Conversations continue here at the CNBC TV18 studio from the India AI Impact Summit. The headliners make their way to us and joining us now is Palo Alto Networks' Nikesh Arora. Nikesh, always great to see you. Thanks very much for joining us here in the studio. It's all about AI. It's all about agentic AI and what have you. What's the big takeaway from the world that you come from to the world that you're in today?
Well, first of all, nice to see you again, Shereen, and thank you for having me. You know, a few weeks ago, we had been talking about LLMs and how they can write letters and create videos, and now it's all about agents. Now, LLMs have become passe. Now you got to talk about agents. Agents are going to take over my job, your job, everybody's job. I was talking to Sam, I said, you know, how long you going to do this? He says, until the agentic AI CEO takes over. I said, no stops in between. And what about the CFO? What about somebody else? No, we're going straight to replace the CEO with agents. So, I think there's a little bit of marketing hype around the idea of agents. I think they're going to take a little while. I mean, I've said this before, but the best example I have found of an agent who is being given actual agency is a self-driving car because there we actually trust the car to drive us and we've eliminated the human. Now, that took a lot of money to build, a lot of years. I was at Google in 2009 when the first prototype came out. It's 2026. So, it's taken us anywhere from 15 to 17 years to build a complex, fully sophisticated, fully sort of delegated agent. I can imagine that we will be building agents much faster than that. But the fact that they're going to show up tomorrow and take my jobs, I'm still not a believer.
Well, you know, I think that will certainly bring down the anxiety level for the common folks that are watching this program because otherwise the headlines are screaming disaster and elimination. But let's talk a little bit about Palo Alto Networks. Just come off your earnings. Good set of numbers there in line with expectations. And plenty of acquisitions that you've closed in on as well.
That's right. Well, look, you think about it. If you believe that 600 plus billion dollars are being spent in capex, capex building the agentic future, we have to build the agentic security future as well, right? We can't just wait because security actually very often has to be ahead of technology because we need to make sure that the technology gets developed in a secure fashion. So, we have to make sure we get ready for this future where people are believing that agents are going to be aplenty. We're going to be consuming way more data than today because we need more data to train the models, train our systems, get stuff to work. If that is true, then you need real-time observability, which is what we bought a company called Chronosphere for. You need identity. Every agent must have an identity, otherwise how are you going to identify it, how are you going to control it, how are you going to govern it. So, we bought an identity company called CyberArk. So, yes, we are preparing for the gigantic future. We announced a third acquisition at our earnings which is Protect AI, which is a technology that allows you to control coding agents so that you can make sure that coding agents are not creating malicious software. So yes, we are doing a bunch of acquisitions. This is what I did 8 years ago when I started at Palo Alto to build the future of Palo Alto at that point in time and it worked out and now we're trying to make sure you build the future of Palo Alto going forward which is a future where we secure AI and secure agents.
But you know from your vantage point as you talk to enterprises today, as you talk to CEOs today, how prepared are we for this agentic future? I just had Dario here in the studio sitting exactly where you were and you know, he's like way out there. We expect 80% of the jobs are going to be gone.
Exactly. That's right. And of course, you know, to his credit, he's also been fairly open about the fact that these are sort of entities that are entirely new. We don't know exactly how this is all going to play out in that situation. What are CEOs telling you? How are they preparing for this? Look, we're trying our own experiments as you can imagine like every company is. Yes, I understand that a certain number of jobs are going to go away, but those are going to be the jobs which are repetitive jobs which should have been automated a long time ago. But we're going to use AI as a mechanism to automate a series of them. Does that impact an organization? Sure. 5, 10, 15%. But I think that job loss will be compensated by the need for more people who understand how to develop AI, deploy AI, train models. I mean, think about it. For cars to work on the streets, there are companies that spend years labeling every tree, every street corner, understanding to make sure that the AI understands what physical objects are in play. So if you believe that logic, we're going to need a lot more people to deliver this agentic AI future to train all these models, train all these robots that we have to. So I'm not worried that we're going to lose jobs. Yes, the nature of what people must do will change.
Yes. And that's been consistent every revolution, right? Every big inflection we've had in the history of mankind, whether it's the industrial revolution or agricultural revolution, every time tasks change, somehow we managed that transition because we had enough time. I think the current feeling is we don't have enough time. But I don't believe that. I think given what I just said about how long it takes to replace a driver in my car, it's going to take a reasonable amount of time to take complex jobs and get them replaced by AI. And that is a time we will use to retrain and reskill people.
