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

Good morning, everyone. I think we're about to get started. Just a brief intro. I'm Mark Newman, Bernstein's U.S. IT hardware analyst. I'm joined on the stage on the far end by my colleague, Daniel Zhu, who is Bernstein's new networking analyst. And Daniel and I are delighted to welcome Arista's, John McCool. John is Arista's Chief Platform Officer and Senior Vice President as well as Special Adviser to the CEO. John has been at Arista for 9 years and prior to Arista had more than 35 years executive networking experience, including being SVP of Dell EMC and 17 years at Cisco. And on my left here, Rudolph is Arista's Head of investor advocacy and has been at Arista for 9 years as well as another 13 years of networking and cybersecurity experience prior to Arista. So I think given the huge shifts we're seeing in the space, it's a great honor to have both of you here today.

Thank you for having us.

So I'll get started, and then I'll hand over to Daniel. So perhaps a very high-level question just to kick off on AI. The AI opportunity is clearly huge. Can you talk about how you're viewing the AI opportunity, how it's evolving and risks and opportunities for Arista?

Sure. So I think, clearly, Ethernet has a place in this AI opportunity. I think if we went back a couple of years ago, there was a question, what role does Ethernet have connecting what we call the back-end network, connecting GPUs and clusters. We've made substantial progress, not just as Arista, but as an industry in the standardization and development of that back-end network. And then the interconnect of these clusters, we call the front-end network also grows tremendously along with the challenges of power consumption. People want to connect multiple clusters across large numbers of physical locations to get power effectively. So that's kind of the framework that we think about the Ethernet opportunity.

Yes. I mean I think the other thing that is clearly manifest is the networks are getting larger, right? I mean you can't throw enough compute at the problem apparently. So now how do you build these larger and larger networks that are still just as efficient. So I think lots of room for innovation there, and we've done well for the last 20 years by innovating and really continue to drive the performance bar higher and higher. And I think in the AI networks, it shows itself up even more.

I guess kind of adding on to that, how are you seeing the evolution of the AI opportunity between front end and back end and also between scale-up, scale-across and scale-out?

Sure. Do you want to start?

Yes, I can start. So one is it's -- we're, I suppose, lucky in the sense that our products are very fungible, right? The same products can work in a front-end network, in a back-end network. And what it gives customers is that flexibility. So it's a little bit harder for us now to parse out like what is front end versus back end. And customers, I think, are increasingly finding that these networks are kind of blurring together, right? I used this analogy in one of our group meetings earlier today, where if you're on your Instagram app and you're trying to look like a Cowboy, is that front end? Is that classic data center? Is that back end? Like where does that line get drawn, right? So that's kind of one thing. The good news is all of that growth has been good and great for our business. Then talking about scale-up, scale-out, scale-across. So a couple of things to remember. Like scale-up is really within the context today of the rack, but it's starting to go beyond a rack, right? And this is kind of the highest speed network, if you will. It connects to the memory, but it is also, in some sense, the simplest network, right, because there's not as much complexity kind of involved there. That's an opportunity that today is really captive since the bulk of the accelerators being used out there come from NVIDIA, it's really an NVLink is the protocol that's used a proprietary protocol. So it's not much of an opportunity today, but there are some efforts called Ethernet for scale-up networking is a term that you hear quite a bit. So we view that as an opportunity for Arista probably in '27 at the earliest. So not something that we include in our TAM right now. Scale-out has been the bulk of that AI revenue. That's where you're connecting across racks, maybe even across buildings that are relatively close to each other, right? And then scale-across is something that's come up fairly recently, let's say, over the last year, but it's frankly a concept that's existed for a while, right? Because if you have data centers across multiple locations, how do you interconnect those data centers. AI brings some unique challenges there, but you've got to think about things like routing and encryption and traffic engineering once you start going. So I think that for us is a very exciting opportunity because it takes a unique set of platforms, which there's not that many companies that have those.

It plays to our strength that we've had in traditional cloud. What we saw in traditional cloud was the same phenomena. How can I connect as many CPUs as I can in the data center? And once we were able to scale ECMP very wide, those data centers provided not enough power. And then we had -- that was actually our impetus to develop routing to add another layer across. The same thing is happening with GPUs and the power consumption is more constrained. So you want to have a broader number of data centers that look like one logical GPU cluster.

Got it. And just double-clicking on that because I think -- sorry, I think the scale-across opportunity is a little bit less well understood by investors. How do you think investors should think about sort of dimensioning that opportunity? And how do the competitive dynamics differ from the opportunities within the 4 walls of the data center?

