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

Great. We're just going to go ahead and get started. So welcome, everybody, to the Arista Networks keynote fireside chat at the Goldman Sachs Communacopia & Technology and Internet Conference. I have the privilege of introducing Anshul Sadana, COO of Arista Networks. He's responsible for Arista's cloud networking operations and global customer advocacy, including sales and market development, product management, systems engineering and customer support. Anshul joined Arista in 2007 from Cisco. My name is Mike Ng. I cover Arista and com tech here at Goldman. We have about 35 minutes for today's presentation, inclusive of Q&A. So if you have any questions towards the end of the session, just raise your hand and we'll get a mic runner over to you. So first, Anshul, thank you so much for making yourself available and participating in our conference.

Thanks for having me over, Mike.

So to start things off, I'd like to talk a little bit about Arista's leadership in data center switching as well as the differentiation in the EOS platform. So could you talk a little bit about the evolution of EOS from a product into a platform, how that differentiates Arista's product against ODMs or white box solutions? And does Arista's ability to write directly to the networking ASIC play into any of this?

Absolutely. So Mike, many, many years ago, when we started, we were told by our competitors, this is the world of Cisco and the seven dwarfs. Welcome to being the eighth dwarf. And not very motivating when you're starting up as a company. But having been in the industry, we knew that the complexity or the issues that come in this space are largely because of software, not hardware. And understanding that is extremely important because many companies have tried to penetrate this space by trying to get there with the next-generation product 6 to 12 months ahead of competition at half the price because they're willing to sell it at low margin or zero margin and hoping that they can break in and get a lot of market share. But you're competing against a pretty mature industry. And selling on price or selling just on 6 months of advantage is not going to get you a lot of wins. We invested in the software stack from day 1. We have a state-based architecture. This has gone to a network-wide architecture with NetDL inside EOS. The way we write our software is highly modular. And any small bugs that show up -- because software has bugs. You have to prepare for that, not say I'll write perfect code and hence I'll never have any issues. You write these modular agents in such a way where if you have a bug, in the worst case, that agent just crashes and restarts. You don't drop a single packet. The data just keeps going through. Starting from that innovation to the programmability, integration and the rest of the stack, automation for enterprises with CloudVision, there's a lot of investments that have gone in. EOS today has about the entire stack, including our test infrastructure and so on, is over 25 million lines of code. That's the core IP we've built. This is not to undermine all the hardware work that we do. The system design is absolutely important, and maybe we'll get a chance to talk about it a little bit more as we go on. But the software side is critical to get it right. And in the end, I believe we've built a better product, better on quality, better on execution, better on integration, more open. And the market needed that. The market maybe didn't even realize the market needed that. But once you bring that option to the table, our biggest customers, be it the cloud, be it the Tier 2 cloud, be it the biggest financials, be it some retailers, be it some hospitals, they come in and give us that feedback every day. They truly value what we're doing with them along with this. And I think that's why we are where we are in our leadership position.

Great. It's a fantastic overview, and I do want to explore those in more detail throughout the conversation. But before we do that, let's talk a little bit about demand trends. Despite tech spending slowing down across much of the sector, Arista raised its revenue guidance to 30% plus growth this year in '23 versus 26% prior. Could you just help us think about how Arista is positioned to deliver that strong growth in 2023 despite some moderating CapEx trends for some of the largest cloud customers? Enterprise has been a notable area of strength. And I was wondering if you could expand on that a little bit and talk about whether that was more backlog fulfillment, new initiatives and enterprise to gain share. How would you best articulate the strength in enterprise?

