I have nothing to lose. I came from nothing. I never even imagined myself to be doing all this.
A $20 billion company, 45 million users, over a billion searches a month, built in 3 years by 400 people.
These numbers don't motivate me. It's hard to get motivated by wealth. You want to get motivated by impact.
This is Perplexity with co-founder and CEO Aravind Srinivas.
No one's ever in a comfortable position that no one can relax.
They forced Google to redesign their homepage, then bid $34 billion to buy Chrome.
More than their own valuation.
Perplexity changed Google.com more than any product manager at Google has ever done. Now you look at AI mode, it looks exactly like Perplexity.
He doesn't do defense. He doesn't do comfortable. His words: attack, attack, attack. That's my motto. Go all in and try your best. Be on the offense all the time.
You know what I hate with podcasts? When people sit on the fence. Aravind has really strong opinions in the show today. He says that Micron will be more valuable than Meta. He says that the resistance to data centers will continue and get worse. He says the biggest problem today is a lack of power. He claims that Perplexity has changed Google more than any Google PM. You want opinions? This is the show for you.
Ready to go, Ara? Dude, I am so excited that we get to do this. We've done one remote and then we did one at Founders Forum last year. So, thank you so much for joining me in person.
Dude, it's a weird start, but just roll with me on it. I asked this of the best founders that I meet: Are you motivated more by the fear of failing or by the thrill of winning?
Because I have nothing to lose. I came from nothing. I never even imagined myself to be doing all this. So my life has already been extraordinary beyond any level of imagination. I was just in India, you know, doing my undergrad and just training neural nets with graphics cards that people in the labs were using for playing video games. It was all for fun, and you know, my path led me all the way here. It was never like a... for my mom, just getting a job was success because we were financially lower middle class in India, which is not even like lower middle class in the UK or the US. And so from there, all we wanted to do was get a job in Google. Being an engineer at Google was considered a win. And so I'm already doing remarkably well compared to that ambition we had as a family. So there's really nothing for me to lose. That's why anytime I try to act like I'm trying to avoid failure, I'm being on the defense. I remind myself that that's the stupidest thing to do. It's better to go all in and try your best. Be on the offense all the time. Attack, attack, attack.
When you review then, what are you not being aggressive enough on today?
Maybe in the early days we were very, very loud on social media talking about Perplexity versus Google, and I used to do that myself a lot. And some people don't like me for having done that. Today I'm a lot more measured in how I talk about our products or competitors and stuff like that. But it's not a lack of aggression or anything. It's just that that is boring. People already heard that enough from me.
Do you regret being so bold in your messaging?
No. So it's not a nuance and maturation of message. It's that it's stale and I need something new.
Not just that, I kind of don't think it's a relevant framing anymore. We worked on search. Perplexity started out as search. We built the first answer engine in the world that people know. Perplexity, even today, if you mention the name Perplexity, people would think, oh, that's an answer engine. We built a lot more things after that. We built a lot of agents, browser agents, deep research, computer. We built so many products after that, but we're still known for that first product. And the mark has already been made. We changed the roadmap of Google. You could argue that I or the company Perplexity changed Google.com more than any product manager at Google has ever done.
Make that argument for me.
Well, nobody ever wanted to ship an answer engine at Google. Nobody wanted to tinker anything on the interface that made them $250 billion a year. And then now you look at AI mode, it looks exactly like Perplexity. There's not even any difference, like the font, the citations, the specific bolding of inline text, inline hyperlinks, suggested follow-ups. The whole experience is literally looking like Perplexity, except it's still not as good.
Is that bad or good for you that they learn from you and adapt?
It's both good and bad in the sense, you know, obviously I knew this around the end of 2024, this is going to happen. So it never caught me by surprise at all. It was just a matter of time. I still am surprised that the quality is still not there because I regularly test every product out there. But I'm happy that honestly they changed Google to be what it should be. And I believe that the frontier is where the money is. The frontier in AI is not about answering questions anymore. It's about actually going and doing work for you. We still have the state-of-the-art deep research in the world. And that's actually where people subscribe to pay for our Pro or Max products, is not for getting answers in the traditional way. They're asking for sophisticated research reports. They're asking for agents that go and do things for you. And so we wouldn't have been able to do all that if we were sitting in 2024 thinking we have everything settled here, we're good and comfortable. No, the answer engine was always a lead gen for the frontier products we build. You need something, right? Like think about it. Every company needs to have one successful product to build the next set of products. And in AI, nobody can sit comfortably thinking they have it all sorted out, including Anthropic. If Anthropic thinks Claude Code is already a win, in 6 or 12 months from now, they won't even be around. And so that's an uncomfortable fact about the whole field.
Would you argue today, you just told me, if you don't mind me quoting you here, you just told me before we started, that you think OpenAI isn't ready for an IPO. Would you have believed you would be in a position to say this two years ago when nobody wanted to deal with any product other than ChatGPT?
Think about it. So anyone, even in such a massive advantage position, can be put in a position where they're no longer the kings. They're fighting from behind, right? So that's the state of the field. It's less about Perplexity or Anthropic or OpenAI not having moats or having moats.
Can I push back on you? I would stand by two years ago, even when they were dominant and they are still a dominant consumer product, but I would stand by it because I don't think they are financially ready when you look at the balance sheet of that.
Maybe I'll decouple that. I'll decouple that.
