Part IThe Story
The first engineer needed health insurance.
That was the whole problem. In the fall of 2022 it was not the kind of problem Aravind Srinivas had been trained to solve. He was twenty-eight, a Berkeley PhD whose résumé read like a walking tour of the three buildings where modern artificial intelligence was being assembled, and he had just left a full-time research job at OpenAI to start a company with three co-founders who did not yet agree on what the company was for. They had hired exactly one person. That person, being a person employed in the United States, required a health plan. Srinivas had never chosen an insurance provider in his life. So he did what he had been doing since he was a teenager in Chennai teaching himself to code from online lectures. He typed the question into Google. What is the best health insurance plan for a startup? Which HR software provides the best value for money?
What came back was what usually comes back: ads from insurance companies at the top, then marketing copy engineered to rank rather than to inform, a stack of links each promising to be the answer and none of them being it.
In another window sat a prototype. One of the half-dozen ideas the founders had been turning over was a tool that would use a large language model to read several websites at once and summarize what they said. It was a demo, a toy. "Let's at least ask this tool that we have," Srinivas remembers thinking. "And it started answering all these questions really well for us."
The order of operations matters. He asked Google first. This was not a skeptic shopping for a fight. This was a man who had read Steven Levy's In the Plex, a history of Google, so many times in a DeepMind office library in London that he later offered to recite its chapters from memory to a Wired reporter, who called Larry Page and Sergey Brin "my entrepreneurial heroes," and who, at the moment of need, still reached for the search box by reflex. The prototype was his second try. Nearly everything that came afterward grew out of that small letdown: the $34.5 billion bid for Chrome, the lawsuits from newspapers on two continents, the line on India's rich list, the long public argument with the Cloudflare CEO about what a robot is. A believer had asked the oracle a simple question and been handed an advertisement.
Weeks later, OpenAI released ChatGPT. Seven days after that, on December 7, 2022, Srinivas and his co-founders put their prototype on the open web and named both it and the company Perplexity. In machine learning, perplexity is the measure of how surprised a model is by what it sees next. Less perplexity means a better prediction. It is a strange word to stake a company on, a word for confusion chosen by people promising to end it. Nobody seems to have minded. Within two months Perplexity reported two million unique visitors.
One more fact about the prototype, worth holding onto for later: it was a machine that read other people's websites and told you what they said.
By the Numbers
The Answer Engine
0.01CGPA margin by which Srinivas missed transferring into computer science at IIT Madras
$3.1MSeed round, September 2022, from Yann LeCun, Jeff Dean, Nat Friedman, Andrej Karpathy, and Elad Gil, among others
Dec 7, 2022Public launch, seven days after ChatGPT
700,000Queries on New Year's Eve 2022, at roughly seven seconds per answer
$34.5BUnsolicited bid for Google Chrome, August 12, 2025, more than Perplexity's own valuation
$21.21BValuation after the Series E-6 round, early 2026
₹21,190 crNet worth on the M3M Hurun India Rich List 2025, its youngest billionaire at 31
A Hundredth of a Point
Before there was a search box there was a bus window.
Srinivas grew up in Chennai, the capital of Tamil Nadu, in a family he has described as more at home with ledgers than with circuits. His mother worked for the central government. His father was a financial accountant. "So I am actually the first engineer in the extended family," he told the Zerodha co-founder Nikhil Kamath in 2025, and Kamath, who knows the stereotype about Tamil families and engineering degrees, sounded genuinely startled: Really? You're a Tamilian, right? Several accounts say his mother had once wanted to attend the Indian Institute of Technology Madras herself and never got the chance. What is on the record, in her son's words, is what she did instead. Whenever the family's bus passed the IIT Madras campus, she pointed at it. "This is where you're going to study," she would say. Srinivas has been careful about the grammar: "It's not even like you should study; this is where you're going to study. So that was the expectation."
Picture a woman telling a child the answer before he has thought to ask the question. Hold onto that image too.
He got in. He did not get the department. The Joint Entrance Examination that sorts a million-plus Indian teenagers into a few thousand seats placed him in electrical engineering, a dual-degree program, while the friends he had trained alongside landed in computer science. IIT gives first-semester students one chance to switch branches if their grade point average clears a threshold. He missed it by one hundredth of a point.
— Aravind Srinivas, The Aarthi and Sriram Show, 2024At the end of your first semester IIT gives you an option to change your branch if you get about a certain CGPA, and I missed that by 0.01. I thought that was the end of the world. I was so mad at myself.
"For one year, I was just depressed," he said on the podcast hosted by Aarthi Ramamurthy and Sriram Krishnan, "and I mostly hung out with them," meaning the friends in the department he couldn't enter. Then he went into what he calls "a whole shell," and the shell turned out to have a door in the back. Online lectures. Competitive programming sites. A roommate introduced him to Kaggle, the data-science competition platform, and he started teaching himself machine learning with open-source libraries like scikit-learn. He learned Python before the computer science students around him had bothered to, because Python was "still not a language that even my computer science friends focused on." He is honest about the motive: "As a kid, it was more out of like vengeance. 'Let me prove the world wrong I'm actually good at this.'"
The vengeance got its first target in his third year. A machine-learning course run by the computer science department was open, in principle, to outsiders. The professor greeted him with a sentence that could stand as the epigraph for the next decade of Srinivas's life: "I don't usually let electrical people take this class. You guys are not going to do well. But maybe you can prove me wrong." Srinivas topped the class.
The pattern is visible already, and it repeats. A gate. A gatekeeper. A margin so thin it feels personal. Then a back route found through self-teaching and stubbornness, and an arrival at the destination by a road the gatekeeper never mapped. The next gate was graduate school. An internship in Montreal with Yoshua Bengio, later a Turing Award winner and one of the founders of deep learning, produced a recommendation letter. Srinivas applied to exactly two PhD programs. "MIT rejected me," he said. "Berkeley admitted me."
