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

Welcome, everybody, to the Arista Networks keynote fireside chat at the Goldman Sachs Communacopia + Technology Conference. I have their privilege of introducing Jayshree Ullal, who's been the Chairperson and CEO of Arista Networks since 2008; and Chantelle Breithaupt, who's the CFO of Arista who joined the company earlier this year from Aspen Technologies where she served as CFO. And prior to that, held a variety of executive finance roles at Cisco and GE. My name is Mike Ng, and I cover Arista Networks and networking technologies here at Goldman. We have about 35 minutes for today's presentation, inclusive of some audience Q&A. First, Jayshree, Chantelle, thank you so much for being with us today.

Thank you so much. Thank you for having us.

It's a real privilege to have you guys here.

It's great to be here, Mike. Can you guys hear me? Yes. Okay.

Great. Jayshree, I want to start out by having you talk about the current state of data center networking. Obviously, there's been a tremendous amount of focus on AI compute infrastructure. And it feels like anything that helps to improved performance, relieve congestion in that AI infrastructure should be very topical, and it certainly is. So would you just talk about the current state of data center networking?

Absolutely. I think data centers just got sexier with AI, right? But just stepping back a little because Arista's origins more than 15 years ago really began with how to build better data centers. And at that time, the whole topic of discussion is should data centers be in the premise on an enterprise, do they move to the public cloud or is there an elastic consumption model between the public and the private or the hybrid, as we often call it. And the conclusion we've come to 15 years later, after having built lots and lots and lots of megawatts and data center capacity and switches and routing, et cetera, is data centers are simply not going away. They're morphing into more and more different capabilities. And I like to think of it as not data centers, but centers of data. And the center of data can sit in the cloud and be something we make available to the hyperscalers or cloud titans, as we call them. They could sit in an enterprise. They could sit in a branch. They could sit in a campus. They could sit in a WAN. And all these centers of data, independent of their location, have to be managed, controlled, automated in a highly visible fashion. So data centers and the leaf-spine architecture Arista has pioneered for many, many years has just stretched elastically, if you will, east to west and north, south.

Great. And then I want to ask about the Arista EOS platform because we often hear about how that's a source of differentiation, particularly among your switching portfolio. Maybe you can just clarify and talk about how EOS differentiates Arista versus ODMs and OEMs [ impact those ]?

Yes. So as you all know, these days, hardware is very central to everything we do, but software has to run on hardware to make networking really work. So Arista took actually 5 years to build what we call Arista EOS, Extensible Operating System. And the premise of the software was we're not going to build a separate silo for data center, campus, WAN, AI, et cetera, we're going to build 1 operating system with an architectural foundation like no one else has yet built in the last 15 years, which is what we call this published subscribed state-driven model. We started out with SysDB, which was making that possible as a database in one switch. Then we move to NetDB, which is making that possible across switches, network-wide database. And today, we have a network data lake architecture, where we can actually not just connect all of these switches, but we can deal with different forms of data. And the data can be structured, unstructured, flow data, you name it, right? So not just connecting, which is a fundamental basis for networking, but analyzing the data, correlating and visualizing it is an important aspect of what we do with EOS. On top of that, Ken Duda and the team have been hard at work. We started with 1 million lines of code. And it's -- I like to think it's 15 years, not old, but 15 years new because we keep [ guarding ] the foundation and adding more features. And today, we move to over 15 million or 20 million lines of code and features and networking capability, recognizing 1 OS, 1 architecture and then many personalities across all the use cases we have. So we're programmable on the data plane, management plane, control plane. And frankly, it would take thousands of engineers and a whole lot of years for anybody to replicate and enable what we have done over the last 15 years. It is many, many millions of man year effort.

Great. And you're seeing a lot of that in the current business momentum and...

That's right. Obviously, there are different types of OSs that we run into. There's always open source like SONiC, and we embrace that. But anytime you're trying to do anything slightly different or complex, you need the advantages of the EOS, whether it's for the AI spine or a WAN optimization or campus or even classic data center. So no doubt, we coexist with a lot of the simple operating systems, if you will. And no doubt, we embrace them, too, because we've always lived in a world of multi-OSs. But a thing of beauty here is to really understand that Arista's differentiation is not just the OS, but that it can manifest itself across so many use cases so many platforms with 1 single binary image. And this is why we've held a real high-quality record here, where the number of critical vulnerabilities is the lowest of any platforms. If we're x, everybody x, everybody else is 100x, not 1x over the last 10 years, and our Net Promoter Score is the equivalent of 95. So you don't achieve that without a maniacal focus on quality, availability, automation, analytics and highly differentiatable set of programmatic APIs that go beyond the classic command line interface. So it's been a labor of work, and it's been over a number of years, and it's paid off.

