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

From the room. I'll just read some brief disclosures. If you need to see any disclosures, please check out morganstanley.com/researchdisclosures or talk to your Morgan Stanley sales representative. And for those who don't know me, the least important person on the stage, Meta Marshall, I cover networking here at Morgan Stanley. We are delighted to have Jayshree Ullal, CEO of Arista with us; and also Chantelle Breithaupt, the new CFO of Arista. So bringing her to the stage.

So Jayshree, Arista has had a remarkable year as AI and the networking needs of AI have come to the forefront. However, for all the value accretion, Arista has not yet seen a lot of that kind of maybe back-end AI revenue. What do you see as the AI networking operating set and what makes you confident that Arista can capture it?

It's a, first of all, pleasure to be here, Meta. I enjoy these conferences because that you ask these questions that make me think. And no matter how prepared you are, you still have to think, right? So I know there's a lot of AI fever going on, but I just want to first put this in context and say, it is a very gratifying moment when I look back at 2023 to see how we have performed as a company in multiple ways. First, as you talked about the back end, but let's talk about the front end, right, the cloud networking and how we are now addressing that market. And just last week, we got some data that said we are the #1 not just in high performance but the #1 in overall data center switching. Never happened in the history of the industry before that a company can capture that kind of performance. And so thanks to our customers. Now going back to your question on the back end. I think there's -- what's going on in AI is the most killer application is obviously training and large language model training requires a significant focus on job completion time. When you're training billions and billions of parameters, this is super important. And you can throw all the GPUs that you want at it, but the fleet of GPUs, while they're connecting all of these training models and all collective [indiscernible] you have to have a very predictive, low latency network to make that possible. So this isn't easy, and we're very much in the first innings of it. We have, as I've described many times, been in early trials and pilots. But the level of activity is very similar to what we experienced with cloud networking 10 years ago. So I think we have a whole decade run here, and we're only in the first innings. So I'm pretty confident that a best-of-breed network vendor like ourselves with our software and platform expertise will excel in this. And at the same time, we have to work with a broad ecosystem, and that's what's going to take time. So I'm confident, but I also know it's taking time. It's taking 1 to 2 years often for our customers to stage these pilots and go and connect these best-of-breed GPUs, the NICs, the switches and really make a seamless AI networking system work.

Okay. InfiniBand has captured the vast majority of kind of that early AI training networking opportunity, seeing remarkable growth over the last year. However, back of the envelope math would kind of indicate that the InfiniBand leader is charging upwards of 2x markup on that technology. That would seemingly create a great opportunity for Arista. But just how do you make inroads against the technology vendor that's bundling product?

Yes. I think there are 2 forms of selling in AI. There's the vertical stack and the horizontal, right? The vertical stack says, I just get everything from 1 vendor, I get the peace of mind, I don't have to think about it. And however, the horizontal stack where you get the best-of-breed components really favors these large deployments. No intelligent customer, and I don't want intelligence -- I don't want to say the others aren't intelligent, they're obviously looking for convenience, would possibly have the fox guard the henhouse, if you will. You've got these super expensive GPUs, you're spending millions if not billions of dollars. Any kind of downtime on that is a problem. And so why wouldn't you put the absolute best network you can, would reduce downtime, increase availability, improve instrumentation, improve security, build AI at scale with a real professional IP network at scale with our EOS stack. So I think InfiniBand or any non-ethernet technology has its moment in time, have happened to have been through several of them, ATM, token ring, FDDI, now InfiniBand. And they certainly are very good in certain use cases. But as I said in the last earnings call, 4 of the 5 customers we talk to have now chosen Ethernet over InfiniBand. We were very much outside looking in a couple of years ago. I think Ethernet is here to stay and everybody is going to compete. InfiniBand's the battle, Ethernet's the war.

And so some of that is just the ecosystem around Ethernet is broader. But some of that is that you guys have made kind of changes over the past year to kind of close some of those gaps with InfiniBand. So just what are some of the gaps that you've helped closed? And what are kind of the other ecosystem pieces that need to come through?