That is how a lot of Silicon Valley gets built: two applications, one rejection, one admission, and decades of compounding that follow from which envelope arrived thick.
Permission to Eat Eggs
Sundar Pichai was born in Chennai twenty-two years before Srinivas. He studied metallurgical engineering at IIT Kharagpur, went west, and rose so quietly through Google's product organization that his appointment as CEO in 2015 came as a shock mostly to people who hadn't been paying attention. To a generation of Indian engineering students he was proof that the route existed. The ambition he represented was not swagger. It was something closer to rectitude.
Srinivas's most surprising debt to him is about breakfast. In his early twenties, raised vegetarian in a household where eggs were off the table, as they are for many Indian families, Srinivas wanted more protein. An interview with Pichai came up in his YouTube feed, and in it the CEO of Google casually described his morning: wake up, read the newspapers, drink tea, eat an omelet. Srinivas sent the video to his mother. OK, she said. You can eat eggs.
It is a comic story and a revealing one. A young man who would go on to argue with the New York Times, the BBC, and Cloudflare about who may read what on the internet once needed a video of a Google executive to get his mother's permission to change his diet. Permission runs all through this story: the professor who didn't usually let electrical students in, the mother who let him eat eggs, the robots.txt files that later told Perplexity's crawlers no. Srinivas's life looks like a long negotiation with people who decide who is allowed into the room. Mostly he has gotten in. Sometimes, his critics say, he went in without asking.
The pilgrimage through Google's world continued. In 2019 he went to DeepMind in London as a research intern. He had rented a flat sight unseen, and it was, in his words, a "crappy home, with rats," so some nights he slept at the office. The office had a library, and the library had In the Plex. He read it over and over. "Larry and Sergey became my entrepreneurial heroes," he said. In 2020 he made it to Mountain View as a research intern at Google itself, working on machine learning for computer vision, on models with names like HaloNet and ResNet-RS. Pichai, that year, ran both Google and its parent company, Alphabet. The boy from the bus window was in the building.
At a TEDAI talk in San Francisco in October 2024, Srinivas played the audience a clip of Larry Page from 2000: "AI would be the ultimate version of Google. So if we had the ultimate search engine, it would understand everything on the web... It would be able to answer any question, basically." Srinivas paused on the date. "I was only six back then." He wasn't really saying I'm killing Google. He was saying something odder and more devout: I'm finishing the sentence Larry started. "At Perplexity, we get accused of trying to kill Google a lot," he told the TED crowd. "But trust me, we're not really trying to kill things."
Then, to Wired, about the man who gave him eggs: "It's kind of a rivalry now. It's awkward."
The Lab and the Door Out
Pieter Abbeel, Srinivas's doctoral advisor at Berkeley, is one of the most cited researchers in reinforcement learning, the branch of AI concerned with machines that learn by trial, error, and reward. He made his name teaching robots to manipulate the physical world, then co-founded Covariant to put that work in warehouses. He is a researcher who could not leave research alone long enough to stop building companies, which made him a fitting advisor for a student who arrived in 2017 wanting both.
Srinivas did not want only to write papers. He kept turning to his lab mate Peter Chen with the same question: "Hey, do you want to start a company with me?" The question went unanswered for years. Meanwhile he did the work, and did it well. He was lead author on CURL, a 2020 paper that brought contrastive representation learning to reinforcement learning from raw pixels, and a co-author on the Decision Transformer, which recast reinforcement learning as sequence modeling in the manner of GPT. His contemporaries cite these papers as real contributions, not résumé padding.
OpenAI found him in 2018. He describes the invitation as coming from John Schulman, the researcher later credited with leading the work that became ChatGPT. "So that changed my life," Srinivas said. "I got into OpenAI and that's when I knew this is the future." He interned there that year on policy-gradient algorithms, went on to DeepMind and then Google, finished the PhD in 2021, and went back to OpenAI as a full-time research scientist. He was on the team behind DALL-E 2, the image generator that made a lot of people realize something was coming.
He lasted about a year. Ask him why and you get a joke that isn't one. "Some people are unemployable," he told a room at Berkeley Haas in 2025. "They just don't listen to what the boss tells them to do. I'm one of them."
The three people who left their jobs alongside him each had a version of the same restlessness. Denis Yarats, who became chief technology officer, had spent years doing research at Facebook AI and was finishing a PhD at NYU. Johnny Ho came from Quora, a company built entirely on the premise that people's questions are worth answering well. Andy Konwinski had co-founded Databricks, a Berkeley research project that turned into a data-infrastructure giant, and so had already done once what the others were attempting. On paper the obvious next step for each was another lab job. They took a different one, and Perplexity AI was incorporated on August 22, 2022.
They tried to hire a few others. In July 2022, Srinivas messaged a young engineer named Aditya Baradwaj and offered him the job of founding engineer. Baradwaj said no. Things were going too well at Alameda Research, the trading firm attached to FTX. Both firms filed for bankruptcy in November 2022. Years later Baradwaj posted the old screenshot on X under the caption "worst financial decision." Srinivas replied with five words that seem to sum up both men's 2022: "e/acc bs effective altruism."
Five Hours Outside Yann LeCun's Door
The first idea was not search. It was narrower, and in the word Wired used, nerdier: software to translate plain English into SQL, the language databases speak. You'd ask a question in a sentence and get rows back from a table. The founders realized the table that mattered most to the people they needed to impress was Twitter. They built a version that could search Twitter, LinkedIn, and GitHub in natural language and surface things Google couldn't find.
