10,000-Word Interview with Perplexity AI CEO: Advertising Is the Greatest Business Model; $20/Month Subscriptions Are Not | Z Talk

Z Talk is ZhenFund's column for sharing insights.

This past April, following a new round of funding in the tens of millions of dollars, Perplexity AI's valuation surpassed $1 billion, making it a unicorn. Founded in 2022, this AI knowledge discovery engine now has over ten million monthly active users. Jensen Huang once said in an interview that he uses Perplexity "almost every day."

CEO Aravind Srinivas was born in India in 1994. He earned his PhD from UC Berkeley and interned at OpenAI, Google, and DeepMind before leaving OpenAI to found Perplexity AI. In a recent in-depth conversation with 20VC, he shared his views on the performance boundaries of foundation models, their commoditization, and the lessons Perplexity AI has learned in building its business model.

ZhenFund actively tracks cutting-edge technological innovation. We will continue to bring you insights and deep content from the world's top entrepreneurs — stay tuned. Below is the full translated interview.

Key Takeaways

• Diminishing returns from model scaling: The performance gains from simply making models bigger will become limited. We'll need to invest more effort in data curation.
• Commoditization of foundation models: Second-tier models that are cheap enough will become interchangeable commodities. But the talent and teams creating frontier models will retain enormous value. The biggest beneficiaries of foundation model commoditization will be application-layer companies.
• Advertising is the greatest business model of the past 50 years: The $20-a-month subscription model isn't good enough. Perplexity AI's dominant revenue engine in the future will be advertising — the greatest business model of the past 50 years, with profit margins as high as 80%.

01

Falling in Love with AI Through a Machine Learning Competition

Harry Stebbings: Aravind, I've really been looking forward to this. First, thank you so much for joining me today.

Aravind Srinivas: Thanks for having me, Harry. I've watched a lot of your episodes and I'm very excited for this one.

Harry Stebbings: I want to start with your background. How did you fall in love with AI? How did you realize this was what you wanted to do and commit your career to it?

Aravind Srinivas: It was more of an accident. I was just a regular undergrad studying electrical engineering and computer science. A friend told me about this competition where you could win prize money if you did well. I needed the money because I wasn't sure I'd land an internship, so I decided to try it.

It was a machine learning competition, but I didn't even know what machine learning was. I just knew you'd get some data, find patterns in what you knew to predict what you didn't, and submit your algorithm to a server you couldn't access. It would score you against the right answers, and whoever predicted best won the prize money. So I looked at this very popular machine learning library called Scikit-learn.

I genuinely didn't know what any of these words meant — decision trees, random forests, none of it meant anything to me. I just brute-force random searched my way through it like an AI would. I ended up completing all the tasks and winning the competition.

I beat people who actually knew machine learning, which gave me a lot of confidence that this might be something I was naturally good at. I remember Sam Altman once told me — this was two or three years ago when I asked him how you know what you're naturally good at. He said, "Anything that's easy for you but hard for everyone else."

That's a great heuristic for figuring out where you might be dramatically better than others. I thought machine learning was a good thing — it wasn't called AI back then — so I started pursuing it. I took all the courses, read Christopher Bishop's Pattern Recognition and Machine Learning. I bought a used copy in India for maybe two or three dollars. I really loved it after reading it — it had mathematical rigor but was also very intuitive.

This led me to a professor at my undergraduate school who worked on reinforcement learning. I asked him to advise me. So when I actually got into AI, it was through reinforcement learning. We'd all written AIs for things like checkers or tic-tac-toe. When you played chess against a computer as a kid, you'd always wonder: how does the computer play? And people would say, it's just AI, don't worry about it. They used "AI" very loosely. But what's the real definition of AI? It's an agent that receives reward signals and optimizes for goals within an environment.

When I started studying RL, all of these frameworks suddenly made mathematical sense to me. Then at the end of the course, my advisor told me about a friend he'd known since grad school named David Silver. His startup had just been acquired by Google for roughly $500 million because they'd written a paper where the model learned to play Atari games from pixels on the screen.

"They've open-sourced the code. Why don't you take it and figure out how to play all the games simultaneously instead of just one?" If you've learned to play Pong, you should learn Breakout faster than starting from scratch. That's transfer learning — that was my first project. I loved all the ideas in DeepMind's papers. I just wanted to stay in the lab, keep reading papers, and try to implement them.

02

To Maximize Returns, You Need to Ensure Data Quality

Harry Stebbings: The first question that comes to mind — when I post about interviews, people message me, and their first question is always about diminishing returns. You put in more compute, you get better model performance. Do you think we're at a point of diminishing returns now?

Aravind Srinivas: I think it's a complicated answer. I could simply say no, and you'd think brute force still works, but that's not really the reality either. It's also not like I can come to you and say: "Hey Harry, take my $500 million, build a massive compute cluster, process trillions of tokens, and get a model better than OpenAI." That's not how it works either.