Yes. Really good questions. There was a formation -- Ethernet's always been around as Bob Metcalfe, the founder of Ethernet would say I would never bet on an Ethernot technology, right? So me too. But the beauty of Ethernet life and the fact that it's familiar, there's billions of nodes installed. There's a lot of tools. There's interoperability, there's standards. But you always have to tweak it to do something better, right? And so in the classic cloud networking world, to go back to analogies, we had to build an active-active topology and make equal-cost multipath workover N-way. N-way could be 64, 128, and that just built an incredible scale of 1 million to going to 1 billion servers nonblocking, never had been done before, and that's how we power the Azure or Meta or many cloud networks today. Something similar needs to happen with AI where you need to look at the problem and say, okay, what is it about these AI clusters. They're data-intensive, they're compute intensive, and you're continuously going through a compute exchange, fetch, reduce cycle over and over again. And so the network has to respond to that in 3 ways. First of all, you've got to have a very nonblocking architecture. But you've got to have a high weightage architecture that can support all of these GPUs coming at the same time. So you can have multiple senders sent to a receiver. And how do you deal with the congestion control, how do you respond to that congestion control. Packet latency is interesting, message latency end-to-end all the way from the GPU to the applications, the network becomes fundamental. And then speaking to the ecosystem, there are things you have to do on the end system side, on the host side with the NIC, where typically you had RDMA or remote direct memory access over Ethernet, a reboot of that is required to do more flexible packet ordering and spring of these packets, dynamic load balancing to make sure you get equal access to all of these GPUs is important. So there's a ton of tweaks to Ethernet that are happening. The beauty of Arista's architecture is, a lot of this we've been already working on. We have congestion control. We have the dynamic load balancing, and we'll be working with a suite of NIC vendors to make the UEC compliance stuff happen.

Okay. So you've outlined $750 million of AI back-end revenue target at your recent analyst...

Up from 0, I might add, right?

That's a tough one -- at your recent Analyst Day for 2025. You've also noted that 4 out of the 5 large AI trials have settled on Ethernet or the latest ones have settled on Ethernet. Just what are the gating items to, when we start to see some of that? Is it Ethernet is an obstacle? Is it -- what part of the ecosystem is kind of that biggest gating item?

As you can expect to have the courage to step up to a number like that must mean I'm in sub trials and pilots. And what has taken in all of these trials and pilots is to first understand the cluster of GPUs they're trying to start with. And most of them don't start with thousands or hundreds. They start with a few hundred. So the first use case we saw a lot of was the 7,800 AI spine, which can connect 576 in a single 7816 architecture or you can do a home DEN to get over 1,000. So you can imagine we're in a lot of trials doing that right now, pushing and moving billions, millions of parameters, right? And this is intense testing that goes on continuously. The next step is they go, okay, wait a minute, I finally got my GPUs. I want to do more. And so how do I scale that to a 2-tier leaf/spine architecture where, again, I can take the spine architecture that's very familiar to most of our customers and then extend it to a 2-tier where I'm adding the AI leaf. And this can go now from 1,000 GPUs to several thousand GPUs. And then you'll hear from us more this year on how we're going to build -- so this is the next step, which gives me confidence in that number. And then if you start having to go to more than 10,000 GPUs to 30,000 to hundreds of thousands, you can build a multi-tier architecture or Arista is working on a Distributed Etherlink Spine that I shared with you guys at the Analyst Day where we can bring a single-stage architecture and sort of bring the best of leaf/spine and a Spline, if you will, and scale up to distributed 30,000 GPUs. So we're working on that. So all of these 3 we hope will contribute in some fashion to that big goal we have.

Okay. So the optics ecosystem has been a gating item to upgrade cycles before. The InfiniBand leader has taken to kind of direct relationships with some of the optics vendors. Are there steps that you can take to kind of speed some of these innovations across the ecosystem?

Yes. Our approach in optics continues to remain really encourage the innovation here and work with all of them. Sometimes the customer wants it with our products integrated, and we have that suite of offerings for 10-gig, 40-gig, 100-gig, now 200 and 400-gig and of course, we'll extend it to 800-gig. One of the most exciting things I think on the optic front is what we've done with linear drive. And there's been a lot of talk about co-packaged optics. And it's a difficult thing to maintain and troubleshoot. But the beauty of linear driver is, especially for smaller distances like within a data center or AI cluster, if you can actually remove the DSP and extend the drive through your own electrical SerDes technology, this can go a long way, and this is going to play an important role in AI as well within -- especially within a data center, but even for kilometers reach between data centers. So I think this is going to be key as an innovation in the optics.