In August 2022, Srinivas and Yarats went to lower Manhattan and waited outside the office of Yann LeCun, Meta's chief AI scientist and an NYU professor, a French-born pioneer of convolutional neural networks and one of the field's most famously combative public figures. They waited five hours and skipped lunch. When they finally got in, they showed LeCun something he had never seen: the accounts that most often replied to his own tweets.
He invested.
So did Jeff Dean, Google's chief scientist, who was in effect writing a check against his own employer. So did Nat Friedman, formerly of GitHub, and Andrej Karpathy, an OpenAI co-founder, and the angel investor Elad Gil. The September 2022 seed round came to $3.1 million. "It was very relatable for them to go and search about their own Twitter," Srinivas told Forbes, deadpan, as if he had discovered a law of physics. Two years later he told Lex Fridman something that rhymes with it: "The most common search is people searching for themselves on Instagram."
That is the founding insight, and it isn't a technical one. The first users of a new kind of answer are the people the question is about. A demo that shows famous men their own reflections in an unfamiliar mirror will raise money faster than one that answers questions about the Federal Reserve. Srinivas does not seem embarrassed by this. He seems pleased, the way a researcher is pleased by a clean result.
The fundraising style never really changed. Years later, with billions at stake, he described the process at Berkeley Haas: "I've never done a pitch deck for any of the other Perplexity funding rounds. I just write a memo, and I tell them, 'You can do a Q&A and ask whatever you want. I'll spend two hours with you, ask me all the questions you have, and we're just going to pull the metrics for whatever you have and show you right there. And anything else that is internal, not internal data, you can ask Perplexity. It already knows everything.' ... I'm not exaggerating." Elsewhere he admitted he wouldn't know how to make a deck if he tried. The demo was the pitch in that waiting room in 2022, and it has stayed the pitch.
Seven Days After ChatGPT
The name, by Srinivas's own later account, was "crappy." Latency was about seven seconds a query, an eternity in search, where Google had spent two decades training people to expect results before they finished blinking. The seed money, a million or two in practice, was gone the moment you paid for inference. The product was a search box that looked a lot like the chatbot everyone was suddenly using, and it shipped a week after that chatbot.
Then came New Year's Eve 2022. On a night when the internet was supposedly elsewhere, Perplexity handled 700,000 queries. Srinivas has called that the moment he knew. A product people use on New Year's Eve, despite the name and the seven seconds and the obvious resemblance to the most famous product of the year, is not a toy.
What set it apart was a rule, and the rule came from the academy. "The cardinal rule in academic research is to base your assertions on citable evidence rather than conjecture," he told the interviewer Vikas Shah. "Perplexity, by design, is restricted to sourcing information directly from the web, eschewing any reliance on pre-existing knowledge within the model." To Lex Fridman he put it more simply: "The principle in Perplexity is you're not supposed to say anything that you don't retrieve." Every sentence gets a footnote. Every claim points to a page a human wrote. ChatGPT at launch had no web access, couldn't cite anything, and hallucinated with total confidence. Perplexity hallucinated too, but it showed its sources, so the hallucinations were easier to catch. In 2022 that difference was enough to build a company on.
Srinivas reached for an image that came straight from his education: "It's almost like Wikipedia and ChatGPT had a kid." In May 2024 he launched a feature called Pages and described it, in his own LinkedIn post, as "AI Wikipedia." The company's tagline became "where knowledge begins."
The numbers moved the way numbers move in a gold rush. A $26 million round in March 2023. In January 2024, $73.6 million at a $520 million valuation, led in part by Nvidia and Jeff Bezos. In April 2024 a valuation over $1 billion. Forbes that spring counted about 15 million users and some 100,000 paying $20 a month. The backers by then included Susan Wojcicki, the former YouTube CEO; Paul Buchheit, who built Gmail; Tobi Lütke of Shopify. Jensen Huang said he used it every day. "It creates snowball effects for you," Srinivas said of the famous names. "People take you more seriously." Then SoftBank's Vision Fund 2 came in at $3 billion in June 2024, a $500 million round led by Institutional Venture Partners in December brought it to $9 billion, $14 billion in mid-2025, $20 billion that September, and $21.21 billion after a Series E-6 in early 2026.
In 2023 the company tried to buy the trademark "Perplexity" from a small software firm called Perplexity Solved Solutions, founded in 2017. The firm said no. On January 31, 2025, it sued. Even the name, it turned out, had belonged to someone else first.
The Weakness of the Enemy
Srinivas, who reads Sun Tzu the way some founders read Andy Grove, gave Lex Fridman the line that serves as his strategic doctrine: "Make the weakness of your enemy a strength." Then he named the weakness. "The core weakness of Google is that any ad unit less profitable than a link or one that discourages link clicks isn't in their interests to pursue aggressively."
That is the whole theory of Perplexity, and he admits he came to it late. "Looking back on the launch of our answer engine, we admittedly hadn't fully considered its potential to disrupt the AdWords business model," he said. "Had we realised this aspect sooner, our enthusiasm for the project would have been even greater." It is a funny admission. The disruptive business case was found after the fact, inside a product built because the founders were annoyed. Google makes money when you click. An answer that makes clicking unnecessary is something Google can build but cannot comfortably sell. Perplexity could sell it because Perplexity had nothing to lose.
Being late was part of the plan. "You can never succeed at recreating the whole index like Google," he told Forbes. "It's too late... It's sort of like finding your way in a maze where you're starting at a huge disadvantage." So he never tried to rebuild the maze. He built a different exit.
It is worth listing what Perplexity is built on. The transformer, the architecture underneath every modern language model, was invented at Google. The models doing the reasoning have come from OpenAI, Anthropic, Meta's open-source Llama, Mistral, and DeepSeek, wrapped in Perplexity's own fine-tunes under names like Sonar and, with a certain national theatricality, R1 1776. When DeepSeek's model became a geopolitical problem in January 2025, Srinivas posted that all DeepSeek usage on Perplexity ran in data centers in the United States and Europe: "DeepSeek is open-source. None of your data goes to China." The browser Perplexity launched in July 2025, Comet, is built on Chromium, the open-source foundation of Google Chrome. In January 2026 the company signed a three-year, $750 million agreement with Microsoft for Azure cloud capacity and access to models from OpenAI, Anthropic, and xAI. Amazon Web Services remained its main cloud.