Right. So last 2, 3 years have been absurd in terms of supply chain crisis, trying to manage that and where we are. Things are a little bit more stable now. That actually gives us more confidence as well that we can deliver. But this is not the increase from the 25% or so we were there pre-earnings call to now the 30% guide. It's largely because of strength in business. Cloud growth is moderating but doesn't mean it's going down or negative growth. It is just slowing down compared to the, what, 110% growth we had last year. And the business there is still very strong. The signals are strong from all of our big customers. Just that with very large numbers, the growth rate slows down. But it's steady state. There's no big surprises that we're expecting from them. The signals have been strong. We covered this in the last earnings call that lead times visibility is roughly about 2 quarters with many of them. That does give us confidence on where we need to be through this calendar year. And cloud, we're not as worried. Just the growth rate comes down when numbers are large. Enterprise, on the other hand, we are a share gainer on both sides of our enterprise business. And when I say that explicitly because many of you in the room have often talked about enterprise equals campus. For us, enterprise has a big data center part. And yes, the world is moving to the cloud, but on-prem is not going away. We have some very, very large customers deploying -- continue to deploy in colos that they run on their own. And the footprints are actually increasing in size, they're very efficient as well as the campus side where we are simply gaining share, right? We started from zero, we keep on gaining. And we talked about the product attributes on EOS. They apply equally to data center and campus. In the cloud world, when was the last time you logged in and you went to, let me do a search on Google and you get this message, "Dear customer, the cloud will be down from 2:00 a.m. to 4:00 a.m. Please come back tomorrow." That doesn't happen. But why do enterprises still have to schedule these outages and maintenance windows and whatnot? And they're getting pressure internally on, "Hey, if you're not competitive, I'll just go Amazon." The internal IT teams truly have to be competitive and keep the lights on, but also keep their business more competitive. We're able to help throughout all of that. The enterprises absolutely value that. Because of that, the enterprise growth has been strong, and we hinted to this for next year as well. We actually believe enterprise will grow faster than the company average even for next year. And good momentum, good signals. Maybe there's something macro events going on. But when you're gaining share in your early phases, you just don't see that because the sales teams are very opportunistic. If the deal has a low probability because some customers lowering their budget, the team will move on to the next deal, which is about to happen.

Great. And it's a good segue. As you mentioned, Arista outlined expectations for at least double-digit revenue growth for 2024 and...

Low double digits.

Sorry, low double digits. And that's despite, I believe, relatively minimal contributions from anything from generative AI. So what underpins your confidence in this outlook? What are you seeing on the enterprise side that gives you that visibility into that continued momentum?

Absolutely. Mike, today, we have over 9,000 customers. So when you talk about enterprise business, it's starting to shape up as a broad enterprise business with many countries, many regions, many verticals, many horizontals and both on the data center side and campus side. But when you look at the sales pipeline, I talk to customers, we're seeing good strength, good signals from all of them. Many of them have fatigue from the incumbent and want to change either because of some licensing discussion that isn't going well or the product quality isn't great, or in general, it was time to do a refresh or an upgrade to the next-gen architecture. This is -- now is the time to make a change. Today, when we walk into a customer, I can't even remember when was the last time I did an introductory deck to anyone in the enterprise space showing -- telling them who is Arista. We don't have to do that anymore. 2018, when we entered the enterprise space, that was the #1 reason my sales team would come and ask me, can you please come? Do the intro, and we'll call you in 6 months when it's time to close the deal, but in between, they'll manage. Now the team just does it on their own or we get a lot of inbound from CIOs, CTOs themselves. I think that momentum is still in its early phases, has lots of strong legs. I think we'll just continue for a couple of years to come before we have to worry about anything. Our execution needs to continue. Of course, that is the risk, but I'm not too worried just given what's in the DNA of our company. When you look at sort of overall trends and AI, generative AI and the hype we live in, if you read all of the newsletters that come out or all the post come out within a 7-day window, it almost feels like we're at the peak of a bubble for AI. If you take a step back, and there was a professor at MIT, he said this in a talk, I had heard a few years ago, where since 1979, the community -- the research community has been saying that the time for AI has come, about to take off. In the bigger picture, we might still be at the start of something massive. It's not yet to a point where AI has already taken over the world and their robots and your Avatar sitting in this room. There's no one here, you can just sit at home and get the meeting summary later. I think that may happen. There'll be twists to it. I think we'll adapt to the technology. But it's very early innings versus the hype we live in. But the cloud is big. The cloud doesn't easily slow down. The growth rate has come down, doesn't mean they go to zero, right? They're still adding lots of compute, lots of sites and all the impacts of COVID are over. So now if someone wants to light up another 30, 40, 50 megawatts of capacity, they have the operational tooling and the team to get there. That business is also in a steady state. I think it's all lining up to reasonable growth next year as well.

Great. And that's a good pivot to talk about AI and how Ethernet switching will play a role in generative AI infrastructure. Before talking about Arista's opportunity specifically, it would be great to ask you about the differences between the role of Ethernet and InfiniBand for high-performance networks. What's the history of some of these protocols and why can Ethernet be better positioned than InfiniBand in some instances for AI?

