Inside Coatue: $70B Hedge Fund’s AI & Retail Strategy

Summary

• Coatue’s Michael Barton sees at least one answer to the “where’s the AI revenue?” bubble worry in ad businesses, though he says it is not enough by itself. Meta was expected to grow 15% and is growing mid-to-high 20s, Google Search went from sub-10% expectations to mid-teens, and Instagram time spent rose ~15% in six months after GPUs were put into the recommendation engine — “that kind of incremental revenue growth, that’s AI. Now, it’s not generative AI, but that is GPUs accelerating machine learning.”
• His formative scar is Melvin Capital’s GameStop blowup: “probably the best performing hedge fund in the world” to “basically down 50% in two weeks.” The lesson was that internet-coordinated retail is both a new risk category with no obvious catalyst, unlike Volkswagen/Porsche squeezes, and a new sourcing channel — “if you go on WallStreetBets, people are posting real work there” — with the same excitement now seen around Opendoor’s ~700% run.
• The AppLovin call is his template for founder-led conviction: at around a ~$20B market cap, meeting CEO Adam Foroughi convinced him there was something special. He messaged his boss mid-meeting — “You have to get in here right now” — and if you believed Foroughi’s claims, “you literally could not make the discounted cash flow analysis, in your worst case scenario, be less than like a 3X” as ad growth moved from roughly 15% to 50% and 70%.
• His explicitly personal call on labor: “any job that exists in the US where you work at a computer at some point will likely be automated, including my job.” The tell is slowing hiring, not firings — and for the Magnificent Seven, flat headcount on 20% revenue growth means margin expansion, accelerating EPS, and possibly “the front of a multi-year amazing run in the stock market.”
• Coatue is eating its own cooking: Barton says “today 85% of what I do basically can be done by AI” and is hiring analysts to reimagine every workflow. The goal is for six analysts hired over three years to become “sector heads with 25 agents working around the clock,” betting rivals will not adopt fast — “we’re gonna be light years ahead.” His self-designation: “the AI captain… maybe the last analyst.”
• The value-accrual debate — Cursor vs. the labs vs. Google (which is “the labs plus the Cursor plus their own cloud, plus their own little mini NVIDIA with their TPUs”) — is deliberately unresolved: “I think all three win for a while.” OpenAI’s $500B round makes sense to him by comparison with Meta near $2T; if Meta goes from $2T to $6T in five years, “what could the 500 go to?” TAM shorthand: $20T of labor spend vs. $1T of software — “that $20 trillion’s up for grabs.”
• The Reddit case shows Coatue’s data edge in action: when AI Overviews displaced Reddit’s link and the market said “this will never grow again,” their tracking showed Reddit appearing in Overviews rising from ~2% to ~15%, above the old-world ~10%. Conclusion: Reddit may be more valuable in an AI world, and its ~$50M/year Google licensing deal may understate what incremental human product-talk data is worth for shopping agents — “my guess is higher.”

Deep dive

1. GameStop broke the old risk model — and made the internet a sourcing channel

• Barton’s origin scar: at Melvin Capital, then “probably the best performing hedge fund in the world” among single-manager long-short funds, the GameStop short took them “basically down 50% in two weeks” because “we didn’t realize how powerful retail could be when they focus all their energy on a single stock.” Prior squeezes — Volkswagen/Porsche — had catalysts; this was “a lot of guys and girls on the internet deciding they were gonna buy it,” and “it completely changed investing and the risk that people think about.”
• The flip side is opportunity: fifteen years ago inputs were 10-Ks, 8-Ks and the Journal; now “if you go on WallStreetBets, people are posting real work there.” Coatue tracks Reddit mentions, Twitter, and internet trends, and sources actual ideas from them — the same excitement around Opendoor that Molly referenced, a stock Barton notes went up “700% or something” on excitement.
• The setup: Coatue runs roughly $60B AUM — about $25B in public equities, plus private markets and credit — with Barton on public TMT: internet, China internet and cloud, alongside a retail product in development.

2. AppLovin: trust the founder, then check whether the worst case is still a 3X

• The find as told: a friend flagged an around-$20B mobile-gaming ad company (“the name is amazing… it’s almost like a meme name to begin with”). Barton had CEO Adam Foroughi in knowing nothing beyond “they do mobile games,” and messaged his boss mid-meeting: “You have to get in here right now and meet this guy.” “I’m busy.” “Trust me.” — “This guy was the most locked in person I have ever met.”
• The mechanism: AppLovin was a pure play on the digital-ad market’s use of GPUs to improve ad engines, with growth moving from roughly 15% to “15, 50, 70.” His stress test: put Foroughi’s claims in a model and “you literally could not make the discounted cash flow analysis, in your worst case scenario, be less than like a 3X” — “the most remarkable thing I’ve ever seen.”

