Nikesh Arora on the Frontier Model Problem: Breadth vs Depth | The Future of Token Costs

I came to the United States with two suitcases, $200, and I was willing to do anything, anything at all to make sure that I made a life for myself because there was no way to go back. I was a security guard. I took notes for the disabled. I flipped burgers at Burger King. I had $200. I had to find a way of paying my tuition.

I think the long-term token pricing should be one-tenth of what it is today. Mythos, I think, ends up being an accelerant to cyber security in technology. You miss one trick, you can survive. You miss two tricks, you're partly impelled. You miss three tricks, you could be obsolete.

Well, the good news is I told you earlier, we're chatting, I have total memory loss. I have no recollection of what I said and what we talked about. So, I have to go back to listen to it. Just happy to be here. It was a great show. We got on blissfully well. It was wonderful. I don't know what that was about.

Oh god, I agree.

So I actually do, and the nice thing about size is they do. But the only challenge is they actually sometimes come quite jetlagged because they come like for the day, and that's a little bit tough.

Maybe.

Maybe I need to adjust. I'm the selfish one. Yeah, Marc Andreessen said something. Marc Andreessen said, I really actually embraced it. Haven't even got the first question, but [__] it. He said you need to embrace how is it all your fault. If you embrace everything in life with how is it my fault, actually a lot of the world changes.

Actually, let's spin that a bit. How do I make it better? Just change the outlook is how do I make it better? What can I do to make this better? That's how I run my day and let run my life or my company. How can I make it incrementally better today? And how can I make it radically better in 3 years?

Not for the lack of trying. So all we can do is try. If it works out, it's great. And if you try a lot, you succeed more often than you think.

I don't think about it like you do because this is what you do for a living. I think about running my business.

I'm here to learn. Teach me, Harry.

I'm here to learn. Teach me.

Do you think that is still the case today though? Like I think brand can be such an accelerant today, especially in a world where it's just so noisy.

See, I spent many years at Google as you know. I spent actually was chief marketing officer for 5 years. And if you look historically in technology, there are companies which have died who had great brands. Remember Sun Microsystems? It's a darling, you know, 30 years ago, it doesn't exist. Why? Because their brand went dead or the product went to hell. What about Yahoo? Remember that company? It's a great brand. Amazing, right? It's way before Google. I think it's probably a fraction, probably even two decimal point number versus what Google is. So I think product helps make brands.

Look, let me speak against my own thesis. There's a spectrum, okay? On one end of the spectrum is when you have real differentiated product, in which case the product helps build the brand, like Google search is now. You Google something, right? On the other end, and it's a commodity. It's water. I don't know why this is called Evian, right? But this is a commodity here. Only brand matters. So it depends where you're on the spectrum. If all you are a commoditized product, yes, brand matters a lot. If you are a differentiated product, then you build the brand in the back of the differentiated product, and you can decide where you want to be in that spectrum.

Yes.

Yeah. Look, I've been listening to some of your podcasts and I've been paying attention what happens to market because I want to understand where all this settles down. Not just because, you know, I want to understand it, but also impacts how I build my own business, right? So, you got these phenomenal frontier models and they keep sort of leapfrogging each other. Every few days there's a net new model that's delivered by OpenAI or by Google or by our friends at Anthropic. And the question becomes, okay, fine, these models are moving into space. What do I need to build? What do I need to do with these models? And you know, as we came with that Mythos moment when everybody was busy chasing Mythos and that was kind of important, you realize even the best model has a high false positive rate. But for some reason in the consumer space we don't seem to care. You know, I was talking to my sister this morning, said I just went to ChatGPT and asked all these questions. It was very helpful. So I guess like what happens is the consumers are way more tolerant of false positives because just it's kind of always the person in the middle, right? There's always somebody who's understanding what the model says and making their own judgment whether they believe the model or not. And somehow people get rid of some false positives. Somehow people don't care about the false positives. Sometimes people believe the false positives. So the consumer is highly tolerant on this notion of false positives and doesn't seem to distinguish, and it just seems to get better and better. I literally had Gemini produce an investment memorandum for something I was looking at. I looked at it, looked pretty accurate, give or take. I tweak a few things, but it seems passable. So on the consumer side, it's a breadth issue, right? It wrote an investment memorandum for me, which is cool. I would have had to hire a banker and a bunch of investment analysts. It would have taken me days, and I did it in 4 minutes. So the breadth is there, which means it's my go-to place. And as you know, in consumer, if you become the go-to brand, talking about brands, it's hugely beneficial, right? Whether it's YouTube, that's the only place to go look for streaming video, or Google, that's the only place to go do a search, it becomes hugely sort of multiplicative from a distribution perspective. So our frontier model friends are chasing the consumer brand, and the false positive doesn't matter. On the enterprise side, false positives matter a lot. They matter because if you imagine a future where an agent's going to make independent decisions and act on it, you have zero tolerance for false positives.

