2023-08-23 16:00:22 ET
Arista Networks, Inc. (ANET)
Rosenblatt 3rd Annual Technology Summit
August 23, 2023 1:00 PM ET
Company Participants
Liz Stine - Director of IR
Jayshree Ullal - President and CEO
John McCool - Chief Platform Officer & Senior VP of Engineering Operations
Conference Call Participants
Mike Genovese - Rosenblatt
Presentation
Mike Genovese
Hi, good afternoon, good morning, depending on where you are. I am Mike Genovese, the cloud and communications infrastructure analyst at Rosenblat Securities. Thank you for joining us for the Third Annual Age of AI Conference.
I am super excited today to have executives from Arista Networks, ANET, a really, really -- company with an exciting history but also an exciting future. And not that long of a history, but it's been a really good one.
And today, we have from the Company, the Chief Platform Officer, John McCool, who I think is as cool as his name sounds. And we have the fun one at the Company, Liz Stine, the Director of IR. So John, Liz, thanks for being here today.
Liz Stine
Thanks for having us.
John McCool
Thank you.
Mike Genovese
All right. Well, let's get into the questions. And for the investors who are joining us here on the webcast, there's a tool on the right hand of your viewer screen to submit questions, which will just be e-mailed directly to me. And I have a 100% hit rate so far of asking the questions that I get on the fireside chats. And I expect that to continue.
Question-and-Answer Session
Q - Mike Genovese
So John, let's talk about '23, the second half of '23. I mean you guys are having a really good year, more than 25% revenue growth, I mean, well more than that actually. And but then you've also talked about push-outs and push-outs, pull-ins, and there's a little bit of caution in the environment this year.
I think web spending is probably supposed -- or Cloud Titan spending, is probably supposed to get even better next year. But help me kind of describe the 2023 environment where there's these push-outs, yet you guys are growing in the -- 30% more. So, yes, how would -- talk about '23 before we talk about '24.
John McCool
Sure. So if you think about -- let's talk about '21 and '22, we saw significant growth in the cloud sector, right? And I think part of that was with supply chain and the constraints, the cloud titans specifically recognized that we were in a difficult environment on supply and recognize that much earlier than the enterprise customers did who might not have direct access to semiconductor companies. So, we saw a lot of growth in those two years, continuing into '23.
But at the same time, starting to catch up on supply in the beginning of the year and started to meet rollout and deployment schedules, and we've been very cognizant of driving those products into the market, consistent with our customers' ability to roll them out and put them into the environment. At the same time, you had the enterprise deployments starting to kick in, and we were able to take some of that capacity and deliver to our enterprise customers who've been waiting one product for some time.
Mike Genovese
I definitely want to get into the enterprise, and I think we'll talk more about the cloud and specifically the Titans. First, I mean, I want to get a perspective from inside the industry because from outside the industry, it seemed like all of a sudden with ChatGPT just really just a few months ago that everything changed and all of the CapEx priorities of the cloud changed and they were all of a sudden overnight doing something completely different than they were doing before. But from inside the industry, things probably look different than what I just described. So just talk about how AI has impacted spending so far. I mean we're going to talk about the opportunity that's ahead of you on AI. But just again, for '23, sort of from your perspective, how is AI really changed things in '23?
John McCool
Sure. I mean, I think we would agree in some ways, the ChatGPT was a defining moment. At the same time, we've been involved in customers who were engaging in AI, doing trials, starting to think about how they would deploy. But I think at that moment, when this became a very public announcement, people had to take a look at the projects internally, how are they doing, think more aggressively about how they were going to monetize them and really kind of announce their plans and tipped their hand if you will.
So, it really became kind of a shift in kind of an internal focused effort to be more external and public. At the same time, I think they were kind of rationalizing what that means to their infrastructure going forward and how they refocus on driving those initiatives. And I think what we've seen more recently is an understanding that that's going to require continued investment in the AI piece but also along with their core networks to be ready to put in those AI clusters along we're just continuing to sustain and run the business.
Mike Genovese
Yes. So, I mean that -- as you think about '24, do you expect then that as I kind of call it that these cloud guys will walk and chew gum in terms of they'll be able to focus on the general core cloud and get what they need done for AI? You don't think AI is going to especially suck up all of the air in the room.