Making today's models bigger and training them on more tokens still has value, but if you want to maximize your returns, you need to put in a lot of effort to ensure data quality and so on, otherwise it's not worth it.

I know of many research labs that trained very large models on massive amounts of data and ended up with nothing to show for it. It depends heavily on what data you train on, how you mix English with other languages, code with math and all the chain-of-thought reasoning, how that interacts with scaling laws, and then how to make mixture-of-experts models more compute-efficient — all of this matters.

I think only the labs that get all these details right can benefit more from scale. And right now there are only three or four labs that can do this. Let me give you an example. When xAI released Grok-1, this open-source model's debut, Arthur tweeted that the model had a lot of redundant parameters because a 300B parameter model was worse than Mistral's 56B parameter model.

You might train a model six times larger and end up with a worse model; you might spend more money and end up with a worse model.

Harry Stebbings: So if you're talking about data curation as the central factor determining model performance. Reid Hoffman was on my show earlier this week, and he said we'll see verticalization of models — you'll use different models for different tasks. Is that what you mean?

Aravind Srinivas: Actually I think that view is flawed too. I used to think that would happen, but let me give another counterexample. Bloomberg spent a lot of money training Bloomberg GPT. They even wrote a paper about training their own foundation model. But that model was beaten by GPT-4 on all financial benchmarks.

Harry Stebbings: How do we know this isn't just a specific case? Maybe they did it wrong, or the team wasn't good enough, etc. That doesn't necessarily refute model verticalization, right?

Aravind Srinivas: What I want to ask is: where does the magic of these models come from? The way you test them isn't the way you train them. The way you prompt and interact with them in a chat window, as if talking to a real person — that's not what they were doing during training. During training they were just trained to predict the next token on the internet. Sure, they got some fine-tuning to become good at chatting, following instructions, and so on, but that's just a small fraction of the compute applied.

So what makes these models magical is their general emergent capabilities. They can perform tasks without specialized training, or quickly understand and execute tasks through simple prompting. This doesn't come from any specific domain — it's emergent from training massive neural networks.

These neural networks are remarkable. You just feed them a very diverse set of data, and they can identify the abstract skills needed to process all of it. It's this abstract skill, this abstract "intelligence," that makes these models so good in actual production use cases.

So when you say you're going to make a domain-specific model, think about how many tokens exist in that domain? Code might be the only domain that actually has a massive number of tokens. You can feed models a lot of enterprise data and say, I have internal data that no one else has. But that doesn't mean these models will gain some entirely new reasoning capability that couldn't be obtained from the internet.

Very few people understand why these models are so good at reasoning. Our understanding of this is still very limited. Is it because they were trained on math or code? If a model were trained only on textbook content, would it still gain reasoning capability? These questions don't have good answers yet.

03

When Model Reasoning Reaches a Certain Height

Will Break the $20/Month Business Model

Harry Stebbings: Do you think models are good at reasoning? I think a breakthrough in reasoning will be one of the biggest breakthrough moments in the next wave. Where do you think we are today in terms of reasoning quality? What will it take to make a breakthrough on the next wave of reasoning quality?

Aravind Srinivas: It depends on what you mean by "good at reasoning." Better than an eighth grader? I think so. Better than 75% of 12th graders? Probably. Can they win the International Math Olympiad or the International Olympiad in Informatics? No, definitely not.

So it's a spectrum. Even among humans, there are many people who are good at reasoning. I'm sure AI is now at the level of a median high school student. Can it reach the level of a median undergraduate? Sure, it seems like we're heading in that direction. Will talking to AI be like talking to great scientists like Faraday or Einstein? That height won't be reached in the short term. Some people call this kind of AI artificial superintelligence.

I think when we reach that point, it will break all these $20-a-month business models. Have you seen Nolan's The Prestige? The magicians compete with each other, and it's like one magician wants to steal a trick from Tesla about making things disappear. He's willing to pay a fortune just to learn that one trick.

I think if models become very good at reasoning, you get something like this. Instead of paying a monthly fee for outputs, you would pay for individual sessions, chats, outputs — and you would pay a lot. Harry, you're an investor. If I come to you — even without any insider information — and I'm so good at reasoning that I tell you, "Harry, this will be a truly important company in two years." This is information you might have to spend two months talking to a hundred people to get. Would you pay $10,000 for that answer?

Harry Stebbings: I'd probably pay $10 million.

Aravind Srinivas: Exactly. So even if you only pay 1% of the return, it's worth it. Someone like Demis Hassabis, they're incredibly smart. Who can give Demis advice? You can count on one hand the number of people who can do that. So if Demis feels there's an AI that can advise him, what is that AI worth? It breaks all your mental models about $20 a month. I think that's what we're looking forward to now.