Let's decouple that being like financial readiness for an IPO versus perception of a dominant leader.
Do you perceive them as a dominant leader right now?
Well, except there's no money there, right? Because it's been commoditized.
So it's always a lead gen. Like, for example, why are they going all in on Codex? Because that's where the money is. And we're doing the same on computer. Anthropic is doing the same on Claude Code. Google doesn't yet have a product in this category, but I'm sure they're going to come after that. Meta is trying to launch Hades for $200 a month. You see that? You see what's happening, right? So nobody...
But there has to be more money than just code, Codex, Claude.
It's not about code. That's the main thing. The money, at least in non-advertising, I'm not talking about advertising revenue. In non-advertising subscription or usage-based revenue, the money is in whatever is the frontier. And today the frontier is about going out there and doing things for you.
Do you not think then that there will be a 100 to 200 billion dollar advertising business for OpenAI?
It's yet to be proven. Let's work through the categories of advertising. Who's the number one advertiser on Google? Amazon. Who's number two? Booking.com. Number three or four, I think, Expedia. So, how much do you think Booking.com spends on Google? 16 billion, something like that. Some crazy amount like that. And how do you book your hotels or flights today? Do you book it on ChatGPT or do you book it on Google?
For me, I actually like discovery. I would like to see the options.
Exactly. Right. So the interface is less about conversations and more about exploration. So when the decision making is more subjective and vibes-based, you don't need an objective answer engine. And you think about the other category of advertising, direct-to-consumer products, fashion. Where is most of that advertising budget going into? It's going to Meta, Instagram, because you're just browsing. You're just like doom scrolling or whatever they call it, right? And so the chat interface doesn't capture that user intent, that user behavior right now, which is why it was never a great fit for advertising. And it also fundamentally corrupts the trust that people have when they go into a product and they want the accurate answer, which is what Perplexity is known for. And then you're like, hey, by the way, you ask for the most highest, best protein shake, but by the way, these are good protein shakes you can check out. It kind of hurts the trust that people have in your platform and your product. And so that's another reason why, if you think about it, like what Meta or I think some other companies in the past have tried to put ads inside messaging apps and emails, and it's never really worked out. It works out in China in WeChat because there's no other way for them to fund the whole thing, you know. So the whole economy and user sentiment, user behavior has been optimized around gamifying. It's not how things work in America. So I'm bearish on advertising to really take off in the chat interface. I'm happy to be proven wrong there, but I'm bearish on that.
There are two areas that I want to unpack there. The first, and just taking them kind of chronologically and how you said them, money's in the frontier. The more I hear this, the more I question it because I think that we dramatically overestimate how important frontier models are to do quite basic work.
Yeah. So frontier doesn't mean a frontier model. Frontier just means whatever is the frontier outcome you can have right now with AI. Greg Brockman recently tweeted, the model is no longer the product, right? And it's funny because you know that as a leader of a frontier lab, he has all incentive to say the model is the product. And that's what Google people tell. I think one of the Google people keeps tweeting that model is the product. I forgot who. And so the reason Greg is right is because if you take Codex or Perplexity Computer, Claude Code, what is that? It's an orchestration system, right? It takes a model, pairs it with an agent harness. And what is an agent harness? Think of it the simplest way of describing it is like rules for how the agent loop should run. What are all the skills and sub-agents and connectors and tools it accesses? And without the harness, you don't necessarily capture and convert the intrinsic intelligence in the model into valuable output tokens. The output tokens, if you're literally just a reseller of model tokens, you have no business because the model will get commoditized. So even if you're a model builder, you don't have a business. As an infra layer, you have some business on serving those output tokens. But as an application layer or a model builder, you don't really have a business. If you're just a reseller of tokens that come directly out of the model, you have business if you know how to take the model, ground it in valuable context, orchestrate it with a really good agent harness, connected to the right set of tools and connectors, whether it's personal connectors or business connectors, and provide the experience to people in one single unified system. And the way we differentiate ourselves at Perplexity is we don't just orchestrate across tools and files and connectors. We also orchestrate across models. That is the differentiation that Anthropic and OpenAI cannot claim because you wouldn't find GPT-5 inside the Claude Code harness. You wouldn't find Claude Opus 4.7 or 4.8 inside the Codex harness. These are competing with each other, right? Whereas you would find both these models inside Perplexity Computer, and that way we can bring down the... we can increase the token value per watt per user. If you assume that whatever decides the dollar, like the price, the dollars is the power watts, fundamentally that's the thing that nobody else can subsidize other than the government. You know that whoever provides the most valuable output tokens with the least amount of power expended to produce them generates the greatest value to the end user and has the most pricing power, has the most value. And so that is the orchestration problem to solve. The one single most important metric in AI is token value per watt per user.
What does it mean for the value of OpenAI and Anthropic if the model is not the product and it becomes a utility, something you can switch into and switch out of?