Sure. Maybe I'll start with the competitive dynamics. These are highly sophisticated networks, need strong routing convergence time, need capability like encryption end-to-end, need multiple optical type of support, so variations in the type of optics dependent on the physical locations of these different data centers. So it's extremely sophisticated. It plays to the strength that we have in our modular platforms, which have a virtual output queuing architecture with deep buffers to deal with the distances involved and plays to our traditional strength. And as Rudolph said, very few suppliers can compete in that kind of environment.

Just a reminder for the audience, we do have an app. You can submit your question there. We'll get to some time at the end, questions for the audience, you can enter in the app or if you prefer to ask live, we'll offer an opportunity to do that with a microphone as well.

So I think kind of speaking a little bit about architectures, I think within AI, we've seen a shift from AI model training being very dominated by pretraining to one that has more of a component of kind of post-training and test time compute as well as inferencing ramping a little bit more. How does that impact sort of the network architecture design that you guys are seeing? And how does that impact your networking opportunity?

Yes. I think that the increase in this pretraining activity brings broader storage and more machines into play, which really drive that next layer of the architecture and that front-end network as well as pulling data from different data centers and that scale-out capability. When we move to an inferencing environment, we think a little bit about the user interaction, which adds a latency component to the question. So you input something, it has to be computed back-end network and then get back to you. So it really interacts across the whole dimension of that network. And we see enterprise customers thinking through their next-generation campus networks and how that will impact them. We had one discussion in the early days of training where one of our large hyperscalers talked about the things they hadn't thought about. So they were so focused on that back-end network. They really hadn't thought about the dimensioning of the front-end network and how that was impacted by pulling in the storage data. And one thing I thought was interesting was on the wide area network. If all the content that we're getting from AI is individualized now. We're not all watching the same cat video. We're watching different cat videos. It can't be cached. So some of these wide area networks were designed with caching capability to kind of eliminate some of the bottlenecks on the back end. So there'll be pressure all the way to the user as this becomes more of a widespread opportunity on inferencing.

There's also things like RAG and stuff like that, that are kind of expanding the aperture of what connectivity is needed, right? So it's not as simple as, look, I've got these or 5,000 or 50,000 GPUs that I'm just trying to interconnect, which is, in a sense, a simpler problem. Once you start having to go beyond just the resources that you control and you own and you're trying to connect the service on the Internet, maybe you're trying to connect to your database that sits in a different location. It just opens up more complexity. And I think one thing that I'd say as a takeaway is complexity has been Arista's friend, right, because more complex networks require higher levels of R&D, higher levels of engineering. And all the way from Andy and Ken, who started this company, I think we've shown that we can innovate better than most.

So is there still a lot of experimentation going on with the cloud customers in terms of how they deploy GPUs and networking to -- in response to power and cooling requirements. Is there still a lot of experimentation? Or are you seeing more of a consensus for the direction forward in terms of how it is done?

Experimentation on one angle, but the intensity of the power consumption and the cooling becomes a question even within a single customer of where I'm deploying and what kind of capability I have in that data center. So if you have a modern data center with liquid cooling, maybe that customer -- and a lot of power, I'm trying to affect the highest level of GPU density within that form factor. I may have legacy data centers that I don't have any liquid cooling. I will still want to air cool, but still have some capability in there. So there's a broader diversity of physical form factors to meet the market needs than I think we've seen. And I don't know whether that gets better. People started to standardize in cloud on not only just leaf-spine to directions, but the width of the rack, x86 GPUs, there just seems like there's going to be more diversity because we're on the edge of that power and cooling dynamic.

And some might take offense to the word experimentation, I guess, but the space of innovation, which is I think the point you're making is not slowing down, right? Like to give everyone an anecdote, we introduced our 400-gig platforms in 2019. We introduced our 800-gig platforms in 2024. So call it, 5 years. It's not going to take 5 years to go from 400 gig to 800 gig. It's definitely not going to take 5 years to go from 800 gig to 1.6. So what that is leading to is a lot of experimentation and trying out different things because ultimately, what they're trying to optimize is power utilization, time to first token, job completion times, being able to generate more tokens per dollar, more tokens per kilowatt, like all of these things are -- so they're willing to try pretty much anything.

The utilization of the GPUs, you want to maximize the utilization of the very expensive assets.

Exactly. Exactly.

