David George - Building a16z Growth, Investing Across the AI Stack, and Why Markets Misprice Growth - [Invest Like the Best, EP.450]
What this covers
The conversation is a working account of how David George runs the growth practice at a16z, with the host, Patrick, pressing on method. George covers how the team makes decisions without traditional committees, how it works with founders years before investing, and how it thinks about missed winners. His view is that errors of omission cost more than bad investments, since a loss is capped at one times the money while a passed-on great company is not. He also describes the favored founder type, the Technical Terminator, who starts close to the product and then learns sales, hiring, and management.
The second half moves across the AI stack. George argues the model industry will look more like cloud, or even aircraft manufacturing, than like a single winner, and that future interfaces will be proactive rather than chat-based. He is blunt that enterprise AI pricing is unproven, with only customer support and coding showing credible paths. The firm screens AI companies on customer acquisition, retention, and gross margin. The Waymo investment serves as a case study: a 2020 bet against a ten-year timeline, followed by a much larger check once cars were on the road.
The more contested material is about markets. George says public investors fall prey to Growth Persistence Modeling Bias, pointing to Apple estimates that missed by about 3X, and that few public companies grow above 30 percent against a portfolio averaging 112 percent. He also takes positions on Winner-Take-Most Market Leadership, on a Business Model Shift favoring startups, and on why 2021 was a late Early vs Late Product Cycle mistake.
David George argues that AI will reshape consumer and enterprise software, with value accruing mainly to end users and the market leader, and that growth investors earn edge through people, product, and market insight rather than spreadsheet forecasting.
- Consumer AI monetization is underestimated, as past consumer internet companies monetized far more per user than forecast
- Winner-take-most dynamics and business model shifts determine which companies win, while the enterprise AI business models remain unproven
- Growth investing edge comes from product, market, and people insights, and early product cycles are the best environment to invest
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
Roughly 90 percent of technological surplus goes to end users rather than the companies that build the technology, as with the steam engine, whose productivity gains were priced to a competitive return on capital and mostly accrued to the users of the machines.
“90% of the technological surplus is going to go to the end users. Just start with that as the assumption”
Investors struggle to model high growth persisting for 5 to 10 years, so they model fade curves that underestimate upside; in Apple's case consensus estimates for 2013 made in 2009 were off by about 3X, meaning growth persistence versus fade can change valuation by roughly 3X.
“consensus estimates were off by 3X. That's a massive number, and that's the most covered company in the world.”
The market still does not fully value growth rates above 30 percent, because investors cannot model high growth persisting for five to ten years, so models assume decay that is unnatural for companies like Google or Visa.
“No one built a financial model for Google or Visa that had them growing 20 years into existence at 15% or 20%.”
Investors in Snap and Twitter a decade ago assumed Facebook and Google would monetize about $20 per user as an upper bound, but ten years later both earned roughly $200 per user in the developed world, showing that the P x Q framing of consumer internet revenue consistently underestimates price.
“we always said, oh, Facebook or Google, they make $20 a user, so that's the upper bound. And fast-forward 10 years later and Facebook and Google make $200 a user in the developed world.”
Errors of commission cost at most one times the money invested, but errors of omission from passing on great companies can forgo very high returns, making missed winners economically more costly than bad investments.
“errors of omission are really, really painful too. And they're more costly, just economically”
Once Waymo had cars on the road by the end of 2024, consumer preference became obvious, with riders in San Francisco choosing Waymo whenever they could, and Waymo now runs about 400 cars in San Francisco yet has overtaken the roughly 50,000 rideshare drivers in Bay Area market share.
“Anyone who was in San Francisco who had the choice was taking a Waymo.”
During the SaaS and cloud wave, the market grew roughly seven times in revenue, and incumbents and startups split the new share about 50/50, so a dramatic business model shift is likely to favor startups more.
“7x more revenue, incumbents grew a bunch, they took half of the new share, startups took half the new share.”
The biggest mistake from 2021 was investing as if in an early cycle when it was late, with COVID a head fake, and 2022 to early 2025 is seen as a good vintage for investing.
