Perplexity CEO on NVIDIA and Who Wins the AI Race

[00:00:00] People here in this conference said we were the most likely company to fail. Okay, let's address that. Yeah, yeah, yeah. It was this time a year ago where there was a vote and they said, it wasn't to fail, they said they wanted to short you guys. Isn't it the same thing? I suppose, yeah. Yeah. Well, here I am sitting here still and we're doing better than most companies ever performing. People would have lost money on that trade. Yeah, and I would bet the same this year. If you guys short us, you're going to lose more money next year. We've obviously heard from a lot of people today,

[00:00:27] such a seismic event. And you, Arvind, I want to kind of just bring up the tweet that you put out a couple of hours ago, basically after Jensen announced that he had resigned because you're someone who obviously, you know, they're an investor in Perplexity, but you know, you're someone who also has dealt with him a lot from business side and, you know, I think are quite close to him. So, you know, the tweet that you put out, you know, thank you, Jensen Huang, for 33 years of bold leadership. You built a small startup into America's most valuable company with relentless

[00:00:56] improvement. Your grit and perseverance should always be remembered. Your personal generosity with time and advice to founders like me has secured a bright future for America and AI. I mean, can you kind of give us a sense as to like what you're thinking today and, you know, what's I imagine kind of a chaotic and unusual day in the world of AI? Yeah. So perplexity and Nvidia have been working together for many years now. Are they one of your largest investors?

[00:01:22] They are one of our largest investors, yes. And it's one of the greatest technology companies in history of America. And we're, you know, everybody here is using Nvidia compute, Nvidia GPUs, Nvidia libraries. And, you know, before any of this controversy, like, there's no doubt about the history of innovation that the company has done. And Jensen is one of the greatest founders ever, right? And part of that is,

[00:01:53] his relentless grit, determination, constantly looking to build the next thing, not being satisfied with where he is. And those are all great traits for any founder to make it. And just like me, I'm sure every founder here wants to build a trillion dollar company. And he's done it from, you know, having a breakfast or lunch with his co-founders at Denny's to actually like building the world's most valuable

[00:02:19] company. That's the sort of story that everybody gets inspired by. And like, I feel like in, amidst all this noise, we should remember to, you know, take the right lesson of like what relentlessness, grit, determination, and long-term thinking can actually help us build. Yeah. I mean, can you talk more specifically about you guys and what Jensen was able to kind of bring to Perplexity to build you guys to where you are today? So we collaborate with them a lot of inference libraries, you know, starting from TensorRT

[00:02:48] and then Nvidia Dynamo. A lot of our inference engineers have collaborated with inference engineers and Nvidia, open sourced a lot of things together. Obviously, like, we've been helped working with them on the NemoTron as part of their coalition program with a few other labs and NemoTron 3.5 VO. We contributed a lot of like mid-training data to that. And we're going to continue collaborating with them on NemoTron 4. We've had pretty good conversations with not just Jensen, but the whole

[00:03:17] executive team there and including Colette. And we're going to continue working on a lot more projects. Have you had the opportunity today? I mean, it's only been a couple of hours to kind of check in with the team there and see how things are functioning at this company. Yeah, I had a meeting with them even today. So things are going as usual and, you know, business has to go on. And even now, like the best GPUs to run your inference is Nvidia

[00:03:42] and best GPUs to run your training is Nvidia. And none of that stops. So we just have to continue building the future. Yeah. And last question on this. I mean, have they said anything to you as a major partner to kind of let you know what continuity looks like and then kind of let you know that like while we deal with this fairly insane, unprecedented situation, like the relationship you have with Nvidia will be unchanged in many ways? They've built a phenomenal foundation. So

[00:04:07] their founder has set them up so well for the future. So we need to continue expanding on that. And that's what we are focused on. So they are. Okay. And then last question. I mean, have you at all had a chance to talk to Jensen in the last couple of days? Not really. Not really. Yeah. Not in the block going on there. Okay. So let's talk about perplexity. You know, I remember first coming across you guys, actually a colleague of mine at the journal who maybe is here tonight, you know,

[00:04:33] kind of introduced us all to this idea of an answers engine. It's not a search engine. It's basically you guys were the first ones to use LLMs as a way to get the answers to the questions you have that previously were like, whatever, the 20 blue links on a page. And that kind of confined you guys, at least in my head to being a consumer company. You guys were consumer in the way that Google is the greatest consumer internet company maybe ever made. Is that a fair way to describe what

[00:05:00] perplexity is today? Are you still a consumer AI focused company? Perplexity is for anybody who wants world class research and accurate AI. That is everybody using Microsoft Word is Word a consumer product or an enterprise product. Microsoft thinks so. Well, that's the thing. Every great consumer product, when it goes up market in the enterprise, it has this advantage that it's already known brand.