Interface. Everyone thinks we're all building the model layer or the race. We're not actually. I would even argue that building models is a way to stay at the frontier, but you have to own an interface in which valuable AI output tokens are generated, the most valuable tokens. It doesn't have to be the product. This is the single most important thing to unlearn for most founders, and I had to do it too, which is to be successful in AI product layer, whether you're a model builder or not, it's not about building something that gets a billion users. That mentality has to completely shift. There are a few power users who are propelling this token economy right now. If you look at all these crazy stories of how there's this one engineer who got Amazon to spend like half a billion dollars in a month because of some stupid way they set up like agent loop inside Claude Code. Okay, maybe that's a mistake, but there are real engineers in Meta, in other companies, spending like 10 million a year per engineer on these coding tools. There are users in Perplexity Computer. There's one user, I think, who spends upwards of like $10,000 a month, something like that. Crazy. And not like wasting it. They're not wasting money. Their business runs using agent loops that are running inside these harnesses. And they use these products in sophisticated ways that I couldn't even conceive when we were building the product ourselves. Even internally inside our own company, there are some people who have set up these kind of like multi-agent hierarchy and agent loops that looks like its own software architecture. And I often just ask these guys to come explain to the rest of the company, hey, what are you doing with these tools? Like you clearly are consuming it way over what we thought the average person in the company would do. And the single biggest differentiation between those who use agents a lot and those who don't is whether they run repetitive cron jobs, like whether you use AI as one-off tasks. You just delegate a task and then it gets done. That's like kind of using it for like deep research or like whatever, right? Like one single task versus the AI is like continuously monitoring something for you. The AI is continuously like triggering based on certain events and going and doing certain things, giving you alerts, like you set up workflows that keep running all the time. Every time you get an inbound email, like it triages, or every time there's a latency spike, it has to identify which part of the codebase caused that, it has to go and do the root cause analysis and then identify the right engineer. All these things, this is where the frontier is. And so going back to my main point, these products are not going to be used by 100 million people, but they will generate revenue that's going to be higher than the advertising revenue of Google or Meta. It's going to happen.
Completely understand what you say there. I do just want to focus in on a specific element there when you were saying like the power users, because I think one of the core numbers is actually Marc Benioff said they spend 300 million on Anthropic, which works out to be about... it'll be interesting to know from him if that 300 million came from, you know, what is the distribution across employees. So it works out to be that was on developers within Salesforce, so it's about 3.8% of developer salaries. What percent of developer salaries do you think will be spent on tokens in 24 months' time? Because that fundamentally changes the value of OpenAI and Anthropic. If it stays at 3.8%, they will not be $5 trillion companies. But if it's 100%, like Brandon at Micron said it will be in a year, they'll be $10 trillion companies.
Well, I think they can certainly be $10 trillion companies, whether it's going to be a full percent of the developer payroll today or not, because there's a lot of non-developer work that'll also be done with agents. And that's actually what we focus on for Perplexity Computer. We're not going after the developer market. We're going after anything that non-developers do. Basically your finance department or your corp dev or your sales reps or your data science teams, your research analysts. I think that's actually an even bigger market. It's not even like... think of it as like Claude Code multiplied by 10. That's the size of that market.
If I push you on developer salary spend, what percent of token spend as a portion of salary do you think we'll see in 24 months?
It's hard to say. I think the costs are going to go down. That's why it's hard to say.
You think the cost will go down? Because this is the kind of challenge that we've had. We thought when we went from chat to agent that costs would go down and token costs would go down. They've gone up.
Help me understand that and how that changes.
I think in software, you kind of want to pay for the frontier. It's kind of like if you know some engineer is awesome. If you know you have like the next Jeff Dean, would you rather hire that person and not hire five people who are medium engineers but not Jeff Dean level with the same amount of budget you have? Yes. Right. Let's say you had a million dollars. You could hire five people worth 200K or you could hire one Jeff and pay them a million. What would you do?
Yeah. So I think you would pay for the frontier. But what stays frontier keeps changing. In 12 months from now, let's say thought experiment, there is an open source model as good as Opus 4.8.
And you still have to pay for inference. You know, nothing is truly free, but it's going to be like, let's say, 10 times cheaper than Opus 4.8. And when you pair it with the right agent harness, you know, and all the connectors, GitHub, everything, all your developer workflows work fine, why would you assume that the token spend is going to be still high? It's not going to be for the same things you're doing today. It's not going to be. But there might be a different set of things you might do with the frontier that you're not conceiving today. My prediction would be software agents that are like completely autonomous software engineers. Today I think we're all using tools like Claude Code or Codex to write code, but not as literal software engineers.
There is a large sway of people that is now bearish on your frontier models, your OpenAI's and your Anthropic's, because they're realizing that you can actually do a lot with open models for a fraction of the price. What you're saying is actually that is true, but we will still pay for the frontier and so they will still...
And I think this distinction, it feels like a contradiction. It's not though. It feels like two things cannot be true simultaneously. But that's not quite the case. In fact, I would argue that the frontier is increasingly going to be a thing that very few individuals might even want. Like you could argue that after a point, it's not even interesting that AI can write software. We've normalized it, right? Let's say that's going to be the case. Instead of companies being built with like tens of thousands of software engineers unlike the past, there'll be a lot more companies with smaller software teams and each of us will be using a lot of AIs. So that's actually good for the world. We'll be seeing a lot of different businesses. We'll be seeing allocation of software labor in places that was never even possible. And whatever is a frontier is going to be things that kind of like AI is going and designing chips, AI is designing drugs, AI is figuring out how to build robots, AI is figuring out how to cure cancer. These are applications where you don't have like 10 million users. It's like a few companies, but the effect of that work will touch a lot of human lives. I think to me, that's where the frontier is headed. You could also see that from the moves that frontier labs are making. Anthropic bought a wet lab, could be for the talent, could be for the infrastructure to run like wet lab experiments, but imagine taking all those tokens and putting it in the mid-training instead of just tokens from GitHub, right? So then that's going to produce something interesting.
