a16z's David George on the Most Controversial Bet at a16z & Do Margins and Revenue Matter in AI?

Summary

• The large-fund debate answered with data: a16z’s best-performing fund in firm history is a $1B fund — Databricks has returned 7x the fund and Coinbase 5x in DPI alone — and its larger funds have outperformed its smaller ones at similar multiples. The structural case: private market cap has grown 10x in 10 years to over $5T, and in a16z’s study of the top 50 IPOs from 2017–2025, 53% of the dollars of gain came at Series C or later. “It’s about the number of winners you capture.”
• George claims public markets still offer a cheaper cost of capital; Harry’s pushback lands — likely Ramp and Lovable are “priced at the same price as Wix and Wix is doing two billion of profit” — and George’s honest non-answer is “we’re not close to those companies.” The real private-market benefit, per George: avoidance of stock-price volatility and employee management (Stripe, SpaceX, Databricks), not cheaper money.
• The labor-to-tech budget shift has green shoots: CH Robinson, a truck brokerage, disclosed a 40% AI-driven productivity gain (shipments per person per day since end-2022) with operating margin up 680bps; Microsoft cut headcount 6%. Meanwhile public software is “guilty until proven innocent… assumed doomed by AI unless proven otherwise” — Harry notes that caps anything without a labor-replacement story (his Monday.com and Duolingo positions included).
• AI revenue counts only with engagement behind it: you can’t see years of renewals at this speed, so a16z’s bar shifted to retention and engagement as leading indicators — and the inverse tell is now a screen: “if we ever see a company that pitches us as an AI company and they have SaaS gross margins, we ask a lot of questions” — it probably means nobody is using the AI features.
• Kingmaking is a flimsy thesis. “If the investment thesis is our investment is going to make them a winner, it’s probably a pretty flimsy investment thesis.” Preferential attachment to the existing leader is real; SoftBank’s capital-as-a-weapon was “a bit of an adverse selection machine” — the companies that opt in are the ones without a reason to win, and the subsidy funnels back to Google and Facebook.
• Models won’t eat the apps — a16z “fully changed our mind” 18–24 months ago; radiologist headcount rose even though AI read scans better, because scans are only 30–40% of the job. The model layer itself goes oligopolistic like cloud: he thinks ChatGPT will lead consumer, the B2B fight with Anthropic is the head-to-head — and Anthropic is George’s lingering error of omission.
• The autonomous-driving investment was the biggest internal fight — George’s team modeled the 2020 price as too high; Mark and Ben overruled with “it’s autonomous driving… the endless market size,” and the compromise (small check, then a much larger one later) worked. Flow is strength-of-strengths underwriting on Adam (likely Neumann): rent is 30% of the average US renter’s disposable income and “the only unbranded experience in anyone’s life.” Robotics, where a16z has made no large investment yet, George thinks could be “the largest category in AI.”

Deep dive

1. Large funds can 5x — a16z’s best-ever fund was $1B

• The opener answers Everett Randall’s claim (relayed by Harry) that you can’t look LPs in the face and promise 5x on large funds: “our best performing fund in the history of the firm is actually a $1 billion fund” — Databricks has returned 7x that fund, Coinbase already 5x in DPI, with GitHub, DigitalOcean and Lyft alongside. Larger a16z funds have outperformed the smaller ones at similar multiples. “It’s about the number of winners you capture.”
• The structural case: private markets grew 10x in 10 years to over $5T of market cap, and disaggregating the top 50 IPOs from 2017–2025, 47% of dollars of gain came seed-to-Series-B versus 53% from Series C onward — “I was actually surprised when we looked at this.” a16z’s LSV funds hold $700B–$1.5T of aggregate market cap; apply ownership assumptions and 3–5x returns are “pretty manageable.”
• The wave argument: mobile/social/SaaS/cloud/e-commerce created $20–25T of market cap, and “if that started from scratch today… so much of that value creation would take place in the private markets.” He never expected Salesforce at $230B or ServiceNow at $175B — “we’re in inning one of this new big tech wave,” so this generation’s winners could be bigger still.