That's great. And translating some of that momentum into the business and financials, I was encouraged last quarter, Arista raised its year revenue guidance to at least 14% from 12% to 14% prior. And it seemed like as much momentum as there as on AI networking, there was classic cloud titan demand, there was campus demand, enterprise. It seems like a lot of things are going right.

Yes, time to be diversified. Actually, regarded the year thinking it would be 10% to 12% growth, you might remember. So began the year sort of cautiously optimistic, and now I'm really optimistic because we can go out of the 10% to 12% into the teens. So it feels good. Maybe Chantelle, you can add some here since -- ever since you've come, things have gotten better.

Okay. Well, I don't know if causation or correlation, but I think that's the reason we did raise the guide for the continuous times is because we started to see optionality across. We began the year thinking we see moderated traditional cloud. We've seen better than expected there, the enterprise and the sense of campus, [ DC ]. So I think all of those things have given us optionality because we wouldn't want to bet just on one horse to guide the year. And so I think that's been a really nice scenario to come into 2024.

That's great. One unique thing about Arista among many is that it's found an incredible amount of success among cloud titans, which I think is pretty unusual, at least among OEM networking companies. Could you talk a little bit about how Arista's engineering, new product organization has allowed for these close partnerships?

Yes. No, I think you've hit something that is so unique to Arista because from the outside, we probably just look like yet another big corporation. From the inside, we really are a nimble, agile, highly innovative engineering company that continues to out-innovate, outpace the innovation, but most importantly, doesn't rely on traditional sales or traditional SEs as our access and communication to some of these large cloud customers, who incidentally have thousands of engineers who are just as smart as ours. So the IQ is very high when these dialogues happen, as you can imagine. And it starts from everything, including how we co-develop a product. Oftentimes, we'll say, "we think we need to do this," and they'll say, "well, what about this?" And it hasn't been unusual for us to customize things. We'll go, "well, that's a dumb idea. Well, maybe not for this use case you have." So we work up-front 2 to 3 years before we actually see the products and the revenue manifest in co-developing, co-designing, co-engineering for each of our major cloud customers. And this is a very important belief inside us. And the reason while many people say our customers are like partners, we truly are joint partners. And a huge shout out to my engineering teams because this is why our cost of sales and marketing is lower than many companies as a percentage of revenue because it's really our engineers who deal very much directly in co-developing this. There's also another whole aspect of this that is, I think, less understood because not only do we co-develop with them, we deploy, we become an extension of their organization. And as their install base with us develops, we're able to give them data on how to deploy it better, how to automate it better, how to develop the analytics and the visibility capabilities. And again, don't forget, these are folks that have a DevOps, a NetOps, SecOps, a CloudOps background. So -- but you take the sum of 1 plus 1, which is our team and their team, you get 10. You don't just get 2, right? And so it's a very unique arrangement that we have not changed. We've, in fact, stepped pedal to the metal, particularly with the advent of AI, where now we're not just looking on the front end of data centers, but we're also looking on the back end.

That's great. And there's clear evidence of the success of these partnerships. We see it in the P&L. But one question that I had is Arista is clearly contributing their IP, their R&D resources to these partnerships. At the same time, how do you make sure that you get a return on those types of things and that Arista doesn't get displaced by a different supplier?

Yes. No, listen, you can never take your success for granted, and you should be clear that we are never sole-sourced with these vendors either. There's always multi-vendors, and we always have to be better and prove that we're better not just from OpEx point of view, a CapEx point of view, but also operationally cost effective. So I would say at any given time, while we win use cases, we're always paranoid about the next use case. But once again, I think we have to earn the merit and earn the seat at the table, and we do through this combination of innovation, operation and then helping through the deployment cycle. So it truly is something we do take together and plan together over a period of time. And you guys see the successes, but I can tell you, many of these that we started together that didn't come through also. We have bought some of them because they don't make sense.