Don't laugh. Is there an asymptote to frontier problems to be solved? I know that sounds ridiculous, but if you are continuously on the chase for the next frontier problem, you get to cancer, you get to climate change. And my word, I hope they solve both in like heaven, that's a huge amount to solve. But if you're on the treadmill of continuously, is there an asymptote to that? Do you see?
There's no mathematical argument to there being a cap on the amount of economic value one can create with AGI or ASI-like systems. And Elon has a good argument for this, like where he says money loses all meaning in a post-AGI economy because you'll be producing an abundance of energy and labor, and fundamentally the economy is grounded to energy and labor. If you can produce an abundance of them, well, what meaning does money have? And so I don't think we run out of things to solve at the frontier. I think we're always going to be creating... like why did people even want to understand the universe? Like why did we want to understand subatomic particles, quantum physics, black hole theory, you know, the origins of the universe, like what is the purpose? But we still went ahead and did it because that's kind of what the purpose of humanity has always been, to understand the unknown. You know, David Deutsch is famous for saying this, right? Like we are the only species capable of being curious about what is already familiar. Like you can stare at a fruit and you know that it's a mango and you know exactly how it tastes, you know how it looks, you know the shape, you know what seasons it grows and stuff, but you can still look at it and ask one more question about it that you haven't asked before. Other animal species cannot. Once they have it in their mental model, what it looks like and touches and feels like, they're going to ignore it. It's no longer interesting to them.
Aravind, you mentioned about agent usage and you said if you do repetitive tasks versus one-off, say cron jobs, you know, I think Sam said it's we're going to have 24/7 AI and yeah, they've talked about a hardware product that's going to come out. Do you think we will have continuous agents running?
Yeah. And I think that's kind of why I believe the orchestration problem I talked about maximizing the token value.
Can you just help me? Sorry. When you say the orchestration problem.
Yeah. So, okay. So there are like four objectives: intelligence, accuracy, and then privacy and cost. You know, these are all competing with each other. So you could argue that you could max out on intelligence and accuracy by building giant, giant data centers and spending a lot of power to run them, and you could miss out on privacy and costs because everything will be centralized and you're going to be paying a lot. You could argue that everything can run locally and so that'll be good for privacy and cost, but may not be frontier intelligence, may not be frontier accuracy. So the solution is to figure out a sweet spot. You know, use local models when necessary, use server-side models when necessary, and orchestrate across local models and server-side models, grounded in valuable personal context. Sometimes the intelligence might already be there, but the system might not work that because the harness isn't grounded in the right set of tools, right? So build a world-class harness that can even make an okayish model appear great and be able to use the right model for the right task and the right part of the task, sub-agents, and even like utilize the compute we all have in our own devices, all that doesn't need to be always on a server. That is an orchestration from a router, an awesome router, a master orchestrator router. Now, if you do that, you can realize the vision of a 24/7 AI without people freaking out about going bankrupt because no one's going to be able to afford a 24/7 AI, frontier AI, running on the server. Imagine you turned it on and you could never switch it off unless something crazy happened. The thing that most people worry about with AI is like, oh, what if it does something crazy? But the real concern actually is the cost. Nobody's going to be able to afford it, a cron job at the fidelity of a few seconds, you know, that runs all the time. And so the bottleneck there is actually orchestration and local compute. And so I believe one needs to build a continuously learning local model that can save you on compaction context windows. So you try to preserve as much compute locally and rely on the server-side frontier only when necessary and keeps learning, keeps adapting, keeps evolving. And that model is not just a model. It's a model plus the harness plus the local chip and the compute and the ecosystem of devices it controls. That system is going to be your own intelligence. Essentially the data center moved to your local device and you get to control it. You get to own it. You don't get to worry about somebody spying on you or looking at all your tokens, very valuable personal tokens. Imagine you have very sensitive deal materials. Let's say you're doing a deal and then a frontier lab has all your tokens that you use to write a memo. Imagine somebody could hack into that server and steal your deal from you. You wouldn't want that, right?
I'm going to be honest. I think there's much more valuable things for people to steal from a London-based VC.
But yes, I can figure you're not just yet another London VC. You have like a $400 million fund last time I read it. So imagine like you're already making your moves for the $4 billion fund, right? So everyone has certain levels of sensitive stuff. And so I think that's where I believe that the 24/7 always-on agent is going to be realized by the company that wants to play the role of the orchestrator, not the model builder, not the frontier model builder, but the orchestrator. And I think that's what we want to do. Computer has been positioned explicitly as the agent orchestrator. The musicians in the orchestra are these sub-agents that utilize these different models. Think of them as the instruments and the tools, the connectors, the models. These are all the instruments and the musicians are the sub-agents and the symphony is the work and the system is the orchestra and Computer is the orchestra conductor. That's how it's been positioned. So what it orchestrates keeps evolving, right? It changes. It changes from, you know, models to files to tools to chips to devices. But it doesn't even matter. You don't care as long as it orchestrates things correctly and maximizes the token value per watt per user. If you can solve this problem, you will capture the most economic value in AI long term. Short term it might look like, oh, like this other lab's revenue is growing exponentially, this, that, but long term this is the one objective that truly matters.
