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

Good afternoon, and thank you for joining us for Arista's 2021 Analyst Day. We are excited to spend the next few hours outlining Arista's vision around long-term growth and innovation. My name is Liz Stine, and I recently joined the IR team at Arista. And while I may be new to the IR space, I've been part of Arista for the last 10 years predominantly on the systems engineering side working with our large enterprise and cloud customers. Much like you, I was initially drawn to Arista because of its engineering and customer-centric culture and leadership: Ken's vision for EOS, Andy's understanding around the future of networking, Jayshree's strategy for pioneering the cloud and Anshul's incredible ability to execute. That's a leadership team that captivated me 10 years ago, and it's the same team that will be speaking to you today. Later on in the presentation, Ita Brennan, our CFO, will summarize our latest view of the company's business model and our financial outlook. Let me walk through the agenda for today's event. Jayshree kicks us off introducing Arista's data-driven networking vision and strategy. Ken will discuss the engineering advances Arista is pursuing to bring about the data-driven network. Anshul will deep dive into product innovations around cloud, carrier and enterprise customers. And Andy will round out the technology discussion looking ahead at the impact of AI and ML on the future network. Lastly, Ita will provide an update on Arista's financial outlook and business goals. We will enhance the presentations with a demo put together by distinguished engineer, Fred Hsu, and some customer testimonials. Overall, we should wrap up the presentations in around 2 hours and end our event with a live Q&A panel featuring the entire executive team. During the course of this event, Arista Networks management will make forward-looking statements, including those relating to our financial outlook for the fourth quarter of the 2021 fiscal year; our longer-term business model and financial outlook for 2022 and beyond; our total addressable market and strategy for addressing these market opportunities; our drivers for growth and diversification; the potential impact of COVID-19 on our business; EOS' architectural advantages and future evolution, product innovation, the impact of supply shortages and manufacturing constraints on our business, including lead time and inventory purchases; our approach to capital allocation and the benefits of acquisitions, which are subject to risks and uncertainties that we discuss in detail in our documents filed with the SEC, specifically in our most recent Form 10-Q and Form 10-K and which could cause actual results to differ materially from those anticipated by these statements. These forward-looking statements apply as of today, and you should not rely on them as representing our views in the future. We undertake no obligation to update these statements after this event. With that, let's get things kicked off with Jayshree Ullal, our President and Chief Executive Officer.

Good afternoon. And for those of you in the East Coast, good evening, I hope we won't steal too much of your night time and dinner time. It's a pleasure and honor to be back for the 2021 Analyst Day from Arista. We took a break last year and this whole COVID pandemic when we last met you at the New York Stock Exchange, celebrating our fifth anniversary. Who would have thought there would have been 5 million lives lost and so much change to our work, home and lives itself? And so for those of you who have suffered those kind of losses, my sincere sympathies on behalf of Arista, and I hope we can make the world a better place with more vaccines and more boosters. Today at Analyst Day we'll have a nice roster, and I'll kick it off with our vision and our next frontier for a data-driven cloud network. For those of you who may be less familiar with Arista, I don't know how many there are, we just completed our earnings call and I think it's safe to say, our consensus for this year will be $2.9 billion, growing 25% year-over-year. We have become and established ourselves as one of the fastest-growing networking companies and we have increased our customer count every year, now exceeding 7,500. We now think we have at least 25% of the Fortune 5000 and Global 2000 customers. And we could penetrate more than and obviously have more to go in our execution here. Our overall market share has moved us now from single digits to 19% share overall with a 32% to 35% port share in 100 gigabit, being the #1 market leader. Never have -- has any vendor or any competitor displaced an incumbent. We're very proud of our Net Promoter Score at 92% because this is an indication of the customer satisfaction to us. Not only are our products innovators, but obviously, they appreciate the support and the quality. We believe in the power of one, one leadership team. You'll see a lot of familiar faces, the same ones you saw at the New York Stock Exchange. The power of one EOS, one cloud vision and one mission, our client to cloud-driven data network. Our TAM is more than one, though, and that is a $35 billion TAM, we believe, in 2025. And while most of this will be through investments we make organically, we have done our share of inorganic investments, including 4 acquisitions in the last few years: Metamako for low latency, Mojo for our cognitive campus and WiFi, Big Switch Networks for advanced monitoring fabric and our most recent one, Awake, that brings us AI and ML technology with AVA as well as security and network detection and threat hunting. Industry recognition is worldwide through a variety of analysts and one we're particularly proud of is the Forrester Wave, which really combines a business-wide view of SDN, all the way from the data center to the campus itself. Now when you look at Arista's principles, cloud principles, they don't any more just belong to the titans and are being permeated across our entire customer base. And the fundamental aspect of it is not only are you providing the best CapEx solution which is reducing the OpEx and time from years to operate your network to months to really real time. So from this sort of access aggregation core oversubscribed architecture, which is extremely manual, we believe in a proactive architecture with repeatable universal cloud network leaf design, leaf/spine design, and a much more automated and a highly analytics-driven network with predictive operations with CloudVision. And then finally, prescriptive experience with an API and AI-driven infrastructure for a single source of truth, highly real-time streaming observability based and more and more AI-driven, state-driven and AVA-driven for root cause analysis and network threat hunting as well. These 5 As have guided our principles and have guided our operation of products, product development and customer deployment. We've also changed the network design architecture from silos of multiple boxes where you have one OS for campus, one for routing or WAN, one for data center and branch and so on and so forth, to really building a flat fact topology with leaf and spine architecture, where the universal spine can be guided by data-driven places in the cloud rather than individual pins. So whether it's your data center leaf/spine or your campus or your routing adjacencies or your software services built into the fabric is really the heart of our software-driven architecture, all with the same universal spine with multiple leaf depending on the use case you are deploying. It is no secret that our secret sauce is, in fact, the Arista EOS. The EOS architecture is 15 years old. But actually, I would say, it's 15 years new. It is still one of the best software architectures in networking with a single image driven across a statewide architecture, with deep programming ability across all of the networks so that not only can you manage one switch, you can do this network-wide using CloudVision. This level of state, single image and programmability has differentiated us, and we continue to build a rich suite of features. Combined with the many merchant silicon options we have, you now have a proactive data plane where you can build a low latency to 400-gig, leaf/spine, macro segmentation, different types of channel encryption and set high buffer platforms, low buffer platforms. You can have a control plane that's more predictive, enabling CloudVision to manage and segment your network, provide levels of visibility, increase the assurance of your network through compliance, through telemetry, through automation. And finally, a prescriptive management plane that allows you to have more and more, not just assurance, but experience across all of your environment. This federated architecture of interweaving our EOS stack and our data plane management and control plane has really expressed our phases of innovation and will continue to do so. Today, on Analyst Day, we are now proud to announce that our best EOS gets even better. Our history of rich innovation with our extensible, open, scalable software, our operating stack, it has now gone through 3 generations of evolution. We started in 2008 with an architecture that was state-based, a published subscribed model, well, have never been done in the industry before, where you could have the level of automation and streaming telemetry and availability that you couldn't before. But in 2016, we built another base of foundation to this architecture with NetDB, a network-wide database, where we can now do this not only switch by switch but across a network of hundreds, if not thousands, of switches. In 2021 or 2022, we are adding the next phase of this, a net data lake. How the net data lake differs from a NetDB is it builds upon the state and the streaming telemetry and adds an array of technologies to address different types of sources of data and multiple types and modes of this data. And so EOS has gone from being an operating system to a stack for the next decade. It builds upon this core foundation and adds a network data lake, where you can ingest different types of sources of data, different modes of data and then layers upon this, a very prescriptive AI-driven data enrichment. You've all heard the AI washing these days. But Arista, just like we didn't fall for the SDN washing, really applies AI where it makes most sense. And you'll hear a lot more about this from Ken and Anshul and Andy later on. What I'm most proud of is the operating system that was best already got better with this migration and foundation of NetDL, rich with 16 million lines of code and features to boot. So when you look at Arista's data-driven architecture, it's important to remember there's a left brain and a right brain. On the left brain is NetDL that can ingest lots of different types of data. It could be flow data, packet data, structured data, unstructured data. And you have an array of open source technologies sitting on top of this time series database and state-driven architecture. So really, NetDL is state-driven, store-driven and query-driven. And then you can enrich the NetDL with the -- that we have built with AVA, our autonomous virtual assistant, to give that prescriptive and contextualized enrichment. Because if you don't have things to learn, you can't really apply it to the data, and learning that data comes from the AVA. Most importantly, I feel Arista is really positioned to this architecture to bring all of these data sources into one common nucleus and network architecture. And the network is a perfect place to address this architecture, upon which you can contextualize and add different kinds of associations to devices, to VMs, to containers and to applications themselves. So it's going to be endorsed by a large number of ecosystem partners who have been working with us in building this. By building NetDL with a single attack and API level surface, we can now ingest not only sources of data from our switches, but also from different partners of ours. Zscaler for cloud security, Palo Alto and VMware for threat vectors, Zoom for quality of experience, Microsoft for sentinel identity. These are all ingestion interactions, upon which we can take action with our NetDL nucleus. And then you can also have reporting capabilities through Equinix and [ Ecolo ] or Red Hat or Ansible through DevOps capabilities, Splunk for data logging and Slack. These are all the power of data-driven networks bringing the best of each of these, and it's a real groundbreaking architecture that gives us a great leap forward to the next generation of networking that goes beyond connectivity to really bringing data-driven networking. You've all heard the ISO model of layers 1, 2, 3, 4. And with this architecture, we're really cutting and transcending all of those layers to bring real-time data-driven networking. Now this networking will also permeate our products that you'll hear more from Anshul on. This combination of NetDL and AVA will be on every single one of our EOS products and platforms, enriched with this capability. It will also be on our segmentation and observability platforms. And it will also connect into our network detection and response, bringing an ensemble of all of these array of technologies to drive software-driven capabilities across platforms, adjacencies and network software and services as well. We've always been looked at in Arista as a data center company. And indeed, our heritage, our market share and our strengths very much began there. But today, as we march towards data-driven cloud networks, it's important to understand that we're actually building different types of data centers. It could be a hybrid cloud, it could be a public cloud through a cloud titan, it could be a crypto cluster, it could be an AI cloud, it could be an observability cloud, a security cloud, a campus cloud. What's important now is data centers are not just for the big cloud titans, but it's really about how you center your data. And your data can be centered in just about any location, a branch, a campus, a data center, a routed brand, you name it. And the network data lake is for all of these centerings of data, bringing your network state, your applications, your AI and our ecosystem of partners and development all into one place. So it's less about the data centers and much more about how you center the data across all of these locations. Now as you know, Arista has always been very committed with a high R&D spend through a rich number of products. And as most people today think of Arista, not just as a networking company, but really as an AI-driven company that's delivering network capabilities on the network. So I'd like to take you through the phases we are going through right now, as we speak, and why we have become such a complete client-to-cloud product portfolio, and we're committed to the evolution of this product portfolio for the next few phases. Phase 1, no surprise, was a data center. And with that, we added CloudVision in Phase 2 to manage the data centers. Phase 3 was adding routing and use cases for both DCI and peering. Phase 4 was adding the campus and adding the appropriate CloudVision and AVA insights for the campus. Phase 5 is bringing that multi-cloud WAN and EOS capabilities as well as adding network software and services with DMF and NDR. Every one of these are going through a transformation, and Arista's commitment to product and evolution with every one of these phases will just get better and better. And so I'm reminded of, as we bring these products, why are we all here. I suppose you're not listening to me just to hear my product direction, but you want to hear the numbers. And at the earnings call earlier today, Ita and I promised to give you a little more of a long-term look. 2025, Jayshree usually doesn't go beyond a quarter. Why am I speaking about 2025? But let me start in the beginning right here. Our first goal will be to double the campus one more time. The campus is an important and strategic part of our enterprise milestones. We now have a seat at the table because we're taking that same universal spine design, particularly post pandemic and developing workspaces and distributed architectures for wired and wireless to come in, to deal with a sprawl of devices and IoT threat hunting that hasn't been done in the traditional campus, which was much more of a wired approach or a WiFi approach, and those were separate. We're bringing them all together, and we believe we can double again in 2022, which means take our 2021 goal of $200 million to $400 million. While we go through this doubling, we also believe we will be looking at different types of customers. A large percentage will be our existing data center customer that will extend to the campus and extend their reach accordingly. But we also see many enterprise customers now choosing the campus first because they may not have a large data center. So we expect this to be a good split of new customers as well as existing customers will extend their campus. Our second goal is a 2022 goal. And you know when we all worked together in the New York Stock Exchange, we explained to you at that time, this was pre-pandemic, how we would achieve mid- to high teens growth. Well, it became difficult to do that. As you know, we had a difficult 2019 and '20. But we're not there. We achieved that in 2021, we expect to have a 20%, 25% growth. But with the demand we have from both the cloud titans and service providers and specialty cloud providers and of course, the enterprise, with the product suite we have with core, adjacent and software services, with the backlog we have that we have yet to fulfill until we get into -- overcoming some of our supply chain issues, we believe we can now hit a $3.7 billion target and a 30% year-over-year growth well beyond the expected targets that you all have been suggesting. We're very proud of this. Obviously, this will take a lot of execution. Especially in the second half of the year, we hope our supply chain will improve. But we think this is -- the demand is there to achieve this, and this is going to be an awful lot of execution to match that demand. And finally, a 3- to 5-year goal. When you look back at our 2014 to 2019 CAGR, we were approximately 30% to 32%, going off of small numbers. There was some fantastic growth in 2016, '17, '18, as you remember. But the cloud titans can be volatile and they can grow from double-digit growth to flat to down sometimes. So being conservative about the cloud titans being volatile and having some great years this year and next year, we're forecasting that our 2020 to 2025 CAGR will be approximately mid-teens or 15%. But we will aspire to be a $5 billion company by 2025. I recognize these are audacious goals, and I can't think of a better and most amazing team to work with me to achieve this. And every single one of them is a huge contributor, along with, of course, our customers, our shareholders and all of our entire extended team. I'm reminded then of how I used to drive a car in 2014. It was my Toyota Camry that a lot of you make fun of. It was a great conversation piece, so much so that my husband finally traded it in for a Tesla. He said, "You're going to take a company public and you're going to drive a Toyota Camry?" But hey, it did a lot of good. It raced down a lot of tracks, including the cloud titans. If I got beyond 65 miles an hour, that car would start reverberating a little bit. And so clearly, I needed a better car. It's a Tesla, which is also, by the way, data driven. As you know, we get lots and lots of data in our cars these days. And now we're racing down the tracks in a much more diversified, balanced momentum with specialty cloud service providers, financials, enterprises and cloud titans. And as I look at 2022 and beyond, I think we can leave our competition in the dust. And here to describe how we're going to leave our competition in the dust is none other than my Founder, my good friend, Chief Technology Officer and Service Provider and Senior Vice President of Software Engineering, Ken Duda. Ken, take it away.