Now take Waymo. In my view, Waymo is the biggest agentic product that is out there because guess what? You've replaced a human being called a driver, right? All decisions are made by AI machine learning. It decides when to turn, when to stop, what to do. But think about the amount of edge case training it took to replace that human agent with effectively an AI-driven agent. I don't know, tens of billions of dollars. So that's what it takes to take one use case and train the hell out of it. And if you think about what happened there, they could have used equivalent of an AI model, but then they built so much context and intelligence and edge case training and proprietary data to make that happen. That data is not available on the internet. You can't stick the next model of Anthropic into your Mercedes and say, 'Okay, drive me home.' It's not going to be able to do it. And that's the depth issue, right? Because you need the depth of the context and understanding and the intelligence in around the model to make it useful for the truly agentic use case. So, I just think there's this constant tension. The frontier models want the consumer attention because that drives post-training for models. It drives the consumer brand of the model. On the other hand, the real enterprise revenue is going to come from use cases that require a lot more context. The one sort of standout use case we all know is coding, right? Coding is a universal activity. Everybody does it. So everybody's data is helpful in training the model, and that becomes sort of a large enterprise application. Hopefully there's a few more out there. But I think that's the tension that I wrote about.

I think more than half the enterprises are still not getting it right on the use of AI perspective. I think we're still busy trying to incorporate AI into our current business practices. So how do I take what I do today, use a little bit of AI, get marginally more efficient because I don't want to do this the old way? I think the opportunity is to rethink your workflow fundamentally with AI. That's where the true benefit is going to come. I think the winners in the long term will be people who actually rethink their companies with AI, not people who adapt their current workflows marginally with AI.

No, I think there's going to be perhaps two or three different categories. Like one category is let's take the workflows of today. We have workflows around ERP. We have workflows around sales team management. We have workflows around human resource management. There are existing workflows which have been SaaSified, as in some software company decided we all have common processes. Let me build a container where these common processes can be marginally customized by enterprise. They can code their workflow into my SaaS application, and we're off to the races. Now, that workflow required a lot of human judgment and human interaction. The software is not intelligent. It's been coded, right? You define input, you define output. I do the input, I know what the output I'm going to get. The idea is imagine workflows where, you know, in the hiring process, most of your workflows are containers. Imagine where AI actually is helping you make judgments. If I put every CV into AI and it says these are 20 people you should interview, this looked at the CV, you should ask this person these following questions, sends a note to Harry saying interviewing this person, ask the following 10 questions because your three other colleagues only asked the following five. We needed no false positives to human beings. So AI could be hugely helpful in informing and making the process more intelligent. But that requires us to give up human control and let AI do 80% of the thinking for us. And that's not how we're doing it right now. All we're doing is let's take this invoice, let's scan it, abstract the data, put it into AI and say look at that, it's happening 20% faster.

I think two different points. I think mass collection of data to inform AI is a little dangerous because people see the outcome what that's going to happen, right? And I've heard of companies where people are, you know, using cameras to track people folding laundry and earning clothes because they want to be able to train physically in the future to do those things. So that aside, I think on the enterprise side we can actually run a business much more effectively and efficiently if we decide where we are willing to relinquish control to AI. Perfect example, marketing, right? Anything that is required to train a marketing model is already out in the public domain. By definition, marketing is public domain, right? If it didn't market it, it's not in the public domain. So, I have the best training data in marketing. I don't need to train an AI model with more marketing content. I may need to train it for tone of voice and what my brand is. And I'm pretty sure an AI model, if I throw my marketing collateral into it, will tell me this is not consistent with your brand. And if I look at all the things you've been talking about the last 10 years, I think some people have done that. They've actually analyzed earning scripts of companies and said the last 20 years, what has Company X done? And when does the CEO start getting away from a topic because perhaps it's not going so well? So this is really smart. They can do that stuff. So it's the best marketing training database in the world. The frontier models. Why do I need 400, 600 people in marketing? Because my biggest problem in marketing is I have 600 people, but I'm not sure they all fully understand how to consistently deliver my tone of voice, my value proposition, and how not to break my brand by having different collaterals in the public domain.

