Sam Altman · OpenAI · 2025-02-09
A short, dense piece where Altman lays out three quantitative observations about AI economics that drive everything else OpenAI does. Worth reading not for predictions but for the unit economics — the claims here are the operating model behind every billion-dollar training run.
1. Intelligence scales with the log of compute.
> The intelligence of an AI model roughly equals the log of the resources used to train and run it.
Continuous, predictable gains from spending more on training compute, training data, and inference compute. Not a curve that ever bends sharply — but reliably climbs.
2. Per-token cost drops ~10× per year.
GPT-4 token prices fell ~150× between early 2023 and mid-2024. That's substantially faster than Moore's Law, which historically delivered ~2× every 18-24 months. If this rate holds, intelligence becomes effectively free for most uses within a decade.
3. The socioeconomic value of linear intelligence gains is super-exponential.
A model that's 2× smarter isn't worth 2× as much — it's worth dramatically more, because the tasks it unlocks have non-linear payoffs. This justifies, in his frame, exponential investment increases. (Whether the value curve actually behaves this way at the high end is the unsettled empirical question.)
- Prices for many goods drop dramatically as the cost of intelligence approaches zero - Luxury goods and scarce resources (especially land) become more expensive - Scientific progress accelerates - New job categories emerge; the categories visible today will look as quaint as 1995 web jobs do now
If you only read one of Altman's blog posts, this is more useful than the inspirational ones. The three claims are testable; you can check claim #2 against price-history data anytime, you can argue with claim #3, and you'll have something concrete to disagree with rather than vibes.