Context
David George is a general partner at a16z who leads growth investing. Jack Altman interviews him on his show Uncapped. The talk starts with a common question, "will open source or frontier models win?", and widens into enterprise and consumer AI, autonomous driving, robotics, and how venture capital works in this cycle. It matters to PMs because it shows how a top growth investor judges which products and layers of the AI stack can last. That includes where AI labs are likely to compete with app builders and where they are not.
The Big Idea
The answer to most "this or that" AI questions is "and". Demand for AI is so far ahead of what exists that frontier labs, open models, app companies and new hardware can all grow at the same time.
David's evidence is that AI revenue today comes from a thin slice of users, mainly coders, and that data center supply is sold out. Diffusion into the wider economy has barely started.
Key Insights
AI revenue rests on few users
- David puts the combined revenue of the biggest AI labs and coding products at roughly $120 billion, with some of them adding more revenue per month than all hyperscalers except Amazon.
- Most of it comes from about 30 million coders, and within that group a power law, where the top million spend a huge amount.
- There are about 1.5 billion knowledge workers. He says AI is under 5% diffused into the B2B economy.
- Outside coding, writing works today. Legal is the next function taking off, which he places about 12 months behind coding.
Supply is the bottleneck
David says demand is "insatiable" and supply is constrained at every step of the chain. In his view you cannot get data center capacity until 2028. If capacity such as 4 gigawatts became 8, he expects it would get used, perhaps at different prices.
He sees two ways this could be wrong. One is that coding is a special case and the rest of white-collar work does not follow. The other is an algorithmic breakthrough that makes models far more efficient. He rates both as unlikely, and notes that more reasoning at inference time keeps improving results, which keeps inference demand high.
Token buyers need to show returns
Jack raised the best counterargument he had heard: most revenue today goes to sellers of tokens, so we still need proof that buyers get real efficiency gains. David's answer has three parts:
- A study from about a year ago found AI coding made people 20% less effective, but it predates the latest models and he expects a rerun would not show it.
- Cutting-edge companies like Stripe report high returns from faster code.
- Large traditional firms such as Procter and Gamble are probably not seeing big gains yet. He says they need a lot of hand-holding, the way Databricks and Palantir give it, and that this is a delivery gap and not a flaw in the tech.
Spend on the front end
David asks where the efficiency effort goes: building new products, or cutting costs. The best companies are, in his words, singularly focused on the front end. A product that works can bring 10x to 100x more revenue, while cost savings have a ceiling. He calls a cost-only focus "shorting your own future".
Labs have a limited blast radius
- David expects the labs to focus first-party products on coding and things near it (the "blast radius of coding").
- They will also build for all 1.5 billion knowledge workers, such as Microsoft Office and Google apps style products.
- Everything else sits lower on their priority list. For legal he guesses it is somewhere between priority number 6 and 15.
- Platforms also risk losing trust if they pick off companies on top of them. He recalls the old Microsoft idea that a platform only works if what is built on top earns more than the platform.
Apps survive on detail and go-to-market
David names two things labs struggle to copy in a vertical. First, the last details of the product matter a lot to users. Second, it needs real go-to-market and on-the-ground effort. He cites Harvey in legal, where end clients now require their law firms to use it. Even Replit and Lovable, which sit "dead in the center" of the coding blast radius, are working very well.
Frontier and open models both win
Today users reveal a preference for frontier models. Reasons David gives: people pay about 10x more for frontier tokens, the consumer surplus is large, and first-party products have harnesses tightly tied to their models, so nobody swaps in another model. Over time, if usage rises 50x, cost will matter and "n minus one" models (one generation behind) will be good enough for many jobs. He points to customer support tickets: most no longer need the top model, and app companies watch this closely because it is their gross margin.
Consumer AI is still early
- A billion people use chatbots, but most use them like a search engine.
- The big shift David expects is from reactive to proactive: assistants that take action for you in a trustworthy way. He thinks voice is now good enough, since it can handle interruptions without relying on a weaker model.
- He believes consumer will bring the best business models: subscriptions plus ads, or a new form of ads not yet invented. The native format is unknown, as a feed would have made no sense to someone in the newspaper era.
