Perplexity CEO Aravind Srinivas: Why Tiny Companies Could Drive US Growth

Perplexity CEO Aravind Srinivas argues that small, AI-native companies will drive American growth, and that AI agents belong on hardware you already own. He makes the case to Molly O’Shea on the October 12, 2026 episode of Sourcery, a 55-minute interview recorded at Perplexity’s new San Francisco office. It covers Perplexity Computer, the company’s agent product, a one-person business that spends about $1 million a year on it, the economics of running models locally, why Perplexity will not build data centers, and how he thinks about agent security after the Hugging Face incident.

TLDW

Srinivas says Perplexity’s revenue is up five times year over year because agents need search, a browser and a way to route work across models, and Perplexity already had all three. He expects future companies to be smaller and more profitable, and he wants Perplexity Computer to be the neutral tool they run on. He is betting that long-running personal agents will move partly onto local hardware, with Perplexity sitting between the laptop and the cloud. His closing claim is that AI products should be judged on depth of use, which he calls recursive effects, and not on user counts.

Thoughts

One customer carries a lot of weight in this interview: a single person who spends upwards of $1 million a year on Perplexity and earns $3 million automating newsletters for other businesses. Srinivas offers it as proof of a “recursive effect,” where the customer makes money, so Perplexity makes money. Look at it from the customer’s side. A third of revenue goes to one supplier, the product is machine-written newsletters that work only as long as recipients keep opening them, and the margin depends on that supplier’s pricing. That is a real business, and it is also a fragile one. It is one unnamed example, and it says more about what a determined operator can do with arbitrage than about what a typical small business will do with agents.

Srinivas says the best engineers are not short of code. They are short of time to work out what to build, and writing the code is “dead simple” once the problem and constraints are clear. I think that is right, and it sits awkwardly beside two other things he says. He wants an AI that tells you your idea is stupid and proposes a better one, yet his product philosophy, borrowed from Larry Page, is that the user is never wrong. And when asked about his biggest challenge, he says models cannot yet do research where the answer cannot be checked, because they cannot handle slow real-world feedback. Deciding what to build is exactly that kind of problem. Perplexity is aiming its product at the place where, by his own account, the models are weakest.

The local compute argument is the most concrete part of the hour. Chat costs about 30 times more to run on your own hardware than in a data center, he says, but at 100,000 tokens of context the data center’s advantage falls to roughly three times, or close to nothing. He adds that 100 million Apple silicon machines at 50 watts each is already gigawatts, and admits that is envelope math. I could not check the cost figures, and the incentive is plain. Perplexity has no data centers of its own and no plans to build any, so a future where the customer supplies the hardware and the electricity suits it very well. He says so himself when he describes Perplexity as the one company with no business model risk in routing work to a laptop. The bus and car comparison is fair, though. People do pay more for the thing they control.

On security his answer is to run things locally, and that covers less than he suggests. Keeping data off someone else’s servers addresses who can see it. It does not address what he describes a minute later: agents inside a sandbox doing things they are not supposed to do, which goes unnoticed because models are tested one agent at a time. An agent on your own machine with your files, your photos and your home sensors is still an agent. The useful part of his answer is the unglamorous part, which is audit logs, detection classifiers and a second agent watching the first.

The line that will travel is that forward deployed engineers exist because the product is bad. It is a good line. It is harder to square with the sovereign, air-gapped installations he describes later, where Perplexity sees no bugs and admits the experience may be rough. Those are the customers who will want an engineer on site.

Key Takeaways

Srinivas says Perplexity’s revenue is growing five times year over year, and that enterprise customers signing themselves up grew faster this year than individual paying users.
The product path was an answer engine in 2022, a multi-step search agent in late 2023, deep research in 2024, browser agents in 2025, then Perplexity Computer, a cloud agent that works in a sandbox with search, a browser and files.
One solo operator spends upwards of $1 million a year on Perplexity and makes $3 million a year automating newsletters that other businesses use to win clients.
He thinks US GDP growth may come more from small and mid-market companies that are capital efficient and headcount efficient. Investors favor Fortune 500 customers because that revenue looks durable.
A trillion-dollar company can be built on US customers alone, he says, citing Berkshire Hathaway and JPMorgan. The way to get there is to be useful to a specific group and let usage build.
Perplexity’s first 10 to 20 hires did week-long work trials. Good fits were usually clear in two days. If it was still unclear after seven, that was the answer.
His product rule comes from Larry Page: the user is never wrong. People should not have to learn prompting, skills files, sub-agents or which model is good at what. If the AI fails, it is a product problem and not a skill problem.
He calls the model labs contributors and not competitors. Each new frontier model adds something Perplexity can offer, and price competition between labs lowers its serving costs.
By his numbers, chat is about 30 times cheaper in a data center than on local hardware, but for long-context agent work the gap narrows to about three times or less.
Perplexity ships two directions of the same idea. Portable Computer, built with NVIDIA, runs fully local and can escalate to the cloud. Hybrid compute on Apple silicon starts in the cloud and moves sensitive files to the local machine.
Perplexity has no plans to build data centers. Srinivas lists the obstacles others face: permits, interconnection queues, memory supply, grid access, gas turbines and local protests.
Sovereign funds are asking for versions of Perplexity Computer that run entirely on hardware they own, where Perplexity sees nothing. The cost is that Perplexity cannot see bugs and can only improve it through software updates.
He expects new models to roll out more slowly after the Hugging Face incident, and says Perplexity has open source monitoring tools named Numbat, for servers, and Bumblebee, for devices.