3. Advertising is AI’s first real revenue line; agentic commerce re-splits the profit pool

• Against the bubble worry — hundreds of billions of committed capex vs. ChatGPT subscriptions — Barton’s answer is that significant AI revenue is already appearing, though “it’s not enough” and more is needed. Meta was expected to grow 15% and is growing mid-to-high 20s; Google Search went from sub-10% expectations to mid-teens. “That kind of incremental revenue growth, that’s AI… GPUs accelerating machine learning to find and serve you an ad for a snowboard that you might not otherwise have seen.”
• Second-order effect: recommendation engines. Instagram time spent was flat for roughly 18 months, then rose ~15% in six months as GPUs were put into the recommendation engine — more time, more ad dollars.
• On agentic shopping he hedges hard: “it’s early. It’s very early.” The OpenAI Shopify/Etsy integration “is not at a place to really be that useful, but you can kinda see where it’s going.” He endorses Tobi at Shopify’s framing of discovery purchases: “those were actually not impulse purchases. I actually secretly wanted those things, but no one had ever shown them to me.” Endgame: agents proactively suggest (“you’re going on this trip, I think you need a new ski coat”), and the merchant’s ~20% marketing spend, of which 2% currently goes to Shopify, may re-split away from Meta-style ads toward Shopify and the agent players — “people are debating this literally every day.”

4. The IQ-100 child grows up: from the 2024 revenue scare to automated computer jobs

• In the summer-2024 AI scare — power, utility, infrastructure and tech stocks down 15-20% in three weeks because beyond ChatGPT “you couldn’t really point to anything else” — an xAI head engineer reframed it for Coatue: “Each model is like a child” whose IQ rises with each breakthrough; at the time “the IQ of the child is about a hundred.” A 100-IQ person has plenty of work in this economy, but a child has to grow up — applications lag the tech.
• A year on, that played out: coding broke out first (Cursor, Windsurf before its acquisition, and Cognition) because lab researchers code — “what’s the first thing they’re gonna try to figure out? How to make their jobs better” — and it’s broadening to Excel models, financial services and call centers. Barton’s explicitly personal view: “any job that exists in the US where you work at a computer at some point will likely be automated, including my job.”
• Of the two camps — 10X-efficient workers means hire more, versus efficiency “orders of magnitude way higher than 10X” — he’s in the second. The tell is not firings but slowed hiring, visible in the college-grad software-developer charts; Molly adds Klarna and Opendoor as turnarounds counting on attrition and, in Opendoor’s case, AI agents.
• The market math: Magnificent Seven companies grew revenue ~20% with headcount in line; flatten headcount and margins rise, EPS growth accelerates, “the stock is gonna go up a lot” — hence “we might be at the front of a multi-year amazing run in the stock market.” His unresolved worry, hedged as stated: displaced workers need new industries, “who’s gonna buy the goods if there’s unemployment?” — though “this will happen a little slower than some of the fearmongers think.”

5. AI-native operators win — including Coatue itself, “maybe the last analyst”

• The Foroughi standard: Foroughi says AppLovin has the highest EBITDA per head of any company in the world, everyone must use AI now (“if you’re not, you’re fired”), and he’s building “not for what the tech is today but where the tech is going” in two years. Companies with that mantra “are the ones that are gonna win.”
• Barton’s admission — “I tell my friends this, and they laugh” — is that “today 85% of what I do basically can be done by AI. It’s not a question of is the tech ready? It’s how do we implement the tech.” Coatue is hiring an analyst class to reimagine every workflow, from two hours of morning sell-side triage to one-click model builds, so that in three years six analysts are “basically sector heads with 25 agents working around the clock.”
• He’s not worried about his seat: hedge funds aren’t people-intensive, and the binding constraint is time to look at ideas. “I don’t think other firms are going to adopt this that fast, and we’re gonna be light years ahead.” His title for himself, accepting Molly’s coinage: “The AI captain.”

