All right, we are going to go ahead and get started with our next session. It's the home stretch here. I'm Matt Niknam, networking and communications analyst here at Deutsche Bank. We're very pleased to be joined by Arista's Chief Platform Officer and SVP of Engineering and Operations, John McCool. John, welcome.
Matt, thanks for having us.
Great. Great. Can you talk a little bit, maybe just to start, around your role at Arista and some of your key responsibilities?
Sure, yeah. So I have two different hats, primarily: hardware design, development, next generation products, and then following them through to manufacturing, supply chain. So pretty much how we decide to build the hardware and how we get it out the door.
So you've been busy the last couple of years, I could add.
A little busy, yeah.
As we sort of look to close out the year, can you talk about some of your top priorities and where you've been spending the most time recently?
I think for us, I mean, obviously, the whole thing around AI has been extremely exciting, and we launched our 800 gig products a few months ago, so you know, making sure they're successful in the market, strong adoption. On the front end, then the team's looking you know, a couple of years down the line and what's coming beyond that and starting to look at architectures and next generation products.
Yeah. Got it. Let's start. You obviously serve several different sort of customer cohorts, but well, let's start with the cloud titans. Can you talk about what you're seeing from your cloud titan customers right now, just in terms of underlying demand? And I assume there's sort of an intuitive answer, but maybe we can sort of unpack that and talk about-
Sure. Yeah, I think, you know, if we, if we were here this time last year, there was a lot of thinking about: How am I going to address AI? What was the meaning of that for my infrastructure? A lot of discussion, was Ethernet even an appropriate technology for next generation AI data centers? Customers have worked through that. We've seen a lot of trial activity, followed by proof of concepts that we're, we're in now. I think there's an increased confidence that Ethernet is the right underlying technology for these new architectures. And then you see them looking ahead, where these lines of...
We had a lot of questions from investors today about, "Okay, what about your AI piece of cloud, and what's that mean for your traditional piece?" I think that's gonna be a hard question to answer a few years from now because they're looking at converged data centers that are both cloud and AI. How do I bring this together in terms of a single, you know, deployment holistically?
Got it. Well, so Arista's laid out a target for at least $750 million worth of AI-oriented revenues next year, and I think that's relative to a much smaller base this year. I think you've quantified it as a low-to-mid single digits as a % of total revenues, just in 2024. So obviously, pretty material growth. Maybe just to start, can you talk about what the target encompasses and the ramp as we think about this over the next 18 months to get to that number?
Yeah, the target question is really important 'cause we defined it, I think, fairly narrowly. When we looked at the opportunity, we felt very confident, given our position with cloud networks and our position in high-speed market share, that we would do well with the upgrade to AI, you know, as we have done in previous cycles. But there's a new addressable market that we call the back end. A lot of people now are calling it the back end interconnect. It's a network that connects the GPUs together, and it's very performance intensive because when you have these clusters performing a task, the task is incomplete till the last GPU sends in its last packet and completes the job, right? So that represented a new opportunity, and it was not obvious that that would be an Ethernet opportunity.
So we talked about $750 million being the back-end opportunity for us, for the piece that we participate in, which is the switching piece. So we see other people that define that back end a little bit more broadly, with NICs and optics. Our piece of that's the switch. The separate piece is a front-end network, so that's the interconnect of these GPU clusters into the larger data center architecture. If you're gonna interface with AI, you're going to send it through a packet that goes into the network over the WAN and into a cluster. That's another opportunity, but that's not part of the $750 million.
Okay. Okay, great. Can you talk a little bit about the visibility you have today around cloud titan spend and maybe how much this is compressed relative to the peak of the supply chain crisis?
Yeah, that's a great question. I think, you know, when we talk about sort of our purchase visibility in terms of procurement and planning being a couple of quarters, on the flip side, there's an architectural technical engagement that's multiyear around what the next generation is. How do you build those products? How will they be constructed? What topologies or software do they need?
