Arista Networks Inc (ANET) Earnings Call Transcript & Summary

Well, why don't we go ahead and get started, try and keep us on schedule here. So extremely excited to host a 35-minute discussion with the Arista team. So we've got Chantelle Breithaupt, and we've got Martin Hull, obviously, CFO, Vice President and General Manager of Cloud and AI platforms for Arista. If there's time at the end, I might ask anybody who has a question, please raise your hand, but I'm going to jump right in. Chantelle, thank you for joining us. Martin, always good to see you.

I'm just going to start here because it came up a lot post the recent earnings. The company put up great results as expected. The debate seems to be the 2% guide, right, on the Q4. So maybe we could start by just talking a little bit and Martin jump into what you're seeing from a supply chain component perspective, how that's maybe affected some of the shaping and the timing of product availability. Just walk us through the puts and takes around that.

Good morning to those in the room and those on the webcast. So I think that I would kind of almost decouple those 2 things, Aaron, and I think it's a great question. So thanks for bringing it up. As regard to Q4 FY '25 and FY '26 in general, we don't see any constraint issues on revenue. Anything that we see in the industry, we've either addressed through our purchase commitments, which we raised to ensure we have the supply we need this year and next year and perhaps beyond. And also to take a look at the fact that we are confident in 20% guide for next year, as early as September at our Analyst Day. And I think that's the earliest you've seen Arista come out with such a bold guide. So we're very enthusiastic about that. So I think the constraint in the industry is across the industry, it's not Arista specific, and we just want to make sure, Jayshree, in the earnings remarks, hey, just heads up. There's some things here. It could be memory, it could be fab capacity. The revenue guide, as Arista, we have a pragmatic style. And I would encourage you to think about the growth, both you see in the P&L and what you see in deferred. You saw deferred revenue grow 87% in Q3. And so I encourage everyone to look at the both combined. We'll see where we end up in Q4 and how we finished the year and how we guide in February for '26. But we're very excited. And for those who have known Arista and Jayshree well, Hopefully, you heard her enthusiasm in the Q3 earnings call.

Yes. I would say that -- I think it was the first question on the call. She said I've never been felt as good about the growth that we see in front of us going forward. Talk a little bit about the deferred. How do we think, I think 18 months gets thrown around. That's more of an average. You expanded your product deferred, which maybe you talk on a total basis, but product deferred another $625 million this last quarter. I think prior quarter was $687 million, extremely robust growth in product deferred. So that pragmatic approach to guidance, how do we kind of pack product deferred vis-a-vis revenue generation? How is that factored into the calculus when you think about guides?

Yes, sure. And so I think I appreciate the opportunity to provide further education because Arista's business model is changing in the sense of how deferred plays. I started January 2024. And I think give or take that timing earlier, deferred was a cloud build-out services, your regular kind of maybe 6 to 12 months. As we progress through '24 and '25, now we have some of the largest, most complicated data center build-out for AI centers. And those take a lot to come together. They're based on acceptance criteria. And so we've moved from 6 to 12 months to 18 to 24 months for some of these larger deployments. And so -- in the modeling, I would encourage you to think about that time shift difference between that. And so you'd lean more towards 18 to 24. Given the amount of growth in deferred, it's related to these AI build-outs and the new products. So that's how I'd ask you to consider that from a modeling and timing perspective.

And just to be clear, you've not -- Martin, power availability, shelf space, component dynamics, have you seen any kind of indicators or dynamics that have changed deployment plans for some of these larger projects?

So I think when we talk about commissioning in a new data center from literally breaking ground, then you're talking about a more than 12-month horizon on the construction cycle. That can get delayed for permitting or power or who knows where in the architecture and supply of a physical building. That's not on our time line. The time line where we get involved is the customer is talking to us about what they want to deploy in that data center and when they want to deploy it. When we get to -- I don't know what the phase is, the building is dry, right? It's got a roof on, it's got power. At that point, we're having in-depth conversations about what architecture, what density, what product choices are going to go in there. That can vary by plus or minus a quarter. But we're not talking about plus or minus a year at that phase because the building is up, powered and they're getting ready for deployment cycles. We can have conversations with customers about -- we're thinking about this, but what about this and now that variability comes into product choice, that variability maybe comes into exactly when and where say, plus or minus 1/4 is kind of that ratio on there.

