CEO & Co-Founder, Perplexity AI
In Conversation with Aravind Srinivas: Live from FF Global 2026
Aravind Srinivas, Perplexity in conversation with John Micklethwait, Bloomberg News.
About Aravind Srinivas
Aravind Srinivas, CEO and co-founder of Perplexity AI, said in June 2026 that the company’s annualized revenue had tripled in the first five months of the year, crossing $500 million in April, and that he was not at liberty to share more recent numbers. He described Perplexity’s valuation of roughly $20 billion as reflecting investor interest in “accuracy and orchestration,” and said the company’s “Perplexity Computer” product orchestrates across models, files, tools, personal contacts, connectors, and local chips to maximize “token value per watt per user.” Srinivas stated that Perplexity is the largest multi-model orchestrator, routing “hundreds of trillions of tokens a month,” and that he aims for the company to be “the most accurate AI and the biggest inference orchestrator on the planet.” Srinivas said Perplexity benefits when frontier AI labs such as Anthropic, OpenAI, and xAI improve their models, because Perplexity routes across all of them. He argued that “the value is in the application layer” and that “the pure API model doesn’t work.” On copyright lawsuits, including one from CNN, he stated that “nobody has any copyright over truth and facts” and said he would let the legal process decide. He described LM Arena as “a terrible cancer on AI,” saying it rewards models with “pretty formatting” over correct answers. Regarding a potential IPO, Srinivas said 2028 is the “upper limit” and that Perplexity might choose to go public sooner, while noting that remaining private offers advantages such as moving faster without quarterly earnings constraints.
Transcript (94 segments)
✨ AI-enhanced transcript with speaker attribution
A
Aravind Srinivas0:10
Hello.
I
Interviewer0:11
How are you?
A
Aravind Srinivas0:12
Good. How are you?
I
Interviewer0:14
Am I right in thinking you're valued the business Perplexity at about 20 billion US dollars? Is that right?
A
Aravind Srinivas0:19
Last time that's what.
I
Interviewer0:20
Okay.
A
Aravind Srinivas0:20
Yeah.
I
Interviewer0:20
So, 20 billion dollar valuation, that is not investors valuing you as a search engine, right? So, what are they valuing? What is the business that they're seeing that they're valuing you at 20 billion US dollars?
A
Aravind Srinivas0:33
So, investors are investing in Perplexity for accuracy and orchestration. Last year, we were orchestrating across multiple different models. With the launch of Perplexity computer, we took it to a whole new level where we started orchestrating across models, files, tools, personal contacts, connectors, and most recently even chips where we can offload some of the tokens and inference through your local devices. So, we think of Perplexity as the only company that's truly incentivized to maximize the output token value per power watt per user. That is the orchestration problem to solve. So, it requires you to balance intelligence, accuracy, cost, privacy simultaneously. It's easy to be good at one thing. It's easy to be really good at cost, but not offering accuracy and intelligence. It's also easy to be really good at accuracy, but being super expensive. It's really hard to do it all simultaneously. So, that is the orchestration problem we're solving and doing it in a way that simplifies everything AI can do and putting it in one unified interface is basically what we tried to do with Perplexity computer.
I
Interviewer1:49
What was the technical moat or the technical moats to get to that point?
A
Aravind Srinivas1:54
The hardest engineering problem to solve in achieving this is building the most sophisticated inference engineering platform. So, you're not just routing to one model. You have to know which model is good for what task. You also have to know how to do a particular task cost-efficiently. You should be good at knowing what to do with any model failing at a task. And so, you're essentially building the most sophisticated applied inference engineering platform in the world. And by bringing in proprietary web search capabilities that Perplexity is known for, turning that into deep research and wide research, and finally into doing anything a human can do on a computer except the AI becomes the computer, that is the engineering problem. And so, the good thing that comes out of solving that problem is creating the most value of the output tokens for power watts consumed. Everyone's freaking out about how many tokens they're spending, and some companies are burning half a billion dollars in a month because of one engineer's token maxing. I think that's not the way things should be headed. We all should get the benefits of AI, frontier AI, without having to worry about whether we can afford it. The whole thing of a permanent underclass doesn't make sense.
I
Interviewer3:17
How deep is that backlash to token maxing right now?
A
Aravind Srinivas3:20
I think the customer essentially speaks the truth. If a lot of people are saying, 'Hey, I don't think my task requires this much dollars spent for the tokens,' I think it's worth listening to. That's actually why some AI inference needs to move local. Open source models are definitely going to play a big role in this future. It's never about only wanting to use open source models or only frontier models. You essentially need hybrid compute between frontier on the server and open source models that can run locally. And that only increases the need for the existence of an orchestrator, which is the role we want to serve in the community.
