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

Welcome to the Rosenblatt Age of AI Tech Conference, fireside chat with Arista Networks. I am Mike Genovese, the cloud and communications equipment analyst. Super happy to be joined by some of the management from one of the very best companies out there. We have the Chief Financial Officer, Chantelle Breithaupt; and Brendan Gibbs, Vice President of Product Line Management. Nice to see you both.

Yes. Thank you. Thanks for opportunity.

Great. And so I'm going to ask questions. We've got 45 minutes, but I also want to let the audience know that on the audience screens, there's a widget in the upper right-hand corner, where you can type in questions that will come to me, and I will ask the team here any questions that I get from the audience. So please feel free to send them in.

But let's get started. So we're going to talk about AI, obviously. And Etherlink, which is the Arista's family of AI-optimized networking, high-performance platforms with advanced software has gone from only 4 or 5 customers in 2024 to more than 100 customers today. Can you kind of break that down more for us by customer type, whether cloud titans, AI labs, neoclouds, sovereign enterprise. And where is the incremental kind of AI fabric revenue coming from going forward versus the first half of the year?

Yes, sure. I can start off, and Brendan will tag team these as we go through and make it conversational. So let me just take a step back. Like how proud we are to say, hey, we had the Etherlink portfolio, June 2024. I remember our 6 months in, we're at our New York Stock Exchange event. And not only did we have a great kind of analyst conversation, but we had a great product innovation reveal. And so I think it's just indicative of how we feel Arista was almost born for this AI moment. First, we had the cloud and now we have AI. So even take the Etherlink portfolio you referred to, Mike, what a great robust set of products unmatched by any of our competitors. We give our all the different customers you mentioned segments, options for their AI journey at a minimum. So we're super thrilled. So now to get to how are we getting to this $12.6 billion, 40% revenue guide that we're very excited about. It's coming from all these sectors. The thing is that we are seeing -- if you're talking about like what's changing in the second half versus the first half, it's more of the same or even versus last year. It's just more of, more Neocloud, more Titan, more AI labs, more enterprise, more campus, although campus isn't quite AI yet, but just generally in the $12.6 billion. And so those first few customers, we were talking at that time of InfiniBand and now we're only talking Ethernet. So the fact we've been able to go from those large customers to a breadth of customers, I think is super exciting in showing that for a pure-play networking company, we can serve any of those use cases and continue to drive the innovation. Going into next year, we'll talk more about scale out, scale across and scale up, and Brendan can give you more of that through your questions. So we're just super excited whether it's trading in inference, front end and back end, scale up, scale out, scale across. We're going to meet the customers where they need and really help them achieve their AI outcomes.

Great. So you touched on it a little bit, but just kind of flesh out more how you would characterize demand right now. Is this a really true -- I mean, it feels to me like it is, right, a genuine expansion of deployment and plans and needs for the network and the data centers. But we also have this supply scarcity issue. And in the pandemic, I think maybe that great if we look back on it for certain companies created pull forwards. And so is any of that going on? And how do you kind of as the CFO look at your dashboard and kind of tell the difference between expanding demand and just pull forward because we're worried about supply?

Yes. I think that mostly it's from the customer engagement, I think, Mike, because for -- if you think about -- we mostly serve large enterprise and larger, right, than that, so you're talking large enterprise, neoclouds, AI labs titans. So for the majority, we've been having ongoing conversations. And so we would get a sense of if they felt they're pulling it in because of supply. I don't think that's the environment. I do think sometimes people are pulling in early because they want to make sure that they come to us as a company to help them in AI environment, so they might pull forward their enterprise renewal because of methods, they might pull forward their scale across because they just have to get to outcomes faster we don't see pull forward at all, but if we did, it would be more from a demand expectation. We don't see it on the supply scarcity. So my dashboard is customer conversations. It's project evolution. It's in the sense of its 52-week lead times. So I think customers are going to lean into 52 weeks unless they're pretty certain that it's something. I don't think it's very easy to pull forward with those kind of lead times. And so we're very clear and very excited about the demand environment. And even this year in the guide, if we can get a bit more supply, we could even go a little higher than the guide.