Yes. And then everybody's got different ways to try and kind of backwards in the math and how we think about the networking opportunity. I think One of the things that's very clear is the networking opportunity is only getting more relevant as these clusters get larger. We go from scale out, scale up, scale across and complexity will continue. When we think about -- pick your number, gigawatt of deployment or we think about Lisa Su's comment last week of $1 trillion of AI silicon TAM. How do you think about -- how has it evolved your thoughts around the networking piece of that opportunity?

It's very difficult to pick down from the power or even the spend on the GPU or the accelerator side of it. It's very difficult to kind of go, okay, so take that number, divide by 4 and divide by another 1,000. So we kind of work from the bottom up. right? If the price of the GPU doubles our halves, I just know what I'm going to do. I don't think that changes the value of the network. And so that divider is the problem. So I tend to work from a bottoms-up approach. How many ports, how many blades, how many switches, how many interconnects in a given network architecture, 9,000 nodes, 16,000, 32,000. Okay, we can do that. . But if I try and work backwards to how much that CapEx is on the compute side of it, I don't know that we got the information on the pricing side of all of that. And we certainly don't have information on the cost of the building. So we tend not to try and divide that down. You guys can all do that. I'm happy for you to try and give me the answers. But we tend to work from the bottom up in terms of how many ports, 400-gig, 800-gig going to 1.6, okay, but I can do some math. We have a product choice as well. That portfolio of products from our X-Series to our R-Series can introduce another level of variability. You mentioned scale across. Scale across is kind of data center interconnect, that incremental revenue is smaller than the total spend on the back-end network itself, but it's another factor. And you can't know from how much they're spending on the building, how much they're going to need for data center interconnect. Those 2 things are independent really.

Yes. And the only thing I would add to it, in addition to what Martin's comments were -- was that if you -- when we were talking maybe 2 years ago, we would get asked the question of a data center build out what percentage are you over spend and I think at that time, we were using basically high single digit, low double digit kind of the 9% to 11%. Our best estimates because we're still finishing some of the larger AI centers is maybe that's moved from 5% to 7% versus the 9% to 11%, but still TBD, but just to give a framework versus what we used to use to try to solidify that.

And just to be clear, with the audience, that is obviously the definitions of what is AI is different depending on what company you talk to, right? That is just switching. It's not anything else. There's no transceivers, there's no optics. There's no, right -- because that definitely differs from if it's NVIDIA or some of your other peers that report these AI networking numbers, correct?

Yes. Well, I would say that the difference between the 5 to 7 to 8 would be if optics were NRO likely. It could be, but everything else we would exclude.

Yes. Yes. When we talk about our AI numbers, we're talking about the switches, the routers, physical devices. We're not counting optics in there. And we're definitely not counting NICs because we don't have any. .

Yes, exactly. The company's over these last several quarters consistently talked about the 4 customer deployments, 100,000 GPUs. Can you just remind us of where we're at on of those deployments and where we're going, maybe or how we should think about where we're going over the next 12 to 24 months, pick your horizon?

So we're very excited. Obviously, we have expansion beyond these 4, but we talk about these 4 because they're indicative as we started to get into this AI cycle, what it meant to us as a company. I think that the 4 are going well, 3 of the 4 are coming into production this year, maybe into January, but basically, we call it December 32 kind of thing. So very close. The fourth one, which is a transition from InfiniBand to Ethernet, that one is expected and on track for next year is the earliest part from the sense of revenue recognition from us. That's right. But we have lots of other AI central opportunities. We talked about basically 25 to 40 other customers between Tier 2 enterprise, specialty providers kind of customers in neocloud, sovereign state. So we're very excited about those conversations as well.