John McCool
That's a good question. I think we've seen over the last six months, kind of that understanding that you have to do more than one thing and customers starting to react to how they're going to deploy and make those investments going forward. And I think we'll see that build out even more strongly through the back half year of '23.
Mike Genovese
Okay. But before we get into sort of the '24 and beyond outlook for cloud, let's go back to the enterprise business for a second because you made that comment earlier where you said -- and then this year, the enterprises started to deploy these large networks. Let's look -- I guess, double-click more on that, right? I mean why did that happen all of a sudden?
John McCool
Yes. I think this might be one that wasn't all of a sudden. We've been working on the enterprise for some time. I think that the enterprise was slower to recognize the impact of COVID on the supply chain and perhaps some disbelief in those extended lead times before the cloud folks did, and they're starting to deploy. And I think at the same time, there's a very unique Arista piece to this is, during COVID, I think our teams were able to get in front of more enterprise customers. Within different market segments, different verticals, we've seen movement of folks from different IT teams.
We're always very strong in financial vertical, media and entertainment, but moving into health care and industrial use cases before we probably didn't have the recognition who's Arista in those kind of accounts as more people have used our products and moved to other companies. I think that there's less headwind for us just being able to insert. At the same time, we've built out the portfolio over the last three years from kind of a data center or data center routing to campus use cases and network visibility. So, we just have more to offer. And effectively, our sales teams are getting more at bats to bring in Arista kit into the environment.
Mike Genovese
Well, there's -- I mean -- yes, there's a -- so let's stick on enterprise for a while before we get back to the cloud stuff. Talk about more of the newer -- of the things that you have of the more complete solution. But then I also want to -- what other pieces do you need to add there? And then the other part of this that goes with it is today, in your enterprise business, how much is direct touch and how much is distribution? And do you think -- how will that change over time?
John McCool
Okay. Maybe I'll take the latter part of that question first. So our enterprise go-to-market effort is very focused on Fortune 2000. And we have to drive a direct preference in those accounts, get them to understand what we have in the portfolio, what our unique advantages are. So the go-to-market is very much Arista-led with partner fulfillment and deployment. So we have gotten a much stronger engagement with the channel as a result of those deployments, and it wouldn't be atypical for Arista to focus on a Fortune 2000 account or region and that partner who's deploying to see other opportunities that might be smaller and pull Arista into those accounts.
We think this approach, we still have a lot of headroom with that focus, either achieving new accounts and new account penetration. But even in large accounts that we're in, they have significant network TAM that we're just beginning to tap into. So we have room to grow even with the accounts that were installed in today. At some point, I think we may look at a mid-market type of approach, which would be more of a distribution channel-led approach, but we're not there today.
Now regarding the portfolio, I think we're recognized early on as a data center player and that allowed our entry into specific verticals that were data center heavy, right? So financials, media entertainment, but as we built out the routing stack, we could then address data center interconnect with their virtual router data center to cloud. With the campus, we broaden that out to POE use cases, connecting IoT equipment, network visibility with one of our acquisitions in Big Switch.
So you've kind of seen this use case build out more recently, the WAN transit portfolio was a piece where we didn't really address the connection of a large enterprise and their backbone into remote offices or branches or retail locations. So, we added that last year. And then we've strengthened that with our introduction of network identity with our AGNI product that we call it, to be able to identify users coming into the network.
The Awake technology, we've coupled with our campus engagement, which gives us a unique way to profile interconnect. People are coming into the campus environment, from our own locations and identify them. So, I think we have a very strong broad portfolio now across all pieces of the enterprise network. Obviously, there's some filling out and building out of the portfolio we can do within that. But I think all the typical use cases at some level we were able to address.
Mike Genovese
I didn't hear you talk about Mojo or WiFi. Did you -- was that just -- we just left like...
John McCool
You said -- there was too many words. And for folks working on WiFi and Arista, I apologize for that. But no, that's been -- I think one of the reasons we did that acquisition when we went into the campus as we realize you had to have both a wireless and fixed wired portfolio to even engage in some of those portfolio bids.