Give or take on 21,000 people.

No.

I don't know. My rule of thumb is that in the next three years, we'll probably have half the people in G&A type activities in companies, things like marketing, things like finance, things like HR because there's a lot of process management there. A lot of process management can be made more intelligent using some version of an adapted future AI application for lack of a better word. So SaaS applications will give way to AI applications. The difference being SaaS applications have no opinion. AI applications will have opinions, and that's a fundamental rethink we need from a workflow perspective.

Everything, right? Like your scholar, whether you want to call an AI assistant, AI marketing assistant, AI HR assistant, is going to say, I looked at your copy. It sucks. That's not good enough. It's not consistent with tone of voice. Here's what I would recommend. This has an opinion that will make my average employee much smarter than they were today. Then I don't need so many of them because they're doing most of the work for you.

No, like I think the places where people could be wrong is how many technical resources we need in the future. I think we need more, not less. I think there's this fallacy people believe we're going to have less people working because AI is going to take over our jobs. I don't believe that. I think what's going to happen is you can't imagine the number of people on my team who want more technical resources, more AI-savvy resources because they want to do exactly these things. They're saying, 'Oh, I've got an amazing project to transform marketing. I've got an amazing project to transform HR. What do you need?' I need more people who understand how to prompt frontier models, build harnesses, bring proprietary data into play, bring modes. I need more compute, more storage because I want to learn everything. So, I think we're going to need more technical resources. I think we're going to need more sales resources because if your product's really good, you need more people to go out there and cover the universe because not enough people know about it. I'm in Europe. I met 20 customers last week. I still see half of them don't know all the stuff we have. I'm like, dude, we've been around for 20 years. Why is my team not out there pounding the pavement telling them everything we do? And the problem is not enough time in a day because I'm too busy dealing with some arcane piece of software at work which I have to go feed. Well, if I had that software be really intelligent, it'd tell me what to do. In terms of wanting more technical resources, tokens, I would put in the more technical resources camp.

Look, I know this whole world of token maxing is kind of gone topsy-turvy with the whole conversations around how many tokens people using. The challenge right now is 90% of the enterprise employees are not AI-savvy. They're not. They have to learn. I can't send them to university, right? There's no course you can take in any school anywhere. They have to be able to learn of their own. I think we're back to a Darwinian moment where everybody has to figure out who's really good. Now, you've seen people like Brian Armstrong and Jack Dorsey go out and say, 'I'm going to decimate my organization. I'm going to start building from scratch.' And they've gone to some version of 30, 40% less people because they've figured out there's no redemption. I can't train these people. I'm going to just find the people who are going to come in and help me do this stuff. That's one model. The other model is sort of gradual. We've been hiring people only through hackathons now, right? And we see natural attrition of 2% give or take a month, and we just replace them with people who actually are AI-savvy people. We hire from hackathons. Give me 12 months, I'll have sort of transformed 20, 25% of my team. Give me 3 years, I'll have hopefully enough AI-savvy people working at Palo Alto. So there are two different ways to get there. I think part of what you're seeing in the token maxing world is people are learning, people are experimenting. The risk is your smartest employee who knows how to use AI really well could be using 20 times the tokens that an average employee uses. And if you get into this whack-a-mole moment saying, 'Oh my god, I'm going to stop people spending too many tokens,' you actually will hurt the best AI-savvy people more than you will hurt the average employee. And by the way, I think the best talent will want to go where they will be best equipped with the most expensive frontier models, the biggest budgets. I think that will almost be like an employee benefit.