None of this embarrasses him. It is the strategy, and he laid it out on Hard Fork in August 2025 with something like glee. The big model labs, he said, "are all hill climbing on exactly the same benchmarks, so that all their models end up being completely undifferentiated, which is essentially the necessary criterion for it being a commodity. And who benefits from that is us." He described the compute footprint with a researcher's precision: "We will not be having hundred GPUs. We'll have tens of thousands of GPUs. We'll not have a million GPUs."
So the company that set out to replace the most powerful intermediary on the internet is itself an intermediary, standing between the user and the models, between the user and the web, between a question and the dozens of pages that might answer it. Srinivas's bet is that the middle is the most valuable place to stand, as long as you are the one people trust. And trust gets expensive.
He once thought advertising was the prize. "If we crack advertising relevance," he told Harry Stebbings on 20VC, "I think we should be worth way more, because advertising can be really, really profitable." In February 2026, Perplexity dropped its AI-integrated advertising and went subscription-first, saying the change was meant to protect user trust in the "answer engine." The weakness of the enemy, it turned out, was also a weakness of the company that had diagnosed it.
Robots.txt
The internet has a door sign. It is a plain text file called robots.txt, and since the 1990s it has been the agreed way for a website to tell automated crawlers where they may not go. It enforces nothing. It is a request, honored on trust, and for decades the web's large crawlers mostly honored it because the bargain worked: the crawler took the page and sent back readers.
In June 2024 the bargain broke in public. Forbes said Perplexity had published a story largely copied from a proprietary Forbes investigation without prominent credit. Its chief content officer, Randall Lane, wrote a column titled "Why Perplexity's Cynical Theft Represents Everything That Could Go Wrong With AI." Srinivas said the feature had "rough edges" and called Perplexity a mere "aggregator of information." A few days later Wired went further. Working alongside an independent developer, Robb Knight, its reporters found a machine on an Amazon server, "almost certainly operated by Perplexity," crawling parts of Condé Nast properties whose robots.txt files told Perplexity's bot to stay out. The machine used an IP address the company did not list publicly: 44.221.181.252. Wired counted at least 822 hits in three months. The magazine also caught the chatbot paraphrasing its stories closely, summarizing others inaccurately, and in one case claiming falsely that Wired had reported that a specific California police officer committed a crime. The headline was "Perplexity Is a Bullshit Machine," which the authors defended as a technical term.
Srinivas's answer was precise in a way that satisfied no one. Perplexity was not ignoring robots.txt, he said, but it did rely on third-party crawlers that might. Asked whether he would commit to stop scraping Wired through those third parties, he declined.
The next eighteen months produced a lot of paperwork. In October 2024, Dow Jones and the New York Post sued, alleging "massive freeriding" and saying Perplexity had attached hallucinated quotes, including some about F-16 jets for Ukraine, to articles that never contained them. The same month, the New York Times sent a cease-and-desist letter. Perplexity's blog post in reply on October 24 is the purest statement of the company's temperament on record: "There are around three dozen lawsuits by media companies against generative AI tools. The common theme betrayed by those complaints collectively is that they wish this technology didn't exist." At the Wall Street Journal's Tech Live conference, Srinivas went to the paper's own stage to make the case in person: "Nobody's coming to Perplexity to ingest their news. People are directly going to The New York Times, The Wall Street Journal. People come to Perplexity for understanding the news in the context of what they already know."
In June 2025 the BBC threatened legal action and demanded compensation and the deletion of retained material. On August 8, 2025, Japan's Yomiuri Shimbun sued over what it called the "free-riding" use of 120,000 of its articles between February and June. The Asahi Shimbun and the Nikkei followed within weeks. In October 2025, Reddit sued in federal court in New York.
In between came Cloudflare. On August 4, 2025, the company that sits in front of a large share of the world's websites published research saying Perplexity was using undeclared "stealth" crawlers to get past firewalls built to block it. Cloudflare's CEO, Matthew Prince, a lawyer by training who has spent his career deciding which traffic gets through, wrote on X that Perplexity behaved "more like North Korean hackers" than a reputable AI company. Perplexity called the report a "charlatan publicity stunt."
On Hard Fork a week later, Srinivas was given the chance to step back and didn't. The crawler and the user agent were two different things, he explained: one indexes the web on the server, the other opens pages on a user's behalf, "literally like a user delegated an AI to open these tabs just like how a human would on Chrome." Then he went after Prince personally. "How would you run a company like Cloudflare that's supposed to protect people from bots when you don't even know what a bot is?" Prince, he said, was "basically trying to be the new gatekeeper."
That word, gatekeeper, is the oldest one in his story. The professor who didn't usually let electrical students in. The 0.01 threshold. MIT's rejection. Every gate he has met he has treated as a dare, and he has mostly been right to, because mostly the gates were arbitrary. Here, though, the person at the gate was the owner of the house.
Casey Newton, the host sitting across from him, named the stakes in the episode's introduction. Chatbots, he said, "represent this kind of new, more extractive version of the web." Srinivas answered with a theory of the future web. The spammers lose. The good writers charge more because readers arrive "out of their own will." There would be a revenue program "closer to Apple News with some protections for users." In July 2024 Perplexity had in fact started cutting checks through a publishers' program with TIME, Fortune, and Der Spiegel. In January 2026, the Wikimedia Foundation disclosed that Perplexity was among the companies now paying for Wikipedia's data through Wikimedia Enterprise instead of scraping it.