“it turns out that the last two years, coming out '22 to early '25, I think were a really good period.”
The model industry will likely resemble aircraft manufacturing, which has high margins due to capital intensity and technical difficulty, rather than airlines, which are fiercely competitive and often go bankrupt.
“Aircraft manufacturing has high profit margins because there's really high capital intensity and it's extremely hard technically.”
Buying 112 percent growing companies at 21 times revenue is less risky than buying a 12 percent grower at 15 times EBITDA, because growth de-risks the investment, and the market still fails to fully value growth rates above 30 percent.
“I think that's way less risky than something where you're buying a 12% grower in PE for 15 times EBITDA. In a weird way, it's less risky, because growth just takes care of so much for you”
Private technology markets have become a real asset class, with about $5 trillion of private market cap that grew roughly 10x in ten years, representing almost a quarter of the S&P 500 and more than half the Magnificent 7 in value.
“Turns out there's 5 trillion of private market cap that is up 10X in the last 10 years.”
In the public software, consumer, and fintech universe where a16z spends most of its time, fewer than five companies are growing 30 percent, while the firm's portfolio averages 112 percent dollar-weighted growth, so private markets hold far more high-growth quality than public small caps.
“The public universe in those sectors, there's less than five companies growing 30%. It's kind of staggering. That's a low number. Our portfolio on average dollar weight is growing 112%.”
The future interaction with AI will not be a chatbot; it will be proactive, with long-form memory, multimodal inputs, and the ability to take actions on the user's behalf.
“I don't think that the future of how we interact with AI is going to be a chatbot, I just think that's way too limiting. I think the big shift will be what is reactive today to something that's proactive in the future.”
The AI model industry is likely to resemble cloud rather than winner-take-all because it will be large enough to fragment and leapfrogging means technical advantages are short-lived, so multiple players will have profit pools, not one dominant model company.
“So on the model side, so far, the way it looks like it's played out is it will be more like the cloud industry. It's not going to be a winner take all.”
The 2022 to early 2025 period was a good vintage for investing because the late-cycle excesses of 2021 had cooled, and David expects the current AI cycle to be an excellent period to invest, despite bubble concerns.
“I think this is going to be a great vintage of time to have been investing.”
The vast majority of market cap creation in technology goes to the market leader, and being number two is often not viable, as shown by the lack of a viable number two to Salesforce, Workday, and ServiceNow, so a growth investor should be wary of second-place companies.
“the vast majority of market cap creation is going to go to the market leader. And this is probably underappreciated.”
Cloud supports multiple winners such as AWS, Azure, and GCP because the market is so vast that each can be a very valuable business, and the clouds remain positive-sum for customers despite anti-competitive egress fees.
“That one is all size of market. It's just so vast.”
The ideal environment for growth investing is an early product cycle, meaning the outset of a new technological change, since a bad capital cycle rarely coincides with it, and the early-cycle years from mobile, cloud, SaaS, and e-commerce made the start of George's and Patrick's careers strong.
“Well, the optimal would be early product cycle, bad capital cycle, but those rarely happen and coincide with one another.”
a16z's first growth fund invested in Waymo in 2020, when the team's analysis suggested a 10-year timeline and stretched returns; the partners pushed the deal because autonomous driving was the mother of all markets, and the firm later wrote a much larger check after cars were working on the road.
“This is autonomous driving. Are you kidding me? This is the mother of all markets.”
The founder archetype George most favors is the technical terminator, who starts as a deep technologist close to the product and then learns sales, hiring, and management, as Ali at Databricks did, so that technical grounding leads to figuring out the next product area.
“they start technical and then you never know if these people are going to become commercially-minded excellent business people.”
The biggest mistake of 2021 was that the industry was actually in a late product cycle and failed to recognize it, with COVID as a head fake, so the ideas and market opportunities were worse than they appeared.
“The biggest mistake from 2021 is that we were actually late product cycle and we just didn't realize it at the time.”
a16z evaluates AI companies on three metrics: ease of customer acquisition, which must be strong even for sold products; customer retention and engagement, as fast-growing experimental products often fall off; and gross margin, which it gives a partial pass for now.
