To, for any research disclosures, please see morganstanley.com/researchdisclosures. Reach out to your sales representative with any questions. I'm Meta Marshall. I cover the networking space here at Morgan Stanley. We are delighted to have Arista, Jayshree Ullal, and Ken Duda, a new special guest joining us on stage, who's President, and CTO as well. Jayshree, you know, welcome back. It's been a phenomenal couple of years for Arista since we last had you on stage. Just how do you think the core, like, value proposition of Arista has changed with kind of AI coming into the framework?
Meta, it's always good to be here. You must be my good luck charm. Every time I come, if we grow like that, I'll keep coming.
Yes.
The value proposition, and Ken Duda, our Founder and President, started this, hasn't fundamentally changed.
Mm-hmm.
Of course, the use cases to make it greater has changed. We always started with the belief that mission-critical networking needs a foundational software, which started in the data center, as most of you know, and we're now the number one market leader there. Great technology, great EOS, great merchant silicon, and just a great product, but more importantly, a great system. We began with this belief that you had to build a great system with our leaf-spine architecture. The leaves got plentiful. Our universal spine connects today to, you know, campus leaves, enterprise leaves, branch leaves with our VeloCloud acquisition, and of course, I have to say AI.
Mm-hmm.
With the advent of AI, we have been able to build this very unique architecture that no one else in the market has, which is an all Ethernet AI spine and leaf. That takes advantage of all the goodness that we brought to bear in the data center.
Mm-hmm.
I call this the centers of data era, where whether it's a data center, campus center, WAN center, AI center, we are able to build that uniformity in our network and yet deal with the heterogeneity that's coming in front of us with different frontier models and different AI accelerators.
Got it. You know, we get a lot of questions about blue boxes, branded boxes, white boxes. Just what do you think that conversation, I'm sure you guys get many of the same questions, like, what do you think that misses about the value proposition of what people get from Arista?
Yeah. I'm gonna tell a little story. Ken, if you don't mind, you'll continue the story, although we won't censor your version versus mine here. When I started with Arista, I said, "Oh, we're gonna build another switch?" I was frankly thinking I'd done enough of that in Cisco. Why would I ever do another switch? What Arista has done is not just build yet another switch, but really build purpose-built network mission-critical platforms that have different use cases. Part of having those different use cases is we had to make sure we could take different Merchant Silicon and ignite it and make it better because silicon leaders need better software to ignite them.
Mm-hmm.
A white box would be a cacophony of an existing silicon, their SDK, open NOSes, and I'm not saying anything negative about that except to say the use cases are more limited.
Mm-hmm.
A blue box, and then of course a purpose-built EOS, is what Ken really sought out to develop. Maybe you can add a few words on that.
Yeah. Look, I think that what we've shown is that in networking, software is tremendously important, that without the right software running across the network, the network just does not have the reliability, stability, flexibility that it needs. By having one operating system, EOS, Arista's extensible operating system, running across the whole fleet, whether it's campus, cloud, WAN, data center, all running the same software, it simplifies things so dramatically for our customers that they're able to qualify one release, have one platform to automate against, and it's just, and its reliability applies across the whole thing. There are use cases for white box, but that requires advanced engineering on the part of the customer. The customer has to figure out how to integrate white box with software into their overall network management suite.
It is not for the faint of heart, okay? We're seeing that the vast majority of companies need a network that works, they can count on, and EOS, one OS is the path to the simplest, most reliable infrastructure.
Got it. I mean, right now there's a lot of new data builders or data center builders in the market. You know, they all might have certain aspirations of what they think that they can do. You might have better insight into what you think that they can do. Just where do you find that kind of, mesh of what is the best product for that customer?
Well, I think we're all struggling with one common problem, power.
Mm-hmm.
Right? If so, if these data center builders can actually find the location, the colos, and have the power to ignite it, more power to them really, right? That's become the biggest scarcity.
Yeah.
Today we're not talking about building a megawatt data center power, right? Everything is being translated into hundreds of megawatts and often gigawatts of power.
Right.
Because of not just these CPUs or compute as we know it, but these powerful accelerators, whether it's GPUs, TPUs or the AMD accelerators. The pressure that's putting on the power.
Mm-hmm
... is really making us work with these builders and models to come up with at least 3 use cases that we see.