So the kid of Wikipedia and ChatGPT ended up paying its parent.
Bids Larger Than the Bidder
On January 18, 2025, the day before a federal ban on TikTok was set to take effect, Perplexity submitted a proposal to merge with TikTok's American business. The structure, as reported, imagined a new U.S. holding company, provisionally called NewCo, in which the federal government could end up owning as much as half after an IPO valued at $300 billion. Perplexity was then worth $9 billion.
On August 12, 2025, the Wall Street Journal reported something stranger: an unsolicited $34.5 billion offer for Google Chrome. Judge Amit Mehta had found the year before that Google illegally maintained its search monopoly, and he was weighing remedies, one of which, in theory, was forcing Google to sell the browser. Chrome has an estimated three billion users. The offer was larger than Perplexity's entire valuation.
On Hard Fork that week, Kevin Roose asked the obvious question: did Perplexity actually have $34.5 billion? "No one has the money in hand to make such a large bid like this," Srinivas said, as if that were self-evident. Before bidding, Perplexity had asked three or four investors whether they would back it, "and they all said yes." Roose pressed: wasn't this just a publicity stunt?
— Aravind Srinivas, Hard Fork, August 2025If the judge rules that Chrome should be sold, we will buy it, period.
Then he described the situation with complete candor. "The reality is it's unlikely to actually be the case that the judge would force them to sell Chrome. And even if the judge forces them to sell Chrome, they're going to appeal it. And it's going to take two years." He closed with a line from a hockey poster: "But you will lose 100 percent of the shots you don't take."
What, then, was the bid? Casey Newton offered a reading. By posting a price, Perplexity undercut any argument from Google that Chrome couldn't be divested because no one would buy it. Srinivas half-accepted it: "We're not saying this should be the ruling. We would rather say in case this is a ruling, we're here." The bid worked as a question posed to a court, to Google, and to the press: What if? His posture toward the answer was the one he has taken toward gatekeepers since he was eighteen: if the answer is no, the question still cost almost nothing.
The bids made Perplexity more famous than any marketing budget it had. They also fit a man who once had to watch a Google CEO eat an omelet before he was allowed one. Now he was offering to buy Google's front door.
The real distribution deals came quieter and bigger. In July 2025, Bharti Airtel offered all 360 million of its Indian subscribers a year of Perplexity Pro for free, and the app went to the top of India's App Store. In December 2025, Cristiano Ronaldo took a stake and signed a global brand partnership. Apple and Meta were both reported to have looked at buying the company. Srinivas said on X that he intended to stay independent and take Perplexity public sometime after 2028.
Meanwhile the product kept trying to stop being a search engine. "It's our transition from answers to actions," he said of Comet, which at launch was available only on the $200-a-month tier and was made free in October 2025. In February 2026 came Perplexity Computer, an agent built to carry out multistep tasks across several models. In April came a version meant to run on the user's own machine. The company named after a model's surprise was now trying to do things, not just say them.
The Inflation Rate of Knowledge
His working life, by his own account, has no edges. "I don't do anything other than working, sadly," he wrote in a 2025 Reddit AMA. "I listen to podcasts and audiobooks whenever I can. I spend a lot of time on X (which is both good and bad)." He gets to the gym three times a week and sees family on weekends. During a three-week internship in Bengaluru years ago he stayed in Koramangala and never explored the city. "I just worked all the time." He has said that during his internships he lived on one cheap meal a day to save up for a MacBook so he could code seriously. At Y Combinator's AI Startup School he told the audience: "Work incredibly hard. There is no substitute for it."
None of this is exceptional for the industry. What is distinctive is the theory underneath it, a theory of compounding that he states like an economist and lives like a student cramming for an exam that never ends. At Stanford's View From The Top series he called wealth a metric that "at some point taps out. On the other hand, there is no end to knowledge." At Berkeley's winter commencement, in the city where MIT's rejection had sent him, he gave the graduates a sentence that could be read off a bank statement: "You must learn with an interest rate greater than the inflation rate of knowledge."
— Aravind Srinivas, TEDAI San Francisco, October 2024Knowledge does not really care about who you are, where you're from or who you have access to. Rather, what matters is the next question you're going to ask.
It is a democratic idea and a sincere one. It also comes from a man who had to wait five hours outside a door to get access.
The same commencement speech asked the graduates, "Who cares what you do next?" What mattered, he said, was "that you do it with great velocity." By then the velocity had made him rich. In October 2025 the M3M Hurun India Rich List put him on its pages at 31 with ₹21,190 crore, its youngest billionaire, and Hurun framed it as a sign of India's shift "from a services-centric past to a deep-tech, product-led powerhouse." A month later, in an email interview with the Hindustan Times, he gave the clearest statement of how he handles criticism: "I've learned that most people who criticise what you are doing aren't doing anything themselves. So who cares?"
On March 30, 2026, he went on the All-In podcast and said something that a lot of people did care about. "The reality is most people don't enjoy their jobs," he said of AI-driven layoffs. "There's suddenly a new possibility, a new opportunity, to use these tools, learn them, and start your own mini business... Even if there is temporary job displacement to deal with, that sort of glorious future is what we should look forward to." The reply that went viral on X was not subtle: "A man worth millions just told the single mother who lost her job that she should be grateful because now she can start a business using his product and called her unemployment a glorious future." A Perplexity spokesperson answered with a statistic, as you'd expect from a company built on citations: Americans had filed 16 million new business applications since Perplexity launched. Eleven weeks later, on 20VC, Srinivas argued that Anthropic's Dario Amodei had done "a disservice to AI" with his messaging about labor replacement.
Both views can be held at once. Most people manage to hold contradictory views. Most people don't have their contradictions transcribed.