Okay. So maybe one of the surprising things in the last earnings call, the GPU leader was just talking about 40% of their data center revenue coming from inference or what we would consider kind of the front end of the network. They're also talking a lot about their Ethernet ecosystem. We traditionally think of the front end of the cloud network as Arista's domain. A big question I get from investors is, why should I not be concerned about kind of all of this commentary about another Ethernet vendor entering the ecosystem with a closed system?

Well, I think you just said the right word, with a closed system. So once again, if somebody wants a proprietary something or a free bundled something, whatever, you can. But without naming customers, I just took down a very buggy Ethernet software stack from vendors like these to replace with us because it just doesn't work. There's something to be said about being in the industry for 15 years and building our operating system with the breadth and depth. And what I mean by that is the breadth of features all the way from layer 2 to AI to IP scale. And the depth in terms of -- this is the third time we are gutting our architecture to build not just a state subscription model on our software, but then moving it from SysDB publish-subscribe to NetDB to now a data lake architecture that can take different forms of data. I think there is no AI strategy without a data strategy and the data strategy starts with your software stack. So we could throw all the hardware in the walls, but without a good reliable software stack, I think it's difficult to build large AI clusters. But certainly, you can do some experiments.

Okay. I want to -- I'm going to move on from AI networking, but I wanted to see if there were kind of questions about AI networking more specifically before I moved on. Okay. We have a couple questions, is it possible to have a microphone? Maybe just shout it and we'll repeat it.

Shout away.

[indiscernible] took us 5 to 7 years to get through the 400-gig cycle...

Yes, actually, I could hear you better without the mic.

[indiscernible]

Yes. So just for those of you who didn't hear, it's like GPUs are going at crazy speeds. It's moved from 400-gig to 800-gig to 1.6-tera. Is the network going to keep up in a way, which is the network interface cards and the switches? Absolutely. And I think the industry has been slow because they haven't had a use case, right? So I think we're going through 3 transitions. The enterprise is still transitioning to 100-gig. The cloud is still transitioning to 400-gig and the GPUs will absolutely transition to somewhere between 400 and 800-gig. And the reason I say somewhere if some people may employ or deploy multiples of 400-gig to achieve that 800-gig. And some people may wait to go straight to 800-gig. Currently, a lot of the pilots we're seeing are more multiples of 400-gig because even though the GPUs have come the SerDes technology and the NICs are still running at multiples of 400-gig. For sure, from Arista's point of view, we will be fully ready to support 400-gig and 800-gig this year with an eye towards 1.6 terabits. So we're ready. Keep those GPUs and packets coming. But you're absolutely right to point out there's this intermediate thing called the NIC that's slowing all of us down. And this is, again, made a reason why I say 2025 is a good year because some people may settle for 400-gig training trial. But if you want that really good seamless 800-gig, going to 1.6, I think this transition is going to happen faster than any of the prior enterprise and cloud cycles in compressed time. Usually took a decade. This is going to take 1 to 3 years.

Do we have another question?

Let's hope your mic works.

Would you expect your software stack, I mean software path hardware solution would have more advantage in 800G or 1.6T compared to your peers try to use maybe open source SONiC solution?

Yes. I think I heard your question to what extent is our software stack and advantage as we move to higher speeds? And what is the role of SONiC? Is that a good paraphrase? Okay. So first of all, our software stack founded, built and bulletproof, I call this 15 years new, not 15 years old because we keep improving it, right? And one of the beauties of the stack is the data plane is programmable, the control plane is programmable, the management plane is programmable. All 3 have to work together. And so our stack is fully ready to support 400 and 800-gig today and will be 1.6T ready when Silicon shows up. And same thing with the management and control plane, they don't always have to run at the data plane speeds. But they have to keep up and distribute the data, whether it's structured data, unstructured data, flow data, contextual data, visibility, correlation, all of that becomes important. We embrace SONiC. SONiC isn't for everyone and not for every use case. And we contribute to SONiC. SONiC is in many ways, a subset of our software stack, where you can, in some use cases, deploy SONiC for smaller, less taxing, less feature-driven networks. And we -- a number of our cloud networking vendors that -- customers that we support either have SONiC or FBOSS running in 1 use case and EOS running in many use cases. So I wouldn't be using it from my most mission-critical neither would many of our customers, but we certainly coexist with that and co-develop with them.

