How Jeetu Patel Runs Cisco Like the World’s Largest Startup

My strong assumption is, moving forward, that we will actually need more engineers, not less. The goal is not efficiency. The goal is to be able to go out and make sure that we benefit from the original insights that AI generates that allow us to solve problems we could have never dreamt of solving before. Do you know how many people in their 50s, how much money they would pay to get back to being 19 if they could? This is a magical time to be alive in tech.

Thanks for having me. It's great to see you.

The way I think about my role is, I just make sure that we help build the best products in the world that all of our customers can use at scale. What Cisco does though, which is even more interesting than what I do, is we think about us as the picks and shovels company during the gold rush. We are the AI infrastructure company for the AI era during this massive movement that's going on right now, which is a secular shift. And we want to make sure that we provide the right level of infrastructure so that AI just runs the way that you want it to, and it runs securely and safely.

You know, I think in these massive kind of shifts that happen in companies, typically the leaders get an undeserved level of credit and the people that do the work don't get enough of the credit. But the reality is, I didn't write a single line of code when we were doing this. It is the team that did that. But what we did was we made sure that we actually set the right direction, hired the right people in the right jobs, made sure that we had the right kind of mechanisms to figure out what markets we're going to go pursue, and do it with everyone feeling like they can win.

So I ended up at Cisco. I was at Box. I was running product for Aaron Levie, who is a founder himself. And so my entire kind of mental model got reinforced again when I came to Box because I had started my own company years ago. And then I went to EMC. EMC was a very large company. I'm like, ah, this is like very corporate. And I went to Box. I'm like, okay, this can be fun because you can actually operate just like I used to operate in my own company. And I think there's two kinds of people. There's folks that are very good functional leaders and they love being in their lane. No one tell me what to do in my lane. I'm not going to tell you what to do in your lane. And then there's people who've run some kind of business and they think of themselves as owners of the business and they are completely comfortable poking in other people's businesses all the time. And I tend to be of that ilk. And so I personally feel like when you have this founder mode approach, it's very good for wartime and it's very good for a time when the market's going through a shift, things are moving at a very fast clock speed. And what you have to do is you just have to make sure that you don't completely alienate the people who might not be in founder mode and kind of help them understand the benefit that if you do go in founder mode, we can actually operate at a very different frequency than what we've been operating so far. And so that's what we started doing at Cisco. And I feel like I tell my peers and everyone at my company, I feel like I'm a late-stage co-founder of Cisco. I was 36 years late coming in, but I feel like I own the company and this is my company. And if anyone from the outside tries to screw with the company, beware because I'm going to make sure I fight for it. And we got to make sure that we take a lot of pride in the craftsmanship and what the company stands for. Everything is your job. And so if something's not working well in sales, I'll go talk to the people over there. Something's not working well in marketing, I'll go talk to the people over there. And I fully expect them to come poke holes in my business.

You know, I also recognize and appreciate the value of seasoned operators who are systems thinkers, because if you had everyone at Cisco who was a founder, it'd be a show. Yeah. So you have to make sure that you get a good balance. The way we've done it is we have, I try to call it a rule of thirds. So a third of my leadership team are people that know how to maneuver Cisco, that are at-scale operators. The second third are people that I bring in from the outside, from competition, from the market, that are very market-in, that might not know how to maneuver Cisco, but they're systems thinkers and they've operated at scale. And then the third third are co-founders and CEOs of companies that we've acquired, but those people we give them charters that are larger than the company that they've sold us.

I feel like Cisco is in wartime right now. In the sense that we have made one of the most successful turnarounds that you've seen from a product perspective. I think we are now squarely in the AI infrastructure game, which is something that no one would give us credit for three or four years ago. But if you look forward the next three, four, five, seven years, the amount of opportunity that exists in front of us is so massive compared to anything that we've experienced before that we have to make sure that we can capitalize on all of that. And that means that we have to operate at, I like to tell my team internally, we should operate like the world's largest startup. That means you have to operate at speed but with scale. If you just operate at speed, then you're just a startup. If you just operate at scale, then you're just the largest company. If you want to operate as a world's largest startup, it has to be with speed and scale.

Let's assume right now that we have sub-2% of humans that are using agents in any kind of power-user capacity, and think about the fact that the entirety of infrastructure that gets produced today is getting consumed. We are short on infrastructure. We're supply-constrained with only sub-2% of the people using agents. Imagine when 25, 30, 40, 50% of the people start using agents. The amount of infrastructure you're going to need is going to be at a very different scale proportion than what we have right now. And we want to make sure that we can help solve that problem. That's problem number one. Problem number two is these agents, in fact, this week was a very interesting week, because these agents are going to go rogue at times. They might actually have a mind of their own. It's actually harder and harder to distinguish between when an agent is operating with malicious intent because of an adversary influencing the agent or poisoning the agent, or an agent just thinks it's the right thing to do and they're just kind of, you know, diligently going out and following your goal command.