Thank you, Jayshree, for that kind introduction, and thanks to all of you for coming and joining us today. I'm really excited to talk to you about the next step in the evolution of the EOS stack, our third generation of EOS state management, which is NetDL. Let me frame this a little bit of context what's going on in the industry, what we see happening. Number 1 has got to be AI and ML. You hear about this everywhere. Everybody is doing it. A lot of people are making claims about it that we're not totally confident they're going to deliver on, okay? But clearly, there have been major -- there's been major progress in this area, and this is something that everybody is looking at. Second is what's happened with SDN as it sort of evolved into intent-based systems. And the best of these is where we've captured this concept of infrastructure as code or network as code as a foundation for automation. And the third major direction we're seeing in the industry is the explosion in the amount of data being generated by the infrastructure itself, by networks themselves, flow data, packet data, quality of service data, the control plane data, application performance data. It's an enormous amount of data out there, all in disparate systems and all very difficult for enterprises to make good use of, thus, the industry context. If you look at the journey that the industry has gone down to get there, if you think back to the '90s and 2000s, we saw a lot of improvements in the control plane, routing protocols, topology discovery, of scaling up and people building larger networks with more routes, more devices, more interconnections. And at that time, we were building our network operating system foundation, SysDB. So on the next step of industry evolution, you may remember a lot of talk about SDN and software-defined networks and the pundits telling us how the control plane as we knew it was going to go away and get replaced by some kind of central controller that was going to set up all the pads in some central fashion. That didn't exactly happen. And that was something that we actually anticipated. And at that time, we built NetDB as a central component all right, but not for the control plane, for the management plane, which we're now able to build on and the next step of our journey. Now we see the industry paying a lot of attention to AI and ML. But how are we going to use the data in the network to make operations more predictable, more computer-driven, more AI-driven to give people better visibility and insight and understanding into what's going on in their infrastructure. And this is where NetDL, the network data lake, Arista is building now is so key. NetDL is a new foundation. It's a transformation of who Arista is as a company. Yes, we make extremely high quality, very reliable ethernet switches and routers. That's not changing. That's not all we're doing anymore. It's no longer adequate to think of Arista as just a data center switching company because we are expanding the scope of what we're accomplishing by gathering an enormous amount of data about what's going on in networks and in infrastructure more generally, enriching, contextualizing that data and using it to improve the customer experience, the visibility that people get into their networks, an understanding of what's going on in the infrastructure. And of course, this state is the foundation for AI and ML. Because without the data, there ain't nothing to learn from, okay? So the -- we see this as part of Arista's growth in the following 3 respects. Development of network data lake is a key enabler of AI and ML. It's now becoming real because we're gathering data from multiple sources, data of multiple types stored in multiple modes where different statistical algorithms, summarizations, query structures apply and then using that data to deliver predictive insight as a basis for analytics and AI. We've also built the AVA, our autonomous virtual assist. And this goes to the heart of how AI applies properly to networks, to infrastructure. Because, again, there's a lot of overpromising going on, but I think we're on to a better path for customers. And then finally, by getting all of the data into one place, a scale-out system with a uniform -- a single uniform API surface on top of that, we can make this data available not just to Arista applications, but to third-party applications as well. So this is a platform on which network and infrastructure intelligence can be built. One of the really neat things about NetDL is that while I describe it as a revolution for the company, it's actually more of an evolution because it's really another step along the path we've been going down for a long time. This is not the throw everything away and start over kind of revolution. This is the take what you've got and build it even better for the next generation. Something I'm very proud of that Arista has done repeatedly throughout our life is rather than entering each new market, we're approaching each new product transition with a whole new network operating system. We have continually gone back and reinforced and enhanced, rebuilt the parts that need rebuilding while shoring up the rest of our foundation to build that same foundation stronger each year. And that's what we're doing with NetDL. The journey started back in 2008 when we shipped our first devices, our first EOS switches with SysDB, a system database that took all of the state of the switch and put it in one place. And it's such a simple idea. No one else has done this. And it gives you tremendous benefits in terms of extensibility, programmability, reliability, recovering from software faults and it went directly to us having the most reliable switches. But then we've extended that. In the early 2010s to 2014, 2015 to NetDB. By taking the state in each switch and streaming it over gRPC, gNMI into a time series database, NetDB, from which we can then build some really key network-wide applications such as network-wide topology, telemetry, provisioning and also compliance, making sure that your every switch is running the version you expect it to be running, has the configuration you expect and isn't subject to security vulnerabilities or bugs that can bring down your critical infrastructure. The next step of this journey is NetDL, extending that network-wide database into a data lake. The key difference here is that no longer is it just data from switches into just a time series database, but it's data from many sources. Yes, still from the individual switches and routers, also third-party devices, x86 sensors placed around the network measuring flow data, measuring traffic patterns, measuring application behavior, but also lots of different third-party systems, including third-party switches and routers, but also third-party application management systems, application performance systems, other tools for understanding what's going on with security vulnerabilities. Lots of third-party systems also stream data into our NetDL, where it's then aggregated into the correct data technology for the job at hand and then presented to applications through a single API surface. And one more level of detail, from every point of the network and this is where it helps us so much to have the same software running everywhere, for every point of the network, whether it's campus and branch, whether it's data center, whether it's in the cloud, whether it's Internet core, your wide area network, your data center interconnect, we're gathering data from all of those locations and ingesting that into a Kafka pipeline. It goes from there into the correct data storage technology. And so far, we have integrated Apache HBase for time series, hierarchical data; Elasticsearch for more text-based data such as syslogs and network events; and ClickHouse for columnar data, data which is more quantitative in nature, where you need to do aggregation of statistical summaries, those sorts of operations. Each has its own unique query and indexing technology, we then unify all of those behind a single API surface, a single surface which is multi-tenant, which scale horizontally, which can be run in the cloud or on-prem, depending on the customer's use case. And to go one more level of detail into what we do with all that data once we've got -- or first, where it comes from, it can come from identity and access management systems. These are your user authentication systems, user databases. You can then use that user information to enrich the network data. So we know for a given flow, for a given packet, for a given event which users are affected, what applications were they running. We integrate with vulnerability management system. So we get information from third parties that track in-progress attacks of various sorts, whether they're on the network or other pieces of infrastructure, bring that into the same data lake, network source of truth systems. So with DNS, with NetBox sort of systems, IP address management. So we -- again, so we can contextualize and give you the enriched data, the visibility into what does that IP address mean, what's running on that port. This includes things like the Kubernetes controller or a VMware -- VM center. Also, threat intelligence, information about security vulnerabilities, Internet performance management. So if we know what's going on, when there are performance problems in the Internet, how can those be affecting your flows in your traffic. Digital experience monitoring to then make that connection between what's going on in the network and what's going on with the user experience, actual applications and what users are perceiving. We take all that data into one unified data lake. And then from that, we can generate information about capacity planning, warn you when you might be running short on capacity or predicting when you will run short on capacity in a given area of your network. We can look at security events and security log management. We can provide visualizations of this data, business intelligence, workflow automation. And then the same foundation supports AVA, our autonomous virtual assist, which is constantly watching your network, your infrastructure, measuring and monitoring quality of experience, proactively notifying you of things that might affect you, including, of course, network detection and response. And then because we have a uniform API surface, we then enable third-party ecosystem partners to provide applications as well. This diagram list out some of the application areas. I've talked about several of these. What I didn't mention here is tax case management data. One of the really neat things we can do by getting all this data in one place along with customer case data is see the patterns in customer cases. Figure out what patterns in the network give rise to what kinds of issues and then not only help resolve cases faster, but also proactively notify customers when conditions in their network match conditions that have caused issues for other customers. So we're pretty excited about the possibility here of NetDL for making networks just much more -- much easier to operate, much more reliable. As one example, link debugging, One constant thorn in the side of the network operator is dealing with links that mostly work, whether it's dirty optics or failing optic, or some kind of a bug on one side of the link or the other. You get packet corruption, you get data loss, which can cause sort of intermittent application problems. With NetDL, we have all the data in one place to figure out what is going wrong on your link. Is it a failing optic? Does it look like a dirty optic? Does it used to work and suddenly stopped working with something replaced on the other side at that time, right? We have access to all that information and can draw those insights. As another example, threat intelligence. We can pull information about threats to the network, maybe there are vulnerabilities in other parts of the infrastructure, not even networking related, that enable attackers to use the network to break into those vulnerable systems and cause all sorts of problems. By subscribing to a live feed of that data, we can incorporate that data in. We now have information about patterns of communication, IP addresses that indicate compromise, and we can then apply those patterns to our time series, our record of what your systems have been doing historically and pinpoint where you're vulnerable, where you're exposed and where you may be actively being exploited. And actually, I'd like to show you a demo -- or I'd prefer to show you a demo of exactly that technology in action, that ability to correlate what's coming in from third-party threat see data, correlate that with the data that we're drawing in from other devices in the network. I think that should really demonstrate the power of this model, the power of state orientation of data-driven networking, of getting all the data into one place and then using that to drive intelligence, operations, decision-making. So I'd like to now turn things over to Fred. And after that, I hope you'll welcome Anshul Sadana, our Chief Operating Officer, who will talk about specific product innovation. Thank you very much.

