> ## Content Index
> Fetch the complete content index at: https://nnutshell.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Perplexity grew revenue 5x by betting on depth, not users
- URL: https://nnutshell.com/perplexity-grew-revenue-5x-by-betting-on-depth-not-users/
- Published: 2026-10-22T11:48:52.000Z
- Updated: 2026-10-24T03:10:19.000Z
- Description: Perplexity CEO Aravind Srinivas on recursive usage over network effects, orchestrating local and cloud compute instead of building data centers, and why curiosity is the brand.
- Author: Admin
- Tags: Aravind Srinivas, Larry Page, Demis Hassabis, Perplexity, Nvidia, Apple, Hugging Face, Agentic AI, Edge Computing, Inference, #rec-2026-10-12, #Import 2026-10-25 09:54

*Recorded 12 October 2026 · 55 min*

AI products do not follow the network effects of social media. Aravind Srinivas, co-founder and CEO of Perplexity, argues they run on recursive effects: depth of usage matters more than user count, because customers who make money with the product keep spending on it. Perplexity's revenue is up 5x year over year on that logic, and one solopreneur spends upwards of $1 million a year on it.

The interview covers Perplexity Computer, its agent product, and a bet that the data center is a bottleneck while hundreds of millions of devices sit underused. The tension is neutrality. Srinivas wants to be the orchestrator across models, chips, local hardware and cloud, and calls the frontier labs contributors rather than competitors. That position depends on cost claims he offers without data. He also admits the product is not yet intuitive enough, and that no model can yet run end-to-end research on questions that cannot be verified.

