Ads Are Machine Learning 1.0: AppLovin Used It to Earn an 84% Margin
Advertising isn't AI's past tense — it's the earliest and most profitable deployment of deep learning. AppLovin used recommendation models to build an 84% EBITDA margin in the $50 billion mobile game ad market, and became its own best investor with $6 billion of buybacks.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Mobile game advertising is a $50 billion market
Adam says outsiders underestimate the scale of mobile gaming: more than 1 billion adults play casual mobile games every day, and these people are household decision-makers. Two years ago AppLovin disclosed $11 billion in annual ad spend on its own platform, and it has since grown at roughly 60% year over year, which by rough math puts it at about $20 billion today. But that isn't the whole market — there are many other ad companies monetizing in this ecosystem, and if you more than double the number, the entire mobile gaming ecosystem spends roughly $50 billion a year on advertising. He notes that not long ago social was also just a $50 billion opportunity.
— Adam ForoughiAdvertising is machine learning's earliest and most profitable implementation
Adam calls advertising ML 1.0: it was the first deployment of all these technologies that now drive AI. Large language models create more economic value today than advertising, but advertising is an extremely profitable implementation of deep learning models. Recommendation systems and large language models have different architectures, but their trajectories are highly similar — research in large language models can transfer to recommendation systems, and vice versa; many large language model researchers spent the early part of their careers looking at ad systems. The beauty of the ad business is that you build a model to predict future outcomes, and the value of that prediction can be immediately converted into money.
— Adam ForoughiSearch ads don't create economic expansion; discovery ads do
Adam divides advertising into two categories. One is bottom-of-funnel advertising: the consumer already knows what they want to buy and is just doing research to complete the transaction — that's Google Search's business, and large language models will compete with Google Search almost entirely in that area. The other is his field: not knowing user intent, creating a demand that didn't previously exist, so that when users see it they think ‘this looks really cool, I want to buy it right now’. He stresses that the transaction in search or in a large language model would have happened anyway, so there isn't much economic expansion there; discovery advertising creates economic expansion, which is also what makes Meta's ad business strong and what AppLovin wants to do.
— Adam ForoughiYour phone isn't listening, but your behavior is tracked
Faced with the question of ‘I talk about something at lunch and then an ad for it appears’, Adam says it's more likely that you previously did something trackable — a search, browsing a website, searching for a product — without realizing it, and then you said something related and started seeing related ads. He states clearly that AppLovin does not track location at all, arguing that tracking precise location to render ads is a very heavy concept, and that the data transfer required to parse microphone content in real time and translate it into ads is unrealistic. But he acknowledges that if you're on a social network, relationship data between you and others can of course drive the ad experience — for example, a friend searched for something and you might see a related ad, which he thinks is fine.
— Adam ForoughiFrom $28 billion to $3.8 billion, then it started buying itself
AppLovin went public in April 2021 with $600 million in EBITDA and a market cap of about $28 billion, peaking at $40 billion. In 2022 the stock fell almost every day, down to a market cap of about $3.8 billion — in a year when the company did $1 billion in EBITDA. Adam's read: market cap is determined by the quality of investors, and at the time a large number of companies went public at the same time, blue-chip investors didn't do the research to figure out what this company with the silly name was, and the result was no demand and lots of supply, with the valuation falling from 50x EBITDA to less than 4x. His response was to stop talking to investors and instead aggressively buy back stock with the large amounts of cash the company generated, and since then it has bought about $6 billion of its own stock, retiring 20% to 25% of shares outstanding.
— Adam ForoughiWhen the stock fell 92%, family called to ask if you were suicidal
Adam says that period was hard, that family and friends would call to ask if he was suicidal, and his answer was that the stock was still at $10 and had still gone up a lot. But the truly hard part was realizing that as CEO, his team was getting the same calls, and they didn't have his composure or his equity. So they rallied the team with a ‘we versus the world’ mindset, and extended a performance stock plan usually reserved for the CEO to key people across the company, telling them: it's hard right now, the house you thought you had is gone, but if you hold on and recover, the upside will be large. Later investors started coming back, the company switched from regression models to deep learning models, the ad algorithm got better, advertiser returns got better, and the business started growing fast.