Mm-hmm.
One is scale up, which is how you scale all of the compute capacity, especially in the back end of an AI accelerator network. Typically a scale-up configuration is a more limited configuration within a rack, if you will, where you might be connecting 100 or 1,000 or a few thousand GPUs. Power is somewhat constrained there. At the same time, the surface area requires you to pack dense amount of cables and optics.
Yeah.
In that scenario. By the way, a lot of times, people will use proprietary technologies like NVLink.
Mm-hmm.
PCIe switching. We're big believers of Ethernet, and we believe we're gonna see a Ethernet for scale-up, ESUN standard coming out this year that'll further that. The other is the scale-out, which is how do you connect all of these racks together, and this is where Arista has really flourished. A lot of our AI business is coming from the scale-out configurations, but that takes advantage of all of the rich protocols, the telemetry, the availability, the visualization, the intense features we've been working on for the last 15 years, and in particular for AI over the last three to five years.
Mm-hmm.
Again, that's a power hogger, and every kW matters there because it's not just the GPU that's contributing the most, but the compute, the storage, front end, back end, and obviously the network as well. What we're seeing more and more of with these colos is they're not able to get the power, and therefore, they're building many more distributed AI centers.
Mm-hmm.
Because if you can only get 1 GW, you take the gigawatt, you know, you build as large as a stadium can build, and then you move across to another one. This distributed scale-out of XPUs is also a huge use case for us because now not only do you have to deal with the different types of accelerators, but you have to deal with isolating them, segmenting them, Traffic Engineering across them. Again, routing across these data centers in a coherent fashion, short, long distances. Power is a culprit in all three, but I think in order to deal with the power, we're coming up with different ways to centralize or distribute the designs.
Got it. You know, with this now massively growing TAM, there's a lot of kind of competitors in the market, whether from white box, branded, silicon, software. Just how are you thinking about maintaining and expanding your ability to kind of capture your share of this TAM?
Look, I'm gonna point to Ken first and say if you don't have a network that works, and an innovative, differentiated product, we can never maintain share. I think what the company has done throughout, not just now, white boxes have been with us since we started shipping products 15 years ago.
Mm-hmm.
-is to always coexist with it, build our value chain across... Like I said, we have 22 Etherlink products in AI that have nothing to do with white box. They all complement that with the use cases.
Yeah.
I think a great product platform is our greatest differentiator. Having said that, the other very important part of how we'll be working is with our ecosystem of partners.
Mm-hmm.
We see a world where it won't be just one homogeneous, you know, AI accelerator, but it'll really be a heterogeneous world with TPUs, AI accelerators. Some of them may be built by specific vendors like AMD. Some may be built in-house by our cloud titan customers themselves. Having a common homogeneous network infrastructure for all of that heterogeneity is going to be very important, as well as the model builders.
Yeah.
I don't see a world where it's ChatGPT is the winner or Claude's the winner or Gemini is the winner. Certainly, those three are going to be prominent, but there's going to be many more. Again, just as we lived in a multi-protocol networking world, we're going to live in a multi-model AI world.
Got it. You know, another question that we often get, I'm sure you're tired of, is just, you know, you've had leading margins for a long time. How do you keep margins against this backdrop, and just where do you think those additional areas are for value capture?
There's two aspects to our margin and maybe even three. One is what is our hardware cost and how much will a customer pay for that? You would be surprised to know that there's significant value add and differentiators in our hardware. We don't just throw a bunch of chips. There's a huge amount of signal integrity. Every power, watt of power we save translates to millions of dollars.
Mm-hmm.
Every latency, megabit, nanosecond we're saving translates to millions of dollars. I think there's a gross underestimation on the total cost of ownership of providing better performance, better hardware, high radix, et cetera, right?
Mm-hmm.
Ken's team I think your hardware engineering team is now doubled or tripled for AI. There's another whole aspect to it now, which is liquid cooling.
Mm-hmm.
Don't underestimate the power of amazing hardware designed over the last 10, 15 years for the right performance, availability, power constraints that therefore the customer is willing to pay value because they don't want. These are not throwaway toys.
Yeah.
These have to be there with them for five to sometimes even 10 years, right?
Yeah.