We have a use judiciously model for tokens. It's not a free-for-all. Free-for-all sounds like you can go token max the hell out of it. We have use it judiciously and we keep track of it to see what people are doing. And if we find somebody who's using it well, we won't constrain them. If we find somebody's gone a little over the top, we'll find a way to cap.

That's still a narrow lens for me. If I abstract, if I step back today, there's not enough compute for what the world is demanding. Unequivocally not enough compute, right? You can't buy compute. Compute is costing 2 to 3x or 4x more than it used to cost 2 years ago. There's not enough compute. That scarcity of compute and that excess cost required to build and deliver compute which allows us to go make AI useful is causing the constraint and forcing pricing, right? Because, and interestingly, I think more than half of the compute is going to feed the consumer, which is a fundamentally loss-making entity right now. Like I don't think any of the frontier models make any money in trying to get you and me to use ChatGPT or, you know, Claude or Gemini every day. It's free. That's a lot of compute. Imagine there's billions of people around the world using for all kind of queries every day. That's sucking away half the compute which is making no return. Guess where the pressure goes? The pressure goes on the other half of compute which is being used for coding and enterprise applications. So now you're saying enterprise applications, coding have to pay until we build transaction models or advertising models on the consumer side because they're not ready. Now you can say, well, that happened in search too, right? Google search was around, that happened. YouTube, people used a lot of YouTube, a lot of compute, but it wasn't paying for itself. The problem is the compute requirements and the cost is now 10x of that when we were in that era versus today. So that's forcing token prices to go up. I think the long-term token pricing should be one-tenth of what it is today. When that happens, you will see that people will consume more. You can decide if 3.8 or 15.8. Not sure we can tell the answer right now because pricing will move very drastically in the next 3 to 5 years. I think at some point in time,

I think so. I think in the next 3 to 5 years we will see reduction in token pricing. I think at some point in time the consumer use of AI will get constrained by these frontier AI companies because they have enough post-training data, more than they need, and each user is inherently unprofitable in their activities they do on frontier AI models.

You know, that's an interesting question. It has to come from somewhere. When I started Google in 2004, we were 2% of the global advertising revenue, and global advertising revenue was estimated between 500, 600 billion. I think online is about 70% by last count of total advertising revenue. And I don't think the overall number has changed by more than 3% a year or 5% a year. So I don't think the total advertising pie is going to increase. You've already taken away 60, 70% of the advertising pie in the online world. Unless you tell me there's going to be an explosion at the top where more people are going to spend more money in marketing, that money that you're hoping to fund consumer AI from advertising will have to come from current advertising revenues. So I don't think that changes the equation drastically to make the consumer profitable. I do think there's an opportunity that AI ends up taking more transaction revenue which has not been into the purview of AI.

Well, think about the marketing chain, right? Like we do advertising. Advertising is inherently inefficient. What's the best conversion rate you get in online advertising? You think?

That's fine. There are, I think the best of breed is 7 to 10%.

Sometimes. But the average probably one and a half to 2%. Which means 85 to 90% of your marketing is wasted. Right. So now if you get really smart, you have memory, you have context, and you get smarter in targeting Harry when he's trying to buy something out there in the world, then your conversion rate goes up, right? If you look at the entire value from the time you decide to buy something till the end, you get the product. You take the case of consumer goods today, I want to say the cost of consumer goods is probably in the 5 to 8% of total, you know, list price. The 92% is distribution and marketing. It's highly inefficient. So you could imagine a world where AI makes marketing really efficient and you get more dollars coming from traditional marketing into the online world because it's coming in the form of transaction.