There is a smaller indignity in the record, one that seems written for a man who modeled his company on Wikipedia. On August 30, 2025, Wikipedia's editors closed a discussion by merging his biography into the article about his company. For more than a year, anyone who searched for Aravind Srinivas on the encyclopedia was redirected to Perplexity AI. On September 16, 2026, an editor posted on the talk page that he "appears to have received substantial additional coverage in independent reliable sources" and asked whether it was time to give him a page of his own again. As of that writing, the redirect stood.
Your Answer Is Wrong
Near the end of the Hard Fork conversation, after the Chrome bid and the Cloudflare fight and the theory of an internet where humans and agents browse side by side, Srinivas brought up his upbringing. He said his parents "still to today don't actually care about all these valuations." Then he offered the detail that undercuts every number in this profile.
"My mom still is like, your answer is wrong."
Roose laughed and said it was good to know that no matter how successful you got, your mother would always give you the real talk. Srinivas kept going, and the second sentence was the one that mattered.
"Yeah, she's always like, I got this in Google, but your thing doesn't work."
Did he escalate it to engineering? "Of course."
Picture it, because it is the closest thing the story has to a resolution. A woman in Chennai who once pointed at the gates of IIT Madras from a city bus and told her son where he was going types a question into the product he built to replace the search engine she had never stopped using. She types it into Google too. She compares the two. Then she picks up the phone to tell him which one got it right.
Part IIThe Playbook
What follows are twelve operating principles drawn from Srinivas's record: the decisions he made, the rules he stated, and the costs he has incurred so far. They are not all virtues. Several are double-edged, and the most useful lessons here may come from what his approach has cost as much as from what it has won.
The Srinivas Playbook
- 1.Treat the rejection as a curriculum.
- 2.Let your own frustration be the first user.
- 3.Demo to the subject, not the skeptic.
- 4.Launch in the wake of the biggest ship.
- 5.Build what the incumbent's income statement forbids.
- 6.Buy the commodity, own the orchestration.
- 7.Replace the deck with the database.
- 8.Ship at eighty and make peace with problems existing.
- 9.Make offers that work as questions.
- 10.Run the company at the temperature of fear.
- 11.Price trust before attention, even if you change your mind.
- 12.Permission debts compound faster than technical debt.
Principle 1
Treat the rejection as a curriculum
The 0.01 CGPA miss is the founding wound, and how Srinivas responded to it is worth studying. He didn't appeal. He didn't try again by the same route. He withdrew into a "whole shell" of online lectures and competitive programming and came out a year later knowing Python better than the computer science students he had envied. The professor who said "I don't usually let electrical people take this class" became a benchmark, and he topped the course. When MIT said no, Berkeley said yes, and he arrived with a Bengio recommendation he had earned by working around the official channel.
The pattern is consistent. Each rejection names a skill the gatekeeper assumes you lack. Srinivas treats that assumption as a syllabus and then makes the gatekeeper beside the point. He admits the early fuel was "vengeance," which is honest, and also a warning. Vengeance gets you through a year in a shell. It is a poor guide for deciding which gates actually deserve respect, a distinction the later robots.txt fights would test.
Tactic: Write down the last rejection you received, name the specific capability it implied you lacked, and build public evidence of that capability through a route the gatekeeper doesn't control.
Principle 2
Let your own frustration be the first user
Perplexity's mission came out of a failed Google search for startup health insurance. The founders didn't run user interviews to find the problem. One of them had the problem, on a Tuesday, with a real hire waiting on the answer. The prototype answered better than the incumbent did, and the founders noticed their own behavior before they noticed a market.
This matters because Srinivas has said the idea maze kills more founders than competition: "I've seen many founders spend at least six months to a year in the idea maze going around and around and not getting anywhere." The way out was not a better whiteboard. It was paying attention to which tool he reached for when it counted. The prototype had been sitting there for weeks. It became the company only when it was the thing that worked under pressure.
Tactic: For one week, log every moment you abandon a tool mid-task, and treat the tool you switch to, including your own prototypes, as data about what the market already wants.
Principle 3
Demo to the subject, not the skeptic
The seed round closed because Srinivas showed Yann LeCun the accounts that replied most often to LeCun's own tweets. Jeff Dean, Nat Friedman, and Andrej Karpathy got the same kind of personal demo. "It was very relatable for them to go and search about their own Twitter." Elite investors have seen every pitch. They have not seen themselves rendered in a new medium.
Note the order. Famous investors created what Srinivas calls "snowball effects": "People take you more seriously." Each name made the next one easier. Huang's daily use became a talking point, and the talking point brought users. Vanity was not the product, but it was a very efficient distribution channel at the stage where every channel is expensive.
P
The Snowball, Round by Round
Each name and each round lowered the cost of the next.
Sep 2022
$3.1M seed after personalized demos to LeCun, Dean, Friedman, Karpathy, and Gil.
Mar 2023
$26M round; iOS app launches.
Jan 2024
$73.6M at a $520M valuation, with Nvidia and Jeff Bezos.
Apr 2024
Valuation crosses $1B.
Jun 2024
SoftBank's Vision Fund 2 invests; valuation reaches $3B.
Dec 2024
$500M round led by IVP at $9B.
Jun 2025
$500M round at $14B.
Sep 2025
$20B valuation reported.
Early 2026
$21.21B after Series E-6; $750M Azure agreement with Microsoft.
Tactic: Before your next investor meeting, build a version of the demo in which the investor's own data, work, or reputation is the input.
Principle 4
Launch in the wake of the biggest ship
Perplexity launched seven days after ChatGPT. The conventional move would have been to wait for the hype to fade so your product doesn't get lost. Srinivas did the opposite. ChatGPT had just taught millions of people to type questions into a box and expect prose back, and it had also shown them its limits: no web access, no citations, confident fabrication. Perplexity arrived as the obvious fix to a problem the public had discovered the week before.