Then there's the software. I'm gonna lean on you to talk about some of the huge differentiators, the combination of which often a customer looks at and says, "Yeah, you're worth the premium, not because I want to pay you more, but because you don't come down, and you have higher quality and the just better products.
Yeah.
The software is absolutely key to where the margin comes from.
Right.
in certain segments of our market. We have many we address many markets, as you well know.
Yeah.
From the hyperscalers to big AI companies, but also into the enterprise, and government, healthcare, and so on.
Mm-hmm.
Especially on the enterprise side, the margin comes from the fact that we have better software.
Yeah.
It is easier to deploy and operate. It's consistent across the whole suite. It works, it's all managed through a single CloudVision management console, which manages the whole estate, from the campus to the data center to the WAN into the cloud. When you're able to deliver that kind of consistent experience, a network the customer can really count on.
It's just, it's so much easier to manage and automate against. You can command a higher margin.
Got it. I just wanna add one thing though to that. Sometimes we don't command a higher margin. I wanna be clear too, right? In very large cloud tighten situations where they may be using an open NOS, certain products are lower margin. You guys get to see the mix of our high, low, and medium margin. It's not always that we have a perfectly high margin, but the combination of it looks good to you. Yeah.
Me. All right. Perfect. you know, you just spent some time talking about scale up, scale out. Scale across opportunity has gotten a lot more attention of late as a lot of traffic becomes north-south versus east-west, and training gets done between data centers 'cause of power constraints. Just how do you see that opportunity for scale up emerging for you guys? Or sorry, scale across emerging for you guys.
Scale across? Yeah. We have been pleasantly surprised by the adoption of scale across. Think of it as a two-step or a one-step removal from connecting the XPUs, right?
Yeah.
Even in customers where we were not connecting their XPUs because they had their own optical switch or they already had a prior design, we're now getting a unique opportunity to connect to their accelerators one hop away.
Mm-hmm.
Through these distributed data centers that they have to build because, as I said, the name of the game right now is multi-tenancy. It's not just one homogenous.
Yeah
AI accelerator, as they wanna access these colos, they don't have a product that can do multi-protocol routing, EVPN, you know, segment routing, traffic engineering, simple things like security and encryption, and also MSS or multi-domain segmentation becomes important because I don't want the traffic from one set of accelerators to talk to another.
Mm-hmm.
I have to do all these AI build-outs. The flagship product for this that Arista has been developing for some time is our AI Spine, the 7800. We launched it late last year. Running at 800 gigabits and providing this kind of real-time throughput with a rich set of features is very unique to Arista.
Mm-hmm
... and describes the value of Arista in a tremendous way. I could just emphasize something Jayshree said.
Yeah.
It's that combination of having the raw performance characteristics that AI workloads require in terms of the load balancing, the buffering, latencies, reliability, delivery of packets. It's a sheer throughput.
Mm-hmm.
Having those sort of low-level traffic characteristics at the same time as you have all the advanced features that she was talking about, the routing, the security, the isolation and segmentation, Quality of Service, having those together on one platform all managed coherently is a good differentiator for us.
Yeah.
It took us about 3 to 5 years to build that.
Yeah.
It's a non-trivial effort.
Okay. All right. I'm not gonna make it out of this room alive unless I ask about the cloud titans. You know.
We want you alive. Yeah.
Yes, exactly. You know, you made waves on the earnings call noting the potential for one to two more 10% customers this year, while still seeing growth in your other two cloud titans. Just, you know, that is a big statement. Just how should investors think about these kind of two more emerging kind of customers?
Look, I'm gonna keep working hard to increase the denominators so that we don't have 10% customers. The reality is these are large purchasers. These are longtime partners. I fully expect customer A and B to continue to be 10% customers. Maybe, maybe their percentages will vary by year. The reason I expect we will have. They might be high single digit or a 10% customer is the spend is just tremendous, right? I know you guys track CapEx more than I ever do. You know, while our CapEx is very small compared to everything else they buy, it is very clear to us that the combination of the front end clouds-
Yeah
... now the AI as a huge copilot assistant to it, is having an effect not just on our titans, but I would also say on our AI specialty providers and some of the neo clouds.
Mm-hmm.
I think our base just got bigger.
Yeah.
