Hi everyone, I'm Nicolai Tangen, CEO of the Norwegian Sovereign Wealth Fund, and today I'm here with someone who is very important for AI, Jayshree Ullal, CEO of Arista Networks. Jayshree has over 40 years of experience in networking, so no one can talk about this better than her. Jayshree, it's great to see you.
It's great to be here. Nicola, I have a bit of a cold, so please forgive me if I speak through my nose.
No problem. Let's start from the beginning. I don't have to sing on this podcast anyway, so it will be fine, right?
Yes, no singing. So let's start from the beginning. What does Arista do? That's a tough question. Arista is building the world's most demanding mission-critical networks. They are high-performance, high-scale, low-latency, high-reliability, high-availability, high-automation, and high-telemetry, meeting the needs of modern data centers and AI. There is very high demand at different levels, and especially in the last decade, Arista has emerged as both a pioneer and a leader in this field.
I've heard that if you think of chips as engines, then networking is the highway between them, right?
Yes, that's a good way to look at it. I believe chips drive a lot, especially in the AI world where processing happens on GPUs, TPUs, XPUs, or whatever you want to call them. But when you process, you have to worry about whether you are utilizing it fully. So if a car's speed is 100 miles per hour but it can only go 30 miles per hour because the infrastructure can't support it, then you are underutilizing the car, or in this case, the GPU. We are the information highway for all accelerator users and mission-critical workloads.
So let's say I'm sitting in my office. I put a prompt into ChatGPT. How do I use your products? How does this service work?
Yes, we mostly work behind the scenes. Like my parents always used to ask, 'What do you do?' because they couldn't see what we actually do. When you type an Arista prompt, at that point we are basically connecting everything to everything. This conversation on this podcast is also not possible without it. So high-intensity traffic patterns, whether from a user, device, workload, machine, server, storage, or increasingly AI killer applications, we connect them. And this happens through a set of switch and routing platforms. We call this the leaf-spine architecture, where multiple leaves connect to an aggregate spine, and nowadays it can also be an AI spine. These are some of the world's largest high-scale platforms. But to do all this, they need power. The platforms and chips need to be powered with the right software. And this is where Arista's specialty, its extensible operating system, EOS, comes in. We power these platforms and networking connectivity with all kinds of data. We can use structured data, unstructured data, flow data, and bring all of them to work with our specially built software.
And how is AI networking different from the systems before? If we go back a bit and look at networking from about 20 years ago, the internet started from the research community. At that time, protocols and algorithms like TCP/IP came from research projects, especially ARPANET. The modern internet is built on this foundation, where research took the form of practical networking systems. At that time, there were many protocols like XNS, AppleTalk, OSI, and IP. But in the 2000s, TCP/IP became the common language of internet communication worldwide. By 2015 to 2020, just TCP/IP was not enough because networking needs became more complex. In this era, people started building the world's largest cloud platforms, like AWS, Azure, GCP, Google, and many others. They needed unprecedented scale to reach many more companies and users. So systems had to be made multi-tenant because they were supporting not just one company but thousands of companies. So a mid-sized company's ordinary CIO stopped building things and started depending on the cloud to run their mission-critical assets, whether e-commerce applications, databases, or something else. Now, as AI has come in the last few years, it brings a new twist because cloud traffic is very different from AI traffic. AI traffic, and we all saw this with the start of ChatGPT and other language models, is creating a complexity and search that is a thousand or million times deeper and larger than ever seen before. So this required even faster traffic patterns, lower latency, and high performance. So what we were doing before AI, we used to call the front-end network, which connected standard CPUs, storage, general-purpose queries, indexing, searching, and things like that. With the advent of AI, we suddenly started connecting hundreds, if not thousands, then millions of accelerators. And as I said, this GPU traffic required a very different kind of traffic. For example, the quality of traffic, the intensity of traffic, the durability of traffic. In terms of peak burst, it is very intensive. So we often call this the back-end network or scale-up and scale-out network. We had to create almost different kinds of parameters and metrics, and until we came along, I would say most of it was like islands of small things. There was an InfiniBand network, a PCI network, a CCI network, and Arista succeeded in bringing Ethernet and IP and standard-based capabilities to the back-end for AI, just as we had in the cloud.
What is the biggest bottleneck in building these networks now?
Yes. I would say the biggest bottleneck for the entire industry is power. Because the power consumption of GPUs, networks, cables, optics, and everything is not like before. Earlier we used to talk about megawatts. Now we are talking about 10 gigawatts in a data center. Look at it from the right perspective. It's bigger than a football field and you are powering it with gigawatt capacity. So our AI providers or cloud providers are not finding that level of power. If you are looking for that kind of power today, it can take you three to five years to find it. Space and power are probably still the biggest problem. But once you find it, you need to prepare it properly, and that's where we come in.
Exactly, and so if this is found, you will have to call you guys. But right now, a lot of money is coming into AI infrastructure. How do you see it? Is this development happening faster than necessary?
