Context

Investor Gavin Baker joins a16z's David George to look at the economics of the AI boom. Baker spent the summer trying to build a bearish case and could not: his standard question to AI leaders, "can you tell me one quantitative data point in your business that's getting worse?", kept returning no. Much of the episode is investor-and-macro material (compute paybacks, orbital data centers, Nvidia's financing moats) that is peripheral to product work. But threaded through it are strong, transferable ideas for PMs: why the AI market may be positive-sum rather than winner-take-all, why demand is far earlier than the revenue suggests, how enterprises will "own their intelligence," and why the most valuable and hardest position to win is becoming the abstraction layer to intelligence.

The Big Idea

The AI market is more likely positive-sum ("an and thing, not an or thing") than winner-take-all, and demand is far earlier than the headline revenue implies, so the decisive product battle is over who becomes the "abstraction layer to intelligence" for an organization, a position that goes to whoever executes best and is lowest-cost, not whoever has the grandest vision.

Baker's frame against the constant "how does this all crash?" question is that frontier labs, N-minus-one models, open source, applications, clouds, and chip companies can all win at once, with the fight concentrated on who mediates intelligence for users.

Key Insights

It is an "and," not an "or"

Baker's central reframe against zero-sum thinking. Every investor conversation starts with "how does this all go wrong, which part crashes?" His answer: the premise is wrong.

• The claim: frontier models, N-minus-one models, open source, application companies, the clouds, the handful of lab companies, and Nvidia at the center can all do well simultaneously. It is not "are the labs screwed by open source or not"; it is all of them working.
• Why it matters for PMs: in a genuinely expanding market, framing every player as a threat to every other leads to bad strategy. Ask where the market is positive-sum (a rising tide) versus truly zero-sum (a fixed pie) before assuming a competitor's win is your loss.

Seek the disconfirming data point

The best transferable habit in the episode is Baker's discipline of actively hunting for evidence against his own view. He went around all summer asking everyone for a single quantitative data point in their business that was getting worse, specifically to make his own sentiment more negative, and could not find one. The method generalizes: instead of collecting evidence that confirms your thesis, go looking for the one concrete metric that would falsify it. If you cannot find it after asking everyone, that is strong signal; if you can, you have learned something early.

Demand is far earlier than it looks

A crucial calibration for anyone sizing AI adoption. Roughly $180B of revenue is being generated on the back of maybe 30 million heavy paying users, and Baker thinks the real number of genuinely heavy users may be under 10 million, against about 1.5 billion knowledge workers.

• The power law: within even AI-forward companies, the highest-spending engineers spend 10x to 100x the median engineer, so headline user counts overstate real diffusion.
• The trajectory: Baker's own firm's internal token consumption rose 100x from March to August, and looked set to jump another 10-20x after adopting one new agent tool, and he stresses this is valuable, not wasteful, usage.
• The takeaway: on a 10-year view, "we're nowhere" on diffusion. Do not mistake today's concentrated, early adoption for the ceiling; plan for demand that spreads from a tiny power-user core to the broad workforce.

Prices could rise, not fall

A counterintuitive point worth holding. Everyone assumes the cost of intelligence falls monotonically, but in a supply-constrained world where demand outruns compute, the price of a token could actually go up (one guest floated 10x).

• Why: there is a large consumer surplus. People choose expensive frontier tokens even when cheaper ones would do most tasks, because the value they get far exceeds the price. If supply stays constrained, that surplus supports higher prices.
• For PMs: do not build a business model that assumes inference costs only ever fall. In a supply crunch, your input costs could rise, and access to compute (not just model quality) becomes a competitive variable.

Reactive AI is giving way to proactive

Baker distinguishes two generations of AI use. The first (what he built with coding tools: podcast summarizers, sentiment trackers) is reactive and knowledge-enhancing: it prepares information but does not do the job. The emerging shift is proactive: an agent that looks at everything you do and returns recommended actions and automations ("what are the recommended actions based on everything the bots learned today?"), which you can then approve to run. He calls the ease of building the proactive version another "ChatGPT moment," and notes it points to effectively endless token demand. For PMs, the frontier is moving from surfacing information to proposing and taking action, and that is where the next wave of value (and usage) sits.

Own your intelligence

A concrete enterprise pattern Baker expects to dominate. Rather than share your proprietary enterprise context (your real IP) with a frontier lab, you take a strong open-source base model, do reinforcement learning and supervised fine-tuning on your own data, and own the resulting model, its capabilities, and its cost.

• The architecture: that owned model runs behind a router alongside one or two frontier models that check each other, transparently to the user, with the most capable model used for planning and cheaper models for execution.
• Why it matters: if intelligence is a core input to your business, you want to control it rather than rent it and expose your data. Baker points to products like Fireworks Nexus productizing exactly this (pick a base model, have it fine-tuned on your data, serve it behind a router).

Verifiable, documented domains mature first

Why coding and legal are taking off ahead of the broad knowledge-work market. Coding is unique in being verifiable and perfectly documented, and legal is similar (well-documented and somewhat verifiable), which is why AI penetrates them first. The vast 1.5-billion-knowledge-worker opportunity is "very messy" to capture because most of it is neither verifiable nor documented. The lesson for prioritizing AI use cases: favor domains where success is checkable and the source material is well-documented, because those are where AI works reliably first; treat the messy, unverifiable domains as harder and later.