Got it. So, I mean, relative to pre-supply chain crisis, are you sort of back to where you were in terms of these lead times and visibility, or is this sort of a permanent change?
Yeah, that's a little bit closer to the original, but a little bit different in terms of the intensity and planning around AI. And just to step back and look at the supply chain in terms of what we procure, we may be in a new normal as well. Lead times have not gone back to their pre-COVID position, roughly on some of the things we purchase, about twice as long as they were, but they're very predictable. During COVID, if somebody said they're going to provide you something, may or may not show up. I think the supply chain's become very reliable.
Great. Great. I want to talk a little bit about some of the trials you're doing for AI. You've talked about several trials around AI clusters you've participated in, moving to pilots this year and ultimately broader deployments next year. How do these typically progress, and where are you today?
Yeah, we're inventing the new typical. We've never done this before, right? New AI, you know, I think people have gotten confidence around the proof of concepts. They're still looking at how to effectively deploy their workload. What does this infrastructure mean in terms of the performance of their results they're getting? Whether they have some opportunities to optimize, might not necessarily be in the network itself, but could be about how they partition their jobs between the clusters. So that's kind of the analysis and the functionality and how well it's working. Also, in these proof of concepts, how do I maintain these clusters? Which is starting to become a more interesting thing around the network. If something breaks, how do I determine what broke? How do I fix it? How do I repair it?
How quickly can I address that?
And these are all sort of hyperscale cloud titan-esque customers, or are there other sort of non-traditionals, maybe tier two clouds that are involved?
Yeah, absolutely. When we talk about the four out of five customers that went the Ethernet direction, there's a mix of customers in there. And we have seen some enterprise customers also, albeit much smaller clusters, starting to take a look at what AI means for their business.
So I assume, sort of, I'm gonna get into the InfiniBand versus Ethernet debate in a second, but who's part of the competitive set here that you typically run into? i.e., of these sort of four out of five clusters that you've won, who else is typically... if you could speak to maybe the competitive set you run into.
Yeah, I think we see our traditional Ethernet competitors that we had forever. Now NVIDIA comes into the mix, both with InfiniBand, and then, you know, they've talked more openly about their Ethernet solution as a piece of it.
Yep. One other item that showed up last quarter. So we saw the product deferred revenue line jump by about $250 million. That typically is relatively stable, so it's a little bit of a sizable jump, amidst some of these customer trials.
Yep.
I guess as we think about looking forward, can we talk about why the jump happened?
Yes.
B, when this may be expected to translate to actual revenue.
Yeah. So just to comment on product deferred, this involves customers who have taken a product but have some acceptance terms. So that could be related to a new use case, how they're deploying the product, maybe some new software functionality that they want to make sure operates the way they believe it will before we recognize revenue or new hardware platforms. And it does tend to follow new architectures, new products, so we are in a new phase with our eight hundred gig products and a lot of these new use cases that you can imagine. No single rule of thumb of when it comes off, and when you look at that balance every quarter, there's some customers who have accepted and some new ones that come into the mix.
Okay, so it could be pretty fluid.
Could be pretty fluid, yeah.
Okay. Okay. So maybe that's a good segue into the InfiniBand versus Ethernet debate. I guess, can you talk a little bit about how NVIDIA Spectrum-X Ethernet solutions affect your view on the AI networking opportunity, if at all? Maybe you can start there, and we'll follow up.
Yeah, I think it's important to think about how we come at the market from an NVIDIA or an Arista perspective. NVIDIA, you know, leading in GPUs, and their deployment and the architecture really coming at: How do I deploy a GPU-based solution and connecting that with InfiniBand or Ethernet? I'm not sure that matters so much what the underlying connectivity is. We've seen them lead for a long time with InfiniBand. We're happy to see they acknowledge that Ethernet has a role. They also provide optics for the solution, as well as the NIC adapter cards, the network interface cards that go into the servers, where they have a very large market share.