Yes,

Well, right now they're figuring out that, you know, all your frontier model companies are value maxing, not token maxing. They're raising money at a trillion dollars. At some point in time, they realize, 'Oh my god, where's the next hundred billion of compute going to come from? Financial markets are not going to bear the cost of another $100 billion at a trillion dollars or trillion and a half because I'm going to need it again the following year.' So you say, 'Well, I need to build a sustainable business model which starts showing some degree of gross margin profitability.' The only lever they have is to take the fastest growing thing that they have in their portfolio from an economic perspective and charge us more for it. That's where you get the price of tokens from. I think the price of tokens is high now. You can imagine if the price of tokens is high, every technologist is trying to figure out how do I make my compute more efficient, right, in the future. So, I'm sure we'll see a whole bunch of advances in the world where memory and compute is going to start getting used more efficiently from a modeling perspective. And I still believe I don't need Fable 5 or Mythos 5 to do 90% of what people do with the AI today. Why is the last model not good enough for certain tasks? All right. I think unwavering is, I think the model overhang in terms of capabilities and especially consumer and most enterprise demand, like, but the models from two years ago were good enough for most of the queries that we understand. That's right. The problem is they were inefficient from a compute perspective. Totally right. So you're seeing the efficiency come in. The problem is the cost of R&D is now being sort of spent in terms of what the tokens have to pay for. So I think token prices come down. I think the amount of compute that we need is going to be huge for the next 10 years. I think the frontier AI models are in a position to capture a significant amount of the future economic value of the use of AI.

Look, infra is making money. Infra is more expensive than it's ever been. That's, you're seeing trillion-dollar market caps in the infra space because of the scarcity of compute and this need for speed.

The question which I don't know the answer to, you can tell me. You spend more time with people here. I have to go do my day job.

While the infrastructure, like infrastructure people are gearing up for large amounts of capacity, large amounts of demand on the infrastructure side. And you come to a point that says, you know what, there's only so many data centers we can build. There's only so much energy we have.

So I think there may be a digestion period at some point in time when we think the demand for compute is there, but the capacity to execute is now limited by physics and the infrastructure players have built up too much capacity for this demand. Now I don't know when that rationalizes, maybe it rationalizes and causes us to go think about a different sort of time frame for putting all this compute out. That doesn't take away for the need of compute. We'll still want as much compute as we can deliver as fast as we can deliver. I think some of the model companies have outstripped anybody else's ability to build frontier AI models of that capacity at that speed in the training. So I think you are seeing perhaps a settling down of who's going to be the frontier model player in the future. The question becomes in the economics what value accrues to the model, what value accrues to the application layer as you said it. And I think the application layer is probably a simplistic term because for the first time you have memory in applications, applications understand context, they understand context as specifically as to what you want, what I want. You're seeing that in the consumer space that has not yet come to the enterprise space. Funnily enough, I think that shows up in the enterprise space, which means the demand for computer memory goes up on the enterprise side. Like how, I haven't, I don't, haven't used all the coding models myself, but over time these coding models have to get really smart about understanding individual context of enterprises and humans. That's how they'll be more effective and more efficient. For that, we're going to still need more compute and more memory. So, I think that will start defining where the value occurs. I think the value gets shared between frontier models and the context that gets created in enterprise play. I think the frontier models are fully understanding that this is where the gap is. I suspect the frontier AI models, if I had a crystal ball, they will spend a lot more time the next year or two building memory around consumption.

More than that. If you look at the consumer interaction that you have with your favorite, favorite frontier model, right, it's starting to remember, oh, you asked about this yesterday, you asked about that. Should I take the question you just asked me in the context of everything I know about you, or should I just limit the answer to as if I don't know anything about you? Now that having context of what I said to you over the last 30 days or last 60 days, 90 days requires you to store a lot of information. It requires a lot of personalized interaction that needs to have. Now if you want to maintain your moat with Harry and Nikesh in the future, the more context you have about me, the easier it becomes for you to give me the answers in the future. And as you start building context on a user basis, you create stickiness and that becomes your moat.

Mythos ended up, I think, ends up being an accelerant to cyber security. Yeah. I think what happened when you saw Mythos came out, it demonstrated that all the training we've been giving these models on how great code is written, the models were able to turn around saying, well, I also know how to find bad code. So what happens, you point the gun the different way and the model says, oh my god, look at all this code that you have, there are so many flaws in it. As we talked about, the challenge is like every model, it also suffers from false positives. So if you're an offensive actor and I point the model against, let's just say 20 VC enterprises and I go look at everything you've done. Somebody's left a websocket open, somebody's done some mess up with IP addressing, etc. So it finds the flaws outside in that allows the bad actor to go figure out how can they daisy chain vulnerabilities and get into your infrastructure. It's not good enough from a defensive perspective because I can't use the model and say go take every vulnerability you found and build a patch and go patch my system and protect me because guess what? It's going to patch 30% of things which are not wrong. Who knows what that's going to do to blow up your infrastructure. So when Mythos came, we looked at it. We treated it with respect. We ran it against our code. We discovered it finds bad stuff much faster than humans can. We found in 6 weeks what would have taken us 5 to 6 years. So we got it. We ran around. We patched it. But you know, cloud code helps build the patch, but you should have to run it through human eval, through testing, through production testing, to sandboxing to see does this patch break anything in infrastructure. Only then after 6 weeks were we able to go patch everything. Now, so what does this mean? This means that every enterprise better fix their stuff faster because if I point the next generation of models against your infrastructure, it's going to find security flaws, security vulnerabilities, misconfigurations, things that you've not been paying attention to. So it creates a bit of an urgency on the parts of the customers to improve their cyber security posture, which I think generally is a good thing for cyber security companies.

