Aravind Srinivas 与 Sebastian Mallaby 在 Sohn 2025

Great. Aravind, nice to talk to you. Your company, Perplexity, is generating a perplexing amount of value, right? I read that in December, I think, you had a valuation of $9 billion. What was it you told me? $9 billion—I thought so.

Now, 5 months later, it's—I mean, it's not closed yet, but it's in the newspapers—$14 billion, right? So, 5 months, $5 billion. You're cranking out another unicorn effectively every month. This must feel pretty wild.

1. AI Creates Real Value

Yeah, it does. I think AI itself is pretty wild. If you look at all the post-money valuations of the top AI companies and startups, it's adding a ton of market cap very quickly. Of course, it's all pretty early days, and there's a healthy amount of skepticism about whether it'll last and all that stuff. But 1 thing is clear: everybody's using it. It's not like crypto, where we have to worry about what's going to happen.

Of course, there are some winners in crypto, too. But the main thing here is that it's not creating paper money and not delivering value to people. There are people who are using these apps—whether it's Perplexity, ChatGPT, or any of the other ones—pretty much on a daily basis for work, life, and many other things. Anytime a product goes down, we hear thousands of complaints. That wouldn't happen if people weren't dependent on it every day.

What does it feel like for you to be at the eye of the storm? We all use it and understand that it's an A+ technology, but you're in the middle of that whole craziness. Does it make you crazy, or do you manage to stay calm? How do you stay calm?

I hope I'm calm right now. I hope I don't come across as a crazy guy. It does drive me crazy at times, obviously.

I think the speed at which these models are gaining new capabilities, and the need to constantly adapt the product to make use of those capabilities rather than get deprecated by them, means you can't just think about the current moment and stay in the present. You have to split your time between what's likely to happen in the next 6 months or 3 months and what you're currently executing today, and juggle both things simultaneously. I think that's the main challenge.

Let's talk a bit about Perplexity. The mission, as I understand it, is to reinvent search so that it's not a search engine but an answer engine. Basically, you're taking on Google. As I understand it, you've got 300 and something people, and they've got, I think the last time I looked at the data, 300 gazillion—whatever number of people they've got. This is really David versus Goliath. How can you win in this situation?

2. Perplexity Challenges Google

By relentlessly building a better product and caring about the user—caring about giving a good answer over making money from ads—and building experiences that are truly disruptive, like giving an answer instead of giving a bunch of links. Going beyond an answer and letting users execute actions: all these things are where the legacy product feels like a legacy product, and the existing business model that's so tied to the legacy product cannot adapt so quickly.

You do catch Goliath in an innovator's dilemma. You delight the users and change their habits, and then over time you start accumulating your own capital, your own user base, and your own revenue. Hopefully, 5 years from now, you won't ask me this question about David and Goliath, because hopefully we'll be pretty big by then. But for now, it's always good to motivate the team. There are users rooting for us because they like this journey. They like this upstart challenging the 20-year-old monster.

Everybody's chugging protein shakes as fast as possible to become Goliath as well.

Yeah. Protein shakes, Diet Cokes—whatever's needed.

There must have been a moment in this journey where you built a feature into Perplexity and thought, “Aha, this is really what we're trying to do with reinventing the organization of information on the internet.” Can you share an example of some real watershed that you crossed?

3. Building The Answer Engine

I would say the first real breakthrough for us was when we started building models that could do multistep reasoning. The first launch was this simple RAG: take the top K links, take the snippets in those links, do a summarization, and reference the summary to the relevant sources through footnotes. It was a very simple idea—nothing that other people couldn't have thought of—but over time we built more depth into the product.

By RAG, you mean combining a large language model with search, correct?

What about questions that can't just be handled in the simple RAG manner? The questions could be ambiguous or complicated. You could break them down into parts, do separate searches for each of those steps, combine the results of each of those steps, and give an answer. That became our Pro Searches.

Then there are so many types of searches. Users don't just want a wall of text; they want tables, charts, and user interfaces with visually rich information that increase the density of information in the answer, so people don't feel fatigued reading long text. That requires you to build custom components and UIs for different types of queries. It requires you to understand what the user really wants and build classifiers across so many different verticals.

All this adds so much more dimension to your product than just a wrapper over a language model and a search index. Then there are research agents like Deep Research. We're building a feature called Project, which is coming out soon. It's going to look more like a likely Manus-style system, doing 30 minutes' worth of work, reading through long websites, and potentially building many web apps for you that you can schedule and build tasks and workflows around.