Our approach is, hey, the network buying center, the network developer, has to be able to analyze what's happening across the network, certainly in the GPU, where that stalling or some network effect could impact performance. You have to be able to debug that. And oh, by the way, you have to connect these clusters into the rest of your infrastructure. So we appeal to that network buying center in kind of the same traditional way we did, from a compute perspective.
What's the customer take in terms of receptivity then?
I think the network buying center really does not want to support a new architecture that's solely focused on the GPU cluster. At the business level, there's a tension between expediency and long-term, I want diversity of GPUs, I want mixed infrastructure, and I'm making my investments in the network for a long time. So that's kind of the dynamic.
I guess we had Martin Hull at our conference last year, and it was a bigger debate around, can InfiniBand, you know, I guess, can-
It was a very technical debate a year ago.
Yes.
Yes.
It certainly was. There were at points of the discussion where I may not have been able to follow as closely, but how has the displacement, if we think back over what's transpired over the last year, how has the displacement of InfiniBand by Ethernet played out relative to your expectations, i.e., faster, taking longer?
Ah! I think when Martin spoke with you, we probably didn't know, and now we feel much more with the four out of five, when we talked about the four out of five customers, I think that was our indication that we felt the leading edge companies that were really focused and forward-looking in this market were adopting Ethernet.
Okay.
Yeah.
Yeah. As we think about it, there's a couple more, just 'cause it is a bigger focus area. If we think about cloud titan spend, there's a lot of questions we often get around CapEx at the bigger cloud titans, and how much of that goes towards data centers, and ultimately within that, how much of that goes towards networking. Can you maybe help us reframe that relationship in terms of their spend, which I think if you look at your cloud titan customers, that's growing about 50% this year, and how that CapEx ties into what you ultimately see?
Everybody wants that magic function.
It's a $1 million question.
A function, if I could come up with that algorithm, it's hard. I mean, you know, when CapEx is going up, it's generally good, but we can't draw a straight line to the networking piece. There's real estate involved, more and more, there's power, construction. When it gets into the IT equipment in the data center, you know, our rule of thumb has generally been networking's about 10% of a compute storage element around it.
D- has the-
Maybe a little less these days as the GPU expenses have gone up.
Got it. Okay, yeah, that was gonna be my follow-up, is, is there any shift of what share of that CapEx you compete for that goes towards networking for these AI use cases?
Yeah, I think as a percentage, yes, but it is driving the whole boat. This is generally good for the networking business. It's driving higher performance, lower oversubscription rates, so it's... Even though the percentage may be shifted by how expensive those GPUs are, it's a good thing.
Got it. Got it. As we think about beyond sort of the eight hundred gig transition and maybe some of that tailwind from AI, is there a meaningful refresh opportunity ahead within your cloud titans base from sort of ten gig, forty gig, hundred gig towards higher speeds like four hundred? Or are they largely at the cutting edge at this point? I guess, like, what's left beyond AI in terms of evolution?
Yeah, I think we had a lot of questions around that. It's gonna be hard after this cycle to discern AI, non-AI-
Mm-hmm
... 'cause everyone's focused on converging their infrastructure to accommodate GPUs and next generation GPUs, along with their traditional infrastructure. So that's one piece of it.
Okay, okay. Just broadly, to tie together the supply chain conversation, last question on cloud titans, but it is maybe a broader supply chain question as well. Are there other areas, particularly within the AI build-out, that have the potential to be bottlenecks besides some of these headlines we hear around GPU procurement?
You know, when we had the 400 gig cycle, the first use case for 400 gig was really interconnect of data centers. And there had been a lot of deployment of servers moving from 10 gig to 25 gig, talking about moving to 50 gig, and the interconnect became a bottleneck. The second piece was people physically couldn't fit that many servers into their data center and needed to have more regional data centers that interconnected them. So I would suspect that the backbone at some point becomes more of a pressure point, especially in training environments, where you have a lot of data that may be stored in a different location that has to come into these clusters.