I
Interviewer4:11
I'm just thinking of that pushback around maxing out on tokens and switching up between open source and the frontier models. What does that do for the business models of Anthropic and OpenAI? Does that challenge their business models? Does that challenge their trillion-dollar valuations?
A
Aravind Srinivas4:26
Well, I think competition is good. No company is immune to competition.
I
Interviewer4:33
Is it a threat to them?
A
Aravind Srinivas4:35
It's certainly good. Frontier model companies only deserve trillion-dollar valuations if they continue to stay on the frontier. If there are multiple people who can provide the same capability at a far lower cost, then you're no longer at the frontier. And so the value accrues to applied inference platforms and orchestration layers. As new frontier capabilities keep emerging from the next-generation models, those companies will command their high valuations. If they do not, then it's obviously going to be difficult. It's an exciting time to be in AI because no one is guaranteed victory. Things change so fast.
I
Interviewer5:19
Mhm.
A
Aravind Srinivas5:20
Whoever is the market leader, whichever company is the one that everyone talks about...
I
Interviewer5:25
Who's the market leader now?
A
Aravind Srinivas5:27
The market leader depends on what metric you measure. If you measure revenue, obviously whoever is doing well right now. But the long-term metric that really matters in AI is the token value per watt per user. That's the metric we want to win on. And it's okay to win on that slow and steady. You have to win on it right away.
I
Interviewer5:46
There's some reporting today, Wall Street Journal reporting that Sam Altman and OpenAI are looking to cut prices.
A
Aravind Srinivas5:50
That's amazing. I'm super glad he's doing that.
I
Interviewer5:53
You're happy to hear that. Is it going to be a price war amongst the frontier labs?
A
Aravind Srinivas5:56
I hope so. And I think that war benefits users and companies like Perplexity a lot. Because that way we can orchestrate cheaper prices, bring down the cost, which will mean that by Jevons paradox, more people are going to use it and people who are using it a lot will use it even more, will automate even more things, will be able to build even more business ideas faster.
I
Interviewer6:17
Perplexity computer, which is what you started off talking about. Where do you see that? Is that an enterprise and a consumer product?
A
Aravind Srinivas6:22
Yeah.
I
Interviewer6:23
Where are you seeing the greatest demand? What are the numbers looking like? What is adoption looking like? What is it doing to revenues?
A
Aravind Srinivas6:29
In the financial services sector. It's natural there because we're known for accuracy and that's one vertical where accuracy matters a lot. Same thing in legal. We're seeing a ton of demand among small business owners, people who are running one to five-person teams. The C-level executives, the founders, they're all really automating a ton of things and building things that they could never build before with teams that are really small. Our power users happen to be executives, founders, people in hedge funds, private equity, investment banks. So, a lot of adoption in these sectors.
I
Interviewer7:10
Are there certain tasks that we should never delegate to agents?
A
Aravind Srinivas7:14
You always want to stay in the loop. You don't want to completely give all the autonomy to the agents yet.
I
Interviewer7:21
Mhm.
A
Aravind Srinivas7:22
It's all about making sure you scaffold it in the right way where you're able to verify its outputs and the guardrails are in place.
I
Interviewer7:32
How do you think all of this, I want to broaden it out to what this all does, what your perception is, what your view is on what this all does to the labor market. You're creating this harness for agentic AI. You want to be the operating layer for AI.
A
Aravind Srinivas7:44
Yeah. So, I have a very optimistic view about the future. I do not like talking about job losses. What I actually want to think about is how can this set of tools and capabilities help us build new businesses way faster? The economy only flourishes if there are new companies being built. If the next trillion dollars of market value is all shared by the big seven or the foundation model layers, that's not going to benefit all of us. But if everybody in this room or our friends can go try to build a new company and turn it into a hundred million dollar company or a billion dollar company, a passive income of ten to a hundred million dollars a year, that's awesome. The benefits are shared more widely. You need to have the tools in place to help you spin up small businesses and get them flourishing. That's why we announced this Perplexity for small businesses program and a billion dollar build to support entrepreneurs. We would give them computer credits and allow them to turn creative ideas into multi-billion dollar companies.
I
Interviewer8:57
Where are you on hiring? Is Perplexity hiring? Are you adding head count as you push out these new capabilities?