And then we have one more question, yes.

With 1.6 terabit optical transceivers coming on at the end of this year, do you see any of your customers afford a 5 choosing Ethernet, are they trialing linear direct now?

Not yet, not yet. So optics at Layer 1 tends to go ahead, and that's why you're seeing the transceivers because once they push the technology, then we all can push the electrical SerDes technology. To get to those kind of speeds that have to go from a 50-gig SerDes PAM4 like I am today to 100-gig to potentially a 200-gig SerDes. So I don't see that in our horizon this year, but I certainly see it in the '26 time frame.

All right. Perfect. So maybe moving -- we'll get to you in a minute. Moving to the non-AI side of cloud for a second. You're coming off of 2 phenomenal years with Cloud Titans, 78% growth, I mean 26% growth as upgrades drove meaningful expansion. You know that you still think that Cloud Titans can grow as you wait for some of these AI use cases to come in 2025. Just what kind of visibility do you have into 2024?

Look, it's their bread and butter. So of course, they have to keep investing here. But -- and you all have seen the projections, most of our customers have increased their CapEx. The question is, what are they spending that CapEx on? And I think the first thing to remember is while they will continue to invest in their bread and butter, and we will grow at the rate they grow. They have all pivoted in some form or shape to also AI, so that big CapEx number has some cloud and some AI, right? So we will get our fair share of the cloud. And it isn't uncommon after 2 or 3 outsized years for them to take a breather or -- and I think this is a breather year for us on the cloud. I've said that many times, nobody likes to hear that, but that's the reality. But it's not a breather year for us on AI, which is where they're pivoting. So I think the sum of the 2 will give us good success with our cloud customers. In terms of visibility, it was -- on one hand, the supply chain was no fun. But on the other hand, it was a lot of fun because we were getting 1 year visibility. I think we're back to 6-month-type visibility and that's why we try not to get ahead of our skis, and we try to go and do this 1 quarter at a time.

Okay. GCP has been kind of a potential opportunity people have asked you about for a number of years. Just how does AI change how you view the likelihood of this relationship expanding?

No. I think GCP continues to be a good partner. The company was formed because we were inspired by Google. They came to us one time way back in our origins and said, give us a non-blocking network. Give it to us a 10-gig and give it to us at $100 support and not a single commercial vendor could do that, and that's how Arista was formed. But because they couldn't get it from anyone, they actually built a lot of their, as you know, inside network themselves using some proprietary OpenFlow methods. I've been very impressed with the progress that Google has been making not just with their TPUs, but some of their Gemini and graph technology. We do have also a great deal of opportunity with them on outside of their intra data center network. Time will tell how well we will do on the AI side, but hopeful.

Okay. Cisco, HP, Juniper all making bigger points about their Silicon strategies, and objects is becoming a bigger bottleneck. Just you have been a big proponent of merchant ecosystems from the beginning, but are you seeing enough development there to keep the portfolio at the forefront as it always has been?

Yes. Look, I'll go back to history again. Back in the day when Cisco started their silicon efforts, there just wasn't any Broadcom or Marvell or Intel or any merchant silicon. So you have to build it yourself. Same thing for Juniper. They had to do the same for their service provider routing products. But I think those companies have now consolidated. You can see that HP and Juniper on one side, 2 big guys. Splunk and Cisco on another side, they're doing networking, but they're going in other directions as well. That leaves us as the pure-play networking innovator, which is very exciting, and we've got multiple frontiers to [indiscernible]. Clearly, the more options we have on merchant silicon and silicon diversity is going to be very important to us. Clearly, Broadcom is an extremely good partner and they literally codevelop with us. They work with us on features. They work with us on pricing. And I have nothing but good things to say. So unless those things change, why would I go off and do something that's not in my sphere of expertise. So I tend to complement what they're doing with things I can do, whether it's accelerating, programmable silicon or FPGAs as needed. And so until I see a real gap -- if not broken, there's nothing to fix. If I start to see gaps or particular opportunities, I'll certainly look to build more of our own.