Got it. Got it. Let's pivot to enterprise. Sometimes I call it sort of the unsung hero. AI gets a lot of the airtime, cloud titans gets a lot of airtime, but can we talk a little bit about some of the key drivers behind Arista's success in enterprise, right? You've obviously had very sizable share gains, and I think you've had a little bit more in the way of positive commentary, even most recently around maybe some newer logo wins and looks you're getting from customers who haven't traditionally been Arista customers.
Yes.
So what's driving some of that success?
In contrast to cloud, where it tends to have some technology cycle element around that, the move to a hundred gig, four hundred gig AI, enterprises make their decisions based on their own depreciation cycle, maybe they're merging with another company and have to integrate their networks. There may be some event that's based on their decision to invest in a piece of the network. I think we have just made, over the last decade, very measured and consistent strides in building out our portfolio. Number two, we have more folks who have used Arista in a different company that have moved around, so the reputational aspect of it is definitely playing a role.
And then, you know, the third piece of this is just the quality of the product and its ease of operation with the single EOS image-
Mm-hmm
... and one management plane has made it easier for us once we insert to expand our footprint.
Any estimate just in terms of where your share is right now, broadly within the enterprise and where that can go?
You know, our share in high speed is very high. We're a market leader in hundred and above. A lot of that's cloud driven, right? But we see even our share of strong franchise wins for Arista being very small penetration in terms of what they're using for us and what the possibility, bringing in campus-
Mm-hmm
... or bringing in some of our new routing use cases.
... how does the sales cycle typically work? I mean, I guess maybe how does it differ in the enterprise to win over a new customer, go-to-market approaches here, relative to the cloud titans business?
I think there's two pieces of it. If it's an existing customer account, the team's very focused on winning new use cases. When they see an opportunity come up, there's a new refresh or there's some event, so there's a lot of activity, even with our existing accounts, to expand. And then looking at Fortune, we're still very Fortune 2000 focused, we're not really mid-market focused. Where do we have an opportunity from someone who is spending on the network? What's their next driver? Where do they need to invest, and can we intersect that? A lot of that has been around campus, but it could be a data center win or a routing win.
The piece that we have to overcome is just having people deploy the product and understand our claims around ease of use and upgrades are real, that you can hear that, but you have to experience it. Once we can get inserted, I think we have really good success in expanding.
Is demand there? I understand that. I know you don't necessarily report these by line item, but is demand within enterprise? I know that a year ago, there was a lot of shipment of product that may have been delayed because of supply chain issues, right, so a lot of stronger growth. Can you just talk a little bit about what you're seeing or maybe how the demand within enterprise or growth within enterprise compares to what you're seeing in cloud titans this year?
Yeah, we've talked about all of them growing.
Yeah.
We haven't gotten to a specific piece. But I, I would say the driver of that has been success built on old success, and the opening up of the parameters of additional use cases outside the data center, and being able to look at us as a full enterprise networking alternate to the incumbent. Before, we were data center people, and now we're viewed increasingly as an enterprise networking company.
As we think about sort of this eventual evolution for more training-oriented AI use cases towards more inferencing, how does that affect your business, particularly within the enterprise?
Yeah, I think it's tough to draw exactly what's going to happen with inference, but a lot of that would move to the edge and potential needs and bottlenecks going out to branch offices, campus environments. If you're really driving more data back to the cloud, do I have enough performance and things of that nature?
Yeah. So from a high level, and given your role as Chief Platform Officer, what's the secret sauce that stands out as a competitive advantage or differentiator that really enables the company to take the kind of share it has, both across cloud titans as well as your enterprise customers?
I think we represented a shift in a couple of ways. One was a more software-centric approach to networking. So the way we actually build our software, test it, and quantify it, has been more like a traditional software company. Networking software was traditionally tested with human beings, and I'm trying to test all the configuration. Ken and his team have done a great job through automation, and also the architecture being more robust to remove quality parameters. The second is, instead of focusing on features and functionality... You know, early days of the internet is like, "How will this even work?" and you needed to invent a lot of features to make voice work and video work.
