Arista Networks Inc (ANET) Earnings Call Transcript & Summary

Great. Thank you. Thanks from everybody. I'm Amy Sutter, Deutsche Bank's TMT Specialist, and I'm very excited to host this discussion with Arista Networks. Obviously, a company that's at the center of cloud networking and increasingly AI infrastructure. And today, I have Mark Foss sitting next to me, who's VP of Global Operations and Marketing; and Hardev Singh, General Manager of Cloud Titans and AI Platforms, very popular topic.

So maybe just to start to get a little bit from both of you. Maybe if you could both just talk about some of the top priorities you're most focused on today. And where have you been spending most of your time? I know it's different.

Personally?

Yes, personally.

First, I wear a lot of hats at Arista, but I'm more focused on the enterprise side of things. And for us, one of our biggest growth areas is the campus market, the enterprise campus market, that's connecting to users. And we're really focused on not only going after large enterprise customers, but one of the focuses that I'm working on right now is enabling channel partners to help us on go-to-market or going after kind of medium-size campuses, and that's kind of a big initiative for us. That's me in a nutshell.

In my role, I work closely with our large customers. So really understanding the requirements, their road maps, the challenges and then really partnering with them to align our road map on products and kind of really been lockstep with them to define our future products. So that's where I spend most of my time.

Yes. And I know we'll go into that more. So I mean on your last earnings call, you raised your revenue guidance by roughly $550 million. And I think investors were very excited about that. I mean how much of that revenue is attributable to these AI cloud titan customers? And maybe just help us understand like your visibility there and the deployment cycles you're seeing.

Yes. We didn't break out exactly what the $550 million was attributed to. I think we said that $50 million of that was kind of our VeloCloud acquisition that we just did. But the other $500 million would be a combination of all of our business be it from a general purpose data center to AI to other use cases, but we didn't exactly break that out. But momentum is pretty strong as we alluded to on our call.

Is visibility different with enterprise versus Cloud Titans, do you think?

It can be. Oftentimes, it depends on -- visibility is oftentimes directly correlated to lead times of the product. And generally, oftentimes, the Cloud Titans will buy our higher-end products, which generally have longer lead times. So generally, you get visibility as to how long your lead times are in there. So I'd say it's more product related than company related. But because the Cloud Titans consume a lot of our high end, we're probably getting a little more visibility from them at this moment, but that can change.

Okay. And maybe just on the AI spend, you've been talking about front end and back end. And maybe there's like historically maybe been a 1:1 ratio, but it seems like the distinction is blurring somewhat. I know I talked -- we did a IR hosted call last week and this is something that a lot of investors were talking about. So maybe if you could talk about that from your perspective, heard of it a little bit. So help us understand when you're working with customers, front end versus back end?

Yes. When we look at back end, it's very crisp and clear. It's the network that's connecting to the GPUs, right? So that's why even the guidance Jayshree or Arista is giving is like we have a clear visibility on that portion of the network that's connecting to the GPUs. Front end, it gets a bit fuzzy. And the reason is, yes, you hear of ratios like 1:1, even 1:2, sometimes 1:3. And the reason is because front end you have connectivity to the general-purpose cloud, you have storage in there, you have connectivity to the WAN, you have some compute there to feed that data into the back end for training or inferencing so -- and then depending on how a vendor is counting that front end, you could have some variability there in fuzzy math, that's why you see that the ratio kind of having this wide band.

And I think for 2026, I think we've said that we're probably only going to give 1 number, we're going to combine the 2. I mean on the front-end side, front end, even our customers have said, "We can't tell the difference" because the same products that go into oftentimes supporting the AI cluster versus the general compute. And our customers say, they don't even know the different. So if the customers don't know the difference than we can't either. That's why we're just going to give kind of 1 guidance number for 2026.

Okay. Great. And then maybe just we've seen large increases in CapEx, especially for Meta. It's been -- had a huge increase in CapEx. And like when you -- like how -- maybe talk a little bit about the relationship with some of your larger customers, what kind of visibility you have into their plans? And then you did -- Hardev talk about like how tight that relationship is. So maybe a little bit -- if you can clarify a little bit like how that relationship works in terms of their planning around their CapEx spend and what you -- what visibility you have and what part, role you have in terms of helping?