If the benchmark for reasoning capability is an AI that can advise Demis, we don't have that today. But there are current AIs that can advise someone in the UK making £120,000 a year. I think we can do that. You have to clearly define what good reasoning is.

Harry Stebbings: I understand your precise definition of good reasoning. When you think about the trajectory of reasoning capability, how do you think about the timeline? Do you think it's up, flat, then up again, or a continuous gradual increase? How do you think about the trajectory and slope of reasoning improvement speed?

Aravind Srinivas: I don't think we've found the secret sauce yet. At least according to the media, they claim OpenAI has something new called Q* that tries to use its own data for self-learning and optimization to improve its intelligence level. xAI recently hired Eric Zelikman from Stanford University, who has written some papers on STaR, Self-Taught Reasoner. Basically, you let the model itself explain its own outputs. You take the correct outputs, have the model explain why they are correct, and use that for training.

You're not just training on outputs, you're training on the process used to derive those outputs. If you can do that, you're basically training a model that can think, reason, get an output, check if the output is correct, then go back, reason again, and iterate. This is what current models lack — they just give you outputs. Future models will start with an output, reason, get feedback from the world, improve their reasoning, until they converge. They will keep trying to improve the output.

I don't know when this will happen. Maybe in a year or two, maybe it will take three to four years. But I think when it is achieved, that will be the true beginning of the reasoning era. We make everything more efficient, and we will pour a lot of resources into it.

The only problem is, this race won't be played by academia like before. For reasoning, for outputs, going back to reason, building a logic, getting another output — just this process requires enormous reasoning compute.

For this you have to pay a lot of money. Even a single experiment will cost you a lot, until you find the algorithmic truth. Running this algorithm will also cost you a lot, because you need to get all the synthetic data for training.

Companies with large amounts of capital will have more advantage. If in the end there are only four or five competitors doing this, and whoever finally gets this algorithm — the first person to get it will have a huge advantage. Once you crack it, you can keep pouring in more compute power and gain a massive lead.

Harry Stebbings: We'll definitely talk about the capital required later, but I want to continue with performance and capabilities first. Why are models with memory so difficult to achieve? Everyone says memory is a challenge. I don't understand why, can you explain?

Aravind Srinivas: There are two things to think about here. What does memory mean? Is it a sufficiently long context that's practical for most use cases, or is it infinite context? For example, an AI that can remember your entire life, every detail — that's infinite memory, and we don't even have an algorithm for that yet. Then there's another kind of AI, like Gmail, which starts with enough storage to be practical and keeps expanding over time. At some point you might need to pay $10 a month. That's more like the trend we're seeing now.

Context length started at 32K, is expanding to one million, then DeepMind released two million. I think that's already good enough. At least we can choose which parts to prioritize, discard irrelevant parts, and continue using memory. That's not hard to do.

However, there's a small challenge that hasn't been fully solved. We've achieved long context, but instruction-following capability isn't good enough yet. Because you have memory, you can put a lot of content in the prompt, but with so much information, the model may hallucinate or get confused. So you need to ensure that after adding all these long-context capabilities, instruction-following ability doesn't degrade.

I don't think that's been achieved yet, which is why these models aren't good enough. For example, they still can't write a complete codebase. But I think all of this will come, it's just a matter of time.

---

Competing in Foundation Models

Almost a Guaranteed Losing Game

Harry Stebbings: Everyone is saying we're seeing commoditization of foundation models, and I'd love to hear your take. How do you think about the end state of the foundation model layer? Are they really being commoditized as people say?

Aravind Srinivas: I think today, the word "commoditized" is somewhat true. Models at the GPT-3.75 level have been commoditized. There are too many such models on the market today, some open-source, some closed-source.

I don't think GPT-4-quality models have been commoditized yet. Today people probably only have one or two choices, like Claude 3 Opus or Gemini. So if there are only two or three choices, I wouldn't call that a commodity. But will it be commoditized in the future? I think so. But when it is commoditized, will there be a 4.5 or 5 version that's even better? That remains to be seen — training is underway.

My prediction is that there will still be a great model after 4. I wouldn't say GPT-4 is much smarter than GPT-4 Turbo. It's more reliable, better, faster, cheaper, but it doesn't have the qualitative leap that 4 had over 3.5. Whether 5 can make that kind of jump over 4 will answer your question about whether these models are being commoditized.

Harry Stebbings: This isn't like the old bad business model where the core product gets obsoleted every six months.

Aravind Srinivas: But is that true? I mean, I've watched your interviews with Altman and Brad Lightcap, but does a product really become obsolete because of a model upgrade?

Harry Stebbings: I think so. Now that we have GPT-4, hasn't GPT-3 become unnecessary?

Aravind Srinivas: But your product was never the model. Maybe we should separate these two. If there are companies developing foundation models to compete with OpenAI, that is absolutely one of the worst spaces to be in.