Coincidentally, this is the year we celebrated the 50th birthday of Ethernet. And Ethernet has come a long way in 50 years. It's a way to connect. It's a way to connect multiple things together. And Ethernet was designed as a best effort medium, not a lossless medium. And the network can drop a few packets here and there, what's the big deal? You should use things like TCP to make sure you have reliability. So we separated the stack and you have Layer 2 and IP is the Layer 3 part, but it's really the Layer 4 protocol stack that makes sure you get some retransmission or losslessness built into the application layer. This structure has now lasted us 50 years. And there are two technology stacks at any given time we've been looking at, Ethernet and Ethernaut. Ethernaut has always lost because you need -- you don't need the best solution all the time, but you need a solution that is broadly adopted. You get interoperability, you get an ecosystem and you get the support from a broad community that everything just connects together. Not a closed-loop system that is so proprietary that you're locked in for the rest of your life. Now on the other hand, InfiniBand, was really targeted to HPC. Early 2000s, Ethernet was behind on speeds. InfiniBand and fiber channels were competing for different types of connectivity. And there was a need for RDMA where essentially you're doing remote DMA from one server to the other for storage data. And InfiniBand added up a lot of tools to do this at small scale. During that era, 128-node cluster would make it to the top 500 supercomputers of the world. So it's really targeted to that top 500 supercomputers, HPC-style architecture, how to operate, how to configure, how to manage. But once you get it right, don't touch it, it will work fine. Now fast forward almost 20 years later, and there's a heavy debate on -- is it InfiniBand versus Ethernet? Ethernet has come a long way in this time frame. We've implemented a lot of enhancements to get RDMA to work over Ethernet. They call it RoCE. We call it Converged Ethernet -- RDMA over Converged Ethernet, which includes some kind of pause or back pressure, some kind of ability to communicate between a receiver and a sender asking the sender to slow down so that you don't drop traffic. The kind of things that would have been built in TCP and retransmission have been built within Ethernet and the nodes itself. We have deployed storage over RDMA with Ethernet with some of our large cloud customers now for almost a decade. And some customers are running more than 100 exabytes of storage globally through these networks. And they haven't had any major issues. They're actually extremely happy with the outcome. You get these massive ECMP networks where one router or device goes down, you don't even miss a beat. You won't even be able to know that something is missing because you lost 1/64 of your bandwidth when you're 64-way redundant in these large-scale ECMP, BGP, Ethernet architectures. We apply that to AI and actually able to get the same outcome. But many of the AI stacks were run by the HPC teams early on, 2018, 2019, and they got comfortable with it and they're scaling on that side. The same customers are looking at Ethernet and saying, this really works well in the rest of my company. Why can't I use it here? And they've done lots of trials or still doing trials, and the results have been great. And the reality is you can run some micro benchmark and throw [indiscernible] at each other and say, my stack is better and this particular corner case will drop traffic. That's largely BS. Real world doesn't work that way. Real world doesn't have only corner cases. It has 99% of use cases are just standard use cases and our customers who have used Ethernet and production workloads with AI applications, including third parties running on their clusters, have found virtually no difference. Like the end customer doesn't even complain. They wanted to stay up and running all the time, don't drop packets and don't interfere because the application team really wants to get on to the next job and the next job and the next shop and get the results and the whole world of how we do inference and licensing of these models will continue to evolve. One hand, our customers have -- some customers of ours have over 100,000 switches in production networks. Not 100,000 ports, 100,000 switches. Each switch has between 32 to 1,000 ports. So lots of lots of links and connections together. And we've tuned it up recently so that you can debug, you can monitor, you can secure these networks, you can scale these networks. And operationally, there is almost no impact to the end user or the tenant. You need the same paradigms applied to AI. Who says that an AI cluster does not need maintenance? Yes, it's running nonstop right now, but guess what, in 2 or 3 years, someone wants to bring down some rack for power maintenance, you run into the same operational challenges. The hackers are smart. They do realize that these back-end networks have nothing in between to observe the traffic. Can you get in? Can you attack? And if you're a cloud provider selling AI as a service, you need to worry about security. How do you secure a stack that was designed for a back-end HPC network that doesn't have any of these knobs or tools. And on the other hand, you have Ethernet and IP, and you can actually do that. So there's a lot of advantages, which is why in the long run, I'm very confident Ethernet will just win. It's only a matter of time. And again, the customer feedback has already been extremely positive. This is not we'll see, we'll try it out 5 years from now. These are customers running with it today, and you've seen some of the early adoption curves of 400-gig Ethernet in AI and 800-gig coming really fast.