My view on the CapEx spend is that from when they announced the CapEx to when it kind of hits vendors like us, there's a lag, right? Most of the CapEx is probably going into securing power sites, construction, and there's a lag there. It could take anywhere from 12 to 18 months before you bring in the network. But working closely with them, the priorities are really around next-gen 200-gig SerDes-based products, 1.6T products, the priorities around reducing power. So what are the other technologies like LPO, NPO, co-packaged copper. So that's where most of the engagement is with them and it's a close partnership.

And from an engineering standpoint, right. And so the visibility can be year out kind of thing?

Yes, can be because these technologies are really kind of cutting edge. Liquid cooling is another very important aspect of technology that we're working on, right? So liquid cooling is there in compute today, right? And I think in the future with these new data centers could be 100% liquid cool. At that point, you have to -- the network needs to be liquid cooled as well. So liquid cooling, 200-gig SerDes-based next-gen,1.6T products, optical interconnect with these new LPO, NPO, CPC technologies, probably the top 3 I would list.

Great. And then maybe, maybe let's talk a little bit about scale out versus scale up networks. And can you maybe just for some investors that don't understand the difference between scale out versus scale up networking model. Could you explain the difference? And then maybe how -- what's your opportunity across both?

Really straightforward. Scale-out is the network that connects hundreds of thousands of these GPU servers or racks, right? That's traditionally been InfiniBand when it started, heavily now Ethernet. Scale-up is the network that is connecting the GPUs within a rack or a server, right? The incumbent player is NVIDIA. We have a proprietary solution, which is NVLink and NVSwitch as the ecosystem of these accelerators increase to, let's say, AMD or custom accelerators from the hyperscalers as that share kind of widens, there is a new opportunity for scale-up networking for Ethernet vendors such as Arista. So it's a new TAM that will open up, but it's still slightly out from a horizon point of view.

Right, right. I mean is -- I think -- so you talk about Ethernet versus InfiniBand and the opportunity being as maybe customers are using custom ASICs or they're using an alternative to NVIDIA. Like is that the main reason then why people would be adopting more Ethernet and scale-up?

Ethernet scales. Ethernet is open.

You can dual source Ethernet too.

The smartest customers don't want proprietary solutions. So -- and now we hear the AI cluster sizes from hundreds of thousands of GPUs to even 1 million, right? You hear some of the cloud titans talk about clusters that could potentially go to 1 million. At that point, Ethernet provides the scale, the openness of the ecosystem. And then it just gets complex, more and more complex from a software development perspective, the features around congestion management, load balancing, visibility, customers really care about that and Ethernet really has been there and continues to solve all those...

Customers in general are just more familiar with Ethernet, right? They know it. InfiniBand, they oftentimes don't have any expertise and they don't want to have the learning curve to do that. And it's also beneficial to have Ethernet everywhere as well versus having to manage an island technology. So it's just -- people just prefer to have the simplicity.

So it sounds like the heterogeneous using different ASICs are alternative to NVIDIA, that's driving some adoption of Ethernet, but then also, there's other things that are driving Ethernet. So we're starting to really hit the knee in the curve, I guess?

Yes. I think InfiniBand was originally -- NVIDIA really kind of bundled the InfiniBand oftentimes with their GPUs and that strategy worked early on, but then customers realize that they have choice and Ethernet interconnect has just the same performance and guess what, I'm familiar with it, and I can manage it easier. So I'm going to move to Ethernet. And they are really large-scale -- there's also a scale limitation on InfiniBand as well. So if you want to go up over 50,000, 100,000 GPUs, you're going to want Ethernet. So that was also a consideration for the large cloud guys, especially who wanted to had aspirations to go above and beyond. They don't want a limitation. There's no limitation on Ethernet.

Yes. I think the debate between Ethernet and InfiniBand is over. I think Ethernet has won there for scale-out. I think the next opportunity for Ethernet is to now go after scale-up.