A
Aravind Srinivas9:02
The most curious and highest agency people do really well at Perplexity. It's not about the number. We've never been interested in token maxing or head count maxing. Both of these things don't make sense. I think what really makes sense is how curious you are, what are the questions you ask and what are the creative ideas you can come up with. This is the best time to, if these values resonate with you, this is the best time for you. There's never been a better time for a curious and high agency person.
I
Interviewer9:33
You recently, am I right in thinking you recently crossed 500 million or you're close to crossing 500 million ARR? Is that...
A
Aravind Srinivas9:39
That was...
I
Interviewer9:39
Could you give us an update? Where is ARR?
A
Aravind Srinivas9:41
I'm not at liberty to share exact numbers yet. All I can say is those are pretty old numbers right now.
I
Interviewer9:45
Those are pretty old numbers. By a significant amount?
A
Aravind Srinivas9:48
Yeah.
I
Interviewer9:48
Okay.
A
Aravind Srinivas9:49
Yeah.
I
Interviewer9:49
Target for 2026?
A
Aravind Srinivas9:52
More digits.
I
Interviewer9:52
More digits. Okay. What are all these massive mega IPOs doing to your world? You talk about 2028. You're on track. You want to IPO in 2028. What is this all, SpaceX pricing today, listing tomorrow?
A
Aravind Srinivas10:05
Yeah. So, we're really happy for all these companies, SpaceX and Anthropic and OpenAI to IPO. We use all their models in our products. They will get better if these businesses grow. So, we're really excited for them. I mentioned 2028 for us as this upper limit. We may decide to do it sooner. I don't know. But what really matters is taking the advantages of remaining private right now. Moving really fast. Not being bottlenecked by the quarterly earnings calls. That is going to be a unique edge we would have if these companies IPO and we do not. We hope to capitalize on that.
I
Interviewer10:44
That's interesting.
A
Aravind Srinivas10:45
Yeah.
I
Interviewer10:46
They go public. They're under more scrutiny, quarterly reporting, but you can charge ahead and do things that they couldn't otherwise do.
A
Aravind Srinivas10:53
What is the biggest constraint for you right now to continue to scale?
I
Interviewer10:58
Honestly, we're not constrained by anything other than compute right now. So, it's really hard to find compute. That is the number one problem stopping us from...
A
Aravind Srinivas11:08
What does that look like? Being hard to find compute? Is this you trying to lock in contracts with CoreWeave? Is this you in negotiations with Jensen and Nvidia? What does that look like on the ground, trying to find compute?
I
Interviewer11:19
Jensen's doing really well. That's what it means.
A
Aravind Srinivas11:22
Yes.
I
Interviewer11:23
It's really hard to find GPUs.
A
Aravind Srinivas11:26
Yeah.
I
Interviewer11:26
So, do you look at Cerebras? Do you look at TPUs? Do you look at...
A
Aravind Srinivas11:31
I mean, we've considered other chips, but Nvidia remains the frontier. GB300s are awesome. Vera Rubens, we haven't even gotten a chance to properly utilize it yet. A lot of agentic computers running on CPUs now. There's innovation at that layer too, where CPUs, there's inference compute that's going to go to local devices. We're very excited about our work with Apple, Intel, and Nvidia there. That's how we're going to fight this bottleneck on compute on the data centers by trying to move inference away to your local devices. We're going to take the data center and bring it to your laptop. We hope to adapt to these pretty challenging times where there's a clear compute bottleneck right now.
I
Interviewer12:21
Do we, are we moving, is this the year where we move more from, does inference become a bigger part of what's happening across chips versus training versus...
A
Aravind Srinivas12:29
CPUs and GPUs, all the agent loops and products like computer or cloud code or Codex, all the agent loops where the agents are actually transforming your files, downloading them, and making web apps, and hosting web apps, all that computer is running on CPUs. So, agents are going to use CPUs more than humans. That's a big change. The frontier model or open-source models, their inference tokens, the code generation, and all that's going to happen on GPUs. But the output of those tokens, which could be the code it generates, that's going to run on the CPUs. You need a sophisticated inference platform that has both these chips at the highest level of efficiency. Your laptops could be that. That's the demo we did with Intel at Computex where I was on stage with Lipu Tan, and we did a demo where, whenever it comes to handling private, confidential, or files, a local open-source model would run on the laptop itself and do the processing. And when it thinks it needs access to the frontier and it's not dealing with sensitive data, it can use the frontier model on the server. I think we are excited about that because that's the best way to handle this compute shortage. It's not a dichotomy between only server-side or only local. You want to be able to utilize both and orchestrate very efficiently across.
I
Interviewer13:55
Yeah.