2. The cost-of-capital spat — George says public is cheaper; Harry says the comps disagree

• Harry’s worry that private companies now compete each other away before ever exiting (Axon eating Flock Safety’s Atlanta business, both private) gets waved off: public-versus-private has little to do with competitive dynamics. The real Flock story is that law enforcement flipped from a terrible category to a wonderful one — and companies staying private “has been to our benefit because we’ve been able to increase our ownership.” a16z has led three Flock rounds and historically does not take chips off the table.
• George insists no public CEO has ever told him “I regret going public” and that publics offer a cheaper cost of capital. Harry’s counter: likely Ramp and Lovable are “priced at the same price as Wix and Wix is doing two billion of profit.” George’s honest dodge, twice: “We’re not close to those companies… I’m not close enough to know how they’re valued relative to their performance.”
• The genuine private-market advantage, per George: “the avoidance of volatility in your stock price and sort of employee management” — Stripe, SpaceX and Databricks have made it work, even at a slight discount to public pricing. Harry’s John Collison paraphrase carries the founder side: “I don’t need some 25-year-old associate to tell me that I need to plan more efficiently.”

3. The asset class grew up while public small-cap rotted

• The count of public companies has halved in 20 years, and the Russell 2500’s ROIC — “the easiest measure of the quality of the company” — has slid from 7.5% to 3% over 30 years. Returns now accrue in privates before companies list; the asset class “is no longer a sort of bespoke small thing. It’s the grown-up leagues” — private tech quality names dwarf US tech private equity.
• His advice to a hypothetical $10B endowment (self-declared “heavily biased,” and conceding many endowments are already overallocated to privates): 8 of the world’s 10 most valuable companies are US West Coast venture-backed tech, so if the next 20 years rhyme with the last, allocation should tilt toward the asset class holding the next dominant companies. Compliance bars return talk, but “the top performing venture funds outperform” top PE — and AI implementation, “the most important thing for companies over the next 10 years,” will make the gap more extreme.
• The operational consequence: companies staying private longer must become multi-product, multi-channel, international — “and with AI, it’s happening much much faster” — so a16z changed its own business around servicing that.

4. The growth fund is internally “the fix-the-mistake fund”

• Brian Kim’s framing confirmed: repairing the venture team’s errors of omission is “very much” the growth fund’s charter, run jointly with early stage (“what Series A’s do you wish you had done?”). By the numbers: about half of what they do is follow-ons from existing venture companies (the ElevenLabs growth round after Jennifer and Brian’s early round), ~15% follow-ons on growth-originated deals (Flock, Figma, SpaceX, likely Waymo), and roughly a third fully net new — always with a pre-existing founder relationship.
• The Deel miss — passing between Anish’s Series A and the Series C a16z co-led — teaches the firm’s core filter, inherited from Ben: invest in “strength of strengths as opposed to lack of weaknesses.” The failure mode is fear of theoretical competition — the old “isn’t Google going to do it?” trope: “if you overweight the fear of future theoretical competition, you can always talk yourself out of making an investment.”
• Harry’s mirror-image confession from passing ElevenLabs and Deel at seed: “I thought I was smarter than markets… I should have just 100% backed up the truck on amazing founder.” George’s other repeat error: passing because the market looks too small — “we always underestimate the size of a market.”