That's great. And while we're on the topic of cloud titans, Arista recently added Oracle as a cloud titan customer, could you talk a little bit about what you saw there to give you confidence in adding Oracle as a cloud titan customer? And then similarly, do you see opportunities to add other cloud titans, Google obviously is one that comes to mind?

Yes. Without naming all the names, first of all, we added Oracle, but we also moved Apple to the Tier 2 cloud. So first of all, definition. When we talk about these things, we want to see mega scale, better scale. And the definition of a cloud titan is you got to at least deploy 100,000 servers, right? And so we saw Oracle heading in that trajectory more so than the others. So Tier 1 or cloud titans are those that have that kind of scale. And then Tier 2 doesn't mean they don't aspire to that scale, but they may not have yet achieved that scale. So we certainly hope we'll add more to that scale, but the current definition includes that. In terms of our partnership with Oracle, OCI and our work with them started more than 5 years ago. I still remember when they -- we all know who the top 3 are, and Oracle wasn't in that list, but -- and you just heard their earnings yesterday, it's so impressive what they've done, and the commitment they've made and the assets they've invested in to make that possible. And Arista has been a strong partner, not only in their cloud efforts but also in their AI efforts. So we believe that just as the classic cloud emerged, the combination of what I now call the AI center, which is the back end can be AI, but the front end is going to be a lot of cloud plus AI is turning out to be a great opportunity at both Oracle and many other customers as well.

That's great. Can we talk a little bit about the demand drivers for cloud titans today and what it could look like in the next year? Specifically, it doesn't seem like there's that much AI back end today, but that's a big opportunity for next year. What are the cloud titans do and is it -- yes.

Let's sort of step back and say what happened here. What happened was Arista and many others were focused on the front end, and we were going from a 10-gigabit refresh to 100 gig to 200 gig in some cases and now 400 gig. And all of a sudden, in the back end, GPUs came alive, largely connected, I would add by InfiniBand to date, right? So for the longest time, we were outside looking in where InfiniBand was more of a high-performance compute technology but was serving its purpose as well. In comes the Ultra Ethernet Consortium that Arista is a founding member in. And we now believe that with Ethernet, you can be equal or faster than InfiniBand. You can have improved latency. You can have better scale. You can have predictable end-to-end messaging. These AI workloads are very compute and data intensive. Congestion control, all of this, packet spraying, multipathing becomes so important. And neither can you have like a single subnet manager manage 40,000 nodes and then bring everything down. Good old-fashioned Ethernet and IP play a big role, right? So this is where we are seeing now that, we don't have to be outside looking in. And I've mentioned many times that in the recent past, we won 4 out of 5 of the major cloud/Tier 2 cloud companies where it's not InfiniBand anymore on the back end, connecting the GPUs, it's Ethernet. 3 of them have now moved from trials to pilots, and we expect them to move into production next year. One of them is slow. It's still in trials and probably waiting for some next-generation GPUs like Blackwell. And the fifth one that I'm hoping we get very close to, I'm pleased to say we intend to go into trials this year. So it gives me great pressure to tell you for the first time that we're going to be 5 out of 5 on our goals. But your point was right, which is these things take time. We began a lot of these endeavors in '21, '22. We started marching into trials in '23, '24. And by the time you sort of see the knee of the curve, we're still in our first innings, and the second innings perhaps starts next year, where we start going into the numbers I mentioned, which is $750 million in back-end production or pilot revenue as the case may be. Now everybody looks at me and says, "Well, Jayshree, that looks like a low number given all the spend." So before you jump on me on that, I want to say a few things. That's a starting point, and it's pretty good from 0, wouldn't you say? And as you often say, there's a glide path, which means we're somewhere between 0 and $750 million already in 2024, right? But there's one more point I'd like to make, which is as a result of these back-end centers, we initially thought the classic cloud would just pivot to these back ends and nothing would happen. But the positive thing I'd like to leave you with is the classic cloud is also changing more to a front end with classic cloud and AI on the front end. So for every back-end GPU dollar I'm getting, I'm potentially going to get at least 2x on the front end, because the pressure and the bandwidth and putting at 800 gig or 400 gig on the back end is going to translate into pressure on the front end. So if you add up those numbers on AI networking, it starts to get meaningful, and it starts to grow meaningfully as well.