And so we talked a lot upfront about people don't love vendor lock in, and all of the reasons why people have been big customers of Arista as we've gone along and your participation in kind of these open ecosystems. Just -- that makes sense a lot for the big public clouds. I think as we go towards Tier 2s or enterprises or even any of your other customer base, they're not experts...

They don't have the staff of 1,000 engineers.

Right. They don't have a staff of 1,000 engineers. And so how do you view the opportunity there particularly when it comes to AI versus some of these more bundled solutions?

Yes. So I certainly think it's early innings on AI for everyone. But we are starting to see the cloud -- the large clouds or the Tier 2 clouds and large enterprises have the engineers, have the thought leadership, they're going to go work with us on it. The enterprise customers also have some very smart people. They just don't have the staff. So over there, I think we will tend to be more educational. We will work with them. The importance of giving them the instrumentation, the manageability, the high availability, the security will be much, much more important, which -- so in the case of the advanced thought-leading customers, the early adopters, they'll figure a lot of this themselves, and they'll have the staff. In the case of the enterprise customers, they're very intelligent, they often don't have the staff. So we will have to give them more Arista validated designs for AI and take them through sort of the small, medium, large clusters and how you build them. But don't underestimate. Their networks may be smaller for AI, but I think they will be up and coming.

Okay. And then moving on to enterprise across this earnings season. We've seen -- many networking vendors have seen a pullback as just order patterns normalize, macro creates a headwind. I know you guys have always said you're not a great enterprise macro bellwether. But just how are you seeing kind of the enterprise opportunity and how are you taking advantage of kind of this environment?

I'm going to turn this question since Chantelle has been very quiet, but I know you had a chance to meet in the last 2 months a lot of our sales leaders and enterprise customers. Do you want to say a few words?

Yes, I would love to, thank you. Thank you, Jayshree. I think that from an enterprise perspective, we're very excited. And it's one of the things I was pleasantly surprised coming into the company, the opportunity that we have. And so as I got to learn, we have a fantastic portfolio that's ready to meet our enterprise customers where they're at. And I think we're making a lot of great progress on kind of a new logo land and expand kind of playbook, which has proven out very well. If you listen to Jayshree's prepared remarks on last earnings call, we estimate about 20% penetration of the global top 2,000 accounts. So I think there's lots of room to grow from a share gain perspective. And it's working, whether we enter through the data center and then move to the campus or enter through the campus and move to the data center, financial institutions, health care, so you know there's some specific industries. We've made really great progress. And so we're very pleased and excited. And things that happen in the market, they take time, the road maps get confused. And so we've heard from customers, give us your road map because we're not clear where others are at. And so we're happy to meet them and we see it as an opportunity.

I think it's perfect timing because clearly, Arista captured a lot of the early adopters, as Chantelle was saying. But now we've got the fast followers and the mainstream enterprise interested in us too, because, as I said earlier, we're pretty much the only pure-play networking vendor available. So it's a good time for us, not withstanding macro and everything else, it's such an incredibly large TAM of $60 billion or at least $25 billion, $30 billion in the enterprise, it's ours to execute.

Got it. I mean you guys have set out campus. We talk a lot about AI, but you've set out rather aggressive campus targets as well. Part of -- you just mentioned disruption in the space. You've mentioned HPE, Juniper. Just how does that change kind of your channel strategy and that there's a lot of channel partners that are now looking for other partners to sell?

It does. But I want to sort of separate our channel strategy for the large enterprises from sort of the mid-market. So I think from a strategy on our enterprises, we continue to invest. We have 9,000 customers who bought our data center products and not all of them have yet tried our campus products, we've got plenty of opportunity there. So I would say our go-to-market strategy doesn't need to dramatically change over there. As we go into the mid-market, we will have to see more changes. And this is where channels come in because while the direct large enterprises know more about us, there are many small channels that don't. So absolutely, we will invest in both. But with an eye towards the larger enterprise because that's where we're well known.