A
Aravind Srinivas13:56
How sustainable is that? Is that the answer to the kind of compute maxing that we've been seeing?
I
Interviewer14:03
We believe so. We believe in this vision that the data center is moving to your device and sovereignty matters. The UK has always cared about sovereignty a lot and you can talk a lot about on-prem, but the highest form of sovereignty is every individual. You own your tokens, all the tokens on agent loops are running on your devices.
A
Aravind Srinivas14:23
Is the UK a good place to invest?
I
Interviewer14:25
Absolutely.
A
Aravind Srinivas14:26
Yeah.
I
Interviewer14:27
Yeah, I was actually just hearing the vision today. I think I was seeing on Twitter that the UK is looking for a next trillion-dollar company.
A
Aravind Srinivas14:35
Yeah.
I
Interviewer14:36
So, I would love to be an investor in that company.
A
Aravind Srinivas14:38
Next trillion-dollar company. Do you know who that is? Can you let us know?
I
Interviewer14:42
Well...
A
Aravind Srinivas14:42
Who's the next? What are you looking at in the UK that's looking most...
I
Interviewer14:46
I mean, I'm very glad to have been a very early investor in Eleven Labs. I think they're doing really well.
A
Aravind Srinivas14:55
David Silver just started his new company. There's a lot of inference chips that are being built here. So, endless possibilities. Honestly, there should be more. We want to support that and we met with the government and told them that we want to actually give them a lot of computer credits for the next business builders here. We have search APIs, model APIs, inference platform. The quality of engineers there is tremendous. In just this week, I met with a couple of them that are likely to join us.
I
Interviewer15:32
So, you're here on a poaching mission as well.
A
Aravind Srinivas15:34
Well, not poaching. They want to come work with us.
I
Interviewer15:37
Okay.
A
Aravind Srinivas15:37
Yeah.
I
Interviewer15:38
Okay. We have a big creative industry in the UK as well. And you faced a number of legal cases around training and use of data. I'm not going to get into the details of those cases, but there is a question about how, if we're aggregating all of this, how creatives get remunerated. How they get compensated. What is your answer to that?
A
Aravind Srinivas16:02
Well, particular cases, obviously, I cannot comment in detail because we have to respect the judicial system. In general, when technology progresses really fast, the laws usually haven't caught up yet with the state of the technology. They were built at a different time. I almost see this judicial process as a way for the courts to do that, to actually update the laws with the current state of technology.
I
Interviewer16:31
As in there should be better protections for the creative industry from...
A
Aravind Srinivas16:35
There should be protections. Ultimately, the users have to be protected.
I
Interviewer16:40
The users or the creators?
A
Aravind Srinivas16:42
I think fundamentally, there should be no monopoly over facts and knowledge. That's what we believe in. If you actually look at copyright law, it states two things. One is there's no monopoly over facts and knowledge, but there should be protection on the specific form of expression of that knowledge. Both have to be respected. Usually people only talk about the second, but the first thing is equally important. That's our position and we believe in that pretty strongly and we will respect the judicial process.
I
Interviewer17:13
Okay. We're running out of time. We've got about 60 seconds left on the clock. What does Perplexity look like in 5 years time? What is your lodestar in terms of ambition?
A
Aravind Srinivas17:22
I want us to be the most accurate AI and the biggest inference orchestrator on the planet. Today, I think we are the largest multi-model orchestrator right now. Nobody routes more multi-model tokens than Perplexity. Hundreds of trillions of tokens a month. That number is just going to multiply tenfold or hundredfold or thousandfold. We want to make sure the value of the output tokens is very high. We want to maximize the token value per watt per user, the most efficient orchestrator on the planet. That's our vision.
I
Interviewer17:57
Okay. And it sounds like you may be pulling that IPO target forward from 2028. 2027 is possible for you.
A
Aravind Srinivas18:01
We want to use our advantage as a private company when these bigger players are actually public and see how that helps us move even faster. So, it's a very good time to actually remain private when these are public. But, 2028 is the upper limit. We might choose to do it sooner.
I
Interviewer18:18
Arvind, thank you very much indeed.
A
Aravind Srinivas18:19
Thank you so much.
I
Interviewer18:26
Arvind, thank you very much for that. Thank you, everyone, for listening to the panels. I just want to say really briefly, I hope you've enjoyed the conversations here today, and if you have, we'll have more like this at Bloomberg Tech London, which is Bloomberg's annual European tech conference. It's on November the 2nd to the 3rd at Tobacco Dock in East London. Hope you'll be there. Enjoy your lunch, and thank you to Founders Forum for having us back.