Exclusive: Cisco President Jeetu Patel On AI, Jobs, Data Centres & India's Future

Welcome to Cisco.

There's a lot of seats at this table.

Actually, we had our silicon meeting here with our entire silicon leadership team and every seat was filled. But it was a fun, fun day. We spent two days over here.

Ever expanding.

So, as you and I have talked about before, we are moving from this world of chatbots to agents. These agents tend to be far more consumptive in their network consumption patterns than humans are. So, for example, we just did a study: an agent consumed about 450% more bandwidth than a human for the exact same task.

They're not sleeping, they're working around the clock. And by the way, what is an agent? An agent is basically a file that actually describes the skills, a file that describes its memory. Actually, you have to keep making sure that those files get uploaded to the model. So, there's a fair amount of kind of network bandwidth because of the chattiness of the system. So, what we have here is an infrastructure that's completely going to need to be reimagined to accommodate agents. And we're not even at the physical AI realm yet. The moment physical AI takes off in a big way, you'll see an even more consumptive pattern on that front. So, the reason it's a networking supercycle is because you're going to see this not just in the data center. Now, what you see is these agents are also running on your desk side. In fact, there's a whole new class of computing called desk side computing. I don't know if it's actually been something you've talked about during your trip over here, but think about an agent not just running in the cloud.

You now have people saying, 'Hey, I have a laptop for the human to work on, and then I have like a Mac mini or a Mac Studio or an Nvidia PC or an AMD PC, and that's what they call a desk side computer.' And that's where the agents run. And you might have hundreds of agents running over there and conducting tasks for you. And that's going to have an enormous amount of bandwidth consumption because for every human, you might have 10, maybe 100, maybe 1,000 agents. And they're working 7 by 24. They're 450% more consumptive in their infrastructure needs. So, think about what that does to the network, to the compute, to the power. It's going to be enormous.

I'll give you some numbers that'll actually just be completely eye-opening for you, which is if you think about our campus branch networking, which is what we call a workplace network. In offices like this, when you have Wi-Fi access points, you've got switches, you've got routers. That business for us grew at about — it was pegged to inflation — it grew about 3 to 4% a year. Last quarter, we had 25% order growth. So, imagine that. And that's not something that's a blip for a quarter. I think we can see sustained level of double-digit growth for a while. And so, that's because of this kind of paradigm shift that's going on in the architecture of the networks itself. So, when you look at the numbers from the financial standpoint, there's a fair amount that's going to be spent on these networks because everyone wants to modernize their workplace.

When you look at it from agents and what they're doing, every human is going to be a manager of agents. And we now have at Cisco our first product that is completely written with AI. No human lines of code. At all. We humans will verify the code, but the agents wrote the code, right? By the end of the year, we'll have about six products that are completely written with AI. By the end of next year, get this, we will have 70% of our products that will be 100% written with AI. And in two years, Cisco will have no legacy code. We will have refactored all of our code.

So, when you start thinking about that level of implication of agents on the workplace, there's going to be a tremendous amount of changes in the campus. Of course, we've been working with data centers, as you know, on the numbers. Three years ago, there was barely any revenue with AI-related data center revenue. Last year, Chuck Robbins, our CEO, said, 'Hey, let's make sure that we actually let the market know what we think we're going to be able to do.' So, we put out a number. We said, 'We'll do a billion dollars in year one in AI revenue for data centers.' We ended up at 2.3. So, we said, 'Next year, what do we do?' Which is this year, this fiscal year for us. We said, 'Let's do 5 billion. Let's double it.' We're going to end up at nine. And so, there's this kind of massive exponential curve of increase in infrastructure.

And then you add on top of that things like security because of Mythos and what's happening over there, and you see different models that are going to be there from OpenAI and from Anthropic that are going to be very, very potent. I think there's going to be a tremendous amount of emphasis placed with every class of customer we have, whether it be a hyperscaler, whether it be a neo cloud, a sovereign cloud, a service provider, or the enterprise, in completely changing their architecture to go out and accommodate for these new patterns of work, the new ways that security is going to be implemented, the new kind of exploits that the agents are going to be participants in and drivers of sometimes, and they'll also need to be participants of the defense against the agents at scale.

There will be no such thing as legacy companies anymore. That's a thing of the past. Because...

