pierre. And maybe as a quick follow-up on that. So Ethernet still represents like a minimal portion of the NVIDIA's networking revenue today, but NVIDIA is doubling down on its Ethernet road map with Spectrum-4. And so do you have any views on how this is going to impact the Ethernet switching landscape for AI and more generally?

Yes. Listen, I have a lot of respect for NVIDIA, the leadership team, Jensen. Jensen and I were in AMD back when he was doing graphics and I was doing networking chips as engineers. So I'm sure whatever they do, they will do very well. That being said, as you described their market, they are a market leader in GPUs and AI and InfiniBand. And they will continue to work with, I think, best-of-breed Ethernet platforms and whether or not they're developing spectrum, customers will choose the best-of-breed platforms, and we fully believe that we will be mostly a friend to NVIDIA and the GPUs and not a competitor. And even when we have to compete, we will compete on the merit of technology and innovation.

Jayshree, so a year ago, before this AI wave who basically took a everybody by surprise, even if some of us were preparing for it in a lab, you hit like a record year with your cloud customers with revenues more than doubling year-on-year to, if I recall correctly, more than $2 billion, and that's basically only a couple of clients. And that was more, I think, like combined revenues in 2020 and 2021. And so today, we are in this situation for Arista in the networking market, like the cloud networking market that we are coming off probably a peak in licensing all general-purpose architecture while AI spending is picking up. So my first question on that front would be, can you tell us about that massive ramp in 2022? What was it about? What was happening at your clients? What were they deploying that required so much networking equipment?

To appreciate 2022, you have to also have to go back to between 2016 and '18. We do go through these vicious cycles of investment from our cloud titans, as we call them. And where they were investing in 100-gig migration at that time and building a lot of centers across all geographies, and we experienced a wild swing up. And when we're doing well, nobody asks why you're doing well, but of course, all parties come to an end and then we saw some slowdown during the COVID time frame in 2019 and '20, as you might recall, where especially one of our cloud titans skipped a service cycle and therefore, skipped deploying more networking and switching. Now whenever I always say, when you go on a diet, you eventually get hungry and then you go from famine to feast. And I think that that's a little bit of what we experienced last year and were -- due to COVID and due to the intense expansion that our customers did between 2016 and '18, they took a little bit of break in '19 and '20, and then they came back roaring after the pandemic and realized that they had to make a lot of investments. That was part of the reason we saw the big jump. The other reason we saw the big jump is that you have to have -- been an ostrich in the sand to not notice the supply chain crisis. So many of them were planning ahead as was Arista, where we were making multi-year purchases. So I think the demand for '22 and '23 even will normalize, and we don't expect the same rate of growth in '24, as I've often said in my earnings. But if you look at us at a 3-year CAGR over any 3 years, we've still done very well and had a CAGR of 20%, maybe going future, it will be lower. And that's the way to always look at Arista, not to look at our sinusoidal waves that go up and down, but to normalize it to what I would call a nice respectable double-digit growth and more importantly, a nice respectable market share gain. So today, at a point where people were concerned, okay, we won the 100 gigabit, and we have 40% market share. Are we going to do the same on 400, are we going to lose? I feel like Arista is very lucky to have customers who will embrace us in 100 gig with the enterprise customers, who will embrace us in 400 gig as these front-end cloud networks start to upgrade from 100 to 200 to 400 gig. And of course, with the AI cycle, I expect some of the pilots and trials to be 400, but many will move to 800 as they go into production. So we're in a very fortunate phase where we've got 3 migration cycles. Hopefully, all of them won't hit a low at the same point or a high at the same point because then we'll have supply issues again. But the inner leaving of these 3 migration cycles will be a fantastic combination for us.

And [ I liked the analogy ] of feast and famine and I can't help asking you because we are in a very specific situation now where there is like a cooling off of spending in the infrastructure, while we are spending on AI clusters. And should we expect like the general purpose infrastructure to come back up again after a period of -- [ I mean, for a period of this ]? Or do you think we see like a structural shift of spending away from the general purpose infrastructure towards AI on a multiyear period as you described it, as you like to think about it?

Yes. I think with the current cloud titans, what I'm seeing in a general trend is that they're pivoting today more to AI, all of them, right, varying levels. But at the same time, most of their business is coming from the cloud. So they can't stop spending on the cloud, but they can certainly swing the pendulum a little bit more than a swap their assets on the cloud, which is what I think we'll see. If you go a couple of years like this, you'll have another feast, famine problem because you'll have so much AI on the back end and you would have built all these GPUs and clusters, you go wait a minute, I need to have AI as a service over the cloud, what do I have to do, I have to connect them. Well, if I connect them, what does that mean? I might be adding 20%, 30%, 40% more traffic on the front end. If I am adding more traffic, what do I do? Oh, gosh, I need a refresh cycle. So I think at the moment, it is very much an AI show and everything is pivoting to AI, and you would think there's nothing happening in the cloud. But the reality is, as we do this pivot, it's going to have a direct proportional effect in the cloud in the latter years, maybe it's north of 25%, but nevertheless, it will happen.