Okay. Great. Okay. And then maybe one big question. I think that I've heard many times before is just the competitive environment. And the one thing you mentioned Hardev, is like your relationship with your large customers and how you're thinking -- working with them from an engineering standpoint. Maybe if we could talk a little bit about white box then versus your traditional competition with Cisco and now HPE, how that's been evolving.

I can start, Mark. I think the landscape is still the same, Amy. It's not changed. There are roles where it's complex from a network architecture from a product standpoint. And the large customers always want sometimes 2, even 3 vendors, right, in a role. So the competition has always existed, whether it's branded vendors or white box, will continue to be that way, and we continue to participate there, show our value and win a fair share.

Yes. Arista is primarily -- our primary value proposition and why our company exists to begin with and our main pillar for our growth has been our software. Arista is a software company and customers really value that. I mean we're really the only vendor out there that has an operating system that was designed and deployed as a 21st century operating system. These other operating systems from other vendors were either 80s or 1990s architectures, so customers see a lot of value there. But as Hardev mentioned, most of these large guys want a dual source. And they'll take these other vendors as a second source. But oftentimes, they want to -- they want a dual source using the same chip. So they'll use the same -- they'll standardize on 1 Broadcom chip and then they'll use two vendors that use that same chip, but then they'll have different operating systems, of course, and different switches. But as Hardev mentioned, there's not a lot of change in those strategies over the last 6 months.

With the acquisition of Juniper by HPE, have you seen a lot of opportunity?

I think what we saw when there was -- when they were in limbo and there was the -- so that year that went on where HPE announced the intention to announce Juniper wasn't actually finalized. There's a lot of uncertainty in the market there, and uncertainty is not really good thing for customers. They don't really -- they're kind of thinking, well, are these products can be around for the long term or not. And so I think now that that's closed, I think that will definitely help them. But for that 1-year period, where it was in limbo, I think that created a lot of uncertainty in the market.

Great. Yes. Okay. Okay. Maybe to talk a little bit about Broadcom and the relationship there. Just on the Jericho4 chip, how does that enhance the scale-out opportunity for Arista? And what advantages does that give you versus competition?

Broadcom is a great partner. Whenever you look at a new generation of speeds to have the full kind of net portfolio, you want to have a mix of different chips, like Tomahawk and Jericho. In the current 800-gig generation, you can think of the Tomahawk 5, right, and the Jericho3. The Jericho4 now complements the Tomahawk 6. Both those chips are based on 200-gig SerDes. So when customers are going to look to build their, not just AI, their classic data centers, you want to have the same SerDes across. We will have the same kind of architecture and we build the product with the Jericho4, which will suit the AI lead, AI Spine, interconnect, Data Center Interconnect that kind of helps us complete that portfolio around that same speed.

Okay. And with Tomahawk 6, which are you seeing trials right now with that?

Very early, the chip has just started sampling. I think the earliest we could probably foresee products probably second half '26.

Okay. And then maybe just let's talk about co-packaged optics. Just like what do you think about the advantages and disadvantages of that opportunity? And where does Arista stand on CPO?

I want to start by saying that CPO as a technology is not new. It has been there for at least 2 or 3 generations. And the promise of CPO, which is same, by the way, for LPO is how do I bring the power and the cost down for my interconnect, right? With these AI clusters, obviously, now that interconnect is very large, you're talking massive size clusters needs thousands of these switches and then tens of thousands of these optics. That's the promise or that's the goal that these technologies are trying to solve. Where CPO has challenges is around serviceability and the whole operations side of things, right? Pluggable optics, you get different speeds -- a port fails, you replace it. You have different vendors to choose from. CPO, a failure domain will go from one port to the whole switch, right? Like 800 gig versus 51T worth of bandwidth, right? So we continue to evaluate their technology. We still feel confident that LPO or newer technologies like near-packaged optics or co-packaged copper are technologies out there that will help the goal, which is to reduce power and cost. We'll continue to evaluate.

And we're agnostic to what -- if the customer demands it, we'll build it and provide it. So we're pretty much agnostic to the technology. But myself personally and a number of other executives at Arista, we went through this exercise 25 years ago, where it wasn't clear when Gigabit Ethernet was coming out, and it was like, is it co-package Gigabit Ethernet or is a pluggable -- and we didn't know then what it was going to be. And it turned out that for Gigabit Ethernet, at least going on to 10/40/100. Pluggables just blew the doors off of co-packaged, and we'll see if that trend holds true in the situation here. But generally, past can oftentimes be a pretty good indicator of the future. But again, we're agnostic if customers want to buy co-package stuff, we'll provide it.

