Inside AppLovin’s $100B Ad Engine

Summary

• AppLovin’s arc is the episode’s spine: down 92% in the first 18 months as a public company, rebuilt around the Axon 2 model, and now a $100B company with a stated path to a trillion. Adam Foroughi’s math: EBITDA run rate is “over a $7 billion” with ~75% converting to cash — he said the comparable figure three years earlier was probably 1/20th of today’s — and a trillion-dollar valuation requires “$30 billion-plus of cash flow a year,” which gaming UA alone can’t support, hence e-commerce and eventually adjacent categories.
• The technical unlock was replacing a tree-based Axon 1 — “hundreds of thousands of if/else branches” — with learnable semantic embeddings feeding a deep neural network. CTO Giovanni Ge says the new model both extrapolates to unseen user-item pairs and runs cheaper, because GEMM operations are what GPUs are optimized for while trees aren’t: “Once we’re able to make prediction more accurate, advertisers see better returns and our business grow.”
• The org design is a major part of the story alongside the model: ~100 engineers, roughly unchanged in three years, and very few product managers in Ge’s organization, while engineering headcount stayed flat as the business scaled. Ge’s framing — “I don’t want our engineers to sit next to AI. I want our engineers to sit on top of AI” — accompanies a new-generation model the team was able to tackle with AI assistance.
• E-commerce, entered roughly 18 months ago, answers the standing bear case that AppLovin only has gaming data. Foroughi’s rebuttal: it’s a billion people, not a billion gamers — casual-game players skew slightly female, 30-50, and include many heads of households — and pixeling advertiser websites follows the approach Facebook used to build a broader data flywheel; the first e-commerce data engine was literally designed on a breakfast napkin at a Vegas conference, with decent ROAS even on a “very premature” model.
• Chatbot/LLM advertising may not be the best fit for AppLovin’s model: if 99% of AI usage is search, chatbot ads will probably look and feel like bottom-of-funnel search ads, while AppLovin’s engine is top-of-funnel discovery. The extension surfaces under R&D on a “three, five, 10-year” horizon are connected TV and open-web video — fragmented environments unlike the Apple/Google mobile duopoly.
• The drawdown playbook is directly relevant to today’s beaten-down SaaS names — and Foroughi doesn’t think most can copy it. AppLovin kept stock comp in fixed-dollar terms, did no investor relations “for well over a year at the bottom” (“nobody buys something that’s dirt cheap… they wanna see a vision”), and deployed every dollar it made and more into buybacks to “become our best investor”; enterprise SaaS without algorithmic growth or cash flow can face “a downward spiral” and may be taken private by private equity.
• Both executives’ hot take converges on taste as the scarce input in the AI era. Ge: “I would attribute the success of Axon largely to what we decided not to do, not actually to what we did” — AI makes building easy, so companies can drown themselves in “bad-taste ideas”; Foroughi adds that legacy organizations may “almost have to replace nearly everyone” to become AI-native, and “there’s no clear answer” for most.
• Ad creative remains a source of manual alpha in an otherwise automated system. There’s no formula — “if you create 30 ads a week, probably one of those might be interesting” — social’s three-second ADHD playbook does not simply transfer to 60-second playable ads, and gen-AI still can’t reliably produce a brand-safe 30-60-second video, so advertisers who invest early in the platform’s format “get alpha.”

Deep dive

1. From 92% drawdown to $100B — and Gio’s contrarian entry

• Foroughi’s opening frame is the whole story compressed: “When we first started, it was, ‘I wanna become a billion-dollar company’… When we went public, we wanted to become a $100 billion company. We dropped 92% in the first 18 months of being public… but post Axon 2, we recovered, and now we’re a $100 billion company.”
• Ge joined in November 2022 and was due to start a couple of days after an earnings call sent the stock down 30% (AppLovin was around $5.5B when he was interviewed). Foroughi’s telling: “I’m like, ‘Damn, we might not get Gio in’” — but Ge saw more equity upside and joined faster. Ge’s own correction is worth keeping: “I wasn’t thinking that way… If I just wanted to stay on the winning team, I would have not chosen to leave. I wanted to find a place where I see opportunities and I can actually make a difference.”
• What Ge found: a lean team where Adam and then-CTO Basil were “deeply involved into the day-to-day,” but junior engineers “were not elevated” — handed low-level tasks with no business context. He took that as the opportunity, not a red flag.

2. A culture where code replaced meetings

• Ge asked Basil for a weekly one-on-one; Basil “didn’t understand what that was because it’s not a thing here in AppLovin.” They soon canceled it — “mostly it was code, and that was much more efficient.” Basil, per Foroughi, personally wrote perhaps 60% of the company’s code before Ge arrived.
• The numbers behind the leanness: ~100 engineers then and now, the company steady around 400 people after being cut from 600 to 400 about two or three years earlier, gaming businesses since sold, and Adjust, a software-as-a-service company, “bought and never integrated.” Foroughi’s engineering principle: “technology moves really fast, and you gotta be humble about what you have… you gotta throw away what you have and re-architect it… probably every couple years or even faster.”
• Product management barely exists in Ge’s organization — “probably fewer than a handful” of product managers. The philosophy: engineers competent enough about business problems “to just write their own architecture… to solve business problems that maybe not even a business team knew existed.”