Liz Stine
And Mike, I think one of the unique things and the differentiating thing about Arista is that we've been able to expand our product portfolio and enter these new markets, all while keeping that common operating system, that EOS in that common management platform or CloudVision. And these are the things that our customers see value in from the fact that you can run the same operating system end-to-end, so that you can manage it and you can get visibility through CloudVision end-to-end from your network rather than come out with a new product portfolio that had its own operating system, our own management tool.
Mike Genovese
Great. So, on the second quarter earnings, I heard the Company. I hope I'm not mischaracterizing this, but I sort of heard the Company say, well, low double-digit revenue growth, probably about where consensus is of 11% is a reasonable target to have basically. I mean that's what I basically heard from the Company. And I think the quarter before that, the confidence wasn't as high to say anything about '24. So, it makes sense that we're closer, we get a little bit more visibility. But is it -- as you think about whatever the growth rate for '24 is and cloud as a driver, enterprise as a driver, just talk about them relatively as drivers in '24 of growth of the Company?
John McCool
Yes. I mean maybe just kind of step back to the comments. Obviously, moving into Q2, a little more visibility, and we did talk about Jayshree, we talked about double-digit growth as our target. As we get into our Analyst Day in November, I think you'll see more color around that 2024 projection. We'll have, obviously, more visibility from our customers.
In terms of mix, I don't think we're ready to call that. I mean, typically, what we've seen going into a new year is we'll have multiple ways to get there, which will be important. I think we do expect our enterprise momentum to continue into 2024. So, I think that's a piece of it, but we'll have a good idea to what cloud looks like and probably more color around this AI and core use cases and how that plays out.
Mike Genovese
Yes. Well, just starting at the sort of the CapEx for cloud and for Cloud Titans in particular for '24, it seems like for the ones who've talked about '24, they either raised it or they pushed out CapEx from '23 into '24. So, the '24 growth is higher than before because '23 is maybe a little lower and '24 is higher. So, it looks to me like -- and I'm not suggesting that the overall CapEx should line up with your revenues. I understand that year-to-year that's different over time, they look the same, but year-to-year, there's different.
But I'm just saying as a general backdrop, what looks to be over 20% cloud CapEx growth to me, I wonder what you guys are thinking internally for '24 seems pretty solid. I know that will definitely get into the AI discussion and the timing of the AI discussion for Ethernet in the back-end and what that means. But it just seems like entering '23, I guess instead of talking about it, I'll ask you, what do you think -- what do you guys think about cloud spending in '24?
John McCool
Yes. I mean we follow the announcements and it's general -- it's good to see it improving rather than decreasing. But as you said, it's been difficult for us over time in a specific duration to do a complete correlation of that spending, which includes outlays for new data centers, building buildings, pulling fiber, you get down to the IT portion of that CapEx.
It's a smaller piece. And within the IT piece, networking is about 10% of the mix, right? So, the timing, I think, is in the correlation of those annual spend numbers is a piece that has to get dialed in a little bit better. But generally, I think we're happy with the overall macro environment. And I think we'll learn more as we get towards Analyst Day.
Mike Genovese
Okay. Cool. This might be -- I don't know if other people start the question about AI here. But actually, let's start by -- talk a little bit more about the Ultra Ethernet form. I mean, is that something that you guys -- I mean, you're a founding member -- was that your idea to start it? Your idea with some other companies, I mean just tell us more about the background of that.
John McCool
Yes. I mean we've been involved in many of these Ethernet industry consortiums to kind of move the base of the standards forward. The Standards Committee work at their own time line, but move much more quickly when one -- more than two, three or four people came together to offer recommendations on how to adopt those standards or new standards that might move the needle forward.
So I think if you look at the collection of folks that are involved some of our competition, some are involved in the GPU space. We've worked together before. This isn't kind of an unnatural act to move the state-of-the-art forward. The opportunity for Ethernet is what we call the back-end network. So if you look at these AI clusters, they're constructed with a number of ports that face the user in some way.
And I think that's uncontested that those ports would be Ethernet ports. What's interesting to the Ethernet community is they are combined together in a cluster for GPU-to-GPU communication to do training and conduct algorithms, et cetera. And that interconnect, I think, represents a new opportunity. Neither InfiniBand nor Ethernet was purpose-built for AI neither was imagined when there was AI. And there's enhancements and capabilities that can be added to make those technologies work better with GPUs.