At some point, that became a known operating parameter, and you had to focus on the people who deploy the network, and is it easy to deploy, is it easy to upgrade? Can I minimize the number of security alerts and just run the network? I think Arista has done a good job appealing to the people who run the network, as opposed to a more theoretical architecture-only approach.
As we think about just competitive backdrop and a little bit more maybe related to enterprise, we've seen some recent industry consolidation around your peer group. Has that affected customer behavior and their level of engagement with you, if at all?
Yeah, I think we're a little taken aback that we're kind of the last pure play.
A pure play.
There's a few pure play, maybe, but of our size, the largest pure play networking company, and we still have a lot of addressable market, so we're not in a position where we feel we need to move out of the networking realm. We still have share we can gain, so we're really happy about that. I think, in particular to HP and Juniper, I think that there are companies who have used them as alternate sources, and we're having more conversations as that's coming together in questions of product overlap, et cetera.
One other sort of, I guess, competitive threat that comes up from time to time is just around white boxing, and customers sort of doing it themselves. How has that risk of larger customers going that route evolved of late? And I would argue maybe it's weakened as a risk after some of the supply chain crisis and dynamics and some of the speed to market you were able to deliver.
Yeah, I mean, I think we see certain customers have invested in supply chain and design capability, have really embedded in that. Some have a mixed model, and some don't do it at all. We haven't seen a shift in their focus, if you will, in terms of their overall make versus buy strategy in a big way. So I think it's fairly consistent and, at the same time, fairly contained. We don't see new people really taking that on.
Got it. We talked a little bit about AI earlier in the discussion. We've seen other parts of this sort of AI ecosystem have margins get negatively impacted as they compete for hyperscale sales, right? Think about AI servers, some hardware companies reporting maybe lower profitability-
I see
... tied to AI. Are there material differences you see for your AI-oriented products as it relates to profitability?
Yeah, not at a product level, but, you know, we have always reported more margin pressure when we move our mix to the cloud. So we don't see this AI being any different than what we had done previously, and margin mix can just shift based on customer as opposed to product.
That's at a gross margin level, I believe, right? Historically, you've talked about enterprise-wide-
Right
-gross margin.
Right. Operating margin, you know, we have less coverage in the cloud accounts, so it—we think it tends to net out on the enterprise side because we have more sales coverage and-
Yeah
more, more pieces there.
I want to talk a little bit about campus. You've obviously. I think you've got a $750 million target for 2025, I believe. Any updates in terms of how that penetration into that market's going, and are you seeing any more sort of cross-sell benefits or opportunities as you continue to invest there?
Absolutely. I think, you know, we broadened out the portfolio in the enterprise in the last year to take on WAN transit and connectivity into the branch. We've strengthened that offering around ease of use and things we've done in CloudVision, based on customer feedback, to make it easier to deploy and simpler. We've done a good job integrating the network detection and response capability that we acquired through Awake, and can actually position good differentiation around segmentation of our campus solution and NDR together, and we talked a little bit about what's happening with HP and Aruba and some of the questions there, so I think we're feeling really good about, you know, what we've done in terms of just pulling the portfolio together, strengthening it, and now have good proof points with customers that we can reference with others.
As we think about maybe what could be a more, I guess, varied go-to-market approach as you maybe penetrate a little bit deeper into campus, relative to the traditional business, how should investors think about the incremental spend necessary to delve a little bit deeper into the TAM?
Yeah, we get that especially when we entered campus.
Yeah.
I think people have been conditioned that campus is a mid-market sell. You know, even our largest cloud customers have enormous employee bases and need campus products. So our center in the customer segment still is in the high end. Our campus offerings are oriented towards very large enterprises and their needs, and any investment in the sales side is to sell the whole thing, not just the campus. So getting more coverage to sell data center routing, Awake solutions into enterprise customers, as opposed to this is a separate-
Mm
-campus go-to-market.
Got it. Got it. Maybe on the topic of some of the adjacencies that you're beginning to maybe tap into, you've talked about areas like some observability, security, as adjacent areas that can expand the TAM. Is this sort of a material revenue driver today, or if it's not, you know, what's the timing around when these become more significant contributors to your growth?