Hi. In this demonstration, I will show how applications can correlate data from Arista's net data lake to find potentially compromised devices in the network. NetDL can ingest and store data from a variety of sources. In this case, I'll be leveraging data from Arista's CloudVision, AVA, DANZ Monitoring Fabric, and multiple external threat feeds, including Zscaler, Palo Alto and Microsoft Sentinel. My workers are subscribing to these feeds, then will take action based on the data received using CloudVision, DMF and AVA together. Here is my application dashboard. On the left, you can see it is receiving both live and historical data from CloudVision's flow collector on traffic crossing the network. On the right is a list of blacklisted IP addresses pulled from an external threat list. This means I can not only find new hosts that are contacting these threat actors, but I can find hosts that may have contacted them in the past. When I execute my data lake query, you can see that it shows me all of the flows, devices and interfaces that match one of the bad IPs. I can click on this monitor button and start monitoring that flow immediately. Over on CloudVision, we see a new change task has been automatically created to start mirroring traffic from the interface attached to the host. On the monitoring fabric filter screen, a new entry has been added to allow the traffic across the fabric over to my AVA sensor. Finally, on the AVA platform, you see threat hunting has begun on this potential threat. Furthermore, this data can be sent over to other platforms such as Splunk or Azure Sentinel for further analysis. In just a few minutes, we've shown how having access to multiple data sources in NetDL enables applications to discover new insights into the behavior of the network. We've gone from managing network state on our switches to streaming telemetry and now multiple sources of data in NetDL. This example just scratches the surface of what is possible.