“One is ease of customer acquisition... The second is customer behavior, customer retention, customer engagement, and then there's gross margins.”
The growth fund uses a single trigger puller decision process rather than an investment committee, where each partner must disagree openly but then commit, so decisions are fast and avoid vote politicking.
“we were going to actually make the decision process just like our venture process, which is single trigger puller.”
Upstarts beat incumbents most often through a business model shift that incumbents find hard to react to, combined with completely reimagined UI and new data sources, and the more dramatic the shift, the more it favors startups over incumbents like Salesforce.
“Business model shift is a super powerful thing that's very hard for incumbents to react to.”
Travis Kalanick at Uber is the counterexample to the technical terminator, because the market was a pure battle with mayors and competitors that required ruthless competitiveness, driven and operationally intense operators rather than product depth.
“one of the elements of people judgment is what is the right founder for the right market? And that market was just a pure battle.”
Traditional growth-lens market sizing undervalued Figma's design market, but the venture team's view that design and engineering were converging justified an investment, showing growth investors need product and market insight to avoid living in a spreadsheet.
“You need those product and market insights or you're just going to live in a spreadsheet and die in a spreadsheet.”
Enterprise AI business models are very uncertain; investor slides claiming that the roughly $400 billion software market is small next to white collar labor spend are hand-wavy, and only a few areas like customer support and coding have credible pricing paths.
“to me that's a little bit hand-wavy, so there's a couple of areas where the business model has progressed in a compelling way to go tackle that directly.”
Every great company has either a unique product or unique distribution, and the best companies have both, because a truly unique product leads to unique distribution, and founders should exploit that position by pursuing enterprise customers after bottom-up adoption.
“every great company either has unique product or unique distribution. The best companies in the world have both.”
A pull business is one where the market demands more of the product, which is the most special outcome; a push business requires going to sell the product, and those tend to get harder as they scale.
“So is the market demanding more of your product? Is probably the most important question that we can answer”
AI customer support can shift from seat-based pricing to charging per successfully completed task because the task is discrete and completion is easy to verify, while coding is more naturally priced on consumption, as developers already pay that way.
“there's a very discreet task with very simple completion analysis that you can do, it's simple to price it on that.”
The growth fund keeps only a tiny reserve for small follow-ons where it is not the lead, and makes zero reserves for large investments, treating every large follow-on as a new investment so that earlier reserves do not encourage lazy decisions.
“We do zero reserving for large investment amounts that we think we're going to make in a company, because I think that would lead to lazy decision-making.”
For exits, the firm uses semi-algorithmic judgment: if the founder is still running the company and the company is the market leader, the firm biases to hold longer, and otherwise biases to exit sooner, while also weighing valuation against performance.
“is the founder still running the company, which we value a lot. And then, a qualitative, are they the market leader that we feel great about?”
The most valuable companies are pull businesses, where the market demands more of the product, which happens organically and is very rare, so the central diligence question is whether the market is demanding more of the product.
“Is the market demanding more of your product? Is probably the most important question that we can answer”
Being number two in revenue or awareness can be fine in a cloud-like fragmented model market, but being number two in the dominant consumer chat interface is probably not acceptable.
“What's not okay, probably, I would think, is being in the number two in something like the dominant consumer chat interface.”
ChatGPT's billion-user growth is organic and brand-driven, and surprisingly does not have a network effect, which surprised the firm.
“it's a billion users, it's organic, it's all brand. And the shocking thing about that one, by the way, is it doesn't have a network effect.”
We have not yet found the dominant consumer AI product; ChatGPT may have the dominant brand, but a chatbot is too limiting, and the future interface will be proactive, with long-form memory, multimodal inputs, and solutions offered without being asked.
“I don't think that we have yet found the dominant product in AI.”
The firm gives AI companies a partial pass on gross margin because a low margin often means customers are really using the product, while inference costs keep falling and token usage rises, so AI businesses will likely settle at around 50 percent gross margins rather than SaaS-like 80 percent.