Who is best positioned to do that?
I believe it's us because you have the incentive of not token maxing, you have the incentive of delivering the most value to the user. Like we, every time any part of the AI stack improves, our product improves. Since the beginning of the year, Anthropic's models have made tremendous progress. But what's also true is that our revenue has more than tripled since the beginning of the year. Tripled since the beginning of the year. And a lot of thanks to model progress made by Anthropic. And we also brought our burn down thanks to OpenAI competing with them and bringing down the cost of the same capability. And now with progress in open source and local models and local chips, we're going to move some of the inference back to the local devices and bring down the cost even more. So every time any part of the AI stack, whether it's chips, models, harnesses, any of these gets better, our system improves tremendously. And if our system improves tremendously, our users love it and they pay more. They spend more and so our business grows. So I think to your question of who's best positioned to win in that world for that objective of being an orchestrator is the one whose product or business benefits from other people's progress at any layer of the stack. And so if Jensen produces a better chip, it's great for us. If Dario produces a better model, it's great for us. If Apple produces a better device, it's great for us. And I love the fact that we are able to be a very positive player at every layer of the stack and not have to rely on any one person to win.
When we look at the different providers that we said kind of server-side versus you on device, when we look at server-side, a lot of people talk about an AI infrastructure bubble, which I think is funny, stupid, and moronic. To what extent do we have a data center supply problem today from what you see?
I think the biggest problem is actually in power. So let's break down what is a data center. Is it like that you just buy a bunch of chips from Dell or Supermicro? No, that's just one part of it. You actually have to go secure land or you have to lease something, lease a property, and you have to buy a bunch of turbines to generate power or you have to work with power suppliers, grid suppliers, and you also have to work on cooling. So there's a lot of other work you got to put in that is far, far slower. You have to get permits to do all these things. And so usually what's happening is there's a lot of lead time to doing this. And the models that are already in use today, these have been trained in the Hopper generation. So the Blackwell generation model, I think the first model that's Blackwell generation category is Mito's. And it's already scary. People are already freaking out about it. So imagine that everyone pre-trains a model on like a million or hundreds of thousands of Blackwells. Now those models are going to be far more powerful than what exists today. And then the Vera Rubins are coming next year in full capacity. Like all the data centers of Vera Rubins will be used next year. That model will be even more powerful. So I think there is a certain physical build-out time that always bottlenecks frontier capabilities. That's why there's a value in that layer. Whoever knows how to do this, puts it, puts together a bunch of GPUs and chips and networking and power and cooling and actually like orchestrating all this software layer on top and is able to convert that into frontier output tokens. That vertical integration has a lot of value. So that's why the markets are pricing infrastructure companies with a higher PE ratio.
than companies like Meta, for example. Even though Meta builds a lot of infra, it's valued as a software company. When we see like, you know, Meta's capex spend and it wanting to increase in the last few days and thinking about raising more and more money to increase capex spend, I get it with a lot of the AI providers like your OpenAI's or Anthropic's because they are making money from their AI products. For Meta, the capex spend correlates to increasing accuracy on ads, which is like a six to eight percent bump in revenue. I get it. But for the capex spend, it doesn't make sense.
Well, I believe like they are understanding what the market's saying. You know, I don't think they're dumb to not see what's being said. I think they're introducing a lot of subscription products from what I'm reading. So they're definitely going to like basically the company needs to not just be a social platform maximizing engagement and turning that into ad revenue, right? And I think that requires them to launch a lot of like agents, subscription-based products, and maybe even a cloud, Meta Cloud, that rents out servers like what Elon's doing at SpaceX. And maybe once they do that, the narrative might change, right? But to go back to my point, it might not be inconceivable that Micron, the supplier of HBM, might be more valuable than Meta in the next six to twelve months. It's already at like a trillion and Meta is like one point three to one point four trillion.
Can you help me understand that? Because memory is already a massive bottleneck. It's increased five x in price in terms of the COGS, right? But people are going, wow, Micron is fully priced at this point. Why is it not fully priced?
Because it's still the bottleneck. Whatever is the bottleneck will command the price. AMD is doing really well because CPUs became a bottleneck again. Agent loops, agent harnesses are all running on CPUs. The tokens are produced by the frontier models on GPUs. But whatever work, like let's say Claude generates a coding script that decides to download five hundred files from different websites and then, you know, munches a lot of data and transforms it in certain ways and generates a plot and then hosts it on a website that you can share with your people. All that compute is running on CPUs. Agents are using CPUs more than humans, right? And so suddenly there's a rise in enterprise CPUs and the beneficiaries of these are like Intel and AMD. So then they get to be the bottleneck. Like whoever's going to be the bottleneck will win. And so infra is the bottleneck right now because there's a lot of demand and we just don't have the supply. And so whoever supplies memory, SSDs for storage, CPU compute, suddenly these are all like interesting, like they're more important than companies that are just building data centers and not knowing how to turn that into valuable outputs.
Do you believe your Nebius and your CoreWeaves will be a sustainable multi-hundred billion dollar company in the future, or is it solving a short-term supply problem?
I certainly think they can be sustainable.