Exactly. Yeah. And so what we have to have is some kind of safety, security, and observability capabilities in place. And so that's the second big problem that we want to solve, is we want to make sure that AI is safe to use for everyone. And that doesn't become this constant fear because of which people don't use AI. In fact, I think it's the other way around. If you don't use AI, you will be less safe. And that is true because most of the attacks are happening at machine scale. So you have to have the defenses at machine scale as well. So we're one of the largest networking companies in the world that provide the infrastructure. We are also one of the largest security companies in the world. We want to make sure that we do that. Then we are also one of the most significant machine data companies in the world with Splunk. And so the combination of those three should give us all of the underlying apparatus that's needed so that we can become critical infrastructure for the critical infrastructure companies for the AI era.

It can't be around efficiency. The goal is not efficiency. The goal is to be able to go out and make sure that we benefit from the original insights that AI generates that allow us to solve problems we could have never dreamt of solving before, which is very different from just getting to be 20% faster at doing this or making sure that you have less people. In fact, my strong assumption is, moving forward, that we will actually need more engineers, not less, as we move forward, because every single time I've seen a step-function improvement in AI, I've seen a bottleneck that's emerged that's a human bottleneck. You know, coding is automated, code review becomes a bottleneck. Code review is automated, judgment on what to build and what not to build becomes a bottleneck. So I do feel like we're going to need to have more people. But what was interesting was getting 32,000 engineers and the rest of the company, one, to taste the honey, and two, not be fearful of the fact that if you use this, then you're going to actually be creating a trap for yourself or you might lose your job. Instead, what we did was we did something completely opposite. We guaranteed our employees that if you don't use AI, you will lose your job. And so it was a very different way to kind of approach the problem. That wasn't a very controversial statement to make because think about today, if you had someone that didn't know how to use a PC or didn't know how to use the internet, would you ever hire them in a knowledge-worker job? You would never. So it's not that controversial to say that if you are seeing a secular shift and if there's this massive platform that's going to get refactored with AI, then anyone not using it is just not going to be that relevant. And so we always believe that there's going to be only two kinds of companies and two kinds of people: ones that are very fluent and dexterous with AI, and the others that are going to be really struggling for relevance. And so we just made it very, very clear for people that if you happen to be in that second camp because you're not using AI, chances are that Cisco is the wrong place for you. And therefore what you should do is go all-in. And we will give you all the tooling and all the support to go all-in. And initially we actually gave everyone unlimited tokens to just start using it because the way I think about it is you have to get, it's like riding a bicycle. You have to first get familiar with something, then you have to get good at it, and then you can become efficient with it. Yeah. But if you try to go efficiency first, you'll never get familiar or you'll never get good. And so we started doing that, got a lot of initial resistance from a lot of folks. The first few months were rough. And the seminal moment for me was I have this one engineer who's an engineering leader in my team, one of the most amazing people. And he is always a skeptic. He's always going to tell me, 'G2, those ideas are crazy. I don't think it's going to ever happen.' And he called me one day in the evening and he was as sincere as he could be. And he said to me, 'Hey, I just want to let you know I was wrong. You were right. The models are moving at a pace now.' And I had some benefit because, you know, Kevin Weil is on our board and I had some, you know, kind of, I knew some people within the ecosystem that every time I talked to someone in a startup world, I was getting a very different set of data than I was getting internally. And so Will says to me finally, 'This has now moved where we have now taken and refactored like, you know, half a million lines of code and we were able to refactor it into like 120,000 lines of code or something.' And so 5x improvement and we were able to do that within like two weeks. You know, and so one, you are right that this is actually going to move really fast. So I can't just think about what we have today. I have to think about the fact of where is this going to be in three months from now, not a year from now, three months from now, and assume that and start making the changes today. But second is I'm also feeling very anxious about the fact that this might mean a lot of bad things for people and we might have to make a lot of changes in our staff. And I'm like, or we can actually motivate everyone to make sure that we can all get better as a result of the use of this. Then, you know, it just kept compounding. We actually became the first design partner with OpenAI and Codex. They gave us some FTEs and then to this date, every, I think it's every two or three weeks now, I sit down with my entire leadership team of engineering and then their FDEs and we talk about like what went right, what went wrong in the past two weeks. And it was just very clear to people that I'm just not going to let this one go. And I think that's in large companies, there's a few things where you have to go top-down. Most things it's going to be pretty hard to go top-down, but a few things that you think are religion, you have to go top-down and you have to just make sure that you're unrelenting on those. And AI was one of those.

