Thank you everyone for coming today. I'm super excited about this on behalf of all of us at Iconic. This has been something that I've been looking forward to and I'm so grateful you're here and have chosen to spend that time with us today. So, thank you.
It's been quite a year as all of you have experienced. But this technology this last 12 months has been something that we've never experienced before. It's not like the internet. It's not like mobile or the cloud. It's evolving at a pace that none of us truly understand and can take hold of. So the leaps between these models we seeing stuff we can do today that we couldn't do a year ago is happening in a month going to a week. And this is a real opportunity and a question for us.
When we gathered last year the question we talked about in this venue was the first and foremost question around AI. How do you move from experimentation to real enterprise impact? So now a year later, what's our set of questions we got to think about? We have a much clearer picture of what this all looks like. But as the technology has advanced, the fundamental question has come together, which is how do you balance near-term ROI with long-term? Where does the LLM end versus the application layer that kicks in? What is simply an API core versus a real business? and who's going to own it. I'm not sure we have all the answers, but how this plays out is going to have enormous implications for the companies getting built, where we put capital in.
We have an incredible group today as you spend time getting to know each other that would value perspective on these questions and actually have a strong opinion on this. The people building the systems, bringing the technology, and some of the world's leading companies that invest in the capital behind it are sitting in this room.
So, we may not leave today with all the answers, but the idea is hopefully we'll walk away a lot more informed than we were at the beginning of this day. This unprecedented pace of AI has helped us navigate, you know, technologies we've seen before, which is the good news. We've had this pace in other things, but very few people in the world have seen this type of torrid AI evolution through the lens of normal technology. and one of them is Nikesh Arora who's going to be here today.
So I've known Nikesh for 15 years and he's been a very good friend and mentor and partner of ours at Iconic. He's been at the center of some of the most transformation companies of our decades. He's been thinking of telecom then the internet mobile at Google, SoftBank, cloud and now AI at Palo Alto Networks. And he isn't an observer. This guy is in the middle of everything. is making massive decisions and massive changes around where this needs to go. So I can think of no better person to have join us today and kick off the day. So Nikesh welcome.
So Nikesh, we've had lots of conversations. One of them around this topic of if there was something you believed about AI 12 months ago that you no longer think is true today. What does that sound like? What does it look like?
Well, it's hard to have beliefs about AI because every 3-4 months you get a curveball and the curveball we all got two, two and a half years ago was this notion of ChatGPT and everybody was holy this thing can do stuff about 70% accurately and that's when you saw Google got sort of caught unawares because Google was busy trying to perfect it to make it 100% right and never let it out of the research labs until it was 100% right and suddenly we found an exciting consumer use case and people out there trying to do 70-80% of the stuff which seems to be above average intelligence most of us don't get 80% stuff right so this seems like pretty cool and we thought we had it all figured out we're going to have LLMs and suddenly we saw this open claw execution and suddenly you had agents that running around that opened all of our brains up to saying holy we actually don't have to be physically present to get something done we can enable a swarm of agents and have them do and that was cool.
And then you know we have I'm sure you Sam later today I saw and then you have Dario and you know we're all trying to figure out who would be king is it Sam, Dario or Demis and now in the last 60 days or so the conversation has shifted to open weight models I won't say open source open source always been around but open weight is a new thing it's like enterprises can possibly train their own models so that we don't have to be captive to the weights of the frontier LLM and suddenly we all feel somehow the idea not that most enterprises know what to do with it but the idea that he can control my destiny seems much more exciting than the fact that I have to be captive to Sam or Dario.
I don't think we're done. I think this is going to go on. Like this morning I was listening to this guy who's trying to build an infant stack which is the cheapest stack to serve as opposed to a fastest stack. And his thesis is we all got headfaked into a fast low latency compute stack because we were coding and prompting which means you want response to come back quicker. He's saying if you want long-running agents overnight, you don't need low latency. You need cheap inference. So my point is every 3-4 months as people anticipate our behaviors the new idea changes and it's too soon to call where everything's going to settle but clearly some things are emerging.
Interesting. So Nikesh if you had a billion dollars to invest into...
You can make that happen right? I hear you have a lot of money to invest.
So, why make it hypothetical? Just give me a billion dollars.
So, you and I have a billion dollars.
Okay. Fantastic. Wonderful.
Expensive coffee down the road.
But what would you not invest in because you think the economics are going to fall apart?
It's fascinating to see where the economics are showing up, right? You have to understand economics, you have to start where the profit pools are because we're not creating new profit pools. The consumer profit pools forever have been in three areas. They've been in subscription, advertising or transaction. Whether you look at Apple, you look at Google, you look at Amazon, no company has built all three. By the way, usually you have a dominant economic sort of engine. Amazon was a transaction engine. You bought on Amazon and they limped along in advertising and subscription. Well, now they're bigger, but still their core business is transaction. Google on the other hand is a large advertising business, but it's not that great. It's a margin zero transaction business and they started building a subscription business. You look at it pretty much across the world if you look at consumer business only through three sources of revenue and there is some version of that that ends up becoming a profit pool on the consumer app side.
We haven't seen AI tap those profit pools aggressively yet. Yet, you can see at the end of the day, we're competing for finite amount of consumer attention and finite in a very large sense amount of consumer dollars and they're going to show up at some place at some point in time. I believe in the next 5 to 10 years is more than likely that advertising and transaction will get much more efficient.
Right? Today you spend a dollar on advertising you end up with a 2 to 5% conversion rate. It's highly likely with AI your conversion rates goes to 20%. Which means you get much more efficient in advertising. Conceptually advertising profit pools could shrink because the money is going to move to transactions. So that's kind of where the consumer stuff doesn't answer your question. The enterprise side is in the end businesses spend money on technology which eventually shows up as a profit pool for anybody serving the enterprise stack.