That's really helpful. I was just wondering if you could talk a little bit about Arista's opportunity in AI today. Like how do you think about the TAM for the top 5 cloud and AI Ethernet switching? What's Arista's potential market share there? And why is it more internal clusters, external on top of the rack? Like how do you think about that?

So Mike, this is going to be a very tricky question because there's not enough research to actually measure this in any objective way. It's -- we can give you [indiscernible] statements. We can give you very exaggerated statements and so on, and this -- every possible way to justify each of these statements. One third-party analyst had published a report recently that Arista has at least 40% share already on the Ethernet part of AI. There's another report out there that says that AI networking will be at least $6 billion to $7 billion TAM in the next few years. At the same time, I think what some of these reports miss out on is the customers are not thinking of is my x86 compute and is my AI as two completely unconnected things, right? The applications are changing. And if any of you have the answer to this, let me know. I would love to know what that mix is between standard compute and AI 5 years from now. And what's going to happen is they're just blend together in this large data center networking TAM. I don't think it will be easy to measure AI separately forever. Early on, you can do that. You can track how many mix or GPUs are going with clusters. But if these sort of servers that customers are building on their own start to keep on expanding in every use case, it's very hard to measure. I'm not trying to dodge the question. I don't have the data. I think we already have the #1 share on the Ethernet side in AI. I think we will continue to do very well. We are very close to almost every large AI clusters in the United States, at least, if not even abroad.

Great. That's helpful. It makes sense, right, like I think a lot of traditional compute will be doing inference at some point in the future and how do you draw lines between that. You mentioned earlier that in some of the testing and piloting using Ethernet, like you're seeing a lot of success. Could you just talk about that evolution of testing, proving out, piloting trials before getting into large scale production and when Arista could start potentially seeing more meaningful revenue contributions from cloud AI deployments?

Absolutely. There's one twist here that has been long lead times in the previous year or 2. We're still fulfilling some of the demand from that time and customers' architectures have changed very rapidly as AI just swept the world. '23 has been a year where customers are doing lots of experiments with Ethernet. Results are positive. '24 is when you'll see them go to pilot. And pilot typically means a couple of clusters running their business-critical applications or AI as a service to other tenants and so on and doing a lot of optimizations and measurements and not so much with a worry that should I use this? Should I use something else? It's much more of, let's just make this perfect, so they keep on tuning it. Again, a lot more operationally how deployments happen, how do you observe for security and hacker getting in, things like that. And then '25, I think it takes off. There's another reason for that curve, which is 800-gig. And AI generally demands the fastest speed because we want the job to complete as quickly as possible. I think Meta had published a paper almost a year ago that about 1/3 of the time, the GPUs are sitting idle waiting for some I/O to complete on the network. So if the network can go faster, you could run the same job with 1/3 less GPUs or you could run 1/3 more for jobs in the same amount of time. How we look at it, these GPUs are tens of thousands of dollars each, if not more, you can save a lot of money or optimize that spend. The customers want to go as fast as possible. 800-gig is coming soon. And often, they like to time changes in the architecture with a next-generation node, which is why I think '25 becomes the year of 800-gig and AI simultaneously.

Great. That's really helpful. And I would like to hear a little bit from you about upgrade cycles, right? You mentioned 800G coming more in 2025. What has happened in the past in the 100G and the current like 400G upgrade cycle? How long does it take for the majority of those speed improvements to be completed? Any thoughts there would be great.

Mike, that's a very interesting question. Going back to our very early days, it took the world about 10 years to go from 1-gig to 10-gig. We went to -- from 10-gig to 40-gig in about 6 years. We went from 100-gig to 400-gig in about 4 years. We're about to go from 400-gig to 800-gig in about 3 years. There's enough demand that if everything lines up, we might end up in a situation where we go from 800-gig to 1.6-terabyte Ethernet in 2 years and largely getting pushed by the newest of applications. Each one of them had a trigger. And 10-gig was more -- the prices had to be more reasonable. Power had to be more reasonable for the technology to be adopted. 40-gig happened because storage needed it. 100-gig happened because the cloud just needed a lot more bandwidth. It costs the same as 40-gig, then why not? 400-gig, the initial applications were things like DCI. 800-gig, there's a lot of push from AI on these large clusters, same for 1.6T. So I think that's the goal we are on. But what we have to keep in mind is these are not easily comparable to the x86 upgrade cycles where the premium generation just dies and everyone only buys the next generation. In networking, you have a lot of legacy, a lot of interrupt, a lot of backward compatibility. So you'll see 100-gig, 400-gig, 800-gig overlap for quite some time. And in fact, 100-gig is a long-term stable node for service providers to connect to the cloud. As an example, we see demand at least for the next decade for 100-gig in those use cases as an example. So 100-gig, 800-gig, I think are long-term stable nodes. 1.6T, there's some electrical issues and power cooling issues to be still addressed. Once that's done, I think it will take off as well. But that's what we should expect on the cycle.