I guess since we've talked a little bit about sort of the technology road map and 800 gig versus 1.6., can you talk about sort of the time line of that rollout and kind of what you guys are seeing with that sort of transition?

Yes. I think you just hit it. It's become more compressed than we saw 100 to 400, 400 to 800 compressing. It's going to go faster. We've seen a lot of announcements of silicon for 1.6. I think that one thing that we have learned, we've gotten this question, by the way, I came into the business, we had really secured a substantial position in market share, 100 gig. And everyone said, what's going to happen at 400 gig? And then what's going to happen at 800 gig. And we've been able to layer on those technology transitions based on a consistent software architecture and methodology and development. We'll always have to add new -- some new things and capability as the speed increases, but it's been pretty straightforward. And I think we've led that deployment. The critical time is from first silicon to thousands of GPU connections. A lot happens within that. So you see us tend to wait until we really have deployments to make announcements around the next generation.

Got it. And continuing with the technology road map a little. We're starting to see a lot more talk about co-packaged optics. But in the past, I think you've talked about how the failure rates of optics is a big bottleneck for CPO adoption and how replacing the DSP and retiming chips of linear drive pluggable optics actually captures a lot of that is a pretty effective substitute. Is that still true with the 1.6 cycle? Or what are you kind of like starting to see more evolution here?

Yes. I'd like to answer this question by kind of coming back to why co-packaged optics began and the concept. There's really 2 things that are intriguing about co-packaged optics that we're trying to solve. One is power consumption. So if I can shrink the distance, but also the number of chips between the switch chip and the wire, either cable or optics, I save power. So that's a good thing. The second problem was in each one of these speed transitions, the laws of physics decrease the distance that I can run on a printed circuit boards between the switch and the front end. So the ultimate shrink is really putting optics on the die. So I think we would say there's some inevitability to the trend of shrinkage, but we've been able to defy that in a couple of ways. The first observation we had is the improvement in the switching chips DSP technology that was designed for co-packaged optics allows you to eliminate the DSP in the optics component itself and preserve the operating model that the hyperscalers use of pluggable optics from multi-vendors. If something breaks, I can replace it. So we were -- and get all the power -- most of the power consumption savings that you have with CPO. So we were able to push out that transition 1 -- and we believe 2 generations. But there will ultimately be a time where to get the laws of physics, bringing that optics closer to the switch chip is inevitable.

One thing I would add is the biggest kind of concern around co-packaged optics is optics fail most often, as you pointed out. And if the optics fails and the optic is actually on the switch or on the board itself, you not take the whole switch out of commission, right, which means every workload that's connected to that switch is now paused or shut down or has to restart, et cetera. So the reliability requirements of the switch have just gone up exponentially, right? The hardware engineering -- and again, complexity is our friend at Arista. We've got the best and sharpest engineers. So we feel really good about engineering for that co-packaged era. But I think what customers are telling us is kick that can down the road as much as possible to John's point.

Fans, power supplies and optics tend to be the things that people want to replace on the fly. So you make in a situation with co-package, maybe there'll be some improvements in the overall reliability of the lasers. Maybe networks will be designed to be redundant of some path failures, so you can delay the serviceability requirements of those switches, but there'll have to be some operational adaptations.

Got it. And I guess continuing with the theme of emerging technologies, sort of optical circuit switching is a technology which has actually been around for a while but remained fairly niche. But at least we're starting to hear more talk about it. And I think in particular, there's talk about scale adoption by one of the hyperscalers. Can you talk a little bit more about what you guys are seeing in the space and sort of how that would impact Arista's business?

Yes. I think you framed it well. I think there's been one customer in particular that's used optical switches and some large-scale deployments. And you can think about them having massive scale. It's not equivalent to switching or routing. It's not a per packet-based decision on which way the packets go. It's more of a construct that interconnect some ways thinking about replacing patch panels, being able to allocate and move different elements of your compute or GPU environment at a more granular level. And as some of these deployments and other customers get larger, they can have the benefits of that kind of technology. So I think there is some elements of TAM growth because more people will be able to use those technologies, but it's still kind of a different aspect than traditional switching and routing. And if you want to...

Yes. I mean I think to Mark's point, this fits in that experimentation bucket, right? Like is there some pound of power savings? Is there some pound of performance improvement I can get. But even at this customer that has been using optical circuit switching for a while, once they get beyond a certain layer in the network, it is all Ethernet, right? Because the flexibility that Ethernet gives you is just tremendous. The variability in the supply chain, the diversity of suppliers, all of that, I think, brings its own advantages. So I don't think we view those as necessarily competing. I think the TAM for optical circuit switching is significantly smaller than Ethernet switching. So...