Great. So I mean, look, I've covered the stock for a long time, longer than just the past couple of years. But I've gone through my own journey with Arista and AI in the last couple of years because when we were just first modeling and talking about the back end, and I was looking at Ethernet becoming like InfiniBand and I thought like -- but where is the value of the software here? And does that change the story. But I don't know if it's front end versus back end and inference versus modeling or if it's actually -- there's a different explanation for this. But the value of the software portfolio and the features and the new stuff you put out has persisted and changed my mind and maybe more of a believer in the last year or so. So can you talk about kind of use cases for software versus Blue Box? And it seems like they're both growing in the company, and there's lots of cases for both, but I'm talking too much. I want to hear this from you.

Brendan, do you want to start?

Yes, absolutely. I mean first of all, I'm mindful of what Chantelle said, we announced our Etherlink portfolio back in June of 2024. That was 800 gig only that was scale out. It seemed so long ago at this point but the portfolio has significantly blossomed the use cases have expanded since then. What we found is that EOS or operating system software has such a attraction for our customers, primarily because of like the most basic value proposition, which is speed with quality. AI is essentially an arms race. All of these customers are racing to get to market, racing to sign up customers racing to innovate at the frontier of models and cloud expansion and GPU as a service, et cetera. So having that software foundation for the network from us really gives them peace of mind and get some quicker time to deployment, time to first token and they just stay up and active. So that's the first level highest level deployment value. What we've seen is that the -- not only the quality as I'm saying, but also the intelligence built into the software, the analytics that we embed to give customers insight to what the heck is going on in their clusters that pays off in dividends over time so we can unlock the greatest performance, the greatest GPU utilization and the greatest insights to what's happening in their clusters. So as they've gone from the initial small highlights of scale-out clusters and they've made them large, EOS has grown with them. And they've -- we've been able to unlock that speed with quality paradigm for them, so they can drive the scale of their clusters a pace. And then as they've found themselves restricted by one physical data center, and they need to scale across the multiple data centers and AI centers. We've grown with them there, too. And there's incrementally even more value than the EOS operating system software for them for that with routing intelligence, with wire speed encryption with more insights across a wide area. So the expansion of the use cases has fit with expansion of our portfolio from 800 gig to 1.6T. And now with more and more intelligence along the way, EOS has remained relevant and even more so over time.

Yes. I just don't want to double-click there on scale across because it is turning into such a big market. And obviously, things like 800G ZR pluggables and the optical line systems and now coming the multi-rail amplifiers, right? Those are supplied by the optical companies. And you kind of had this impression well, these ZRs will just plug into available switching and routing for companies like Arista. But I guess there isn't -- there aren't just ports lying around ready, they actually need to deploy, it seems like many new switches to do this scale across. Is that -- am I thinking about that correctly?

Well, in some cases, yes. So if I think about a scale-out cluster because that's where it's run from going back in time and it's not that long ago, you've had these modular high-end spine platforms with tons of ports. And that's one of the hallmarks of differentiation of Arista is these high port radix modular platforms. In the old world for scale out, all the ports just went south at the leaf nodes. But as the data center gets large across multiple geographically dispersed data centers and scale across comes into play, now you just turn some of the port sideways, so to speak, to go for data center interconnect. And so what you have is instead of a spine remaining local, your spine becomes geographically dispersed. And so in some cases, yes, you're right, you need to deploy new platforms. In some cases, it's taking advantage of the high port capacity to maybe plug some of the ports in with ZR and ZR+ coherent DWDM pluggables. So the great thing about us is that we work with a broad ecosystem of pluggable providers we offer the true differentiation for that scale across domain of the encryption-based hardware platforms with all the routing software intelligence and diversity of optical plugs whether it be from any of the coherent providers, we can work with any and all of them.