That's very helpful.

Let's maybe now look out beyond 2024. So in a world in which people spend over $100 billion on AI servers and switches in 2024 and still have to spend to upgrade their traditional data centers, I think the key question now becomes what happens next. Are we headed towards a generative AI bubble burst? Or on the other hand, one could argue that a lot of the clusters required in 2024 will be used for training. So there could be a wave of spending on inference infrastructure on the back of that. So could that actually result in sustained spending on AI in 2025 and even in the second half of the decade?

Yes. So to answer your question, I'm going to first step back and say there are actually 3 types of AI networks we see. The first is super small ones. I don't even know if you call it training or inference. They're just within a small server cluster, maybe just 100 GPUs and sometimes, Arista doesn't even see it because you could just be connecting it in with a CXL or PCIe or NVLink back end and they're doing generative AI applications too on a much smaller pod level. And that's one that I think will continue and especially as the generic AI explosion of applications happens and more and more apps are coming on, that will happen. Then there's a second use case, which you rightly pointed out, in my view, the inference use case, and I think both the inference use case and the training use case lend themselves beautifully to Ethernet. It's not so much that I think one technology is better for the other. But the size of the GPU cluster will vary for those 2. So in my view, the GPU clusters for a generative AI inference application is much more distributed, smaller in size. And you could probably do a lot, a lot of capability with just 1,000 to 5,000 -- well under 8,000, 10,000 GPUs. So you're absolutely right. I think you're going to see a set of clusters that are inference-only that don't need the billion or trillions of parameters we talk for large language modeling and quantization and et cetera. And then there's -- what the world is doing today. Today, majority of the focus is, let's prove the large language models. Let's build large GPU clusters. And they range anywhere from 1,000 to 8,000 today. And I think if the GPUs were available, they'd be north of 32,000. So today, we're more constrained by the availability of GPUs than the training models our customers want to build. And they want to prove that if they can do the high end of accelerated computing and build that mainframe like AI with the large language models that can support a trillion parameters, then it's a lot easier to go to the 2 other use cases I just mentioned. So I think you'll start to see -- and again, I don't think this is thousands of customers. For Arista, it will be single-digit customers in the near term. But every one of them is looking at these large language models and how to offer AI as a service to their customers, whether it's cloud titans, Tier 2 cloud providers and even some of the very large enterprises. And they're putting in the CapEx to prove the highest level of training. And then once they do that, which is going to be, as I described, pilots and trials in '24 and -- '23 and '24 and really large productions in '25, I do believe the world will swing the pendulum a little and come to the smaller clusters as well. So I think the AI wave will continue, but there will be different use cases.

Jayshree, we're getting to the end of the time we had. So let me ask you maybe one last question, and we always like to hear you talk about like 3-year failures and 3 years old, but what about a decade from now? So if you were like to -- yes, difficult call because who would have guessed we would be where we are today 10 years ago. But how do you think like where is the current trend getting us in terms of what likely the data center infrastructure of the world looks like in 5 to 10 years from now in terms of scale, in terms of AI versus traditional backbone infrastructure and in terms of technology deployment and all the silicon photonics innovation, how much will have been deployed in which part of the infrastructure versus like the most additional Ethernet networking?

Yes. Well, look, it's been hard to predict the next 3 years, let alone the 10. But if I sort of look broadly at the vision, I think quantum computing, quantum capabilities and networking in optics now has the killer applications to drive it. Before it was just file, print, share and some basic SaaS applications. And so none of us could see bandwidth really jumping by leaps and bounds. Now we can. So in 10 years, I can absolutely see terabit Ethernet in the horizon. I can see that today's gigawatt centers become zettawatt centers. But more importantly, I think just as it will go up in scale, it will also go down and miniaturize. And the ability to go up and down and have data centers that are super intelligent and capable and actually become centers of data, whether it's in a branch, in a campus or in a core, that's been Arista's vision. The ability to scale that up and down will be critical. So clearly, we always look at the high end. But as we develop the high end, that same technology will happen in smaller levels or smaller clusters across different use cases in the network as well. And I think some of these things that we say is hard today like all optical switching, liquid cooling, immersion of GPUs in cooling environments, almost an aquarium of GPUs, doesn't seem that far away. It seems -- I conclude by saying artificial intelligence is becoming real intelligence.

Jayshree, thank you so much for being with us and for closing the day with us. It's always a pleasure to catch up with you. I hope we find another excuse to have you around next year. And I'm sure there will be like a long list of topics to discuss with you. So thank you very much.

I look forward to that. Pierre and Antoine, it's always my pleasure. Thank you for having me back a second time. It's been AI last year and AI this year, and let's see what it is next year. Thank you.