And that ultimately underpins right, the $2.75 billion guide for AI, front end, back end for 2026 versus the $1.5 billion for '25. How has your -- in these customers, how has it evolved in terms of your breadth of deployment? Any kind of anecdotes that you can share of like you've won these customers, but it's expanded your upsell, if you will, opportunity in these?

So there's 2 axes on that one. So that is within the, let's say, top 4. Within the top could be 5, it could be 6. Within the top customers, we're having more deployments. They started in the second phase, third phase, and we're keeping enrolling. Now these are customers who -- many of the household names, they're not going to slow down to in AI deployments. And so that continues to grow on multiple locations, both in the U.S. and internationally. And then there's a set of next customers. These are large customers in their own right. They're not necessarily in that top tier. But for each one of those, there will be a first phase and then typically a second phase and occasionally a third and then maybe they'll stop because they've achieved their business, their business goals. So we are seeing -- we've talked about it on the earnings. We're talking about 15 to 20 customers in that vanguard. They're not all neoclouds. They're not all sovereign wealth funds. There are some AI as a service providers in there. So we're seeing that broader base of customers who are putting their first deployments in calendar year '25 that will go through a pilot to production. Calendar year '26 is a second generation or an expansion of that. So it's getting broader and the ones that we're in first with are getting deeper.

So the $1.5 billion to $2.75 billion would you say that, that's majority heavily driven by just your top 4 hyperscalers? Or is that...

No, it's across the customer set. It's across the 4 and the 15 to 20. And there's possibly pull-ins from like you were mentioning, there could be campus, if there's more inference and agenetic happening in their own companies. And I just want to take a moment to friendly remind that's revenue recognized, not orders, right? We talk about revenue. So we are talking about $1.5 billion to $2.7 billion revenue recognized.

Great point. You touched on it and I think scale out is where the predominant majority of your business is in the AI, the scale across is like what might have been called DCI, not too terribly long ago, has now become scale across. Scale-up is still opportunity set, kind of TBD more '27 than it is '26. Can you help us maybe -- how do we think about Arista in the context of the scale-up opportunity?

Well, I'll start in the sense of positioning and then Martin can speak how it works from a product and technical perspective. So at Analyst Day in September, we talked about Arista having $105 billion TAM growing from the $70 billion we had said just a year before. So very excited about the TAM growth. And in that, there is no scale up TAM. So this is a net new TAM to that number, just a position. We're still exploring what that TAM could be waiting for the Ethernet conversation for this up conversation. And maybe, Martin, you could talk about some of the things we have to go through to get to that.

Yes. I'm trying to think where to start. So you mentioned the scale-out network. So the scale-out network is a multi-tier, typically 2-tier infrastructure to allow these thousands to tens of thousands of GPUs to all talk collectively across 1 single infrastructure. For that, we've got our portfolio of X-Series and R-Series products today that address that well, we put them in the 2-tier networks and they're good to go. Those products are purpose-built for AI but those are also the same products that customers can use on their front end networks for building the client connected network. We think about this next generation of scale up. It's a vertically integrated, tightly integrated network architecture. So effectively, it's a new set of products that are built with the customers to physically integrate into their physical infrastructure. So it's going to take 2 or 3 key differentiators there. Next generation silicon, next-generation products, and then those customers coming to us and working with us on the requirements, the definition, the delivery and then also the release of those products. So that's why it's different to the scale-out but we're perfectly positioned for that in that we have assessed the technology. We have the best engineers in the industry, I would claim, right? We've proven what we can do with these customers. And so they are coming to us and asking us to partner with them on developing these next-generation solutions to scale up.

And I think you've brought up a good point with me in the past, like it's important to understand like excluding the leading vendor of GPUs, everything else is attached Ethernet, right, scale up across the entire fabric architecture, right? So Helios from AMD might be a very well positioned rack scale solution for a scale-up opportunity or pick your other piece looking out there?