Our opportunity and TAM doubled in the last few years from $60 billion to $105 billion. You would expect us to cross the $10 billion mark this year. You know, we were $9 billion last year. I think we've signed up to north of $11 billion now. That won't be enough. I know we have to go into the teens and beyond. Naturally, that kind of results expects a customer base therefore that will spend.
Okay.
We're certainly very intimately involved in co-designing and making that possible.
Got it. Kind of a couple follow-up questions that I've gotten you know, since the quarter and since that announcement is just, you know, would all four of these customers kind of be the same use case? Does it highlight the breadth of Arista's portfolio? You know, is it the same as these kind of 100K clusters you guys have been talking about for the past couple of years?
No, not at all. I think if I look back two years ago, the benchmark was how big and bad is your GPU?
Mm-hmm.
We too got strung along by that. Today, I would say it's not how many 1,000 GPUs am I building in cluster, it's more what's the aggregate going to be? They may build small ones and then go scale across and therefore have 1 million GPUs, or they may start out with small racks and scale up. We fully expect to see all use cases for AI shine in this numbertwo
Mm-hmm.
Scale up, scale out, and scale across. More importantly, we also expect to see the front end of the cloud connecting to them as well and get some refreshes. It isn't just isolated AI, it's a combination of AI and cloud networking as well.
Okay. You know, another question that we get is just is it always going to be the end customer that counts as the 10% or some of these model builders that might be using kind of different data center builder partners? Kind of who is the customer and who's kind of mandating that networking decision?
I think we've taken a pretty pure approach to it. We don't count the influencers, we count the end customer.
Okay.
Right? The influencer could become a large end customer over time.
Yeah.
Or decide to do something different themselves. We're pretty clear that the channel isn't the customer, the end customer is the one.
Okay. You know, We've talked a lot about the neo clouds. There's a lot of data center builders popping up, obviously kind of require some level of prioritization from you guys just in terms of.
Yeah. Sure.
kind of, how to judge some of those opportunities. How are you doing that?
I think this is an excellent question. I'll kick it off. You know, we do have a vast number of requirements clearly from the AI sector. Don't underestimate our commitment and investment to the enterprise.
Mm-hmm.
We've got a very large set of customers now over there. We have over 10,000. Our specialty providers also include service providers. We keep them all in balance, and we look at this as not just a near-term priority, but who and what are we in three to five years so that we don't get seduced by the hottest thing we can do now and forget them.
Yeah.
Ken's team works very closely with Todd Nightingale, our co-president, and myself and the product management teams to prioritize, prioritize. Because no matter how many engineers you add, it's never enough, is it?
Never enough. The other thing I'd like to add here is this is another example of what an advantage it is to have one operating system.
Yeah.
When I'm building new platforms for AI use cases, when I'm building new software features targeting a hyperscaler operator, I can then leverage those in other segments as well.
Mm-hmm.
We're getting that sort of, that alignment between our different efforts. We are able to do both at the same time because we operate as one big software team on a common software platform.
Got it. We get a lot of consternation from investors around Nvidia and Spectrum-X, particularly at Meta. Just are there any meaningful changes to how you see your relationship with the cloud titans and maybe kind of how you see Nvidia as a, as a participant within the networking market?
Sure. Well, first of all, shout out to Jensen. He's just done a tremendous job of building the world's best accelerators. That being said, we are going to work with all accelerators and build the world's best networks.
Mm-hmm.
Especially if they're Ethernet. I think the two can coexist, and I'm really grateful to see this whole bunch of accelerators crop up because it gives us greater opportunity to connect as networks. You know, Nvidia is very much a partner when it comes to the accelerator side. Nvidia is very much a competitor if you look at their Mellanox division, and they're trying to push NVLink or InfiniBand or even their own version of Spectrum-X, where they have a natural tendency to bundle everything together. It's only normal, right? I like to focus on the positive side of this because there's so many more accelerators to connect to than worry about the vending. More specific to Meta or any other partner, our position and our partnership has not changed. We continue to co-develop with Meta.
Our partnership is as strong as ever. You know, we see occasional blips depending on their purchase behaviors when they skip a service cycle or whatever. You've seen that in the past. You know, I think Meta is just being smart about working with multiple accelerator vendors and making sure that their network partner of choice continues to be one they can closely work with.