Okay. I mean, I guess, like if, I mean we're starting to see more and more CPO products come out. Do you just think that -- how long do you think it takes for customers to maybe look at adoption? Is this like a year?

It depends on the customer. I mean, generally, they'll trial things for like 3 or 6 months oftentimes, it depends but...

I think there are two aspects. One is the technology being ready in terms of product offering. And second, to see once it's deployed in some certain scale to see how the TCO performs from operations and serviceability and failures. So I think there's two-step kind of phases there to kind of then come out and say, yes, this technology works and then you would probably see the takeoff of large deploy.

And so when you start to see like successful deployments, that's when Arista would say like, well, maybe we need -- this is what are asking for. So we'll...

Yes, we base everything on what customers want. I'd say, at least a couple of years out.

Okay. You talked about 25 to 30 enterprise and Neocloud customers adopting AI so far. Can we talk about the ramp of those customers? What are some of the main workloads use cases that they're going after?

It's of a wide variety. If I had to say a trend, a lot of these guys are providing kind of AI as a service model, it'd be like maybe a startup or someone who's just wanting to build out an AI cluster that they could provide -- that they can actually charge for a service. That's one model I'm seeing, Hardev probably see others as well.

Yes. GPU as a Service, sometimes building capacity in certain regions for large customers, sovereign clouds, maybe government mandated to -- where you have a customer already for what you're going to build, so lot of activity there. The architecture is very similar to, let's say, the large customers or the cloud titans, just that the scale is smaller. We see a lot of inferencing use cases. We see diversity of GPUs. It's a fun space, and it's exciting.

What are some of the challenges that you see? Is it maybe because they're not the cloud titans that are so experienced with large data centers or perhaps funding? Or just curious what some of the challenges that you might be seeing? And does it offer opportunities to really go in there and help from an engineering standpoint that might be differentiated?

One thing that stands out, and Hardev probably has a lot more information than I do, is the ability to get power. So they have -- there was one large -- one that we talked about on our conference calls previously that they were had an ability to get power and they ended up getting some delayed fundings. That's a big issue. There's a lot of other...

Yes. And these customers also care about the performance out of their network, right? I mean these are large investments and to get the best performance out of the GPU they really want the best-of-breed network. And I think they're looking at Arista to provide that capability, that differentiation, the software stack that we have and the bunch of features we've developed around load balancing, telemetry and then they view us as, okay, these guys are participating and running one of the largest some of the largest AI clusters and we'll get a fair share on the table to participate on those opportunities. And if there's some places you will see bundling and stuff and those opportunities we don't participate in, but otherwise we do.

Okay. I was going to ask where are we in the upgrade cycle to higher speed ports like 100-gig?

There's always many upgrade cycles going on. They never end. But I guess, starting on the lower speed. I mean, the transition to gigabit to 10-gig is almost done. There's very little gigabit going on anymore. There is a transition going on from 10 to 25 currently, but it's still 10 gig is still at least 4x the size of 25 gig, but that's happening. Then there's the 40-gig to 100-gig transition, which is complete. There's pretty much 40-gig is now pretty much dead. There's no deployments going on for 40-gig. 200-gig is, there's primarily one customer in the world that's deploying a lot of 200-gig. There's not a lot of other customers beyond this one customer, which is only 200-gig, that's kind of a one-off for 200-gig. But then on the higher speeds, we're definitely seeing some -- 400-gig, I think, is fairly close to 100-gig. There was a transition from 100-gig to 400-gig. And that's fairly close now. I think you're starting to see 100-gig peak a little bit. 400-gig is starting to, I think, have more momentum there. And then the one that's growing the fastest right now is 800-gig. 800-gig is only been shipping for like a year and pretty much all of the new AI -- if someone's bidding on a new AI project, it's always 800-gig. So 800-gig is growing the fastest, but it's the newest and the smallest. But anyway, there's a lot of dynamics going on there.