So Cisco does software, hardware, and silicon. And each one of them have a different kind of adoption cycle. So on the hardware side, you're starting to see good kind of design patterns that are starting to emerge now. But it's not quite as much as software. Software at the clock speed is really fast. Silicon, there's a level of intentionality of not going all-in by the team because they're very concerned about making sure that the data and the IP does not leave and then before you know it, someone else has replicated it. Right. You know, and so on the silicon side we had gone a little bit slower, but now I think with the models getting better, with open-weights models being there, having our own infrastructure, all of that, I think it's going to change, but we have to compress the cycle time for silicon. The thing that's interesting in our business is we are a full stack, right? Like we build silicon, we build photonics, we build systems, hardware, software, the entire stack. Each one of the elements and layers of the stack have a different planning cycle. So silicon is a five-year planning cycle. What I do today, I have to make sure that I know what's going to happen for at-scale production five years from now and then make the appropriate investments today.

Think about the entire manufacturing cycle of not just the silicon, but silicon to systems to SDKs to taking it out to the market. You have to make sure that you start, because in order to build the silicon, you have to build the IP for the silicon, like SerDes and all those pieces, and those take, actually, you have to build them up ahead of time. So like there's a sequencing that needs to happen in silicon that cannot happen within a matter of like, you know, you can't just say go build a chip in 12 months and you can build a chip, it just won't be that good. Hardware tends to be, you know, 18 to 24-month planning cycle, maybe 30 months in some cases because of supply chain, because we think about it, we're doing this at scale for very large volumes of units. You go to software and operating system. That is a 12-month planning cycle, maybe 18-month planning cycle. Then you go into models and agents and apps. Now you're into a three-month planning cycle, maybe a weekly planning cycle. And we have to make sure that that co-design full stack actually has all of those things working in harmony. And if you can figure that out as a company, you have a very strategic moat in the sense that you've been able to figure out all these different layers that work well together from an IP. That's very hard for other people to do and replicate.

Absolutely. All of them will get compressed. But when I say planning cycle, I don't just mean how long does it take to build it. I mean from idea to at-scale production with billions of dollars in revenue.

So in data center buildouts, by the way, everyone knows the role the GPU plays. And what had happened was when you just started seeing these transformers come about, these transformers would fit on a single GPU, and so the training runs would be pretty simple. And then you had more parameters, and those GPUs needed to scale, and so you had eight GPUs within a server, and so you now went into a server, but those GPUs were all interconnected within a server. Then you had racks of servers, GPUs got interconnected with a rack of servers working coherently like one, you know, kind of giant machine with things going. And then racks started scaling to rows because you now had multiple rows of racks that needed to be worked together within a large data center. And that required what they call scale-out networking, not just scale-up networking for rack-level scaling. And now what you have is because there's such a huge power constraint and power shortage, you can't always bring all the power to where the data center is. So what you do is you go to wherever the power is available. And so you might have data centers that are hundreds of kilometers apart that need to then work like one completely logical unit.

As if they were in the same building, as if they were the same computer.

That requires a whole different family of networking called scale-across networking. For within a data center, you might have scale-out networking, and for within a rack, you have scale-up networking. We are largely focused on scale-out and scale-across, and we do not just the silicon and the switch trays for connecting the GPUs together, but we also do the photonics, which is the optics. We started our business selling to hyperscalers two years ago for AI workloads, and we thought we'd give the market this tree of guidance and said we'll do a billion dollars in orders taken.

So the next year we said, well, let's double it. Instead of 2.3, let's do like 5 billion. Last year, which was a Q4 that just ended a month ago, we ended up at 9 billion.

Yeah. And so we basically went from zero to 9.3 billion in a matter of two years.

Right. And so that's the speed at which this is moving. We still have so much share to go out and capture. It's a complicated business. $63 billion in revenue, hundreds of products that we have to kind of make sure that we maneuver in the right way.

But you're right. It's starting to get to be meaningful and it's exciting because you're right in the thick of things and the market is so huge. A gigawatt of capacity would cost about anywhere between, it used to be 35 billion, now it's like close to $50 billion is what a gigawatt of data center capacity would cost to build. And about 10 to 15% of that tends to be networking. Largely it's in the GPU, but there's a huge amount of complexity in, you know, you have to build out the shell and you have to build out the power permits for that. And you walk into a data center, it impresses upon you on how complicated, when you make a query on your prompt to say, give me a pizza recipe, what happens behind the scenes to get that to you within a second or two.