I almost think there are five companies standing in this space now. Google, Anthropic, Meta, Mistral, and maybe after xAI's new funding round, you can include them too. But this is a very hard game to play. I really admire Mistral — despite having 10x less capital than the others, they're still in this space.

Harry Stebbings: Aren't you in the same space?

Aravind Srinivas: We do post-training on models, we don't train foundation models. We can use any model on the market today, shape them, and make them good at our product features. When you say there's a Llama 3-70B, that's just a foundation model. The first part is just trained to predict the next token; then there's the supervised fine-tuning and RLHF part, which trains them to be very good at chatting, instruction-following, summarization, translation, and all these skills.

The second part is what adds the magic to the product. Without it, you wouldn't have such a good chatbot. But the first part establishes the basic intelligence for these models.

We don't do the first part because playing the first part is almost a guaranteed losing game. Every time you finish a large training run, you burn a lot of money, you have a great model, then you watch it get destroyed on the leaderboard by the next new version. Then you have to catch up, and you have to spend more money again.

So how do you recoup all that money? You might do it through APIs. If others are just offering better models at cheaper prices and faster speeds, no one will want to use your API. That's why I think this is a hard game. It's hard not because training these models is difficult. Of course, the science and talent behind it are hard to assemble. But from an ROI and business perspective, the competition is fierce.

Harry Stebbings: Isn't this exactly what we mean by model commoditization? You reach a stage where you realize, oh no, everyone has reached this stage, we have to do it again, and again, and your model becomes obsolete.

Aravind Srinivas: Yes, exactly. That's why I think second-tier models — not the frontier ones, but ones cheap enough to build a business on top of — will ultimately become commodities.

But there will also be very intelligent frontier models that are far ahead of that second tier. I think today there are still only three or four players actually in this game.

Harry Stebbings: In the end, will it be three or four players, or one?

Aravind Srinivas: I think the answer to that really depends on who cracks bootstrap reasoning first. We touched on this earlier — models using their own outputs to reason and improve. Whoever cracks bootstrap reasoning first, if they pour all their capital into scaling that reasoning, I think it ends up being one player. But if they've been playing conservatively, then it won't be one player.

Harry Stebbings: Who do you think is most likely to be that player?

Aravind Srinivas: Most likely OpenAI or Anthropic. I could make a strong case for both. OpenAI because they're so far ahead. Anthropic because they're a fundamentally superior algorithmic company — they reached OpenAI's level with less capital than OpenAI, they're better at post-training and so on. On the other hand, OpenAI has the advantage in capital and speed.

So it's a question of what matters more? Smart minds with a certain amount of capital, or good minds with strong aggression and massive capital. If it's the latter, that's OpenAI. If it's the former, that's Anthropic.

Harry Stebbings: If OpenAI hadn't partnered with Microsoft, would you pick Anthropic?

Aravind Srinivas: Anthropic is also partnered with Amazon and Google, so they're all well-funded. xAI has money and people too, but they're way behind on the timeline.

Harry Stebbings: I think the major cloud providers will realize they need to acquire these models in some form, they'll keep their core cash-cow cloud hosting business, but they'll acquire these models and add them as free features to what they already offer. You'll see Anthropic, Cohere, Adept getting acquired by these major cloud providers. Do you agree with my prediction for the next three to five years?

Aravind Srinivas: I don't think so.

Harry Stebbings: Why?

Aravind Srinivas: I have no prediction for Cohere, but I think for companies like OpenAI and Anthropic, their value isn't in the models they possess. That's a very superficial valuation.

I think the second-level valuation is that their value lies in everything that goes into building these models: the people who have all the tacit knowledge needed to train these frontier models, who can innovate algorithmically, and the compute they've accumulated. That's why these companies are valued so highly, even though their revenue and valuations don't seem to make sense. The point isn't just the output, but also the machine behind the output. When we think about valuation, we're thinking about how easy it is to reassemble the whole system.

When you say models are being commoditized, so OpenAI and Anthropic aren't that valuable, I disagree. Because these are the people who will produce the next model. Are these people being commoditized? No. In fact, it's becoming the opposite of a commodity. Companies are spending a lot of money trying to keep them. The knowledge stays with them because people aren't publishing openly anymore.

I was recently hanging out with a really great researcher, and he even joked that the best research is research that hasn't been published yet. There's nothing left to read in the archives. Even those people at Stanford who wrote all these reasoning papers — now Elon Musk is paying him a lot to work, and he's not publishing papers anymore.

That's what's happening. The value isn't in the model, the value is in the people who produce the model. That's why I feel these companies are valued highly, have a lot of negotiating leverage, and won't be acquired.

These people don't want to work at big companies, and big companies need their output to stay in business. Microsoft needs GPT to drive sales and make Azure the number one cloud service. AWS needs to maintain its leadership in cloud. I don't think OpenAI and Anthropic will be acquired.