Great. That's really helpful. And then as you think about history, when there have been these speed upgrade cycles in the past, have they presented share opportunities? And what have been the key deciding factors in those transitions?

I think it's largely our competition thinks that they have an opportunity to gain share against us because there's a speed transition coming. But remember, it's the same EOS software running across the board on all of our products. If you look at the 7800, which is also referred to as the AI Spine, it's a complex system, and there's a lot of details that go in there. Just to give you sort of -- just imagine how you're running a network and then you have this massive switch, lots of interfaces, lots of data going through and you debug a few packets that were dropped and you want to find out why did they get dropped. Generally, on each interface, you have a few counters. You look at counters, what happened? There's an error, I can see and so on. And on a 7800 fully loaded system with just 300 -- 288 ports on an 8 slot or 384 ports a 12-slot system, we have over 1 million counters. The good news is most of them are 0 all the time. But when they're not 0, there's an error going on. And you need a lot of sort of correlation of what happened, which flow and so on to impact and gear to the customer. And our competition that can claim, I have the same density, still has a decade worth of work to catch up on the software side. So not that easy. Also, customers are not doing a complete transition like 100-gig. I think 100-gig was the last time where it was a complete switch out, stopped buying 40-gig. Only buy 100-gig, same cost, same power and so on, with the opportunity presented to us and we grabbed it. I don't think that same thing applies here. It's independent decisions for different parts of the network.

Great. I wanted to switch gears and maybe just talk about the Cloud Titan's customer vertical. You mentioned that some moderating growth, but it's still certainly growing and it's still very strong. Maybe you could talk a little bit about what's been driving some of that moderation? Do you see an increased threat from white box in that specific customer vertical? And then on the flip side of that, are there opportunities to expand business among other Cloud Titans beyond Microsoft and Meta to grow the Cloud Titan vertical even further?

Absolutely. So -- now as many of you know, I spent a lot of time personally with the Cloud Titan. So I love this space. It's been amazing to see the cloud grow. And Arista's Cloud Titan growth rate is slowing down largely because the Cloud Titan's growth rate is slowing down. And it's not correlated directly month-on-month or quarter-by-quarter. But in the bigger picture, it ends up roughly in that range. The -- there are a couple of views on how big the cloud market gets. And when I ask each of our cloud customers where are we on this journey? And these are visionaries that are involved in planning the next decade or longer, and they have to decide which -- how many gigawatts of power they need to build for and so on. The general view I get is, for the next 10, 15, 20, 25, 30 years, the cloud will just keep growing. The growth rate may keep on changing. And it's so volatile on an each cloud customer basis, but on an aggregate, it's on a steady state. And the slope of the line will keep on reducing, but it just keeps growing. So there's a massive opportunity still ahead of us. We might think cloud is huge, but cloud may not even be halfway of what it needs to be. And there are more and more newer use cases that keep coming up. And I think we not only sort of benefit from it, we, to some extent, drive that along with our customers. As you look at the scale and the next generation of these architectures, the complexity is mind boggling. If you look at any of the videos, if you've been to a site of any of the Cloud Titans and just look at the number of fibers that come into the large facilities, it truly is mind boggling how these things work and how it's so automated and efficient. Now you want to take it to the next level. And someone shows up and says, wait a minute, I want to build it on my own. I can build my white box. You want to get to 800-gig as quickly as possible to stay competitive. You want good software stack. We partner very closely with our silicon partners like Broadcom and our team is extremely good at programming these chips. In our previous lives, we've built our own ASICs as well. So we know how these things work. We often bypass the SDK and start programming it ourselves. That gets you much higher quality, that gets you efficiency. But more importantly, as you get to 800-gig and 1.6T, building these systems is very complex. You're trying to cool 1,000 watts, 1,500 watts in a small 1-rack unit or 2-rack unit switch. It defies the laws of physics. On one hand, the customers are talking about liquid cooling. On the other hand, someone says, I build it myself because I can do it with four people. It doesn't work that way. And these companies at a business level are fairly smart. They don't want to take so much risk that they miss a cycle. So this drama between build versus buy, I don't think is going to end, but I think it largely maintains status quo with customers recognizing they need good partners like us. And we are so entrenched, I don't know of a single Cloud Titan that we work with that has a road map discussion without including us. So we are talking about 2025, 2027, 2029, how will they evolve this architecture and are very much ingrained with it. Lastly, despite all this noise option, everything is equal, everyone can build a chip -- a box with Tomahawk 5 or with Jericho or competition shows up with their own chips. Quality matters. When you are running these networks with so many devices, so many switches, so many routers and you have data to show for 10 years which consistently outperformed any other alternative out there, whether it's internally built or externally, that carries a lot of value to their business teams. They're able to stand up in front of the largest companies in the world and say, we have a great infrastructure, come to our cloud and we stand behind that, which is why I think we'll continue the co-development efforts, that will only strengthen. The partnerships will only strengthen. And it's like, yes, we can debate on a monthly basis who gets how much share, but I don't see this changing in a larger scheme of things from where we are on status quo. If anything, building next generation is harder, which means, objectively, there should be a little bit more in our favor. Will customers actually go there? I don't know. We have to execute and plan and get that on their -- we'll find out in maybe 3 years.