Yes, that's true. This change has happened very quickly. If you look at the old cycle of networking, it used to take about 5 years to go from 100 megabits to 1 gigabit, then to 10 gigabits and 100 gigabits. But now, the migration from 100 gigabits to 200, 400, 1.2, and further to 3.2 terabits is happening in about 12 months each time. So the changes that used to happen in 5 years are now happening in 12 to 18 months. This means that where we just reached 400 gig last year, this year we are already thinking about 800 gig and 1.6. There is a huge demand for speed and scale because you have to aggregate all this throughput and it's all coming at us together for predictable latency.
You were at Cisco during the internet boom, right?
How can you compare what you saw then to this?
The internet boom happened in a very, very steady way. There was a time when a peak came in 1999-2000, which was due to the dot-com boom and bust. I would also say it was due to Y2K. But mostly, it took 15 years for that internet to develop. I was there from the first acquisition of Crescendo in 1993 when the company was less than a billion dollars, and when I left in 2008, it was a 40 or almost 50 billion dollar company, where the Catalyst switch, which I think is still Cisco's most successful product line, was contributing at least a third. But the point was that it happened slowly and steadily. But there was a peak. There was a surge for two years. This seems very different for many reasons. The biggest reason is that the 15-year process we were talking about is now shrinking to just five to seven years. The second thing is its scale aspect is multiplying rapidly. It has become almost 1000 times more. And the third thing, perhaps the most important because people like to compare it, are we in a bubble? I think there are many responsible companies, and therefore we are not in a bubble, or if we are in a bubble, we are in a long bubble. This is not like the dot-com companies that came, got funding, and then shut down. These are responsible companies like Microsoft, Google, Meta, and Oracle that are taking the responsibility to build all this. Along with this, there are also new clouds like OpenAI and Anthropic that need to be mentioned. So I would say this looks more like a tremendous mega trend than a bubble that is going to be here for some time.
During the dot-com time, we had overbuilding. So how can we be sure that this won't happen again?
Because this is all not available. Neither so much power nor so much capacity. It will take time to prepare it. Our desire is much greater than the ability to execute it. So it will still take about 3 years of execution phase, not like 1999, when everyone rushed to create companies and thought if you build it, they will come. This time the situation is reversed. Here the thought is, please build it quickly. We need it now. That's the biggest change. But again, this doesn't mean I won't correct myself and say it's not a bubble. Rather, it's important to say that it's happening very quickly and suddenly on a large scale. But I know it's much more than the steady state we are all used to. But I don't think it's being done for a one-year factor. This is a three to five-year job and responsible companies are doing it, voting with their money because they see the demand. That's the biggest difference, that there is demand.
Speaking of responsible people, your biggest clients are hyperscalers, right? Cloud companies, Microsoft Cloud, items...
So how concentrated is the customer base?
Well, they have always been concentrated. As an example, Microsoft and Meta have each been more than 10% of our revenue since going public in 2014, and I am not sad about this concentration. I am happy and grateful for that partnership. So they remain 10% of our revenue. But obviously, as the denominator gets bigger, we are diversifying into many specialty cloud providers and enterprise campus. But if I didn't tell you that they are very, very important for Arista's strategy and the partnership has never been stronger, it would be a joke.
When it comes to building AI capabilities, which country are you most impressed by?
Which country? We are definitely the United States and we are leading. Some things are happening in other countries too. But in terms of scale and the companies we are working with, most of them are in the US today.
Are there any other countries that you think are doing the right thing?
Absolutely. I think we are seeing some of this in Nordic countries, in the Middle East. There is something similar in the UK. Some of it is in Korea and India. So there is no doubt that it is also reaching many other international locations rapidly.
When you see where the investment is happening, are you worried that this will divide the world even more?
No, most of the investments I am seeing are not dividing the world but connecting the world more. So thankfully, they are not government-sponsored. So they are all about how to enhance AI and technology and communication. A month ago I was in the Middle East and I am very impressed with the speed of their progress, where they are investing a lot of their oil earnings into necessary AI investments and capabilities. That's why I see it not as a division but as unity.
Speaking of being impressed, you now have a 21% share of the market for AI data center networking, which is tremendous. How did you get there?
Is that all we have? I expected it to be more. But then we had to work. We have traveled from zero to 21 very beautifully. I think if you look at the high-speed market, we are probably around 41. But if you look at all switching speeds in general, we are probably around that number you mentioned. Look, this is work done with love. The company was formed in 2008 and it started shipping products. So we didn't do this suddenly. We did it slowly but surely. And really, I think to optimize the culture of this company, which Andy Bechtolsheim and Ken Duda started, I have to say they laid the right foundation for our software and our hardware infrastructure and kept making it better. So our market share actually comes from the fact that our customers appreciate our innovation. We are built by engineers for engineers. We focus a lot on quality architecture and support our customers by doing the right work for them. Because we sit at the center of a mission-critical environment where if we go down, everyone knows. If we don't go down, we remain the world's biggest secret, and I have no problem with that. So I believe we have done it in a very methodical way and haven't launched any product in a hurry before it was ready.