Now on the flip side, the models might not produce any breakthroughs. Scientifically, you can't just keep stuffing more and more tokens into the model and keep seeing good results. If, say, a year from now OpenAI doesn't have a better model, then their advantage disappears. Because you have to produce something new, and if they can't, the value drops.

I think both of these things could be true. But I believe these people will still make breakthroughs, so I have a different prediction. But time will tell who's right.

05

Every Company That Raises Money

Will Eventually Have to Build Its Own Business Model

Harry Stebbings: We just touched on access to capital, and obviously OpenAI has slightly more than Anthropic. What struck me was, I heard that Mistral AI's new funding round is roughly equivalent to Microsoft's free cash flow in 30 hours. Microsoft generates $330 million in free cash flow per day. The relative amount of money you raise is trivial compared to Microsoft's free cash flow. In a world like that, how do you compete?

Aravind Srinivas: That's why you have to build a business model. First, let's separate two things. If Microsoft generates that much cash flow, why can't they just poach all the OpenAI scientists or Mistral scientists to work for them? They could use that money and get 10 of those people to come work for them. Say I'll pay you a lot, no need to work at OpenAI anymore, just build AI here at Microsoft, whatever GPU I give OpenAI, I'll give directly to you. But that's not happening, right? There's a reason for that.

People want to work with other best people. So getting just one person isn't enough — you need the whole team. That's why when the whole board drama happened, everyone was joking that Microsoft acquired OpenAI for cheap by poaching the entire team.

I think that's the crux of the problem. Cash flow doesn't change the dependency problem. If they could get these models from someone other than these two companies, things would be very different. They could just take open-source models and make the same money with less spend. That's bad news for foundation models.

As for the way out? I think you have to build your own business model. Fundamentally, every company that raises money will eventually have to build its own business model, or hope their algorithmic advantage lasts forever. I'll bet on the people who are seriously building business models. OpenAI is making money with its product. They're at $2 billion in annual revenue — that's higher than Snowflake. They're not as capital-efficient as Snowflake, but in terms of recurring revenue, they're in the same ballpark as Snowflake, and growing faster.

So this shows that if you don't just settle for training these models, but also go to market with a product and generate revenue from it, you have a shot at independence and self-sufficiency.

Harry Stebbings: Are you focused on building a business model today? You said we're not going to rely on £20 per user anymore, but you're currently at the stage of charging £20 per user per month.

Aravind Srinivas: I think that business is actually not that great — the margins aren't high enough. If you could get to YouTube's level, then sure. Netflix, YouTube with 50 million to 100 million paying users, that's absolutely a great business.

I don't think we're at a stage yet where these AIs are so important to people's lives that 100 million people subscribe to them.

If they could reach that point, if you could develop a product that's not just AI but has many other features, people pay a substantial monthly fee for it, and retention is close to 100%, that would be an amazing business. I think we'll try to do that too. But all these great subscription services advertise for a reason — for margins.

Whatever we say about Google, the greatest business model of the past 50 years has been click-based advertising. It's an incredibly good business model with 80% margins.

Harry Stebbings: When you and the team discussed adding advertising as a revenue engine, what did those internal discussions look like? How did the conversation go, and how did you arrive at the final decision?

Aravind Srinivas: Larry Page and Sergey Brin's paper on PageRank said that advertising is fundamentally incompatible with providing good search results to users. I've read books that said they delayed introducing ads as long as possible until they succumbed to investor pressure. We said, let's be practical — this is the highest-margin business model ever, but we don't need to chase Google's level of margins.

You don't have to aim for 80% margins. As long as you can get a reasonably high-margin business without betraying user expectations. Be happy, but don't be greedy.

How do you advertise without destroying search results? If you can guarantee that, I think it's a great idea worth exploring. That's why we have other places to show ads, like the "Discover" feature and so on — it has a bunch of interesting posts for people to read every day. This will eventually become an infinite scroll feed, the format Instagram and TikTok use for ads.

Advertising is a great business model, and it's great when what you're offering is relevant to the user. Not a single person has told me Instagram ads are terrible. The key is relevance. If you can nail personalization and relevance, the ad experience is excellent.

Harry Stebbings: Do you think you've nailed relevance?

Aravind Srinivas: Not yet, but we hope to. If we did, I think our valuation would be higher. It's a chicken-and-egg problem. You only get the code when you have a lot of users. So advertising is interesting — it can't work well when you don't have enough users. When you do have a lot of users, if you get all the details right, it can work really well.

I spoke with Marc Andreessen a few months ago, and he told me there are three tiers of advertising. At the top is Google, in the middle is Meta. Google benefits from everyone else's advertising. Because ultimately, once you discover a brand, you go to Google and click on their link.

Below that are companies like Twitter, Reddit, and Snap. He said the gap between these tiers is massive. It's like the peak of a mountain versus the base — advertising is currently dominated by companies like Google and Meta.