Mike Genovese
Yes, I mean just some of the standards for Ethernet, could you talk -- that are needed that you kind of want to develop quicker. Can you just talk about some of those features?
John McCool
Yes, I think the general problem statement is you have a high-value asset now connected to the network with the GPUs incredibly expensive, large amounts of memory and you want to get full utilization out of that asset. So, you don't want them to stall. So, it drives the bandwidth requirement higher, which was already on the Ethernet road map to 800-gig and such. So that's good.
But there's things that you can do around balancing the load, dealing with congestion inside the switches, giving visibility to the end user on the utilization of various GPUs that are connected in the cluster. So, they are the kind of aspects that would be added that the group is looking at. The other piece is just making it simpler to deploy Ethernet for these use cases. So making it a little bit more plug-and-play than it would be today. So they are the set of enhancements, I think that the consortium is going to address.
Mike Genovese
Okay. I mean, so I think you said -- I'm trying to remember from the 2Q call that -- they are currently trials back-end -- I mean with hyperscale customers, or in trials for back-end Ethernet?
John McCool
So, I think if we look back at Q1, with all the excitement of AI, I think we actually wanted to characterize where the state of the industry was. I think there were a lot of people wondering why is an Ethernet there today, our deployments happening in Q2. And I think that what we did in this announcement was try to lay out the framework for how we see things progressing. So, we talked about first half of 2024 trials. Second half of 2024 being more proof-of-concept, so people are actually deploying their technologies. And then in 2025, really being the opportunity for meaningful revenue transition in AI, assuming that AI is the ubiquitous use case that everybody believes it's going to be.
Mike Genovese
So are these going to be 800G Ethernet switches coming into these networks?
John McCool
I think there's a definite interest in 800G, but there's -- our 7,800 product today and some of our Tomahawk series products today are being used in kind of AI clusters. So there's not a binary kind of event here. But with the new capabilities that will be added in 800-gig, things will get better, and I think more interesting in terms of bigger deployments.
Mike Genovese
Yes. I mean, I think that -- well, I can only speak for myself. I want to say the financial community, but maybe I could just speak for myself that we're getting better at sort of counting the number of GPU ports that are going to need to be connected optically to other ones and what these things cost and sort of figuring this out. The size of the -- the size of the Ethernet opportunity, though, for you guys, I guess I don't really know how to think about it. And do you have any advice for the back-end opportunity, number of switches, capacity that's needed or compared to before? Obviously, back-end didn't exist before. So, I don't know how do we -- I don't know what we do compare it to you. But my question is, can you help on this kind of a model question?
John McCool
Yes. I think this math is hard. If I look at kind of the rollout of Ethernet into telephony, right, you could count the number of the phones on the desks and this is how many and what's the replacement rate. The challenge here is the AI use case itself will be very diverse, depending on what application you're trying to run if you're -- so how -- what are you actually building.
And the use cases will vary in the number of interconnect between the GPUs and how many front-facing ports at the front of it. So we believe there's going to be a lot of diversity in that answer. And there's a lot of I think quantification of the value to the end customer and how much they can monetize it that also has an effect on those.
So we're far from standardization in this industry. We're in the very wild west days of people building things out, sorting out what's optimal for them, and it could be very use-case-driven customer-by-customer on how they see the value.
Mike Genovese
I guess my question now, like I have a question here, which is -- are there more than three -- and I don't just mean for you guys that you're selling to. But you mean for the entire industry, are there more than three important customers here? Is it all -- is it -- I mean it might not even be all the Cloud Titans as far as I can tell. But -- and then beyond the Cloud Titans, are there important AI customers? I don't want to sound ignorant, but I really -- so far, I see this focus to three Cloud Titans in particular, and I want to know if I'm wrong about that.
John McCool
I think this could go more broad than just the cloud titans. I mean, that's obviously where our focus is. We think they'll have strong advantages of scale right off the bat. We think they have interesting use cases. A lot of people believe that actually having the data is important because what's the value of AI if you don't have the data to feed it. So that's where our focus is. There could be smaller use cases in enterprises or government that require smaller clusters and maybe they're just interconnected with NVLink, let alone InfiniBand or Ethernet. So, I think there'll be a lot of diversity. But specifically for Arista, we're focused on that cloud segment.