Could you elaborate on that opportunity a little bit more? Just so, I mean, I'm clear, is this about an inferencing opportunity once the training clusters are put into place? Is it -- yes.

Yes. No, this is a good point. I'm still talking only about training. Once you get to inferencing, it becomes harder to distinguish the back end and the front end because could be a GPU, it could be a distributor GPU, it could be a low-end GPU, it could be a high-end CPU. So if you just stay on training, initially, you're going to spend a lot of your dollars on back end, enabling all of these clusters at billions of parameters going to trillions of parameters. We have job completion time, and every second matters, right? This is what we're doing. My point is as a result of doing this, you now have to worry about the front end where you're connecting storage, wide area, general purpose compute or even like you mentioned, some of the inferencing applications. So the heavy-duty back-end training is going to have a broader effect on front-end networking and what we call holistically AI centers.

Great. That's very clear. Last year, Anshul was here, and we talked a little bit about InfiniBand versus Ethernet. And I love it when I hear you on the earnings call and talk about how Ethernet is going to be it Ethernaut. And you've been saying that for a long time I feel like a...

Yes. I've kind of been through this movie for decades, not just years, right? You have to remember, I've been there when there was SDDI, which is now a 4-letter S word, ATM, token ring. I've seen it. All and every time Ethernet wins. So it's simply a matter of when, not whether.

Right. And if we're past the InfiniBand versus Ethernet debate, within Ethernet for AI networking, could you talk about some of the competitive dynamics there? Obviously, Arista's [indiscernible]...

Listen, whether it's front end or back end, we're always dealing with competition. We have to earn the right to be there. I think one of the nature of competition on AI for Ethernet that we cannot lose track of is there is a lot of bundling that can go on between -- in an AI rack between the GPUs and the NICs, which Arista doesn't build and the network, the scale out and scale-up switches, if you will. Now we're a firm believer that best-of-breed always wins, but sometimes it's just convenient for people to put it together and say, okay, I won't deal with it. But when the dollars get really large, trust me, they do. And a good network more than pays for itself if you build it right because you can actually improve the utilization time of the GPUs if you build a good network. So you don't always want an average network because then you're going to be idling expensive GPU time. And if you're idling 30% of your GPU time, and let's say you have $100 million in GPUs, which can sometimes be $1 billion, that's $30 million of GPU idle time. I can build a network for much less. So the net cost of the network is low, and it keeps track, and it guards the henhouse, if you will, from getting the maximum utility on your GPUs. So while some people may bundle for good enough, I think most folks when they start spending sufficient money and are trying to get the best job completion time for their training workloads, we'll go with the best-of-breed network, and this is where Arista finds itself highly competitive but highly differentiated.

That makes a lot of sense. And there's the bundling aspect and then there's the white box aspect. One thing that Arista said in the past is that Arista switches with EOS can actually have a lower TCO relative to white box despite a lower up-front cost. Maybe you can talk about that?

That's right. Yes. I'm sure Anshul covered that last year. It's a very important fact to realize that how many of us build our own PCs today. We could. It would be cheaper. I mean we did that back in the Radio Shack days and all of that, right, but yet we all get a Mac or a PC from someone. Why? Because the economies of scale when you bring all of the components together, whether it's a storage drive or the screen or the keyboard, everything. And it's not too different here. We embrace the white box, but the economies of scale on building it yourself only comes to factor if you have 1,000 engineers to put on it, and very few companies do, right? So while it looks like the white box may be cheaper, by the time you're done with the TCO of the hardware, the drivers, the software, the applications, and most of all, maintaining all that the blue box, in this case, the Arista, would be much more of a cost-effective TCO by several percent, not just slight percentages, right? So that said, it depends on the use case. If you really have 1 compute application, you're not trying to push it a lot, which is not true with AI, you might go with that, right? But if you really have a multitude of use cases and tremendous pressure on the capabilities you need on the network, you're not going to take any chances on that.

That's great. We touched a little bit about the at least $750 million of AI revenue. I was wondering if you could also just talk about the AI for enterprise opportunity.

