Summary

Consumer AI’s monetization ceiling is a myth, George argues, and history proves it: private-market investors once capped Facebook and Google at “$20 a user” — a decade later they make ~$200 per developed-world user. ChatGPT (likely) reached Google’s scale roughly 4x faster, has ~a billion users and monetizes fewer than 50 million of them; the interface will shift from a reactive chatbot to something “proactive… long-form memory… multimodal,” and the ad format that wins will be as unforeseeable as feed-based ads once were.
On enterprise AI he’s deliberately the skeptic: the “software is $400B but white-collar labor is huge” slides are “a little bit handwavy.” His base case — “90% of the technological surplus is going to go to the end users,” like the steam engine, which wasn’t priced based on replacing 50 laborers. The clearest business-model progress is in customer support (outcome pricing) and coding (consumption); everything else is “pretty TBD” — yet the next generation of enterprise winners can still be bigger than the last.
The episode’s core market call: growth above 30% is still not fully valued because it’s hard for investors to model persistence — 2009 consensus for Apple’s 2013 was off by 3x on “the most covered company in the world,” and a growth rate that decays 80→75→65 instead of 80→65→50→40 is “a 3x difference in your valuation.” His portfolio is growing 112% dollar-weighted at a 21x revenue entry — “way less risky than a 12% grower in PE at 15 times EBITDA.”
Most tech markets are Glengarry Glen Ross: “First prize gets Cadillac. Second prize gets a set of steak knives. Third prize, you’re fired.” The vast majority of market cap goes to the leader — “there’s no number two to Salesforce.” Exception: the model layer looks like cloud or aircraft manufacturing, not airlines — multiple players with profit pools. Being #2 in model revenue is fine; being #2 in the dominant consumer chat interface is not.
The likely Waymo trade is his formative lesson in overriding spreadsheets: in 2020 he resisted (“it’s going to take 10 years… the valuation is going to be really high”) and Mark and Ben said “Don’t care… this is the mother of all markets. Stop overthinking it.” A small 2020 check became a much larger one at end-2024 once “consumer preference slapped you in the face” — the likely Waymo overtook ~50,000 SF Bay Area Lyft drivers in market share with only ~400 cars. Robots, by contrast, face “endless degrees of freedom” — he thinks the robotics timeline is longer than expected.
Venture has quietly become a grown-up asset class: eight of the ten largest market caps are tech (seven West Coast venture-backed), private market cap is ~$5 trillion, up 10x in a decade — nearly a quarter of the S&P 500 and more than half of the Mag 7 — while public markets have shrunk by half in 20 years and fewer than five public software/consumer/fintech companies grow 30%. The scarce asset — high growth — now lives private.
How a16z grades AI companies: pull over push (“Is the market demanding more of your product?” is on a post-it on his monitor), durable engagement (Harvey usage step-changed when reasoning models landed — “lawyers need to reason”), and — inverted from the SaaS era — low gross margins as a badge of honor: “I’m an AI thing and I got 75% gross margins… no one’s using the AI stuff then.” Endgame maybe 50% margin businesses, not 80, but big enough not to matter.

Deep dive

1. The chatbot is not the endgame — consumer AI’s upside is open-ended

George’s starting posture is humility: “I don’t think that we have yet found the dominant [product] in AI” — even though ChatGPT “has grown faster than anything in the history of technology,” reaching Google’s scale roughly 4x faster, with ~a billion users. The big shift he sees: from reactive to proactive — long-form memory, multimodal, offering solutions before you ask. ChatGPT “probably has the best chance” of capturing that, but the chatbot form factor is “way too limiting.”
The monetization math is his favorite historical rhyme: more than 10 years ago, evaluating Snap and Twitter, everyone anchored on Facebook and Google making “$20 bucks a user… that’s kind of the upper bound.” Ten years later they make ~$200 per developed-world user. ChatGPT monetizes fewer than 50 million of a billion users; really active AI users already spend ~30 minutes a day in the products, versus ~50 on Instagram and ~70 on TikTok.
Patrick’s light-bulb moment was mundane — Deep Research picking a baseball bat for his 9-year-old (length, drop, specifications) where “Google… would be a total mess. Amazon, no chance because of the ads.” From there it’s “an execution problem” of building capabilities to execute that stuff on your behalf on the web — hard (Instagram famously found native shopping too hard), but shopping is just one category.