My view is that if we can correct the fundamental mistake Google made early in its journey — not over-relying on a single revenue source, but instead diversifying across subscriptions, advertising, APIs, enterprise services, and more to generate sufficient revenue — I think we have a shot at building something where shareholder and user interests are more aligned.

Jeff Bezos once quoted a line that shareholders and users should gradually converge in their interests. If that's not happening, it's not a customer-centric business. Google made this mistake because they couldn't align user and shareholder interests. Wall Street loves when Google serves more ads, but users hate it.

06

The Biggest Beneficiaries of Foundation Model Commoditization

Will Be AI Application-Layer Companies

Harry Stebbings: You mentioned OpenAI's revenue hit $2 billion, with most of it coming from enterprise — their enterprise business is doing quite well. When did Perplexity decide it should build its own enterprise division?

Aravind Srinivas: The first observation that led us to build this division was: what tool do enterprises use most today? Google. You use it at work every day for search. All the data you search through is your company's internal data, but nobody cares because you need it, you can't live without it. You pay with your time and your data.

But in an AI-native search world, things change — people are constantly worried about data leaking to AI. If data leaks to a traditional search engine, they don't care. But if the search engine incorporates a lot of AI, they get concerned.

So we said, okay, if you want to use Perplexity at work but your employer won't allow it, we'll solve that for you. We'll offer an Enterprise Pro version with compliance, security, and data governance features — essentially the same product with all these security capabilities built in. That's our Enterprise Pro.

But that's just the beginning. You need more features. These features need to be more enterprise-specific, not just consumer-oriented. That's what we're going to build. We want to build it in a differentiated way and rethink what internal search means.

For example, instead of just integrating into enterprise tools like Slack or Notion, we can think more deeply about fundamental problems like ranking. Why is this harder for enterprises than for consumers?

If we can build a unified interface where all proprietary data, external data, internal data, and all different models — open-source and closed-source — converge on one platform, where you can transform your outputs into easily readable pages, organize them into knowledge bases, and index them yourself — that would be a great enterprise product. I think we'll push in this direction. Not saying we'll definitely succeed, but we'll try.

Harry Stebbings: That's incredibly admirable. Are you nervous about developing enterprise products, considering these GTM strategies are so different? Enterprise is a massive, elusive beast — it's a challenge. You also mentioned the size of OpenAI's sales team. How do you think about building the enterprise GTM function? Will people buy Perplexity Enterprise, or OpenAI Enterprise, or both?

Aravind Srinivas: My sense is that AI is still so early that nobody is locked in or loyal to any particular AI company — they don't even have lock-in effects yet. For example, having your data run on just one tool. I haven't even gotten into questions like why migrating from Snowflake to Databricks is hard. Because the SQL formats themselves are so different. Once you've written all your SQL queries in one format, it's hard to change. That's not the case in AI. Custom prompts you wrote for ChatGPT can easily transfer to Perplexity.

I think enterprises are still willing to experiment and try different tools. That said, without differentiation, the bigger brand with the bigger team will have the advantage early on.

But is the game over? No, the game has just begun. I think this is entirely about packaging. If most of your product's value is the model's value — if the value lives only in the model and everything you build around it isn't specialized — you'll lose.

But if you build enough value around the model that it's hard to replicate without coordinating a bunch of other hard-to-achieve engineering inputs — not just LLM-based, but involving many human factors — then we can hardly imagine a world without that value. People will want it. That's the specific value of search.

Why are Google's AI Overviews bad? They have the world's greatest index and the world's best model, but it's still not good enough. Or why do many people still think ChatGPT's browsing is inferior to Perplexity's, despite all their updates over the past year?

Harry Stebbings: Why is Perplexity's browsing better than ChatGPT's?

Aravind Srinivas: I think it's a lot of small details. I strongly believe that people who can coordinate models and data sources, build excellent user experiences, and keep innovating here will survive this packaging war. It's going to be hard — you start a company, and everyone constantly worries you'll go out of business.

But as you accumulate users, as you figure out the business, I think the biggest beneficiaries of foundation model commoditization will be application-layer companies.

Harry Stebbings: Why do you say that?

Aravind Srinivas: If models get commoditized, model prices drop. And companies that directly use these models to reach users, that leverage these models — as long as they package them into great product experiences, create practical value, and build direct relationships with customers — they'll have a bigger advantage. Because they can sell a commodity at a premium price.

That's a great business. If models get commoditized, I'll be happy. If they don't, I still want to figure out how to benefit from it. That's why this is a great company despite all the difficulties.

This isn't something where you just hire a senior VP of Product from Twitter or Meta and have them set your product strategy for you. It's not easy. They don't have the mental judgment for what happens when the next AI model gets better, how to rethink the entire product strategy. Similarly, it's not about hiring great AI talent and having them build the product, because they'll always think the model is what matters and keep trying to do everything through the model.