Yes. So clearly, networking for AI is consuming a lot of our bandwidth, both in terms of performance bandwidth, but also our own investments and the 4 to 5, soon to be 5 out of 5 customers we were talking about. But there is a whole tale of enterprise customers that Chantelle and the team have been working on. And before I hand it to here, so I think the key here is to remember there's networking for AI and they're building some small clusters, but there's also AI for networking on how we use AI to get an advantage to do more root cause analysis, security, zero-trust observability. And this is why using AI to go beyond behavior and signature to really look deep in the packet to get better analysis is as important on the enterprise as building the GPU cluster. Do you want to say a few words here?

Yes, sure. Thank you. And so back to the topic of enterprise AI to your question, Michael. Last week, I was in New York having dinner with several of our customers across industries, and I was asking them specifically because this is something we want to better understand as a company and to inform our stakeholders. And they had -- the conversation, the optimism has moved from everyone needs an AI strategy, the enterprise, they answer to their Board that they're thinking about it to actual specific use cases. And I think that's where the goal is coming through, at least in the sense of optimism. And so some of the industries would be media and entertainment, financial, health care, automotive, et cetera. And so they went through some of their use cases. And I said that's very interesting. Now tell me what the sales cycle insight is to earn that CapEx, right, to proceed? And so what they described as they try some of the pilots out in the cloud to do some pilots. And then they bring it back in, and they demonstrate their use case, have a conversation with their Board, Then they decided they're going to remain off-prem or on-prem. And that, they said like a 2- to 3-year journey. So that's where they're seeing specific use cases, 2- to 3-year journey across the different verticals. And I think that if 1 customer in a vertical finds the golden nugget, that vertical could have some accelerated growth. But until that, it's going to be this 2- to 3-year cycle time.

Great. And is it very clear that more of this enterprise AI infrastructure investment is going to happen on-prem? Or does it still feel like a little bit of a debate in your mind?

It still feels like a debate. I mean I think if they can't get the dollars to invest in a significant way and more often they can't operate it, they will go to the cloud. But if it gets to be a significant use case where they need ROI, they'll bring it back into the premise.

Right. Great. One area of enterprise that with a clear area of strength for Arista anyway was WAN and routing, which was particularly notable given that the industry or the market is not doing that well. So why is Arista outperforming the market in that vertical?

Well, I think when you go back to history, there's a belief that you needed separate switches and separate routers. And there was always this debate just like we talked about InfiniBand versus Ethernet on, switch when you can, route when you must. One of the things Arista has done has just made Layer 2 switching and Layer 3 routing technology equal citizens. We built it. We built our own BGP stack, OSPF, we do all of the routing, multicast, all of the rich features required and the enterprise required in the service provider. It's been a labor of love. It's taken us 5 to 10 years to get there. But the big difference between Arista and just about anybody else is, we don't believe you need a separate router box at 10x the price for your wide area links. Why? Because wide area links are moving to Ethernet 2, by the way, and you can do all of the traffic shaping and Metro Ethernet capabilities on it, whether you're switching at Layer 2 or Layer 3 or Layer 2 over Layer 3, technologies like tunneling, right? So that belief has really translated into an advantage to the customer because now they don't have to buy 2 boxes. They got 1 platform, Arista switch platform, very common one being the 7280 or 7800 and you can have multiple personalities on it. You can have a switch facility. You can have a routing personality. You can have an automation security, you name it. And I think as that's transformed the TCO, and it's transformed the way people think of the wide area and they don't have to have 2 different investments.

Great. That's really helpful. And is Arista doing particularly well in any area of wide area networking. I've heard that Arista does really well in like Data Center Interconnect, as an example. And then related to that, are we seeing more distributed training right across multiple data centers as well? Is that a theme that is beginning to emerge when you see it's happening?

So clearly, we're doing well in the routing use case connecting data centers. When people have the rights to their own cable, fiber, et cetera, not only are we doing well in that, then they add their levels of encryption on it as well. And whether it's MACsec or TunnelSec or IPsec, and this has been Arista's advantage, both in cloud -- Tier 2 cloud, et cetera. I do want to highlight, though, we're doing very well now more recently in enterprise and service provider routing as well. And the enterprise is new to us. And this comes from some new products we launched that has additional encryption and IPSec capabilities on top of the routing. It's still new. It's early days, but I think the momentum there. And the consolidation once again of switching, routing, encryption, all in a single platform with the right traffic shaping and latency management capabilities is really helping. What was your second question?