I think there are some, I don't like, look, I don't know particularly which of those is going to win. And there's also other players like Crusoe and Firebird and a bunch of companies. It's all about being resourceful. You got to take power from areas where there's a lot of natural resources and the cost to bring up the data center is pretty cheap and the time to bring up the data center is cheap and your service is reliable. Like if somebody commits to buying one hundred thousand GPUs from you, the service should be pretty good. And you should be able to secure the supply ahead of time. Plan well. And I think some companies are even innovating at the power layer. You know, generating their own power is one way to bring down the margins. And so I think there's certainly like value in that layer because it's hard to replicate work. That's how I see it. You could argue that OpenAI can do all the work that CoreWeave is doing and that's kind of what they wanted to do with Stargate. But why is CoreWeave more successful at building data centers than OpenAI?
It's hard to do. It's operationally intensive.
Yeah, operationally intensive. You got to focus. You got to like spend most of your time securing permits, like figuring out power, figuring out like bottlenecks in the supply chain here and there, and constantly plan ahead and like test all these systems carefully, deal with like random physical issues that arise in like running a data center. There's something called TCO, you know, cost of operations. You got to factor that in. So that said, I don't think there's value if you're just like a server renter, if you're just a GPU server rack renter, if you're just leasing it to different companies on certain hourly pricing rates, there's not a lot of value. You have to actually build some software on top, kind of like how AWS did. It's called Amazon Web Services, not Amazon Servers, right? So you have to have some software orchestration on top that allows you to get software margins on top of what you're doing. And I think that's why you're seeing moves like Nebius, like going for the AI model inference, like taking open-source models or hosting your models. And that's a business model of certain other companies like Fireworks and Baseten and all that, but you could imagine neocloud just going for that business.
That was exactly my question. So I just had the co-founder of Nebius on the show, and the really clear takeaway was the challenge that he has, which is there's a huge amount of money that wants just capacity and compute.
With the awareness that he needs to build a full-stack product if he wants to have a long-term sustainable business. That was the core realization for me. When I look at the inference layer, like you said, Fireworks or Baseten, how do you think that plays out? Do we have standalone hundred billion dollar companies in inference alone, or do we see that commodity?
I mean, it's just, it's all about working backwards. Like, what does it take to build a hundred billion dollar company? Assume like ten billion in revenue, exactly ten billion in revenue, thirty to forty percent gross margins, good amount of net income, good cash flow. Okay, ten billion in revenue is not that inconceivable for a company that can both do AI hosted inference and server capacity and data center build-outs very operationally well. It's all about like, you know, there are some factors beyond their control, like open-source models continuing to be awesome. If open-source models stop to actually be good, where the gap between them and the frontier is like more than twelve months or like fifteen months, eighteen months, then I don't think these companies really have a business model because they're not going to be able to host, they're only going to be able to rent capacity to OpenAI and Anthropic.
And so that's exactly what Roman at Nebius said. He said if consolidation happens and there's Anthropic and OpenAI or two or three dominant providers, that is the biggest threat.
That's correct. Yeah. And so, but you got to make a leap of faith assumption that, you know, like the models from China or NVIDIA is making good progress on their models in Llama. So there's going to be enough factors in the market to keep consolidation as an outcome from stopping from happening. But you don't control your own destiny if you're those companies. That's basically the problem.
Totally get that. Okay. So we can have standalone companies that are hundred billion dollars in inference alone. So I'm just pillaging you for your knowledge. When we look at the model selection companies like an OpenRouter or like Factory AI just released that kind of model selection or model routing product, which did very well on launch, is there hundred billion dollar companies in the model selection and routing business?
Probably not. I think you can't just be a provider of router. You have to use the router to produce something meaningful. Actually, most of the business value of OpenRouter is less than the router, even though the product is called OpenRouter. It's not routing across models there. It's actually just routing across different endpoints of the same model. So why, okay, so maybe let's ask this question. If you wanted to use Claude Opus or, I don't know, like GPT-5.5 Developer, why would you not want to just use it with your own API key versus using it inside OpenRouter? Number one argument. The single simplest argument as to why you would want to do that is model fallbacks. Sometimes your API keys might not have the rate limits, or even if you have the rate limits, there might be an error on OpenAI servers that don't guarantee you the response time you need to run your application. And OpenRouter would go and earn the, you know, they would pay for capacity for like one year ahead with the funding they have and secure the rate limits and multiple endpoints across multiple different providers of OpenAI models, be it Bedrock or Azure or OpenAI themselves. And so that routing is valuable. It's essentially an infra problem they're solving, which is reliable token supply. It's not actually, oh, like they're lowering the cost by deciding if this prompt should go to like GPT or Claude or something like that. That's not what they're actually selling to the developer. That's not actually the business model. And then for a lot of these Chinese open-source models, there's not, you probably don't want your API tokens from going to, like, let's say you don't want your API tokens going to China. And so you, and let's say you don't have the bandwidth to work with like different inference providers or verify who's good and you know who's not. You're just trusting OpenRouter to take care of all that, and then you know they're going to like supply the tokens to you. So it's routing not at the level of like, oh, like deciding which model is cheap for what task. It's more like a reliable token supply, and I think there's some value in that layer definitely, otherwise they...
wouldn't have these many users and these many trillions of tokens being routed a month, but it's not like, you know, a high gross margins business. It's the way the business model works for them is actually they would secure a discount from the model providers by guaranteeing a lot of supply, but they would still charge the user their listing price on the API, and that difference is their margins. Do you understand?
I totally get you. We spoke about bottlenecks and you said about HPM, high performance memory and Micron and the value that, you know, they have to say and what it can be. What bottleneck will we have in three years that we're not discussing today?