- On his own automations he says only about two out of ten give him real value today.
Autonomy is underrated
- David says rides in Waymo cars are between 10 and 14 times safer than human driving, based on millions of miles. He calls it "barbaric" to block that.
- His economics: owning a car costs about 80 cents per mile all-in, while Uber or Lyft cost about two dollars and change per mile. If autonomous rides fall below the cost of owning, he expects a demand explosion.
- Uber and Lyft grew the market 10x over taxis (from about $100 million to about a billion dollars in San Francisco, in his figures). He expects another 10x at least for ride hail.
- For car ownership he guesses buyers would pay at least $10,000 for full autonomy, across 17 million new US cars a year.
- Waymo cars feel everywhere but there are fewer than 10,000 of them.
Mental Models & Frameworks
Product cycle versus capital cycle
David scores each period out of 10 on two separate axes:
- Product cycle: how big the technology wave is. 2021 was a 1 (late SaaS, post mobile). 2010 was an 8. Today he gives a 9 or 10.
- Capital cycle: how cheap it is to invest. 2023 was around a 7. Today he says about a 6, "okay but not great".
- The two rarely line up. He says the product cycle is what drives returns over ten years, so he leans on that.
The three waves ahead
Past product waves (mobile, social, e-commerce, SaaS, cloud) together created about $25 trillion of market cap. David lists what is ahead now: AI, autonomy, robotics, AI in health, and defense modernization. He thinks the combined result is "way more than 25".
Spending as the new lever
In SaaS, pouring money in did not make a company better. At Salesforce, Workday or ServiceNow it would have mostly gone to sales reps. He cites the Vision Fund as an example of too much money into companies that could not use it, though he thinks it was a right idea at the wrong time. In AI, more money into training makes models better, which makes capital a competitive tool. This is why he thinks the power law is more extreme today.
Believe it ten times more
David quotes a line he first heard maybe from Thrive: you can be contrarian by believing what nobody believes, or by believing what everyone believes but ten times more. He says a16z growth leans on the second. His version: pay fair prices for great companies and expect the upside on the revenue side, not margins.
Decision Principles
Principle: Ask what could go right
- When: judging a high-growth company, especially at a later stage.
- Why: David says the alpha is on the revenue side. Markets often miss how long a company can keep growing fast, and good outcomes can be larger than any financial model.
Principle: Back founders who keep expanding
- When: deciding whether a company can sustain growth beyond its first product.
- Why: David says when things go right, it is often the founder who finds the next product. He cites Databricks, where revenue accelerated seven years after a16z's first growth investment, and SpaceX, where Starlink was not yet generally available and the economics looked bad when they invested.
Trade-offs & Nuance
Frontier quality versus cost
Right now heavy users want the frontier and pay a premium. David also says that at 50x usage, cost starts to matter and a model one generation behind becomes good enough for many tasks. Both can be true: the top tier keeps growing, and cheaper tiers grow faster as volume rises. Application companies feel this first because inference cost is their gross margin.
Labs versus app builders
Labs have the scale advantage. David notes that OpenAI and Anthropic together have raised about $350 billion, more than the entire downstream ecosystem. Even so, he argues they cannot build every vertical product with all the polish and sales effort it needs. His own filter for app investments in or near the labs' territory is to ask harder questions about the product roadmap and the company's right to win.
Technology readiness versus adoption
Customer support is his example. The tech has been better, faster and cheaper for two years, and the companies are growing very fast, yet penetration is still small. Selling, implementing, and messy integration and brand rules take years, and labs will not build that motion themselves.
Job impact needs thought
David expects a big productivity boom and says history shows people find higher-order work. He notes about 60% of today's jobs did not exist 70 years ago. He also says we will have to think hard about whether workers are trained to make the transition, especially if it happens fast.
Common Mistakes
Mistake: Treating AI as a binary bet
If you ask whether open or closed, labs or apps, Nvidia or new chips will win, you assume a single winner. David's view, shared with his colleague Eric Vishria, is that growth is large enough for several winners. The better question is which part of the market each player is best placed to serve.