Are we seeing more distributed AI training at least for the hyperscalers?

We are seeing, let's -- it's kind of not related to the WAN, but stepping back a little, we are seeing that power is probably the most difficult thing to get to build any kind of AI center, right? And if they get lots of power, then they can build large GPU clusters, back end, do all of that. But a lot of our customers, especially the early adopters, they're being opportunistic about whatever power they get and therefore, building distributed training smaller data centers or distributed training AI centers with the power on will they get, right? Rather than saying, let me go build this big, highly centralized [indiscernible], let me distribute it. It isn't their ideal environment because the larger the training cluster, the better, the more improvements you can get on training workloads and number of use cases. But given the necessity of how much power is available, this is becoming a key trend.

Great. Shifting gears and maybe talking more about the financials gross margins. So last quarter, Arista reiterated its outlook for 62% to 64% gross margins for the year. What drove some of the margin outperformance in the first half, you're pacing ahead of that? And what happens in the second half? Why is 62% to 64% the right long-term target?

Yes. I think that. The way to look at it is there are two drivers I had mentioned in the prepared remarks on the last call, and thank you for the opportunity to reiterate those. So from the perspective of mix as a topic in the sense that mix has a large contribution between cloud enterprise, cloud service provider, et cetera. So that gave us some tailwind in the first half of the year. The second is related to more inventory-related reserves. And so coming out of COVID, John McCool and his team have done a fantastic job rightsizing the inventory, getting the component supplier liability in a good space. And so from that perspective, some tailwinds in the sense of not having to repeat those reserves. Those don't continue every quarter. So the lack of those, which has continuing and the mix in the second half, gave us that guidance at that time.

Great. And I'd be remiss not to mention Arista did raise its EBIT margin targets for the full year to 44% from 42% prior. How much did vertical mix by customer kind of impact that margin outlook? Anything to consider as we think about the EBIT margin implications?

Well, I think I'll always reserve the right for Jayshree and I to have a year of investment where we need to, whether it's enterprise, systems, so we'll always reserve that, but I think this 42%, 44% -- the one thing I wanted to mention when you mentioned EOS is a differentiator, the one thing I want to sneak in there as the CFO, it's a beautiful cost-to-serve model because it's 1 EOS, and it's very easy to serve than not often we have to serve it. So I think we have a lot of examples of high sales leverage, productivity, EOS being very easy to serve. And that allows us healthy operating margins. And so we'll fluctuate among those amounts. But in addition, in that guide, we're still investing in sales and marketing. We're still investing in great product introduction. So it's a really nice mix, and we'll continue to operate in that range.

Great. And Chantelle, if I could ask about capital allocation, the Board authorized a $1.2 billion share repurchase program. What are your capital allocation priorities? How are you balancing those buybacks with investing further in the business?

Yes. I'd say the top 3 are -- we're still getting some pretty strong interest income by investments. So that's one of them. The second one is the organic investments and the things I just mentioned, perhaps some tuck-in acquisitions if we need them, but that's a maybe. And then the third one are the share repurchasing. Now you've seen that some things can move the markets pretty quickly. So we're just watching how those things happen. So the buybacks are always when is the right opportunity and where it makes sense financially. So those are the three that we're considering in '24.

That's great. And Jayshree, we're just on time. Maybe I can ask a closing question to you. What are your strategic priorities over the next few years? A few years from now when we look back, what will become more obvious even if it's not obvious now?

Well, I think Arista has always been a very strategic company about what we do and equally important, what we don't do. So I think what we do, we're going to continue to push the envelope on networking and build the greatest centers of data, whether it's AI center, data center, WAN center, campus center, you name it. So that will be a real strategic priority. I think the other big one for us that we haven't focused as much on, hopefully, we'll get a chance in the next few years, is I think security is too important a topic to just let the security vendors do. So we'll be infusing more and more security into our platforms and turn security from noun into an adjective. And the same thing with observability, you can't secure what you can't see. And so I think these two kind of go hand in hand, and that's going to be a strategic initiative from an innovation perspective.

I'm excited to see you execute on that.

It would be a little harder, but give us a decade.

That's a great way to cap it off. Jayshree, Chantelle, thank you so much for joining us on stage today.

Thank You. Thank you.

Thank you so much.