So, what does a company... By the way, this is an interesting thing because legacy used to be where I, as a company, have actually enjoyed so much success. And products have been used so much by so many people that I need to now make sure that I slow down my innovation so that I don't go — because I've accumulated tech debt over time. And so large companies had a harder time in moving as fast and with the level of agility that a startup would.

Now, what's happened is we're starting to see that when you actually do write this code and you do have success in the market, you can continue to keep refactoring the code. So, for example, in our security business, we had our firewall — I'll give you a couple of examples here. This is just mind-boggling. So, we have our firewall management console. And what we did was we said, 'Let's refactor the code here.' 8 million lines of code, within six weeks the teams were able to refactor it to less than a million, million and a half lines of code. So, firstly, you've compacted the code. What does that mean? Performance is faster. It's more efficient. It's actually got less security vulnerabilities. The way in which we've designed the product can be updated faster. All of these things have a profound effect on the customer.

And now all of a sudden, our scale of being with all those customers, but being able to go out and innovate at a very, very fast speed becomes our very asset. And so I feel like as time goes on, you're going to have very little legacy code because everyone's going to write code using agents. And what you're going to then do is — we have about 32,000 engineers right now. I think in two to three years, they're not going to be writing code. So then the question is, what are you going to do? Well, they're going to review the code, and in a world where you can write anything, what you choose not to write and what you choose not to build is actually far more important. So the product judgment and the product intuition is going to be really important.

The other thing that these engineers will do is they will actually be the front lines for our customers because they will help our customers adopt better. So every engineer will become a forward deployed engineer at some point for a certain percentage of their time. And so you'll see the configuration of the entire workforce change where a product manager is going to actually become a product developer of some sort, and the first prototype will actually be built by the product manager, and the design of the engineer is not even involved. The engineer gets involved in the architecture. And then the engineer is involved in the front lines with the customers. And so I think you'll actually start to see a workforce shift which is very different from what we've been used to. But I don't think it's going to be around this notion that there's going to be fewer jobs.

What I've started to see and observe is every single time we've seen a massive step function improvement in AI, actually the bottlenecks have been largely human. So when we have much, much higher volumes of code, now my bottleneck is code review. So I say, 'Okay, I'm going to automate code review.' My bottleneck then is going out and building a test harness. So I automate that. My bottleneck is going to be what kind of products do I build, and do I need to do the right kind of market research, and do I need to do deep thinking and apply intuition to it. So I think there's going to be a snowball effect, and I think you're going to see a need for more people.

I think over time you will have a shortage of people that are dextrous with AI and have fluency in AI. Now, by the way, the gap between someone who is really fluent in AI and someone who is not fluent in AI is not going to be like a 10% productivity gap. I think it's going to be 50x or 100x. What does that mean? That means it's going to be very hard to justify someone who's not AI fluent to be in a job. It would be like saying, 'Let's hire someone that doesn't know how to use the browser or the internet.' It would be an absurd concept in knowledge work. So, I think that'll be an area that we'll have to make sure that collectively, between public-private partnership, we will have to make sure that we raise the dexterity level of people with AI.

But if you look at every single person that we hire today that is fluent with AI, they get paid really well. In fact, I would say, if I was in my 20s getting paid that kind of money, I'd be in trouble. So, they get paid really well. They have a tremendous amount of shortage of those people. And the value that they add is enormously high compared to what they're getting paid. And I think you're going to see more and more of the workforce get there. So, there's going to be more of a demand, but there will be this painful period where the retraining of the workforce will require some displacement, and there will be some people that will lose their jobs. I think every job is going to get refactored, and there'll be entirely new industries created.

So, that's actually a multi-layered question. Let's parse it out. The first question that gets asked is, 'Are we in a bubble? Are we building too much?' The second question is, is this notion of tokens that we're consuming, which is actually skyrocketing in cost, commensurate to the value that's being generated, and is the cost going to keep going up with these agents, and is it predictable in nature? So let's talk about the first one. I think right now there's about 0.1% of humanity that's using agents. These agents are very consumptive. The more humans that use agents, the more agents are going to use infrastructure. I don't think we're anywhere near the peak point of infrastructure build-out. You're going to have trillions of dollars being spent on infrastructure build-out. So where are we in the cycle on the infrastructure build-out? We're in the first inning. You know, we're very, very early in the cycle. There'll be gigawatts and gigawatts. In fact, Elon talks about terawatts of data center build-out that'll happen. So I think that's going to continue.