And does EOS is important for that entire now network that's one domain but across multiple data centers?

It's important and more so than ever, absolutely because now you need to arbitrate whether a workload stays local or remote. And if it goes remote, how do you properly preserve and prioritize the transmission so that you don't have any sort of congestion. It's difficult but easier to manage congestion and have no packet loss locally, but now if you're going across hundreds or thousands of kilometers, it's even more difficult. And if any problem comes up, it's going to be seen immediately. So the routing intelligence and analytics becomes even more paramount, even more complicated, but we're -- we've risen to the challenge, and we've got scale across now at about 1/3 of our business is part of the most recent guide.

Right. So that's getting to the point, which is this is creating incremental demand for switches. But I mean it sounds like -- I mean, you need to use up the existing ports or anything you have left, but then put in new -- put in new switch.

Yes. And we're identifying it as a distinct use case. And for sure, it is. But also another way to look at this is just gigantic clusters. What used to be smaller, inside of one physical building is now gigantic, like I say, where the spine becomes distributed across many data centers. So we identify that as a scale-out cluster or a bit large across many sites. But yes, it's a new use case. It's expansion. It's more ports, more systems, more software intelligence and more value.

Yes. So when this -- when we first -- on the street, started talking about this market, which wasn't very long ago, right? I think that the vendors are trying to jump to the top of the mind that are really going to benefit here on the optical side, Ciena and on the switching and routing side, Arista. Now we've also seen on the optical side that it's helping Nokia, and it's helping Cisco as well. And -- but on the routing and switching side, in terms of actually moving the numbers it seems that it's really working for you guys, maybe a little bit at Cisco, but I mean, I guess my question is, a, do you think this helps all of the routing and switching vendors? Or is it somehow more particularly focused on you because of the 3 major hyperscalers who are ahead on scale-out are your customers, right, to long-time customers, I think 1 newer customer? I mean, is this something that's going to help all switching and writing vendors? Or does it so somehow more share of it go to you and you get more than your share?

No, it's definitely a scenario where not all vendors are created equal in this context because before you can take a data center and an AI deployment and scale it across, you actually need to win first that back-end scale-out cluster. And Arista is one of the most successful vendors deploying large-scale multi-customer back-end trending clusters period. And so when it comes time to take that cluster and expand it across multiple geographically dispersed sites with scale-across. That's absolutely within our bailiwick. It's more than just putting on a cap that says scale across on the top and saying, "I've got a routing box too." I've got a coherent pluggable like I'm a scale across vendor because they didn't win the scale across deployments -- or sorry, scale-out deployments to begin with. So it's not just scale across is a perimeter routing, one box or one DWDM transport node, it's the whole cluster writ large. So there's very few vendors who have had those type of large-scale back-end scale-out clusters and Arista is one of them, especially from an Ethernet perspective, we're absolutely leader. So yes, it's not a all vendors are in an equal scenario. This is very much advantages us because of our success in scale-out and now our intelligence and capabilities for scale across it just becomes natural.

Great. I guess, as long as we're talking about these different scale out, which is what you've always done. The scale across is already moving the needle. And I just need more help understanding we have scaled up for switching and routing, right? We've done a lot of work on what does it mean for optical and the optical content that goes in and replaces the current electrical content in these racks. I don't really yet understand ESUN and what that opens up for you guys. And I know that it's not happening tomorrow. I mean, pretty soon, but not this year. So just help me understand not just the timing, but actually what it is. And when it does happen, how big of an opportunity it is?

So I can take a stab and then Chantelle, I'd love for you to jump in too.

Yes, go ahead.