Thanks, Fred, for a great demo. Hi, everyone. I'm here to talk about our products. How do we connect clients to the cloud? Our portfolio is rich and wide. We started with data centers. We look at cloud connectivity, at cloud EOS, multi-cloud. We've expanded to cognitive campus networking. We've extended EOS to support various routing use cases. And now we have a key focus on software and services. We're the only company to offer one consistent operating model for connectivity for cloud customers, service providers and enterprises. This end-to-end deployment and automation comes because of our EOS stack. Whether it's multi-cloud, cognitive campus, universal cloud networking, you need these rich APIs. You've just heard about NetDL from Jayshree and Ken. You need that data, you need the analytics, you need the inference, and then you can help the network engineers run their networks better. We do this all with one consistent EOS model across the board. Let's look at EOS NetDL a little bit more. This predictive analytics or prescriptive insights is so critical because as a network engineer, often they know they have a problem but they do not have the context. It takes a lot of digging and looking at logs and counters and flow data to find out what's going on. When we know there are better tools today, we need to leverage these and AVA on NetDL allows the network engineers to do that. Now let me start with Cognitive Campus. First of all, it's clear that the office space is changing. The new office, post COVID, we call it the cognitive workspaces because gone are the days where you're going to have tempered high walls in your cubes or your own private offices all the time. We now live in a world where everyone wants more open spaces, more shared spaces, and you need space where people can come and go as they like. But as they're interacting, we also need to support the new way of working. Everyone is on a video conference all the time. You need security. You need a lot more bandwidth. You need to worry about hackers in IoT in a different way. And that's what the Arista Cognitive Campus solution offers. When you look at our portfolio, whether it's the WiFi offering, all the way from smaller 2x2s to WiFi 6 to WiFi 60; look at our cognitive campus leaf switches, whether there are 24 ports, 48 ports, 1 gig, 2.5 gig, 5 gig, MACsec, no MACsec, 30 watts of power, 60 watts of power, we now cover it all. We even have switches that support up to 96 ports of PoE in a fixed form factor rather than forcing the customer to go stacking. For the campus spine, we have specific models that are preferred by our customers. The reality is you can use any of our products from the entire portfolio based on your use case. This portfolio has come a long way already. The one product I do want to highlight today out of this is the Arista 750 Series. This is a modular campus leaf switch, supports up to 384 users or client connections. While it has a lot of features that EOS brings to the table, it also supports 100-gig uplinks. Remember, as I mentioned, when you have that many video conferences going on simultaneously, it cannot rely on legacy oversubscribed hierarchical networks. You have to move to these cloud designs even for the employees to have a great user experience. One unique aspect about this product is to disaggregate data plane from the control plane. It broke the traditional model. The supervisor runs software, that's its only job. And when you want to upgrade the supervisor, you don't have to also forcefully upgrade your data plane. That's on a fabric card at the back of the switch, very much the learnings we have from our cloud and data center products. In addition, this product is designed with security in mind. It integrates with AVA, Awake and supports MACsec on every port. The 750 Series is designed for large enterprises and is already very positively received by our customers. Now let's talk about software and services. When you look at software, before we get into specifics of any offering or a product, why are we building software for networking? Network operations have matured over the years based on the type of customer as well as their scale. Legacy traditional enterprise customers don't really have homogenous designs that scale out. We call them the anti-patterns, we have simply one of everything, and it takes years to make a change. Large enterprises are a little bit more systematic, but it takes them quarters to make a change. And then all the way to our cloud titans that can make changes in hours. How do we help our enterprise customers largely automate? Before I show you the solution, this is what the schedule for a network engineer looks like for one of our customers on a regular basis. The network engineers have to do code reviews, decide which software should the network be running. They have to test it. By then they're done testing, they have to find all the bugs that still exist. We call that the bug scrub process. Then you have to look at all the security vulnerabilities and go address those. Other than you're done analyzing all of that, that product or offering is likely end of sale or end of life, so you have to start over. This is backwards. That's not how we live in the 2000 era today. NetOps done right means you automate from day 1. There's a dashboard that CloudVision provides to our customers, where you have essentially security vulnerabilities or the status of the network right there. CloudVision is connected to arista.com and every few hours, it gets a new update. So you do not have to do this manually. No one else does this for their customers. This offering simplifies our customers' lives to a point so they can work and focus on their strategic road map rather than day-to-day tactical work of blocking and tackling. CloudVision not only runs on-prem, on service, but we also run it in the cloud. This is extremely important for large enterprises that have a global scale. To managing 70 different offices and 10 data centers, you need a fairly large team if you have to go manage and run everything manually on site. We do this for our customers from the cloud back end at pretty much every scale. When you look at software and you look at automation, you cannot ignore network observability and security. Everyone wants to have a zero trust framework. But to have a zero trust framework, you need the infrastructure and the tools in between to monitor what's going on in your network, observe it and take the correct action. With the DANZ Monitoring Fabric, we can automate traffic steering, analytics, recording, all of the work that security and networking teams need to do together in a fully automated manner. We have taken this from security teams requesting tickets internally within their own company that takes months to handle because you have to set up new topics, tiering rules and so on to something that can be done in minutes. All of this goes into the sort of analytics tools framework, whether it's with Awake Nucleus or whether it's with our technology partners. We've spoken about AVA. It's an AI engine used to detect malicious intent. But why do you need yet another tool in security? Isn't this space already overcrowded and full? Well, what happens is most of the activity being detected by security teams or network teams today relies on logs or alerts that they have to see. In fact, some of the most famous breaches in our industry over the last decade weren't occurring because there was no tool to detect the attack. They were occurring because these alerts were hidden amongst 200,000 other alerts that the teams hadn't yet gotten to. So we need AI and ML to automate that. The challenge is the threat actors know how to hide. They know what tools people are using. And when you rely on these alerts, you can easily hide. With the Arista Network, whether it's for data centers, cognitive campus or routing use cases. With NetDL, we can now capture that data. And because we do packet captures, the threat actors cannot hide because the only way to attack is to traverse the network. And if you traverse the network, those packets will get captured, sent to Nucleus and Arista AVA engine will detect and flag all of the malicious intent or activity going on in your network. And this becomes extremely important in the world of cognitive campus and IoT for many of our customers who are driving our road map here. Talking about software and services wouldn't be complete without talking about services. We have a great customer experience. We've been on this journey since day 1, since 2008. Today, our NPS score is 74, which is outstanding. As you know, an NPS score greater than 0 is supposed to be good. NPS score of 74 is unheard of in our industry. Our customers [ start up ] 92%. But there is some specific feedback from our customers. Compared to competition, our customers believe we are superior or best-in-class 99% of the time in overall support. Many enterprises years ago told us, "Hey, I know you have great technology, but can you really support us?" I think we've proven not only can we support enterprises across the world, but we can support them better than anyone else has done so far today. We do not outsource stack, our engineers pick up the phone and handle the issue as soon as it comes in. And the feedback continues to be extremely positive. I'm very appreciative of the entire support team, the engineering team that does this for our customers day in and day out. And here is some feedback directly from one of our top customers who said, "I work with many, many vendors and Arista has, by far, the best stack in the industry." That, I think, is very satisfying for us and the entire team, and we believe we'll keep on executing on this part. So now let's talk a little bit about data centers and cloud titan designs. Before I show you what customers are doing today, this is how we started 10 plus years ago. We have these scalable leaf/spine designs. In this example, you have 4 spine switches. If one of them fails, you only lose 25% of your bandwidth. You do not have ordeals in a wide outage. This is not the legacy model of 1 plus 1 redundancy everywhere. This is a distributed architecture. This has worked beautifully. We've overcome challenges you have to scale, you have to enhance features. We've done this, as Ken mentioned, with the same EOS, every generation after generation. And today, these designs are implemented by many of our customers in a scalable fashion for 50-, 100-megawatt data centers running 200,000 to 0.5 million servers or nodes in 1 building all using this architecture. You also have the advantage where you can drain some of these sets of spines without causing any tick up, any impact to the user. As a result of that, maintenance, upgrades and so on, all become quite easy. And this is one of the reasons the cloud does not commonly go down. It stays up because you have these operational benefits to -- that's in your toolkit. The same architecture now gets scaled up for data center interconnects. When you look at data center interconnects, it's the same logic. You have a regional spine connecting multiple data centers in a region. These lengths could be as long as 100 kilometers. This is why ZR for 400-gig has become so important because without that type of an optical interconnect, you cannot withstand the distances and continue doing this architecture. We're able to provide 4,000 to 6,000 terabits of throughput to interconnect many of these data centers in a seamless fashion. You might ask, why isn't everyone then moving to 400-gig? Because you have all these great networks, you have these good products, why can't everyone just upgrade? Well, not so easy. When we went from 1 gig to 10 gig, it was very simple, 1 lane just plotted faster. We went to 40 gig, 4 lanes at 10 gig; 100-gig, 4 lanes now at 25-gig. But then what comes next? There's 400-gig, correct? Not really. There are many, many iterations and versions of this that our customers are choosing, depends on the use case. Servers are not yet pushing 400-gig from each node. You don't really need 400-gig for the common case going to the server. You might need it for the uplinks, you might need it for the spine, you might need it for DCI. And that's really how our customers are choosing. So when we see our customers are upgrading to next gen, that includes various forms of 100-, 200-, 400-gig, mostly based on 50-gig SerDes. So for us, these are all new products. But not everything goes to 400-gig. And if you remember the DCI network, unless you first upgrade DCI, you don't -- can't really upgrade the rest because you'll have bottlenecks in the path, which is why the upgrade starts from the DCI and over the next couple of years, trickles down to the rest of the network. Let's look at cloud-grade routing. When you look at our use cases for routing, we have enhanced them since 2016 with features, hardware scale and a lot more bandwidth, not just for data centers, but also backbone co-routing, clearing routing. And over the time, they've gone in more and more use cases and expanded that for our customers. In fact, if you look at where we are today, we've not only focused on routing for targeted use cases the way they exist in the past, but also for what's happening in the cloud. The cloud does not connect to the Internet through 1 or 2 links, it connects in a distributed fashion and essentially the cloud edge, whether it's at an Equinix or another meeting point. And we have evolved EOS to handle many of these use cases, whether it's for cable companies or smart cities or 5G interconnects and more metro edge use cases as well. The proximity that you need to your cloud is so important that these architectures will continue to scale and be distributed. And I think Arista EOS as well as our hardware platforms fit very well in these use cases. Talking about platforms. First of all, the EOS stack continues to evolve, but it's 1 EOS. We now support 6 architectures that EOS runs on from x86 to different types of chipsets for the data plane. It supports multiple silicon families. If there's a good chip out there, it's a very likely chance we already support it in EOS. Using this foundation, we have evolved our data center and routing portfolio. We cover use cases from small switches, 25-gig, 100-gig connections. You need features, you need [ ACO scales ], you need routing scale. We have that in MPLS. You need more [ VPN ] enhancements. Beyond EVPN, we have that. We need deep buffer for storage or for edge connectivity, we have that, too. Different flavors. And because we already cover this space really well, you'll see us continue to evolve this in a road map, but it's not as if we have a count. Our customers love this data center portfolio, and many of them are deploying it for routing use cases as well without any other modifications. Talking about hardware products, the need for more bandwidth has never been greater, and you'll hear a lot more from Andy about AI. But this is one use case where the back-end networks are getting saturated already with 400-gig, and we have the 7816, 460 terabits of throughput, virtual output queue architecture, deep buffers and every feature, security, ZR optics and option that you want is already supported and a rich road map to get to 800-gig coming soon as soon as the 400-gig cycle is done. Because of platforms like these, I believe we are already well suited to where we need to be. No talk on our platform is really complete without addressing white boxes. So a lot of talk about these white boxes. It's been there, I believe, since the start of Arista. We actually don't mind white boxes out there. Many of our customers build their own switches. But it's no longer build versus buy. The analysis they do is can they do build and buy because they get the advantage of the Arista engineering team, they get the advantage of the Arista manufacturing and supply chain team. They get the advantage of being dual-sourced on software because if they're running just their own OS and something goes wrong, they don't have a backup. They don't have a wait and escape route. Today, we support EOS or SONIC or FBOSS on many of our products. And our customers like this evolution. In fact, our partnership with our cloud titan customers has never been stronger than where we are today. In summary, we started with data center, our core market. We've expanded our portfolio and done well in data centers from cloud to enterprises. We've then enhanced and gone after cognitive campus and routing segments. We've done well there, already a breadth of offering on hardware and software products and features and more to come. In addition, our software and services segment is continuing to evolve and grow very well as well, all supported by DMF, CloudVision, AVA and now NetDL as well. I believe our platform is already market-ready to enable growth in every segment. Thank you. With that, we'll take a short break. And after that, you'll hear from Andy Bechtolsheim, our Chairman and Chief Development Officer. [Break]