You know, you've been part of many transitions, Nikesh. And you're calling me old. You said that. I didn't. But the point is every time you have this conversation with CEOs, they'll tell you this time it's different. What is different this time?
The pace. But that is also not different because if you think about it, you know, since the internet came about, every next wave was a little faster than the last one. The internet came about, then video came about with YouTube, etc. Mobility came about with the iPhone, every time you got a new inflection, even within technology curves, the adoption curve was faster. And right now people are flabbergasted. The adoption curve of AI is under three years. I remember 3 years ago when I came to India to go speak at BHU and ChatGPT had just come out and I was just surprised it was actually making recommendations about talking about things. So from there we've come a long way already in three years and there's billions of people using it. This was not what happened with the internet. So I think the adoption curve has people under the impression this is going to move fast. But I think the adoption curve is there where we're willing to tolerate a lower level of accuracy. If I write a letter to someone it's not perfectly accurate, I'm okay. If I make a video it's not, I'm okay. But am I okay if a control system is managed by AI and it makes a mistake? That's a problem. So I think getting the precision right is going to take longer and we are under the impression that precision is going to happen faster. It's not.
So this belief that you can throw out existing IT systems and replace them with an agent.
Look, I think the belief is that a lot of the work that was done by humans in analyzing data, in trying to figure out how to build more modules and existing software, yes, that's going to go away because I can throw the data into a large data lake and run a large language model against it. It'll give me the insights much faster than humans can program all kinds of dashboards and systems. I think what you're seeing in the software industry is the fear that that activity is going to be taken over by AI, and rightfully so. But you know it still takes work. Software is not just a system, right? Software is an articulation of a workflow which is coded by human beings saying this is how I do it. Hence software should echo me now. Yes, you can rewrite it, of course, but you have to rewrite it with AI. It's going to take a little while and it's going to take a while to go replace it everywhere in every company, right? Because I have to go not just change my software, I have to change the way I work.
You know, you just talked about that $600 billion plus capex that we are seeing companies commit themselves to. And I was talking to Altman and he said, you know, he believes that we're in that moment where we will say, man, we should have invested more. We haven't invested enough. And I know there's divergent views on this, but from your perspective, where are we? Are we overspending? You know, how far out before we actually see any kind of return on investment? And is there a pullback now or at least a start of a pullback in the valley on the kind of spends that we're seeing?
I think I'm with Sam on this one. I think we need a lot more capacity than we have. Let's take the analogy. If I had told you many years ago, let's build fiber as fast as we can. Let's do 5G before 4G and 3G. You'd have said, are you crazy? Yeah, we haven't got any adoption in 3G. Guess what? Today, everybody's consuming video while driving around the street and we still want more bandwidth because I'm not happy with my bandwidth. Are you?
In the back of a car in New Delhi, I'm not.
But that's taken billions and billions of dollars of infrastructure. We're still unhappy with bandwidth because we found applications that consume the bandwidth much faster than we can build bandwidth. If you take that analogy, right, I remember getting an email, I couldn't open a video 15 years ago and today I'm streaming video on a plane. So we have found a way that applications consume these things. So I think AI is going to have a similar fate where there are way more applications. I want lower latency response from AI. I think it still takes longer to produce a report on something that I like sooner. It takes five minutes to produce a 70-page report on endpoint security. When I ask it, I want it in one minute. I want it in 5 seconds. What does that need? That needs more compute. I want it to make the entire video. I want it to give it a prompt and have it make me a five minute video. It still requires a lot more compute to get there. So, I think there are applications where there are expectations. I could do with it faster. I could do with it better. All that requires more compute. And I still believe 80 to 90% of the compute will be consumed by consumers.
Which is where people are busy doing whatever they do on social media making videos. That activity will be AI enabled. People will use AI. AI will be your therapist. It'll be a nutritionist. It'll be everyone. It'll be your friend. It'll be your partner when you want to be when you're lonely. So that's a lot of work it needs to do. It needs a lot of compute.
Well, speaking of a lot of compute and you know I'm glad you brought up the telecom example there because I think the larger question that's being asked today is also one of monetization. I mean you know in a country like India where we've actually dropped rates to the lowest cost in the world the question is of the average revenue per user and how do you move the needle on that and that's the question that's being asked even as far as AI spends are concerned.