We're building a new browser that lets you act and provide your full personal context to the model. That way, the model isn't just giving you information from the web; it contextualizes it in whatever is relevant to you that day, allows you to execute tasks in the browser, and becomes a true personal companion while you're on the web. Any site, any task—it's always there to help you.

I want to double-click on what you just mentioned in passing in a very modest fashion. You said, “a new browser.” This guy has not only taken on Google in search; he now wants to build Google Chrome and make that better, too. It's 2 battles at once against Goliath.

4. The Omnibox Is The Battleground

It may look like that, but I think the browser is the real battle. We kind of earned our right to play that battle by building a better search product in the beginning, but the real battle is honestly the omnibox—the search box. Whoever owns that is going to get everyone's traffic at the end of the day.

Otherwise, imagine a world where I'm having AI app A, AI app B, and AI app C. I open all these different tabs, type the same prompt across all of them, and check the answer. All of that is going to go away, just as there were 20 different search engines and eventually the one that built the browser and integrated search right into the omnibox won.

I think the same thing is going to be true in AI, where the one that blends navigation, exploration, information-seeking questions and answers, tasks, and agents all into 1 single search box is going to win. You've got to start building for that moment today. It's going to take at least half a decade to make it truly work, and that's what our company is going to do in the future.

Somebody will write a book about Perplexity. It will be called The Omnibox.

Sure.

5. Google Faces The Innovator's Dilemma

You referred to this already when you mentioned the innovator's dilemma, and it's worth going deeper on that because this is the really wild thing. Google is a company that has been working on AI—in fact, leading in AI—for more than a decade. In 2012 and 2014, it bought 2 other companies. One of them was led by Jeff Hinton, who won a Nobel Prize for his contributions to AI last fall. Another was founded by likely Demis Hassabis, who also won a Nobel Prize last fall for his contributions.

These Nobel Prize winners were inside Google building AI. Google understood that AI mattered. On top of that, Google completely understood that it could be disrupted by AI. It knew all about stories like the famous Xerox Park story, where Xerox Park invented the graphical user interface, invented the mouse, invented Ethernet, and built a prototype PC. It circulated it to its own staff but never shipped a single PC to any customer outside. It let Steve Jobs take all these ideas, put them into Apple, and Apple was off to the races.

This is the classic story of the innovator's dilemma, where the incumbent company invents something but doesn't do anything with it. Google knows all about that story. So here you have this Goliath, which understands AI, understands that it can be disrupted, and understands that disruption happens because you have an innovator's dilemma. Yet you're saying they can still be caught in this trap that they saw coming. You've worked at Google. Why is it that way? Where does Google make all its money from?

Google has a diversified revenue model today. Anyone who looks at their transcripts of earnings calls can see them presenting revenue from search and ads, YouTube, cloud, and so many other sources. But the profit margins—they report 32%, or something like that, with 30% to 35% operating margins—the majority of those operating margins is coming from search advertising, which is close to 80% to 90% profitable.

So that pays for everything else. YouTube is not a high-margin business because you have to pay creators a lot. Storage is expensive. Cloud is not a high-margin business. That's why they never even went for building a cloud business compared to Amazon or Microsoft, for whom building a cloud business was a no-brainer idea, because it's way lower margins than search advertising.

Same thing with what is likely Waymo. It's going to be pretty good, but it's never going to have margins as high as search advertising. And same thing with AI, by the way. AI may never have margins of 80% to 90% like search's 10 blue-link advertising, which is okay. When you have no business, building something with positive margins at scale is a great idea, but when you have an existing business with 80% to 90% margins at the scale of the world's population, going for something with way lower margins, even though it's potentially going to kill you, puts you in a trap.

So that's basically what's happening. If that money from search advertising stops funneling all of Google's research and development efforts, it's going to be hard to pay for everything else. That's why they're letting go of a lot of people. They're trying to cut down on all wasteful spending and let go of a lot of bets. They're trying their best to adapt, but it's very hard to change things when you have hundreds of thousands of people working in a company.

It's no longer your game to win. It's your game to defend and protect, and to constantly think about 10 different effects of every action. Your brand is at risk. AI is not perfect; it's going to make mistakes. It's especially going to lead to hallucinations. It's going to say, “You have to use glue to stick cheese to your pizza.”

For Perplexity to make that mistake, nobody cares. It's just one mistake. One in 10 mistakes for us is fine. How did the startup get 9 out of 10 right? That's how people would think. For Google, one out of 10 mistakes at the scale of 10 billion daily queries is a billion mistakes a day.

