Aravind Srinivas and Sebastian Mallaby at Sohn 2025

Summary

• Perplexity’s valuation went from what Mallaby thinks was $9B in December to a not-yet-closed $14B five months later — Mallaby’s math: “another unicorn effectively every month.” Aravind’s defense of the froth: this isn’t crypto’s paper money — people depend on these apps daily, and “anytime product goes down we only hear thousands of complaints.”
• The headline call: “it’s the first time in like two decades that Google is extremely vulnerable” — possibly since founding. The mechanism is margin math, not technology: search advertising runs 80-90% profitable and pays for everything else, while “AI may never be as high margins… ever,” so defending means cannibalizing the world’s best business.
• The browser, not search, is the real battle: “whoever owns [the omnibox] is going to get everyone’s traffic.” Just as 20 search engines collapsed into the one integrated in the browser, the AI that blends navigation, Q&A, tasks and agents into one box wins — but “it’s going to take at least half a decade.”
• Brand risk is brutally asymmetric: a 1-in-10 error rate makes a startup look impressive, but at Google’s 10 billion daily queries it’s “a billion mistakes a day” that destroys a two-decade brand. His verdict on Gemini: “It kind of works. It’s not like Google where it just really works.”
• On the A+ technology / C-minus business model problem: paying for an answer feels weird when open source is free, but a true assistant that has absorbed your context and workflows is “worth way more than $20 a month” — and margins expand structurally, because operating costs fall as models commoditize while value rises as the assistant gets better and absorbs context.
• The moat against a crowded field (4-5 US contenders, China, likely Mistral, the Middle East): “obsession and attention to detail… are still in scarcity.” If money and talent were sufficient, “likely OpenAI would be dead” — Microsoft has no innovator’s dilemma. Shipping past the threshold where users bear with mistakes “is still art.”

Deep dive

1. $9B to $14B in five months — “AI itself is pretty wild”

• Mallaby opens with the number: Mallaby thinks Perplexity was valued at $9B in December; five months later a still-unclosed round in the newspapers says $14B. Aravind grants “healthy amount of skepticism on whether it’ll last,” but insists the difference from crypto is usage: “it’s not creating paper money” — people use these apps daily for work and life, and “anytime product goes down we only hear thousands of complaints.”
• His operating reality at the eye of the storm: model capabilities improve so fast you must “adapt the product to make use of those capabilities rather than getting deprecated by them” — splitting time between the next 6 months or 3 months and today’s execution simultaneously.

2. David vs. Goliath: from simple RAG to a likely Manus-style agent

• Roughly 300 people against Google. The playbook: “building a better product relentlessly,” caring about a good answer over making money out of ads, then going beyond answers to actions — catching Goliath in an innovator’s dilemma until, hopefully, “5 years from now you don’t ask me this question.”
• The product arc as told: launch was simple RAG — top-K links, snippets, summarize with footnotes, “nothing that other people couldn’t have thought of.” The first real breakthrough was multi-step reasoning: break ambiguous questions into parts, search each, recombine (Pro Searches); then custom UIs and classifiers per vertical; next a “Project” feature “more like a likely Manus-style agent” doing 30 minutes of work, and a browser carrying your full personal context.

3. The browser is the real battle — whoever owns the omnibox wins

• Aravind’s correction to the two-front-war framing: “the browser is the real battle… whoever owns that is going to get everyone’s traffic.” Typing the same prompt across AI apps A, B, C goes away, just as 20 search engines collapsed into the one that put search in the browser box. The winner blends navigation, questions, tasks and agents into one search box — “it’s going to take at least half a decade to make it truly work.”

4. Why Google walks into a trap it saw coming

• Mallaby’s setup is the puzzle: Google housed two Nobel laureates (Hinton, likely Hassabis), knew the Xerox Park story cold — how does it still get trapped? Aravind’s answer is margin math: 30-35% operating margins, with search advertising at 80-90% profitability paying for everything else. YouTube, cloud, likely Waymo are all lower-margin — and “AI may never be as high margins… ever,” so pursuing it means trading down at the scale of the world’s population.
• The brand-risk asymmetry: one mistake in ten makes a startup look good; at 10 billion daily queries it’s “a billion mistakes a day” — glue on your pizza, amplified on social media, until “googling it means you basically got the answer” collapses into “the same brand as every other upcoming AI startup.” On Gemini: “It kind of works. It’s not like Google where it just really works.”
• The call, hedged as he made it: “I’m not saying everything is guaranteed to succeed for us… but it’s the first time in like two decades that Google is extremely vulnerable.” On incumbent executives sitting on Perplexity’s cap table: “it’s a good hedge for the individuals.”

5. A+ technology, C-minus business model — and where the moat is

• Mallaby’s pushback — worth keeping: OpenAI burns enormous capital while DeepSeek (likely) and Meta ship free open-source models, so where’s the pricing power? Aravind concedes only three business models exist today — consumer subscriptions, enterprise, API; ads “still pretty early days” — but rejects the framing: paying for an answer feels weird; paying for a true assistant that absorbed your preferences, context and workflows is “kind of like hiring a person” and “worth way more than $20 a month.” Costs fall as models commoditize; value only rises.
• Against a crowded field — four or five US contenders, “a whole bunch in China,” likely Mistral, the Middle East — the moat claim: “obsession and attention to detail and just caring about the problem are still in scarcity.” Otherwise “likely OpenAI would be dead” — Microsoft has no innovator’s dilemma. Passing the threshold where users bear with stochastic models’ mistakes “is still art”; if it weren’t, “the one with the most money will always win. But that’s not the case right now.”