2. Enterprise AI: 90% of the surplus goes to end users, and the business models are TBD

Where George diverges from peers: consensus-bullish on consumer, “slightly more skeptical about what their ultimate business models will be” in enterprise. The confident slides — software is only ~$400B but look at white-collar labor — are “a little bit handwavy.” The most developed pricing models are customer support, a discrete task with simple completion analysis where you can shift from Zendesk-style seats to per-completed-task pricing (“you know what it’s worth”), and coding, which is consumption-driven in a developer world already trained to pay that way.
His prior: “90% of the technological surplus is going to go to the end users. Just start with that as the assumption.” The steam engine wasn’t priced based on replacing 50 laborers — competition drove it to a fair return on capital and users kept the productivity gains. Same logic as the iPhone: what would you actually pay? “90% consumer surplus is probably low.” Yet even so, “the next generation of business companies can still be much bigger than the previous generation.”
The startup path through incumbents runs on three things: business model shift, reimagined UI, and new data sources — the more dramatic the shift, the harder it is for incumbents to react. That’s the Decagon thesis in customer support: “the odds are so stacked in their favor” because incumbents find the model change hard to react to, and it’s better-faster-cheaper by an order of magnitude. In the SaaS/cloud wave, market revenue 7x’ed and split roughly 50/50 between incumbents and startups; a sharper business-model break should tilt it toward startups this time. “My hope is that’s what happens, but we’ll see.”
His Salesforce sketch is the tell: today it’s “basically a sophisticated form checker… brutal, painful to use.” The AI future logs you in with “these are the five customers you should be doing business with… I’ve drafted a call script… I’ve already taken a bunch of action on your behalf” — powered by unstructured data pulled from every interaction rather than just Salesforce’s existing database. That new data could give startups an advantage, though Salesforce’s sticky database remains an incumbent advantage.

3. Robotics is the biggest market — which is exactly why the timeline discipline matters

“These are the ones that are the biggest market opportunity… if you knew it was going to work in 5 years, you’d put all your money into it.” George happens to think robots “will take a little bit longer,” informed by likely Waymo: a car “needs to basically stay in a lane, avoid collisions, go a certain speed limit, find places to park” — and that took Waymo 10 years, and the industry roughly two decades from the DARPA challenge. A home robot faces “endless degrees of freedom.” Generative AI techniques will compress the curve, but not to zero.
The investing posture, given uncertainty: the early-stage team meets every robotics company and waits for a seed/Series A-able team; the growth fund waits for it “to work.” What does working mean? “I think we’ll know it when we see it… customers pulling their products that we will have not seen before.”

4. The autonomous-driving story: “Stop overthinking it” — and the 400-car surprise

a16z first invested in Waymo (likely) in 2020, its first outside capital ever (previously pure Google funding), as the only VC in the round. George’s 2019 test ride showed unprotected lefts and construction avoidance — though it stalled at a parking lot and needed an override. He fought the deal: “I said no, I don’t like this at all… it’s going to take 10 years, the valuation is going to be really high.” Mark and Ben: “Don’t care. This is autonomous driving… the mother of all markets… stop overthinking it.” They compromised on a small check.
Five years later, at end-2024, “consumer preference slapped you in the face” — anyone in San Francisco with the choice took a Waymo — and the firm wrote a much larger check into the new round. The kicker: Patrick guessed Waymo runs 10,000 cars in SF because “they’re everywhere.” Actual number: ~400 — optimally routed, fully utilized vehicles that overtook the ~50,000 Lyft drivers in the Bay Area in market share.

5. Edge comes from people: the “technical terminator”

George’s one-line philosophy: “I like to pay fair prices for great companies” — the art is “recognizing where greatness may lie where other people don’t.” Lots of people can model margins and unit economics; edge comes from product, market, and people insights, and people is the hardest. His favorite archetype: the “technical terminator” — starts technical, grounded in product, then learns business as a sponge. Ali at Databricks is the canonical case: one of seven on the open-source project, not even CEO at first, who now “knows more about sales ops and hiring processes and reporting lines than probably any of our CEOs.”
The pattern repeats — Zuckerberg, Elon, George (likely Kurtz) at CrowdStrike, Dave at Roblox (“a little quieter” on the surface, but “ruthlessly competitive and he really cares about market cap creation”), Dylan at Figma (“one of the nicest guys in our industry, but he is brutally, ruthlessly competitive”). Among the AI generation: Michael (likely at Cursor), and Shiv (likely at Abridge) — a practicing cardiologist commuting from Pittsburgh who showed George where he’s putting a bed in the office: “I just want to be working all the time when I’m in town.”
The counterexample underscores the point about founder-market fit, not technicality: Travis at Uber, where the market “was just a pure battle — you fight mayors, you fight competitors” and demanded someone ruthlessly competitive and operationally intense. On passing decisions generally: errors of omission are “really, really painful” and economically costlier than commission — you can only lose 1x on a bad deal but forgo unbounded returns on a miss.