Yes. So at this point in time, if you're doing a scale-up network, you're probably using an NVL generation technology. We've seen at the OSC conference last month, the introduction of the ESUN technology. That's an evolution of the scale of Ethernet. So ESUN is Ethernet scale of network than scale up Ethernet, move the letters around get another terminology. But ESUN is a multi-vendor with customer participation in building a specification that everyone can work to. So you do get this choice of products, choice of vendors so that when you're moving into the late '26, '27 generation, you have vendor diversity, supply chain diversity and derisking of these technologies. So that's where we're heavily engaged. We were one of the leading partners of that ESUN initiative. So tightly engaged with the technology and the customers to make sure we bring it to market. It's not single vendor technology.

That kind of maybe ties or segues a little bit to the blue box narrative, blue box white box. I mean that's -- white box has always been a persistent topic of discussion competitively maybe help us, one, what have you seen from a competitive dynamic vis-a-vis white boxes? And what is Arista's blue box strategy? And I'm guessing it does tie back to maybe scale up over time. But you've been participating -- just walk us through kind of white box, blue box dynamics for the company?

Yes. We'll tag team this. There's probably a lot of conversation here. So again, we hold the position that nothing has changed from a white box dynamic and from the perspective, there are places white box can absolutely grow. The whole market is growing. And so there's room for everyone to grow. And they grow in some of the customers we're not even in. You take -- some of them are with Amazon and Alphabet, and they've been in that white box even before Arista started. So there's going to be growth there, and we recognize that. In our hyperscaler conversations where white box is usually where you have the ability to have the engineers team to support Sonic or FBOSS on it. So those are at the larger customers. So white box, I think apples and oranges compared to Blue Box and Arista branded. One example I'll give you where Blue Box is great and where we use it today before I pass it over to Martin is a lot of our largest customers, they need to and they should have dual source vendor strategies, right? You never want to place any 1 large company on 1 vendor. But where Blue Box can be very useful as a dual source strategy where it's Arista hardware underneath and some of the Arista hardware is running EOS and some of it's running Sonic or FBOSS, and that allows the dual-source strategy, but it has Arista hardware underneath. And so that gives them the flexibility to say they have dual source, which is great. And that's one example of a use case that's existing and a great application. Blue box generally is in our guide. Sometimes we get a question, what's the margin impact from blue box? For us, it's in our guide, it's in our actuals, and so nothing is really changing there from a material perspective. And then, Martin, maybe you want to talk about some of the other blue box.

Yes. So Arista blue box isn't really new. It's something we're talking about now, whereas it's something we've been delivering with some of these large customers for multiple generation technology. They came to us and asked us to partner more closely on developing a product, a box for them. And they wanted their own technology inside, but they also wanted the benefit of all the Arista design, manufacturing, supply chain diagnostics, the 1-800 number or call for support. So when we take all that packaging together from the fundamentals of an engineer designing something, mechanical engineering, thermal, packaging, everything that goes into making the Arista product today is what they want to tap into. They didn't want to get an off-the-shelf white box they wanted an Arista product, but with their own code running on top of it for the software stack. So it's something we've been doing for a while. There's many examples out there. We partnered with a couple of these large customers. And so now we're talking about it more broadly as a way to help the investor community and customers generally understand what it is that differentiates Arista from an off-the-shelf product. So design, manufacturing, supply chain continuity, so multi-sourcing, the diagnostic software that we run even on the design level before we run it on the manufacturing line for the test, the release, the firmware that we're running on some of the FPGAs that are inside these systems is all Arista. It's that Arista value. That blue is the Arista blue capability. So it's very it's very difficult to say where does the line start and stop. So we've put some documents out there to kind of show where these layers are. You're not just taking off-the-shelf Broadcom silicon, putting on a standard reference design and shipping it. There's a lot of value. And we've seen examples where an Arista product compared to a standard one, we have lower power, we have better thermals, we have higher reliability. These are tangible benefits to the customer taking an Arista product even if they don't run Arista's operating system, EOS. Now of course, we prefer that they do that. And then the other benefit is if they take this blue box and they run their own network operating system, they can, at any point, refresh it to run EOS. So it gives them some investment protection in terms of that technology. They're deploying it in one role. They want to redeploy in a different role, they can start to run EOS on it. And I think that's something that these are all the value propositions of blue box as against a standard white box.

