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

Everyone, welcome to day 3 and the final session of Citi's Global TMT Conference. We have definitely saved the best for the last. My name is Atif Malik. I cover U.S. semiconductors and networking equipment names. It's my pleasure to welcome Chantelle Breithaupt Chief Financial Officer; as well as Tyson Lamoreaux, Senior Vice President, Cloud and AI Networking at Arista. Welcome, guys.

Thanks so much. Thanks for having us.

Thanks.

I'll kick it off with my questions first. If you have a question, save it towards the end, we'll send the mic and you ask your question. Chantelle, deferred revenue has been a hot topic this year on Arista. Arista consistently achieves strong revenue growth has guided to 40% growth this year, and yet the complexities of AI use cases have resulted in strong growth in Arista's product deferred revenue. What is the right way for investors to think about Arista's growth? Is it the growth in the revenues plus the change in deferred revenues?

Yes. I think it's a great question. We're super excited to in the last earnings call, raised in our revenue guidance to be 40% growth, $12.6 billion. So we're very excited to be able to deliver that kind of guidance. And so if we think about it, deferred revenue is a topic. I think that if we just level set what deferred revenue is, if we're talking about the product deferred revenue, that means equipment that's been shipped, invoiced and cash collected, but we have acceptance criteria that keeps us kind of skimming the game with the customer to get to the kind of net results they want to get to. And if you think about what goes into here, it used to be the cloud, if you go back a couple of years to the cloud deployments were put in there and now AI has basically put that on steroids from a complexity to your point. And so from a -- from a how to interpret it perspective, I think if you look at Arista, there's a few things I would keep my eye on. I would keep my eye on P&L growth, what's happening in deferred revenue, but I wouldn't do a quarter-over-quarter kind of thing. I would do a trend over 4 quarters, maybe 6 quarters, because things will come in and out generally because it's an acceptance based criteria. I would look at the purchase commitments and maybe what's sitting on the RPO table and all four of those things kind of give you a picture of what's happening and what's to come. I think that from the perspective of deferred revenue as well. And sometimes that people don't understand that some of these deployments can take 18 to 24 months until we hit the acceptance criteria. So it's important to understand the timing of how long this could be and when we would realize it from a P&L perspective.

Great. Chantelle, just for perspective, how far are we into this current cycle, the deferred revenue? And how does it compare to what Arista experienced in prior cycles?

Yes. So even though we're now going to be part of the S&P 100, we're only about 12 years old as a public company, right? And so we've had, I would say, two cycles, the cloud cycle and the AI cycle. And so what you're referring to, Atif, is during the cloud cycle, which kind of the 2019, 2020, '21, you saw deferred revenue raise because we had this deferred construct because we had the use case of cloud. 400 gig and then coming into 800 gig. And then we kind of went through a little bit of cloud and then AI took off. So this AI cycle is the second. The cloud one was about 3 years. We're in, I would say, going into year 2.5 to 3 on AI proper from a materiality perspective. But I think there's many years to come on this cycle. And so we'll have to see because we're talking about scale up, scale out, scale across training, inference, front end, back end. We don't see AI going away anytime soon. So this cycle could be longer.

All right. With that topic out of the way the next big topic has been the supply environment. Supply obviously remains challenging. And on your last results, you came out a lot strongly on supply. Where are you seeing the most pressure in the supply chain?

Yes. So we started the year, and we'll -- I'm very transparent. Q1 was not our favorite call because we did have to talk about the supply chain environment and I think that's what you're referring to. But we wanted to be honest and transparent. A lot of this -- A lot of this pressure is coming from right at the top of the fab capacity. And I think anyone in the industry should be talking about there is tightness no matter where you sit in that hierarchy. But as we work through to the Q2 call coming into August, we did find some vendor arrangements with fab capacity, memory. So those things, I think we've sorted through '26 and '27. But you have whack-a-mole on some of the other component parts, PCBs, capacitors, et cetera. So there's things that are more peripheral short term, and then there's structural things like fab capacity. We feel very well positioned. When we talk about purchase commitments now, that's a year lead time for chips going into Q3 next year. So you were already talking that $9.6 billion at the Q2 and purchase commitments is already going into second half next year. So well positioned, but we're -- I don't think anyone is out of the woods when it comes to the full supply chain for networking data centers generally.