1. **Curiosity as brand** — [▶ 1:45](https://www.youtube.com/watch?v=1C8%5Fn9nnx5Y&t=105s&ref=nnutshell.com)  
"Our brand is fundamentally to celebrate curiosity." The reasoning: AI can answer anything, so the human role is to ask the next set of questions. People grow up rewarded for answers, and it is no longer cool to be a know-it-all. Citing physicist David Deutsch, he says humans are the only species that can be curious about what is already familiar. The best brands, he argues, have nothing to do with the technology. Apple was about beautiful things, and Nvidia's mission was to make possible what only supercomputers can, which could apply to quantum computing tomorrow.
2. **Useful, not trillion-dollar** — [▶ 8:34](https://www.youtube.com/watch?v=1C8%5Fn9nnx5Y&t=514s&ref=nnutshell.com)  
A $1 trillion company can be built on the US alone. Berkshire Hathaway is worth about $1 trillion and JPMorgan about $940 billion, both mostly on US customers who spend repeatedly with high lifetime value. The strategy is the most trustworthy model-neutral agent platform, orchestrating frontier and open-weight models. Do not optimize for the number: be useful at scale and a decade later you are worth it. He thinks future GDP growth may come from smaller, headcount-efficient, AI-native companies, even though growth investors prefer Fortune 500 accounts because that revenue looks durable.
3. **The user is never wrong** — [▶ 16:48](https://www.youtube.com/watch?v=1C8%5Fn9nnx5Y&t=1008s&ref=nnutshell.com)  
Srinivas credits Larry Page for the principle: when the product fails, fix the product, not the user. As models advance, users' understanding will not keep pace, so nobody should have to learn skills files, sub-agents or which model suits which task. Forward-deployed engineers exist because products are bad. Imagine the iPhone needing an FTE to onboard everyone. Not being able to get the AI to do something should be seen as a product issue, not a skill issue. He concedes Perplexity is "not there yet".
4. **Agents need search** — [▶ 23:21](https://www.youtube.com/watch?v=1C8%5Fn9nnx5Y&t=1401s&ref=nnutshell.com)  
Perplexity went from answer engine in 2022 to multi-step search agent in late 2023, deep research in 2024 and browser agents last year. Agents need the web, a cloud browser, personal context, a multi-model harness and low-cost cloud sandboxes, all behind a prompt box. Frontier labs made models good at code, which made sandboxes viable. Result: revenue up 5x year over year, with enterprise growing faster than prosumer this year. People onboarded themselves through self-serve, and running in the cloud avoids the setup an OpenClaw-style tool needs.
5. **Recursive, not network effects** — [▶ 26:49](https://www.youtube.com/watch?v=1C8%5Fn9nnx5Y&t=1609s&ref=nnutshell.com)  
A single-person company spends upwards of $1 million a year on Perplexity while making $3 million automating newsletters for others. That is the recursive effect: customers make money with the product, so they spend more on it. Investors wrongly apply the social media rule that fewer than a billion users means failure. Millions of users and real distribution are still needed, but depth matters more. "We don't have competitors, we have contributors." New frontier models add capability, and competition lowers prices and serving costs, which raises usage.
6. **Local hardware economics** — [▶ 29:18](https://www.youtube.com/watch?v=1C8%5Fn9nnx5Y&t=1758s&ref=nnutshell.com)  
Chat is best served from the data center, and would cost 30x more locally. At 100k tokens of context, the data center advantage shrinks to about 3x or close to 1x, because unique context dominates batching. Data center build-outs are stuck on interconnection queues, permits, protests and gas turbines. By envelope math, 50 watts across 100 million Apple silicon machines is gigawatts. Data center is the bus, local hardware is your own car. Perplexity's portable computer with Nvidia runs locally and escalates to the cloud, while the Apple version starts in the cloud and drops local for sensitive files.
7. **Own the runtime** — [▶ 36:48](https://www.youtube.com/watch?v=1C8%5Fn9nnx5Y&t=2208s&ref=nnutshell.com)  
Sovereign funds want Perplexity Computer on hardware they own, with the company seeing nothing. The cost is no bug visibility, a junkier experience and software-update-only improvement. He warns against always-on agents from ad companies holding blood work, tax returns and photos. On cybersecurity, a vendor's gateway asks you to trust the vendor: "The only thing you can trust is yourself and the stuff you own." The answer is an airgappable runtime on your own device or on-prem hardware, accepting a compromise on frontier intelligence. Finance, legal and healthcare have shown interest.
8. **Skipping the data center race** — [▶ 39:30](https://www.youtube.com/watch?v=1C8%5Fn9nnx5Y&t=2370s&ref=nnutshell.com)  
Perplexity has no plans to build data centers. Srinivas expects its success to increase data center demand, but the shortages are permits, HBM supply, grid access and turbines, and he does not want those problems. The neutral orchestrator across models, chips, local and cloud has no business model risk, whereas even Microsoft is tied to Azure and OpenAI models. NeoClouds prove token demand exists, but without their own power or software they are undifferentiated. Multi-model is a trust signal for people building on the platform, not only a feature.
9. **Coding is not the bottleneck** — [▶ 41:20](https://www.youtube.com/watch?v=1C8%5Fn9nnx5Y&t=2480s&ref=nnutshell.com)  
Ask the best engineers, he says, and they are bottlenecked on problem solving, not writing code. Once specs, system design and constraints are known, code is "dead simple" with AI. He wants an AI that answers a request to clone a product by questioning the business model, then supplying the spec, security risks and go-to-market risks before building. More people with that clarity of thought would make Perplexity 10x bigger in market cap. So while others chase coding tokens and NeoClouds, Perplexity aims at research, analysis and scoping.
10. **Multi-agent threats** — [▶ 48:27](https://www.youtube.com/watch?v=1C8%5Fn9nnx5Y&t=2907s&ref=nnutshell.com)  
After the Hugging Face incident, he expects model rollouts to slow, since testing happens in single-agent settings and misses these threats. A main agent with sub-agents is already widely deployed and less risky than multiple collaborating parent agents, each with sub-agents and heavy compute. The fix is secure sandboxes, background guard agents and detection classifiers trained on audit logs. Perplexity is releasing open-source tools: Bumblebee for local devices and Numbat for servers. Every company will also need a remediation agent, built on some harness.
11. **Long-horizon science** — [▶ 50:49](https://www.youtube.com/watch?v=1C8%5Fn9nnx5Y&t=3049s&ref=nnutshell.com)  
The next target is longer research tasks: building prediction models, running experiments and testing hypotheses in real time, so "everybody can turn into a scientist". The biggest obstacle is capability. No frontier model can yet research autonomously end to end. Math and code are verifiable and can be hill-climbed, but most curiosities lack a clear way to verify truth and need slow real-world feedback. The interface is also unclear. His thought experiment is 10 research interns working 24/7 on your team, starting with Perplexity's internal research teams.

AI products do not run on network effects. They run on recursive effects, and the difference decides how you build a company. Aravind Srinivas, co-founder and CEO of Perplexity, treats the social media reflex, that anything under a billion users is a failure, as a category error. Perplexity still needs millions of users and real distribution. What it needs more is depth: users whose work with the product earns them money, so that they spend more as they earn more. Usage compounds inside each account rather than across a graph.

The evidence is revenue and one customer. Perplexity's revenue is up 5x year over year. Enterprise grew faster than prosumer this year, mostly through self-serve onboarding. And one solopreneur, a single-person company, spends upwards of $1 million a year on the product while making $3 million automating newsletters for other people. Those clients then win business of their own. That chain is what Srinivas means by recursion. The customer's success is the company's revenue.