Hello. I'm Andy Bechtolsheim. And I want to talk to you about what is happening with AI in the cloud in terms of what our customers do with AI and how this impacts the network requirements. Now as the backdrop here, there's obviously billions and billions of smartphones and other devices out there connecting to millions and millions of service like cloud, and in many cases running applications that invoke some type of AI component. If you look at the growth in cloud traffic over the last couple of years, it's hard to come up with a single number, but it's generally believed that it averaged out to be about 50% per year, driven by the large increases in CapEx investments being made in servers and storage and networking and so on. Now behind the scenes, of course, there's all these innovations going on in what AI can do, ranging from image recognition, speech recognition, translation, all the way to self-driving cars and health care. And these AI applications have unique requirements which is to have very, very large training sets, ranging from terabytes to petabytes to even exabytes of data. And these learning operations are continuously evolving based on the updated data and the -- and cause just a tremendous competition load. Then there are billions of users that want to be served with these AI applications, translating perhaps to 1 million requests per second that have to be serviced in real time, meaning tens of milliseconds to be effective. This requirement exceeds easily any computational task that was ever done on this planet, and it has been a major challenge to the industry in terms of how to deliver that. So as an example here, I want to use -- there's one benefit when people have solved -- solving these problems, which is that AI does not need high-precision arithmetic. It works perfectly well with what Google calls brain FLOPS or b FLOPs, which are 16-bit instead of 30-bit-ers -- 64-bit resolution, which means you can do much more computational units per chip. And people have developed, and as many chips are being developed, to do special-purpose AI chips using a very large number of these brain FLOP-type units to achieve the maximum throughput at the lowest power consumption. Using Google TPUs as an example, and all the data I'm quoting here area publicly data from Google, they have achieved a tremendous amount of growth in their scaling of their TPU clusters where TPU clusters consist of movable chips in just a few years. Back in 2017, the [indiscernible] is logarithmic, so this is a logarithmic scale. Basically, every 2 years, the performance of the cluster went up by an order of magnitude, and it was a combination of the chips going 2.5x faster and the number of chips in the cluster growing by a factor of 4. The latest Google TPU cluster vision for [indiscernible] that was just announced does an exaflop of TPU brain FLOP performance, which is 100x faster than the cluster -- the first cluster they had just 4 years ago, which was just 10 petaflops. And we can be sure they're working on a new one that will be 10x faster in a few years. Now -- and this is a picture of the exaflop machine here. Now what makes this so interesting is that the performance of these AI applications scale extremely well with the size of the cluster. This is some older data from Google that demonstrates near linear scaling for the ResNet image recognition application where it ran on 1,400 minutes on a single device and 22 minutes on the 64 cluster. Similar results have been reported from other vendors. This is an example from a start-up company from the called Graphcore that shows that as you double the size of the cluster, the throughput roughly doubles. So building larger cluster is great because you can reduce training times from weeks or months to hours. And obviously, you have the much higher throughput available to service a very large number of requests. Now click -- what this means for the AI network is the traffic required to support these large clusters is also scaling an order of magnitude roughly every 2 years, which is actually much faster than the rest of the cloud network. And there are distinct traffic patents here that are different than normal TCP/IP. There's a lot of short memory, remote access or RDMA requests and there's some large sort of elephant flows to bring data in and out of these machines. So there are 2 ways to build these networks. Google has chosen to go with a custom proprietary fabric that interconnects directly to the TPU chips and that works very well for them. Other customers have chosen to stay with Ethernet, which is the standard interface and also allows the devices to connect directly to the rest of the cloud network. For the Ethernet fabric for these applications, one needs predictable latency and very, very high throughput and no packet loss because the kind of transactions that are going on here, again, are not TCP/IP network transactions. And what we've seen so far with customers that have deployed these solutions is a very good application performance comparable to, say, the traditional InfiniBand network which is another not exactly standard solution that doesn't tie directly into the rest of the cloud network. Now in terms of the ideal product for this application, it is actually our Arista 7800 big buffer virtual output queuing switch which delivers an outstanding throughput of up to 460 terabits per second and no congestion VoQ fabric and the big buffers that eliminate the backdrop for the RDMA application. To sum this up, the cloud is looking at Ethernet as a way to integrate the very high demands of the AI applications. As we talked about every -- the chips are going fast all the time, the clusters are getting larger. By the way each chip, even today, requires typically 10x the throughput of a traditional Intel-type CPU. And the clusters are getting bigger all the time, driving a large number of 400-gig and future Ethernet gig ports. And again, last not least, AI clusters need predictable latency, no packet loss for which our VoQ switches are the best architecture. And with this, I would like to introduce the most important person for this conference here, which is our Chief Financial Officer, Ita Brennan.