The 700,000 queries on New Year's Eve, despite a seven-second latency and a name its own CEO calls "crappy," showed that the wake was carrying it. Timing like this is not luck you wait for. It requires having something ready to ship when the giant moves, which is why the prototype mattered more than the pivot.
Tactic: Keep a shippable version of your product on a short leash, and when the category leader makes a major launch, release within the week as the answer to whatever the leader's launch reveals it can't do.
Principle 5
Build what the incumbent's income statement forbids
"Make the weakness of your enemy a strength." Srinivas's reading of Google is not that it lacks talent or technology. Google invented the transformer. His point is that Google can't aggressively deploy any answer format that discourages clicking on a link, because clicks are the business. The opening isn't technical. It's a matter of accounting.
His own later admission that the AdWords disruption only came into focus after launch is useful. You don't need to know at the start which incumbent revenue line you threaten. You need to notice, once users start showing up, why the incumbent hasn't already copied you. The answer to that question is your moat for as long as the incumbent's income statement stays the same.
Tactic: List the three features your largest competitor could build in a quarter but hasn't, and for each, trace which of their revenue lines it would cannibalize. Build the one with the biggest line.
Principle 6
Buy the commodity, own the orchestration
Perplexity has never tried to win the frontier-model race. Its products route across OpenAI, Anthropic, open-source Llama and DeepSeek derivatives, and its own fine-tunes. Srinivas's reasoning is cold: the labs "are all hill climbing on exactly the same benchmarks," so their models become "completely undifferentiated, which is essentially the necessary criterion for it being a commodity." Every price cut by a model provider goes straight into his margin.
The engineering effort goes elsewhere, into what he calls the "much harder challenges... outside the models": retrieval, latency, browser control, orchestrating tools, internal evals for agent reliability. His compute budget, "tens of thousands of GPUs... not a million," is sized to that bet. The risk is plain: a company that owns no model depends on several that might become competitors. He has hedged it by spreading the dependency across many suppliers, not by eliminating it.
Tactic: Inventory every component of your stack, mark which ones have three or more credible vendors racing on price, and move engineering hours off those components and onto the layer that connects them.
Principle 7
Replace the deck with the database
Srinivas has never made a pitch deck for a Perplexity round. He writes a memo, offers two hours of open Q&A, pulls live metrics on request, and tells investors that anything non-confidential they can "ask Perplexity. It already knows everything." It's partly theater, since the product is doing the pitching. It's also a real filter: investors who need slides to be persuaded aren't the ones he wants.
The method only works if the numbers can stand being looked at directly. It shifts the burden from storytelling to transparency, and it signals the confidence the Hindustan Times interview described: "You have to show them a good product, and you also have to show them that you personally are tough, committed, and capable."
⚖
Fundraising: Convention vs. Srinivas
How he restructured the most ritualized meeting in startups.
| Conventional practice | Srinivas's approach | Effect |
|---|---|---|
| Polished 15-slide deck | Written memo, no slides | Signals depth |
| Curated metrics in an appendix | Live pulls during the meeting | Builds trust |
| Founder controls the narrative | Investor drives two hours of Q&A | Filters investors |
| Product shown in a demo video | "Ask Perplexity. It already knows everything." | Product is the pitch |
Tactic: For your next raise, send a two-page memo in advance and spend the meeting answering questions with live data, and treat any investor who asks for the deck as information.
Principle 8
Ship at eighty and make peace with problems existing
At Berkeley Haas, Srinivas named what he considers a CEO's most important skill, and it isn't vision or hiring. "I used to think every problem just needed to be fixed instantly. As the company has scaled, I have learned to make peace with some problems just existing." Perplexity ships products at about 80 percent and improves them in public.
The principle has two halves that pull against each other. The first is velocity: in a field that moves weekly, polish is a luxury that costs you the window. The second is triage: some problems aren't worth fixing yet, and an executive who can't stand that will drown. Mid-2025 accounts say Srinivas still personally triages production bugs, so "peace" doesn't mean indifference. It means ordering problems. The cost appears where an 80 percent product meets people with legal standing. The "rough edges" defense of the Forbes incident was the 80 percent rule meeting a party that didn't accept it.
Tactic: Sort your open issues into those that hurt users, those that hurt third parties, and those that merely bother you, and ship past the third category on purpose, but never past the second.
Principle 9
Make offers that work as questions
The $34.5 billion Chrome bid and the TikTok merger proposal looked absurd: offers larger than the bidder, contingent on court rulings and government action Perplexity didn't control. Srinivas admitted on air that Chrome was "unlikely" to be divested and that any ruling would be appealed for years. He bid anyway, because a public price changes everyone else's options. Google can no longer claim no buyer exists. The court has a reference point. The press has a story that costs Perplexity nothing.
"You will lose 100 percent of the shots you don't take" is a cliché, but the analysis underneath it isn't. An offer can be a probe that asks every party what they believe about the asset's value and the likely ruling. The condition is credibility. Srinivas had lined up "three or four investors" before bidding, so the question was serious enough that it couldn't be dismissed as a joke.
Tactic: Identify one asset in your industry whose ownership may change through regulation or distress, line up the backing to make a credible conditional offer, and make the offer before anyone else posts a price.
Principle 10
Run the company at the temperature of fear
"You should always assume that a model company will copy it," Srinivas told Y Combinator's AI Startup School. "You've got to live with that fear and you have to embrace it." Elsewhere: "There's real benefit from embracing that fear and sleeping with that fear and waking up every day and feeling excited about what you're going to build." Every launch cycle of Google's AI Overviews brings a round of "Perplexity is dead" posts, and he treats them as weather.