I think power will remain the bottleneck.
Yeah, I think it feels like that to me.
Unless something dramatically changes in the way data center buildouts happen. I actually believe that there will be a lot of resistance to building data centers. It's because people incorrectly think that data centers consume a lot of water or eat up a lot of power, which is both untrue. Satya even made the statement that it's like a can of water or something in terms of how efficient these companies are.
Do you think that's why they're putting up resistance to them? I don't. I think it's because it's a symbol of job losses, increasing wealth in...
It's a lot of things. It's a lot of things. It's a lot of apprehensions, fear about like what's going to happen, channelizing in so many different ways. Sometimes it's channelizing through hatred for wealth inequality and like wanting to tax people. Sometimes it's channeling through like concerns for the environment and like climate change. Sometimes it's channelizing in a way where you're all like, oh, like the price of the grid is going up because you guys are building all these data centers, and then or like I'm paying more for my phones and laptops now because the RAM prices have gone up because you guys went and bought all of it. So I think there's a lot of different ways in which it's getting channelized, but the common sentiment is like a pretty bad sentiment about AI.
Do you think it will be meaningful to the development of those data centers? I think right now 40 out of 100 are not being developed because of public resistance.
Yeah, so that's where the power bottleneck is, and you could see maybe certain countries see the opportunity for this and allow these model builders to go build data centers there. Elon's going to space to do that, so that's going to be an interesting experiment because there's a lot of energy from the sun that can be harnessed there, and there's a lot of natural resources in other countries. Regulations might be more friendly.
So we're still going to see data center buildouts. It might not happen in the US. But the fact that you have to solve physical problems, like you actually have to deal with the supply chain, the permits, securing power, like making sure things work and getting the lead times lower and lower. You're not solving problems like cloning some SaaS apps here, right? Or like you're building a go-to-market team or like doing better marketing against the competitor's products. Like yes, those are also hard problems, but these are like much harder problems where like you're not in full control of your destiny and you need a lot of capital and connections and like the right people, sometimes even like political help to unlock progress. And so that's why this will continue to remain the bottleneck in my opinion. And there's a lot of risk as well because if you do encounter another DeepSeek moment here where there's a vastly more efficient model that's been built with a very different vertically integrated architecture and you built out all this capacity and you're like, damn, I overbuilt. There's something far more efficient that can run on people's local devices, their MacBooks, their Windows PCs. Yeah, like you're probably freaking out then. And so you hope that...
How likely do you think that is though?
It's probably like 20%, 30% chance. The reason I think there's some possibility is that because of the export controls, so DeepSeek is not building with the Nvidia stack. They're building with the Huawei stack, and because there are export controls on not just the Nvidia GPUs but also on HPMs, these architectures that DeepSeek is building are far more memory efficient. They made innovations on the KV cache to be really small enough that you can host it on the SSDs and you don't need high bandwidth memory for inference time, and they're going to have a completely different architecture for inference, completely different architecture for storage because they're not allowed to use the 3D NANDs. So their architecture is going to look... it's not just the model architecture. The model architecture is already pretty different. They've made innovations on the attention layer. They made innovations on the training algorithm so that it doesn't consume a lot of interconnect capacity. So they've made a lot of... basically their whole stack is getting vertically integrated to their hardware and their chips and their fabs and so on, and so that's a very different bet from what America's making.
Do you think the export controls have helped or hurt us?
Jury's still out. Short term it's helping because the only reason, my belief, the only reason why there's even like a 12-month gap between open source and frontier is export controls. It's definitely helped, and definitely like companies like Anthropic lobbied very hard for it. But there is a chance that because of that they now get really good at the physical layer. And one advantage they have is they can actually build data centers a lot, a lot, a lot faster. Power is not a problem. Permits are not a problem. People are not a problem. Labor is not a problem. Expertise is not a problem. And so by forcing them to go out there and build all this, you're converting them into a far more potent competitor.
Do you think we still dramatically underestimate China's capabilities?
I think so. Because if AI is like not just digital, that's also physical AI. You got to build fabs, robots, chips and harness the energy really well and package it into local devices. I think they have a lot more advantages than America.
How important is it that we have our own TSMC in the US?
So TSMC is actually... there is a fab of TSMC in Arizona. Like not a lot of people talk about this, but TSMC is investing like $150 billion into building American fabs, and they've already invested $40 billion or something like that, $60 billion last time I checked. So there is a TSMC in Arizona that's coming up. There's also Intel, and that's why, you know, American government owns 10% of Intel. Nvidia and SoftBank own 5% each. So there is a lot of investment going into an American fab as well as TSMC is investing into its American fabs. Elon's building Terafab. Like I think people have woken up to the importance of building fabs, but this is also why China is particularly very, very competent because given the capabilities of China that we just mentioned.
Very articulately. I know it's a ridiculous question, but sort it. If I were to say to you, your job is to make sure America stays competitive, what would you do to ensure that you retained competitiveness in an increasingly strong China?