Got it. And we've kind of been talking about a few technologies that sort of are real, but maybe less disruptive than some investors might hear. Are there any technologies that sort of are on your radar that you think kind of are being underappreciated by Wall Street?

I'd say we touched on this earlier. The scale across opportunity, I think, for us is very, very exciting, right? I mean to kind of repeat a little bit of what we said earlier, like it takes a unique set of technologies. It requires these deeper buffers. It requires the virtual output queuing. It requires encryption. And that's something that I think probably the biggest misconception I've heard out there is people assume that a switchboard is a switchboard is a switchboard, right? But to John's point from earlier, a low latency switchboard is much more different than a scale-across switchboard. And so that nuance, I think, sometimes gets missed.

Got it. And I guess if we zoom out and kind of like talk a little bit more about the platform, right? I think at the Analyst Day, I actually asked Jayshree, why is it that Arista was the only -- essentially the only networking company that sort of entered the market after the dot-com bubble and saw sustained success. She highlighted that data center switching has really been kind of the foundation of Arista's strength for a long time. That's a category which has been one of the fastest-growing and most important components in the data center and probably all the more so now. So can you talk about how sort of you're leveraging that strength in switching to building a platform that supports other products and services?

Yes. Maybe just to amplify Jayshree's answer. In 2010, '11, it's hard to imagine how small these hyperscalers were. So it was a classic case of an underserved market. And a new competitor coming in with hyper focus, not trying to take enterprise networks or service provider networks and serving the emerging hyperscalers. Their problem was they had very smart people, sophisticated operators, but we're trying to connect together millions of compute nodes with a very small team. So how can I put my agent on your networking device like I do with my servers on a Linux operation and build a centralized management stack? How can I have resilience and scale and being able to scale out without bespoke different operating systems? So they were sort of the architectural premises that we developed at Arista. The hyperscalers had their own management stack, a centralized management stack. They encouraged us to be able to stream data to that management stack. We took that concept and built something called CloudVision, which is a management stack that we use for people who don't have the wherewithal to build their own. So back to your sort of question on AI, this being able to scale and manage is happening as we speak into campus environments. If you think about a university, it's very easy to think about a new student coming in with 3 devices they roam with, right? Maybe a pad, laptop and perhaps a watch. And they're moving around campus. You imagine them all coming to see a sporting event that seats 20,000 people, all of a sudden, you're up to 60,000 Mac addresses in a very small place. It starts to feel like connecting GPUs and CPUs. We've taken in our WiFi stack concepts that have been developed and standardized in the data center like EVPN and VXLAN, and with our VESPA architecture using Wi-Fi 7, being able to build these large Layer 2 roaming networks that fit into campus. It is a similar problem except these people move around. They're not standard like racks. The operating principles of a single management stack and a single operating system in your data center and campus are consistent with what we've done in the data center. So very much plays this whole consistency and scale piece that we've focused on.

Yes. One maybe thing that I'd add is our customers -- what customers really liked about us in the data center also was just the cost of ownership of an Arista network is fundamentally lower, right? Because it fails less often, it's easier to maintain, it's easier to operate the level of automation, et cetera. And so if you think about it, who has even less money to spend on networking, right? It's the enterprises because they might have one networking person. If you're a large bank, for instance, which many of you probably work at banks and you've got 5,000 ATMs, you don't want to be sending a tech to reboot a switch at every ATM every few weeks, right? I mean it's just not cost effective. So you want that reliability, you want that centralized management, you want that automation. And I think that's the advantage we've been able to bring from our cloud learnings, if you will, to the rest of the world, right? So we call it the modern operating model, and it's been very, very effective for customers.

It's hard to imagine in networking, customers were very scared to upgrade because they can run into regressions, something that used to work, I've upgraded it, now I broke it. And with kind of the concerns around security and security alerts, being able to move quickly is really important. It was really critical in the data center, but it applies to the entire network.

Got it. And now that we're talking sort of a little bit more about enterprise, where are we seeing sort of enterprise AI adoption inflect the most? And how do those use cases differ from sort of what we've seen kind of with the cloud titans?