Chapters

1:30 A Brand Built on Curiosity

Srinivas says Perplexity’s brand celebrates curiosity because asking questions is the one human quality that matters more as AI gets better at answers. He cites physicist David Deutsch on humans being curious about what is already familiar. The meeting rooms are named after scientists, an idea he took from Demis Hassabis at DeepMind. His favorite is Ramanujan, as a symbol of intuition. He argues the best company brands are about people and not technology, pointing to Apple and NVIDIA.

8:18 A Trillion-Dollar Company on America Alone

US customers spend heavily on products they love and keep starting new kinds of businesses, he says, so serving them is enough. Perplexity Computer is meant to let new companies hand work to agents, mainly because hiring the right people early is so hard. He describes his own trial-week hiring as a survival mechanism and not a strategy. He then makes the central claim that future companies will be smaller and richer. Asked about the American dream, he says no other country invests in outsiders this way, and that ambition here is contagious.

16:28 Stop Blaming Users for Bad Prompts

His entrepreneurial hero is Larry Page, for the principle that the user is never wrong. Engineers tend to tell users what they should have done, when the product should have anticipated it. He compares AI today to computers before Steve Jobs made them usable by ordinary people. Companies paper over hard products with forward deployed engineers, and he wants the routing between models to happen inside the product. He concedes Perplexity is not there yet.

22:55 How Perplexity Grew Five Times

Agents doing computer work need the internet, he says, the same way a person cannot work with the Wi-Fi off. Perplexity had search, a browser and multi-model routing before agents had a name. He credits the frontier labs for making models good enough at code that an agent can use search and a browser as code inside a sandbox, with files to hold intermediate work. The interface stayed a prompt box, and businesses began onboarding themselves.

26:51 The $1 Million Solopreneur and Recursive Effects

The one-person newsletter business is his example of recursive effects: the customer’s income depends on the product, and Perplexity’s income depends on the customer. He says people wrongly apply social media logic and treat anything short of a billion users as failure. He still wants millions of users, with deep usage among them. The same logic covers the model labs, whose progress and price cuts flow through to Perplexity.

29:20 The Data Center Is Coming to Your Laptop

Data centers win on short, repeatable requests because batching many users together is efficient. Personal agents with long, unique context lose most of that benefit. Add slow data center construction and hundreds of millions of capable devices already plugged in, and local hardware starts to matter. He also raises privacy: people may not want a frontier lab holding their home sensors, photos and messages. With NVIDIA the product starts local and reaches up to the cloud, and with Apple it starts in the cloud and reaches down. Finance, legal and healthcare customers are the most interested.

36:11 Sovereign Deployments and No Data Centers

Sovereign funds want the whole system, with a fine-tuned open weights model, running on their own hardware. The conversation sets that against an advertising company offering a personal agent that can read your blood work and tax returns. On data centers, he says there is no shortage of builders, only of permits, parts and power, and he would sooner solve a different problem. He argues even Microsoft has less reason to push local computing, because its stock is tied to Azure.

41:25 Ask AI to Think, Not Just Write Code

O’Shea asks about hype cycles in coding agents and neoclouds. Srinivas says neocloud growth proves demand for tokens, but another neocloud is not differentiation. Perplexity is not chasing coding. He wants an AI that, asked to clone a product, questions the business model, proposes what you probably meant, lays out the spec and risks, and then builds it. He says more people who think that way would make Perplexity worth ten times as much. Being neutral across models, he adds, is a trust signal as much as a feature.

45:47 Agent Security and the Hugging Face Incident

His first answer to rising cyber threats is a version of the product that runs on hardware you own, including fully air-gapped. He is skeptical of security vendors whose fix is to route everything through their own gateway. On the Hugging Face incident, he separates one main agent with sub-agents, which is common, from several parent agents collaborating, which is riskier and rarely tested. He wants guarded sandboxes, classifiers trained on audit logs, and agents that handle remediation.

50:45 Turning Everyone Into a Scientist

What he wants next is research over longer horizons: building prediction models, running experiments in the world and refining hypotheses. The limit is capability. Models improve quickly in math and code, where correctness can be verified, and struggle where answers need real-world rollouts. Perplexity will start with its own research teams, framed as what you would do with ten tireless interns. His hottest take closes the interview: AI products have recursive effects, not network effects.

Notable Quotes

“It’s possible that the future of American GDP growth comes more from smaller and mid-market companies that are going to be far more productive, capital efficient, headcount efficient and create value way faster using these tools by being more AI native and less bureaucratic.”
Aravind Srinivas, on where growth comes from next

“The FDE is needed because product is bad.”
Aravind Srinivas, on forward deployed engineers

“He’s making 3 million a year by automating newsletters for others.”
Aravind Srinivas, on the one-person company spending about $1 million a year on Perplexity

“Think of the data center like a bus. Shared ride is obviously cheaper, but that didn’t stop people from wanting to have their own car.”
Aravind Srinivas, on why local AI hardware will sit alongside the cloud

“I don’t want an AI that just spits out slop that makes a vibe coded clone.”
Aravind Srinivas, on wanting AI to challenge the request before building it

“Most of the times you test these models in single agent settings and so you don’t discover these threats.”
Aravind Srinivas, on what the Hugging Face incident exposed

“AI is a recursive product. It’s meant to have recursive effects, not network effects. So, the depth of usage and the quality of users matters a lot more than the quantity.”
Aravind Srinivas, giving his hottest take at the end

The revenue, cost and customer figures here are Srinivas’s own and were not checked independently. Watch the full conversation here.