But the question that you have to ask then is how is the token consumption and how are the economics working on the token consumption? Right now, I think there's a gap between the cost of the token and the value that that token generates for most companies. And I think it's important to understand why that is. The way that AI is being used, the way you have to think about it is it's like riding a bicycle. The first time you use it, you're not that good at it. So you have to get familiar with it first. That's step number one. Then you have to get good at it. That's step number two. And then you can get efficient with it. Majority of the world is in the familiar state, you know, getting familiar with it. There's some that have gotten good at it. Very few have gotten efficient with it. But the problem is in the familiar and good state, that's very consumptive on tokens.

Blowing up budgets. And so I think what'll end up happening is you'll see some thrashiness right now where people are going to say, 'I'm going to stop.' There's going to be some fits and starts. It's already happening. And that's inevitable because you're going to need to make sure that you plan out these cycles effectively, and not everywhere do you need to use a frontier model with a trillion parameters or 10 trillion parameters. You might not need something that heavy for basic usage. And so what you will see also is this kind of intelligent routing layer that says, 'I'm smart enough to know, Shereen, that when you ask a basic question, I'm going to route this to a model that's going to be much less consumptive on infrastructure. And when you need something that actually has all the context of the world, maybe I need to route that to a frontier model.' And that percentage balance is something that I'm going to think about.

And then what you'll also have is this entire observability stack and tokenomics. In fact, we just announced that for our customers where we have a full stack that says, 'Hey, how's your infrastructure utilization? Is the GPU being utilized effectively?' That's step number one, layer number one. Layer number two, are the agents behaving the way you want them to behave? And if not, can you appropriately assign guardrails on it? And then step number three is, what are the tokenomics looking like? Is this agent that's gone rogue, that's consuming a lot of tokens all of a sudden, going to give you a bill of $24,000 that you didn't want? And you got to make sure you stop that.

And then I think customers are going to need to get a little bit more predictive as well. Because if you think about tokens today, and let's say that every employee at Cisco, 90,000 people, we use $200 a week. Let's say everyone takes a couple weeks off where they don't have the agents do any work. That's 50 weeks a year times $200. At 90,000 employees, that's $900 million. Right? $900 million that no one had budgeted for. So you will have to have some mechanisms of controls that you put around the economics of it. But I think it's going to be inevitable that there's going to be more token usage, there's going to be more demand for agents, there's going to be more demand for infrastructure. And that's not going to be something that just stops, in my opinion. I think it's going to happen, but there's also going to be more demand for humans. But you will start seeing the growth in the economy and the value being created by this way outstrip the cost in the long run. In the short run, we have to create that state of equilibrium, which we aren't at yet.

I think firstly, the curve of innovation is exponential now because of this reinforced self-learning loop that we've created. So, because of that, anything that seemed like it was years away, you will actually see it be compressed within a matter of weeks, months, and quarters, right? And so, one thing I have been 100% wrong on in my overly ambitious state that I've been in with AI for the past few years is I have always been wrong on time, but I've been wrong on time erring on the side of conservativeness, even though it seemed like I was irrational when I was saying that something was going to happen really fast. So, I feel like it's going to happen sooner than we think and longer than we want, right? In the meanwhile, what do we do? In the meanwhile, we have to figure out efficient mechanisms within the organization to make sure that the use of tokens is more efficient and can actually create the returns that we all are expecting from it in a multitude of different ways.

Use smaller models. Use open-source models. Make sure that you're actually using intelligent routing. Ensure that you can have more effective queries. The way that the models process things, the way that the bandwidth is consumed, all of these things will have to get more efficient over time. And I think the model companies, I think the infrastructure companies are doing all that work right now. And even just a year ago compared to now, the efficiency in the infrastructure and the cost of tokens is tremendously high. I think over the course of the past 18 months, if you look back, I don't know the exact data, I think the cost of tokens is down 1,000x. The problem is the demand for tokens might be up 100,000x. And the value has not caught up yet because a lot of people were in the early phase. But they're not going to stay in the gaining familiarity phase forever. They're going to get to good and efficient very quickly. And as they get there, you will actually start to see the equilibrium state being reached.