Mike Genovese
Well -- and do you see the opportunity with both of your because I was kind of excluding, I think, your largest customer from this. And I might be wrong about that. But to me, it seems like the ones who are really building this infrastructure are the ones who sell it to other enterprises, right? I mean they sell workload processing capacity to other enterprises. And I think your largest customers aren't really in that business. But -- so my question though is for AI networks, is there a huge opportunity with both of your biggest customers? Or is it more one than the other?
John McCool
I don't want to get too customer specific, but I would say, certainly, there's definitely infrastructure as a service type of opportunity similar to what we saw with just computing. But there's also internal use cases in large customers on how they're going to monetize AI for their own business.
Mike Genovese
Yes. Yes, that's helpful. All right. Well, look, I've been trying to avoid the question because I'm a little bit bored by now, but maybe some people on the phone aren't, so I have to ask it. Some people on the line about -- because I really see it as a matter of when, not if, and InfiniBand. But maybe just give us a little bit of color of how InfiniBand did get this much traction so far in the networks that have been built? Is it just because it was available and there and that's why we saw it? And tell us the reasons why Ethernet has to be the technology of choice for the long term, I think?
John McCool
I have to give you the brief history of time here just to answer that question. So just for people's background, InfiniBand also started with a consortium, but wasn't standardized by IAAA. And that consortium was folks that were involved in high-performance computing. I believe Intel, Mellanox, HP were involved in that. And at the time that happened, Ethernet wasn't very motivated for the data center or high-performance computing. It was operating at 1-gig and wasn't addressing some problems.
So two things that InfiniBand did that were very important for the industry. One, they adopted 10-gig interfaces before Ethernet, and they had a road map to 40-gig and beyond. The second one was something called RDMA. So this is moving data from an application out to the network, and they worked with the Linux community to make that happen. Ethernet came along and co-opted those technologies. So, you saw Ethernet move to 10-gig, 40-gig and at the interface and bandwidth level, both InfiniBand and Ethernet are moving relatively in lockstep.
So 800-gig coming on Ethernet, I'm sure InfiniBand has some kind of road map to 800-gig because the investment is so high on those interconnect technologies. They're being standardized across the board. Also, Ethernet adopted RDMA through a standard called RoCE, RDMA over converged Ethernet. So that's built into operating systems today. Ethernet also developed the ability to have a lossless fabric, which was a key part of InfiniBand. So some of those more potent type of high-performance things have been co-opted integrated into Ethernet already.
That said, Mellanox was acquired by NVIDIA. With the crypto craze, InfiniBand became an inter-connective choice for those crypto clusters and a very convenient way to roll out GPUs. So I kind of think there's a technology piece of this, but there's also a consumption piece of this question. And one supplier being more vertically integrated and selling kind of a solutions-based approach versus people who want to consume from multiple vendors who may be cooperating but also our competitors and have a more diverse environment.
Mike Genovese
Okay. So that's great background. And actually, that was not boring. Not to me. Maybe to some people.
John McCool
Somebody's fallen asleep, right.
Mike Genovese
Yes, somebody fall asleep. But I actually work up. So -- but now the question of why do so many of us -- I think including you too, I think that it's not a matter of if, but when that Ethernet is -- as the volumes ramp, as the models get larger as the data centers get larger, it almost inevitably has to be Ethernet, so why is that?
John McCool
Yes. I think, first, you have to believe that the AI thing is real, and it's going to scale and be really big. So if you have that as your backdrop, right, then you want to interconnect it's ubiquitous. You want multiple vendors, you want a rich ecosystem at all points of the infrastructure down to the, and you have some cloud customers today building their own GPUs, right? So we do believe there's going to be some diversity there. The InfiniBand technology is really owned by one company. There's merchant silicon or multi-vendor approach to that. We also think people will be interested in doing the same things they do today with compute.
So have multi-tenancy, so they can cut up their network or GPU investment into multiple buckets virtually and then either lease them to enterprises or even for their own use, they might have different use cases and want to just build out one large cluster and be able to flex that over time. They might want GPUs in diverse places like an edge environment. So, all those things that we've seen that you need it from a technology point of view to deploy are already existent in Ethernet. So, that's kind of our view. If this isn't big and it's just going to be a small number of clusters in proprietary then maybe what we have is good enough. But I think this has legs.