“Now, it's like a badge of honor to have low gross margins, because we're like, 'Oh, at least people are using your AI products.'”
Salesforce is essentially a sophisticated form checker that is painful to use, and the AI-native future is a proactive system that drafts calls, monitors customers, and takes action, so the stickiness from its database weakens as value shifts to unstructured data from every interaction.
“It's basically like a sophisticated form checker with some analysis and it's brutal, it's painful to use.”
a16z first invested in Waymo in 2020 as its only VC investor, but David initially opposed the deal on the grounds of a ten-year timeline and high valuation, and it was Marc Andreessen and Ben Horowitz who pushed the small investment through.
“We originally invested in 2020, they came to us to raise outside capital for the first time.”
Robotics is the biggest market opportunity, but it will take longer than people expect; Waymo took about 10 years from DARPA-era autonomy progress, and the whole autonomous driving industry took roughly two decades to reach its current point.
“but it took Waymo 10 years, and if you go back to the DARPA challenge, the whole industry, decades to get this point, two decades to get to this point roughly.”
The growth fund runs as a scoreboard business where the team is expected to win, with the Yankees principle meaning high performance, collaboration, and hard work, because returns for founders, customers, and LPs depend on it.
“There's a scoreboard in this business and our expectation is that we win.”
Growth deals are won through years of relationship building, not dramatic moves; a16z helps non-portfolio founders as if they were already investors by recruiting candidates, bringing customers, and showing deep understanding of the business, which can make non-investor deals easier because the brand is strong.
“The reality of the growth stage business is we win deals based on years of relationship building.”
David deliberately blocks two hours on Tuesdays and Thursdays and an hour and a half twice a week in afternoons so he has time to think and learn, since he otherwise doesn't learn as much as he should.
“unless I block off a bunch of time, I actually just don't find that I'm spending the time learning as much as I should.”
Financial modeling of margins and unit economics is widely available, so growth investors get edge from product insights, market insights, and people insights, with people judgment the hardest and most valuable.
“lots of people know how to do that analysis, that's out there, so where can you actually get edge? You can get edge from product insights, market insights, and people insights.”
The Figma investment shows the tension between growth and venture lenses: the growth team judged the design-tool market too small to justify roughly $2 billion, while venture partners argued the designer-to-engineer ratio and the engineering-design melding made the market far larger, and George ultimately accepted the market-size risk for an exceptional business model and founder.
“The risk I don't want to take is quality of business, quality of founder. But you really had to have a nuanced view of the market”
The core investing style is to pay fair prices for great companies, with the art being recognition of greatness before the market fully prices it, because most growth investors know how to forecast margins and unit economics, so the real edge comes from product, market, and people insight.
“You can get edge from product insights, market insights, and people insights.”
a16z decentralized decision-making from a group of 25 to 30 decision-makers to smaller fund-level teams, because deep sector expertise improves decisions and go-to-market, while large groups make cross-fund tradeoffs too hard.
“you can't scale an organization with 25 or 30 decision-makers around the table, it's too hard.”
Just over half of the growth fund's investments and about 70 percent of its dollars go into companies where the firm already had venture or growth-originated investments, giving deep game film before the growth check.
“So if you just do it by dollars, it's a little over half that are preexisting venture investments.”
For robotics, growth investors wait for early-stage teams to find a team worth a seed or Series A and then look for signs of it working, which will be recognized when customers pull products in ways not seen before.
“I think we'll know it when we see it, there will be things that start happening and customers pulling their products that we will have not seen before.”
Push businesses that must be sold get harder as they grow, especially in consumer and ad-driven models where Google and Facebook accumulate better economics at the expense of advertisers, with TikTok as a rare exception because it pushed aggressively early.
“If you have to go sell or market your product, the bigger you get, often it gets harder.”
George wants his time split so that about 80 percent goes to learning about new companies and founders and 20 percent to known portfolio companies and internal management, and he blocks two hours on Tuesdays and Thursdays for thinking because open-ended browsing otherwise does not turn into learning.