Thank you. Thank you, everybody, for making the time to participate to our conference, and I'm going to immediately put up the notes I took in room time and give some concluding remarks. So we've had a great day. I think early in the day, we really laid out like a very interesting framework where we are in a world in which model size, compute requirements are very clearly on a trend to grow 10x every year, and there is no end of sight to that. It's the size of the models. It's the pace at which services are adopted, the pace at which they are going to be deployed and the pace at which models are being specialized on very specific use cases and sometimes down to very specific single use. At the same time -- so we have models going 10x. The technology is being adopted at a pace that has never been seen in the history of humanity. So 5 days to 1 million users for Chat GPT. The last record was 75 days for the iPhone. And before that, it was like several months for Instagram. So there is also an element of pace of adoption that is frightening. But I think from that, my conclusion is we're definitely going to have to get the best of both worlds. We will have a lot of innovation because there is no way the technology of today can address such a fast-growing need for more compute and for more technology infrastructure. So we will have a lot of additional innovation. I would expect the pace of innovation to accelerate because demand is there. We cannot manage that demand without innovation. And at the same time, I think innovation is never going to be enough to address this demand. We'll also have to spend more. And I am confident that the contribution to value creation to GDP growth of AI will make ample room for always increasing levels of spending on the technology infrastructure. So that's very, very encouraging. And it makes me feel very bullish about our industry going so up. Then the second thing I wanted to share with you is we really had a very good overview of what the enablers of this AI revolutions are from an infrastructure perspective. So I think we had a huge focus on optics. Some of it was by design. We had a couple of people contributing to the day, specifically on optics, but even someone talking about hardware engineering ended up teaching the fact that optics was going everywhere and was the most important enabler of the AI revolution. So at every level of the infrastructure, there is a lot of optics innovation, optics disruption and things fascinating to watch. The second one is the networking. You heard Jayshree just now. There is no good technology -- good network technology for AI today, neither InfiniBand nor NVIDIA are good. And so a lot of innovation is in progress. Jayshree and her teams have been in the lab since 2021 working on AI clusters. NVIDIA is working on evolving InfiniBand to [ Ethernet ] as well. Going to be a very interesting space to watch. And of course, there is this very interesting statistic that came in from Marvell shared that 25-terabit switch that will cover 500 traditional servers is required for every 7 GPUs -- so for not even every GPU server. So it's going to be very interesting to see that. And then, of course, last but not least, we've talked about chips, about customized chips. I think Kevin made a very strong case for like the growing ecosystem and infrastructure that Marvell and Broadcom and others are bringing to the industry to get everybody to be capable of developing more customized chips. And we've always believed in this study at New Street, we looked at NVIDIA like that 5 years ago. As compute grows faster than Moore's law the only way you can accelerate efficiency faster than Moore's law is by specializing your architectures. That's the history of the GPU. And we see around the corner that is interesting. What we see is that the feedback from [ Kevin ] is very clear, GPUs are very good in some instances, and GPUs are very good in some instances. That's the #1 thing I would take. The second one is GPU's exceptional feature is actually the fact that they are general purpose. They can accelerate a very, very broad variety of workloads. That's a very important feature in terms of staying power for GPUs. Whatever the innovation, we believe that GPUs will maintain that particularity of other accelerators. And then the last one that we noted is really this idea that, yes, the game between the TPU and GPU is interesting. There is a good competition. And yes, TPU v4 is better than the a1, but is not as good as H100. And we'll have to see whether the TPU v5 is going to reveal in the next few months. So the one thing that is certain that this competitive landscape is not going to stand still, it's not going to see like a massive shift, and it's going to be a competitive environment. And let me then bring like one last perspective in this concluding remarks. I think it was great this year to have like very rich content on the downstream of the industry like AI users. We had a fantastic session with Suchi from BCG and Arun from Intel and myself about like how to characterize like the take-up of use cases, and we had like an absolutely incredible group of people in biotech talking about what they do and how they are adopting AI. I thought it was very insightful. I pulled like 3 things from this conversation. So number one is as AI gets adopted AI, we have to remember that AI is a tool, AI is not replacing humans. Humans stay in the loop and AI is enhancing humans, making them more creative, making them more productive. So I do believe AI is much more of a driver of value creation and economic growth and good than job destruction as it is too easily pitched these days. The second one is that implementing AI in real use cases is very challenging, moving from having fun chatting with Chat GPT to making like a real copilot that adds value to a developer or to the way you manage your inbox is a nontrivial problem. And I think it's very important to focus on that and you can count on New Street for researching that in years to come. And lastly, in this challenging implementation and adoption of AI, I really think we are -- we've gone through like a very important catalyst in the last 12 months, in the last 9 months. which actually is a language model and conversational interfaces with AI. I do think that moving from an AI telling you, you have a 90% chance of failure to an AI having a conversation with you that what it is as a potential indication of failure is going to be a game changer in the way people adopt AI and use it. So very encouraging as well. And with that, I'll leave it there. And thank you again for participating. All sessions have been recorded, will be available at replay -- on replays. Feel free to reach out to them to share them around you, and we'll see you next year for our next Big Idea Conference.