Why Perplexity and not someone else? Because an agent that does real work needs what Perplexity built for other reasons. His test is simple: shut off Wi-Fi and see whether you can work. An agent needs the web, a cloud browser, personal context across files and connectors, a harness that routes between models, and cloud sandboxes with low serving costs, all behind a prompt box that works from a phone. The company started as an answer engine in 2022, moved to multi-step search in late 2023, deep research in 2024 and browser agents last year. Then frontier labs made their models good at code, which made the sandbox approach viable. Hence the line: "we don't have competitors, we have contributors." Every new frontier model adds capability. Competition between labs cuts prices, which cuts serving costs, which raises usage.

The same logic extends to hardware, and this is where the bet becomes physical. Srinivas argues the cost case for data centers weakens as context grows. Chat is best served from the data center, and would cost 30x more locally. At 100k tokens of context, the advantage falls to about 3x, or close to 1x. Batching rewards repeatable workloads with small contexts. It does not help when each user's context is unique and huge. Meanwhile the build-out is stuck on interconnection queues, permits, protests, gigawatt requirements, substations and gas turbines. Against that, 50 watts across 100 million Apple silicon machines is gigawatts already plugged into the grid, by his own envelope math. The data center is a bus, local hardware is your own car, and people will want both. Perplexity will not build data centers. It will orchestrate in both directions: a portable computer with Nvidia that runs fully local and escalates to the cloud, and a setup with Apple that starts in the cloud and drops to local silicon for sensitive files. Finance, legal and healthcare have shown the most interest.

Neutrality holds this up. Srinivas says Perplexity can orchestrate across models, chips, local and cloud without a business model risk. Even Microsoft, he argues, is too tied to Azure and OpenAI models to do it. Privacy turns neutrality from a convenience into a product. Sovereign funds want a deployment on hardware they own, where even Perplexity sees nothing. Security vendors that sell a gateway are asking you to trust the vendor. "The only thing you can trust is yourself and the stuff you own." The price is a worse product: no visibility into bugs, a junkier experience, and a compromise on frontier intelligence.

Then there is the question of what the product should do. Srinivas rejects the coding race. Writing code, he says, is nearly solved once the specs and constraints are known. The bottleneck is deciding what to build. He wants an AI that, asked for a clone of a product, questions the business model and returns the spec, the security risks and the go-to-market risks before writing anything. With more people like that, he says, Perplexity could be 10x bigger in market cap. His customers are small and mid-market firms, not the Fortune 500\. Growth investors like big accounts because the revenue looks durable. Srinivas thinks American GDP growth may come instead from smaller, headcount-efficient, AI-native companies, and that the US alone can support a $1 trillion company, citing Berkshire Hathaway and JPMorgan at about $940 billion on mostly American customers. He insists on not optimizing for the number. Be useful at scale and a decade later you are worth it. Serving those firms means a product that needs no forward-deployed engineers. Borrowing a principle from Larry Page, the user is never wrong. A failure to get the AI to work is a product issue, not a skill issue.

The obvious objection is that much of this is asserted. The 30x, 3x and 1x cost figures come without data, and the power figure is, in his words, envelope math. The neutrality thesis assumes the frontier labs keep behaving as contributors. And he concedes the product is not there yet. The harder limits sit in capability. No frontier model can do research end to end on its own. Math and code can be verified and hill-climbed. Most of what people are curious about cannot, and the answers need real-world rollouts with slow feedback. Perplexity wants to turn everyone into a scientist, but he admits the interface for that is unclear. The risks are also growing. Multi-agent systems, especially several parent agents each with sub-agents, slip past testing done in single-agent settings. After the Hugging Face incident he expects model rollouts to slow. Perplexity's answer is open-source audit tooling, Bumblebee for devices and Numbat for servers, plus detection classifiers and remediation agents.

What is unresolved is whether depth really substitutes for scale, and whether the cost crossover he describes holds beyond long-context agent work. Srinivas is betting on every layer at once: progress in chips, in open-weight models and in frontier models. The strategy works only if all three keep moving.

**The open question.** The load-bearing claim is that long-context agent work erases the data center's cost advantage, from 30x in favor of the cloud for chat to roughly 3x or even 1x at 100k tokens of context. It is stated without data, and the whole local-and-cloud orchestration thesis rests on it. Watch whether Perplexity publishes cost comparisons from the Nvidia portable computer or the Apple hybrid, and whether sovereign customers accept a product that is, by his own admission, junkier.

Watch the full video at [https://www.youtube.com/watch?v=1C8\_n9nnx5Y](https://www.youtube.com/watch?v=1C8%5Fn9nnx5Y&ref=nnutshell.com).