— Adam ForoughiA week in New York took the market cap from $28 billion to $55 billion
When the new model launched in April 2023 the company still wasn't communicating with investors, so the outside world didn't know. By around September 2023 the stock was at $80 with good results, and Adam judged the market cap high enough that it couldn't buy back that aggressively anymore, so he went to New York and started talking to investors. In that one week the stock went from $80 to $150, and the market cap went from $28 billion to $55 billion — just because he went out and said ‘our company is still here, we survived’. He describes sitting in meetings and being able to read that the person across the table was texting a friend to buy. Over the next two and a half years the stock went from $9 to $750, and the market cap reached $250 billion. His conclusion is that public market investors and private investors aren't that different, they all chase trends, just often later than you'd like.
— Adam ForoughiAfter privacy rules tightened, users complained ads got worse
Adam says in these areas you want regulatory clarity, and once the rules are clear, technology can respond. Five years ago you could precisely target users on iOS; today if users say no precise targeting, you can only bundle them into groups and serve worse ads. The interesting result of this change is that after Apple's changes the company received a flood of user complaints saying ‘show me more relevant ads, you're showing me a bunch of garbage now’. So he offers a duality: on one hand you need privacy regulation so tech companies know what to do, on the other hand consumers do want relevant ads — when you watch a 30-second ad in a game to get an extra life, you're getting something with monetary value, and you don't want to spend those 30 seconds watching garbage.
— Adam ForoughiIt bought game studios for training data, then sold them off
Adam explains that the company's motive for buying game studios was data: building the first deep learning model required data, and game studios are usually unwilling to share data with third-party companies, so they bought their own studios, fed the training data into the first model, built a model that was very successful in the market and started growing fast. Once it worked, third parties started coming to them proactively, and they spun off and sold those game businesses. This clarifies outside speculation about whether AppLovin was going to become a game publisher — what it bought was data, not long-term game assets.
— Adam ForoughiThe typical shopper isn't the person on Twitter chasing new tech
On agents and agentic commerce, Adam thinks one part of the world will use agents to optimize fixed shopping behaviors, like handing a supplement subscription to an agent to optimize monthly and deliver on time. But discovery platforms don't work that way, and the typical shopper isn't the kind of person who uses agents deeply and hangs out on Twitter chasing the latest tech. He says his audience is ‘the New York Times audience’, and there are still lots of people using Yahoo products today. The typical shopper wants to find products, wants to browse, wants to go through the transaction flow, wants to compare prices, wants to track shipping; even if you tell them afterward that an agent could have saved them 20%, on a $50 transaction it doesn't matter, because the dopamine of going through the process is what they enjoy. He thinks the industry pays too much attention to the Twitter bubble and forgets that ordinary shoppers aren't like that.
— Adam ForoughiEvery morning you wake up feeling you're about to be killed, so you run faster
Faced with the question of how a small company wins when Meta and Alphabet have the best engineers and have done advertising for decades, Adam says one thing that got them here is that they never think they've won, and every morning they wake up feeling they're about to be killed, so they have to work harder. The company stays lean and has gathered a lot of domain experts truly focused on the mobile game experience and on converting it into transaction behavior. He thinks that as long as you're very focused and stay lean, you can run faster than the giants, and that gives you the ability to challenge them.