Yes. That was very thorough. That's great. Sticking on the competitive dynamics. AI, you've got the largest GPU vendor wanting to participate, obviously, deeply in the networking stack, a full stack solution, walled garden approach. You've got, obviously, your biggest competitor in the networking side that talked about AI. How would you characterize the competitive landscape in these AI fabrics, be it scale out, scale across?

I'll start. So on the front-end network, the front-end network is de facto Ethernet. Nobody would think about putting an InfiniBand technology into a front-end network. It's a part of the network that over the last 10 years or so, we've incrementally taken market share away from who was the #1 vendor in the data center. So we've done that to the point we've now surpassed them in market share for the front-end network. This back-end network is very much a net new opportunity for all of us. It's a TAM that's expanding rapidly. So within that, you're going to get some market share dynamics. We saw 2 years ago the question about InfiniBand versus Ethernet. I think that question has largely gone away, not to say that InfiniBand will never go to 0. But largely speaking, people have decided that Ethernet is the right technology for the back-end network. We're now having a new debate about scale up. We'll leave that to one side for now. But on that back-end network, you've now got a level playing field for Ethernet technologies. Just described how we were successful at the front end of the network. There's no reason to think we can't be successful at the back-end network. And given the level playing field, you take best-of-breed, you take the software quality, the features, the ability for our engineering team to partner with our customers and make sure we're building the right products at the right time. and then that track record. So you kind of take all that together, and that's -- it's a broad answer rather than saying, well, this feature is better than that feature, and we have a product and they've got a radix and we've got a radix. Those are all the detailed answers, but I like our chances on a level playing field. And our networking technology, our networking products are best-in-class.

Yes. That's perfect. Before I go to model stuff, I'm going to ask this other follow-up to Martin. There's been some debate out there about one of your M&Ms, your largest cloud titans, has gone from a disaggregated scheduled fabric architecture and talked about a non-scheduled fabric architecture. Obviously, they've done stuff with their mini pack solutions and stuff like that. How does Arista play in a nonscheduled? Maybe walk us through because I think you guys have obviously switch portfolio that addresses nonscheduled fabrics. What do you think about that?

As I said before, we have a portfolio. We've got the X-Series and the R-Series. Both of them are optimized for these large-scale AI deployments. At any point in time, any customer can choose from that portfolio. We don't force them down any path. We've seen in previous generations, these large customers, you can give the names out, have deployed a mixture of the X-Series and the R-Series at multiple tiers. For the latest generation, they chose to go with the DSF, which is taking the R-Series products and basically spreading it out, so disaggregated distributed scheduled fabric. The next generation, for whatever their technical reasons are, I mean we talked about it at OCP, they're going to deploy NSF. That's possibly a timing as well as achieving some scale capabilities. The distributed scheduled fabric, I think they published results that show it's performed extremely well. And say, by offering a choice, we can win or we can win. And it's not a -- there's no losing in here.

Yes. That's perfect. Now sticking to maybe some of the model stuff. Analyst Day, I think it was September 11, very thorough. I think one of the things that I took away was that you set a guidance, which you mentioned earlier, was quite strong relative to the way you've set in the past, 20% growth. If I do the numbers, right, the $2.75 billion of AI, the $1.25 billion of campus, it really implies that the rest, the non-AI non-campus business doesn't grow. Why?

Yes. So I think that we are very intentional with some of our goals to make sure that communities like yourselves understand what we have line of sight to and what our North Stars are. So we set a North Star for the AI growth, the $2.75 billion in revenue and the campus $1.25 billion, basically 60% plus growth on both of those targets. We laid those out for the company because we want to be very clear where we're going from a strategic perspective. We also, from a style, never assume 100% of everything is going to work in our guidance. And so it doesn't mean we don't want or anticipate the rest to grow. We'll see as we get into next year with 2 quarters of visibility, other things that can add to that number, but we will never assume 100%, give ourselves some optionality. If we hit both of those targets and things continue to grow, we'll see what -- how next year progresses.