Yes. Maybe I'll start and you add in here. I think definitely, we see different verticals with a different approach. I think the consistency is an enterprise may have certain amounts of data that they want to monetize or take advantage of and build out their own AI pieces. It could be in the financial vertical, maybe around security and credit card and credit card handling and fraud. In health care, it could be detection of anomalies around your health or radiology. So we see these vertical applications emerging in AI, specifically where a company may have a lot of data that they want to take advantage of.

Yes. I mean I think the health care example is a great one, right? Like we have customers that will have conversations with us about, look, we want to take advantage of this AI wave, but we're concerned about -- our patient privacy, right? So like how can we ensure data is segmented correctly and there's encryption on the wire, and we've got threat detection on the wire, et cetera. So we've got a whole slew of services now that -- and software capabilities that we've built within the switches that can provide that, right? I'd say the biggest trend we're seeing maybe with customers is fewer of them are trying to build their own training clusters because I think they recognize that the amount of CapEx investment that it takes for that is incredibly high, right? But what they also realize, and this touches on something John said earlier, is inferencing is very much a latency-driven game, right? If you're going to ask a question, and it's going to take 2 hours to come back with a response, you might as well not have it, right? So how do you optimize that? Well, you can try and bring AI to the edge, if you will. So you're already starting to see questions about, okay, is there going to be an accelerator on each of these devices you have running in front of you? Can we bring the inference compute as close to the edge, right, whether it's putting it in a data center that doesn't belong to you, but closer to you or putting it in your own data center. So those are some of the trends we're seeing. But I mean, we talked about over 100 800-gig customers that we have now and just from about a year of the product being out there, right? Obviously, there's not 100 hyperscalers. So there's a small number of hyperscalers in there. There's some of these new clouds and specialty clouds, but a fair amount of enterprises in there in the verticals that John mentioned, right, insurance, financial services, health care, the educational sector is up there, manufacturing, some of those folks are trying to think about the whole AI factory concept.

Got it. And we spent a lot of time on AI, and we kind of have touched on this already, but campus networking has obviously been a huge area of focus for Arista. Can you talk a little bit more about sort of value proposition and kind of the go-to-market in scaling that business?

So yes, I think it's important to think from an Arista perspective, campus is part of our enterprise go-to-market. So we have a very intimate engineer-to-engineer relationships with the large cloud providers. But we've had from the beginning of the company, starting with the financial vertical an enterprise go-to-market. The only thing they could sell was low latency switches when we started. Then we added more broad data center products, and we've been continually to add to that. We don't necessarily go to a customer or target a customer because they're a campus player. They buy a lot of networking equipment. Networks is important to them. And then our sales team will look for whatever opportunity exists within that account. Maybe their WiFi is coming up first or maybe it's a data center opportunity, but they have more arrows in the quiver, if you will, to go build out an enterprise network. So that's how we go about it. We were pushed from our data center customers. Why aren't you going into campus because they want the same benefits of a single operating system, the same operating model into the campus pieces. So we took that to heart. When we were ready, we built our campus LAN switches organically. We did add a Wi-Fi component that I would say we've [ Aristafied ] into the architecture and going about it that way.

Yes. I mean I think it's important to keep in mind that the campus dynamics are also slightly different. I mean the refresh cycles can be 7 to 10 years. Data center, typically, even 3 to 5 years, you see a refresh cycle. So that's kind of one thing. It's also very rare that someone says, okay, you know what, come in and replace my entire campus my switching, my WiFi, my SD-WAN, right? It's much more piecemeal, right? Like where they'll say, okay, you know what, you won this business, but we're going to start with you in one building and then maybe a few months later, you get the second building. So it's a much more slower grind in that sense, right? With that said, I think we're very happy with our growth, right? I mean we told the Street that we were targeting $800 million exiting the year. It was $750 million, and then we acquired VeloCloud, which John can talk in more detail about since he's heavily involved there. So we bumped that up to $800 million. We were able to meet that target. And now for this year, we said $1.25 billion, right, which we actually got this question earlier. I mean that's pretty aggressive growth in a market that isn't growing as fast, right? And the reason we can do that is we're only about 3% market share in that campus market. So it's much easier to grow at a faster clip when you're a share taker, right? And it's -- I think we feel very confident about being able to become, frankly, one of the larger players in the campus in the next year or so. Again, there's one big player in the room and the name talks to the C, ends with an O. And then there's HPE Juniper, which obviously with -- especially with the merger now, they're in the probably in the teens percentage-wise market share. But beyond that, it's been quite fragmented, and we feel like we've got a pretty good opportunity to kind of grow that business. So it's an exciting opportunity. It's a different sales motion. It's a different pace than the data center. But in many ways, I think customers care about the same things, right? They want a network that just works, right? You're not having to reboot stuff. You're not having to throw people at it over the weekends, things of that nature.