Well, look, the solutions are there. Solutions are there. The challenge is getting, sometimes getting the attention and focus of the customer saying, listen, I got to go fix my stuff because it's important. What this has done is it lit a fire under the security practitioners around the world saying this thing is not good, this is going to weaponize the bad actors, I better make sure my defenses are in place. Now remember, the way cyber defense is done is just fundamentally, just cyber security is two fundamental things, right? One thing is if it's bad and I'm at the gate, I'll stop it, which means you have to have somebody at the gate. Now we have 150 million sensors in the world where we stand at the gate protecting our customers. So if I can find a way of infusing AI at the gate and taking all these vulnerabilities and finding a way to protect you, I'm good. I don't have to chain the gatekeeper because there's no cloud endpoint agent that exists out there. There's no open endpoint agent that exists out there that I can replace Palo Alto or the other people in the space with. The problem is not at the gate. What happens is despite all the perimeter defense you put in, things come in, things leak in, people make mistakes, people passwords get breached, you know, there are vulnerabilities, people get in through. Then the question becomes, all right, oh [__], the bad actor's in my infrastructure. How quickly can I find him and get rid of him or her? That becomes an AI task. That becomes the same conversation we're having so far is I need context. I need intelligence. I need to know what this means. So creating that context, that intelligence within the enterprise of what this intrusion means and how to protect against it becomes a challenge. This is the AI cyber security challenge, something we, you know, again, this is not trying to hear to pitch my book, but we spent 5 years trying to build that capability inside enterprises. So in the net, it ends up being an accelerant. Does that mean I have everything I need? Not everything. Does that mean I need to get AI models to start helping me? Yes. So we're going to infuse more AI into our defense infrastructure.

I think we're going through a discovery process. I think this notion of guardrails has not been built robustly enough because these models seem to be easy to get past. Remember the early days when you used to read about somebody had this conversation with a model and found a way to, you know, obey the guardrails because ask questions differently and the model was able to get sort of jailbroken. It became a hobby amongst people and like they had all kinds of conversation with models. The same challenge, how do you make sure that you can put enough guardrails around AI model that you built to make sure it's only used for the purpose that you had, right? That's the challenge that there is. I think the guardrail needs to get better. To the extent the government feels the guardrails aren't robust enough, is trying to tell us it's a national security issue, I need to go fix the guardrails. But I think it's a simple matter of trying to treat the guardrails as a real problem and solve it.

I'm hoping it is.

Oh, he has questions for me.