6. Most markets are Glengarry Glen Ross; the model layer is the exception

The firm’s adopted metaphor for tech markets is the Alec Baldwin scene: “First prize gets Cadillac. Second prize gets a set of steak knives. Third prize, you’re fired.” His strong view: “the vast majority of market cap creation is going to go to the market leader” — and peers underrate this (“even the number two player is going to be really viable — maybe, but more often than not that’s not the case”). It’s obvious in network-effect consumer businesses; less obvious but just as true in enterprise: “There’s no number two to Salesforce… Workday is Workday, ServiceNow is ServiceNow.”
The model layer looks different: early in technological shifts, “markets tend to fragment in ways that we don’t foresee.” Technical leads keep leapfrogging, so he expects a cloud-industry structure — multiple players with profit pools — framed early as “is this going to be aircraft manufacturing or airlines?” Aircraft manufacturing has high capital intensity and technically difficult production; airlines “all go bankrupt in the fullness of time.”
Why did cloud sustain multiple winners? “It’s all size of market — it’s just so vast.” And despite partner Alex Rampell’s line that “the best businesses in the world don’t have customers, they have hostages,” cloud customers are actually well served (egress fees aside) while the clouds remain great businesses. The caveat that keeps the Glengarry rule alive: being #2 in model revenue is fine; “what’s not okay, probably, is being the number two in the dominant consumer chat interface.”

7. Venture became a grown-up asset class — and the good growth moved private

“We’re a grown-up industry now… no longer some little bespoke asset class.” When he and Patrick left college, one or two of the world’s ten largest market caps were tech; now it’s eight of ten, and seven of the eight are West Coast venture-backed. Private market cap is ~$5 trillion, up 10x in 10 years — almost a quarter of the entire S&P 500, more than half of the Mag 7.
The scarcity argument: in a16z’s public universe (software, consumer, fintech), fewer than five companies are growing 30% — “kind of staggering” — while the growth portfolio grows at 112% dollar-weighted. Public markets have shrunk by half in 20 years, and small-cap public quality “is so much lower than what is available in the private markets.”
Competition has barbelled: on one end the large multi-stage firms with strong venture practices — “the fiercest competitors… trying to play the same game as us,” fighting to hold seed/Series A winners tightly. On the other end, deep-specialist boutiques — the Gucci or Prada store to a16z’s Walmart/Amazon — Nat and Daniel, Elad, and the crossover funds he respects.

8. Deals are won in the two years before the call comes

No sensational war stories: “The reality of the growth stage business is we win deals based on years of relationship building.” A recent founder of “one of the best companies in the market” called and said he’d transact without going to market — “fruits of my labor, two years of this.” The trade-off that creates: a clean look at a price below what the founder would get in the market, but “can I bear the price?” The two years of earning it: helping with candidates and customers “as if we were already investors.”
The Figma saga is the template. Joining from GA in 2018, George found Peter Levine “apoplectic” — “We need this tomorrow… we did GitHub early, how did we not do this one?” The firm ran the full-court press: summit invites, “Mark and Ben bear hugs,” bear-hugging Dylan’s crypto interest, placing a board member. Dylan’s answer: “I’ll let you know when.” Then COVID hit, markets cratered, and he called: “Now’s the time.”
The internal debate is the lesson: George’s growth team ran the traditional lens — “the market for designers is not that big… I don’t think the price makes sense at $2 billion.” The venture side was “losing their minds”: the designer-to-engineer ratio was basically 2:1 at modern tech companies, front-end engineering and design were melding, and “thinking about this as the market for design is way too limiting.” Ben called the meeting; George slept on it and concluded the risk he’d take was market size — never founder or business quality. “You need those product and market insights or you’re just going to live in a spreadsheet and die in a spreadsheet.”