Okay. Your COO, Anshul, recently announced he was going to take a leave of absence last week. Can you just speak to kind of workarounds in the near term and I have a lot of questions from people as -- any details we can...

Yes. Well, first of all, Anshul's just a tremendous asset to the company. He's been with us 16 years from the young age of 30s to middle age of 40s. So he's been with us 16 years. And he's been working nonstop. I think he deserves the break deservedly so, and I'm looking forward to seeing him. I missed him already in the first week of -- all his work has come to me. So I got to find some delegation. So I took a leave of absence when I was at Cisco for several months and came back rejuvenated. So I hope he will too. But I think it's going to be good training for many of us in Arista, where we still function like a start-up, and we have single points of approval, whether it's Anshul or Chantelle or myself, and we'll be distributing that responsibility and at the same time, looking forward to his return. I think developing leadership 2.0 is important for Arista in terms of scale. And Chantelle is a very good example of that. I'll be looking to hire somebody who works for Anshul as a cloud sales leader. We'll be looking to bolster our AI strategy. We have had the good fortune of 15 years of continuous leadership for most part, except the CFOs. And while I would love for that to continue another 15 years, I have to be realistic about whether it will or not, people still getting older and wiser.

And I would -- the only thing I would add as a peer to Anshul, who I have had the pleasure to meet and he's a pleasure to work with, is we're all here to support him as peer, so we will make sure he feels comfortable in his leave of absence that we will help him and have his back.

Okay. I wanted to open it up to questions again. Did you -- yes, we still had a question back here.

I just had a quick question on the vertically integrated player in the space has talked about scaling their proprietary NVLink solution up to kind of 500 processors and the interconnect between each processor as many multiples of what's available on either InfiniBand or Ethernet. Could you just talk about how you think about a solution like that? Is that a mainstream type solution? And how does that impact Arista's opportunity?

I think I understood your question. Can you just repeat the beginning? Did you speak about a...

Yes. I was talking about -- essentially NVIDIA has planned to scale their InfiniBand up to a 2-layer architecture and up to 500 processors.

Yes, yes, yes. Okay. In order to answer your question, I'm going to step back a little and describe the anatomy of an AI network, if you will, and maybe that will put it in context, right? So often, you just have a 2-socket or 4-socket CPU to do all the clearing and indexing and all of that, that remains. Then you have a fleet of GPUs, and this is where our most successful GPU vendor comes in to do all the data crunching for millions, billions of parameters, training, inferencing, et cetera. And then usually, you have to talk between that server out into the network through a transmit layer typically a NIC, right? And then you either scale up or scale out, right? A lot of the scale-up either happens with NVLink or InfiniBand, 2 proprietary technologies for a limited number of machines and processors. And usually, it's sort of scale-up of tensor traffic. We're really not involved in those kind of rail-based designs within a server, right? The real part that Arista comes in is once you have to collect and integrate a whole slew of servers, so from server-to-server communication across an AI network. So what companies do within a server and a rail-based design, whether it's PCIe or CXL or NVLink or they connect an HPC networks with InfiniBand. That's going to keep happening. That's a compute-oriented approach, where we really come in as a sort of an AI networking approach across the servers.

Do you have a question?

Jayshree, earlier you mentioned startup and culture and a single point of approval. Can you just expand on that? And second point, when I think Arista, I think Andy, so can you spend a little bit of time talking about Andy, his involvement with the company, the culture and the business that he's built? And how -- if he was here, what would he say for 2 or 3 minutes?

Yes. Well, it's hard to be Andy. So I wouldn't try to do that. But I will tell you, the culture at Arista is one of the strongest attractions for me to be here at Arista always. We're very much a founder-led company. We're built by engineers, both Andy and Hugh are very active founders, very committed to the company. He's -- because he's off the Board right now, and I don't think he really enjoyed being Chairman or being a -- having fiduciary responsibly, that's not Andy. He is getting deeper and diving more into AI architectures, the optics, the silicon. So if he were here, he would tell me, my God, I've never seen this kind of progress in silicon and AI advancements and he's very active on work on that. Ken Duda, other founder, is very committed, and you may have seen also the introduction of Hugh Holbrook, our Chief Development Officer, who is employee #4 or 5 in the company. These are guys who just live and breathe Arista. And we're very much an engineering-led company. And if he were here, he would tell you that for all the hardware that Andy builds, Ken is the creator of the software, and Hugh brings these platforms together from an AI perspective. What was the other question?