Got it.
I could expand on that just a little bit. I wanted to point out that in the context of Nvidia, they've been very successful at the backend networks interconnect GPUs. Our cloud titan customers, certainly including Meta, have much more broad and diverse networking needs than that.
Mm-hmm.
If you're talking about the backbone, talking about the data center spine, you're talking about the access networks, that's a place where I think Arista is well differentiated and has a very strong position.
Yeah. That's a really good point. Backend and front end have to come together, and almost always Arista is the chosen one in the front end.
Exactly.
Got it. I know you've kind of talked about this, Jayshree, that it's getting a little bit harder to kind of distinguish what is AI, what is traditional networking at some of these customers. Do you think that kind of, you know, people just get too obsessed with, okay, what is the AI number versus just kind of what is the cloud number in general?
Yeah. I agree with you. At first, I tried to be very pure and say, "Only the first connection to the AI accelerator is backend." I found that people were counting optics as AI.
Mm-hmm.
Everything is getting whitewashed with AI, right?
Yeah.
We're still trying to be very careful, and look at AI as at least the first and second hop. I do think people need to pay attention to if the AI spend is really high, chances of them doing a refresh on the front end or the cloud may get delayed.
Mm-hmm.
They may do it ahead of time, like one of our cloud titans did, before they do the AI refresh. Not all of these happen together. They happen in waves. If we have a first wave with AI and we don't get much cloud spend, remember, there's time for that later.
Right.
Right? They do go together, but they don't happen at the same time. That's an important thing.
Okay. Got it. Maybe talking about optics for a second. You know, we've heard a lot about Optical Circuit Switching, and the opportunity for optical vendors. You know, at least some of them talking about kind of partially being at the expense of spine switching. Just where do you see kind of Optical Circuit Switching fitting into the market? You know, are they more symbiotic than people think about?
I think they're more symbiotic. There may be a couple of use cases, but we have seen optical switching mostly and predominantly. Then a couple of customers in scale-up.
Mm-hmm.
We don't see any optical switching in the scale across, although you can put coherent optics if that's what you call, but that's not quite optical switching. Where there are 7800 spines, there almost always is a ZR+, or some sort of coherent optics that we're connecting to.
Mm-hmm.
Other than that, we've not seen any disruptive designs that make optical the only one choice. There's a simple reason for that. People need not just the plumbing to connect, but the intelligence to get the performance, the packets, the forwarding, the Quality of Service, the security. They wouldn't risk their expensive AI clusters to just do layer one switching. They want much more.
Yeah. Maybe moving that conversation to CPO. There's been a lot of talk about whether we're gonna, you know, we've had little introductions kind of on the scale outside. There's kind of talk about there being larger introductions on the scale upside. Just where do you see kind of CPO timelines? How do you know... Andy has obviously been very involved in this over time.
Yes, he's been very vocal. I wanna say on behalf of Andy, he's not anti-
Yeah
... co-packaged optics. He's pro good co-packaged optics. That's what we haven't seen for the last 10 years. It's largely been in the labs, right? First and foremost, I think it's important to understand that the versatility of optics, particularly within a data center or across, is highest with pluggable optics. If you can build the right form factor like Arista pioneered with OSFP and continue that over 800 gig, 1.6T, 3.2T, I think you'll see that as the 80%, maybe even today 90% of the use case. It's so reliable, it's pluggable, it's versatile, you can change your mind, you can mix, you can match, et cetera. Now, there are places for co-packaged, both copper and optics.
I think co-packaged copper in a more limited radius within a scale-up, if you can be two to three meters, would be good enough and more cost-effective, less power. co-packaged optics, we will become real fans of it if we can make it open, and standards-based. We think that's super important. Otherwise, you end up having one vendor situations where one vendor is doing one type of CPO and another, and another. We're not fans of that, but we will really embrace CPO when it's more open and standards-based.
Okay.
If I could just emphasize-.
Yeah
... one thing that Jayshree just said there. The openness and interoperability is core to networking, and so many companies in our space attempt to capture the market with a single vendor proprietary solution. I can give you a long list of these. IPX, SPX, NetWare, DECnet, LAT, S-SNA, you know, Apple, AppleTalk, LocalTalk, NetBIOS, NetBEUI. The list goes on and on. It's a graveyard of technologies.