3. Axon 2 was partly built on a 12-hour flight with no internet

• Ge planned to start two months later than his original date after a Europe trip; Basil talked him into “a week of work” first. That week converted him — he coded through the vacation, including a 12-hour flight to Italy right as ChatGPT-3.0 had just come out but was “definitely not enough to write code for you.” No internet meant reading library source code directly: “a man is given infinite amount of time… but I have no access to external help.”
• The human detail as told: an Italian in-law kept asking “Giovanni, are you okay? Are you losing your job?” — “No, no, no, I’m just very, very passionate about what I’m doing.” By the time he returned, the Axon 2.0 training infrastructure was ready and they started training the first-generation model.
• Ge’s meta-point: everyone asks how he changed AppLovin, “but very rarely people ask how AppLovin changed me” — watching Foroughi and Basil “gave me a new definition of what it means to be hands-on.”

4. What Axon 2 actually is: trees out, embeddings in

• The old Axon 1 clustered user-item combinations into a tree — “hundreds of thousands of if/else branches” — which couldn’t model relationships that shift with time, weather, promotions, and holidays. Axon 2 encodes otherwise meaningless IDs into “semantically meaningful embeddings” via learnable tables, passes them through a deep neural network to study user-item interactions, and extrapolates “to unseen data and unseen user-item pairs.”
• The underappreciated kicker: it was also cheaper to run. Neural nets are built from standard GEMM (general matrix multiplication) operations that GPUs are “highly optimized for,” while selection trees “are very hard to optimize on a GPU machine.” Accuracy up, infrastructure cost down, advertiser returns up.

5. The closed loop: 60-second playable ads and lipstick math

• These aren’t banner ads: users sit with a playable mini-game for ~60 seconds, and roughly half the ads are opt-in ads watched for in-game rewards. Ten ads a day ≈ ten minutes of engagement — “like a mini-game serving website” — and every interaction is a feedback data point. Crucially, “just a download for us means nothing”; the loop closes only when the advertiser’s data shows regular engagement and revenue.
• The business model was born of a confessed weakness: “I realized I don’t have an ability to sell all that well” — so instead of a sales force, build a system where the advertiser measures profit and scales spend themselves, “desperate to work with us, not the other way around.” That skews the platform toward businesses “you almost have never heard of before” with “very large P&Ls.”
• Why performance metrics make engineering tractable, in Foroughi’s example: a $10 lipstick with a $5 cost leaves a $5 spread — “they can’t spend more than $5 on selling the lipstick… Engineering, here’s where we’re at. We’re either good or we’re bad.”

6. E-commerce began on a napkin — and answered the gaming-data bear case

• The pushback from analysts, investors, and others covering the business: “you only have gaming data, so how could it work?” Foroughi’s answer: “it’s a billion people on the other side, not a billion people who are only playing games and doing nothing else.” Like Facebook’s approach, they started pixeling advertiser websites; the casual-gamer base turns out to skew slightly female, 30-50, and include many heads of households.
• Ge’s origin story — which Foroughi heard for the first time on this podcast: Basil and senior leaders at a Google conference in Vegas sketched the first e-commerce data-flow engine on a breakfast napkin, had a prototype before arriving in San Francisco, and launched with test advertisers within a couple of months. “Even with very premature model, we’re able to see people were making a purchase. The ROAS was decent, and we’re like, ‘Okay, this is gonna work.’”
• Ge’s generalization: with a great engineering team “a lot of times you don’t have to guess” — build the prototype fast and let the result settle the debate.

7. Creative is where alpha still lives — and there’s no formula

• Foroughi, 21 years in advertising: “I don’t know that there’s a formula to know how to create a great ad… None of us can really predict what a consumer’s gonna respond to.” It takes shots on goal — “if you create 30 ads a week, probably one of those might be interesting” — differentiated concepts, and nothing transfers across brands.
• Social UGC playbooks do not simply transfer here. Social is “really high ADD” — three seconds to capture attention — while AppLovin’s Solitaire/Mahjong user has “differently constructed minds” and 60 seconds of dwell. Advertisers who came into the roughly 18-month-old e-commerce vertical early and learned the format “create alpha… much larger campaigns at successful metrics” versus those who ported social creative wholesale.
• On gen-AI ads: short clips work, but a 30-60-second brand-safe video that “doesn’t mess anything up” isn’t there yet. Ge distinguishes two product paths — fully automatic (AppLovin’s aim, since “we’re serving all the advertisers”) versus collaborative human-in-the-loop tools, which are more widely adopted today. On acquiring tools like Higgsfield: it’s built for influencers, not a brand such as Wayfair — “there’s nothing in the market that we’ve seen that solves that problem,” so the team has to develop it too.