The useful part is the assumption that copying is certain, not just possible. That changes how you plan. Moats built on features erode on a known schedule, so the only lasting advantages are speed and identity: "Realize that your moat comes from moving fast and building your own identity around what you're doing because users at the end care." Fear, in this framing, isn't an emotion to manage. It's an input to the roadmap.
Tactic: For each major feature you ship, write down the date by which you expect the largest platform in your category to copy it, and plan the next release to land before that date.
Principle 11
Price trust before attention, even if you change your mind
In 2024 Srinivas told 20VC that cracking advertising relevance would make Perplexity "worth way more." In February 2026 the company dropped AI-integrated ads entirely and went subscription-first, saying it wanted to protect user trust in the answer engine. The reversal is the lesson. The whole case against Google rested on ad incentives distorting answers. Putting ads inside the answer would have brought that same distortion into Perplexity's core product.
It is unusual for a company at a $20 billion valuation to walk away from the revenue model that could most easily justify it. Srinivas has framed the company's purpose as "provide truthful, accurate answers," and has said the way things go wrong is "we lose sight of our core objective mission." Choosing subscriptions is choosing the customer who pays to be told the truth over the advertiser who pays to be seen. That he once argued the other way makes the choice more credible, not less.
Tactic: Write down the single revenue source that would most compromise your product's core promise, and decide now, in writing, the conditions under which you would refuse it.
Principle 12
Permission debts compound faster than technical debt
This is the cautionary principle, and it may be the most important. Perplexity's speed came partly from treating the web's permission signals (robots.txt, firewalls, trademarks, publishers' terms) as obstacles, the way Srinivas has treated every gate in his life. The bill is long: Forbes's accusations, Wired's logs of an undisclosed IP address, lawsuits from Dow Jones and three major Japanese newspapers, a New York Times cease-and-desist letter, a BBC threat, a Reddit suit, a trademark case over the company's own name, and Cloudflare's very public "stealth crawling" report.
The defenses have been clever: the distinction between a crawler and a user agent, the argument that media companies "wish this technology didn't exist." Some may win in court. What they can't do is reverse the compounding. Each fight made the next counterparty less willing to settle and more willing to sue. The company's eventual moves (a revenue-sharing program, paid Wikimedia Enterprise access, the end of advertising) look like payments on a debt that would have been cheaper early.
!
The Permission Ledger
A partial record of conflicts and settlements over content access.
| Date | Counterparty | Issue | Type |
|---|---|---|---|
| Jun 2024 | Forbes | Story allegedly copied without prominent credit | Accusation |
| Jun 2024 | Wired / Robb Knight | Undisclosed IP ignoring robots.txt | Investigation |
| Jul 2024 | TIME, Fortune, Der Spiegel | Publishers' revenue-sharing program | Deal |
| Oct 2024 | Dow Jones / New York Post | Copyright, hallucinated attributions | Lawsuit |
| Oct 2024 | New York Times | Cease-and-desist over scraping | Threat |
| Jan 2025 | Perplexity Solved Solutions | Trademark on the name itself | Lawsuit |
| Jun 2025 | BBC | Demand to stop scraping, delete data, pay | Threat |
| Aug 2025 | Yomiuri, Asahi, Nikkei | Copyright; 120,000 Yomiuri articles | Lawsuit |
| Aug 2025 | Cloudflare | "Stealth" crawlers bypassing blocks | Investigation |
| Oct 2025 | Unlawful scraping | Lawsuit | |
| Jan 2026 | Wikimedia Enterprise | Paid API access instead of scraping | Paid access |
Tactic: Treat every third-party permission signal your product touches as a line item on the balance sheet, and settle the cheap ones early, before a public fight turns them into lawsuits.
Free playbook
Get The Business Model Playbook
58 business models, one visual page each: how the money flows, the metrics that matter, and who runs it. Free when you join the Faster Than Normal email.
Free. No spam. Unsubscribe anytime.
Part IIIQuotes and Maxims
In their words
— Aravind Srinivas, Lex Fridman Podcast #434, June 2024The principle in Perplexity is you're not supposed to say anything that you don't retrieve.
— Aravind Srinivas, Y Combinator AI Startup SchoolYou should assume that if you have a big hit, if your company is something that can make revenue on the scale of hundreds of millions of dollars or potentially billions of dollars, you should always assume that a model company will copy it. You've got to live with that fear and you have to embrace it.
— Aravind Srinivas, Dean's Speaker Series, UC Berkeley Haas, 2025I used to think every problem just needed to be fixed instantly. As the company has scaled, I have learned to make peace with some problems just existing. That is the number one skill you need as a CEO or a founder.
— Aravind Srinivas, View From The Top, Stanford GSB, 2025You can probably focus on wealth and your net worth, you can try to use that as a metric for your progress in life, but at some point it taps out. On the other hand, there is no end to knowledge.
— Aravind Srinivas, UC Berkeley winter commencementYou have to compound what you know. You must learn with an interest rate greater than the inflation rate of knowledge.
Maxims
- The margin is a syllabus. A rejection by a hundredth of a point tells you exactly what to go learn on your own.
- Your annoyance is a market. The tool you reach for when the incumbent fails you is the product nobody has built yet.
- Show people themselves. The fastest demo puts the audience's own name in the results.
- Ship in the giant's wake. The week after a category-defining launch, the public knows both what it wants and what it's missing.
- Read the enemy's income statement. The features an incumbent won't build are visible in its revenue lines.
- Let the suppliers race. When models commoditize, the orchestration layer collects the savings.
- Data over decks. If your numbers survive a two-hour live interrogation, you don't need slides.
- Eighty percent, then the world. Ship before you're comfortable, and order your problems instead of trying to fix them all.
- A bid is a question. A credible offer for an asset you probably won't get still sets the price.
- Ask before the gate closes. Permission is cheapest when requested early and most expensive once a court has to decide.
