Hi everyone, I'm Nicolai Tangen, the CEO of the Norwegian sovereign wealth fund, and today I'm here with somebody who is absolutely vital to AI, namely the remarkable Jayshree Ullal, CEO of Arista Networks. Jayshree has more than 40 years background in networking, so nobody is better able to talk about this than her. So Jayshree, super to have you on here.
Great to be here, Nicolai. I have a bit of a cold, so apologize if I sound nasal.
No problem. Let's start with the beginning.
I don't have to sing on this podcast, right? So we should be fine.
No singing. Let's start with the beginning. What does Arista do?
That's a loaded question. Arista is building the world's most demanding mission-critical networks, which is high performance, high scale, low latency, high reliability, high availability, high automation, high telemetry. So it's a lot of demands across a lot of spectrums, and Arista has emerged, especially over the last decade, to be both the pioneer and the leader in this.
I've heard that if you think of chips as the engine, then the networking is the highway between them.
That's a good way to look at it. The chips fuel a lot, especially in this world of AI processing with GPUs, APUs, XPUs, whatever you call them. But as you process this, you have to worry about whether you are processing it to full utilization. If a car is rated at 100 miles an hour but can only go 30 because the infrastructure cannot support going farther, then you're grossly underutilizing the car, or in this case the GPU. So we are the information highway to all accelerators, users, mission-critical workloads, you name it.
So let's say I sit in my office and do a prompt on ChatGPT. How do I touch your products? What happens?
We are very much behind the scenes. This is something my parents always used to say: 'What do you do?' because you can't see what we do. When you type in a prompt, Arista basically connects everything to everything. This conversation on this podcast is not possible without it. The traffic patterns of high intensity, whether from a user, a device, a workload, a machine, a server, a storage, or now more and more the AI killer applications, we connect them through a set of switches and routing platforms. We call it leaf-spine architecture, where you can have multiple leaves that connect into an aggregate spine, and these days you can also have an AI spine. These are some of the world's largest high-scale platforms, but they need to be powered with the right software. That's where Arista's hallmark has been its extensible operating system, EOS, where we fuel these platforms and networking connectivity with more and more types of data. We can deal with structured data, unstructured data, flow data, and make that all work through our purpose-built software.
And how is networking for AI different from what came before?
Very different. If you go back even 20 years, the internet was born through TCP/IP and algorithms from the research community, but then there were multiple protocols. TCP/IP became the worldwide language of internet communication. In the 2015-2020 era, that wasn't enough. People were building some of the world's largest clouds, like AWS, Azure, GCP, and they needed unprecedented scale to touch so many more companies and users. It had to be multi-tenant. Now comes AI in the last couple of years, and that's adding another twist because cloud traffic is very different from AI traffic. AI traffic, as we saw with ChatGPT and other language models, is conducting a class of complexity and search that is a thousand or million times deeper and broader. This required even more intensity of traffic patterns. Before AI, we were doing front-end networks connecting standard CPUs, storage, general-purpose query, indexing, searching. With AI, we suddenly had to connect hundreds, thousands, or even hundreds of thousands of accelerators. This GPU traffic is much more intensive in fidelity, intensity, durability, and peak bursts. We had to build almost a different set of parameters and metrics. Until we came along, the majority of this was little islands of things: an InfiniBand network, a PCI network, a CX network. Arista was able to bring Ethernet and IP and standards-based capabilities like we had in the cloud into the back end for AI.
What's the biggest constraint now in terms of building out these networks?
The biggest constraint for the industry at large is power. The power consumption of these GPUs, the network, the cables, the optics, everything is unprecedented. In the old days, we used to talk about megawatts. Now we talk about tens of gigawatts in a data center. To put this in perspective, that's larger than a football field, and you're powering that with gigawatts of capability. Our AI providers or cloud providers are unable to find this level of power. If today you're searching for that kind of power, it could take you three to five years to find it. The space and the power are probably the biggest problem right now.
But once you do find it, you have to outfit it correctly. And that's where we come in.
Absolutely. So if you find it, you have to call us. But enormous amounts of money are flowing into AI infrastructure right now. How do you view this? Is it a bubble? Is it too fast?