[00:05:27] And it solves a real practical utility problem. What perplexity solved is the utility of like accurate AI. When we started, we had a bunch of investors who are now working at like big tech companies actually, who actually told me that it's a bad idea to give accurate answers because people actually like it if AI makes mistakes and AI hallucinates because people actually find that entertaining. And that was true. A lot of how the original... Your investors told you AI is better than what it's supposed to show up.

[00:05:55] One of my investors, not one of my investors, is very well regarded, told me that. Obviously, one of his arguments was that it's actually a good idea to make AI hallucinate because it's entertaining to watch AI make mistakes and share that and screenshot it and it goes viral on Twitter. And that'll give you a lot of new users. Okay. And obviously, we did the opposite of that and we've

[00:06:21] been pretty good too, right? And I think fundamentally, the problem we really solved is getting AI to be accurate. And obviously, that's immediately realizable in the form of an answer engine. But you can keep building on that accuracy foundations for deep research, browsing agents, every form of agenting action is searched. You're searching for the right tool. You're searching for the right file. You're searching for the paragraph within the file. You're figuring out which files to use in a big repository

[00:06:51] of code. You're figuring out how to like orchestrate different models, different connectors, how to put it all together to a final report or an answer or a web app or an artifact. Each of these individual actions is picking something in a large search space. So with good accurate AI, you can build the best possible agents. And building safe agents is also the same problem. You just have to be accurate. And so we're

[00:07:17] building on the foundations we started in 2022. And that's allowed us to not just build an answer engine. But go beyond that, to building a deep research experience that's world class, state of the art, browsing agents last year, sandbox cloud based agents this year, and most recently, persistent co workers in the form of teammates. Yeah. When you look at the revenue for Perplexity and like the mix, tell me what I'm wrong here. But you guys sort of began with subscription based, right? People were

[00:07:44] paying for like the pro version of Perplexity, which uses, you know, multiple models. And then you guys, I think do some sort of token usage based products now. And I imagine that's kind of part of your shift more towards enterprise customers, is that you kind of are token based. Yes. A big part of our revenue growth this year has been coming from consumption based or token based. And we charge for usage basis on our computer agent. And every subscription plan gets a certain amount of bootstrap credits.

[00:08:14] Mm hmm. On the $200 a month plan, everybody gets a certain number of credits every month. And so and we have a quota system for the $20 a month plan as well. But the demand for actually automating a lot of your workflows in an enterprise or a business context is very high that people run out of those credits pretty quickly and are happy to pay for credit based like usage. A lot of usage comes from repeat

[00:08:40] automations. It's not coming from the same user going again and again and chatting one more time and adding to their credit consumption. It's actually coming from them automating a lot of their workflows that they would have otherwise hired a lot of people for. Analysts, SEO, marketing, growth marketers, ad campaign management, chief of staff or admin like usage, corporate development strategy, competitor tracking. These are all like repeat automations that you once set it up. And your whole

[00:09:09] business relies on this. So the revenue is sticky, the net dollar retention is very high. And our job is to actually keep lowering the cost for doing the same task that was more expensive six months ago by continually orchestrating and routing to the best possible model for the job and lowering the price. And we often see that that takes a hit on our usage based annualized revenue in the short term. But the usage expands enough and then total revenue keeps going up. What would you say, I mean,

[00:09:34] what is the main use right now? Like the most amount of use of perplexities tools coming from what? Research, monitoring, automations, almost all forms of knowledge work outside of coding. We have pretty much a really good market fit because all of it comes down to like orchestration, routing and accuracy. And our harness is really good for that. Yeah. And some analysis pointed out like yesterday that

[00:09:59] increasing percentage of their AI usage is around persistent agents, right? That they can delegate to like a slack coworker and perplexity has the best harness and product experience around that. And because that's because we really own the underlying infrastructure powering these agents, we build our own sandboxes, we build our own model routing and we host our own models when we delegate to like open weight models that are fine tuned. And so this is the way in which like we get to be