At a high level, there's multiple industry analysts, but I've seen it in as high as like a $20 billion market just for Ethernet scale-up networking by 2030. That's an industry separate analyst report, but that's compelling as a market size that we want to go after and pursue. So we're putting our money where our mouth is, so to speak. We're investing in pursuit of scale-up networking for Ethernet. We see there's a huge opportunity for the non-NVIDIA ecosystem of Ethernet attached XPUs where you need some sort of facility to create that kind of coherent memory on the back end between all of the different XPUs in a rack. You need some sort of high-speed interconnect, and we think Ethernet switching is really well designed for that. And so I think that, first of all, it's an attractive market, it's potentially a large and lucrative market. We're helping lead the definition of that market to begin with, where we helped invent and popularize the whole concept of Ethernet scale-up networking one of the software leaders and architects from our company is the co-chair of the ESUN Network Working Group as part of OCP. And not only do we have the industry standardization, not only do we have the Ethernet prowess for kind of consistency of AI fabric from scale up to scale out to scale across, but we're also bringing to bear a lot of the hardware engineering prowess that we've got from our modular chassis over the last couple of decades, we're bringing that to bear at this new level of complexity for scale-up networking within a particular rack. So the opportunity is significant. We think we've got a lot of software and hardware differentiation and unique expertise to bring to bear.

Yes. And the other thing I would add, that's a great like technology description, Brendan. So thank you. The thing I would add is think about when we were talking 2 years ago, kind of the Ethernet and InfiniBand, there's going to be some segmentation and scale up, we believe they'll be proprietary NVLink. There'll be some kind of white box equivalent for the lowest needs. And then there'll be a value-add Ethernet, which we hope is us as the branded vendor. So we see that segmentation maintaining. And we got to the point of announcing to be in scale up because we had almost like a crowd sourcing from our customers, "Hey, we said can you think about getting into this space as an Ethernet leader." So there's a bit of a pull and a bit of an innovation, let's get there, sort of a combination between us and some of the customer base. And so that's where ESUN was born, and I think that's a great thing. It will not be material in our '27 financials because we do have to get the ESUN standards and do pilots and trials. But once it gets going, in '28. We hope it's a great new segment. And when we update our TAM, we'll give a shot as to what we think that can be.

Okay. Great. I got a question from the audience, and I'm just going to read it. And the audience -- the questions are coming from the audience are different than the questions that I would ask. So...

That's fine. Yes.

Let me just read it. Historically, Arista has discussed the customers who prominently use white box will continue to pursue that strategy. Is this still the case? And what could happen that would allow you to win meaningful business at these customers?

Yes. So generally, generally, the white box environment is not an environment we'd want to go after because it's not -- it's kind of overpaying for technology for places where it's low use in the network, right, Brendan. So the white box use case remains the white box use case. Having said that, there's probably somewhere a little bit in the middle where it's -- the operating system that sits on top of that will be a difference as we go and kind of maybe into these mythos environments. So maybe that's where it gets a little squishy for lack of a better word in the sense of where that will meet. But we're not looking to enter the white box business that has its own use case in any of our customer designs. Brendan, anything you'd want to add?

Yes. The only thing I would contribute is we should think about why because some of the customers went for white box to begin with. There are some that went white box even before Arista was a company they're going to persist with that strategy. But I think white box persisted for those companies because it was high volume and it was low entropy. Like the complexity didn't change overnight. It was a well-known commodity, and there wasn't a lot of technological change or innovation. So they could just stamp out in high volume and save themselves a couple of pennies. As we think about the AI trade, that doesn't necessarily hold true, meaning that there's so much entropy of technological innovation, so there's an opportunity for some of these customers to think differently. That's what we're hoping. We think that the speed with quality value proposition that I referenced for our branded EOS-based operating system and platforms has utility in such a world where innovation speed is off the chart, where white box may not be as attractive because you need not just a white box hardware platform, but you need an open source software, need all of the features added to that by somebody. And then you need to find quality and you need to find time and speed and maybe that doesn't all work together. So it's possible that in the future, as Chantelle said, we're not going after white box, but some of the value drivers for white box may no longer be as valid, and we would love for them to join our ecosystem.