It's tough to say that they're broken out as a separate contributor or also something that drives the base sell of the campus to begin with. We'll go in and talk about our campus environment and, you know, what you can do from a connectivity point of view, and then we see problems around, "Well, I have a high-performance application running over there, and I'm concerned about its performance." "Oh, yes, we have this observability piece." It's built into CloudVision, so if you invest in that platform, that's just a piece of the whole solution that you can bring in. So it's not like a disaggregated set of products that we have to go sell. It's all centered on the same stuff you would buy just to provide connectivity.
Got it.
And I think that's been kind of a key differentiation, is we're not competing against some point thing. We're showing how the network can be used to provide that point of observability or security piece in it.
I was gonna ask, so obviously Cisco, huge incumbent in the space, they've talked more recently, a little bit more around leveraging that scale and incumbency and networking alongside maybe some of the security and observability or newfound scale in security and observability after the Splunk acquisition. You know, I guess you still are a relative share gainer within enterprise, but does this maybe broader platform approach from Cisco, has that presented any sort of incremental challenges in your selling motion?
Not really, no. I think that, that's a separate kind of solution that might be even broader-
Mm-hmm
than the network, where we don't really compete. We're just focused on the observability in that network
Yeah
and how you can look at that. So it's more. I believe our approach is much more adjacent.
Yeah
to that network buying center.
Yeah. The one piece, the one customer vertical we didn't talk about, and it doesn't come up as often these days, but service provider, and I guess you can even weave in sort of tier two cloud. What's the latest you're seeing from these customers in terms of demand?
Different dynamic between the two.
Yeah.
We lump them together-
We lump them.
But really, what's happening is very different, right? We see tier two cloud looking at what's happening in the large cloud providers, making investments and, you know, figuring out how to do that, and we're very well engaged with those customers, and it seemed to have really good fit in our product. On the service provider, coming from running their legacy systems, we weren't traditionally focused on replacing an incumbent that's providing two decade-long MPLS support and trying to insert in that business. We definitely look for new opportunities when they're wanting to move to a more cloud-based solution. We've had good success in the data center with a lot of those cloud providers, some of the MSOs that have done more significant rearchitecture. So, you know, we continue to invest in the routing stack. It's more full, but that market hasn't moved.
as robustly as some of the others, I would say.
On the Tier Two cloud, is there a bigger AI opportunity there, or is that $750 million mainly within the sort of cloud titans realm?
It includes a mix of people you know, and people you don't know.
Okay.
So leave it at that.
Got it. It's an interesting question. I have two last questions. First one, what do you think is the most underappreciated aspect of the Arista story by investors from your seat internally?
You know, we touched on it three or four times in this conversation.
Yeah.
And I'm not sure it's underappreciated, but the cloud and whatever new tech it is, is a very broad-based appeal of many different companies that are playing in that space, where our enterprise wins and share gains are an Arista-specific piece. And I tend to think that it's underappreciated 'cause it happens over many product introductions, many customer wins. It just is a little engine that keeps going. And, you know, even today, talking about, you know, we had recent wins in enterprise, if you talk to the team, they'd feel like, you know, I laid the seed for that three years ago, and I've been working on that ever since. So I think that that momentum, because it's not a one big pop, it's just the continual growth, that's a little bit underappreciated.
Great, and last question, sort of tie this all together: if we're sitting here a year from now, what would you like to look back on as key milestones and achievements for Arista over the next twelve months?
I'm already looking back on our June announcement of our eight hundred gig product. There was a ton of people that worked many years on that, and it all came together with three product announcements at the same time. So that was definitely a key milestone. I think that we'll see, if we're here next year, deployments moving from POCs to, you know, actual, real wide-scale production environments in AI, and continued franchise wins and growth within some of our key enterprise customers.
Great. That's a great place to end it.
Thank you.
Thank you.
Thanks for the time. Appreciate it. Thank you.