And can you remind us any purchase commitment you made to secure the supply?

So we ended -- so we've grown -- if you look like 3 quarters ago, our purchase commitments were about $3.6 billion, and now we're at $9.6 billion at the end of Q2. So over 3 quarters, basically tripling it. And that's leaning into demand we're seeing. So you should see it as a demand. It's getting some of these memory things sorted. And then there are some things that are a little bit longer than 12 months. But from that perspective, we feel very good that the supply chain team learned a lot during COVID. We learned a lot during the beginning of AI, got the contract manufacturing components sorted. So I think it's a demand signal at this point and we feel very confident about that.

Tyson, a lot has been said about the differences between AI traffic and AI networks versus classical cloud computing network and Arista has developed its Etherlink portfolio to address the opportunity. What commonalities are there between Arista's classic cloud solutions and hardware and software versus the AI solutions?

Yes. I think Chantelle and I were talking about this a little bit earlier, too. And Arista's like heritage and foundation we believe, really put us in a very well-suited position for the AI era. And in a way, the company was born and bred to be prepared for the AI era. And so there are clear distinctions between cloud, traffic, architectures, hierarchies, oversubscription, interface speed, upgrade cycles, refreshes and how the compute is interconnected. AI is much more demanding. The traffic profiles are significant. They're very highly coordinated. The reliability is essential at every level of the piece of the infrastructure because you don't have the ability, at least with current generation AI technology to provide ways of working around failures as easily as you do in cloud-type primitives. So the demands are just very, very high here. And I think the commonality for us from the kind of where the company forged its foundation was rooted in reliability, security, performance and scalability. All of those dimensions are more critical for AI than they were even in cloud. I would highlight too, in particular, that kind of route around our heritage and hardware development, the ability to deal with list generation, intercept latest generation technology, overcome very hard electrical engineering problems, signal integrity problems, power integrity issues, things that are exacerbated as you move up into faster link speeds as we're now kind of transitioning to the 1.6T generation. These are very difficult challenges. You're emitting a lot of heat because you're consuming so much power, so your ability to cool that and be power-efficient is super critical. Something Arista has really kind of built its heritage again in hardware engineering around. So that lends itself very well for our Etherlink platform and AI Fabrics. And then the software is another piece on the reliability side and the security side that's essential. One of the things that's come to front of mind for a lot of people is -- last year really is Mythos and Glasswing. And the notion of AI driven attack vectors and an ever-increasing rate of security vulnerability, discovery, fixes, patching and the cycle that needs to basically run continuously now to keep your network secure, One of the benefits of our software is that it was built on a foundation of always staying up so you can patch the software and deploy it and you can automate the entire workflow seamlessly so that your data plane is never interrupted. The work stays up, traffic continues to transmit across the network. That gives us a real relative advantage against competitors as well as open NOS and other type of solutions. So we feel very good about that foundation that we've been able to carry forward into the AI era. And get leverage out of it. And then the reliability aspect I mean if you're talking about reliability that is a full point better than the next best in the history, that kind of 1% improved availability translates in a very significant way when you start talking about amortizing availability across all of your accelerators, which are the lifeblood of AI, right? This is the biggest expense. This is where you're spending the most money on. You don't want that capital sitting around idle. You want putting to work, you want your high utilization. So the speeds and feeds, cutting-edge technology, reliability, security story come together very, very well. So served us well in cloud, I think it's serving us even better in AI.

Great. Historically speaking, you guys had concentration at two hyperscalers. And you've done a very nice job in diversification at the hyperscalers with potentially three or four this year, depending on and how things shape but the hyperscaler concentration is improving. On the non-hyperscaler side, the neoscalers and enterprise, can you talk about your adoption cycle over there?