5. Venture risk at mature prices is fine — for about five people

• Harry’s charge: the market is taking venture-stage probability at previously mature-company prices. George’s carve-out: it works when “some degree of likelihood of success is very very very high despite a very early stage” — Sarah’s Character AI round at a growth price, because backing likely Noam meant “a pretty safe downside and an extremely high upside.” The population warranting that underwriting: “the list is five people, I think.” And they “almost never make an investment saying, oh, we’ve got the liquidation preference.”
• Harry then runs the math on paying ahead ($10B for Sierra his example): a company at $50M ARR must 5x, then 4x, then 3x to reach $3B ARR on “pretty optimistic” rates, and at a 6–7x public multiple that’s only ~3x on today’s price. George won’t concede the frame: “I’ve been historically surprised at how good the best companies can be and how fast they can grow” — the winning AI apps grow 3x faster than predecessor SaaS, and not every winner trades at six times; “high valuations from the outside, I think, in many of those cases are warranted.”

6. AI disrupts SaaS through business model first — and the “new incumbents” are tougher prey

• His rank-ordering of disruption vectors against today’s SaaS incumbents: (1) business-model shift — Decagon pricing customer service per completed task (“if you are going to compete with a seat-based customer service thing, lookout”); (2) UI and workflow; (3) access to data. When all three change at once, “you’ve got a really good chance for a startup to come and beat the incumbent.”
• He rejects the eat-all-labor maximalism (“we have that on slides, too”): in practice “there’s massive surplus that gets delivered to end customers and you can still create much bigger companies than the previous generation.” Harry’s condition: labor budgets must actually convert to technology budgets, “cuz if we don’t… we’ve all just overpaid a [__] ton.” Harry’s caveat is that the shift must be product-pulled, “slapping the customers in the face,” not CIO-mandated AI initiatives.
• On the 50-vendor customer-support pile-up that baffles Harry: the category is better-faster-cheaper with today’s model quality — “you don’t need to believe any future state.” Roughly half of SaaS/cloud markets are winner-take-vast-majority and half fragment like payroll; either way Decagon’s “growth is staggering, the market pull is staggering,” and it converts most executive briefings into a deal.

7. Fast revenue still counts — if engagement backs it; SaaS margins in an AI pitch are a red flag

• Does 100M-ARR-in-months revenue (Gamma under Grant his example) mean what revenue used to? Yes, “if it is high retention and high engagement” — with no renewal history at this speed, engagement is the leading indicator, and “the bar for assessing that is way higher than it used to be.” The magic combination is organic acquisition plus engagement: ElevenLabs, ChatGPT, xAI, Abridge, Harvey — “the market is just absolutely starving for their product.”
• Triple-triple-double-double isn’t dead: “the number one way to measure a company is ultimately return on invested capital,” and required momentum is relative to your peer set — in fast markets, “momentum gives you a chance to build a moat.” Harry’s rejoinder is opportunity cost: a solid compounder is fine, “but is it the best place for my precious dollars and for my LP’s precious dollars?”
• On margins: history says they rationalize upward, though today is muddy — token costs down massively while reasoning drove usage up. He expects model-market structure “sort of like cloud… relatively oligopolistic” with reasonably high margins, and even a 50%-gross-margin app generation is “totally fine” if it delivers value. The tell runs the other way: “if we ever see a company that pitches us as an AI company and they have SaaS gross margins, we ask a lot of questions — it probably means people aren’t actually using the AI features.”

8. Kingmaking is a flimsy thesis; preferential attachment is real

• The direct answer to Harry’s kingmaking question: “if the investment thesis is our investment is going to make them a winner, it’s probably a pretty flimsy investment thesis.” What is real is preferential attachment — increasing returns to scale even without network effects: the more you lead (Salesforce, Workday, ServiceNow, CrowdStrike), “the more resources come your way and the easier things get.”
• The SoftBank Vision Fund critique, with credit given first (early to AI via Nvidia, good picks like Slack): the flaw was believing “capital as a weapon was a viable strategy.” It’s nearly impossible in enterprise (you physically have to hire reps) and mostly fails in consumer (TikTok, maybe Uber, the exceptions) — and it’s “a bit of an adverse selection machine,” since companies opting into subsidized winning “maybe don’t have as good of a reason to win in the first place,” with the money funneled back to Google and Facebook.
• The retail barbell as firm strategy: scale players (Amazon, Walmart) and specialists (Chanel — “this is where Europe really thrives”), with death in the middle — department stores without scale. Harry: “Do you mind being, like, Walmart?” George: “We’re happy to call ourselves Amazon.com. Customers love it.”