Now you spent 15 years at Cisco building their data center business. So I just want to know what it was like to work during Cisco's peak level?
It was much more than just the data center business. I loved my time at Cisco. When I joined Cisco, the company's revenue was about 650 million and the market cap was 6 billion. So it wasn't a startup but a young company. I got an opportunity, and at that time the company was mostly making routers. During that time, I got the opportunity to contribute to building and developing the Catalyst brand of switches. First, there was the Catalyst 1000. Then 5000, 6500, 4000, 3000. So this switching brand was zero when I first came. And I still remember the first year it was 40 million. Then the second year 365 million and then the third year it became 1 billion. And things like this cannot happen without the World Wide Web and the internet, no matter how brilliant your products are. So I felt proud to be part of a rocket ship. It was a great company, but at some point I felt, wait, I didn't want to stay in such a big company. I wanted to do something smaller. But I remember my Cisco days and those 15 years very fondly.
What did you learn most from that rocket ship?
I think the same thing applies here too, which is customer focus. Cisco has a culture of customer focus. I think Arista doubles down on this and also on innovation, engineering, and quality. But I have attended calls at 2 AM on behalf of Cisco. As an executive sponsor, if something was serious, I would pick up that phone. You know, even though I value my sleep a lot, at that time it was more important. So I would say the culture of customers is very important.
Now you could probably have been the next CEO of Cisco. So why didn't you want to change the spaceship?
You know, I don't know if I would have been the right CEO for Cisco because I was an accidental executive at Cisco. I never planned to become the executive I was. I liked products. I liked working with engineers to build products for customers, and I was so immersed in technology that I didn't think I had greater capabilities or the desire to become CEO of Cisco. So actually, when I left Cisco, the truth was I thought I would do something completely different, like clean tech or energy or solar or battery. And I really wanted to come back to Silicon Valley and connect with my roots, which I liked, meaning becoming an entrepreneur. And that's why I joined Arista when its revenue was zero and it had 30 engineers. And the most important thing I liked about it was that I enjoyed working with those people.
So let's say I have a startup now. My revenue is zero. 30 people are sitting in one office. So how do I recruit a superstar like you in the midst of all that uncertainty?
Yes, no, I think Andy and I, this was a long time ago. We knew each other. So this is Andy Bechtolsheim, the German co-founder of the company. When I was at Cisco, we acquired his company, Granite. So we had worked together at that time. Besides, when he founded Sun Microsystems and I was making chips at AMD, we also worked together. Actually, we have known each other since 1987, meaning for about 35 years or even more. So I think when he saw that I was leaving my job and I didn't know where I was going, I just wanted to clear my mind, he said, 'Oh, I have this startup.' I said, 'Oh,' and my first reaction was, 'Oh no, Andy, I don't want to do networking and switching again.' But as I understood what the founders were trying to do and what the engineers were building, I got interested because it was networking and switching. But with a twist, it wasn't like the internet and Catalyst switching I knew. It was something completely different.
What was that thing that sparked your interest in it?
Its software and software architecture were very unique. What do most networking companies do? They have five operating systems for five use cases and 50 images for each, totaling 250 things to manage. But here there was only one operating system and one binary image. It had a lot of class foundation, which we called a state-driven publish-subscribe model. In this, if something fails, it doesn't just fail. It automatically recovers and the agent comes back. And this was exactly the opposite of a Christmas tree light, where if one light goes out, all the lights go off. Here we had built a system where if one light failed, we would just change that LED and the whole system would keep running. The customer wouldn't even know. Because of this software, we got the opportunity to work with so many mission-critical and high-profile customers. And then we used that technology in different kinds of use cases. That technology attracted me, and so did the people. But to be honest, there was also a little nervousness because the year I joined was 2008. It was the time of the financial crisis. One of our early customers was Lehman Brothers, and you know what happened to them after that.
I know. So my entrepreneurial career at Arista didn't start well, which has now been going on for 17 years and is strong. It's not that we haven't seen ups and downs. We absolutely have. Financial collapse was one of them. Our biggest competitor, Cisco, suing us was another. COVID and the entire supply chain crisis was another. So to get to where we are today, we have faced many difficulties, and it's good to be here.
What does it take to compete with a big company like Cisco? You are a startup and you think we have to beat them. So what do you think is needed for this?
I don't think most of my colleagues would tell you that I ever felt I was competing with Cisco because in the beginning we were so focused on very specific and unique use cases that Cisco didn't even want to be involved in or pay attention to. Like high-frequency trading and low latency. These were use cases that Cisco didn't focus on. And our first 100 million in revenue came mostly from there. Our second use case was cloud, and Cisco called it MSDCs, or massively scaled data centers. But they didn't pay attention to that sector. They didn't even try to understand what customers want and why they do what they do. Actually, in the early days of AWS, Amazon, and Google, they didn't use Cisco or Arista because they built their own systems. They were very dissatisfied with their requirements and the solutions available in the market. So once again, I felt we were doing something that large corporations, especially Cisco at that time, were not doing.
