Mindset & The Google Story:
- Nothing to lose, so attack - From lower-middle-class India where a Google job was the family's idea of making it; "attack, attack, attack."
- Motivated by impact, not wealth - $20B, 45M users, a billion searches a month - none of it moves him; impact and process do.
- Changed Google.com more than any PM - He forced Google to ship an answer engine it never wanted to build.
- Answer engine was always lead-gen - Search was the doorway, never the destination; it funded the agents, deep research, and Computer.
The Frontier Is Where The Money Is:
- Frontier = doing work, not answering - The money moved from answering questions to agents that go and do real work for you.
- "Model alone is no longer the product" - Quotes OpenAI's Greg Brockman; resell raw tokens and you get commoditized to zero.
- OpenAI "isn't ready for an IPO" - Dominant in consumer search, but that's commoditized lead-gen with no money in it.
- Everyone races to agentic work - OpenAI on Codex, Perplexity on Computer, Anthropic on Claude Code - "that's where the money is."
- No one is safe, including Anthropic - Six months without a new capability and "it's bad for them"; no one can relax.
Orchestration & Token Value Per Watt:
- The harness is the real product - Tools, connectors, and sub-agents turn a raw model into work that's actually worth paying for.
- "Token value per watt per user" - He calls it the single most important metric in AI; max it to win.
- Orchestrate across rival models - GPT-5 and Claude both run inside Perplexity Computer - something OpenAI and Anthropic can't claim.
- A few power users drive the economy - Forget a billion users; a small base of heavy spenders sustains the whole business.
- Every layer's gain helps them - Jensen's chips, Dario's models, Apple's devices all compound; revenue tripled this year.
Why Advertising Fails In Chat:
- Ads poison an answer engine - Ask for the best protein shake, get a "btw, try these" nudge, and trust dies.
- Ads chase browsing, not asking - Booking.com pours billions a year into Google and fashion floods Instagram, where people explore, not ask.
- Objective buys go to agents - Spec-driven buys (a mic) become an agent's job; taste-driven ones (a table) keep you browsing ads.
- Agents already outrank humans online - Cloudflare reported agent traffic has overtaken human traffic on its network.
- The real money isn't ads anyway - Agentic tools will out-earn Google's and Meta's entire ad business - without 100M users.
Power, Memory & The Infra Bottleneck:
- Power is the bottleneck, not chips - Land, permits, turbines, and cooling dwarf the lead time of just buying GPUs.
- "Funny, stupid, moronic" - His verdict on AI-bubble talk; the real limiter is how fast you can physically build.
- ~40 of 100 data centers stall - Blocked by public resistance rooted in job-loss fear, inequality, and rising grid and RAM prices.
- Micron could pass Meta in 6-12 months - Whatever's the bottleneck commands the price; memory is up 5x, Micron is already ~$1T.
- Agents quietly revive CPUs - Agent loops run on CPUs more than humans do, lifting AMD and Intel as demand surges.
- "AWS, not Amazon servers" - Renting bare GPUs is a commodity; the money is in the software layer you build on top.
Export Controls & China:
- 20-30% chance of a DeepSeek shock - A leaner, vertically integrated model could leave today's massive buildouts overbuilt.
- Export controls create the 12-month gap - The only reason open source trails the frontier; Anthropic lobbied hard for them.
- But they may backfire - Forcing China onto the Huawei stack breeds memory-efficient, vertically integrated designs America isn't building.
- China is underestimated on physical AI - Fabs, robots, chips, and energy packaged into devices may let it out-build America.
Aravind Srinivas of Perplexity just gave the clearest articulation of where AI is actually headed, validating what I'd already begun to believe from using these tools every day.
On 20VC with Harry Stebbings, he doesn't do the soft "AI will change everything" routine. He takes hard, specific positions and backs every one: power, not chips, is the real bottleneck; Micron will be worth more than Meta; advertising will never work inside a chat product. A few of these are aggressive enough that some won't age well, but that's exactly why it's worth your time.
He says value no longer comes from the raw model. It's the orchestration around it, the tools and connectors and sub-agents that turn a model into actual work. The frontier has moved from answering your questions to going and doing the work for you.
A couple of his other calls landed just as hard. The whole token economy, he argues, runs on a small handful of power users rather than a billion casual ones. And ads, in his view, will always poison an answer engine, because the second a product slips in a "btw, you should try these," the trust you came for is gone.
His advice for building: stop obsessing over moats in year one, move fast, and build at the orchestration layer. Because every time the chips, models, or devices get better, your product improves for free. Perplexity was voted "most likely to fail" at an AI summit a few months ago. It has since tripled its revenue. His motto, in his own words, is just: attack, attack, attack.