Mike Genovese
Yes. How should we think about the fact -- I mean when Cisco just reported, right, and they talked about $500 million in orders, I think that they were talking about for chips, right? I mean, for Ethernet fabrics, I think they're really talking about something that would compete with your supplier before we compete with you. So, I just -- I mean, the fact that Cisco is more vertically integrated in their business and that they've always been one of your key competitors. How do you think about your focus on what you do versus, I guess, a little bit broader focus?
John McCool
Yes. I mean, there's portions of their business, we don't compete in at all. We just beat in that networking core segment, right? And I don't think that we have seen the strategy being much different than it has been. A lot of announcement around Silicon One, but that's consistent with how Cisco has done business for years, developing their own silicon and integrating that into products.
The optical -- they always had an optical business. Optical has become more of an integrated business with optical components. We don't compete with them in the optical space, we support multi vendors, including Cisco. And in terms of execution, I don't think we've really seen them be a chip supplier, they're still competing with us on a system-by-system basis because that helps.
Mike Genovese
Okay. And then I guess the -- I mean the relationship with NVIDIA, how would you characterize how you interact with them? Because I think that in the future, in this open environment, they'll be supplying a good portion of the GPUs and a smaller portion of the networking and you'll be supplying a bigger portion of the networking, and you guys will probably have to get along, right? So how do you think -- I mean, what's your relationship with them?
John McCool
I think that's a good way to frame it. I think we got to get along. We want to support anything that's connected to the network. And to the extent that they want to sell GPUs in an open environment, we're happy to engage. And I think we don't have a -- there's no friction in the relationship. Obviously, there's this InfiniBand versus Ethernet piece today, but I don't think that we're focused primarily on different things as companies, right?
Mike Genovese
Yes. Well, I guess I don't know the Company as well. I don't cover that sector, but it seems to me that they're -- I mean, they're clearly enjoying benefiting from the Mellanox acquisition. But are they -- do you see them taking steps to kind of extend the lock-in period? That's -- I don't, but I'm wondering if you if that's something you guys see out there with if you're competing against a vendor who is trying to hold on to this monopoly longer or if they're really -- I mean, if -- I think if they were supporting their customers, which they are, they're willing to -- it's -- I guess it's not up to them, but I mean my question is really -- I'm stumbling over my question here, but how do you see them operating out there? Are they going to play nicely?
John McCool
I mean, look, I think they're competing on more of a vertically integrated approach, which doesn't just include the network, but all the way up the stack to what they're doing with Kudo and really trying to take advantage of the interest to deploy quickly and roll out racks and a system that you can build applications on. And that seems to be their focus and InfiniBand comes along with it to the extent customers want to push against that and open this up to the multi vendor. I think that's going to be a very customer-driven approach, and I would anticipate that we will get along.
Mike Genovese
Okay. Great. Let's back it up and take it to a higher level. We got a question from the field. We're not right up against time, but we're getting towards the end. So I think that this question kind of gives us a chance to combined a lot of what we've talked about and go back up to kind of the summary view. But it is from a longtime shareholder who wrote in and said, Arista is clearly winning in its markets from enterprise to the Titans. But what is the main reason for this is? Is that the radical simplification of the solution versus competitors? But I sometimes struggle to explain this to my other portfolio managers and people not as close to the market. So this guy wants to tell his colleagues what makes Arista special and why you guys are successful in the market?
Jayshree Ullal
I struggle as well because the answer is so boring. It's hard to imagine, but deploying networks is difficult and been fraught with a lot of frustration by the people who operate the networks. So as we sell into either a cloud provider or an enterprise, there's somebody who has to get up every day, run that network, make sure it's secure, deal with upgrades, deal with ads, moves and changes. And their life has not been that great.
Networking was very slow to evolve to a software-based approach with APIs. So you couldn't add your own management stack or do integration on the switches. It was also fought with a lot of bugs. So if you upgrade with the story that we get from multiple customers, I want to do an upgrade because there might have been a security vulnerability in the kernel. And that upgrade didn't go well.
First, the system crashed, I had downtime. Then when I finally got up and running, a feature that previously worked no longer operated. And we just focus on that problem by the way we have constructed our software, the way we automate the testing. And I think that the opportunity and challenge for our sales team is it's easy to say that I have better quality, but everyone says that.