I wonder if I could just follow up on -- you mentioned HPE Juniper, obviously, a big merger there. Has that created an opportunity for Arista? Or what's the -- how is Arista benefiting? Or is there any risks from this acquisition that you see?

It's a discontinuity in the market. So that always presents an opportunity. There's been customers who used Juniper and HP as alternate sources. So now they don't have an alternate source, particularly around the Wi-Fi piece. There seems to be 2 things moving forward. So customers think about that. So it's an opportunity to create a conversation or have a conversation. But it hasn't fundamentally shifted our strategy. We're not sort of reacting to that event. We're still focused on the bulk of the market share and what we're bringing to market as opposed to changing our direction.

And do you see the combined scale and breadth that the new HPE has in networking as a potential threat to Arista potentially them coming in, gaining a bit more share in your areas or...

No, no. I mean what we've seen, it leaves us really as the largest pure-play networking company that can focus on every need in an enterprise. And that's a pretty big TAM for us. And as we said, particularly in the enterprise on the campus side, very low market share, so we can be a disruptor in that space. While on the other side, there's some rationalization and there's other pieces with storage and compute that require some focus for networking. So I think overall, we view as an opportunity.

Yes. I mean I think, again, this market, even if you go back to the data center market, size is not always an advantage, right? And frankly, it's one of things like Jayshree always keep us honest, right? Even though...

We disadvantage it.

Yes. Even though we've grown, we've got to keep that nimbleness and that start-up kind of mentality, which is why we tend to operate very efficiently and have stayed focused on pure-play networking, right? Because when your attentions are divided across storage and across compute and things like that, it's very easy to kind of move resources to whatever is the flavor of the day. And customers don't have patience for that, right? So I think in some sense, it helps create opportunity. It hasn't -- to John's point, it hasn't fundamentally changed the dynamic. And the largest player by far is still Cisco, right? So I think the opportunity is still there for, frankly, them and us to take share.

Great. So I had a follow-up question on memory, a bit of a hot topic these days.

We waited 30 minutes to get there.

Well, we have to ask about networking, too. So your CEO, Jayshree, talked about memory being the new gold and pricing environment being horrendous and lots of talk about exponentially higher pricing. My question really is how is Arista positioned versus other OEMs here? Are you able to secure enough memory? Or are you memory constrained right now?

We're not memory constrained. But as she mentioned, pricing has gone up, and we've made some targeted changes on our pricing to address that and also changed our internal price structure to make sure we have continuity of supply.

Yes. I mean it's -- if you noticed our purchase commitments have been going up. So it is -- and this is not specific to memory. I mean memory is a relatively small part of the switching BOM, right? We're not a server manufacturer where switching -- I'm sorry, memory is a bigger part of the BOM. So impact-wise, it's less of an impact. And also, I think one nuance that sometimes maybe the Street doesn't entirely appreciate, the more complex the switch the more memory. So for instance, the chassis switches, there's more memory in those. But as a percentage of the BOM, it's actually a lower number, right, because there's a lot more into that switch than just memory. So what -- my point being is we are not maybe as impacted as a server maker. With that said, it's clearly an industry-wide constraint. I think our customers get it. So to John's point, where we need to, we have the ability to do these targeted price increases on these more memory-intensive SKUs and the customers get it. It's always a negotiation. I mean this is not a -- we'll say and they just do it kind of thing. But at the same time, I think because we've got that free cash flow, because we've got a lot of cash on the books, it's allowed us to make these investments to get in line? Because right now, I think the challenge is how do you get in line for that supply. But we don't believe -- pricing aside, we don't believe we are supply constrained in terms of being able to meet the demand we anticipate and our guidance is still valid.

Do you have a rough guidance on what portion of the bill of materials is memory?

We haven't -- we haven't disclosed that, and it varies.

We never disclose that.

Servers is going from 10% up to like 30% now.

Just because of the price. Just because you have the price.

PC is also from 10% to 30% roughly.

Yes.

So I mean, you're definitely in the single digits, I think.

We haven't disclosed that. But the percentages are going up just by the nature of the price increases. I would also just add in terms of being able to be in line, our relevance in the cloud market with top-tier customers helps some people understand where we fit in memory. They're also cognizant of the golden screw. If there's a lot of GPUs, but they can't connect them with switching, that wouldn't be great. So I think that suppliers are aware of that position. And obviously, we have a strong balance sheet that helps us make those multiyear kind of commitments where needed, right?