— Adam ForoughiAn 84% EBITDA margin that no one can compete away
Adam says AppLovin's EBITDA margin is the best in the market at 84%, so he doesn't think there's much leakage. He explains the business model: advertisers come in with a transaction model, say selling lipstick, and the company gives them an arbitrage — they buy consumers from AppLovin, and once the consumer completes the transaction it immediately covers the acquisition cost. The consumer spends $20 on lipstick, the advertiser pays AppLovin less than $20 minus cost of goods sold, the advertiser is happy and spends more, and this performance model is very scalable. The company isn't the whole chain, isn't the advertiser itself, but empowers advertisers to reach consumers, while operating extremely lean and being extremely algorithm- and automation-driven, so there aren't many leakage points. Faced with the argument that an 85% EBIT margin means you might be over-earning and competitors would be willing to take 60% or 50% to grab share, Adam's response is: these technologies are very complex, and if you can innovate and have differentiated data, you can build an advantage — just as Anthropic shouldn't have run away with the language model field, but once a model reaches scale and is adopted by a large community, it becomes a moat others can't cross.
— Adam ForoughiA Chinese team is an advantage; I'm often the dumbest person in the room
On Chinese teams, Adam says Chinese people are very humble, very hardworking, very sharp, and whether in China, the US or anywhere else in the world, working with them is working with the smartest people in the world. One of his goals when starting a company was to work with excellent people and figure things out together. He says when he sits in a room with some of the people on his team, he knows he's very likely the dumbest person in that room, and that makes him excited to come to work.
— Adam ForoughiIn their own words · checked verbatim
advertising is like ML 1.0 but really was the first implementation of all these technologies that now are are driving AI today
Adam Foroughi3:02
the transaction via search or LM was going to happen anyways. If the LM didn't exist and Google ads had never come to existence, but Google search existed, that transaction, the closed loop would have happened. So there's not actually a whole lot of economic expansion that happens from that.
Adam Foroughi6:03
I would get phone calls from family members, friends, are you suicidal? And I'm like, look, we we got stuck at a penny. The stock's still like 10 bucks. It's still up a lot.
Adam Foroughi12:07
in that week, the stock went from 80 to 150. And I think it was like 28 billion to 55 billion from you being in New York. From me just going out and saying, "Hey, our company still exists. We survived this."
Adam Foroughi14:08
we'll get a lot of complaints after that change that Apple made from users that say, "Serve me more relevant ads. You're showing me a bunch of spam."
Adam Foroughi16:09
We never think we won. We think every day we wake up and we're probably going to get screwed right now and we better work hard.
Adam Foroughi20:11
the power of a model that then reaches a point of scale and gets adopted by a large scale community becomes something that is a a moat that is hard for other people to overcome
Adam Foroughi21:12
when I sit in a room with some of the people on my team, I know I'm probably the dumbest person in that room. And that gets me excited to show up.
Adam Foroughi23:12
Figures
| AppLovin annual ad spend on its own platform (disclosed two years ago) | $11 billion | 1:02 |
| AppLovin platform ad spend year-over-year growth | about 60% | 1:02 |
| Mobile gaming ecosystem annual ad spend (estimate) | about $50 billion | 2:02 |
| AppLovin market cap at IPO | about $28 billion | 10:07 |
| AppLovin lowest market cap in 2022 | about $3.8 billion | 10:07 |
| AppLovin 2022 EBIDTA | $1 billion | 10:07 |
| AppLovin stock buyback amount | about $6 billion | 12:07 |
| AppLovin shares retired as a percentage of shares outstanding | 20% to 25% | 12:07 |
| AppLovin stock price move over two and a half years | from $9 to $750 | 15:09 |
| AppLovin EBIDTA margin | 84% | 21:12 |
Glossary
- EBIDTA
- A measure of a company's core profitability; Adam uses this phrasing multiple times in the transcript.
- ML 1.0
- Adam's term for advertising as the earliest and most profitable deployment of deep learning technology.
- agentic commerce
- A model in which AI agents make shopping decisions and complete transactions on a person's behalf.
- discovery
- A form of advertising that creates new demand through recommendations when the user has no clear purchase intent.
How to listen
Founders, investors and engineers watching ad tech, recommendation systems and AI commercialization; anyone who wants to know how a small team builds a moat in a field full of giants.
If you're not interested in ad tech details, the part from 8:07 to 9:07 about whether phones eavesdrop on you can be skipped.