The paranoia I have if I look at self-driving, there are broadly two or three approaches out there, right? One was my car is not going to have a human in it. It's called Waymo, right? I'm going to keep pounding it. I'm going to keep pounding it, training it until it learns by itself to drive and no human should be holding the steering wheel ever when my customers are in it. Right? You see, it's out there. Many cities have, I saw one down the street from here. So that's one way of doing product development. And what you do is every edge case gets discovered. You build training around it. Every experience is a learning experience. You build training around. You keep training. You keep training. You keep training. Get to a point of total autonomy. The other version is I'm going to start taking segments of the driving and start automating that segment where I get really comfortable. So my car drives 50% of the time. The other 50% the human gets in on the edge cases. That's my Tesla, right? Tesla used to drive just the highway for me and now it's getting better at in other streets. It's slowly getting better, but still I'm holding the steering wheel very often. It'll tell me you're not paying attention. Look at this. Look at the camera. Otherwise, you can't drive. And I know FSD gets better. But that's another way to get there, right? Okay. I'm going to keep training. I'll fix this. I'm going to start working the edge cases as my business continues to evolve. The third one is I'll infuse some degree of self-driving into my car because I've come from the traditional model of having amazing cars, great V8 engines, beautiful, sleek cars, and I'm not into the technology aspect. You see them out there on the street. We have all three versions out there. My fear is, do we all need to stop and start thinking the Waymo way as enterprises, or is there room for the Tesla approach to self-driving in our businesses, right? Because we have an existing set of customers to satisfy who are not going to take kindly from me saying, you know, guess what, I change my product, it's right 80% of the time, and I'm going to take another three years to train the product to be right 100% of the time, right? So my fear is, am I pivoting fast enough in my product strategy that over time my products become more self-driving than they are today, or do I need to go faster and can I get there by automating or AI-enabling certain parts of my product where I can apply the models that are capable to do certain aspects of cyber security today and keep doing the other through machine learning and managing edge cases, machine learning, or is it time for me to pivot? My view right now is you have to have the Tesla approach if you're an enterprise that is building AI-infused capability, but you can't have the approach of traditional car manufacturers which are trying to stick a little bit of AI and sort of AI-washing their cars and saying I'll get there eventually.

Different courses for different horses, right? I don't think in our business we can go implode the organization because I don't think the underlying application software is there. I don't think all the things that we've talked about that we need to get done are ready from an AI software perspective. I don't want to build a lot of software that is proprietary to me for things that should be available for everyone. I don't want to build an AI marketing stack. I don't want to build an AI HR stack. I don't want to build an AI ERP stack. I'm hoping that somebody goes and does that much more effectively and efficiently for the world at large. Perhaps it's the next iteration of Salesforce, next iteration of SAP or the next iteration of Workday that is going to help me do that because I do believe it needs to become more intelligent. We talked about it needs to be more AI-enabled or AI-controlled. I do want to build things that are particular to

me. There I have all the information, all the intelligence, the context and the memory from an organizational perspective. I think the right way to get there is to hold people accountable. And I run a meeting twice a week now called AIO. It's kind of funny. It's like old McDonald had a farm. AIO because everybody in my company wants to do AI. So I use it as a converging function, as a convergent function to do brainstorming across my team. How do we think about this? Why are we building this? Because I have to AI.

Everybody [clears throat] comes and shares how they're adapting to the new world of AI. What are they doing from a product development perspective? How are they thinking about it? How are they including agents in their products? How are they going to go build the backend infrastructure? How do we think about how they're using tokens? How do we think about the capability of resources? Remember, for me to transform a 21,000 people organization, I have to get the hearts and minds of the leaders to make sure we're all swimming in the same direction or pulling in the same direction. So this is my way of ensuring the top 15 or 20 technical leaders in my company are pulling in the same direction, whatever the direction be. And as you know, if there's no expert, then a group of smart people do better than an expert. I don't think there's an expert for future enterprise design yet in terms of here's the blueprint. Like you said, as a VC, you're saying, holy [__], where's the value going to accrue? Is it going to be in models? Is it going to be in the application layer? Trust me, we have to have a point of view in that stuff and what's going to emerge in the future before I can start transforming my company. Like, how should I build my products? Are your leaders AI-built? I interview CROs as well, some of the best CROs, some of the best CPOs. And in all honesty, the bigger the company, the less AI-built, AI-matched they are to the extent where one the other day, public company doing 5 billion plus in revenue, CRO goes, you know, we don't have that AI talent internally, but we'll bring it in.

Yeah, like I discovered this in 2004 when I was in Europe. I ran Google Europe. I used to go around meet CEOs. You know, there was this fad where CEOs would hire this 24-year-old sherpa. They were called the web sherpas or internet sherpas. Right. It's like chief internet officer. Remember those companies had chief internet officers at one point in time.