Thank you, Andy. I'll now pull all of the numbers piece of what we've heard today together as part of our financial presentation. So again, we'll talk a little bit about growth and diversification. And I think really for Arista and the stage of our growth and of our maturity, those 2 things go hand-in-hand and go together. So we'll talk about that a little bit. We'll talk about investment and where we want to invest our cash and our money, and we'll talk about capital allocation. And then finally, we'll take another look at the business model and how we see that going forward from here. One of the tenets of Arista right from its founding was this concept of profitable revenue growth. Not only did we want to grow business, we wanted to grow the business profitably. And I think one of the key reasons why that has worked so well is the fundamental foundation of EOS and the idea that you've built an architecture on a foundation that allowed you to evolve and enter into adjacent markets, enter into adjacent product sets, without having to redo the entire operating system from scratch. So instead of having a huge lift every time we wanted to expand into an adjacency, it was more of an incremental build and you were still leveraging that fundamental EOS architecture as its space. And that has allowed us to do a lot more, more efficiently, drive more leverage, and it will continue to do that, and you saw that again today with NetDL. So we're expanding our TAM very efficiently and really taking adjacencies that are coming to us as time evolves and as our capabilities have evolved. This is our TAM chart, and you can see we're addressing a larger TAM, and that TAM is growing. If you look at the blue piece of this chart, it's very much the -- this is the core data center networking, switching part of the market, and it's growing. And we're addressing pretty much 100% of that market at this point. The next block up is what we call our cognitive adjacencies. This is where we put campus, this is where we put some of the routing capabilities that we've added to the product. Campus, we're addressing more of that campus market as time goes on. And therefore, the market has actually returned to growth. But in addition to that, we are adding more of the parts of the market that we can address. Then on the top, we have our services and software. This, again, is the services business, which Anshul talked about. Very important part of our business, growing very nicely. It's a recurring revenue model, a very high gross margin recurring revenue model. And on top of that, we're layering these software capabilities, right? We're layering some of the additional visibility tools, some of the monitoring tools. And finally, some of the security capabilities that we're adding over time. And all of this is bringing us to this growing and very healthy TAM. I think I've talked to most of you at some point in time about this concept of building blocks of growth. For a company to grow consistently, the company needs to have sufficient building blocks, and sufficient building blocks of sufficient size and scale to enable the company to grow in the business to continue to grow because not every element of the business is going to grow linearly over time. And really, having these building blocks allows you to have some pieces of the business perform well at certain times. And yet -- and others not so well and still have the business grow and continue to grow. So that's been a focus for us of late. We've been driving product diversification around some of the routing, campus, monitoring and other capabilities. And we've also been driving customer diversification. So we've now gotten to the point where we believe we have these very strong areas of diversification, all of which can contribute to our growth as we go forward. And we'll look at each of these separately and see where we stand in kind of that evolution as we go through each of these. So we'll start with enterprise, and I'm very deliberately skipping and starting with enterprise just because I think it's the unsung hero of the Arista story of late. We tend to focus on cloud because that's where a lot of the more interesting and exciting activities are. But really, this enterprise business has been growing very steadily for us. You can see that it's been growing step by step. We have a 30% CAGR over the last 5 years, and it's been growing on a very consistent basis. It even grew right through to FY '20 when we had -- when you started to see COVID and COVID was impacting different parts of the business. So again, very consistent growth. It starts with focusing on the data center. And it's a beautiful land-and-expand model, right? Because you really focus -- we win an opportunity in the data center. Once we gain access to that customer into that opportunity, these product and these capabilities are demonstrated, and then we get the opportunity to go and work on the adjacencies. And finally, work on services and software subscription and other areas, right? So will see -- we see this repeated time and time again. And we're early in this kind of -- in this evolution. When you think about our enterprise business, the most mature piece of that enterprise business is probably the financial. And yet when you sit and look at the financial customer base and where we play and how much we play in those accounts, there's still a lot to do even there, right? So it's a very early part of the business. Look at some of the newer verticals where we didn't play before, say, industrial, commercial, and we're only at the very beginning of some of those verticals. So there's a lot for us to do and a lot of runway for us to continue to grow leveraging this piece of the business. And we'll continue to do that over time. So enterprise, we talked about overall enterprise growth of 30%. Within that is obviously the campus business, which has been growing very aggressively. We've grown that 100% now twice and Jayshree and the team have just signed up again to grow that again to 100% this coming year. So very aggressive growth on the campus, very good progress there and leveraging just this whole go-to-market and land-and-expand model that we've managed to create for the enterprise business. Then there is cloud. Everybody likes to talk about cloud, right? We spend a lot of time talking about cloud. It is a very important part of our business. It's near and dear to our hearts. We'll continue to invest in that part of the business. It's an area of the business that challenges us all the time, challenges both from a technology perspective, from a fulfillment and operations perspective. They're continuing to cause us to change and adapt and respond because they're a leading edge and they're driving new initiatives. And I couldn't think of a better team to kind of interact with those customers and to help evolve that. We've seen healthy growth in the cloud up until probably 2019. All we had seen was up into the right with this cloud vertical. We did see some pause in spending at the end of 2019 and heading into 2020. So I think it is a cyclical business to some extent. But it's a cyclical business that over time is growing and the investments that they're making in the network are very, very significant and varied. So our -- the role that Anshul and the team have played here is focus on the data center, but then we've been adding these adjacencies here, too. It's been routing, DCI, WAN, edge, continuing to find new ways that we can interact with these customers and new pieces of their business that we can support and win for. And you'll see us continue to do that. So if you look at the top chart on the top right-hand corner, you can see that we're just at the very beginning of this next product cycle. So we expect that we'll see some more aggressive growth in the cloud business over the next year plus. And really, honestly, it's constrained more by supply than it is necessarily demand at this point. We'll see some deferral of revenue, et cetera, that will temper some of that growth as we head into next year. But ultimately, that business is actually returning to a reasonably aggressive growth level, and we see that on the demand side and on the booking side. And then there's the providers. This is really 2 pieces of our business combined, right? We have the specialty cloud providers and then we have the service providers. The specialty cloud providers is obviously a very good overlap to what the large cloud customers need and want from us from a product and a support perspective. We've seen some volatility in that part of the business but really more because these customers were figuring out exactly what their differentiation was going to be, how they were going to compete in their markets. But once that's been established and we've seen that, we've been through that over the last year or 2 years, you start to see them return to growth, and we have seen that piece of the business also contribute to growth over the last period of time. And the other piece of that business is the service provider business. So the service provider business is probably the newest piece of our business. We started to really interact with the service provider business with the routing capabilities, and we brought the switch routing products to market. And we found some very good use cases that could leverage that technology early on. And that drove some good growth in this piece of the business for us as well, right? As we've gone forward, we're finding that we need to develop some service provider-specific software features, et cetera, to support that business, and we're continuing to do that. Again, the leverage both for us and for the customer, of them -- of adopting the technologies that we're building for the cloud and for the specialty cloud customers, it's very real and very tantalizing as something that we should all be able to capture, but it takes time, and it's requiring us to continue to do some development on the software and the service provider-specific routing features and software. But we're doing that, and we'll continue to do that, and we're remaining engaged here because we really think this could be an important part of the business. We're at the point where I think we get a good seat at the table. Whether we've reached the point where we can consistently put together new opportunities and have this grow consistently over time, I think that's still a TBD and something that we're very much focused on and working on as we go forward. Let's talk a little bit about gross margin. Now this is an area that has a lot of focus right now. And honestly, a lot of complexity. We have our normal complexity of customer mix where there's quite a difference between the gross margin depending on the size of the deal that we're transacting. You're now layering in some supply chain -- or a lot of supply chain complexity in terms of just expedite fees, costs, et cetera. And to offset that, we have introduced a customer price increase that we'll be rolling out to customers here over the next while. The timing of some of that is not perfect, right? So we've been recording some of those costs in the P&L already. We'll continue to attempt to match as much as possible the customer pricing with the cost increases and other things that are flowing to inventory. That may not be perfect, but it does definitely help and contribute to that over time. There's obviously a product benefit to this as well, the more software and the more services that we do, the more growth that we have in those parts of the business, and that also helps the gross margin line and helps accrete to the gross margin line. So I think when I think about gross margins, I'd encourage you to -- we think we can stay in our 63% to 65% range. But don't assume that we'll be at the upper end of that range just because we've been there for the last 2 quarters. The last 2 quarters have had a particularly high enterprise mix. We were deferring some cloud stuff, et cetera, this quarter. So it's somewhat masking some of the supply chain impacts that we see. So I would think about that very much as a range. And a range that if we are heavy on the cloud content, then we would be at the lower end of that range. But I think for the year, it's still a good, solid range as we look forward. Let's talk about investments. Again, we want to and we will continue to invest in the business. From an R&D perspective, that's very much hiring and growing the team. Whether that's organically or that's through an acquihire of some sort, if we could find the right teams, we are looking at ways to continue to grow that team. You've heard some of the initiatives that are on the table today, we need to continue to hire to fund those initiatives. There's a lot of good returns to be had on continuing to grow that team and continuing to develop products with that team. So you'll see us continue to do that. Sales and marketing. I think we found a reasonable cadence now where we're adding roughly 30%, plus or minus, to the overall sales team as we go each year. We're starting to invest very targeted investments in the channels and addressing some channel activity as well. So we'll continue to do that and maybe even accelerate that a little bit if we have the opportunity to do that. And then finally, on the G&A side. We obviously run a very efficient G&A organization. We will make some investments here around some tools and some IT capabilities to help us continue to scale as we broaden the breadth of the business and we include some of the other parts of the business other than just cloud and that core enterprise area. So we'll see some further investment there as well. So really, more of the same, but maybe a little bit more accelerated than what we saw this year. We talked a lot about go-to-market and how Arista's model of 8% to 10% of revenue can't possibly support an enterprise business. And I think I just keep reminding you that the vast majority of our dollars that we invest in sales and marketing go to that enterprise business, right? So it's really a much higher percentage of the enterprise revenue that we're investing there, and we're targeting kind of large and medium-sized companies. So those 2 things combined, we feel, makes that a reasonable level of investment. Like I said, we'll go a little bit faster. So you should think about it maybe in this 8% to 10% of overall revenue perspective. But the overall model in and of itself, we don't believe needs to change fundamentally. Let's talk a little bit about capital allocation, and what do we do with the cash that we're generating because again, we have a very high cash generative business. It's a profitable business, and it's growing, it's returning to growth at this time frame. So we still believe we need to maintain a very healthy balance sheet. I mean it is an industry and we have competitors that are much larger than we are and can be aggressive in various ways in the marketplace. We will always want to maintain a very healthy and solid balance sheet. We want to invest in the business and have the ability to invest in the business. And this has taken all sorts of different forms over time whether it's hiring. Today, it's purchase commitment. Today, it's having the freedom to be able to go out into the supply base and make commitments for purchase commitments to the tune of $2.1 billion, $2.2 billion. And know, and have suppliers know that we have the means of which to do that. It's not going to consume the cash all at once, but it's having that ability to be able to make those commitments and to be comparative in the industry in terms of being able to make those commitments. I think that's very important. It's just another demonstration of where -- why we feel so strongly that we do need to retain a healthy cash balance. M&A. We found the right M&A opportunities. We obviously would go ahead and progress with those. And -- but we know that we're generating so much cash. There probably is some of it that we can return, and we have been returning cash to shareholders. You've seen us execute on our first $1 billion. We're almost done, and the Board did just approve adding a further $1 billion to that authorization. So we will continue to return cash to shareholders as well as executing on those other items. Then obviously, a dividend at some point in our lives will make sense, but probably not yet. So you should think about the return of cash being very much focused on the buybacks and the share buybacks, and we have upped that authorization, and we'll continue to execute on that. So then to get to the business model. 2021 here is obviously a forecast at midpoint of our guidance that we gave you today, right? That gives us roughly a 25% plus or minus growth rate. As we look at 2022, we believe having looked at this in multiple different ways, that 30% is a good, solid starting point for us from our growth rate for 2022. It is supply constrained, right? I think supply is shaping the growth rates in the business right now, more than probably we've ever seen that before. So it is a supply-constrained number. We've tried to reflect the risks that we see from a supply chain perspective, and this is kind of the number that we're coming up with. The gross margin, again, 63% to 65%. Again, please don't hang out at the upper end of that just because we have. I think you should be solidly in the middle, and there'll be quarters where we'll be in the lower end of that. We'll continue to invest, and I know you're going to tell me, we've never managed to grow OpEx of 30% plus. But we grew OpEx in 2021 at 20% plus, and it was a very difficult year to spend money. It's was very difficult year to hire and retain people. So we are hoping we can do a little bit better in 2022. We'll see if we exceed or if we reach the 30% range, maybe, maybe not. But we do want to aggressively pursue resources in the areas that we've talked about as we go through 2022. Looking longer term. It's difficult always to come up with kind of a long-term growth rate. I think when we look at the different pieces of the business and now we've given you the building blocks, so you can look at those building blocks as well. We feel there'll be periods of accelerated growth where cloud is active, and we're in a very positive cycle with cloud. And there will be times when cloud is less active. And that will kind of shape the curve somewhat. But now we're in a much better place with our other building blocks where we think they can grow consistently. So I think at this point, we would say there's a strong case we made that this could be a mid-teens grower over time. We've talked about looking at a CAGR from 2020 to 2025 that you could have a good solid mid-teens growth rate, again, as a base case that you could feel really good about. And then obviously, we'll update as we go. Because nobody has perfect visibility that far out. Again, I think the rest of the metrics stay reasonably the same. As we grow, we'll look to invest and we'll continue to invest in the business. So that's kind of our view, our bookends, if you like, for 2022, and just a view of what we think the longer-term growth rate could look like. And then obviously, as time passes, we'll continue to update that, and we'll continue to share more with you on that. So that's our business model slide. That's the last slide that I had from a financials perspective. We're going to let you hear from some customers now while we prepare for the Q&A, and then we'll bring you back to join us for a live Q&A. Thank you very much. [Presentation]

All right. Well, welcome to the Q&A portion of our event. Joining me on stage is the entire executive team that you just heard from. Due to limited time and out of consideration for everyone on the call, please limit your questions to a single question.

Our first question is going to come from Meta from Morgan Stanley.

I just wanted to get a sense of how NetDL works for some of your cloud tools. And of course, how does it work with some of the other cloud tools that your customers may be using? So just how are they interacting with the data that they're getting from AWS or Azure or GCP? And how can they kind of have a singular point of network management between those kind of 2 different networks that they're running?

The NetDL vision is exactly to integrate data sources from multiple points of view. So to be able to take data from AWS, from GCP, from the networking stacks, especially, but also application management and performance management aspects of those services and bring them into a common data lake so that we can then give you the information in context. We can make those connections between what's happening in the cloud and what your users are actually experiencing. That is absolutely the vision.

Okay, Ken, one of the features you have with NetDL is Cloud Tracer today. You can actually look at the round trip latency of the traffic coming from a public cloud into an Arista premise that may be running cloud EOS with NetDL and measure the total latency and what's the best round-trip and where do you take your workloads, what's the best route, et cetera.

That's a great point. I think Cloud Tracer is an example of an early application.

Great. Thanks, Meta. Our next question will come from Paul with Cowen and Company. Paul, are you on mute?

Not any more. Can you hear me?

All right. Yes, we can hear you.

Well, first off, I just want to congratulate Alex and then finally, who can finally leave [indiscernible] behind him. Now for my question. Hopefully, you'll answer it, I'll give it a shot. To get to that mid-teens growth rate through 2025, can you give us some insight on how you get there from a customer perspective? On what type of growth you're expecting from enterprise? What sort of growth you're expecting from cloud? What type of growth you're expecting from communication service providers?