Mistake: Only chasing cost savings
Using AI mainly to cut costs gives a capped gain, while building the next product can multiply revenue. David says the best companies are focused on the front end, and that cost opportunities will always be there for later.
Practical Application
Map your product against lab priorities
Write down where the big labs are heading: coding and the products near it, plus horizontal tools for all knowledge workers. Then check whether your product sits inside that area. If it does, the roadmap and your reason to win must be sharper. If it is a vertical like legal, the lab is less likely to prioritize it.
Test on a cheaper model
- Do: run your core user flows on a model one generation behind the frontier and compare quality.
- Then: note which flows need the best model and which do not.
- Why it works: David expects usage growth to push cost pressure, and your gross margin may depend on routing work to cheaper models.
Own the last details and delivery
If you build on top of AI models, invest in what labs will not do: fine product details for a specific user, plus implementation and integration work. David says these are the modes of defensibility for application software.
Aim at proactive, not reactive
For consumer products, ask which tasks an assistant could do for the user without being asked, and in what trustworthy form. David calls the shift from reactive to proactive the big consumer change still ahead.
Choose front-end ambition
When you pitch AI work inside a company, frame it as a way to grow revenue with a new product and not only as a cost cut. David's view is that the cost side has a ceiling and the product side does not.
Questions to Consider
- If the large AI labs launched a first-party product in your category tomorrow, what would still make your product the better choice for your users?
- Which of your product's flows would still work well on a model one generation behind the frontier, and what would that change in your gross margin?
- Is your company spending its AI effort on cost reduction, such as cutting support or engineering hours, or on building a new product that could multiply revenue?
- What tasks could an assistant in your product do proactively on a user's behalf, without being asked, and what would make users trust it?
Bottom Line
David George argues that AI is early, supply is sold out, and revenue rests on a small group of users, so many layers of the stack can win together. For PMs the lesson is to build where labs will not go, meaning the details, delivery and go-to-market, and to aim AI efforts at growing revenue rather than only cutting cost.
Case Studies Mentioned
Harvey and AI in legal
David's firm is a large investor in Harvey. He says legal looks about 12 months behind coding in diffusion and is now at the point where end clients demand their law firms use Harvey. Early worries about hallucinations are, in his words, 18 months old. Harvey made a bet on what happens when models get 100x better. He also notes Kirkland said it would spend $500 million to build its own technology stack, which some read as a threat to Harvey, but he sees the market as large enough anyway. Lawyering is reasoning, and the data (case law) is well documented, which fits what reasoning models do well.
Databricks and founder expansion
The first growth fund investment, at a roughly $6 billion round, was in Databricks. David says seven years later revenue has accelerated as CEO Ali Ghodsi found new products, and the company is positioned to benefit from AI. His takeaway: founders who keep finding the next thing are the asset class.
Palantir and the value of narrative
David uses Palantir to show how story affects results. He says many holders cannot say what ontology means, yet the company has a cult following and a high valuation, and CEOs now see it as a trusted place to implement AI. A high valuation helps with raising capital, hiring and keeping employees, and he says a messianic founder with a product people love is a big asset.
Tools & Products
| Tool / Product | What it does | Why it was mentioned |
|---|---|---|
| Claude Code, Codex, Cursor | AI coding products | Examples of first-party, harness-tied products that make users prefer frontier models. |
| Replit and Lovable | AI app building tools | David's example of apps that thrive even in the center of the lab's coding territory. |
| Harvey | AI product for legal work | His example of a vertical app growing fast, with clients now requiring law firms to use it. |
| Waymo and Tesla | Self-driving rides and cars | Examples used to show how good autonomy already is and how little it has spread. |
People to Follow
David George
General partner at a16z who leads its growth investing. He discusses AI diffusion, the product and capital cycle, and why founders are the key asset in late-stage investing.
Jack Altman
Host of Uncapped and a venture investor at Benchmark, which recently raised its first growth fund. He pushes on the bear cases in the conversation.
Notable Quotes
"If the premise of your question is, is this going to be successful or that going to be successful, the answer in AI is probably and." (David George)
"It's like shorting your own future if you are just focused on the cost side." (David George)
"Believe the same thing everybody else believes but believe it ten times more." (David George, on a line he first heard elsewhere)