9. Models won’t eat the apps — and the model layer likely goes oligopolistic

• The firm’s biggest change of mind, 18–24 months back: “we sort of all thought at first the models will just do everything and subsume everything. We fully changed our mind” — application companies will be built on models “in pretty much every direction.” The radiology proof, as told: neural nets beat radiologists at scans before this wave, yet radiologist counts went up — scans are only 30–40% of the job, and “the model companies aren’t going to go do the work to automate the other 60 to 70%.” Harry’s nuance (OpenAI doing customer support, Google shipping a Lovable competitor) gets the AWS analogy: hyperscalers offer everything, yet independent infrastructure companies thrive.
• Market-structure call: like cloud — “if you could own all of AWS, Azure, and GCP as independent companies, that would suit you pretty well.” Anthropic is the error of omission that lingers: “they’ve done a really good job.” He expects OpenAI and Anthropic to diverge — he thinks ChatGPT will lead consumer, the B2B fight with Anthropic is the head-to-head, and Google will play some part — and a16z has underwritten later OpenAI rounds “very much with the mind of consumer.”
• On when OpenAI’s entry price stops making sense: “we have to constantly reassess this.” The humility case: Databricks at $6B in 2019 — the largest deal in growth fund one — “our investment case never would have predicted what they became”; Google and Facebook monetized users at one-seventh today’s rate a decade ago. The pattern they hunt: a theory the core market is bigger than consensus (Stripe, SpaceX with Starlink, likely Waymo) plus founders who find the next product — Anduril from one program of record (border towers) to, unpredictably, autonomous fighter jets.

10. The autonomous-driving fight, Flow’s controversial bet, robotics’ potential

• His biggest disagreement with Mark and Ben: the original likely Waymo investment in early 2020, when a16z was the only VC fund in the round. George’s team produced analysis showing the price was too high; Mark and Ben’s response: “it’s autonomous driving… this is the endless market size.” Resolution: a smaller check then, a much larger one in the latest round. He cites an op-ed he thinks was in the New York Times by a medical professional: likely Waymo’s data now shows 7–10x safer than a human driver — results that in a clinical trial would fast-track approval; “it would be irresponsible to block this.” Autonomous driving and robotics are “maybe the mother of all markets coming on AI.”
• Flow decoded through strength of strengths: Adam (likely Neumann) “has some of the strongest strengths of anybody, any entrepreneur in the market” — spiking on product and hiring. The insight: the average US renter spends 30% of disposable income on rent, the highest spend of any category, “and yet it’s the only unbranded experience in anyone’s life.” The team has sort of proven out the value prop, and now it’s scaling — with Harry supplying the Calm founder’s heuristic: “How often do you meet a founder like Adam?… then write the [__] check.”
• Quickfire residue worth keeping: the best picker at a16z is Dixon (“the clearest articulation of what our early stage strategy is”); Mark “can see the future — give Mark any 10-year prediction… most of the time they’re right,” while Ben is “probably the best management coach or understanding of executive dynamics that I’ve ever encountered.” Harry’s own changed mind: a16z and YC — “every great European company is a YC company.”
• Most memorable first meeting: Shiv of Abridge, a practicing cardiologist who “knows his end market, knows his product, knows the technology and yet is a total total killer” — Harry maps the archetype onto Winston at Harvey: domain authenticity plus tech-founder aggression. Next-decade excitement: proactive personal health management (“one of these large consumer categories that hasn’t really hit yet”) and robotics — no large a16z bet yet, but George thinks it could be the largest category in AI, B2C, B2B.