[00:10:27] the orchestrator of all possible knowledge work. Yeah. I saw, you know, you had an exchange with Dylan Patel of Semi-Analysis about this because he was making this case that the reason he loves using computer, I believe, or our teamwork teammates, as a harness versus just because I believe the underlying model for them is Anthropics models. But he just said straight up, which is very Dylan, like Anthropics is not good at building product. Maybe the technology itself is good, but like they are,

[00:10:54] when it came to actually developing a useful harness, what you guys have done is so much more effective for what they need. I mean, do you think that's the role for companies like yours now is to just be so much better at building harnesses that there's inherent value just in that and not have to basically let the foundation model companies build the models, but they just suck at harnesses and product and they're just not going to be threatening to you? We believe in a world of multiple models, right?

[00:11:19] Just as Anthropics making great progress, so is Google. Carbon is pretty much at faber quality and OpenAI is behind today, but that's not going to be the case for too many weeks or months. And so as this competition of the model layer arises, the value accrues to the one who can orchestrate all of these in a really solid harness that stays model agnostic, lightweight, so not a lot of hard coding. So you're not building around the limitations of models. You're actually benefiting from the

[00:11:48] progress in the models and world-class infrastructure. You need to own your own sandboxes. You need to keep it safe. Enterprises need to be able to trust their agents to be running on your platform. It needs to respect role-based access. Agents cannot go rogue inside your sandbox. So all that stuff matters tremendously. Connectors fail a lot in most other agent products, not on ours. So solid connector infrastructure and of course our search advantages, being able to spin up browsers on the

[00:12:15] sandbox. These are all things outside the model that you have to be extremely good at for providing a reliable experience to a demanding organization like Semi Analysis. Yeah. So let's actually talk about the progression of your products because I find it very interesting and I'm sure I'm going to mess up the timelines so just cut in when I get it wrong. But you obviously have your answers engine and then I believe last year you build Comet, which is a browser. And I mean I'd love to know an update on usage

[00:12:40] and your commitment to continuing that but I think through Comet you guys end up developing computer, which is your kind of cloud-based agent. Yeah. Yeah. And then from there teammates come through that. That's right. So we really recognize that you need to be where users are, which is why we built a browsing agent because most people spend a lot of time on the browser. Sure. But what we realized is people like this conversational interface using things from their phone, being able to delegate their tasks

[00:13:08] from the phone itself. And for that you actually just have to put the browser on the cloud. Mm-hmm. Except like the infrastructure to run everything inside a sandbox and have a coding agent orchestrate the whole thing did not exist when we first launched Comet. It started more towards the end of last year when Opus 4.6 was really, really good. So that's when we took two months or three months off around the time people here in this conference said we were the most likely company to fail. Okay. Let's address that. Yeah. So it was this time a year ago where there was a vote.

[00:13:36] And they said it wasn't to fail if they said they wanted to short you guys. Isn't it the same thing? Uh, I suppose. Yeah. Yeah. Yeah. Well, here I am sitting here still and we're doing better than most companies ever performing. People would have lost money on that trade. Yeah. And I would bet the same this year. If you guys short us, you're going to lose more money next year. Uh-huh. But the main thing is we actually recognize... Why do you think people thought that though? I mean, I don't want to put you in other people's head, but there is sometimes a misconception about you guys and you're very out there in the news and you're very good on Twitter and, you know,

[00:14:04] people see you getting involved in various kind of high, you know, there was perplexity was going to buy Chrome, perplexity was going to buy TikTok. Like it's an interesting strategy just as a reporter to see it happen. But I think it causes people to maybe misunderstand a bit about what you guys do in the solidness of your business. Well, towards the end of last year, we realized that coding agents, everything has to be code mode, right? And other companies also realized that like Cloudflare has written blogs about this. Basically, you should just expose all the primitives outside of the model

[00:14:35] in the form of code primitives, including search. So you should move away from rag to just exposing search primitives like crawling, fetching, you know, chunking, all of that in form of like actual like commands that the model can call. And when models were fundamentally becoming really good at code, almost every task just becomes a coding task. So you have to essentially rebuild a company, your backend, your harness, your infrastructure around it. And we are really good at doing these