Great. That's fantastic. So we have about 2.5 minutes left. I did want to see if there are any questions from the audience. Well, if you do, just raise your hand and we'll get a mic over to you. Anshul, I was wondering if you could talk a little bit about the campus opportunity, right? You guys have a target to get $750 million of revenue from campus by 2025. Who are these customers typically? What verticals may be best suited for Arista to address from an enterprise perspective?

Enterprise is a very large group of companies. And any company that has employees who come to offices, even if it's 1 or 2 days a week, need infrastructure. And the way we work today has changed a little bit post COVID. At the same time, the level of collaboration and online collaboration has increased 10x. So customers that we have, especially for campus, are continuing to upgrade, continuing to build quite a bit. We are able to see customers in the big financials that are already good customer on the data center side. We have customers in retail. We have customers in health care. We have customers in manufacturing. We have customers in transportation and just keeps on increasing. So today, it is quite broad-based. Our campus journey is still getting started with many of them. Many of them we'll deploy on one site and say give it a little bit of time, everything goes well. I'll do a full refresh with you or give you more than what share that we're planning on earlier and so on. And I think that just gets us on a good growth trajectory for many years to come. This is a pretty large TAM, right? You're talking about at least $14 billion, $15 billion just in swing plus WiFi. I think there's a question out there.

Great. Can we get a mic over there? Great.

Just had a question regarding Ethernet competition from NVIDIA. They're clearly pushing InfiniBand very, very hard right now. But on the last Q&A -- last earnings call, they did say, we're also going to offer an Ethernet option. How do you see that evolving in that space with them making probably the most serious player being in Ethernet for a while since Arista's own emergence, obviously?

It's a great thing that NVIDIA is acknowledging they need Ethernet too because they need it at scale. And I think Jensen actually said this at one of their big events. When you look at the spectrum chipset versus the offerings we have, where we have a very broad portfolio between every possibility of the Tomahawk chips as well as the Jericho chips with very deep packet buffers, and that's an offering that they don't have or almost nobody else in the market has. But going back to the tooling and the software and everything else we've done, we have a decade of an advantage in that stack, right? We've been running RDMA over Ethernet for a decade now. So the level of maturity we can provide is already there. When you get into the hardware details of how these are cooled, can you use linear drive optics across the board or not? It's a very big deal. Linear drive optics save you at least 15%, 20% on cost per optic and about 25% of the power per optic. That alone is more than the cost of the entire switch for many of our customers. We bring those advantages, right, on day 1. I don't know they can do that easily. So lots of benefits when you go into details. Just don't look at a high level of, hey, they have 32 ports, you have 32 ports, everything is equal. That's never the case.

Great. So we're about a minute over. I just want to say thank you, Anshul. It's been such a privilege to be able to moderate this session with you. Really appreciate all your thoughts.

Absolutely. Thank you.