Yes. I think, Paul, the challenge is it's not one scenario, right? I mean it can't be. You're talking about a long-term view of the world. I think how we've looked at it is saying, what are the different scenarios, what are the different growth scenarios, what are the different pieces that would get you there, right? And that's -- we're trying to come up with an answer that is robust enough that we can find multiple different ways to get there. I mean there's obviously an assumption in there that, like we talked about, that maybe cloud is a little more cyclical. So we're looking at that to say, okay, can we continue to achieve that even in that scenario? There's another scenario where maybe this 3 providers business is more -- is stronger than we expect. It's looking at various different scenarios. And saying, when you look at all of those, what's a reasonable growth rate that you think you can achieved in multiple different scenarios as opposed it is that we think we have 1 path to get there.

Yes I think -- it's important to understand there isn't 1 recipe to get to our CAGR of 15%. It's possible that we may have a single-digit growth in the cloud down the road. And at that time, we would need the specialty cloud providers or service providers or enterprise to kick in with double digits. I think the beauty of this is, even if it's not linear 15% across all the verticals and all the products, we can see many parts to get there.

Jayshree, I trust to the point that your customers are giving you larger visibility, plus you have the benefit of 400 gig, [ 600 ] gig upgrade cycle. I trust that from the past experience with 100 gig upgrade cycle and the insight you're getting from Facebook, Microsoft and others, that this visibility extends well beyond '22, out into '23, maybe even '24 in terms of how those plans translate into rough numbers.

Paul, like I said in the earnings call, and Ita will echo this as well as Anshul, our visibility on 2022 is great, which is why we've given you some very aggressive growth numbers. Our visibility into 2023, '24, '25, less so. But we're hoping our strength into 2022 will carry us into the next few years. And obviously, there's a whole lot of execution. So visibility, very good for the next year, less so after that.

Thanks, Paul.

Thanks so much, Paul. Our next question comes from Jason of William Blair. Oh, are you on mute?

Yes, I wanted to ask about how you plan to monetize that deal. Is there a specific add on? I guess I know that it's part of the core platform here, but is there -- are there specific ways that you feel you can monetize that deal?

I'll take that and maybe Ken and Anshul. NetDL is monetized through every EOS-based system and switch, right? So we -- and then on top of that, we plan to expand this to ecosystem of partners. So today, our monetization of NetDL would largely be an expression of the EOS and the platforms, and sometimes maybe the software and services that we can expand on top of that through services or Ava or network detection NDR platforms, et cetera. The greater monetization is more a vision and aspiration, it's not short term, where we can actually work with our ecosystem of partners. And examples of that would be the work we're already doing with Slack on notification of some of our automation alerts or the work we're doing with Splunk, where they can -- the demo you saw from Fred Hsu, where you can see some of the digital optical monitoring on a Splunk dashboard as alerts. So we haven't yet formally done those kind of Arista marketplace, but the possibilities are very much there.

Great. Thanks, Jason. Our next question comes from Ittai of Oppenheimer.

First of all, congrats on the target for the $5 billion. I remember you're running at about $500 million at the time of the IPO. So you can see all the growth over a decade, it's quite impressive, and congrats on the work there. I just had a small question for you, guys, maybe one for Ken or Ita on just -- on the guidance, I guess, should we assume the 10 points out of the growth next year is purely price increase-related? And then for you, Ken, taking a high risk here, talking about the NetDL. We always see the development of many ops-related data lakes with performance monitoring tools. And then we also have the security guys with [ XGR ] and [ security ] solutions. So I guess the question is, what makes NetDL a source of gravity that we think customers will want to put the data there versus take your data and just ship it to their own existing data lakes and not use your service at all?

Do you want me to take the pricing one first? I think look, it depends on the timing, right? I mean we're going to look to adjust pricing on a go-forward basis. There's a -- there's some backlog, et cetera that will be adjusted, not adjusted, et cetera. So I think I wouldn't take the whole 10% and say that's pricing. I mean, there will obviously be some benefit from the pricing, but it's not necessarily that it will be applied across 100% of the business.

I thought it was 1 question per customer, wasn't it?

Yes, there's always a 1A and 1B, Ken.

There you go.

Yes. There you go.

So look, I don't think there's anything about NetDL that requires customers to give up any other data lake solution which is working for them. The critical thing for us is that we've extended our architecture to be able to ingest state from all of these other systems and to then correlate and draw inferences and make that -- those inferences suitable to our -- useful and actionable to our customers. If they find value in that, they can then stream state from any of their existing systems into NetDL and take advantage of those. And they -- again, it's not either or. So they can get the value there as well as making whatever existing use they are making of their existing system. So I don't regard this as competitive as much as accretive, but it increases the value to the customer of their data because we can ingest it and increase the visibility of insight [indiscernible].

All right. Thanks, Ittai. Our next question will come from Amit of Evercore.

This is going to make my 1A and 1B really awkward to ask, Ken, but I'm going to do it anyway. So I guess I feel like without historically you've been upside to the cloud if you haven't [indiscernible] the cloud titan companies and historically be like once you have this good upside, you don't put more thought. Digestion then tends to be more challenging, what I'd like to see the uplift over here that you see. Can you just talk about maybe 2 vectors I want to understand. A, how do you get confident that this is truly a strength to the end demand by the hyperscale vendors versus perhaps these customers are just prebuying ahead of all supply chains issues that are out there? Do you have any confidence that what you're selling to them is getting deployed versus getting stored, I guess? And then secondly, can you just talk, about as I think about the 5-year road map you laid out, what does that mean for your services and software mix as you go forward? And does that growth faster in the overall revenue CAGR number you talked about?

Sure. Amit, I can answer the cloud titan question and then I'll defer to Jayshree and Ita on the other one. When you look at the cloud titans, their business is strong, right? With COVID, the need for the cloud increased, and many companies are migrating more and more workloads there. As you very well know, we've worked with them very closely. I'm quite aware of the schedule for various sites, where the needs are. And to a great extent, we do work, collaborate with them, and we have a very good handle on how much is inventory on hand for them versus what's needed at each data center to go live. So with all of that in the background, we don't believe that this current increase is because of them trying to just build their own inventory and stock or stockpile on products. They do need it to grow their business. If you look at each of the cloud titans, their business are growing. The revenue is growing. The investments are growing. As a result, they need more compute. They need more storage. They need more GPUs and AI, but they also need more networking, which we are here to provide.

Yes. I would also refer you to Andy's talk and looking at the sort of the bandwidth requirements of modern applications. So it is not getting -- we are -- Moore's law marches along, and we are finding ways to use that capacity, and the cloud guys have to build for that. And so there's no time for them to leave switches just sitting on the shelf. They stuff they're buying, they deploy.

And then back to the services software question a little bit. I think the services business continues to scale, obviously, as the revenue scales. We will eventually get to a point where some of the installed base will start to retire, and we won't have that same renewal momentum that we have seen in the past. Some analysis we've looked at it. I mean I think it can grow -- continue to grow over time in that kind of mid-teens, maybe a little bit faster based with the pure services business. The software business, that's really an enterprise. If you think about the different parts of the business, that's really an enterprise story. So I think as we scale the enterprise business, we're seeing that software element grow as well. And it's been growing somewhat in line with the enterprise growth rates that we see. I think as we add more solutions, more capabilities, then you can probably accelerate that. But I think for now, it's been growing roughly around -- at the same rate of the growth of that enterprise piece of the business.

Great. Thanks so much, Amit. Our next question comes from Rod of Goldman Sachs.

I've got 2, I guess, one technical one for Ken, and then I've a question on the 30%. So play up to 30%, that's an easy one, which is how would you break that down across the 3 verticals? Any color you can give us on that? And then for Ken, I wanted to ask about NetDL and its competitiveness. It seems like, for example, accessing that by a bunch of third parties and maintain performance would be difficult. Can you just talk to us about how the sensible -- that technology is, it seems like it probably is, but maybe without giving the farm away, give us a little bit more detail on why it's going to be tough for people to duplicate that?

Yes. So I think on the mix of the business, we have been operating with a heavier enterprise mix over the last couple of quarters, and you've seen that, right? Even longer than that probably. You saw that in the growth charts that we had. So I think there's no doubt that the cloud accelerates as we head through this next period. How long it lasts, what that slope looks like, we'll have to see over time. But I think it definitely accelerates. So I do think you see more -- one scenario for sure is that you see more aggressive growth from the cloud customers than we've seen in the past, and they kind of returned to that kind of trend. We talked about the 35%, 39%, they returned to a healthy position in that range, while enterprise and the other pieces continue to grow.

And I think the other unsung heroes are still the specialty cloud providers who are starting to build their own specialized cloud. So that's going to be very strong as well. So we believe that actually all our sectors will grow double digits. And the question is, are they going to grow at 15% or 20%? Or at that 30% we cited for the overall or more?

Great. Thank you so much, Rod. Our next question will come from Alex with Needham & Company.

I was hoping you could talk a little bit about the competitive landscape in the context of Cisco saying -- they're talking about 160% growth in their cloud customer orders to 30% of their service provider off the last call. But more importantly, as we move forward and start thinking about the direction of the company, how you're going to reach coders, how you want to tie into the DevOps community and how important is that relative to your business as we go forward? Are you tying into the HashiCorps? Are you going to be competing increasingly with the cloud players and maybe MATRIXs of the world? How do we think about that DevOps world and Arista?

Alex, I can touch on the DevOps interaction with our customers. When you look at the Tier 2 cloud, cloud specialty providers, many of them have actually come from cloud titans. They're very used to that world of automation, CICD pipeline, and they essentially write software, right? They're not classic network engineers. When we look at Arista EOS, when we look at CloudVision, NetDL to some extent is driven by this type of customer base where they're looking to integrate. They're looking to integrate their systems. And since day 1, we support open APIs, the APIs CloudVision uses to talk to our switches is the same API that's been available to our customers. As a result of that, it integrates really, really well. In fact, in the last couple of years, we've now exported or supported all of Ansible integration directly with CloudVision. It's not meant to be a monolithic sort of locked-in system. It's designed to be integrated with what customers are doing with the overall infrastructure. So I think EOS, CloudVision, NetDL, the APIs we have, fit really, really well with that ecosystem. And as you very well know, that is growing very well beyond the cloud to large enterprises, financials, media companies, and they all love our products for that reason as well.