Yes, we're very about the other non-hyperscaler environment. So we talked about we do have two, 10% customers that you were referring to. We do anticipate a third and potentially a fourth. So that diversification continues. From a Nelcloud perspective, one thing back to Tyson's commentary, the neoclouds are looking for a few things. What makes them competitively advantage is their token cost optimization and so they come to us to say you have great experience both on AI, front end and back end and cloud. Can you help us get to token optimization as quickly as possible with your fast portfolio. And so all the different architecture designs help them get to their token cost optimization. If it's a fair fight in the sense of it's a best-of-breed conversation, and it doesn't come down to commercial arrangements, taking the customer [indiscernible] different avenue, we have a very good chance of winning. Even some of the neocloud customers who have gone with a competitor, perhaps for different reasons, have come to us perhaps for their second installation as they realize the best of breed will get them to those token cost economics, which is how they stay in business. We're very excited in both local and international growth from that perspective. And enterprise, enterprise, I think, is just getting started. We have about 20%, 22% market share in data centers. So even if we just continue doing classic data center, there's market share to go gain but why else would enterprises to accelerate their next refresh for us. Back to the Mythos environment. If you think about where some of these enterprise players are they are very nervous about if they're currently with many different operating systems, how they're going to get on this patching treadmill when it comes to making sure there's care. So we've had customers come to us that are new logos ahead of their refresh to say we're very interested in your EOS because we only want to patch once, and we want to make sure that it's done properly because we have the lowest CBEs in the industry by a magnitude of 30x to 1x. And then the other reason the enterprise is coming to us is just total cost of ownership. With one EOS with CloudVision, they can get rid of the 50 to 60 to 70 people, perhaps 200, they employ just always be constantly updating, taking the system down, repatching. So from a TCO perspective, enterprise is also very excited. So market share and TCO is a conversation there. And then you go to campus of the enterprise. Campus, we have 5% market share going from $800 million, '25 to $1.25 billion in '26. That growth rate is definitely larger than the market growth rate. So that's a share gain story, and that will be a long, slow, rinse, repeat, that we're very excited. And we win campus first deals, which is a great validation of our portfolio and our brand recognition. And so there's land and expand going back to the data center, the underway. So there's lots to be excited about, I think, in those two categories.

Security seems to be hit a nice flywheel for you guys. Scale across AI applications have emerged rapidly and are expected to be approximately 1/3 of AI mix. What is driving the growth in scale-across and what could shift the scale-across mix higher?

Do you want to start?

Sure. Yes. I mean, I think I'll start probably with the back end of that, like what could drive it higher constraints are driving a lot of this in general, and there's some architectural choices and there's some desire to distribute data centers either because you can't get the power and data center space in a single facility that you'd like or you want some resiliency kind of built into your infrastructure, so you choose to distribute. But another big driver here is actually coming around not just driven by constraint but in force constraints, regulatory regime change and like top-down government initiatives. So I'd highlight like a good use case that really exemplifies it is, if you look at South Korea, the government there is actually driving by policy, distributed generation and local consumption of power. And so by its very nature, what's going to happen to the data center infrastructure inside of South Korea is, it's not going to concentrate into a small set of locations or cities where all the capacity is kind of all your eggs are in one basket, its actually going to distribute around the country, and naturally, that is going to drive much more demand for scale-across solution because your single largest facility is going to be capped effectively based on the local power available to it. And so as things federate, you naturally have to interconnect them and interconnect those clusters. That is something that can shift and change and shape the scale across TAM quite a lot. I think I'll let Chantelle talk more broadly about the TAM and how we think about it.

Yes, I'd be happy to. So we are revisiting our TAM generally, like at the last Analyst Day in October '25, we gave you a $105 billion dam. A lot has changed even since less than 1 year ago. So we'll be revisiting scale up, scale across and hopefully come back to you later this year with a revised TAM. So we see that expanding for all the reasons Tyson listed. But also to provide a framework perhaps to help the audience just how the scale-across kind of gets us to Arista being a good choice. We're very well positioned. The customers, we feel, are looking for a combination of great software operating system, hardware and AI experience because this isn't just a DCI. This is a DCI with a much more complicated structure and high large clusters, right? So if you have AI experience, software and hardware, I think that's what wins in scale across it. And we don't feel our competitors have all three. Some might have DCI, some might have AI clusters, but we don't see [indiscernible] to have all three. And that's why we feel incredibly well positioned for scale across. And so wait to see on the TAM. But if the TAM increases, we were very excited about that.