8. Engineers on top of AI, accountable for AI-assisted decisions

• Ge’s signature formulation: “I don’t want our engineer to sit next to AI. I want our engineers to sit on top of AI” — as the boundary of what AI can do moves from syntax fixes to autocomplete to entire PRs, humans migrate to what it still can’t do, notably long-horizon planning.
• Do people still need to learn to code? “At this moment, yes” — though “I’m not sure my answer will keep the same in the future.” AI “very often still makes bad coding decisions,” and engineers who can’t code can’t catch them. The house rule for core systems: “you as a human is held accountable for every decision that your AI made for you. ‘Oh, I don’t know, AI did this’ — this is not acceptable.” Prototypes and temporary dashboards can be less robust; core infrastructure stays “absolutely clean.”
• Foroughi’s complement: AI commoditizes code, so alpha requires smart people who “direct the AI a specific way” — same as ad creative, where “if AI starts writing all the ad creative, all of it’s gonna eventually look the same.”

9. Why many companies may struggle with the AI transition

• Foroughi’s blunt estimate: “probably 99% of people that use AI today just use it as a search engine.” AI-native startups are fine; AppLovin stayed lean with a high talent bar so its people “should technically be AI native… If they’re not, they’re probably not gonna survive here.” But legacy organizations with people who “push back at the usage because they want to justify their jobs”? “You almost have to replace nearly everyone and rebuild the culture… there’s no clear answer.” Token-max budgets just produce “a bunch of slop” — “everyone’s paying Anthropic and OpenAI for the same stuff.”
• Ge’s diagnosis of the same failure: because AI makes building easy, people execute “a lot of bad decisions and wasteful, bad-taste ideas” that offset AI’s benefits. His line: “AI really helps us to solve the problem, but it actually doesn’t change the problem we have to solve.” And his hot take: the differentiator isn’t building ability — “it’s the taste. The taste to know what to build and when not to build.”
• On rebuilding for the AI era, Ge rejects the premise: “I don’t really want to rebuild… we are definitely building on top of what we have today.” The business grew while engineering headcount stayed flat because growth “was perfectly in sync with the advancement of AI” — and the team was able to tackle a new-generation model that would be even more powerful than Axon 2 with AI assistance.

10. If chatbot usage is mostly search, its ads may resemble search ads — AppLovin’s next surfaces are CTV and open web

• The framework: search is bottom-of-funnel (the user already knows the shoes they want); AppLovin is top-of-funnel discovery, the Instagram window-shopping model — “we wanna start the funnel and then close the loop.” If 99% of AI use is search, chatbot ads “are probably gonna look and feel a lot like search,” which is someone else’s excellent business.
• The R&D vectors on a “three, five, 10-year time horizon”: connected TV and video placements on the open web — both potential discovery surfaces, both fragmented rather than duopoly-controlled, since frontier labs don’t look like they’re becoming operating systems.

11. The trillion-dollar math and the drawdown playbook

• The arithmetic as stated: EBITDA run rate “over $7 billion,” ~75% converting to cash after SBC — with the comparable figure three years earlier probably around 1/20th of today’s — and a trillion-dollar valuation needs “$30 billion-plus of cash flow a year” at a potentially strong multiple. Mobile gaming UA “is not big enough” for that, hence consumer/e-commerce, and beyond that the trillion-dollar template: dominate a product, then “extend to adjacent categories and execute exceptionally well.” No hyperscaler ambitions — “we don’t have much CapEx.”
• The 92% playbook, for today’s crushed SaaS names: keep stock comp in fixed-dollar terms (if a company normally issues 2-3% of its equity and the stock falls 90%, maintaining the same dollar value could require roughly 30% dilution, which “you’d never recover” from); skip IR entirely — “we actually didn’t do any investor relations for well over a year at the bottom… nobody buys something that’s dirt cheap. They wanna see a vision”; and deploy every dollar the company makes and more into buybacks: “we’re gonna become our best investor.”
• His warning on who can’t copy it: enterprise SaaS platforms are “robust but not algorithmic” — accelerating revenue means hiring go-to-market into a falling stock, potentially creating “a downward spiral.” Companies in that position may be acquired by private equity for a levered restructure and taken private.

12. No excitement, low ego, and no current plan after AppLovin

• Foroughi’s emotional register, in his own words: “I don’t get excited… Everything is a grind.” Going public at $30B wasn’t a celebration but “holy shit, we gotta make those people money.” Ge admits the mirror image — a “desire to not disappoint” that makes it “hard to press the pause button and just enjoy that moment.” Foroughi’s generalization: in highly competent people “no moment is good enough… the downside is it’s hard to reach fulfillment.”
• On singularity, Ge’s honest uncertainty: “Sometimes I have to believe that singularity moment will never come… As a human, I have to defend the value of humans.” Foroughi’s grounding: most of the world isn’t the Twitterverse — “people actually like to shop… humans need human interaction.”
• Hiring filters: intelligence plus low ego (people who take feedback and proactively question themselves; high performers can be vulnerable precisely because no one around them questions them), plus Foroughi’s third test: how they react to adversity. And on what comes after AppLovin: “not a lot of people in the world have had the privilege to be able to work with people and together build something from zero to $100 billion-plus… I can’t really sit there thinking about what’s after this.”