You actually have to experience. So they'll encourage a customer to adopt Arista in a use case, please call our customer support. That's the second question. Call them up, ask them some questions. And it's just a better -- it's a better user experience at the fundamental level.
Mike Genovese
Yes. And it seems like -- I mean you really obviously focused on the highest-performing networks, and I'm still kind of struggling to understand how kind of out to the edge in the enterprise you go. I mean there's a focus on the Fortune 2000 who probably are the most likely to have their own data centers, if they do that or the biggest campuses. So -- and I guess just over time, should we expect you to keep moving further and further out to edge solutions? Or I mean, as you mentioned earlier, you have so much opportunity just in what you already do. But help me think about how far to the edge and how kind of much -- to me, that implies much more distribution, right? And so that doesn't really seem to be your focus in the near term. But if we really look over the long term, do you think you'll be much more?
John McCool
We see that more of a function -- the distribution of the account sizes that we go after. So if you go after a Fortune 2000 company, typically, they're driving standards to be used globally across all their all their operating space. And they'll want from a provider like us to address as many use cases as possible so they can reduce the number of vendors that they work with. So the edge is definitely important. If you think about health care, you have large hospitals, but they're connected to regional centers, the strip mall down the street from me has a Sutter Health, a little building there, that edge deployment, so it's vital. It's key to their business. Now as you go to sort of less global, more regional mid-market providers, there could be a different go-to-market approach, but we're not saturated with that Fortune 2000 yet.
Mike Genovese
Yes. And just a couple of things just on the business model to wrap up here. So, it seems like -- I mean when we think about leverage going forward, '24, '25 revenue growth gross margin expansion or operating leverage. I mean what do you think will make the most impact on the operating margin?
John McCool
I think that's difficult to break down. I think we talked about gross margin and really kind of driving back our gross margin part of that's mix, but we're also recovering from some supply chain pieces. So there's definitely opportunity there. I don't have any other color on that.
Mike Genovese
Yes, yes. And so I imagine that the search must have started for Ita. What's just the time line on the transition for...
John McCool
I think Ita is committed to help us do the transmission. I don't think we announced any specifics around the time line at this point. Liz, I don't know if you have any...
Liz Stine
Yes, I think it was more around the intent to retire in 2024 from the earnings call, but she will be helping us find her replacement.
Mike Genovese
And then, I mean, you've got a very, very strong balance sheet. You guys generate cash in your business, inventories seem like they have to come down from a high point. I mean they haven't come down quickly yet. But at some point, they really do have to drop, and that will be cash. Just -- I mean, I guess, you're the Chief Platform Officer. So maybe capital allocation is not your area, but we have Liz here. I mean what should we think about you got -- like are you guys going to become a bank or what are you guys going to do?
Liz Stine
I mean I think that it's important for us to maintain a healthy balance sheet. We do have some aggressive competitors. We've got to fund the business. We also invest in the investment business, whether that's R&D, whether it's sales and hiring and targeted sales and marketing, et cetera or into M&A, although kind of trying to find that perfect M&A, I think everybody has always asked us if you do something big. I think bigger is a little bit more of a challenge.
And I think that -- we've been pretty -- we've given you some good ideas around our M&A strategy with the types of M&A that we've done, kind of the smaller tuck-in network adjacent because we're really focused on the network, right? So we want to have the right to play in those spaces. And then obviously, return cash to the shareholders. And we do a share buyback. We still have some to execute for the current base, and then we'll obviously work with the Board on what that looks like to be Board.
Mike Genovese
Well, great. I really enjoyed talking to you guys for 45 minutes. I learned a lot, and it was interesting.
Jayshree Ullal
And the Ethernet versus InfiniBand was important.
Mike Genovese
No, it was important. We could have another conversation about it another time, I think. But John thanks so much. Unless there's anything else you guys want to say before we go. I think we call it here.
John McCool
Appreciate the time and the opportunity, Mike. Thank you very much.
Liz Stine
Thanks very much, Mike.
Mike Genovese
Okay. And thanks for everybody for joining us, and I look forward to seeing you at more firesides. Thanks, bye.
For further details see:
Arista Networks, Inc. (ANET) Rosenblatt 3rd Annual Technology Summit (Transcript)