“I want most of my time spent on the new stuff because I need to be learning about those new markets.”
Growth-stage relationships are built by helping companies before investing, which gives founders reasons to call the firm first, reducing competition and giving an investor a clean look at a top company.
“the founder... called us and he was like, "Hey, I'm ready to do this. I'll just talk to you."”
Deep research solved a complex baseball bat purchase that would have been a mess on Google and useless on Amazon because of ads, which showed that the models will improve and that the real challenge is the execution of acting on a user's behalf on the web, requiring guardrails and a lot of product plumbing.
“Deep Research was really, really, really good at it, and it solved my problem for me. So, light bulb went off for me at that moment, one, the models are going to get so much better, and two, to me, it's an execution problem”
Harvey's customer engagement took a step change as models improved their reasoning, because lawyers need to reason, so usage rose sharply once models became good at reasoning, an example of a model capability unlocking product engagement.
“Harvey is an example of a company where, as the models have gotten better, customer engagement and usage has actually really grown.”
Driving is a constrained task of staying in a lane, avoiding collisions, obeying speed limits, and parking, while a home robot must handle endless degrees of freedom such as making coffee and doing laundry, which makes general robotics much harder than autonomous driving.
“Endless degrees of freedom. Make a cup of coffee, go do my laundry”
The 2022 to early 2025 period was a good vintage for growth investing because valuations were more attractive and private companies stayed private longer than expected, which allowed the firm to convert them at attractive prices.
“I think this is going to be a great vintage of time to have been investing.”
No one predicted feed-based advertising before feed products existed; it turned out to be probably the best advertising format in history, which is why it monetizes at high rates and why users like it.
“we never would've predicted what a feed-based advertisement is, no one would've known what that is because we didn't even know what the feed-based product was.”
When George started his career there were one or two technology companies among the ten largest in the world, but now eight of the top ten are technology companies, and seven of those eight are West Coast venture-backed tech companies.
“Now, it's eight of 10 and seven of the eight are West Coast technology venture backed companies.”
GitHub was so product-driven that for a long period it never talked to customers, yet it sold to Walmart for about $400,000, which George said implied far higher revenue potential had it called customers directly.
“for a long period of time, they never actually talked to customers.”
Alex Rampell's line that the best businesses have hostages, not customers, does not fit cloud, where customers are well served and the market is positive sum, although egress fees make leaving difficult.
“The clouds are anti-competitive with egress fees. They make it really hard to leave and get your data out and all that stuff, but that's minor.”
The iPhone example shows the consumer surplus is far larger than the price, since a user would pay much more than $1,000 for an iPhone, so a 90 percent surplus assumption is probably low.
“The sky's the limit. 90% consumer surplus is probably low, if the iPhone costs $1000 or something like that.”
Anduril shows pull because a confluence of AI capabilities, autonomy know-how, government navigation skills from Palantir and SpaceX alumni, and a desperate geopolitical need together create demand that the market is pulling for.
“We've reached all at the same time this confluence of AI capabilities, autonomy, know how, and how to navigate governments”
Bob Swan advised reviewing the prior year's calendar for two hours each year-end with the goal of cutting 30 percent of commitments, which pushes responsibility to teams and creates leverage.
“they spent two hours looking at their calendar from the year, and then they had an objective of cutting 30% of stuff that was on their calendar.”
Growth-stage competition has become a barbell, with large multi-stage firms that have strong venture practices as the fiercest competitors on one end and deep specialists on the other, analogous to a superstore versus a boutique retailer.
“the market has become a barbell. So we're faced with the large multi-stage firms that have very strong venture practices on the one hand”
Team members, starting at entry level, are evaluated on their contribution to collective investment judgment as part of promotion criteria, which makes junior people chime in and improves decision making.
“As part of the team's promotion criteria, evaluation, et cetera, I put in their contribution to collective investment judgment.”
Instagram tried native shopping and failed because it is too hard, so agentic shopping will require significant guardrails and product infrastructure before it works.