It has happened very fast. When I look at the cycle of how networking progressed from 100 megabit to 1 gigabit to 10 gigabit to 100 gigabit, that process happened every five years maybe. But the migration from 100 gigabit to 200, 400, 1.6 terabit, and eventually 3.2 terabit is literally all happening within 12 months each. So what used to be five-year cycles has become 12- to 18-month cycles. We just finished 400 gig last year, and we're already thinking of 800 gig and 1.6 terabit this year. The insatiable demand for speed, for scale, and for predictable latencies is all hitting us at once.
You were at Cisco during the internet boom. How does this compare?
Much greater. The internet boom happened in a very steady fashion. There was a peak in 1999-2000 due to the dot-com boom and bust, and Y2K as well. But for the most part, that internet took 15 years to develop. I was there from the very first acquisition in 1993 of Crescendo, when the company was less than a billion dollars, to when I left in 2008. It was a $40 or close to $50 billion company, and the Catalyst switches, which I think are still Cisco's most successful product line, were contributing at least a third. The point is it happened slowly and steadily, but there was a bubble for two years. This one feels very different. First, that 15 years is collapsing into five to seven years. Second, the scale is exploding in magnitude, a thousand times more. Third, and perhaps most important, people like to ask if we are in a bubble. I think it's a lot of responsible companies, so we're not in a bubble, or if we are, it's a prolonged bubble. It's not startups that came in, got funded, and collapsed. These are responsible companies like Microsoft, Google, Meta, Oracle, not to mention the neoclouds like OpenAI and Anthropic. So more than a bubble, it feels like an explosive mega-trend that is here to stay for some time.
We did have overbuilding at the time of the dot-com boom. How can we be confident that it doesn't happen again?
Because it's not available. The power and the capacity are not. Nobody can really overbuild. It's going to take time to build it. The desires are greater than our ability to execute. So I think it'll still be a three-year execution rather than what happened in 1999, where everybody hurried up and said 'build it and they will come.' Here, it's not 'build it and they will come.' It's 'please build it, we need it now.' That's the biggest change. But I wouldn't correct myself and say it isn't a bubble. It is happening in a furious, sudden state, greater than the steady state we've been used to. But I don't think it's being done for a one-year factor. It's three to five years, and it's being done by responsible companies who are voting with their money because they see demand. That's the biggest difference: the demand is there.
Talking about responsible people, your biggest clients are the hyperscalers, right?
Correct, the cloud companies. Microsoft and Meta, for example, have been more than 10% of our revenue each since we went public in 2014. I'm not going to apologize for that concentration. I'm pleased and grateful for that partnership. They continue to be 10% of our revenue, but as the denominator gets larger, we are diversifying to specialty cloud providers, to the enterprise, to the campus. But I would be kidding if I didn't tell you they are very integral to Arista's strategy, and the partnership has never been stronger.
Which country are you the most impressed by when it comes to building AI capabilities?
Undoubtedly the United States. I think we're leading. There are pockets of things happening in other countries, but in terms of scale and the companies we're working with, the majority is in the US today.
Any other countries that are doing the right things?
Definitely. We're seeing pockets in the Nordics, the Middle East, the UK, Korea, and India. No question it's coming to many other international locations as well.
When you look at where investments are taking place, are you worried that it's going to split the world even more?
No. The majority of investments I'm seeing are not splitting the world but connecting it more. Thankfully, they're not government-sponsored. They're all about AI and technology and how to increase communication. I was just in the Middle East a month ago, and I'm impressed with the rate of progress there. They are investing a lot of their oil currency into important AI investments and capabilities. So I don't look at it as a divide; I look at it as a unity.
Talking about impressed, you now have 21% of the market for AI data center networking, which is incredible. How did you get there?
Is that all we have? I was hoping it was higher. We've climbed our way beautifully from 0 to 21. If you look at the high-speed market, we're more at 41. But at general speeds of all switching, we're probably more like the number you mentioned. It's been a labor of love. The company was founded and started shipping products in 2008. We haven't done it in a bolt of lightning; we've done it steadily and surely. To epitomize the culture that Andy Bechtolsheim and Ken Duda started, I would say they built the right foundation for our software and hardware infrastructure and kept improving it. Our market share comes from the fact that our customers appreciate our innovation. We're built by engineers for engineers. We put a maniacal focus on quality, architecture, and doing the right thing for our customers. We're sitting in the heart of a mission-critical environment where if we go down, everybody finds out. If we don't go down, we're the world's well-kept secret. That's fine with me. We've acted in a very methodical fashion and not rushed a product to market before it was ready.