And the trend I see, again, with the large customers is that all this traffic from AI is having a knock-on positive effect on the classic, if you will, data center, right? So it's forcing some of the speed upgrades to accelerate.

Okay. Okay. Yes. I guess also then to talk about how important your observability in zero trust security solutions are for customers in AI use cases?

Yes. Security is very important, but observability is key. I think once going back to the same point, as these cluster sizes become large, having good telemetry, good visibility into where I'm having network issues, with the size of clusters things are going to fail, right? Meta put out a paper where they're like the biggest link flaps or ports coming down was due to optics failures, right, where due to HBM memory corruption errors on the GPUs. So having visibility there and then to quickly get the ports back up, really reduces the overall job completion time for them, which is a direct correlation to revenue for them. So they really care about visibility, observability. And that's where with Ethernet and Arista, we have a great portfolio of features in our software stack to address those challenges.

So is it typically then where observability always comes from the networking aspect of the...

Yes. Networking plays a very key role. And CloudVision, which is our management platform provides that visibility for the customers, not only in the network, we go all the way down to the NIC on the GPU. So you get an end-to-end visibility of the AI cluster to the interconnect to the backbone in one dashboard which is CloudVision.

Right. Okay. So maybe let's talk a little bit about sovereign cloud and service providers. So there was no sovereign cloud contribution expected in 2025. I mean how do we see that market developing? And what do you think helps gain share?

Definitely sovereign cloud opportunities out there, and we're involved in those. I think if you compare the sovereign cloud spend to what like a cloud titan were to spend on AI, it kind of dwarfs it. So I think it's not about the number of opportunities, more of like what the scale is. Generally, I don't think there's a sovereign cloud that I've heard of that's planning on going over 100,000 GPUs, but there is a number of others that are doing that. So I think it's not about the number, but it's more about the size.

I think you said no contribution from Arista from sovereign clouds in 2025. That's not true. In fact, if you see -- when Jayshree spoke about our -- we started with 5 customers to 4, we still very confident on our $750 million number and a lot of that is kind of coming from these 25 to 30 other customers out of the top 4 that we talk about and where sovereign cloud, a part of that.

Okay. Yes. And then maybe on service providers, which parts of your product portfolio are best aligned with their needs from what you're seeing today?

Service provider side, it's kind of our fifth out of 5 verticals that we do. We generally do very well in the service provider space if they're building kind of a data center architecture to support their infrastructure, and we do very well there. They also use us for routing use cases. And oftentimes, they'll use us for like an MSP service. That's where the VeloCloud stuff comes in. We'll oftentimes use our WiFi sometimes for a service as well. But there's a variety of use cases. You had something Hardev?

Yes. The classic service provider, in my view, is slightly on the decline from a TAM perspective. But I think our portfolio of products with our 7,800 routers, both fixed and chassis form factor play well into that space, and we continue to compete in that market.

How is market share for -- and you continue to gain share like -- primary share?

Do we break it down and talk about...

I think it's broken down by region or by, yes, customer type.

And then, I think -- are you seeing, I think tariff uncertainty has kind of come and gone. But are you seeing any impact from a customer buying perspective, is it still part of the conversation with customers?

It can be because it's a very fluid situation. We -- obviously, there's a lot of uncertainty 6 months ago, and that's one of the reasons why we didn't change our full year guidance for last or 2 quarters versus last quarter because we were uncertain what's going to happen in tariffs. Fortunately, we are covered under the U.S. MCA exemptions and under the communication equipment exemptions for tariffs right now. So our tariff impact has been material, so we haven't had to make any changes there, which has been good. But as we've seen, the tariff situation can change very quickly. So we're definitely watching it closely. And customers will ask about do we see any risk there. But again, we just say, what I just told you, it's very fluid, and we're just -- we'll continue to react when it happens.

And how is the supply chain? I mean we've come in and out of an inventory situation also. I mean, is it pretty not having an impact in terms of delivering product to customers. Maybe it's -- I think from a cloud titan perspective, you might have talk about like their power issue is like their bottlenecks. But it doesn't seem like there's many supply chain issues.