Yes. That's perfect. And the reference to 2 quarters of visibility, that's an enterprise.

Yes, that's everything basically underneath the hyperscalers.

I got you. Okay. The other question that's come up on the model has been this quarter, I might have gotten the math a little bit wrong, but like the gross margin on the product line takes a little bit of a step down. So immediately, when people see that, they're like what's going on, why? I think the short answer is mix, but maybe I'll let you address that.

Well, the majority of the time mix is what drives -- customer mix is what drives our gross margin conversation. Obviously, the hyperscalers have a different volume purchasing power than the enterprise and everything in between is a mix of those. So on a high mix cloud AI titan quarter, you're going to see a little bit lower gross margin and a higher enterprise mix a little bit higher generally. The other thing, as we tried to articulate at Analyst Day, the other thing that moves through there is our E&O activities. The Arista model, we have long lead times in our supply chain. So we lean in almost a year in advance. And we have educated procurements, but it's not 100% forecast driven with 2 quarters of visibility. There will be times when we don't get the mix right, and we do everything we can to mitigate the E&O, but that's the other factor that can come in. And this year, you saw not a lot of E&O. And so our gross margin has been elevated at that 64% to 65% range, some of the quarters. So those are the dynamics, very open and transparent about it. But very excited even to kind of report those growth rates, the margin rates and the operating margin rates this year and next year.

Yes. The campus opportunity, you've brought in -- the company hired Todd Nightingale to really -- it sounds like drive the campus, right, that I think it's $700 million to $800 million this year, growing to $1.25 billion. Can you -- I think you're only 5% market share of the campus market. What's changing there? What gives Arista the opportunity to win? What -- how do we think about maybe beyond $1.25 billion because it is a large market, $18 billion, $20 billion in the campus market.

No, absolutely. We're very excited. I think the one thing that I've seen, and I'm very proud of Arista is once we set an intention, we very much try to execute and overachieve on that. You've seen it with the hyperscalers and the data centers specifically now going into campus. So what has allowed us to kind of come out with this declarative state? We've not finalized but very well finalized the campus portfolio. The VeloCloud acquisition was a great part of that SD-WAN kind of conversation. So very happy with the portfolio. We think it's ready now to get more than 5% of that market share. As well as the fact that we have someone like Todd coming in who can spend more dedicated time, given his background as to how we're going to approach this market through land and expand and new logo acquisition. So I love time between Todd and I and Martin and the team to talk about what are we going to do with that? And so where do we see this kind of validation? We are winning campus-first deals now that are material, especially for campus market. So we're very excited. And those could be without the data center, so we can win a new logo acquisition in campus and then land and expand over to the data center if we're not already in that position. Super excited about that. 5%, we do see that $20 billion kind of market TAM perspective. So very excited, and Todd is very much focused as well as he owns the supply chain of Arista, so he can work on making sure we have the right products and lead times to make sure we win those refreshes as they come due.

Yes. So first part of it is portfolio. So throughout this year, we've incrementally released new technologies, new solutions with identity. Pulling in the VeloCloud has given us more of a complete solution. And then the second factor is time, right? A lot of these large campus opportunities come around once every 5 years, once every 7 years. So it's not for want of trying, but if the customer is not in the buying phase, then there's no opportunity. So for the very largest customers, as they come up for a technology refresh, it gives us an opportunity to engage in an RFI and RFP or get into a lab or a qualification exercise and then hopefully be successful. But if the customer is not in that phase, you sit on the sidelines and wait. So I think we're identifying that within the campus and the enterprise more broadly, the next 2 years, there's a lot of refreshes coming up for renewal. And as Chantelle said, we are very encouraged by the customers we're winning, winning first as a campus opportunity, and that gives us a chance to go and talk about other areas of their infrastructure.

