Adam Foroughi, Applovin CEO: Surviving a 92% Drawdown, Ads as ML 1.0 & the $50B Game Ad Market

Summary

• AppLovin sits on a mobile-gaming ad market Foroughi sizes at roughly $50B a year — “it was not very long ago that social was a $50 billion opportunity.” Over a billion adults play casual mobile games daily; AppLovin disclosed $11B of annual ad spend on its own platform nearly two years ago, has grown roughly 60% year over year since (“gross that up to a nice round number today, you get $20 billion”), and the thesis now is extending game-to-game ad intent into e-commerce shopper behavior.
• The host framed the internal-culture episode around a 92% stock drawdown; Foroughi’s capital-allocation response was a masterclass: IPO in April 2021 at about a $28B market cap on $600M of EBITDA, reaching $40B before collapsing to $3.8B in 2022 while producing $1B of EBITDA — sub-4x. He stopped focusing on investors and bought roughly $6B of stock, retiring 20–25% of shares outstanding; at peak, those repurchases were worth over $50B, and shares later ran from $9 to $750 in 2.5 years, touching a $250B market cap.
• The re-rating catalyst was ML 1.0 → ML 2.0: swapping a regression model for a deep-learning ad model launched in April 2023. Because the model improved advertiser performance and let customers scale, the company grew quickly; investors took notice when Foroughi resumed meetings in New York around September 2023. “In that week, the stock went from $80 to $150” — roughly $28B to $55B — after he told investors the company had survived.
• Foroughi’s ad-market split is the key framework for the OpenAI-ads debate: LLM advertising “is almost going to exclusively compete with the Google Search business,” while discovery advertising creates net economic expansion. Bottom-of-funnel transactions would have happened anyway; showing a consumer something “they had no idea existed” is what powers Meta’s business and what AppLovin aspires to.
• On competitors compressing his stated 84% EBITDA margins, his answer is that complexity, differentiated data and scale create a moat: “if you can innovate and you have differentiated data, you can build an advantage.” He analogizes to Anthropic’s position in large language models despite the theory that it should not be running away with that market.
• On surveillance, he says he does not think advertising companies can track location, says AppLovin does not track it, and calls parsing microphone audio for ads unrealistic. He attributes many seemingly uncanny ads to trackable actions people forget. On agents, he sees them fitting repeat behaviors such as supplement subscriptions, but says the typical shopper still wants to browse, compare and enjoy the transaction; “we really over-index on the Twitterverse.”

Deep dive

1. A $50B ad market hiding inside 100,000 mobile games

• Foroughi’s framing of why few people know the company: no VC funding at the early stage meant “we just had to build quietly,” and “the goofy name” did not help. What it actually is: an ad platform monetizing a universe where over a billion people play casual mobile games daily — “these are all adults, heads of households.”
• The sizing chain: $11B/year of on-platform ad spend disclosed nearly two years ago, roughly 60% year-over-year growth since then → about $20B today; more than doubling that across other monetizers → roughly $50B annually. The reward mechanic — users watch ads to receive rewards — creates the possibility of intent, and the growth opportunity is extending that game-to-game recommendation behavior into shopping.

2. Advertising was ML 1.0 — and discovery ads survive the LLM era

• Ads as an early deep-learning implementation: “advertising is like ML 1.0.” Recommendation systems and large language models are related and, in many ways, follow similar trajectories; research can port between them, while advertising can translate the value of a prediction immediately.
• The load-bearing distinction, prompted by the host’s question about OpenAI’s ads: bottom-of-funnel search ads close a transaction that “was going to happen anyway,” so LLM ads “almost exclusively compete with the Google Search business.” Discovery — Meta’s model and AppLovin’s aspiration — creates a recommendation and transaction that did not previously exist, plus “a really fun moment for the consumer” as they wait for the package.
• On ad quality’s arc since 2005 — “complete garbage. It was all spam” — Facebook paired available data with better technology so ads became more like content. In AppLovin’s domain, Foroughi says “people love the ads”: users engage with playable minigame previews appearing inside other games.

3. Surviving the collapse: buybacks as offense

• The host framed the culture challenge as a 92% drawdown. Foroughi’s diagnosis of the collapse was that market prices depend on the quality of the investors owning a stock. A COVID-era IPO wave meant blue-chip investors did not research “this goofy-named company,” leaving little demand against substantial supply; the stock fell “literally every day” in 2022, from a $40B peak to $3.8B while EBITDA reached $1B.
• The response was to stop focusing on investor relations and deploy cash into roughly $6B of buybacks, retiring 20–25% of shares. “You do have an opportunity on the other side of it,” he said.
• The human cost was real: “I would get phone calls from family members and friends — are you suicidal?” His team received similar calls without his ownership cushion. The response was “an us-against-the-world mentality,” plus a performance stock plan extended across key employees: “we understand you thought you had a house and now you don’t.”
• The recovery mechanics: the April 2023 deep-learning model improved advertising performance; because the business is performance-based, better advertiser returns enabled customers to scale and drove rapid company growth. Foroughi resumed investor meetings around September 2023, when the stock moved from $80 to $150 in a week. His takeaway from the full $9-to-$750 cycle is that public and private investors often follow trends later than ideal; the best investors identify them early.

4. Privacy, “creepy” ads, and why agents won’t replace discovery

• The host’s theory — geolocation grouping friends at lunch, then targeting them after one person searches — gets a qualified rebuttal: “I don’t think advertising companies can track location. We don’t track location at all.” Parsing microphone audio and turning it into ads is “not realistic.” Foroughi’s mundane explanation is that users performed trackable actions such as searches or browsing and forgot them. He allows that social-network relationships can influence ad experiences: if one person searches, friends may see something relevant, and “there’s nothing wrong with that.”
• On Apple’s tightening of privacy rules, his view is that regulations should be clear so technology companies can adapt. After users are grouped more broadly, some complain that the resulting ads are spam and ask for more relevant ones. He also argues that relevant ads create economic value: the digital ad economy contributes to GDP, and better technology can accelerate GDP growth.
• On agentic commerce, agents fit consistent repeat behavior such as his supplement subscription. But the typical shopper “wants to window-shop,” compare, track the order and enjoy the transaction; even saving 20% on a $50 purchase may not outweigh that experience. “We really over-index on the Twitterverse.”

5. Competing with Meta and Google through an 84% margin, lean machine

• Asked how a small company competed with a decade or two of Meta and Google advertising expertise, Foroughi says, “We never think we won.” The company stays lean, concentrates subject-matter experts on mobile gaming and transactional behavior, and tries to move faster than larger competitors. The game studios were a data play: AppLovin bought them to seed training data for its first deep-learning model, then divested them once third parties began supplying data.
• The host’s leakage question mostly bounces off AppLovin’s stated 84% EBITDA margin. In Foroughi’s example, a lipstick advertiser pays AppLovin less than the $20 purchase price minus cost of goods sold; the advertiser covers the customer-acquisition cost immediately, then scales under a performance-based model. The admitted gap is that AppLovin is “not the full chain” and is not itself the advertiser.
• Why competitors have not competed the margins away: these technologies are complex, and differentiated data plus a model that reaches scale and adoption can become a moat that is difficult to overcome. Foroughi compares that dynamic to Anthropic’s position in the large-language-model market.
• On the engineering offices in Palo Alto, Beijing and Singapore, he describes Chinese colleagues as humble, hardworking and sharp, and says that when he sits with some team members, “I know I’m probably the dumbest person in that room. And that gets me excited to show up.”