Great. And then on the scale up, the market has been waiting for Ethernet standards, but a consensus is emerging that scale up is a huge opportunity. Broadcom talks about Ethernet being a leading candidate. How do you see the mix between scale out, scale across and scale up evolving?

Well, I can talk about the mix and then maybe we could talk a little bit about scale [indiscernible], perhaps your questions will lead us there. So a lot of this depends on what timing we're talking about. So from scale up, we have 0 in our guidance and 0 revenue for Arista for scale up in '26. We feel that we'll start to get some kind of maybe trials and pilots similar to our InfiniBand Ethernet conversation later next year because the ecosystem has to come together. This is Ethernet scale-up is going to take some ecosystem parts coming together because it's a bit more complicated. So we're talking '28 is when we get the revenue there. So that's why I say the percentage of will depend on the timing. Scale across, I imagine will happen a little sooner because we'll see that '26, '27 and '28. So both will have great growth trajectories, but the timing of it, it makes it difficult to answer your relative perspective. We're hoping all three are running very well for us. We get to '28 when all three are online, let's say. But we'll have to see the timing of all those three, how they show up.

And then, Tyson, can you talk to the XPU road map? And when do you guys intercept that in terms of revenues and how does that fit in this whole kind of NPO CPO road map?

Yes. Yes. We feel very good about the like road map in general for our products and the base building blocks that we have in place for advancing our product portfolio, kind of generation over generation in [indiscernible] NPO and CPO. I mean we are customer-driven in terms of our product road map. And so if customers want a CPO switch, they want an NPO switch, they have certain requirements that need to be fit. We're going to go build that switch for them. I think, you've got some alignment right now with XPO and kind of the scale-up timing. There's some coincidence to that like but they're also not just purely time coincidental. There is form factor and bandwidth density that comes with scale up. And so as the Ethernet standards around ESUN or ratified, you get chip production, chip production leads to the ability to build a system. You're talking about very dense systems here. We're going to build a scale-up solutions to kind of focus on customers who would like either co-packaged copper for the lowest possible cost, but they're going to be constrained in terms of their reach because you've got a bandwidth kind of over distance problem, the longer your link is less bandwidth you can attain over it. That gives rise very nicely to an NPO solution, which is what our XPO architecture is focused on. The ecosystem is coming together very well. The module makers are building all kinds of products, and they're going to come together very well in that time frame. It's a portfolio or a product that we think it fits very nicely in our portfolio through optionality, serviceability in the field, things that our customers are telling us they want. It's also going to help them start to get into that multi-rack high density scaling that they require that copper isn't going to reach into. So we see this as a nice kind of two, three generation solution that lands in between kind of copper running out of steam and when CPO can really be built at scale. But we expect some early adoption on CPO for sure. we're ready to build as soon as customers are telling us that they want it and need it. I think you'll follow -- you'll find a pretty common adoption there, like early adopters really drive the front of it. There's a whole ecosystem that has to be built out. It's a completely different manufacturing technology and architecture. The supply chain has to come together for it. Yields are going to be very challenging and very early on. So there's a lot of customers that are being pretty pragmatic about it in terms of their expectations of when they can intercept and what value it's going to create and when and when the costs are going to stabilize and the volumes are going to increase. So we like the road map we have right now, which is this continued copper as long as we can, bring NPO and XPO into the fray, build out a lot of the modularity in that generation, you're able to then feed into CPO and then CPO coming on. So I think XPO, early deployments second half next year really ramping in '28 and probably 12 months behind that is where we start to see some of that CPO ramp occur.

And do you guys have a view in terms of the mix of ESUN versus UL-Link, like where is this kind of market going?