Mm-hmm
... because open and interoperable wins every time in the market. We're gonna see InfiniBand go the same direction, by the way.
Okay.
This is where taking CPO from a vendor proprietary technology to an open, pluggable, interoperable standard, like all of the other optical standards are.
Mm-hmm
I think is gonna make the difference between it being viable and non-viable.
Got it. That alphabet soup of technology graveyards is we'll see how.
There's a few more for those of you...
Yeah
... who are pre-2000
Yeah. We'll see how the transcript captured all of those. All right. We spent the vast majority of the time talking about the data center market. You know, you guys have also been scaling the campus opportunity, and the enterprise opportunity continues to be important. Just how are you seeing? You know, is the primary opportunity on the campus side still with existing data center customers? How are you taking advantage of kind of the HPE- Juniper dislocation?
Yeah. Actually, the dislocation started post-COVID. It wasn't a vendor-specific one. We chose to enter at a time where suddenly no offices were being built, and there was, you know... We used to talk about the carpeted headquarters and the remote branches, and none of that mattered anymore. It was one homogenous system where everybody was working remotely, and the user had to carry their credentials, whether they were in a hotel or Starbucks or home or in an office in the post-COVID era.
Mm.
I think that was a unique disruption that therefore gave us an opportunity to really build a wired and wireless homogeneous system.
Mm-hmm
... and take advantage of that same leaf-spine architecture. Before I turn it to Ken to give more details of some unbelievable innovations you have done to make layer two function better, which is a large part of campus, what I would say is Initially, we saw a lot more of our existing customers-
Mm-hmm
... adopt our campus. We're starting to see a nice blend of both. I would say we see about 40% new customers and 60% existing right now. The shift to more and more new customers is happening.
Okay. Ken?
Yeah, no, I mean, we're seeing also plenty of campus-first customers, meaning customers whose first purchase from Arista is for the campus. Campus is not a follow-on or just an add-on. It's a first-class initiative, and one of the reasons for that is the innovation that we've brought to campus. By integrating the wired and wireless management, being able to deploy into a, whether it's a branch or a small campus or even a large one, the wired and wireless together. Too many of our competitors view them as these completely separate things, where wireless comes with its own complexity, its own set of controllers you have to deploy. We have a controller-less solution.
We've taken the wired side and wireless side, and made them work better together, both from a management point of view, but also just from a sheer control plane point of view, using the wired routed infrastructure as essentially like a Mobile IP infrastructure for mobile campus clients. We can support hundreds of thousands of campus clients in a completely flat IP space, so anyone can go anywhere across a large campus with seamless fast routing across the whole infrastructure. We've really raised the bar, I think, on campus mobility.
We've also done some pretty incredible things like SWAG, the Stackable Open Architecture.
Yeah.
You know, I think we've again, brought more open standards, and innovation to a somewhat stagnant set of technologies in the campus.
Yeah.
It's been kind of fun to see that. Security. A lot of segmentation and zero trust security inside the campus as well.
Okay. in Q3 on Q3 results, you noted that there were some kind of constraints on the part of your customers that kind of limited some upside. Just what are you seeing as the biggest bottleneck that your customers are facing today?
What did I say in Q3?
that just that there were some bottlenecks to your customers that were installing equipment, that was just kind of making it back towards you.
look, I don't know if I was referring to supply chain or what the context was, but while we have solved many supply chain issues, I think the one in front of us in Q3, and actually we realized it more acutely in Q4, is memory.
Yeah.
I, obviously, we're in a conference where I'm sure that's discussed a lot. Memory goes through these cycles, but I think because of the high, adoption of memory in automotive sector, in AI sectors, specifically servers, we're seeing some real shortages there. It's affecting our customers. We're doing everything in we can to lean in and overcome it. I think it's gonna be a two-year problem that actually separates the wheat from the chaff. We're making significant investments and purchase commitments in chips, in silicon, in memory to make that possible.
Got it. Any questions from the audience as we wrap up? All right. They're all too busy. All right. This has been a great conversation. Jayshree, Ken, thank you so much for being here today.
Thank you.
Thank you.
Thank you for having us.
Thanks, everybody.