He has this huge layer in the middle of data center businesses. I coined a new phrase last week called neocalers because I said neocloud are going to get cheap. So people gave me... so it's okay. Neocalers will survive. So the two or three people who are fans of Nebus and Corv got vacated. So they didn't kill me on a Twitter yet. So we'll see. But there's a whole layer in the middle which are the data center providers.
I don't know how that lands. I feel like you're building inefficient compute because you have a license for land and power. And some people will make it and some people won't. I'm pretty sure that Google doesn't pay the same price to buy chips and memory that somebody building 50 megawatt does. Has to be different, right? It's a scale business. If that's true, fundamentally, somebody's building compute at one and a half times the cost of what hyperscale is building it at.
So when let's assume at some point in time the world normalizes and you can pick your horizon. I said two or three years and people beat me up. So let's say five or seven years so you don't beat me up. But in five to seven years if compute normalizes then the price has to go down. If the top line goes down and your cost can't go down because you already put the money in the ground then your margins shrink. So at some point I think the data center business becomes a regular IR business.
And I see company after company, company after company says, 'I'm going to raise equity. I'm going to give it to the data center providers. I'm going to build data center.' That's the worst way to fund a business. The worst way to fund capex is equity. But we're busy giving money to people building data centers as equity because somehow we believe that's going to... equity is the lowest on the liability table when you're a capex intensive business. The first guy who comes take the chips. It's not the equity guy. The guy takes the chips back is Blackstone. He'll lend you the money.
Or Jensen will take it back if you lend you money. Anyway, so I already lost my...
Where I put my billion dollars?
I would put my billion dollars in people helping companies do AI transformations because every CIO has no idea how to make that transformation happen. And I'm sure a bunch of people are trying to do that. And I would put my billion dollars somebody who was going away for three years. Say I'm not going to have a product for three years. Nikesh, give me a billion dollars. What are you going to do? I'm going to take a product that is done in a certain way today. I'm going to rip it out and build it from scratch.
And I apologize, but I'll say one more thing. Every one of us, including people who run large businesses or products, has this existential question. The existential question is in five years from now, do I want to be Waymo, Tesla or Mercedes? The reason I use that analogy is a Waymo was not drivable for the first 9 years. But today there's no human being sitting in the car. It's all doubt. A Tesla has all of us driving them. It can do 50-60% of FSD or 80% of FSD, but I still see people with a steering wheel in hand. It keeps saying grab your wheel to you. But it's more AI enabled than not.
And Mercedes probably AI washed.
Right. There's like a Grok agent you can talk to saying, 'How's it going? How's your day? I'm AI. I'm going to help you understand your car.' Okay, great. Now, if you take that analogy, there are people taking software companies and they're putting little AI agent. So, hey, how would you like to extract your data today? Okay, great. How can you help me? I'm going to just call my cloud for oh, sorry, cloud AI agent and help me answer the question.
So, the question is, which one do you want to be? I don't want to buy software which has been AI washed five years from now. I at least want to buy software that is AI enabled or reimagined with AI. And if I look at people in the startup business, they have this drug where they've been told, get your first beta customer in 6 months, show me ARR in 12 months, raise more money, show me more ARR in 24 months, raise more money. And all you end up building is Frankenstein.
That's helpful, right? So then if I push on economics in this angle for a second, you've argued that frontier labs may be caught in this AI business model trap. So help me understand more about that. Consumers generate the data but don't pay enough and enterprises who can pay but demand reliability for these models but don't yet provide that level that they need. So five years from now who actually captures the value?
The good news is once you get escape velocity, you become too big to fail. When you become too big to fail, so yes, I did say there's a business model trap. The business model trap in my mind is as follows. OpenAI, Anthropic both allow us to use pretty much the product for free if you're a consumer, at least on the chat side, not as much the coding side, but there's a lot of free. And the value I guess to them is that that free allows them to train so that they get a better frontier over time. So we're sort of you know if it's free you're the product. So we're the product that's fine.
The challenge I suspect right now is 50-60% compute is going towards serving free. As a result when they look for money to go build their next data center from capex perspective they end up charging enterprises twice as much which means we're all paying twice as much. Now, interestingly, going back to what we just talked about that twice as much is pretty non-discriminatory. I could be running long-running persistent agents, I still pay twice as much as opposed to I could be learning low latency compute, I still spray twice as much.
So, I think over time, you're seeing that routers are being touted as the people who will allow us to arbitrage from a commercial pricing model. I am not a huge fan of routers conceptually because I think at some point in time we barely understand one model trying to parse your intelligence across seven different models and trying to figure out what's the right answer is holy right so anyway my view is over time and you're already seeing that over time what's going to happen is I suspect both Dario and Sam and even Google should build their own routers to allow us to move between their own models which could be cheaper because they're the most logical people to hold context and harness and memory in one place and abstract the model layer below that.
I think third parties trying to abstract harnesses and context and intelligence and then trying to route them across the board to 12 different models is going to be a task from hell. So I think price will come down for the enterprise world. I think we'll all pay a lot less for tokens, but I suspect we're going to end up using a lot more tokens. So I think the math shouldn't change for our friends in the LLM business. I think the maths will head fake a bunch of people who are trying to build this middle layer because if you're building a product purely for economic reasons generally it's a dangerous it's an unstable place. It's called an arbitrage model. This industry is not stable enough to arbitrage.
That's interesting. So pricing power given this two-sided problem, right? One is that token prices need to come down but enterprises need to actually efficiently work and benefit from this. Where does the power sit? Do you think it sits with the LLMs ultimately or the enterprises that say I'm going to only pay you X?