6. Stocks reprice in seconds; the long term is a collection of quarters

• OpenAI’s DevDay as exhibit: getting named on stage meant “bang, you’re up five” — Mattel, a toy company, up 6% “in a second” — even though Barton argues that being included in ChatGPT’s agent layer “might not actually be a good thing.” His generalization from tech waves: even the most bullish person on AI probably underpredicted GPU demand, and disruption runs faster than expected too — “normally it ends up being better than you think to the upside and worse than you think to the downside.”
• The method marries horizons: he models Meta’s revenue, EBIT, profit and free cash flow out to 2031 and runs DCFs, but “the long term is simply a collection of quarters” — Netflix’s end state was knowable, yet every hiccup was a 20% drawdown, so the craft is not being massively sized before the hiccup and sizing up after the overcorrection. He traces the industry arc from Julian Robertson’s multi-year fundamental style through credit-card-data quarter traders and sector-focused, market-neutral managers to today’s hybrid.
• Dispersion discipline: even with the Nasdaq up ~17% this year, AI infrastructure and power names — including Constellation Energy — are up ~50%, while Microsoft and Meta are up ~25% — book sizing within the winners drives outperformance. Philippe’s “single best quality is his risk management”: cutting gross from 100% to 50% invested “very quickly” with strong timing during drawdowns, including the tariff-board period.
• Getting into the book is its own skill: 95% of the work is the thousand-line model and expert calls, but you must compress it into “a three-sentence pitch that when he hears that pitch, he’s almost ready to buy the stock before even opening the model.” Thomas, he says, is the best he’s ever seen at it — a skill “I’m still developing.”

7. The value chain: Cursor vs. the labs vs. Google — and why OpenAI at $500B pencils

• The through-line of his process: “the best way to figure out what’s gonna happen in tech is to actually talk to the practitioners” — CEOs, but also OpenAI, Anthropic and the researchers. “They will tell you what they think.” In eight or nine years he’s “never seen a moment where a few private companies are impacting so much public market cap” — you now need the whole chain, down to NVIDIA’s allocations, because cloud revenue is “100% dependent upon how many chips you get.”
• An Anthropic reinforcement-learning friend’s case study frames the accrual debate: Cursor at one end, “the most loved, most used coding agent”; the labs in the middle; and Google — “the labs plus the Cursor plus their own cloud, plus their own little mini NVIDIA with their TPUs, plus a search business, plus data on everything.” On paper Google wins — “well, they’re also the slowest.” His honest non-answer: “I don’t know the answer yet… I think all three win for a while.”
• On OpenAI, in which Coatue is an investor, the $500B round “makes sense to me”: 800M weekly active users, time spent by his estimate near Instagram’s, versus Meta near $2T — and he thinks Meta could 3x in five years, so “if the 2 goes 6, what could the 500 go to?” — plus unmodeled optionality (social, cloud), talent density and “the zeitgeist.” His TAM shorthand: $20T of labor spend versus ~$1T of software — “that $20 trillion’s up for grabs.”
• The hedge: picking winners below OpenAI is “really difficult” — every startup bet waits on the next OpenAI launch, as with the n8n-style “here’s our version” moment.

8. The data edge: inflections pull IRRs forward — the Reddit case study

• What Coatue tracks: inflections in growth and margins that prove a thesis early — “IRRs get pulled forward” — via credit-card data, email traffic and a Thursday KPI review of every covered company, owned or not, which doubles as a macro read (ads were strong in Q3, then slowed a week ago: consumer weakness or shoulder period?).
• The Reddit trade shows the edge: when Google’s AI Overviews displaced Reddit’s link and user growth hiccuped, the market’s reflex was “this will never grow again.” Coatue’s illustrative data: Reddit appeared in ~10% of old-world Google searches but only ~2% of early Overviews — yet as Overviews went from 5% to 50% of searches in two months, Reddit’s appearance rate rose to 15%, above the old world. Takeaway: Reddit moves into “the AI winner camp,” the multiple re-rates, and the stock went up a lot.
• The licensing kicker: OpenAI trained ChatGPT on Reddit data and “may have not asked for permission”; Google did the same kind of thing. Google pays Reddit ballpark $50M a year, and ChatGPT does the same. The view that the fee would not grow was widely held, including by the companies. But if there were rumors of Mark Zuckerberg hiring people for $100M a year to build a shopping model, what will Google or OpenAI pay for Reddit’s valuable corpus of human product conversations that shopping agents may need? “My guess is higher.”