Summary

A 24-minute All-In Podcast interview with AppLovin CEO Adam Foroughi, introduced as 'probably the best founder no one's ever heard of'. It is a founder interview rather than a news segment, and it argues one idea: that advertising was machine learning 1.0, that the recommender systems behind it and today's large language models are two branches of the same research, and that a model which reaches scale becomes a moat in its own right. Around that sits a first-hand account of a 92% share-price collapse and the recovery from it, a flat denial that ad platforms listen through phone microphones, and a sceptical read on agentic commerce. Every figure is the CEO's own, about his own company, in a friendly setting with no adversarial questioning — treat them as reported, not audited.

• The market Foroughi describes, in his own figures: over a billion people a day play mobile casual games, whom he characterises as adults and heads of households rather than teenagers. AppLovin disclosed roughly $11 billion a year of ad spend on its own platform in January of the prior year, and he claims roughly 60% year-on-year growth since, putting it near $20 billion today. He estimates the whole mobile-gaming ad ecosystem at about $50 billion a year, and notes that social advertising was a $50 billion opportunity not long ago.
• The thesis, stated plainly: 'advertising is ML 1.0'. He argues ad systems were the first commercial implementation of the techniques now driving AI; that recommendation systems are structured differently from large language models but follow the same trajectory, with research porting in both directions and many LLM researchers having started in ad systems; and that advertising's structural advantage is feedback — when you predict an outcome you learn immediately what the prediction was worth. He concedes the economic value of an LLM to society is far greater than advertising's, but says advertising is deep learning's most profitable implementation.
• The recovery was a model change, and the dates matter. He describes moving from 'ML 1.0 to ML 2.0' — a regression model replaced by a deep learning model, launched April 2023. Because everything on the platform is performance-priced, a better algorithm means better advertiser returns, which means advertisers scale spend, which he says is what restarted growth.
• The drawdown, with his figures: IPO April 2021 at about $28 billion on roughly $600 million of EBITDA, peaking near $40 billion; by 2022 the stock fell every day to a low around $3.8 billion market cap, in a year the company claims it did $1 billion of EBITDA — a multiple falling from roughly 50x to under 4x. His explanation is investor mix rather than business performance: a COVID-era IPO glut meant blue-chip investors never did the work on 'this goofy named company', while private-equity and insider holders were sellers, leaving heavy supply and no demand.
• What he did about it, which is the part most likely to be quoted elsewhere: he stopped talking to investors entirely on the grounds that they were not buying, and redirected the cash into buying the company's own stock — 'let's become our own best investor'. He claims roughly $6 billion of buybacks retiring 20-25% of shares outstanding, and that at peak that holding was worth over $50 billion.
• On running a company through a 92% fall: he says family members phoned to ask whether he was suicidal, and his concern was that his staff were getting the same calls without his ownership stake or standing. The response was an 'us against the world' framing plus extending a performance stock plan — normally a CEO instrument — across key people, on the explicit argument that they had lost a house on paper and would make it back on the upside.
• The re-rating, and how fast it moved: with the stock near $80 in about September 2023 he resumed investor meetings in New York, and claims it went from $80 to $150 in that week, roughly $28 billion to $55 billion in market cap — on, as he puts it, going out and saying 'our company still exists, we survived this'. Over two and a half years he claims $9 to $750 a share, and roughly $3.8 billion to $250 billion in market cap. His reading: public and private investors behave alike, following trends late, and the good ones are simply earlier.
• On margins and the moat, including an explicit Anthropic analogy. He claims 84-85% EBITDA margins, which he says are the highest in the market, and is asked why competitors have not undercut them at 50-60%. His answer is that the technology is complex and differentiated data compounds, and he reaches for AI to make the point: 'by that token Anthropic shouldn't be running away with the large language model space, but the power of a model that then reaches a point of scale and gets adopted by a large scale community becomes a moat that is hard for other people to overcome.' That is an operator's claim about why model businesses defend themselves, offered as analogy rather than analysis.
• On what LLM advertising actually takes, which is the most useful distinction in the interview. He splits advertising into bottom-of-funnel — the user already knows what they want and is researching the purchase, which is Google's search business and what an OpenAI ad product will compete with — and discovery, showing someone something they did not know existed. His argument is that bottom-of-funnel creates little economic expansion, because the transaction would have happened anyway; discovery does, because it creates demand that did not exist. That is Meta's business and the one he says AppLovin is moving into.
• Asked directly whether phones listen, he says no, and gives reasons: AppLovin does not track location at all, precise location tracking to render an ad is 'a really heavy concept', and the data volume involved in transmitting and parsing microphone audio is 'not realistic'. His explanation for the experience is trackable actions you have forgotten making — a search, a browse, a product lookup — plus social-graph effects on platforms that do hold relationships. Worth noting he is answering for his own company, not for the industry.
• On privacy regulation he wants rules clear rather than absent: once they are clear, technology adapts. His reported consequence of Apple's targeting restrictions is that users complained about being shown worse, spammier ads, and his argument is that consumers genuinely want relevant ads — particularly when watching 30 seconds of advertising to earn an in-game reward, where the ad is something they have paid attention for. He also claims ad relevance now contributes materially to GDP, and that better technology means faster growth. That last claim is asserted, not supported.
• On agentic commerce he is a sceptic, and has a commercial interest in being one. He accepts agents will handle consistent repeat purchases — his example is putting a supplement subscription on autopilot — but argues discovery is not that. The typical shopper wants to window shop, compare, transact and track, and 'the dopamine hit from going through it is what they enjoy'; told afterwards that an agent could have saved 20% on a $50 purchase, he does not think they would care. His framing: 'we really over-index on the Twitterverse and forget that the average shopper is not that', describing his audience as 'the New York Times audience' with plenty of people still using Yahoo properties daily.
• Two operational details worth keeping. The game studios AppLovin bought were a data play, not a strategy: studios would not share data with a third party, so it bought its own to seed the training data for the first deep learning model, then divested once third parties came in. And the engineering is split across Palo Alto, Beijing and Singapore, with Foroughi describing the Beijing team in strongly positive terms and saying he is often the least capable person in the room.
• Interviewer claims, separated from his: that AppLovin is the most valuable of the thousand-plus COVID-era IPOs, and that it will 'print something like $6 billion in cash this year'. Both are asserted by the hosts, not by Foroughi, and neither is substantiated on the show.

Why it matters

This is the archive's clearest first-hand account of deep learning making money at very large scale outside a frontier lab, and it reframes something the Handbook has been treating as new. If advertising was ML 1.0, then the pattern of a model improving, capturing more value, and funding its own next iteration has a fifteen-year track record in a commercial setting — and Foroughi's claim that recommender and LLM research port in both directions is a concrete reason to expect talent and technique to keep crossing. The more consequential claim for the Handbook is his answer on moats: that a model reaching scale and broad adoption becomes hard to dislodge regardless of competitors' ability to build one, which he applies to Anthropic by name. That is the exact question the Handbook currently holds open, and here it is answered in the affirmative by an operator whose own 84% margins have survived a decade of competitors who could have undercut them — which is evidence of a kind, though from an adjacent market. His discovery-versus-bottom-of-funnel split is also the most useful frame yet for judging what an LLM ad product can and cannot take from whom. Set against all of that: he is a seller of discovery advertising telling us discovery advertising survives agents, his figures are self-reported about his own company on a friendly show, and the hosts never push back on a number. The entry is valuable for its frames, and its figures should carry the source with them.

JtomF4bGxHs-transcript.txt