And to answer your question, Alex, on competition, the cloud titans never told us they'd give us 100% of the share. They are looking for silicon diversity. There are other use cases like optics or routing layers or switching layers, particularly with the supply chain issues, they're always looking for more than 1 vendor. We just want our fair share of it, and I'm sure others will get some of it too.

Thank you, Alex. Our next question will come from Tal of Bank of America.

I have a question for Andy. You -- in your presentation, you spoke about basically 2 things. You spoke about the need for units, more units, more processing power, et cetera, I mean you shared with us the hybrid cloud kind of processing needs. But then there was also [ edited ] features. The second issue was expansion of the market. So when I look at your targets for the next few years, I want to kind of connect the bottoms up to the top down and ask, how do you think -- when you think about campus switching new opportunity, routing new opportunity, but when you think about your core markets of switching, how do you think about it in terms of growth? Is it a question of the market is going to grow faster because of the added features that you spoke about? So do you expect market growth acceleration? Or do you think about it in terms of share gain? Meaning the market will continue to grow at the same rate, but because of the quality of the product, the share gain will accelerate? Or a third option, which is added value, meaning you can grow and the growth will come from areas that are not counted in classical switching, the way we count the numbers. So I'm just trying to understand what does it mean for the market? And what does it mean for your market share?

So what I was talking about is just the growth driven by the new applications in the cloud, the AI-type applications where a single AI chip can easily consume 10x the bandwidth of a current CPU chip. So even though the AI chips are less numerous in numbers, they're growing quickly, but they have 10x the bandwidth requirements. And thus, they're causing a very significant increase in overall metric demand bandwidth in the cloud. So I was discussing an industry trend. I use Google as an example because they have been fairly public about their remarkable achievements. But you can be assured there's lots of AI investments going on in almost any large cloud company. It's probably one of the best-kept secrets in the industry, what they're doing with this AI, but it clearly has become a business driver for them, right? It has enabled them to monetize these troves of data they have in a better way than before. And you're seeing that in the results that these companies have been able to generate. So AI has become a big investment, certainly in the large cloud, to the point where it is becoming a significant driver of network bandwidth, which is where our networks come in. So I was really describing the overall scenario that the cloud bandwidth is expanding at a very healthy clip which supports our business forecast.

And does it mean -- and so that's why I wanted to ask you. Does it mean that the market growth will accelerate because of the investment in AI? Or will the market growth remain the same, and you're basically better positioned, so your market share will go up? I'm trying to translate it into margin growth and market share gains.

I can't speak to that. But what I can say is today's general purpose CPU have 50 gigabit, maybe 100 gigabit connection per server. On the AI side, they're looking at 400 gig going to 800 gig per chip. So the densities and the amount of bandwidth required for these AI clusters is just off the charts compared to traditional network servers. And even though it's a smaller percentage of the total data center, it's having an outsized impact on the overall, the interest demand.

Great. Thanks, Tal. Our next question comes from Aaron of Wells Fargo.

I'm going to try and couple 2 questions really quickly together, and A and B, as Amit said. So, there seems to be a lot of architectural evolution happening inside the data center, and one of the things that we've started to see is this role around taking certain attributes of the stack and moving them off the TCU and into a data processing unit, a smart [indiscernible], and some will talk about networking layers as being part of that. I'm curious of how you see -- this is for Andy here or the team, does that evolution means something for the Arista strategy going forward as it accelerates some of the opportunities around AI, et cetera? And then also in the commentary on today's call, you alluded to the data center investment. Are you building the data centers out? Do you have to make those data center investments for NetDL as you're capturing a little bit of data? Is that how I think that -- or a.m. I taking something differently?

Yes, if I could address the data center question first. We have acquired some land in Santa Clara, and we are planning to build our own data center building for our own internal use. We're not building this for customers. This is purely used for software testing, 24/7 regression testing. It's going to be a fairly large data center, and we concluded it was much more cost effective to bring it in-house than leasing space from third parties.

If I can add something, Andy. In the past, a lot of people just asked us, what's the size of your QA team. And today, our answer is 8 megawatts. We don't have any people, that's where Ken's auto test really runs, the test EOS and we need expansion in that area for sure. I can take the second part of the question, which is all of these smart mix or DPUs and other changes that people are trying to bring to the deals in the market. Largely speaking, I still consider them as offloads for the server. But they were on the [ net ], the [ net ] became too heavy. And when the system becomes monolithic, it's hard to evolve to the next gen, have the cadence that you want. All of our large cloud customers have looked at these technologies don't really want that integrated deep into the network because they want the network to remain simple, scalable and they already have their hands full. You saw the DCI chart. Sometimes you're looking at 10,000 links that you need to monitor and make sure work error free or you switch without dropping any packets. And the cloud customers, I think, will keep on focusing on those kinds of operational enhancements rather than making the network more complex. So DPU, smart mix, appliances, I think that's all about the server market and where things might shift.

And the good thing about that is it feeds more bandwidth into the network market. So it works well for us. It's symbiotic.

Thanks, Aaron. We have -- our next question comes from Samik from JPMorgan.

I guess I just wanted to talk a bit more about your enterprise business or your enterprise wins and what you're seeing there, particularly as we think about what you mentioned as the land-and-expand motion. Are you getting these findings at the displaced incumbent completely over time? Or is the customer looking to use you as a second or alternate supplier in a lot of the cases? Just trying to think about what's your fair share when we think about enterprise data center. And in the long run about campus, what's a fair share for Arista, given that you've already kind of established yourselves as a leading player in cloud data centers? And just a quick second part, when you think about the $5 billion target, what's your vision of what campus looks like as part of that?

Samik, just to answer your question, I mentioned this. We have now got a seat at the table with enterprise customers. I don't think they had a real alternative until now. The dominant player was one, right? And unlike servers or storage and networking, there hasn't been that clear alternative or option. But today, I think Arista's seat at the table goes beyond just saying data center or campus or routing or observability or software services. It's really about building their next-generation network. And that's why we're so enthusiastic about this next frontier, which is driving all these data sets. They are asking, especially with the pandemic. They're asking to plan their network differently. They're asking for help in getting from that legacy to the new. And they're seeing that there's all these disparate software architectures when they're dealing with the current network, and we could give them much more of a simple, easy, one image, one platform, one cloud vision, one OS, different use cases. So we are coming to the table with different use cases, better performance, better innovation, superior quality, better support. And there hasn't been a platform like CloudVision to gather and stream and automate all this data in one place. So I think we're really getting a seat at the table, not just for individual markets, but to help them design their next-generation network. And then your other question was on campus?

Yes, and $5 billion [indiscernible]. Yes.

It would be difficult to keep doubling. But by 2025, when we hit that $5 billion, I would hope we could be 15% of that number.

Great. Thanks, Samik. We have time for one last question. Our last question is going to come from Rod. I hear we skipped your 1B. So...

It was just a [indiscernible]...

I won't get away with it.

Yes, I just wanted to come back to Ken on the technology nodes or the rest of the team, if you want to comment. And I want to check 1 thing, Andy said, too. Andy, you told me in 2018, I think it was 100 gigs per second per GPU. And now you're saying 4 to 8x that. Is that rate growing as [ speeding ] GPUs data? So do we assume that drives networking growth in and of itself? Or are we kind of peaked out in terms of utilization here? But I'd also like to hear this competitive question answered.

Yes. I mean if you look at the latest TPU/IPU GPU chips, they are roughly 4x faster than the chips they had 3 years ago. So this is consistent with the increase in performance of these devices.

Andy, if I can add to this because Rod, you and a couple of others have asked. The current workloads which require CPUs or their storage and lots of application and so on, will continue to grow. The AI workloads are new. And I don't think there's a perfect model out there today to project what the time will look like 5 to 10 years from now. It could be much bigger, it may not be much bigger. We just don't know. It's too early. But these AI networks do drive a lot more bandwidth and are more similar to storage 5 years ago, where you have a completely isolated back-end network for all that traffic to go through within the cluster.

That's a really good point. The AI network is almost a parallel network to the data center network. So with high bandwidth, with predictable latency. Ken, you get to answer the 1B question.

Okay. With regard to the competitive situation around NetDL. Your -- part of your question was what's stopping somebody else from building this? And I think the answer at a high level is, well, nothing, but they have to actually do it. And it does -- it is very difficult for our competitors with their menagerie of operating systems, their Noah's arc of forks and branches and releases, and they've been cobbled together, a huge collection of acquisitions. It's -- they don't have the uniform data model. They don't have a uniform streaming format, and they don't have a coordination across those different lines of business or market segments and engineering execution side to come up with a unifying architecture effectively. And so I think we do have a structural competitive advantage that we've been state-oriented since day 1. We have 1 EOS that runs across every one of these use cases. And we have a team committed to the vision of creating better software. Doing software right, it takes a little extra time, it takes a little extra time. And we're not fragmented into different waring tribes with our different market and revenue targets. We take is a unified approach. So I think we have some structural advantages that makes us more effective at tackling this problem than maybe some of our competitors. But certainly, there's no -- we're filing our patents, of course, that sort of thing. But there's no -- I see -- I don't see any fundamental competitive barrier for others doing the same thing. Just like there's no fundamental competitive barrier to competitors creating highly reliable switch operating systems just so far they haven't done that. And so we're expecting that if we execute well, we will be very successful in this area.

Thanks so much, Rod. And thank you to the executive team. That's going to wrap up our Q&A session for this event. A recording of this event will be available on our website as well as some of the company's slide decks and additional resources. This concludes Arista 2021 Analyst Day. Thank you so much for joining us.