And your strategy is to pass on these price increases to customers and try to maintain your...

Well, some of it. I mean we've been able to absorb some of it. I think Jayshree said this or Chantelle said this on the call, right? Like Jayshree actually signaled the memory thing at our last earnings call, right? And I think everyone is a little bit surprised why she brought it up. But in many ways, I think it shows that she was almost questioned about where this is going. Now it has gotten actually worse since then. So we've been able to absorb for a while. But at some point, we can't, right? I mean the good news is our customers are some of the biggest buyers of memory, right? Like if you talk about the hyperscalers, they need more memory than pretty much anyone else. So they get it. But no one likes a price increase, we don't like a price increase either. So it is always a tricky thing. But the question is really about our margin guidance. We feel very comfortable with that 62% to 64% gross margin guidance for the year that we can fit right within that using a variety of mitigations, right, whether it is absorbing some of it, obviously, purchase commitments and then some times price increases are needed.

Great. Thanks for that. Before I -- before we hand over to audience for questions, just one question I want to ask you, is there something -- what do you think that Wall Street is missing in your company? Is there something you'd like to highlight that you think Wall Street is not paying enough attention to?

When we have these conversations and even the discussion today, we think about from an investor point of view, TAM and opportunity, and we tend to segment things, right, front-end network, back-end network, campus. From an Arista perspective, it kind of -- especially the enterprise, loses track of our value proposition. Our value proposition, if we walk into an enterprise customer is, I can now deal with or execute any use case you have today, whether it's distributed enterprise wide area network connection, GPUs, connecting together your virtual hybrid networking environment. And I can do this all with a single operating system that is proven to be extraordinarily low in defects versus our competition and simplify your operating model. So we go to market with that value proposition to an enterprise customer. They give us some thing about what's coming up in the next 18 months in terms of their changes, right, whether it's a management network in the data center or now with Velo, we're interconnecting a wide area network in a distributed fashion. So I just wanted to make that point and connect these different network use cases and area to really what our sellers are going to market with.

Yes. I mean maybe from my perspective, what I would add is I think networking is incredibly technical and complex. And maybe we need to do a better job of kind of simplifying that for the audience. But I think people sometimes assume like I was saying earlier, right, a port is a port is a port, like it's just a pipe, like how much complexity can be there. But there is a lot. And because of that, I think people assume sometimes that networking is also a zero-sum game, right? Like if -- like the question earlier about optical circuit switching, just because optical circuit switching is seeing some growth doesn't mean that Ethernet-based switching is declining because of that, right? It's not a zero-sum game. And I think especially in this AI era, like it is definitely a rising tide kind of setup. So you're going to see all kinds of companies, I think, do well because of that experimentation thing because of the pace of innovation, et cetera.

Great. Appreciate that. Any questions from the audience? I don't currently see any in the app, but feel free to stick up your hand, and we've got a microphone here and can come to you. Any questions? It's all super clear. We have one right here.

Question, things that are out of your control. What is the biggest thing on the medium-term horizon that is out of your control that will constrain growth? Is it power? Is it memory? What are you doing about this? How are you thinking about it? What mix can you put in place?

And power is definitely up there, right? Because I think customers want to build these as quickly as possible and power is probably becoming the biggest constraint. So to some extent, you're right, we don't control it. But what we do control is can we get more power efficient and power efficiency out of their clusters, right? Because the network is not a big consumer of power, but it is a consumer of power. And so some of the stuff that John talked about earlier in terms of reducing the power utilization by optics -- from optics by 50%, right, with our linear pluggable optics. And again, we don't sell a ton of optics, by the way, which is I think another thing that the industry misses, but we are definitely thought leaders in innovating when it comes to optics. Andy Bechtolsheim is considered essentially one of the smartest minds when it comes to optics. And he's been pushing this idea of linear pluggable optics, eventually co-packaged optics, et cetera. So that's kind of one thing. But even if you look at the switches themselves, right, look, I love this comparison where even with the same Broadcom chip within our switch, right? So let's take a Tomahawk-based chip, put it in our switch, put it in another branded vendor switch. Our switches tend to be 15% to 30% more power efficient than the competitors. So that 15% to 30% that you can now save, you can apply to more GPUs.

Why is that? How can you save power?