Cursor won by being product-first

A pointed contrast in philosophy. Most players at the frontier framed themselves as building "a digital deity," an AGI or ASI. Cursor "just wanted to make great product," meeting customers where they are and the technology where it is, then legging its way up into more autonomy from that practical starting point. Baker's read is that being the most product-focused was itself an edge, and that the same practical, engineering-problem mindset suits how the best operators work. For PMs: a concrete, meet-users-where-they-are product can out-execute grand-vision competitors and still reach the ambitious end state, just by a more practical path.

Mental Models & Frameworks

The abstraction layer to intelligence

The single most important product concept in the episode. The prize is being the layer that mediates intelligence for an organization and its users, "who gets to be the abstraction layer to the organization and the users of intelligence," which Baker calls maybe the most vied-after position in the history of business.

• Why it is hard: he compares it to running a national retail chain, which sounds easy (just stock 1,000 stores across 50 states with the right products, staffed by friendly people who do not steal, at 100% turnover) and is brutally hard to execute. Being the seamless, continuously-updated, trustworthy intelligence layer is similarly deceptive.
• Who is fighting for it: the labs, Microsoft, Databricks, Palantir, Snowflake, Salesforce, Workday, the inference providers, and vertical agents like Harvey and Legora, all colliding for the same position.
• How to use it: ask whether your product is trying to be an abstraction layer for intelligence, and if so, what makes you the lowest-cost, best-executing, most-trusted option, because that is what the position ultimately comes down to.

Vertically integrated but horizontally open

Baker's read of Nvidia's strategy, and a general model for an ecosystem leader: be deeply integrated in your own stack while remaining open to work with everyone, including competitors. His practical corollary for anyone building near a dominant platform: do not go head-on (every 1% of share can be worth ~$100B, so pick a niche and take your 1%), and be nice to the ecosystem leader rather than taking shots (the extended Michael-Jordan-on-the-court analogy). The reason is concrete: the leader's offering is the most financeable, which lowers its customers' cost of capital, a real moat that is hard to beat by antagonizing it.

Inferring preferences under constraint

A sharp signal-extraction idea. When everything is supply-constrained, you cannot read true customer preferences from sales, because customers will take whatever they can get. Baker's move is to read preference from the structure of the deals instead (who invests in whom, who takes a residual-value guarantee, who gives warrants), because the deal terms reveal what each side actually believes. The general lesson: when demand is so hot that usage no longer discriminates, look at the structure of commitments people make, not the raw volume, to infer real preference.

Decision Principles

Principle: Go looking for the bear case

• When: you are forming or holding a strong thesis about a market or product bet.
• Why: confirmation is cheap; disconfirmation is informative. Baker deliberately sought the one metric getting worse to talk himself out of his own optimism. Adopt the habit of asking, of yourself and everyone you talk to, "what is the single data point that would prove me wrong?" and weigh how hard it is to find.

Principle: Don't hand your context to a lab

• When: deciding how to build AI into a product where your proprietary data is the real value.
• Why: your enterprise context is your IP. Sharing it with a frontier lab to get better answers can be hazardous to your competitive and financial health, especially as data-retention policies shift. The alternative is to own the model: fine-tune a capable open-source base on your data and serve it behind a router with frontier models.

Practical Application

Ask for the disconfirming metric

Before committing to a product or market bet, explicitly ask what single quantitative data point would show it is getting worse, and go hunt for it across everyone you can talk to. Treat "I asked everyone and could not find one" as real evidence, and treat finding one as an early warning worth acting on.

Size demand from the power-user core

When estimating adoption or TAM for an AI feature, do not read headline user or revenue numbers as the ceiling. Recognize the power law (top users spend 10-100x the median) and that heavy real usage may be a tiny fraction of the eventual base. Plan for diffusion from a small power-user core outward, and watch your own internal token growth as a leading indicator.

Decide whether to own your intelligence

If intelligence is a core input to your product and your data is your edge, evaluate owning your model (fine-tuning a strong open-source base on your own data behind a router) instead of routing everything to a frontier lab. Weigh the control over capability, cost, and data against the effort, and consider a router that uses the most capable model for planning and cheaper models for execution.

Sequence use cases by verifiability

Prioritize AI use cases where success is checkable and the source material is well-documented (like code and, to a degree, legal), because those work reliably first. Treat unverifiable, poorly-documented workflows as harder and later, and do not assume the messy broad market will fall as easily as the verifiable niches did.

Questions to Consider

• Are we framing our market as winner-take-all (an "or") when it may actually be positive-sum (an "and"), and are we treating competitors' wins as our losses when the pie is expanding?
• What is the single quantitative data point that would prove our current thesis wrong, and have we actually gone looking for it rather than collecting confirming evidence?
• Are we sizing demand for our AI feature off headline user or revenue numbers, when heavy real usage may be a tiny power-user fraction of the eventual base?
• For the workflows we are automating, is success verifiable and the source material well-documented (so AI will work reliably now), or messy and unverifiable (so it is a harder, later bet)?
• If our data is our real IP, are we handing that context to a frontier lab, or should we own our intelligence by fine-tuning an open-source model on our own data behind a router?

Bottom Line

Baker's case is that the AI market is positive-sum ("an and thing, not an or thing") and demand is far earlier than the revenue suggests, so the decisive contest is over who becomes the abstraction layer to intelligence for organizations, a position won by execution and low cost, not vision. The transferable moves for builders: hunt for the data point that would prove you wrong, size demand from the power-user core outward, sequence use cases by how verifiable they are, own your intelligence when your data is your edge, and, near a dominant platform, take a niche rather than going head-on.

Notable Quotes

"This is not an or thing. It's an and thing." (Gavin Baker, on who wins in AI)

"The truth shall set you free, but only if you tell it." (Gavin Baker, on the AI industry telling concrete benefit stories)