And remind me again, I just forget, how many enterprise customers does -- I think it's like a rounded number, 10,000...

Well, we said 10,000 customers for Arista, 10,000 plus.

Enterprise customers.

Well, we said customers, but you can.

Yes, small list of hyperscalers. But I guess the metric that would be interesting would be is like how many of the -- to your point, if you have an enterprise data center footprint, it clearly gives you an opportunity to go into the campus opportunity...

That's right. Land and expand our new logo. This is what Todd and we're working on this year, next year planning, et cetera, how do we get to that $1.2 billion, which methodologies.

Okay. And VeloCloud is in that $1.25 billion.

Yes, it will be.

Yes. Okay. In the few minutes I've got left, I think I have to ask you about component constraints. I know we touched on it a little bit earlier, but how are you guys mitigating? Is there -- is it DDR4 that you're seeing some constraints on? Is it -- and I guess you've got a lot of purchase commitments. I think your inventory plus purchase commitments were like $7 billion. So not concerned about supply. How are you mitigating the price inflation risk? Do you see that at all in your gross margins?

So we haven't seen a material amount, and I give kudos to Todd, Mike Kappus and his team for being very proactive multisource strategy. So we have been mitigating where there's price inflation either through our own activities or through the actual negotiation with our vendors. So not a lot of materiality there, Aaron, to your question. I think from the perspective of the supply chain constraints, I think that was part of what you're asking. We don't see any restraints constraints for '25 and '26. We're trying to get ahead of anything that could potentially become a topic. But Martin, if there's anything memory fab that you wanted to cover?

No, I think there's a recognition in the last few weeks since we put our earnings out that there is a worldwide tightness on some components, right? We're just one company within that. We can't be immune to it, but I don't know that we're impacted. So visibility and then taking the right corrective actions, and that is putting in place volumes, multi-sourcing and then time frames, like we need this volume over this time. Are you able to support us? So commitments is how we do that.

Yes. And I would say just going back to your starting point, Aaron, the purchase commitment increase that you saw, which was sizable, is more related to demand than it is to buying into component shortage situation, just to be clear.

I think the number is $4.8 billion, I think, was the purchase commitment number, if I'm right?

A little bit higher than...

I'm sorry. $7 billion in total, right. I guess that -- the question I was trying to get to is like how do you think about managing -- if it's demand, I'm going to ask the backwards question, how do you manage -- how do you think about inventory turns, right? And thinking about from that perspective?

Yes. So generally, at least since I've been in the role January 2024, inventory turns have gone between 1.1 and 1.3. So we've been fairly steady. And I'm actually working with Todd to see if we can increase our turns. That's the goal. However, we are working through quite a frothy period. So let's say it stays in that range. The purchase commitments are meant to flush out in the time frame that doesn't really impact that inventory turn calculation over time.

Okay. In the minute we've got left, I'm just going to -- I'm going to put an open-ended question out there. When you're speaking, Chantelle, with investors, or Martin, what do you feel like are the 2 or 3 things that are just not fully appreciated in the Arista story?

Well, I think there's -- obviously, we work with very smart people such as yourselves in the room. I think it's just a -- it's a shift in the sense of looking at deferred and P&L growth, knowing that, that deferred is revenue over a time frame. So I think that's important, especially given that this quarter is the first quarter that product was the majority of deferred. It's the first time I've said that in the prepared remarks. Usually, it's been services, so now product. So I think one is the kind of the revenue outlook, just looking at the guide, P&L and deferred. I think the second one is just recognizing the way that Martin very well articulated what blue box is and the white box and Arista branded and blue box are a little bit apples and oranges. And I think the third one is there are some things we can't control, such as announcements by things in the industry where the whole industry is impacted. We will keep producing great product. We will keep our head down and work close with our customers. And we'll just continue to execute and show you versus all of these kind of interwoven things with investments and commitments between customers and vendors. That's not the Arista style.

Yes. Perfect. With that, I think we're right on time. Thank you so much for joining us.

Thank you. Thanks for your time.