Aravind, thank you so much for joining us.
Of course, thanks for having me.
And I wanted to start kind of talking about the business. You launched, I think Perplexity was really the first new competitor in search like since I was in college. And this year you pivoted into becoming really kind of an agent company. And again, you're sort of basically competing with the hyperscalers in a very competitive business. And I'm curious, like, how will you guys, how will your product teammates, how does it find space in agent world?
So Perplexity's moat is accuracy and orchestration. We're the only model-neutral, accurate agent orchestrator out there. Of course, hyperscalers all want to do it, but they want to lock you into their cloud or they want to lock you into the models in which they have equity in. We have no such incentives. We are the only company whose incentive is user-aligned, which is get the work done, do it in the most trustworthy and secure and safe way, and keep the answers and tasks accurate. And that's what users want. And every other competitor in the space always has this problem of wanting to get a return on their capex, either the capex spent in building your models, capex spent in building your data centers. So therefore, you're forced to actually like recoup all that investment. Example, we just crossed a billion dollars in annual run rate today.
Thank you, Bloomberg, please.
Well, the key realization we had when we looked at that was we spent 5x lower than the company that's going to IPO this week to get to that number. Our burn to get to a billion dollars in annual run rate is 5x lower, and that's because we are much more capital efficient. We are an application-layer company. We try to do what the users really want, which is solving their problems. They don't care about which models. They don't care about how many tokens. They really care if the work actually got done. And so that's our moat. It's accuracy and orchestration.
And how do you, what is the, you know, as you orchestrate these different models, if I'm using that verb correctly, where is this going to settle out? Like, how much of the information you send is going to go to Anthropic and OpenAI? How much is going to go to Chinese open-weight models? Where do you see, where is that now, and where do you see that going?
So for fast search, quick fast search, we've built our own model that's serving like 80 to 90% of our traffic right now. And all of that is served with our own compute clusters, our own post-trained models, our own inference stack. And that's allowed us to save at least $100 to $200 million annually. And so all that winnings from there that we get, those extra margins, has actually increased our gross margins. So we are now at 56% gross margins.
Yes, on paid users. Now, a lot of companies don't report gross margins over free users as well because they count that as sales and marketing expenses or R&D expenses. If you count even the spend on free users, we're actually at 17% gross margins. It used to be minus 219% in Q4 2025, and it flipped to 17% in Q3 2026, and it's actually only going to go up as we train our own models. And in order to do that, we announced our frontier research program today, where we're going to commit at least 10,000 GPUs to post-train models that are very good at multi-agent orchestration and continual learning. So we want to build models that are capable of harnessing the best of frontier labs. We're not competing against frontier labs. We think the future is about a system that orchestrates multiple models for what they're good at, combines their best properties, and does what the user really wants, which is getting their work done.
I like this is a new tradition that companies are going to release their financial figures on stage here. We appreciate that, and hope to have you back next year to do the same thing. You, I think, worked with Nvidia among other companies, and have talked about the extent to which you think some of this can move locally to people's own devices. How can you just talk a little bit about that partnership, how you see that developing?
Yeah, so towards the middle of this year, we started having this idea that people are going to care about AI systems, especially more and more capable agents, to make sure that it's deployed in a safe and secure way and they have full control. And you can argue about how, like, oh, I've done this security thing, I've made sure my agents are good, my sandboxes are great. But ultimately, the only way in which you can fully feel secure is if the whole system is deployed in hardware you own and it can be run in a completely air-gapped, isolated manner. So that's fully local on-prem AI.
What share of companies, of users, do you actually think care about that, about fully local on-prem?
Sovereign wealth funds, regulated institutions like banks, credit card companies, exchanges, insurance companies. Pharma, bio, anyone who cares about their IP or anyone who cares about doing right by their customers. Example, credit card companies who have a lot of your data and don't want all that to be going into a frontier lab that'll just take all those tokens and post-train their next model on it. We saw something like that happen with the Millennium price problems. So anyone who cares a lot about their data or are regulated will want a system that at least a version of that should be able to run fully local. Now, you should still want access to the frontier, right? It shouldn't be something you compromise on intelligence. And so we have the ability to orchestrate up from the local model to the cloud through the escalation framework. So the model decides, and you give it access, you give it the rights to do that.
OpenAI and Anthropic are, you know, going after those same customers and telling them that they're not going to do what you just said they are going to do with their data. I mean, should people not trust them?
Well, it's up to them. I'm just saying that, you know, it's really up to people who they choose to trust and not trust. But you can only provably have full control if you have it inside your own premises.
That's why it's called on-prem, right? And so people want a version of that that they don't have to rely on trust or like vibes. It's a real solution. It sits here. It's just like your refrigerator at home, you know. And I think we're going to see systems that can run in a few megawatts of power and draw a lot of capabilities and keep your data isolated. Nobody else can look at it. And I'm sure, like, not just institutions but some individuals would also want that.
I mean, one of the, the journal, I think, wrote about this project under the headline, 'Is this how the AI boom ends?' And because one of, I mean, one of the potential consequences of what you're saying is that we don't need as much compute as people think. I mean, I haven't quite seen if you've totally... What do you, where do, how do you see that? Are we building too much compute?
I mean, I think the answer to whether we're building too much compute is unknown, but I believe that local is not actually going to compete with, or local, like, the rise of local AI is not going to lead to any AI bubble pop or anything like that, because you can still consume frontier through local as a gateway. That's why we built this escalation orchestration, where sometimes the local model is not smart enough to do your task, and it gets your permission to orchestrate up to the cloud, take the help of the frontier, and come back and get your job done. And so that's very useful. And so as more and more people start using local AI, it's probably going to lead to an expansion in the use of frontier AI as well. And same thing, there are some people who like to be on frontier cloud system as a default and orchestrate down to local if they feel like their data is too private. And that's the mode called hybrid that we released on Apple Silicon, and we found pretty good adoption there as well. So you have to give people the choice. And eventually, Jevons' paradox is going to play its role, and as you make the cost of AI lower, you're going to see increasing adoption by businesses, especially small businesses that find adopting frontier AI pretty expensive, and start running a lot of automations and loops using that. And that's what we really need to do to diffuse AI widely in the society.
And there is sort of this debate, thanks for bringing up Jevons' paradox, which I was going to struggle to explain, and maybe you can explain better than I can. It does require, you know, just the notion that, you know, the more consumer spend costs go down, the market expands. It requires the market to expand by many, many multiples, and, you know, the number of zeros matters here. And I guess just to push on that question about whether we're building too much compute, I mean, are you seeing in the demand for your services, for others, is it growing fast enough?
Yeah, we don't have any issues with, like, you know, the amount of money we're actually burning for the return on investment. That's why I pointed out, like, we're one of the most capital-efficient businesses in terms of how much we actually burn to get to the revenue we got to. And that's only possible because we don't actually have any massive compute commitments. We have a pretty big GPU cluster. Pretty big. Like, the key differentiation between us and everybody else in the market is, on the one end of the spectrum, there's frontier labs that have so many GPUs, like millions of them, gigawatts, and on the other hand, there are like these startups that don't have that much to do their own models. We're in this amazing sweet spot where we don't have too many employees. The headcount is low enough that you get to have enough compute per person, per researcher. Yet we don't have the capex commitment that we need massive return on investment on trillion dollars worth of commitment.