I think take physical infrastructure a lot more seriously and continue funding it, and not like have all these... you know, I wouldn't say meaningless. It's more like not propagate fake news around data centers about how data centers are polluting and contaminating water or like they're sucking up all the water, and actually be fact-driven. And so, you know, I hope our product helps there. Like you can go to Perplexity and ask any question and get fact-checked on your assumptions. But yeah, like it's very important that we educate the public about what's actually going on in a language they easily understand and not fear-monger, okay? Like not be like, oh, like all their jobs are going to go away like this, that. Like there's going to be lots of amazing companies that are going to get built with far fewer people getting multi-billion dollar, multi-hundred million valuations with like 20, 30 people and propelling like trillions of dollars of new GDP. Like let's talk about how to enable that. Let's talk about how to build that and create a more positive future together, right? Instead of, oh, like 90% of the jobs are going to be gone. Like you're all going to get screwed over by our models and like, and it's our moral duty to tell you all this like blah, blah, blah. Like that doesn't make any sense to me. Like you can't win by saying that and also like complaining about not being able to build data centers fast.
Do you think we've done a complete disservice by having the marketing message that Dario has had that all jobs are going and it's all doom and gloom?
Yeah, I think so. I mean, I think, you know, they have contradictory messages in their own different social engagement so far, where the most recent one I heard was there is no evidence that AI is taking over jobs. But so I think there needs to be a consistent communication around this. And I also think that very little is being spoken about how AIs can help you build companies in a very, very different way, like the current AI. It's already true that so many things you would hire people for, you can do it with agents. But one way of looking at it is like, oh, like what happens to all the jobs? But the other way of looking at it is like, hey, like I never had the chance to go build out a company on this idea that I've been having all this while, and maybe me and a group of friends can come together and build this. And can you guys figure out a way to give us compute credits? Or, you know, Amazon gave a lot of compute credits to a lot of startups. Like when we started Perplexity, we had like around $200,000 worth of Amazon credits and GCP credits and Azure credits. That almost like together cumulatively, this was worth like a million dollars in compute credits. Now in today's world, it's going to be like a million dollars of compute credits. And we're doing that. Like we're funding this thing called a billion-dollar build where we're giving a million dollars of compute credits to any group of people who have a credible path to building a billion-dollar company, and I want like a thousand such companies to be built.
What did you think of Sam Altman giving $2 million of tokens to YC companies initially?
I think we should do more of that. Yeah, that's the right thing to do. Like we should do a lot more of this because you want new companies to be built. And even if they're worth multi-hundred million, right, it's good. If there are thousands of them, like that's a lot of new GDP.
I spoke to Ann Bertsch before the show and she said how AI-pilled the team is for you. How big is the team today?
400 people. How big will it be in two years' time?
I don't know. It's hard to say. Maybe 800 or 1,000.
So, will companies follow the same headcount trajectory that they have always followed and we will just solve new problems? Or will they be dramatically more efficient with a much fewer number of people?
Definitely, they'll be dramatically more efficient, right? And that's why I am a believer in building a lot more efficient companies and being an example for all these companies ourselves. Like people should look at Perplexity and be like, oh, like with 400 people, you can build like a multi... like I don't know, like a $20 billion company. And so that means with like 40 people, I could probably build a billion-dollar or $2 billion, you know, and that's totally doable. Totally doable. And so for us, maybe that means is with 4,000 people, we could be worth $200 billion. We could be worth $2 trillion with like 10,000 people. You know, I think that doesn't mean it's bad for all the hundred thousand people we did not hire for a typical $2 trillion company. I would rather have those 100,000 people be split into groups of like hundred thousand groups like that, and each of those thousand groups are worth a few billion dollars. That's awesome. And I think a lot more people need to be entrepreneurial. There are people who would be bad employees in any company because they're just like difficult to work with. They don't listen to like instructions. So like they don't follow like roadmaps or they're not like easy to collaborate with. But maybe the flip side of that is those are the kind of qualities that founders typically have.
There is a population and a very large population that are not AI-native people that are not using AI to improve workflows, improve efficiency. What would you advise them?
Get started. First steps, get started and channelize your curiosity. Right? You don't need to use AIs to do your existing work. If your existing work is boring to you, you probably won't enjoy it even if you use AIs to do it.
You got a lot of heat for saying people don't like their job. So...
I didn't say... if you actually listen to my interview, I did not say that. So people want clickbait articles and they take something I said in one sentence out of context and make it into a headline.
I specifically said this. Hey, like there are a lot of people who don't enjoy their jobs. Does any... By the way, the fact that that thing went viral is not because I was completely wrong. I think a lot of people resonated with the fact that I was actually honest in saying a lot of people don't enjoy their jobs. And that has nothing to do with your economic position or standing in society. You might even be like really wealthy but doing a job that you completely don't enjoy and like destroying like the peak years of your adult life working on something that is horrible or like depressing. So like my point is that if that's you and if the reason you could never leave your job is because you were always worried if you... how would you build a company from scratch? Like there are all these things to figure out, how you have to hire a lot of people. Oh, you have to like set up an office, this, that. Like that's changed. For the first time in history, you can get started on an idea with like one or two other friends and maybe have a real genuine shot at building a billion-dollar company.
Totally get that. Everything that we've discussed today has been on the back of unprecedented demand up and to the right. We need more memory. We need more data center supply. We need on-demand and server-side. Everything's like up and to the right. Seeing some cracks in, and Uber saying, I'm not sure I'm getting the productivity gains that I thought. Microsoft aligning with them, putting a $1,500 token budget. Do you think we will have a continuous up and to the right acceptance that productivity gains are unwavering? We have to do this, or will there be falterings along the way?