Another question from the audience, which follows up on this as well as the last question and the end of the last question that you commented on Chantelle. There's some chatter that the nature of scale-up domain lends itself a bit more to the white box model. Is this true? And how would you characterize opportunities that are attractive to you in the scale-up domain?

Yes. So I'll start in the sense that we do see that segmentation coming in the sense of a white box equivalent scale up, a proprietary NVLink that eventually transitions to Ethernet just like InfiniBand has done and then the value-add Ethernet. And then maybe, Brendan, why don't you take us through like what's that value-add Ethernet scenario?

Well, the bear case, so to speak, like Chantelle is noting where you might want this white box would be where there's not that many features. The bull case for us would be where our customer values consistency of operations, where they've already got EOS deployed for their scale-out network. So having consistency stamped out the same way for scale up gives them a very rapid deployment. It's already qualified. It's already proven to work. So turning it up for kind of the new world scale up makes it even easier because we shouldn't pretend all of the people in a non-NVIDIA ecosystem going with new XPUs, it's all new, stating the obvious, adding yet another new thing to the mix of an open source operating system with white box add extra risk and complexity. So the opportunity for us in that segment who might want branded is that we can give them that speed with quality to get deployment and just remove one variable from an otherwise net new and complex thing.

Yes, for sure. And if you think about all these customers to whoever asked the question and Mike, all of these customers are racing to get the most compute to get the best economics they can. And part of that is utilization. So whether it's scale up, scale out, scale across, we feel open source Ethernet is going to get them the best utilization, the best outcome, the best economics. And I think that's the ebb and flow you're going to see as they work through their business models. So we're very excited to help them.

Great. Great. Let's move on to some other questions. Just 1.6, which I think is trialing right now with large customers. You said volume production in '27. I think there's a lot of things that all have to come together to make a new platform like this work. So -- basically, just any more on the time line of 1.6? And are you -- and then as we talk about NPO and CPO and the potential for open standards, NPO, CPO, is that -- how does that layer into the 1.6 product over time?

So I would say from a timing perspective, as you noted, trials are going on now volume for next year. From a technology perspective, it's relatively straightforward in that we can still, with this generation, have both air-cooled and liquid-cooled options. So there's no sort of blocker from a technology impediment perspective where everybody's got to go liquid, not everybody is ready. We have air-cooled systems we've already announced it publicly. For those people who want liquid cooling, again, we've announced that as well. A lot of the liquid cooling requirement is going to be driven by the compute nodes. So we're not going to be the tip of the spear for that, people will have deployed in ones in liquid cooling for the compute nodes, that will likely be deployed before any of the networking gets deployed to interconnect them. So that's, again, likely not an impediment. It fits into the overall build-out that Chantelle and Jayshree have been very circumspect describing for a long time, which is data center build-out and powering and facilities build-outs, they're all out of our control, but a lot of that will be done before we get there. So I think 1.6T is here. The optics are here for that from a local connectivity perspective and 1.6T ZR for scale across probably next year. With regards to NPO and CPO, it's a technology set of options we're looking at very closely. We actually announced earlier this year at OFC, we announced something we're calling open CPO. CPO, of course, is co-packaged optics. We also announced at the same time, technology called XPO, which is our pluggable high-speed liquid-cooled of 1.6T option. So in that case, again, we're trying to innovate for the industry, and we're trying to give choice. If a customer wants CPO and they're just bent on that decision, then Hallelujah, in that case, we would offer open CPO, which offers serviceability and multi-vendor options for co-packaged optics. So that you're not locked into any one solution, and it's not a value proposition to throw the whole switch out if something fails, serviceability, we think, is critical there. But XPO is a complement to that in the sense that it's fully serviceable, it's pluggable. It's liquid-cooled, like I said, and it offers every sort of laser available from short range plus LPO and ZR, ZR+. So we believe in choice and we think that with 1.6T with open CPO and XPO are giving our customers tons of choice.