Well, we've seen this movie before, and sometimes it's fine to have seen this movie before. And the task of the chief internet officer was to make sure the organization was ready for the internet because I'm too busy in my traditional business and I don't know who Amazon is. I don't know who Google is. So this wonderful 20-year-old who understands this stuff is going to help be my savior. And then CEOs would wash their hands of the internet because they have this wonderful team of people who would be frustrated because they can't get anything done because nobody's giving them attention. The risk of that happening is true with AI as well. I'm so busy doing what I did yesterday. I have no time to think about tomorrow. So meet my chief AI officer was probably a researcher at some amazing university before and has low execution skills. So until I can get my leadership to understand and agree the extent of the AI challenge and the AI opportunity, we're not going to make progress.

Look, everything when you go talk to a product manager or any large company, I'm pretty sure they have a product roadmap that exists in their heart, in their hands, which is 6 months or 12 months long. I got to go fix all these things. I'm like, what's interesting is in the 6 to 12 month thing, there's nothing called agents in there. How come like the world is talking about identifying everything and your product roadmap doesn't have that? I'll get to it once I get this done because it's what customers want right now. I'm like, no, that doesn't work. How do I get you to do more agentic work? How many people are you going to free up doing development the way you're doing it today so I can use that excess resource to go make new things happen? So those are all important conversations. I can have them one at a time across 14 or 20 people or I can have them twice a week with them and people demonstrate how they are. And what's fascinating, remember, you have to make sure your leaders are ambitious. You have to make sure they're competitive. You have to make sure they want to win. You have to make sure that they have a learning mindset. When they watch their peers around them do cool [__], they want to show up with cool [__] the next time. So for me it's getting 14 people together and saying, hey Harry, tell me today what have you done for AI in the last three days since I last talked to you in your organization. And whatever motivates you, whether the fear of Nikesh asking you 3 days again what you did, or your sort of inherent learning ambition, or it's your team pushing you, you will show up with something. And then you'll see what the other guys are doing. You'll say, oh my god, I'm doing a lot or I'm not doing enough. So it creates a little bit of Darwinian competition amongst them. It creates this urge to go embrace this new technology. And I think hopefully I get 14 people fully motivated and then they go do that with the next set of people because I need to transform from the top down, not from the bottom up on this topic. There's a bottom up experimentation, of course, right? People using tokens to see who's really good at that allows me to find the best talent. So you got to find a way of transforming 20,000 people over the next three years in that direction. Bottoms up, top down. You got to get into these organizations.

Okay. Is that that went viral?

So I think what is true is we've only been chasing the enterprise dream for AI for the last what, 12 months?

And if you think about everything that happens on a weekly basis, we see new things come which we don't quite fully understand and grasp, right? We're all busy trying to get our arms around LLMs and how they're going to be great for chatbots to talk to our customers as an enterprise. And suddenly agents showed up like, oh my god, I got to figure out agents are going to start working internally. Agents going to do a lot of stuff. I'm pretty sure you could still have an agent fest and have everybody tell you what an agent is. You'll still walk out and say, I'm not quite sure that his agent or her agent is the same as what the last guy said. So because AI is moving so fast, I don't think the products are fully there yet. Like the enterprise products at the application layer don't exist in their entirety because we haven't been tested against the enterprise ask. So FDE is a short form for saying my product's not fully there because it's evolving as the technology evolves. I'm going to send some people across who are going to sit in your office and build my product while I adapt it to your needs. That's what it is, right? That's what we saw from Palantir. That's what we're seeing from all these companies. So what you're saying is I'm going to send my product engineers or developers to your enterprise. They're going to build my if you get it, if you do it right. Again, FDEs have different versions. Some people are just trying to get you to consume AI, which is actually not an FDE. It's just a technical sales consultant who's trying to help adoption. On the other hand, an FDE truly is somebody who actually brings the code back from a customer side and goes back to your product and say, listen, I built this at the customer side because they had this need. We should incorporate this into our product because everybody's going to need it. That's an FDE in my mind. So I think FDEs are needed for the short term because remember all the enterprise AI startups are hungry for revenue. For some reason we've created this notion that don't worry, just keep selling it. There's a huge sort of pent-up demand around AI applications. Sell it before the product's fully ready. That's what you're seeing.

I think that's the case. I think as we think things evolve in the next 12 to 24 months, people will switch from one set of products to another because something will emerge as a better product. Look at the coding conversations, right? How many coding companies have you heard of the last 24 months? I think we had Windsurf, we had Devin, we had which is now Cognition. Windsurf got sold.