What you need is the perfect combination of design, product, AI, and search. That combination isn't easy. That's why we're able to do things others can't.

07

Advertising Will Be Perplexity's Dominant Profit Engine in the Future

Harry Stebbings: Were there any surprises during your fundraising process?

Aravind Srinivas: The fundraising process was brutal. I think most people probably assumed that if you're in AI, people will just write you term sheets without any due diligence.

It was actually quite difficult. Everyone asks the sharp questions you see on Twitter. "What if OpenAI does this?" "What if Google does this?" "Why don't they just stop giving you models?" "How are you going to build your own model?" "How are you going to build a really good search index?" "How are you going to compete in enterprise sales?"

These are questions everyone asks, and you don't have a clear read on the future. You have to give them strong arguments, but at the end of the day, they're just arguments — not actual things.

One factor working in our favor: we had great execution. We've been around for less than two years, but relative to our team size and funding, we've shipped a lot.

Harry Stebbings: Of the capital you've raised, how much goes to compute?

Aravind Srinivas: Most of it.

Harry Stebbings: Like 50%? Or 75%?

Aravind Srinivas: Let me say two things. First, we haven't spent that much money. It's not that most of what we raised went to compute — it's that of what we did spend, most went to compute. Whether we're buying GPUs and serving or post-training models, or paying for APIs from companies like Anthropic or OpenAI, that's fine.

That's why we don't train our own foundation models. Because if we did that too, the money would be gone already. The way it works is, you have to pay three years in advance to get a large compute cluster. Not all the money at once, but you have to commit to three years to get thousands of GPUs if you want to compete in this game.

On the other hand, we benefit from model commoditization. We don't do our own foundation models, so all our remaining money goes to user acquisition. And user acquisition isn't just marketing — it's actually more like Amazon Prime: offering a lot of great features at an amazing price, retaining you through excellent product execution, and ultimately building a large enough user base and brand loyalty.

That's the business model we're pursuing. In such a world, the power of advertising can be very strong.

Harry Stebbings: Every business has a core profit engine. They might have ancillary businesses, but usually one dominates. When you look out five years, what do you think Perplexity's dominant profit engine will be? Consumer subscriptions? Advertising? Enterprise?

Aravind Srinivas: I'd predict advertising. If we can crack the code on ads, it's advertising. If we don't crack it, if our users don't grow to that level, or if we grow but can't figure out how to do ads well, then it'll be one of the other two. Either way, we can be profitable. I think with advertising, we can do quite well.

Then you might ask me, hey Aravind, why do you even care about profits? Sam Altman doesn't care. But he doesn't care because he's not interested in product as a business — he's trying to build AGI. He's already said in a public interview that even if we spend $50 billion on AGI, it's fine. So that's a different company. That's why we shouldn't be viewed as an OpenAI competitor at all. We're not an AGI lab.

You could say Perplexity and ChatGPT are products in a similar space, competing for some mindshare and users. But even then, in two years it'll be very clear — you won't keep asking, what's the difference between Perplexity and ChatGPT? You ask that today, but in two years, I don't think you will. If it's still like that by then, one of us is just copying the other.

Harry Stebbings: What do you think is the best question you've never been asked?

Aravind Srinivas: I think someone asked me, why are you doing what you're doing? This is something you don't even know yourself. I think a lot of people give made-up answers, like, oh, I had an existential crisis, I'm going to save humanity — I've seen entrepreneurs answer like that.

The reality for me is, you look up to some people, you want to become them, and you try to shape your career based on your role models. But in the end you find yourself genuinely liking certain things, and you shape it in your own style. At least that's how it was for me.

I've always been a Larry Page fan. I've always wanted to do something at that scale of ambition. Though that's not why we built a search engine. We started from something completely different. I don't actually have a clear answer to this question, but I really like it because it's something worth asking yourself repeatedly.

Like "Why are you doing this?" Jobs had this idea, right? If you internalize death, normalize death, and ask yourself every morning while standing in front of the mirror, "If today were my last day, would I still do this?" If the answer to that is yes, then go all in today.

If the answer to that is frequently no, you really need to rethink your life priorities. For me, at Perplexity, every day is yes. Despite the pain, the stress, the toll on my mind and body — I think it's worth it.

Harry Stebbings: You still look very young, so don't worry, it's not aging you. Everything's fine.

Aravind Srinivas: Thanks. I've cleverly hidden my gray hairs.

08 Rapid Fire

Harry Stebbings: In the past 12 months, what have you changed your mind about the most?

Aravind Srinivas: My long-term view of people. I've seen some people who don't change efficiently and quickly, but given enough time, they can genuinely transform themselves. This is something I didn't have the right attitude about initially — I always thought the people who act immediately are the best, but you know, different people have different ways of showing their true talent.

Harry Stebbings: What do you think is the biggest misconception in AI today?