9. Firm design: single trigger-puller, zero reserves, game film, Yankees

The growth fund deliberately rejected the investment-committee model (“you battle to get the votes… the smoke comes out and here’s the decision”) for the venture-style single trigger-puller, with openly expected disagreement and then commit. The claim: it removes the temptation “to politic for a vote,” so risks and rewards get fully explored. His first “IC” was breakfast with Mark and Scott before he’d even joined. The team is ~10 investors — kept small because the early-stage teams feed it — and juniors are evaluated from day one on “contribution to collective investment judgment.”
Concentration mechanics: a little over half of investments (and ~70% of dollars) go into companies where a16z already has a position — “game film… it’s not just numbers.” Reserves are near zero by design: small follow-ons only; every large check “is a new investment,” because reserving “would lead to lazy decision-making.” The largest positions read: Databricks, SpaceX, likely Anduril, OpenAI, xAI, Flock Safety, Figma, Stripe, Coinbase — most held across multiple funds, with no target metrics by sector or inside/outside the portfolio.
The culture is explicit: offer letters come with a signed culture document, Ben personally runs onboarding, and George’s growth-fund principles include “there’s a scoreboard in this business and our expectation is that we win” and “we are the Yankees and we’re going to act like it” — meaning performance standards, not arrogance. His pre-join perception that Mark and Ben were “semi-celebrities — do they really work hard?” inverted immediately: “man, it is a competitive place… no one is resting on their laurels.” On why no buyout fund: “all that we stand for is helping the next generation of companies go beat the incumbents… buying the incumbent and squeezing as much as they can out of their customers is culturally antithetical to what we do.”

10. Why markets misprice growth — and why 2022–early-2025 is a great vintage

The mechanism: “It is just so hard for any investor to build a five or ten-year model where high growth persists. It’s just not natural.” Nobody modeled Google or Visa growing 15–20% twenty years in. The killer stat: 2009 consensus estimates for Apple’s 2013 were off by 3x — “the most covered company in the world.” Modeling 80% growth decaying to 75 then 65, versus the standard 80→65→50→40→terminal, is “like a 3x difference in your valuation.” His conclusion: “above 30% growth, the market still doesn’t fully value the growth rate.”
The portfolio expression: last year’s activity entered at 21x revenue with 112% dollar-weighted growth — and, acknowledging revenue multiples are flawed, “if I could invest for the rest of my career in 112% growing companies at 21 times revenue, I would do it in a heartbeat. That’s way less risky than buying a 12% grower in PE for 15 times EBITDA, because growth just takes care of so much for you.”
Cycle placement: the perfect setup is “early product cycle, bad capital cycle — but those rarely coincide.” Forced to choose, product cycle is everything. The under-told lesson of 2021: “the biggest mistake is that we were actually kind of late product cycle and we just didn’t realize it — there was a bit of a head fake with COVID… the ideas are just worse.” Today’s LP questions are all bubble questions, but his answer: 10 years from now there will be a bunch of great companies, “we’ve got to be on the field” — and 2022 through early 2025 “is going to be a great vintage,” helped by companies staying private “longer than we expected.”

11. Pull beats push — and low gross margins are the new badge of honor

The post-it on his office monitor: “Is the market demanding more of your product?” When yes — Roblox with two network effects, likely Anduril where AI capability, likely Palantir/SpaceX-alum government know-how, and “a desperate geopolitical need” converged — you get the most special companies. The surprise in likely ChatGPT’s billion organic users: “it doesn’t have a network effect — that was one of the more surprising things for us.” Push businesses “don’t tend to get easier over time… the bigger you get, often it gets harder,” especially ad-fueled consumer; TikTok is the exception that pushed early and aggressively — largely via Facebook ads, perhaps a decision Facebook gets to think about “forever.”
The AI screen has three parts. Ease of acquisition: Cursor’s viral growth, or Abridge, where hospital systems must be sold but “are dying for this because the doctors love it.” Durable engagement, since “there are some head fakes… things that grow really fast and then fall off”: Harvey’s usage took “a step change” right as reasoning models landed — “lawyers need to reason, and turns out models got really good at reasoning.” And gross margins, where the SaaS-era 70%+ bar has inverted: “we get these pitches — ‘I’m an AI thing and I got 75% gross margins.’ I’m like, well, no one’s using the AI stuff then.”
The margin pass isn’t unconditional — “there’s a big difference between 30% and 70% gross margins, so we do care” — but the expectation is inference costs fall over time (even as reasoning has exploded token usage, so margins haven’t improved yet), and model providers won’t hold enough market power to squeeze the app layer. Endgame: “maybe they end up as 50% margin companies as opposed to 80 — but the size of the impact and the usage is probably so high that it’s fine.”
The unique-product-to-unique-distribution flywheel: “Every great company either has unique product or unique distribution. The best have both — such unique product that it leads to unique distribution.” GitHub sold Walmart a $400K contract and “no one ever talked to them on the phone” (“what would they have paid if you just called? Probably $4 million”). Cursor’s go-to-market notes read “immediately to POC… immediately to full sale” every time — until infrastructure partner Martin replied to a 100-person email thread with just “product market fit,” and internally “PMF is now PFMF.”

Michael Burry Speaks | Michael Lewis The $700 Billion AI Productivity Problem No One's Talking About