[00:15:02] things, by the way. A lot of companies are unable to adapt in new times. So they built a whole company around chat AI. They don't know how to put work around agents. And then they have to like host open claw, and like they have no idea what open claw is. Which companies do you think of? What would that happen? A lot of companies I know, including Microsoft, just launched their version of agents by just forking open claw and running it on Azure. Yeah. Right? You need to have the ability to actually know that open claw is even an outdated

[00:15:28] harness at this point because a lot of things that built around are unnecessary as models get even better coding. So that's basically what we're good at. And that's what Dylan pointed out too, which is you need to know how to build harnesses, how to build infrastructure, what to not build as models get better, and own all the primitives outside of the model like search, browsing, sandboxing. And that's basically what we did. Right. Well, people were like having fun chatting here. Right. We were building and that paid off and we're doing pretty well.

[00:15:57] Right. The other thing going on right now obviously is the impending Anthropic IPO. And it's interesting to watch. I mean, you've been at the beginning of, you know, this is a search tool and a consumer tool. OpenAI seems like they had to go through a growth, you know, maturation phase where they realized that going all in on consumer was not going to be, I mean, like you said, you know, agentic coding is really where everything is at right now. I mean,

[00:16:22] can you give me a bit of the landscape right now of the frontier model companies and, you know, OpenAI seems like they're desperately but quickly catching up. Anthropic seems to be in the lead. Like just on a step back sense, where are you seeing that world? I think people are increasingly realizing that AI is a deeply recursive product more than a network effect product. There are a few people who use

[00:16:47] it a lot. 80% of the revenue is probably coming from 20% of the users. You know, there's a deep parallel there. It might even be 90-10. Yeah. So, and the reason for that is like agency and curiosity are not widespread. There are only few people who really know how to take the most out of these things and automate a lot of these things without being wasteful. It's not the same as token maxing. Like everybody can just set up hundreds of loops and consume a lot of tokens here today. But there are very few

[00:17:15] people who actually know how to get the right ROI out of it that saves you a lot of costs on so many other things. Applying agents to optimize your infrastructure, lowering your costs, building an entire new database. We did that. We completely replaced our usage of AWS DynamoDB, which is where we hosted our entire trillion web page index. We built a new database from scratch using our own coding agents and that's saving us like several tens of millions of dollars a year now. Yeah. So, that's kind of

[00:17:43] the way that we have agency and curiosity are not widespread. So, you may not actually need to have a billion users, a hundred million users anymore. All that matters is whether people are using your product for frontier use cases. Mm-hmm. Are they actually delegating a lot of tasks that they would otherwise do themselves on a computer to your agent? Yeah. Are they? So, there are only a few companies who know how to build such products. We're one of them.

[00:18:09] Anthropic obviously pioneered this. OpenAI is pivoting to this and they've been doing really well with Codex as well. Uh-huh. But that took them a while to realize and they had all this advantage they accumulated with, you know, consumer, mindshare, with chat GPT that may not necessarily be useful for getting people to pay for tokens. Yeah. Well, what do you think about consumer though? Like where does that world go these days? I mean, you're just wondering that 90% of the revenue is going to come from or whatever your split was. Like,

[00:18:37] what's the future of consumer look like? Is it even relevant or does Google just kind of suck up what interest there is and that's kind of where it goes? Well, I feel like consumers is like kind of the building the yellow pages equivalent in AI. Like Google AI overview or AI mode is that. And obviously, some people are joking, okay, you put it's like the Google shipping perplexity on the Google search bar. Yes, but it may not help you make money. In fact, like the latest earnings call showed that it actually loses money for them. On the other hand, being extremely

[00:19:07] good at deep research, knowledge work, automations, coding, uh, allows you to make a lot of money. There was this Twitter user. I think his name is Signal. Some people might know him. That's like G. N. Yeah. Yeah. He actually wrote a very good tweet once saying, in consumer products, retention is actually sought after. You know, you don't want a product with low retention. Right. Except when it's AI because higher retention means high token consumption.