Aravind Srinivas: Short-term thinking. Anytime someone releases an update, everyone thinks another company is finished — usually the crowd on Twitter. Even among well-informed people, the biggest misconception is that because most people in the world aren't using chatbots, this is just a bubble. But they'd be surprised to find it's not a bubble, it's not overhyped — it's actually underhyped.

When you apply it to familiar workflows and forms, it makes a huge difference. The chat UI is a new UI. We're not used to it yet. We're all used to WhatsApp and Signal and so on, but that's different, that's more of a messaging service. On the other hand, Word, Docs, Gmail, Google Search — or not even those specific products, but more the form factors, the UIs — you're deeply familiar with them. When AI is presented to you in a workflow that feels so natural, it will have massive impact. That hasn't really happened yet.

Harry Stebbings: Have you looked at WhatsApp's integration features?

Aravind Srinivas: I have, and it's not the right approach.

Harry Stebbings: Why?

Aravind Srinivas: I use WhatsApp not to search for things, but to message people or reply to... I usually have 20 to 30 WhatsApp notifications, and by the time I get through them, I just want to leave the app. I'm not clicking on the WhatsApp icon to search for something.

Same with Instagram. I'm just there to look at pretty pictures, not to search for who won the NBA championship. User intent when opening an app matters. That's also why their repeated attempts to launch "Stories" and "Reels" failed. Initially, "Stories" was a separate app trying to copy Snapchat. That didn't work, then they tried many different variants. What eventually worked was the bubbles at the top. That worked because it was built on existing user flows — users were already opening Instagram to check what others were up to.

You have to seriously consider not just why you're adding a feature, but what is the existing user intent in your app? How do you ensure the new feature ties into that existing intent? This is extremely important.

Harry Stebbings: What's your vision for the future of browsers?

Aravind Srinivas: You could reimagine the browser. There's a reason we haven't done a browser. I don't think browsers get disrupted because users get answers instead of links. People still want to browse and go to a new website, enter a specific site, input details, fill out forms and so on. None of that gets disrupted by traditional chat interfaces. Just because you can type things into Perplexity's search box, say the integration is complete, I don't think you make the browser better or anything. It becomes more efficient, but you still need a lot of traditional browser functionality.

What will change, though, is you open a browser and just say, start playing Aravind's podcast, and it goes to Riverside, auto-fills your login — that would be magical, that would change everything. Or on Amazon, buy me this thing.

Then you can go further and ask, what's the future of operating systems? What's the future of Mac? Windows? The browser is also an operating system, right?

Harry Stebbings: What do you think is the future of operating systems?

Aravind Srinivas: Movies like Her could become real. I'm not talking about voice, but the operating system itself could be AI — a fully native AI operating system, not organized in traditional ways, where you just talk to it and it works for you. That's an amazing vision. And that's exactly what hasn't been achieved yet. GPT-4 can't do it.

Harry Stebbings: What do you think is the hardest part of your job right now that nobody knows about?

Aravind Srinivas: Constantly dealing with contradictions. The brain isn't good at handling contradictions — it actually exhausts us. When we can't reach consensus on something, yet startup CEOs face contradictions all the time. Should you take risks or double down on what you have? Should you move fast or build the company in a scalable way? Is it time to try this feature just because it's not what your competitors would do, or keep doing what you're good at even though your competitors are doing the same thing? You have to constantly navigate these contradictions across different dimensions, and it's exhausting.

Harry Stebbings: Second to last question. You know, we as investors analyze in advance why a company might fail. If you were to write down why Perplexity might fail today, what would it be? Compute issues? Google innovating and beating you?

Aravind Srinivas: Poor execution. Competitors don't beat startups — startups beat themselves. It's not that Google Drive beat Dropbox. People say that, but Dropbox could have built a great enterprise business, they just didn't move fast enough compared to others. Startups beat themselves.

So if we were to write our own cause of failure, it would be the CEO not being decisive enough, the company not executing well, lack of focus, inefficient capital allocation. It largely comes down to decisions made, correctness of decisions, speed of decisions and execution, and whether we stay focused. If we can't consistently do these things, then I think we'll fail.

Harry Stebbings: Last question. It's 2034. What do you most hope Perplexity will be? If we do another episode then, what will the company look like?

Aravind Srinivas: That's a good question. I just hope it becomes the facts and knowledge assistant you can't live without. You could ask me, in 10 years, will people still want facts? We always have to ask: what will still be true in 10 years? If you commit to that, then you're doing the right thing.

I feel that even in a world where AI dominates and human autonomy is reduced, people will still want to know what's true and what isn't. We're working on that. If we can become the go-to assistant for facts, accurate information, and knowledge, then I think we'll be doing very well in 10 years.

Harry Stebbings: Aravind, I've really enjoyed this conversation. Thank you for your time.

Aravind Srinivas: Thank you, Harry. This was great.

Translated by Stone

Edited by Wendi

Recommended Reading