I was going to add just to -- if I can, on the culture, because I'm 3 weeks in, in the sense of the role or 2 months if we count the beginning. So the thing I really appreciate about this start-up culture for a company of this size is that, it hasn't been overcomplicated. I've worked at probably 2 of the biggest, most matrix companies, and I've learned from them. But I can tell you there's an elegance to not overcomplicating things because it drives accountability very clearly. And so I hope you guys appreciate that's a really nice part of a $6 billion startup culture. But to me, that's one of the things I've been pleasantly surprised with from that perspective.

And it's also about keeping it flat, right? We don't have 5 people to get approvals for...

Right. That's right.

And a lot of it just went to Anshul or me or Chantelle or [indiscernible].

Another question here?

Wesley Wise here. You've put out some good kind of indications on where cluster sizes are going into, say, 2025 -- sorry, 2024, I think. And if some of these trials are successful, assuming they will be for AI, Ethernet, back end, you might have mentioned, say, 30,000 clusters and even on perhaps past earnings call up to 100,000, and there was like a reference architecture showing up to 165,000 at your Investor Day. What should we think about as kind of how big could these get into the second half of '25 is 100,000, 200,000 even higher...

I think on the extreme training side, 100,000, 200,000 would be -- would start to become a normal number. But on the distributed training or inference side, I think 10,000 to 30,000 would be plenty. So you're going to see 6 of one and half a dozen of the other because not everybody is going to build the mainframe of AI and a lot of them may build the client server equivalent of AI. So you will see both.

Is that one [indiscernible]...

2025, 2026, yes.

Do we have any last questions?

Jayshree, you guys have the best $6 billion networking company in the world by far...

$5.9 billion, I think.

Yes, for now. The one concern that I do have is you also have the best product mix, customer mix, deal size, transaction per customer. Can you maintain -- when you look forward and you look at customer mix, product mix and deal size, as you move to the enterprise, that's going to change as the market switches from back -- front end to back end and all these transitions, you may lose some more of the high-end chassis that you're crushing the market. You're not going to go employ a chassis. I guess my question -- my concern for the company is that it's doing great, is that it is so great that the incremental growth may not be as great. So Anshul, you saw, always talk about product mix, customer mix as to what drives our business. My question is, when you look at product mix, customer mix and your margins and your growth rates, do you think that this can still be a 10% grower at 40% plus margins?

Yes, I do. And the reason I do unless we go do some wild acquisitions, then the 40% is off the table because we might spend some money there. But if we went the way we mostly do right now, the organic way. First, I want to say, we still have 3 major markets we're pursuing, right? And this is heavenly for somebody who didn't have 3 major markets to pursue at all times, which is we got the cloud. Now we got the AI, and we got the enterprise and -- including service provider and routing markets to go along with it. At the same time, our architecture is uniform across all these 3, and you talked about the great products, which is really a packaging mix for each of these of different kinds, right? So I'm so lucky to have 3 transitions going on at the same time, 100-gig, 400-gig and 800-gig whoever has that. And so if you believe my $60 billion TAM, and I'm here, I am sitting at $5.85 billion or $5.9 billion, shouldn't I be growing to achieve more of that TAM, especially at a time when I'm pretty much the only pure networking company with a best-of-breed platform. So on theory, I should, right? On execution, I need to keep doing better. Not too many companies have rapidly gone from $5 billion to $10 billion and proven that. I happen to work for one that did. And undoubtedly, we'll have to invest in different types of investment models. I think Chantelle is nodding her head vigorously there and hanging on or loving our productivity model for cloud and AI won't be the same as enterprise. So naturally, we will put in a lot of investments in the enterprise and go-to-market. And it got a little slowed down. I would have liked to do it faster. But during the supply chain crisis, it made no sense to invest more when we couldn't ship more, but we'll get back on our front foot for that.

All right. We here at time. But Jayshree, Chantelle, thank you so much for being here.

Thank you, Meta. It's a pleasure. Thank you, guys.

Thank you, Meta. Thank you.