Because we just -- it's kind of the stuff that John touched on earlier, like for instance, better signal integrity on the boxes, being able to design the hardware so that you're being as efficient as possible, getting rid of some power hogging components because you've got that better signal integrity. So it's just a variety of things.

You can lower the output power to the effect of the channel and minimize that, more direct control over the SDK on the software side to do some optimizations, thermal integration with how we do heat sinks and cooling to run the fans at a lower power. Lots of little things that add up to some pretty substantial changes. And sort of back to the memory, not 100% direct control, but definitely, we have an opportunity to influence that with some of the things we mentioned around our engagement with the large cloud customers and our momentum in the market is influential in helping us get some memory.

Thanks for that. Any other questions? We have one question over here.

They're making you walk the microphone today.

The other end.

So there was a large AMD meta deal announced yesterday. Is that additive to your TAM? I don't know if you can comment on that specifically. But then more broadly, in these -- once a large compute deal like that gets signed, let's say, is someone like a risk that consulted kind of as that discussion is happening? Do you come in later in the process? How does that sort of time line work of when the networking piece would come in?

Yes. I would say, look, I think we've been talking about the diversity of GPUs playing to Ethernet as the environment for running AI. We largely believed even when things were back in the InfiniBand days that, that would happen. AMD didn't even have a GPU in the market at that place, but we've just seen transitions like this that people want a diversity of endpoints and they want a consistent way to operate and connect those endpoints. So I think it's validation that there's going to be a multi-GPU environment. It's an external event. It's probably more long term in terms -- you see us making longer-term supply agreements. I think people are having to think through longer-term agreements in this environment to make sure that they have the capacity to grow. So I think it's all consistent with what we have baked into the model.

Yes. I mean I guess, in general, the biggest advantage, I think, the diversity of suppliers in any area gives -- it eliminates strategic locks that single vendor might have had prior, right, where they can influence buying decisions beyond just the GPU. So like someone saying like, hey, if you buy my GPUs, the network comes for "free" or you get the GPU sooner because you bought the network. So I think that -- it opens up the market, right? I mean the good news is all of these third-party, whether it's the branded accelerators or the homegrown kind of ASICs that are getting built by some of the larger players, they're all Ethernet-based, right? So that is a market that we can compete with. We feel very good about winning in a best-of-breed kind of pipe. And it eliminates some of these maybe more go-to-market type impediments that we would have run into otherwise. So I think it's a good thing. It's directionally something that we've always planned for because I think customers always tell us that, look, we don't want to be locked into one vendor for anything. You guys included, right? Like even with us, right from the early days of the cloud, we've known we've had to coexist with other networking vendors. And it's not at all surprising that these customers would not want a single source something like compute either.

All right. Any other final questions from the audience? We're almost out of time. Maybe have time for one more if there is any other question. I don't see any hands, last chance. Daniel, do you have any other question on your side?

Yes. I mean I think sort of -- I think kind of we kind of just go back to sort of that high-level sort of like AI discussion, right? Jayshree has really talked about sort of addressing a $100 billion TAM. And as part of that, Arista has guided to AI networking revenue doubling from $1.5 billion in 2025 to -- I think the number is $3.25 billion in 2026. And I just want to get a sense of like how much of that expansion is coming from hyperscalers spending more versus Arista being able to capture kind of a bigger share of that opportunity compared to competitors than you might have anticipated before?

It's hard to parse it. It's a combination of both. Definitely hyperscale growth and our architecture growing within that is a key component. But I think we've also been effective at picking up some of the new companies and people that are starting to emerge in AI. I don't know if you have anything.

Yes. I mean we've talked about potentially even 1 or 2 more 10% customers, right? So I think we are definitely seeing a diversity of customers coming to the table. But also like we're very happy like Meta was referenced earlier, right? Like we're very happy with our partnership with our existing 10% customers. We don't name them anymore. They're customer A, customer B, but we still love them, just to say, right? And so yes, it's -- I think business is good on all fronts and not forget the enterprise, right? I know this was more AI focused, but even within the enterprise, we've got customers, like we said earlier, that are starting to be relatively significant, not as large as the hyperscalers, but the significant amount of dollars associated with AI coming out of the enterprise in the nontraditional places that you might imagine.

Great. Well, we're right out of time. Thanks, everyone, for joining us, and thanks very much for joining us on the stage today.

Absolutely. Thanks, Mark.

Thanks, Mark.

Thanks, Daniel. Take care.

Thank you.