The supply chain continues to get better post-COVID. I think one of the big changes is that the chips that we use, those -- the lead times have come down on that, but they haven't come down to post to pre-COVID levels yet. I think that's kind of one aspect of the supply chain which is still an issue for us. But supply chain is never over. There's always some kind of a shortage or there's always some kind of a surplus of things, and it's always something we work on a daily basis.

Yes. I think we are well positioned from a manufacturing supply chain in terms of diversity. We have contract manufacturing in Asia, Malaysia, Vietnam and then Mexico. And then we use 3 or 4 different contract manufacturers, and there's fungibility there in terms of product. So I would say we are well positioned to address any kind of foreseen increases in demand.

I want to check if there's any questions from the audience, just in case. We do have a microphone. Okay. And then maybe on the margin side, is there a big -- how should we think about AI product margins? Is it more dependent on the customer or the product?

Definitely, the customer. Generally, Arista does volume pricing. So the largest customers could have the most spend at the best prices. And right now, AI is a pretty concentrated market. It's generally the largest customers that are buying a lot of AI. So yes, it's definitely customer versus product.

Okay. And then maybe you mentioned this -- we'll talk about the campus opportunity.

Yes. My favorite subject.

Yes. So you raised your target, kind of maybe talk about what's driving that increase there?

Yes. For us, the $50 million increase was really because of our VeloCloud acquisition that we completed July 1. So that extra $50 million was really because Velo really rounds out our branch solution. We have the PoE switches and the WiFi, which can go in the branch, but the WAN optimization for the branch was a piece that was really missing and Velo really fits in there and really smooth that out for us. But the campus is just -- I'm so excited about it because in 2025, some of the analysts were saying that data center is a $30 billion TAM, but they're also saying that campus is also a $30 billion TAM. They're the same size of TAM. However, from a market share perspective, Arista is small, low single digits right now in terms of share of campus, while we're kind of 30% share in data centers. So we have a bigger runway, I guess, growth -- multiyear runway for growth in terms of campus. So it's an area which people don't talk about a lot, but it's an area of our biggest TAM. So when people ask about it, I get really excited about it.

I guess like in terms of market share opportunities, how do you see that playing out? Is it a go-to-market as well or if it's VeloCloud also maybe accelerate the opportunity to gain share, please expand on that.

Yes. So VeloCloud definitely helps us. I would say that but half of the TAM of campus market is large enterprise, meaning Fortune 500 Global 2000. And we focus on direct touch to the end customer for those opportunities, which we're really good at because the same guys that sell data center, which is great. On the small side of things, we don't really focus on that market, which is maybe 20% or 25% of the market. But the other 25% of the market, which is medium business, that's an area of focus for us as well, but we're -- our go-to-market is more on the channel side. So one of my big jobs right now is to focus on enabling channel partners to help enable them to bring business to us kind of on the medium business side of campus. And that's really big because there's essentially 1 million campuses on planet earth. So your sales force can't scale to have that coverage. You need channel partners on the ground to help you. And that's been a big focus of ours.

And I guess, what would you say on the relationship with the channel? Like where are you kind of on a trajectory of like first inning to ninth inning, in terms of those relationships and seeing the benefit in market share?

Yes. No, for channel enablement, it's still -- I can still fairly early. I think a big chunk of our business is -- the channel oftentimes plays a fulfillment role for a large enterprise. But our focus is to get them to do the heavy lifting and enable them to go after the mid-market and be very independent. And we're having some success there, but it's -- I think it's still early. And you can tell by our share on a number basis, there's a lot of room to go, and that's a big opportunity for us.

Great. Well, thank you. I know we're getting kind of close to the end, but I know you have the Analyst Day coming up. Anything you want to leave us with to entice to be sure we're...

Well, there'll definitely be some technology updates, how we differentiate in the market. And then maybe some insights into next year guide or...

Yes. Probably we'll probably talk about 2026. I think there's a lot of questions on blue box. I know for a fact that we're going to be talking a lot about blue box and go in depth on that. So that will be exciting.

Yes.

Great. Well, thank you guys so much for taking time today.

Thank you. Thank you for having us.

Thank you.

Thank you.