Even if some percentage of that gets amplified on social media, your entire brand that you have built over 2 decades—about how Googling it means you basically got the answer, you basically got the information—completely gets broken and destroyed. You have the same brand as every other upcoming AI startup: “Yeah, it kind of works.” That's how people think about Gemini now. It kind of works. It's not like Google, where it just really works.

I think that asymmetry in brand risk, the innovator's dilemma, and disrupting your own extremely high-margin business, combined with the fact that they're so large and it takes a ton of people to sit at the same table and agree on everything to make any moves, is the reason that we even have a chance to do something here. I'm not saying everything is guaranteed to succeed for us. If that were the case, then it would be pretty obvious where all the investor money should flow.

But I think it's the first time in 2 decades that Google is extremely vulnerable.

Possibly the first time since they were founded, pretty much. I mean, more than 2 decades. It's ironic, isn't it, that a couple of people from the incumbent companies are on your cap table? They backed you at the beginning. They probably didn't realize they were backing their own nemesis.

It's a good hedge for the individuals.

6. Finding The AI Business Model

Let's talk about something that you referred to in passing before. AI is an A-plus technology, but the business model might be C-minus, C-plus-plus, or we don't know. People are trying to find the right business models.

One could argue that, even if you take OpenAI as an example, even if the product is super impressive, it's making an enormous loss and burning capital. The prospect that one day it will be able to raise prices when what is likely DeepSeek and Meta are producing open-source models that are free—how much pricing power do they have?

You can be A-plus on the technology and still not really have a business model. As you survey the AI field—and you must think about this every day—who do you think is closest to discovering a good business model, and how would you stack-rank them?

I would say that there is no business model in AI right now other than subscriptions from consumers, enterprise subscriptions, and an API, if you have some models or some kind of unique offerings that people use to build their own products. These are the 3 things that everybody is pursuing.

Some people might try ads in the product itself, but all that is still pretty early days. I would push back on this rising-prices thing. It feels a little weird to pay for an AI that just gives you an answer, because eventually you're going to say, “The open-source model is going to do it, and it's going to be cheap and free. Why would I pay for it?”

But I think it's going to be much more natural to pay for an AI that is a true assistant. It really has so much context about you. You spend some time actually training it—not exactly training a model, but telling it your preferences—and it kind of understands and gets all that. You already put in some amount of investment and time, and now it got used to you: your style, your context, and your data.

It's doing workflows and tasks for you every day. That's going to be much easier to rationalize paying for, and it's going to be something worth way more than $20 a month. It's kind of like hiring a person, except that person does so many different tasks for you. They're a jack-of-all-trades.

I think that would be the model that makes AI an extremely high-margin business, because the cost of operating that kind of assistant will go down as models get commoditized and capabilities improve. But the value it provides is only going to go up because it increasingly gets better and more reliable, keeps absorbing context again and again, and keeps learning from your past interactions. You cannot just live your life without it.

7. Execution Is The Real Moat

But I guess what we need to focus on is: what is the moat? You can have a product that is completely wonderful. People love it and get enormous value from it. You have this personal assistant; I get all that. But if the other guy is offering a similar thing and charging not very much, and then there are multiple competitors, we've got, just in the US, let's say, 4 or 5 contenders. Then there's a whole bunch in China, and that's before you count what is likely Mistral in France, the ones in the Middle East, and so forth.

It's a crowded field. Because the price is high, everybody wants at least some piece of the pie.

The pie is so big. I would say there are many people, but obsession, attention to detail, and just caring about the problem are still scarce. If that were not the case, then I wouldn't be sitting here. Our company would be done. Google would have already solved this problem. likely OpenAI would be dead because they have all the resources, all the talent, all the money, and a billion-user surface area.

Leave Google—let's say Microsoft. They don't have the innovator's dilemma. The amount of energy it takes to actually care about every single detail, handle all the corner cases where these models fail, and ship a product that is just there—no AI product is 100% reliable because these are all stochastic, nondeterministic models, so you can never be sure if something will really work—but just passing the threshold where it's already so useful and people are willing to bear with mistakes is still an art.

If that weren't the case, then, with a lot of competitors, the one with the most money would always win. But that's not the case right now. That's why people like us still exist, and there's a chance to truly win this game if you have the best execution capability.

Okay. We'll be following your journey with excitement. Great to talk to you.

Thank you.