[00:19:33] Mm-hmm. And if you're free, who's paying for all of that? Right. Right? So you actually, retention might look good on your charts but not on your bank account. So you have to be in a position to monetize that and, uh, that's why token-based revenue is good. Right. And that's, again, getting into this territory of business use cases and enterprise, which is why, like, you know, we, we saw this coming end of last year and that's why we moved our strategy from focusing on, like, mass market

[00:19:58] consumer to high value, high agency, uh, business owners, solopreneurs and entrepreneurs and, like, up-market enterprises. Yeah. So what, what is the business model for consumer then? And, and on top of that, like, um, you look at these agentic tools that are, everyone seems to be, you guys included, trying to get more users to, to, to adopt. The retention rates are fairly low on a lot of them. Right. I think I've read that muse is, is quite weak and instinct is mediocre. Yeah. That's because

[00:20:25] only VCs book a lot of business class flights and hotels every week. Yeah. Normal people don't. That's it. Yeah. It's as simple as that. It works so well for them. Well, it's exactly. So it's, it's a product for a hundred people, a hundred people in the world. The rest of the people when repeat use cases for agents are actually in the works workspace, which is why we called our product teammates. You don't need teammates for your life. You need teammates for your work. Yeah.

[00:20:49] So I think, I think that, that, that realization we had last year itself that if you want to actually build agents and if you want to have high retention, you need to have people pay for it. Right. And, and, and that's, that's, that's why you've got to focus on. So are you, are you pretty bearish on, let's say muse? No. Um, or you think it would just become an enterprise tool? No, I think,

[00:21:13] I think, look, I actually think the muse strategy is very good. Um, they don't rely on click based ads. So they can use all the understanding they gain about every user on the free usage and make your Instagram feed a lot better. They claim they won't do that today with their privacy policies where every sandbox session is private, but eventually that's the only thing that can pay for it. And so I am,

[00:21:37] I'm actually, I actually believe Meta is extremely well positioned to monetize muse and keep it free. In a way that you think the other players in that space are not. Right. Yeah. Uh, last question for you here, cause you always tend to be pretty bold and outspoken about the path and trajectory for your company. I've seen you in the past say you expect perplexity to IPO in 2028. I don't know how recently you've said that. So maybe you've readjusted, but can you give an update on where you see

[00:22:04] perplexity going and is the public markets going to be receptive to? So I'm not in a position to share any numbers today, but all I can say is we're doing pretty well enough that we could IPO anytime from now, uh, anytime from now to 2028. And I feel very bullish about our company and, uh, you know, we made all the right decisions last year when people were thinking we didn't know what we were doing

[00:22:29] and that's really paid off this year. We've hired really good people. Our infrastructure is solid and we've been very successfully post training our own models and lowering our serving costs. Mm-hmm. And we don't rely on inference providers to do any of that. We just use our own computer. You have your own compute, which again brings up your relationship. And we pay for that. Like we are in a position to pay for procuring compute too. Mm-hmm. So that's, that's, these are, these are part of the changes that we make. But you're not one of those companies that thinks, well, we've developed,

[00:22:57] we've built up a great amount of compute. Maybe we can resell that in some way. No, actually I, I, you know what? I have people who are supposed to be having a lot of compute constraints technically based on their business model coming and selling me compute, you know? You could probably make a lot of money from that. So, you know, the thing is like, He's one of those people anthropic? No, no, no, they're, they're doing well. But, uh,

[00:23:22] uh, compute is, like you have to be in a position to like warrant compute. And if you over build compute, then, uh, you don't have any like margins other than reselling it. Right. And, uh, that's why it's very important to be providing a real application to the end user. You know, uh, there was some quote Twitter thing I saw was everybody can be building infra. Uh huh. Because then you are only infra and no, no use it, no need, uh, demand for that.

[00:23:52] I mean, Elon's done okay with that. It's been a decent business for SpaceX. Yeah, but that's, he, Elon's great. You know, he's singular, but I'm saying everybody can just be an infra company. So I think you need really well class applications and those applications need to be in a position to monetize. And it, it, in AI at least in a genic space, that's where the real revenue growth is coming from that happens to be in the business and workspace. And that's what we are going after. Right. Well, look, Arvind, I think we all understand for Plexi a little bit better now.

[00:24:20] Always more to, always more to talk about. I appreciate you for really, uh, trying to help explain the company better than I can. So, uh, so we do what we can. And then I guess anyone who shorted them last year, uh, that was a bad trade. So sorry about that. Maybe better luck next year. Thank you. All right. Thanks Arvind.