You spent 15 years at Cisco building up their data center business. What was it like being at Cisco during the peak levels?
It was more than the data center business. I loved my time at Cisco. When I joined, Cisco was approximately $650 million in revenue and $6 billion in market cap. It wasn't a startup, but it was a very young company. The company was largely building routers, so I got a chance to contribute to creating and building the Catalyst brand of switches: first Catalyst 1200, then 5000, 6500, 4000, 3000. This switching brand went from zero when I first came in to $40 million the first year, $365 million the second, and a billion the third. Those things don't happen without the worldwide web and the internet taking off, no matter how great your products are. I felt privileged to be part of a rocket ship. But at some point, I felt I didn't intend to be in such a large corporation; I wanted to do something smaller. I fondly and nostalgically remember my Cisco days.
What was the most important thing you learned from being on that rocket ship?
The same thing we have here: focus on customers. Cisco has a culture of customer focus. Arista doubles down on that as well as innovation, engineering, and quality. I've taken calls on behalf of Cisco at 2:00 a.m. as an executive sponsor. If something was serious, I'd pick up the phone, even though I value my beauty sleep. The culture of customers is super important.
You could probably have been the next CEO of Cisco. Why did you want to jump spaceship?
I don't know if I would have been the right CEO for Cisco. I was an accidental executive. I didn't plan to be the executive I was. I love products. I love working with engineers to enable products for customers. I was too steeped in technology interests to think I had broader capabilities or desire to be CEO at Cisco. When I left Cisco, I thought I would do something entirely different, like clean tech, energy, solar, or battery. I wanted to get back to being an entrepreneur in Silicon Valley. So I joined Arista when it had zero revenue and 30 engineers. The most important thing I loved was working with the people I enjoyed.
So let's say I have a startup with zero revenue and 30 people. How do I recruit a superstar like you into total uncertainty?
Andy and I go back. We knew each other. Andy Bechtolsheim, the co-founder of the company, we had bought his company Granite when I was at Cisco. We worked together when he was at Sun Microsystems and I was at AMD building chips. We've known each other since 1987, 35 years or more. When he saw I was leaving and I truly didn't know where I was going, he said he had this startup. 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 was intrigued because it was networking and switching with a twist. It was not the internet and Catalyst switching like I knew it. It was a completely different ballgame.
What was that twist that made you interested?
The software architecture was so unique. Other networking companies usually have five operating systems for five use cases times 50 images, so you're managing 250. Here we had one operating system, one binary image with a lot of class foundation that we call the state-driven publish-subscribe model. If something fails, it doesn't just fail; it automatically recovers that agent. It's the opposite of a Christmas tree light where if one fails, everything goes down. Here, if one light failed, we could replace that LED and keep everything up, and the customer would never know. That software has earned us the right to be in all these mission-critical, high-profile customers. I was drawn to that technology and the people. But I would be kidding if I didn't tell you there were moments of anxiety. The year I joined was 2008, in the middle of the financial collapse. One of our earliest customers that I won was Lehman Brothers, and you know what happened to them. So I had an ominous start to my entrepreneurial career at Arista, which is now 17 years and going strong. We went through ups and downs: the financial collapse, our largest competitor Cisco suing us, COVID and the supply chain crisis. We fought a lot of adversity to be where we are.
What does it take to take on a giant like Cisco as a startup?
Most of my colleagues would tell you I never felt I was taking on Cisco. Initially, we were so focused on very specific and unique use cases that Cisco didn't want to be in or pay attention to, like high-frequency trading and low latency. That was a use case they just didn't deal with, and our first $100 million came largely from that. The second use case was the cloud. Cisco used to call it MSDC, massively scalable data centers, but they didn't really pay attention to that sector. In the early days, AWS, Amazon, and Google didn't use Cisco or Arista because they built their own, dissatisfied with what was commercially available. So we were fulfilling a white space that the large corporations, particularly Cisco, were not doing at that time.