“Instagram famously tried to do shopping natively, it's just too hard.”
a16z does not run a buyout fund because buying incumbents and squeezing their customers is culturally antithetical to a firm whose mission is helping the next generation of companies beat the incumbents.
“culturally, buying the incumbent and trying to make them last as long as possible and squeeze as much as they can out of their customers”
The kindest thing George reflects on is the sacrifice his parents made driving him to soccer, baseball, and basketball, which shaped him and which he now appreciates through raising his own children.
“the sacrifices my parents made for me are extraordinary, they're incredible.”
Technical founders who combine relentless intensity with product understanding, such as Dave at Roblox and Dylan at Figma, look quiet or nice on the surface but are ruthlessly competitive and care about market cap creation, which is why backing them pays off.
“he's ruthlessly competitive. And he really cares about market cap creation”
Inference costs are expected to keep falling as technology history suggests, and although token usage has risen with reasoning, gross margins have not yet improved, but they are likely to improve over time.
“The cost of inference has gone down at the same time that reasoning happened. And so token usage has gone way up.”
In a first meeting George asks the founder to spend five minutes on strategy and vision, then asks questions for about 20 minutes, because founders consider it a compliment when the investor has done research and asks smart questions.
“I keep the introduction super brief. I like to jump in and say, 'Hey, why don't you please spend five minutes explaining to me the strategy and your vision?'”
Robotics is potentially the largest market opportunity, even though the industry is currently obsessed with LLMs and diffusion models.
“Robotics is the biggest market opportunity, we are all obsessed with LLMs and diffusion models.”
When a company has unique product, unique distribution, and a founding team that wants to optimize that advantage, the result is magic, which a16z partner Martin Casado summarized by saying 'Product fucking market fit' after Cursor's sales notes.
“So when you see that, that's unique product, that's unique distribution, and you have a founder, founding team or full set of employees who really wants to optimize it, that's magic.”
The firm uses the Glengarry Glen Ross sales contest scene, where only first place wins a Cadillac and the rest lose, as a metaphor for technology markets, where the market leader takes most of the value.
“So we've adopted that as a way of describing most of the technology markets that we live in.”
ChatGPT has grown faster than anything in the history of technology, reaching Google's scale about four times faster and reaching roughly a billion users, while monetizing only a tiny fraction of them.
“OpenAI and ChatGPT has grown faster than anything in the history of technology, I think they reached the same scale as Google, something like four times faster, a billion people using it”
Abridge, despite needing enterprise sales to hospital systems, is easy to adopt because doctors love it and it saves them time, so hospitals are eager to buy.
“hospital systems are dying for this because the doctors love it, and it's really good. It saves them a lot of time”
David is drawn to venture-style growth investing because it combines learning from exceptional founders and new technology with the competitive scoreboard of putting up returns for founders, customers, and LPs.
“I love to learn and I love to be around great founders as they're exploring really interesting things.”
When the firm correctly passes on an investment, it is typically because the business quality is weak, and the firm feels strongly that market leadership is required, as there are no common patterns beyond that.
“when we get it right on not doing an investment, it's typically for the right reasons.”
Even though most value goes to end users, the next generation of business AI companies can still be much bigger than the previous generation because the capability gains are so large.
“the next generation of business companies can still be much bigger than the previous generation of companies given the capability gains.”
The growth fund meets about 30 companies a week and David personally about 10; he wants his time split so that 80 percent goes to learning about new companies and new markets, with only 20 percent on known companies and internal management.
“trying to learn about companies, spending time with entrepreneurs I want to be 80% of my time.”
The growth fund meets about 30 companies a week, and George personally meets about 10, showing the scale of deal flow in late-stage AI investing.
“We, as a growth fund, probably meet 30 companies a week”
Active AI users spend almost 30 minutes a day in the products, versus about 50 minutes a day on Instagram and 70 minutes on TikTok, and AI is monetizing only a sliver of users, so consumer surplus is large and could support a massive company.
“really active users spend almost 30 minutes a day in the products. For context, users spend 50 minutes a day on Instagram, 70 minutes a day on TikTok”